fix(cli): enter setup on a requires-model config and hint zero auth login

Two residual first-run gaps after the catalog model-defaulting fix:

- An active provider whose model has NO catalog default to fall back on
  (custom openai-/anthropic-compatible endpoints) still bricked bare
  `zero` and `zero setup` — the exact commands that could have repaired
  the config. Tag that resolve failure with ErrProviderRequiresModel and
  extend the interactive TUI's fixable-error fallback (previously only
  ErrNoActiveProvider) to drop into the setup wizard for it. The sentinel
  is attached via a message-preserving wrapper (multi-target Unwrap), so
  headless commands (zero config / zero exec) print the exact actionable
  text unchanged and every other resolve error stays fatal.

- `zero login` / `zero logout` don't exist (it's `zero auth login`);
  first-run users try them and got only "unknown command" (reported in
  the wild). The unknown-command error now answers: did you mean
  "zero auth login"? Unrelated unknown commands get no suggestion.

Tests: the sentinel comes from a REAL Resolve over a real config file
(message-prefix asserted, no key leak); the TUI fallback is red-green
(reverting the condition reproduces the exact brick with the old error);
the existing no-active-provider fallback and other-errors-stay-fatal
tests still pass; login/logout hint with a no-false-positive check.

.github/workflows/release-artifacts.yml

@@ -54,8 +54,10 @@ jobs:
       # is published under the pushed git tag. If they disagree, install.sh/install.ps1
       # and the npm postinstall all build a download URL that 404s. Fail fast so a
       # forgotten version bump can never publish mis-versioned, un-installable assets.
+      # Matches on the ref (not event_name) because release-please dispatches this
+      # workflow at the tag ref — its GITHUB_TOKEN-created tag cannot fire `push`.
       - name: Check tag matches package.json version
-        if: ${{ github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v') }}
+        if: ${{ startsWith(github.ref, 'refs/tags/v') }}
         shell: bash
         run: |
           pkg_version="$(node -p "require('./package.json').version")"
@@ -78,13 +80,15 @@ jobs:
           path: dist/release/*
           if-no-files-found: error
 
-      # On a `v*` tag push, publish the packaged archive + .sha256 as GitHub Release
-      # assets so the documented `scripts/install.sh` / `install.ps1` path (which reads
-      # /releases/latest) works. Uses the preinstalled `gh` (no new action dependency);
-      # each matrix OS creates the release if absent (best-effort) then uploads its own
-      # assets with --clobber. NOTE: maintainers must push the tag — this never tags.
+      # On a `v*` tag ref (pushed manually, or dispatched at the tag by the
+      # release-please workflow), publish the packaged archive + .sha256 as GitHub
+      # Release assets so the documented `scripts/install.sh` / `install.ps1` path
+      # (which reads /releases/latest) works. Uses the preinstalled `gh` (no new
+      # action dependency); each matrix OS creates the release if absent
+      # (best-effort — release-please normally creates it first) then uploads its
+      # own assets with --clobber. Tags come from merging the release-please PR.
       - name: Publish to GitHub Release
-        if: ${{ github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v') }}
+        if: ${{ startsWith(github.ref, 'refs/tags/v') }}
         shell: bash
         env:
           GH_TOKEN: ${{ github.token }}
@@ -92,3 +96,89 @@ jobs:
           tag="${GITHUB_REF_NAME}"
           gh release create "$tag" --repo "$GITHUB_REPOSITORY" --title "$tag" --generate-notes 2>/dev/null || true
           gh release upload "$tag" dist/release/* --repo "$GITHUB_REPOSITORY" --clobber
+
+  # Publishes the npm wrapper (@gitlawb/zero) once every platform's release
+  # assets are live — the package's postinstall downloads its binary from the
+  # GitHub Release, so publishing before the assets exist would ship a package
+  # that fails to install.
+  #
+  # Auth is npm OIDC trusted publishing: no NPM_TOKEN secret. The package's
+  # npmjs.com settings must list this repo + workflow file as a trusted
+  # publisher, and provenance is then generated automatically. Note npm only
+  # allows configuring a trusted publisher on an EXISTING package, so the very
+  # first @gitlawb/zero publish is a one-time manual `npm publish` bootstrap.
+  publish-npm:
+    name: Publish npm package
+    needs: package
+    if: ${{ startsWith(github.ref, 'refs/tags/v') }}
+    runs-on: ubuntu-latest
+    # Manual approval gate: this environment carries a required-reviewer rule
+    # (Settings -> Environments -> npm-publish), so the job pauses until a
+    # maintainer approves it in the Actions UI before anything reaches npm.
+    environment:
+      name: npm-publish
+      url: https://www.npmjs.com/package/@gitlawb/zero
+    permissions:
+      contents: read
+      id-token: write # OIDC token exchange for npm trusted publishing + provenance
+    steps:
+      - name: Checkout
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
+
+      # No registry-url here on purpose: setup-node's registry-url writes an
+      # .npmrc line referencing $NODE_AUTH_TOKEN, and npm hard-fails on the
+      # unset env var. Trusted publishing needs no token at all.
+      - name: Setup Node
+        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4
+        with:
+          node-version: 24
+
+      # OIDC trusted publishing needs npm >= 11.5.1. Use the npm bundled with
+      # the pinned Node toolchain (Node 24 ships >= 11.5.1) rather than pulling
+      # npm@latest from the registry inside the publish path — an unpinned
+      # install there would be avoidable supply-chain exposure. Fail fast if a
+      # future runner/toolchain change ever regresses the bundled npm.
+      - name: Check npm supports trusted publishing
+        run: |
+          npm --version | node -e '
+            const version = require("fs").readFileSync(0, "utf8").trim();
+            const [major, minor, patch] = version.split(".").map(Number);
+            const ok =
+              major > 11 ||
+              (major === 11 && (minor > 5 || (minor === 5 && patch >= 1)));
+            if (!ok) {
+              console.error(
+                `::error::bundled npm ${version} is too old for OIDC trusted ` +
+                  `publishing (needs >= 11.5.1) — bump node-version in this job.`,
+              );
+              process.exit(1);
+            }
+            console.log(`npm ${version} supports trusted publishing.`);
+          '
+
+      # Probe the exact URLs postinstall will build (internal/release/release.go
+      # asset-name scheme) as an anonymous client. This also blocks publishing
+      # while the repo is still private — private release assets 404 without
+      # auth, exactly as they would for an installing user.
+      - name: Verify release assets are downloadable
+        shell: bash
+        run: |
+          version="$(node -p "require('./package.json').version")"
+          base="https://github.com/${GITHUB_REPOSITORY}/releases/download/v${version}"
+          for asset in \
+            "zero-v${version}-linux-x64.tar.gz" \
+            "zero-v${version}-linux-arm64.tar.gz" \
+            "zero-v${version}-macos-x64.tar.gz" \
+            "zero-v${version}-macos-arm64.tar.gz" \
+            "zero-v${version}-windows-x64.zip"; do
+            for url in "${base}/${asset}" "${base}/${asset}.sha256"; do
+              if ! curl -fsSLI -o /dev/null "$url"; then
+                echo "::error::${url} is not publicly downloadable — npm postinstall would fail. Is the repo public and are all assets uploaded?" >&2
+                exit 1
+              fi
+            done
+          done
+          echo "all release assets for v${version} are publicly downloadable."
+
+      - name: Publish to npm
+        run: npm publish --access public

.github/workflows/release-please.yml

@@ -0,0 +1,35 @@
+name: Release Please
+
+on:
+  push:
+    branches:
+      - main
+
+permissions:
+  contents: write # push the release-PR branch, create the tag + GitHub Release
+  pull-requests: write # open/update the release PR
+  issues: write # create/apply release-tracking labels on the release PR
+  actions: write # dispatch release-artifacts.yml once a release is cut
+
+jobs:
+  release-please:
+    name: Release PR / tag
+    runs-on: ubuntu-latest
+    steps:
+      - name: Run release-please
+        id: release
+        uses: googleapis/release-please-action@a02a34c4d625f9be7cb89156071d8567266a2445 # v4.2.0
+
+      # Tags created with the default GITHUB_TOKEN never fire `on: push: tags`
+      # workflows (GitHub suppresses workflows triggered by workflow-created
+      # events), so release-artifacts.yml would not run for a release-please
+      # tag. workflow_dispatch invoked through the API is exempt from that
+      # suppression — dispatch the artifacts build at the new tag explicitly.
+      - name: Build and publish release artifacts
+        if: ${{ steps.release.outputs.release_created }}
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          gh workflow run release-artifacts.yml \
+            --repo "$GITHUB_REPOSITORY" \
+            --ref "${{ steps.release.outputs.tag_name }}"

.release-please-manifest.json

@@ -0,0 +1 @@
+{".":"0.1.0"}

CHANGELOG.md

@@ -6,12 +6,32 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 aims to follow [Semantic Versioning](https://semver.org/spec/v2.0.0.html) once the first release is
 tagged. Until then, source builds report the version `dev`.
 
+## 0.1.0 (2026-07-02)
+
+
+### Features
+
+* publish zero to npm via release-please ([#367](https://github.com/Gitlawb/zero/issues/367)) ([8eccc26](https://github.com/Gitlawb/zero/commit/8eccc2669887bc38d35bc16a315c888e4d9ec43a))
+* **tui:** FILES sidebar panel with click-to-select and file drill-in ([#365](https://github.com/Gitlawb/zero/issues/365)) ([142c548](https://github.com/Gitlawb/zero/commit/142c548c89a8652ce300e64ddf1228ee36df7606))
+
+
+### Bug Fixes
+
+* **auth:** propagate credentials to every provider-build surface and pin children to the live provider ([#366](https://github.com/Gitlawb/zero/issues/366)) ([6e0a665](https://github.com/Gitlawb/zero/commit/6e0a665118fe0e09c4b07d482dd18f86045acd2b))
+
 ## [Unreleased]
 
 ### Added
 - `SECURITY.md` with a private vulnerability-reporting path, `CODE_OF_CONDUCT.md`, this changelog, and
   GitHub issue/PR templates.
-- `--theme {auto|dark|light}` flag for the TUI (previously only the `ZERO_THEME` env var existed).
+- Interactive `/theme` picker: bare `/theme` opens a popup that live-previews each palette as you move
+  and applies on select (Esc reverts).
+- Ten built-in color themes alongside the `dark`/`light` built-ins — `dracula`, `nord`, `gruvbox`,
+  `tokyo-night`, `catppuccin`, `one-dark`, `solarized-dark`, `rose-pine`, `everforest`, and
+  `solarized-light` — selectable via `/theme <name>`, `--theme <name>`, or `ZERO_THEME`. Every palette
+  is contrast-audited to WCAG AA. The built-in light theme was reworked for legibility.
+- `--theme <name>` flag for the TUI, accepting `auto` or any registered theme (previously only the
+  `ZERO_THEME` env var existed).
 - "Accessibility / Appearance" section in the README documenting `NO_COLOR`, `ZERO_THEME`, `/theme`,
   and `ZERO_NO_FADE`.
 

README.md

@@ -254,9 +254,9 @@ zero update           check for newer releases
 | Control | Effect |
 |---|---|
 | `NO_COLOR=<anything>` | disables color output |
-| `ZERO_THEME=auto|dark|light` | selects the startup theme |
-| `--theme auto|dark|light` | selects the TUI theme from the CLI |
-| `/theme auto|dark|light` | switches theme inside the TUI |
+| `ZERO_THEME=<name>` | selects the startup theme (`auto`, `dark`, `light`, or a color theme like `dracula`, `nord`, `gruvbox`, `tokyo-night`, `catppuccin`, `one-dark`, `solarized-dark`, `rose-pine`, `everforest`, `solarized-light`) |
+| `--theme <name>` | selects the TUI theme from the CLI (same names) |
+| `/theme` | opens the theme picker inside the TUI (live preview; `/theme <name>` switches directly) |
 | `ZERO_NO_FADE=1` | disables streaming fade animation |
 
 Meaning does not rely on color alone; diffs, permissions, and statuses also use

internal/agent/compaction.go

@@ -280,6 +280,25 @@ func isContextLimitError(message string) bool {
 	return false
 }
 
+// isStreamTimeoutError reports whether a streamed error is a provider stream
+// idle/stall timeout (providerio surfaces these as "idle timeout after …" /
+// "no output for …" / "stream stalled"). Such a turn produced nothing, so when
+// no content was forwarded yet it can be safely re-issued on a fresh connection
+// (see the stall-retry in the agent loop). Substring + case-insensitive to
+// tolerate the prefixed "provider stream error: …" wrapping.
+func isStreamTimeoutError(message string) bool {
+	lowered := strings.ToLower(strings.TrimSpace(message))
+	if lowered == "" {
+		return false
+	}
+	for _, needle := range []string{"idle timeout after", "no output for", "stream stalled"} {
+		if strings.Contains(lowered, needle) {
+			return true
+		}
+	}
+	return false
+}
+
 // compactionState carries the per-run state the agent loop needs to compact a
 // conversation safely. It is created once per Run and is a no-op whenever
 // options.ContextWindow <= 0.

internal/agent/decode_args_test.go

@@ -0,0 +1,99 @@
+package agent
+
+import "testing"
+
+// decodeToolArguments must (a) decode a normal single-object payload unchanged,
+// (b) tolerate a weak model packing multiple concatenated top-level JSON objects
+// into one arguments string by decoding the FIRST object and ignoring the rest
+// (the minimax-m3 "invalid character '{' after top-level value" failure), and
+// (c) still reject a genuinely malformed/truncated first object.
+func TestDecodeToolArguments(t *testing.T) {
+	t.Run("single object decodes normally", func(t *testing.T) {
+		var a map[string]any
+		if err := decodeToolArguments(`{"path":"x"}`, &a); err != nil {
+			t.Fatalf("err = %v, want nil", err)
+		}
+		if a["path"] != "x" {
+			t.Fatalf("decoded %v, want path=x", a)
+		}
+	})
+
+	t.Run("concatenated objects decode the first, ignore trailing", func(t *testing.T) {
+		var b map[string]any
+		if err := decodeToolArguments(`{"path":"first"}{"path":"second"}`, &b); err != nil {
+			t.Fatalf("multi-object should decode the first, got err = %v", err)
+		}
+		if b["path"] != "first" {
+			t.Fatalf("decoded %v, want the first object (path=first)", b)
+		}
+	})
+
+	t.Run("distinct tools' args concatenated (real minimax case): first object used", func(t *testing.T) {
+		var c map[string]any
+		if err := decodeToolArguments(`{"plan":[1,2]}{"cmd":"go version"}`, &c); err != nil {
+			t.Fatalf("err = %v, want nil", err)
+		}
+		if _, ok := c["plan"]; !ok || c["cmd"] != nil {
+			t.Fatalf("decoded %v, want only the first object's keys", c)
+		}
+	})
+
+	t.Run("truncated first object still errors", func(t *testing.T) {
+		var d map[string]any
+		if err := decodeToolArguments(`{"path":`, &d); err == nil {
+			t.Fatal("truncated JSON must still error")
+		}
+	})
+
+	t.Run("non-JSON trailing garbage still errors (not masked)", func(t *testing.T) {
+		var g map[string]any
+		if err := decodeToolArguments(`{"x":1}xyz`, &g); err == nil {
+			t.Fatal("a valid first object followed by non-JSON garbage must still error")
+		}
+	})
+
+	t.Run("recoverableToolArguments: only whole-JSON trailing is tolerated", func(t *testing.T) {
+		cases := []struct {
+			in    string
+			ok    bool
+			first string
+		}{
+			{`{"a":1}`, true, `{"a":1}`},
+			{`{"a":1}{"b":2}`, true, `{"a":1}`},
+			{`{"a":1}  {"b":2}  `, true, `{"a":1}`},
+			{`{"a":1}xyz`, false, ""},
+			{`{"a":`, false, ""},
+			{`   `, false, ""},
+		}
+		for _, c := range cases {
+			first, ok := recoverableToolArguments(c.in)
+			if ok != c.ok || (ok && first != c.first) {
+				t.Fatalf("recoverableToolArguments(%q) = (%q,%v), want (%q,%v)", c.in, first, ok, c.first, c.ok)
+			}
+		}
+	})
+
+	t.Run("historySafeToolCalls recovers the first object, not {}", func(t *testing.T) {
+		out := historySafeToolCalls([]ToolCall{
+			{ID: "c1", Name: "read_file", Arguments: `{"path":"README.md"}{"path":"AGENTS.md"}`}, // concatenated -> first
+			{ID: "c2", Name: "read_file", Arguments: `{"path":"x"}xyz`},                          // garbage -> {}
+			{ID: "c3", Name: "read_file", Arguments: `{"path":"ok"}`},                            // valid -> unchanged
+		})
+		if out[0].Arguments != `{"path":"README.md"}` {
+			t.Fatalf("concatenated args should record the first object, got %q", out[0].Arguments)
+		}
+		if out[1].Arguments != "{}" {
+			t.Fatalf("garbage-trailing args should record {}, got %q", out[1].Arguments)
+		}
+		if out[2].Arguments != `{"path":"ok"}` {
+			t.Fatalf("valid args should be unchanged, got %q", out[2].Arguments)
+		}
+	})
+
+	t.Run("whitespace-only errors (caller guards empty)", func(t *testing.T) {
+		var e map[string]any
+		if err := decodeToolArguments("   ", &e); err == nil {
+			t.Fatal("whitespace-only must error")
+		}
+	})
+}

internal/agent/loop.go

@@ -5,6 +5,7 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"io"
 	"sort"
 	"strings"
 	"sync"
@@ -20,6 +21,12 @@ import (
 const maxTurnsAnswer = "Agent reached maximum number of turns without a final answer."
 const maxTurnsFinalAnswerPrompt = "You have reached the tool-turn limit. Do not call tools. Give a concise final answer now: summarize what you completed, what you found, and any remaining blockers."
 
+// maxStreamStallRetries bounds how many times a turn that timed out (idle/stall)
+// WITH NO OUTPUT yet is re-issued on a fresh connection before giving up. Only
+// the no-output case is retried (a partial turn would duplicate), so this is a
+// safe recovery for a stalled/dead pooled connection.
+const maxStreamStallRetries = 2
+
 const (
 	toolResultMetaControl       = "control"
 	toolResultControlSpecReview = "spec_review_required"
@@ -203,55 +210,74 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			}
 		}
 
-		collected := zeroruntime.CollectStreamWithOptions(ctx, stream, zeroruntime.CollectOptions{
-			OnText:          options.OnText,
-			OnReasoning:     options.OnReasoning,
-			OnUsage:         options.OnUsage,
-			OnToolCallStart: options.OnToolCallStart,
-			OnToolCallDelta: options.OnToolCallDelta,
-		})
-		if collected.Error != "" {
-			collectedErr := errors.New(collected.Error)
-			if isImageRejectionError(collectedErr) {
-				result.Messages = copyMessages(messages)
-				return result, fmt.Errorf("model %s rejected the image: %s. The model may not support image input — try switching to a vision-capable model (claude, gpt-4o, gemini)", options.Model, collected.Error)
+		// forwardedVisibleText flags forwarded final PROSE (OnText) specifically —
+		// the only output whose re-stream on a stall retry would visibly DUPLICATE
+		// answer text the user already read. Transient working indicators
+		// (reasoning previews, tool-call "writing…" previews) are deliberately NOT
+		// counted: a turn that streamed only those then stalled — the common
+		// gpt-5.x / ollama "froze mid-write_file" case — is safe to re-issue, since
+		// the TUI resets the tool-call preview on the next tool-call-start and a
+		// stalled turn is never appended to `messages` (it returns on the error
+		// before the append), so the retry re-sends clean context with no
+		// conversation-state duplication.
+		forwardedVisibleText := false
+		forwardingOpts := zeroruntime.CollectOptions{OnUsage: options.OnUsage}
+		if options.OnText != nil {
+			forwardingOpts.OnText = func(s string) { forwardedVisibleText = true; options.OnText(s) }
+		}
+		if options.OnReasoning != nil {
+			forwardingOpts.OnReasoning = func(s string) { options.OnReasoning(s) }
+		}
+		if options.OnToolCallStart != nil {
+			forwardingOpts.OnToolCallStart = func(id, name string) { options.OnToolCallStart(id, name) }
+		}
+		if options.OnToolCallDelta != nil {
+			forwardingOpts.OnToolCallDelta = func(id, fragment string) { options.OnToolCallDelta(id, fragment) }
+		}
+		// recoverStreamError applies the same non-stall recovery the initial stream
+		// gets to ANY collected error — including one from a reissued (stall-retry)
+		// stream: an image rejection gets the friendly wrapping, and a context limit
+		// gets one compaction + reactive reissue (omitting visible callbacks, since
+		// any pre-error output was already forwarded). It returns the possibly-updated
+		// collected and a non-nil stop error when the run must end now.
+		recoverStreamError := func(collected zeroruntime.CollectedStream) (zeroruntime.CollectedStream, error) {
+			if isImageRejectionError(errors.New(collected.Error)) {
+				return collected, fmt.Errorf("model %s rejected the image: %s. The model may not support image input — try switching to a vision-capable model (claude, gpt-4o, gemini)", options.Model, collected.Error)
 			}
-			// REACTIVE compaction: the streamed error may also be a context
-			// limit (some providers surface it mid-stream). Compact and retry
-			// the same turn once before giving up.
+			// REACTIVE compaction: the streamed error may also be a context limit
+			// (some providers surface it mid-stream). Compact and retry once.
 			if compacted, retried, retryErr := compactor.recover(ctx, provider, messages, request.Tools, collected.Error); retried {
 				messages = compacted
 				if retryErr != nil {
-					result.Messages = copyMessages(messages)
-					return result, retryErr
+					return collected, retryErr
 				}
-				// Reuse the SAME active-mode partition (exposed) from this
-				// turn rather than the bare toolDefinitions: exposed depends on
-				// registry+loaded (not the messages), so they stay valid after compaction.
-				// Routing through an empty-loaded partition here would re-hide every
-				// already-loaded deferred tool.
+				// Reuse the SAME active-mode partition (exposed) from this turn rather
+				// than the bare toolDefinitions: exposed depends on registry+loaded (not
+				// the messages), so it stays valid after compaction.
 				retryRequest := zeroruntime.CompletionRequest{
 					Messages:        copyMessages(messages),
 					Tools:           exposed,
 					ReasoningEffort: options.ReasoningEffort,
 				}
-				// Pre-content reconnect of a mid-stream disconnect: route the connect
-				// through the reconnect helper too (AUDIT-L1).
 				retryStream, retryStreamErr := streamWithReconnect(ctx, provider, retryRequest, reconnectNoticeFor(options))
 				if retryStreamErr != nil {
-					result.Messages = copyMessages(messages)
-					return result, retryStreamErr
+					return collected, retryStreamErr
 				}
-				// Omit OnText on the reactive retry: when the original error
-				// surfaced MID-stream, partial text was already forwarded to the
-				// user. Re-streaming the retried response on top of it would
-				// duplicate output. OnUsage IS kept so token telemetry/budgeting
-				// still counts the successful retry. The retried text is captured
-				// in collected.Text and becomes the turn's assistant message.
 				collected = zeroruntime.CollectStreamWithOptions(ctx, retryStream, zeroruntime.CollectOptions{
 					OnUsage: options.OnUsage,
 				})
 			}
+			return collected, nil
+		}
+
+		collected := zeroruntime.CollectStreamWithOptions(ctx, stream, forwardingOpts)
+		if collected.Error != "" {
+			updated, stop := recoverStreamError(collected)
+			collected = updated
+			if stop != nil {
+				result.Messages = copyMessages(messages)
+				return result, stop
+			}
 		}
 		// Check ctx first: on cancellation helpers.go sets collected.Error to
 		// ctx.Err().Error(), so returning errors.New(collected.Error) would lose
@@ -260,9 +286,58 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			result.Messages = copyMessages(messages)
 			return result, ctx.Err()
 		}
+		// A stream idle/stall timeout is safely re-issued when the turn committed NO
+		// answer text — no forwarded visible prose (forwardedVisibleText) and no
+		// collected final text (collected.Text). This covers two cases:
+		//   1. Nothing streamed at all before the connection died (the original
+		//      macOS stale-pooled-connection hang past the response-header timeout).
+		//   2. The model streamed transient reasoning and began a tool call (e.g. a
+		//      large write_file) then froze mid-arguments — the common gpt-5.x /
+		//      ollama heartbeat-pause stall. That partial tool call is NEVER executed
+		//      (a turn with collected.Error returns before dispatch below) and NEVER
+		//      appended to `messages`, so a retry re-issues clean context; the only
+		//      re-render is transient previews, not duplicated answer text. This is
+		//      why the gate no longer excludes collected.ToolCalls: an incomplete
+		//      tool call from a timed-out stream is discard-and-retry, not output.
+		// A turn that forwarded real prose is NOT retried (it would duplicate visible
+		// answer text) and falls through to the error return below. Capped +
+		// exponential backoff, with a user-visible notice per attempt.
+		for attempt := 1; attempt <= maxStreamStallRetries &&
+			isStreamTimeoutError(collected.Error) && !forwardedVisibleText &&
+			collected.Text == ""; attempt++ {
+			if notify := stallRetryNoticeFor(options); notify != nil {
+				notify(attempt, maxStreamStallRetries)
+			}
+			if err := sleepWithContext(ctx, backoffFor(attempt)); err != nil {
+				result.Messages = copyMessages(messages)
+				return result, err
+			}
+			retryRequest := zeroruntime.CompletionRequest{
+				Messages:        copyMessages(messages),
+				Tools:           exposed,
+				ReasoningEffort: options.ReasoningEffort,
+			}
+			retryStream, retryErr := streamWithReconnect(ctx, provider, retryRequest, reconnectNoticeFor(options))
+			if retryErr != nil {
+				result.Messages = copyMessages(messages)
+				return result, retryErr
+			}
+			collected = zeroruntime.CollectStreamWithOptions(ctx, retryStream, forwardingOpts)
+		}
 		if collected.Error != "" {
-			result.Messages = copyMessages(messages)
-			return result, errors.New(collected.Error)
+			// Route a reissued stream's non-stall error through the SAME recovery as
+			// the initial stream (image-rejection wrapping / context-limit compaction)
+			// rather than returning it raw.
+			updated, stop := recoverStreamError(collected)
+			collected = updated
+			if stop != nil {
+				result.Messages = copyMessages(messages)
+				return result, stop
+			}
+			if collected.Error != "" {
+				result.Messages = copyMessages(messages)
+				return result, errors.New(collected.Error)
+			}
 		}
 
 		// Carry the turn's terminal stop reason so a final answer cut off at the
@@ -615,8 +690,20 @@ func historySafeToolCalls(calls []ToolCall) []ToolCall {
 	for i, call := range calls {
 		safe[i] = call
 		args := strings.TrimSpace(call.Arguments)
-		if args == "" || !json.Valid([]byte(args)) {
+		switch {
+		case args == "":
 			safe[i].Arguments = "{}"
+		case json.Valid([]byte(args)):
+			// already a single valid JSON value — keep as-is
+		default:
+			// Recover the first object for the concatenated case so the replayed
+			// transcript matches what executeToolCall actually ran; otherwise (real
+			// corruption) fall back to an empty object.
+			if first, ok := recoverableToolArguments(args); ok {
+				safe[i].Arguments = first
+			} else {
+				safe[i].Arguments = "{}"
+			}
 		}
 	}
 	return safe
@@ -641,10 +728,53 @@ func toolSchemaJSON(registry *tools.Registry, name string) string {
 // is a non-nil ABORT error only when the call demands the whole run stop (a
 // canceled/timed-out ask_user prompt) rather than continuing — every ordinary
 // success or tool error returns a nil abort error.
+// recoverableToolArguments inspects a tool call's raw JSON arguments. It returns
+// the raw text of the FIRST JSON value when the payload is one value optionally
+// followed by additional WHOLE JSON values (the weak-model concatenated case, e.g.
+// `{A}{B}`). ok is false when the payload is genuinely malformed: a bad/truncated
+// first value, OR trailing NON-JSON garbage after a valid first value (e.g.
+// `{"x":1}xyz`) — that case must still error, not be silently accepted. Sharing
+// this between dispatch and replay-history keeps them consistent: only a cleanly
+// recoverable first object is ever used.
+func recoverableToolArguments(arguments string) (first string, ok bool) {
+	dec := json.NewDecoder(strings.NewReader(arguments))
+	var head json.RawMessage
+	if err := dec.Decode(&head); err != nil {
+		return "", false
+	}
+	// The remainder must be only whole JSON values (or nothing); any decode error
+	// other than EOF means trailing garbage — reject so corruption still surfaces.
+	for {
+		var rest json.RawMessage
+		if err := dec.Decode(&rest); err != nil {
+			if errors.Is(err, io.EOF) {
+				return strings.TrimSpace(string(head)), true
+			}
+			return "", false
+		}
+	}
+}
+
+// decodeToolArguments decodes a tool call's JSON arguments into v. It tolerates a
+// weaker model that packs MULTIPLE concatenated top-level JSON objects into one
+// arguments string (`{A}{B}`) — the cause of "invalid character '{' after top-level
+// value" failures with small models like minimax-m3 — by decoding the FIRST object
+// and ignoring the trailing WHOLE JSON values, so the primary intended call runs.
+// A genuinely malformed payload (bad/truncated first object, or non-JSON trailing
+// garbage) still returns the real parse error; a standard single-object payload
+// decodes exactly as before.
+func decodeToolArguments(arguments string, v any) error {
+	if first, ok := recoverableToolArguments(arguments); ok {
+		return json.Unmarshal([]byte(first), v)
+	}
+	// Not cleanly recoverable — surface the genuine parse error.
+	return json.Unmarshal([]byte(arguments), v)
+}
+
 func executeToolCall(ctx context.Context, registry *tools.Registry, call ToolCall, permissionMode PermissionMode, options Options) (ToolResult, error) {
 	args := map[string]any{}
 	if call.Arguments != "" {
-		if err := json.Unmarshal([]byte(call.Arguments), &args); err != nil {
+		if err := decodeToolArguments(call.Arguments, &args); err != nil {
 			return ToolResult{
 				ToolCallID: call.ID,
 				Name:       call.Name,

internal/agent/loop_test.go

@@ -1053,7 +1053,10 @@ func TestRunSanitizesMalformedToolCallArgumentsBeforeRetry(t *testing.T) {
 		turns: [][]zeroruntime.StreamEvent{
 			{
 				{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "call-1", ToolName: "read_file"},
-				{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-1", ArgumentsFragment: `{"path":"README.md"}{"path":"AGENTS.md"}`},
+				// Genuinely malformed (truncated) args: still error -> sanitize -> retry.
+				// (Concatenated multi-object args are now tolerated; see
+				// TestRunRecoversFirstObjectFromConcatenatedToolArgs.)
+				{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-1", ArgumentsFragment: `{"path":"README.md"`},
 				{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "call-1"},
 				{Type: zeroruntime.StreamEventDone},
 			},
@@ -1103,6 +1106,53 @@ func TestRunSanitizesMalformedToolCallArgumentsBeforeRetry(t *testing.T) {
 	}
 }
 
+// A weak model (e.g. minimax-m3) that packs MULTIPLE concatenated JSON objects into
+// one tool-call slot must RECOVER the first object and run the call, not fail with
+// "invalid character '{' after top-level value". The trailing object(s) are dropped
+// (the model re-issues them on a later turn if still needed).
+func TestRunRecoversFirstObjectFromConcatenatedToolArgs(t *testing.T) {
+	root := t.TempDir()
+	if err := os.WriteFile(filepath.Join(root, "README.md"), []byte("hello readme"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewReadFileTool(root))
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			{
+				{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "call-1", ToolName: "read_file"},
+				{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-1", ArgumentsFragment: `{"path":"README.md"}{"path":"AGENTS.md"}`},
+				{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "call-1"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+			{
+				{Type: zeroruntime.StreamEventText, Content: "done"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+		},
+	}
+
+	result, err := Run(context.Background(), "read", provider, Options{Registry: registry, MaxTurns: 2})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("final answer = %q, want done", result.FinalAnswer)
+	}
+	var toolResult string
+	for _, message := range provider.requests[1].Messages {
+		if message.Role == zeroruntime.MessageRoleTool {
+			toolResult = message.Content
+		}
+	}
+	if strings.Contains(toolResult, "Failed to parse arguments") {
+		t.Fatalf("concatenated args should recover the first object, not fail to parse: %q", toolResult)
+	}
+	if !strings.Contains(toolResult, "hello readme") {
+		t.Fatalf("expected README.md contents (the first object's path), got %q", toolResult)
+	}
+}
+
 func TestRunDefersSelfCorrectFeedbackUntilAfterToolBatch(t *testing.T) {
 	// A single assistant turn with two tool calls where the first mutates and
 	// self-correct fails must keep both tool_results contiguous, with the feedback

internal/agent/reconnect.go

@@ -40,6 +40,21 @@ func reconnectNoticeFor(options Options) reconnectNotifier {
 	}
 }
 
+// stallRetryNoticeFor builds a notifier for the loop's content-stall retry (a
+// stream that connected and produced only transient output, then went silent).
+// Distinct wording from reconnectNoticeFor's "connection lost" because the
+// connection was fine — the model stalled. Surfaced through OnReasoning, the
+// non-content channel, so it never folds into the answer text. Nil when there
+// is no reasoning sink (the retry still happens silently).
+func stallRetryNoticeFor(options Options) reconnectNotifier {
+	if options.OnReasoning == nil {
+		return nil
+	}
+	return func(attempt, max int) {
+		options.OnReasoning(fmt.Sprintf("\n[no output — model stalled; retrying %d/%d…]\n", attempt, max))
+	}
+}
+
 // streamWithReconnect issues request via provider.StreamCompletion and, on a
 // transient disconnect error, retries the connect up to maxStreamReconnects
 // times with exponential backoff. It returns the live stream on success, or the

internal/agent/stall_retry_test.go

@@ -0,0 +1,168 @@
+package agent
+
+import (
+	"context"
+	"sync/atomic"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// stallProvider connects successfully (HTTP 200) but the stream emits a stall/idle
+// timeout error for the first stallBefore calls, then succeeds with "done". The
+// pre-stall emission simulates different partial-output shapes:
+//   - partialText: forwarded PROSE (StreamEventText) — must NOT be retried
+//     (re-issuing would duplicate visible answer text).
+//   - reasoningText: transient reasoning preview — IS retried.
+//   - partialToolCall: a tool call started (+ arg delta) but never ended, the
+//     "froze mid-write_file" case — IS retried (the incomplete call is never
+//     executed or committed).
+type stallProvider struct {
+	calls           int32
+	stallBefore     int32
+	partialText     string
+	reasoningText   string
+	partialToolCall string
+}
+
+func (p *stallProvider) StreamCompletion(_ context.Context, _ zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
+	n := atomic.AddInt32(&p.calls, 1)
+	ch := make(chan zeroruntime.StreamEvent, 5)
+	if n <= p.stallBefore {
+		if p.partialText != "" {
+			ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventText, Content: p.partialText}
+		}
+		if p.reasoningText != "" {
+			ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventReasoning, Content: p.reasoningText}
+		}
+		if p.partialToolCall != "" {
+			// A tool call that starts and streams a partial argument fragment but
+			// never gets StreamEventToolCallEnd, then the stream errors — exactly
+			// the gpt-5.x "began the big write_file then froze" shape.
+			ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "tc_1", ToolName: p.partialToolCall}
+			ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "tc_1", ArgumentsFragment: `{"path":"x.html","content":"<!doctype`}
+		}
+		ch <- zeroruntime.StreamEvent{
+			Type:  zeroruntime.StreamEventError,
+			Error: "provider stream error: no output for 6m (the model produced nothing)",
+		}
+		close(ch)
+		return ch, nil
+	}
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventText, Content: "done"}
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone}
+	close(ch)
+	return ch, nil
+}
+
+// A no-output stall is re-issued on a fresh connection and recovers — this is the
+// macOS stale-pooled-connection hang turned into an automatic recovery.
+func TestRunRetriesStreamStallWithNoOutput(t *testing.T) {
+	p := &stallProvider{stallBefore: 1}
+	result, err := Run(context.Background(), "go", p, Options{Registry: tools.NewRegistry()})
+	if err != nil {
+		t.Fatalf("a no-output stall should retry to success, got %v", err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("final answer = %q, want %q", result.FinalAnswer, "done")
+	}
+	if got := atomic.LoadInt32(&p.calls); got != 2 {
+		t.Fatalf("want 2 calls (1 stall + 1 retry), got %d", got)
+	}
+}
+
+// A stall AFTER partial output must NOT be retried — re-issuing would duplicate
+// the already-streamed text. It surfaces the error instead.
+func TestRunDoesNotRetryStallAfterPartialOutput(t *testing.T) {
+	p := &stallProvider{stallBefore: 1, partialText: "partial"}
+	_, err := Run(context.Background(), "go", p, Options{Registry: tools.NewRegistry()})
+	if err == nil {
+		t.Fatal("a stall after partial output must NOT be retried; want an error")
+	}
+	if got := atomic.LoadInt32(&p.calls); got != 1 {
+		t.Fatalf("partial-then-stall must not retry, got %d calls", got)
+	}
+}
+
+// A stall after only transient REASONING was forwarded (no prose text, no
+// completed tool call) IS retried: reasoning is a "thinking" preview, not
+// committed answer text, so re-streaming it on a clearly-signalled retry is
+// acceptable and recovers the common "reasoned then froze" stall. Only
+// forwarded final prose (OnText) blocks the retry (collected.Text is empty
+// here, so this exercises the transient-vs-prose gate specifically).
+func TestRunRetriesStallAfterOnlyReasoning(t *testing.T) {
+	p := &stallProvider{stallBefore: 1, reasoningText: "thinking hard"}
+	result, err := Run(context.Background(), "go", p, Options{
+		Registry:    tools.NewRegistry(),
+		OnReasoning: func(string) {},
+	})
+	if err != nil {
+		t.Fatalf("a stall after only reasoning should retry to success, got %v", err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("final answer = %q, want %q", result.FinalAnswer, "done")
+	}
+	if got := atomic.LoadInt32(&p.calls); got != 2 {
+		t.Fatalf("want 2 calls (1 stall + 1 retry), got %d", got)
+	}
+}
+
+// A stall mid-way through an INCOMPLETE tool call (started + partial args, never
+// ended) IS retried — the exact gpt-5.x / ollama "began the big write_file then
+// froze" case. The incomplete call is never executed (a turn returns on the
+// stall error before dispatch) and never committed to history, so re-issuing is
+// safe and recovers to success. This is the primary scenario this change fixes.
+func TestRunRetriesStallAfterIncompleteToolCall(t *testing.T) {
+	starts := 0
+	p := &stallProvider{stallBefore: 1, reasoningText: "planning", partialToolCall: "write_file"}
+	result, err := Run(context.Background(), "go", p, Options{
+		Registry:        tools.NewRegistry(),
+		OnReasoning:     func(string) {},
+		OnToolCallStart: func(string, string) { starts++ },
+		OnToolCallDelta: func(string, string) {},
+	})
+	if err != nil {
+		t.Fatalf("a stall mid-incomplete-tool-call should retry to success, got %v", err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("final answer = %q, want %q", result.FinalAnswer, "done")
+	}
+	if got := atomic.LoadInt32(&p.calls); got != 2 {
+		t.Fatalf("want 2 calls (1 stall + 1 retry), got %d", got)
+	}
+	if starts != 1 {
+		t.Fatalf("the incomplete tool call should have been forwarded once before the stall, got %d starts", starts)
+	}
+}
+
+// A persistent stall surfaces an error after exhausting the capped retries.
+func TestRunGivesUpAfterMaxStallRetries(t *testing.T) {
+	p := &stallProvider{stallBefore: 99}
+	_, err := Run(context.Background(), "go", p, Options{Registry: tools.NewRegistry()})
+	if err == nil {
+		t.Fatal("a persistent stall must surface an error after exhausting retries")
+	}
+	if got := atomic.LoadInt32(&p.calls); got != int32(1+maxStreamStallRetries) {
+		t.Fatalf("want %d calls (1 + %d retries), got %d", 1+maxStreamStallRetries, maxStreamStallRetries, got)
+	}
+}
+
+func TestIsStreamTimeoutError(t *testing.T) {
+	timeouts := []string{
+		"provider stream error: no output for 10m (the model produced nothing)",
+		"provider stream error: idle timeout after 5m0s (upstream stopped sending data)",
+		"stream stalled (upstream kept the connection alive but produced no output)",
+	}
+	for _, m := range timeouts {
+		if !isStreamTimeoutError(m) {
+			t.Fatalf("want timeout-classified: %q", m)
+		}
+	}
+	notTimeouts := []string{"", "context length exceeded", "rate limit error: slow down", "model not found"}
+	for _, m := range notTimeouts {
+		if isStreamTimeoutError(m) {
+			t.Fatalf("must NOT classify as timeout: %q", m)
+		}
+	}
+}

internal/agenteval/benchmark_test.go

@@ -455,11 +455,15 @@ func TestHarnessTimeoutCancelsBlockedAgent(t *testing.T) {
 
 	// The timeout is longer than materialization of the small fixture even on
 	// slower Windows CI runners, so the agent is reached before it blocks until
-	// the deadline fires.
+	// the deadline fires. 10s wasn't generous enough in practice — it failed
+	// on an otherwise-unrelated PR merge (materialization alone consumed the
+	// full budget on a contended Windows runner, so the agent was never
+	// reached); widened for headroom, since the assertions below still bound
+	// worst-case test time via that same deadline, not this constant.
 	report := harness.Run(context.Background(), suitePath, suite, BenchmarkInput{
 		TaskID:   "document-stream-json-verify-events",
 		WorkRoot: t.TempDir(),
-		Timeout:  10 * time.Second,
+		Timeout:  30 * time.Second,
 	})
 
 	if !agentReached {

internal/cli/acp.go

@@ -43,8 +43,11 @@ func runACP(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int
 	conn := acp.NewConn(deps.stdin, stdout)
 	acp.NewAgent(conn, acp.Deps{
 		ResolveConfig: deps.resolveConfig,
-		NewProvider:   deps.newProvider,
-		RunAgent:      agent.Run,
+		// deps.newProvider is wrapped in fillAppDeps to apply the stored API key,
+		// so ACP is authenticated for apiKeyStored profiles like every other
+		// surface — no ACP-specific credential handling needed.
+		NewProvider: deps.newProvider,
+		RunAgent:    agent.Run,
 		// Build the SCOPED registry + sandbox engine per workspace, exactly like the
 		// exec surface, so ACP shell/file tools are confined — never run unconfined.
 		BuildWorkspace: func(workspaceRoot string, resolved config.ResolvedConfig) (*tools.Registry, *sandbox.Engine, error) {

internal/cli/app.go

@@ -43,12 +43,17 @@ import (
 var version = "dev"
 
 type appDeps struct {
-	getwd                 func() (string, error)
-	stdin                 io.Reader
-	userConfigPath        func() (string, error)
-	resolveConfig         func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error)
-	resolveMCPConfig      func(workspaceRoot string) (config.MCPConfig, error)
-	newProvider           func(config.ProviderProfile) (zeroruntime.Provider, error)
+	getwd            func() (string, error)
+	stdin            io.Reader
+	userConfigPath   func() (string, error)
+	resolveConfig    func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error)
+	resolveMCPConfig func(workspaceRoot string) (config.MCPConfig, error)
+	newProvider      func(config.ProviderProfile) (zeroruntime.Provider, error)
+	// exportActiveProvider pins spawned children to the run's provider (production:
+	// config.SetActiveProviderEnv, set in defaultAppDeps — deliberately NOT filled
+	// by fillAppDeps, so tests never mutate the process environment unless they
+	// inject it). nil ⇒ no export.
+	exportActiveProvider  func(providerName string)
 	probeProviderHealth   func(context.Context, providerhealth.Options) providerhealth.Result
 	detectLocalRuntimes   func(context.Context, provideronboarding.LocalDetectOptions) []provideronboarding.DetectedLocalRuntime
 	newSessionStore       func() *sessions.Store
@@ -99,9 +104,10 @@ func Run(args []string, stdout io.Writer, stderr io.Writer) int {
 
 func defaultAppDeps() appDeps {
 	return appDeps{
-		getwd:          os.Getwd,
-		stdin:          os.Stdin,
-		userConfigPath: config.DefaultUserConfigPath,
+		getwd:                os.Getwd,
+		stdin:                os.Stdin,
+		userConfigPath:       config.DefaultUserConfigPath,
+		exportActiveProvider: config.SetActiveProviderEnv,
 		resolveConfig: func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error) {
 			options, err := config.DefaultResolveOptions(workspaceRoot)
 			if err != nil {
@@ -118,9 +124,14 @@ func defaultAppDeps() appDeps {
 			return config.ResolveMCP(options)
 		},
 		newProvider: func(profile config.ProviderProfile) (zeroruntime.Provider, error) {
+			// Resolve the OAuth login ONCE: the bearer resolver and the login key it
+			// bound must describe the same login (the key is passed on to the Codex
+			// account-header resolver so it never re-selects independently).
+			resolver, loginKey := oauthLoginForProfile(profile)
 			return providers.New(profile, providers.Options{
 				UserAgent:     userAgent(),
-				OAuthResolver: oauthResolverForProfile(profile),
+				OAuthResolver: resolver,
+				OAuthLoginKey: loginKey,
 			})
 		},
 		probeProviderHealth: providerhealth.Probe,
@@ -217,9 +228,9 @@ func runWithDeps(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 	if err != nil {
 		return writeAppError(stderr, err.Error(), 1)
 	}
-	// --theme {auto|dark|light} selects the TUI palette non-interactively (populates
-	// tui.Options.Theme, which resolveThemeMode prefers over ZERO_THEME). Re-split
-	// --add-dir afterward so it may appear on either side of --theme.
+	// --theme <name> selects the TUI palette non-interactively (auto or any registered
+	// theme; populates tui.Options.Theme, which resolveThemeMode prefers over
+	// ZERO_THEME). Re-split --add-dir afterward so it may appear on either side of --theme.
 	theme, args, err := splitLeadingThemeFlag(args)
 	if err != nil {
 		return writeAppError(stderr, err.Error(), 1)
@@ -391,6 +402,14 @@ func runWithDeps(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 		if _, err := fmt.Fprintf(stderr, "unknown command %q\n", args[0]); err != nil {
 			return 1
 		}
+		// First-run users reach for `zero login`/`zero logout` (reported in the
+		// wild); point them at the real command instead of a bare usage pointer.
+		switch strings.ToLower(args[0]) {
+		case "login", "logout":
+			if _, err := fmt.Fprintf(stderr, "did you mean %q?\n", "zero auth "+strings.ToLower(args[0])); err != nil {
+				return 1
+			}
+		}
 		if _, err := fmt.Fprintln(stderr, "Run zero --help for usage."); err != nil {
 			return 1
 		}
@@ -493,6 +512,17 @@ func fillAppDeps(deps appDeps) appDeps {
 	if deps.now == nil {
 		deps.now = defaults.now
 	}
+	// Wrap newProvider ONCE, after all defaults are filled, so every surface that
+	// builds the runtime provider gets the credential-store key applied. The
+	// resolver is pure and leaves apiKeyStored profiles keyless; an unwrapped
+	// build sends unauthenticated requests. config.ApplyStoredAPIKey is a no-op
+	// when the key is already set or the profile isn't stored-key, so this is
+	// safe and idempotent for every profile kind and every injected newProvider.
+	baseNewProvider := deps.newProvider
+	userConfigPath := deps.userConfigPath
+	deps.newProvider = func(profile config.ProviderProfile) (zeroruntime.Provider, error) {
+		return baseNewProvider(applyStoredProviderKeyAt(profile, userConfigPath))
+	}
 	return deps
 }
 
@@ -508,7 +538,21 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 
 	resolved, err := deps.resolveConfig(workspaceRoot, config.Overrides{})
 	if err != nil {
-		return writeAppError(stderr, err.Error(), 1)
+		// A resolve failure the setup wizard can FIX is not fatal for the
+		// interactive TUI: drop into the wizard with an empty config so the user
+		// can onboard/repair, instead of exiting with an error they can only fix
+		// by hand-editing config.json. That covers a missing/unresolvable active
+		// provider and an active provider without a model (custom endpoints have
+		// no catalog default) — previously the second shape bricked bare `zero`
+		// and `zero setup`, the exact commands that could have fixed it. Any
+		// other error (malformed JSON, directory conflict, etc.) is still fatal,
+		// and headless commands (zero config / zero exec) still fail with the
+		// actionable message.
+		if !errors.Is(err, config.ErrNoActiveProvider) && !errors.Is(err, config.ErrProviderRequiresModel) {
+			return writeAppError(stderr, err.Error(), 1)
+		}
+		resolved = config.ResolvedConfig{}
+		forceSetup = true
 	}
 	userConfigPath, err := deps.userConfigPath()
 	if err != nil {
@@ -641,6 +685,7 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 	return deps.runTUI(context.Background(), tui.Options{
 		Cwd:                  workspaceRoot,
 		Theme:                theme,
+		SavedTheme:           resolved.Preferences.Theme,
 		UserConfigPath:       userConfigPath,
 		DoctorUserConfigPath: doctorUserConfigPath,
 		ProjectConfigPath:    projectConfigPath,
@@ -719,20 +764,50 @@ func tuiSandboxSetupCommand(backend sandbox.Backend, deps appDeps) func(context.
 	}
 }
 
+// buildProvider constructs the run's provider at STARTUP — it is called only from
+// the two launch paths (interactive TUI and headless exec), never from mid-run
+// rebuilds (escalation, wizard, ACP go through deps.newProvider directly).
 func buildProvider(resolved config.ResolvedConfig, deps appDeps) (zeroruntime.Provider, error) {
 	if !config.HasProviderProfile(resolved.Provider) {
 		return nil, nil
 	}
-	// Load the API key from the encrypted credential store when it isn't already
-	// resolved from inline config or an env var. This keeps the resolver pure (no
-	// I/O) and is a no-op for env/inline-key providers and when nothing is stored.
-	profile := resolved.Provider
-	if path, perr := deps.userConfigPath(); perr == nil {
+	// deps.newProvider is wrapped in fillAppDeps to apply the stored key, so this
+	// (and every other newProvider caller) needs no per-site key handling.
+	provider, err := deps.newProvider(resolved.Provider)
+	if err != nil {
+		return nil, err
+	}
+	// Pin spawned children (sub-agents / swarm members inherit the environment) to
+	// THIS run's provider from launch, not only after an in-session switch. Without
+	// the launch-time export a child re-resolves config.json at spawn time, so a
+	// provider switch persisted by ANOTHER zero process mid-session would silently
+	// move new children onto a different provider (and different credentials) than
+	// the parent is running. Runtime switches (/model, /provider, wizard,
+	// onboarding) re-export on commit, keeping the pin current. Injected via deps
+	// (nil in tests) because it mutates the process environment.
+	if deps.exportActiveProvider != nil {
+		deps.exportActiveProvider(resolved.Provider.Name)
+	}
+	return provider, nil
+}
+
+// applyStoredProviderKeyAt loads the API key from the encrypted credential store
+// when the resolver left it empty (the resolver is pure — no I/O — so an
+// apiKeyStored profile reaches this layer keyless). No-op for inline/env-key
+// providers and when nothing is stored. This is applied once, centrally, by the
+// deps.newProvider wrapper in fillAppDeps so EVERY surface that builds a runtime
+// provider (headless exec, the ACP builder, exec's mid-run escalation switcher,
+// and any future caller) is covered without a per-site invariant to forget.
+func applyStoredProviderKeyAt(profile config.ProviderProfile, userConfigPath func() (string, error)) config.ProviderProfile {
+	if userConfigPath == nil {
+		return profile
+	}
+	if path, perr := userConfigPath(); perr == nil {
 		if store, err := config.ProviderKeyStoreAt(filepath.Dir(path)); err == nil {
 			profile = config.ApplyStoredAPIKey(profile, store)
 		}
 	}
-	return deps.newProvider(profile)
+	return profile
 }
 
 func newCoreRegistry(workspaceRoot string) *tools.Registry {
@@ -1033,9 +1108,10 @@ func splitLeadingAddDirFlags(args []string) ([]string, []string, error) {
 	return addDirs, args, nil
 }
 
-// splitLeadingThemeFlag strips a leading --theme {auto|dark|light} (space or =form)
-// from the root argument list and validates it. The last occurrence wins. A value
-// outside the allowed set is a loud error rather than a silent fallback.
+// splitLeadingThemeFlag strips a leading --theme <auto|theme-name> (space or =form)
+// from the root argument list and validates it against the registered themes. The
+// last occurrence wins. A value outside the allowed set is a loud error rather than
+// a silent fallback.
 func splitLeadingThemeFlag(args []string) (string, []string, error) {
 	theme := ""
 	for len(args) > 0 {
@@ -1043,7 +1119,7 @@ func splitLeadingThemeFlag(args []string) (string, []string, error) {
 		switch {
 		case args[0] == "--theme":
 			if len(args) < 2 {
-				return "", nil, errors.New("--theme requires a value (auto, dark, or light)")
+				return "", nil, errors.New("--theme requires a value (auto or a theme name; try --theme auto)")
 			}
 			value = strings.TrimSpace(args[1])
 			args = args[2:]
@@ -1053,12 +1129,10 @@ func splitLeadingThemeFlag(args []string) (string, []string, error) {
 		default:
 			return theme, args, nil
 		}
-		switch strings.ToLower(value) {
-		case "auto", "dark", "light":
-			theme = strings.ToLower(value)
-		default:
-			return "", nil, fmt.Errorf("--theme must be auto, dark, or light (got %q)", value)
+		if !tui.ValidThemeArg(value) {
+			return "", nil, fmt.Errorf("--theme must be auto or a registered theme name (got %q)", value)
 		}
+		theme = strings.ToLower(value)
 	}
 	return theme, args, nil
 }

internal/cli/app_test.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"fmt"
 	"io"
 	"os"
 	"path/filepath"
@@ -157,6 +158,80 @@ func TestRunNoArgsLaunchesSetupTUIWithNilProviderWhenNoProviderConfigured(t *tes
 	assertAgentOptions(t, launchedOptions, 12, agent.PermissionModeAsk)
 }
 
+func TestRunNoArgsEntersSetupWhenResolveReportsNoActiveProvider(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	cwd := t.TempDir()
+	setCLIUserConfigRoot(t)
+	userConfigPath := filepath.Join(t.TempDir(), "zero", "config.json")
+	var launchedOptions tui.Options
+	launched := false
+
+	exitCode := runWithDeps([]string{}, &stdout, &stderr, appDeps{
+		getwd: func() (string, error) {
+			return cwd, nil
+		},
+		resolveConfig: func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error) {
+			return config.ResolvedConfig{}, fmt.Errorf("%w: active provider %q not found", config.ErrNoActiveProvider, "ghost")
+		},
+		newProvider: func(config.ProviderProfile) (zeroruntime.Provider, error) {
+			t.Fatal("newProvider should not be called without a resolved provider")
+			return nil, nil
+		},
+		userConfigPath: func() (string, error) {
+			return userConfigPath, nil
+		},
+		registerMCPTools: func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			return noopMCPRuntime{}, nil
+		},
+		runTUI: func(ctx context.Context, options tui.Options) int {
+			launched = true
+			launchedOptions = options
+			return 0
+		},
+	})
+
+	if exitCode != 0 {
+		t.Fatalf("exit code = %d, want 0 (setup TUI launched), stderr=%q", exitCode, stderr.String())
+	}
+	if !launched {
+		t.Fatal("TUI was not launched; expected fallback into setup instead of a fatal error")
+	}
+	if !launchedOptions.Setup.Visible || !launchedOptions.Setup.Required {
+		t.Fatalf("Setup = %#v, want visible required setup", launchedOptions.Setup)
+	}
+	if launchedOptions.Provider != nil {
+		t.Fatalf("Provider = %#v, want nil for no-provider fallback", launchedOptions.Provider)
+	}
+}
+
+func TestRunNoArgsFailsWhenResolveErrorIsNotProviderRelated(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	cwd := t.TempDir()
+	setCLIUserConfigRoot(t)
+
+	exitCode := runWithDeps([]string{}, &stdout, &stderr, appDeps{
+		getwd: func() (string, error) {
+			return cwd, nil
+		},
+		resolveConfig: func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error) {
+			return config.ResolvedConfig{}, fmt.Errorf("invalid config JSON")
+		},
+		runTUI: func(ctx context.Context, options tui.Options) int {
+			t.Fatal("TUI must not launch on a non-provider resolve error")
+			return 0
+		},
+	})
+
+	if exitCode == 0 {
+		t.Fatal("exit code = 0, want non-zero for fatal config error")
+	}
+	if !strings.Contains(stderr.String(), "invalid config JSON") {
+		t.Fatalf("stderr = %q, want the underlying config error", stderr.String())
+	}
+}
+
 func TestRunNoArgsLaunchesTUIWithMCPState(t *testing.T) {
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer
@@ -1397,7 +1472,79 @@ func TestRunThemeFlagRejectsBadValue(t *testing.T) {
 	if launched {
 		t.Fatal("the TUI must not launch on a bad --theme value")
 	}
-	if !strings.Contains(stderr.String(), "--theme must be auto, dark, or light") {
+	if !strings.Contains(stderr.String(), "--theme must be auto or a registered theme name") {
 		t.Fatalf("expected a clear --theme error, got %q", stderr.String())
 	}
 }
+
+// A custom endpoint saved without a model previously bricked bare `zero` and
+// `zero setup` — the exact commands that could have fixed it (the resolve
+// error escaped before the wizard could open). The interactive TUI now treats
+// the requires-model failure as "needs onboarding", same as a missing active
+// provider. The error comes from a REAL Resolve over a real config file, so
+// this exercises the production sentinel wrapping, not a hand-built error.
+func TestRunNoArgsEntersSetupWhenActiveProviderMissesModel(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	cwd := t.TempDir()
+	setCLIUserConfigRoot(t)
+	brokenConfig := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(brokenConfig, []byte(`{
+		"activeProvider": "gw",
+		"providers": [{"name": "gw", "provider_kind": "openai-compatible", "baseURL": "https://gw.example/v1", "apiKey": "sk-x"}]
+	}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	launched := false
+	var launchedOptions tui.Options
+
+	exitCode := runWithDeps([]string{}, &stdout, &stderr, appDeps{
+		getwd: func() (string, error) { return cwd, nil },
+		resolveConfig: func(string, config.Overrides) (config.ResolvedConfig, error) {
+			return config.Resolve(config.ResolveOptions{UserConfigPath: brokenConfig, Env: map[string]string{}})
+		},
+		newProvider: func(config.ProviderProfile) (zeroruntime.Provider, error) {
+			t.Fatal("newProvider must not be called without a resolved provider")
+			return nil, nil
+		},
+		userConfigPath: func() (string, error) { return filepath.Join(t.TempDir(), "zero", "config.json"), nil },
+		registerMCPTools: func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			return noopMCPRuntime{}, nil
+		},
+		runTUI: func(_ context.Context, options tui.Options) int {
+			launched = true
+			launchedOptions = options
+			return 0
+		},
+	})
+
+	if exitCode != 0 {
+		t.Fatalf("exit code = %d, want 0 (setup TUI launched), stderr=%q", exitCode, stderr.String())
+	}
+	if !launched {
+		t.Fatal("TUI was not launched; a fixable requires-model config must fall into setup, not exit fatally")
+	}
+	if !launchedOptions.Setup.Visible || !launchedOptions.Setup.Required {
+		t.Fatalf("Setup = %#v, want visible required setup", launchedOptions.Setup)
+	}
+}
+
+// `zero login`/`zero logout` don't exist (it's `zero auth login`); first-run
+// users try them (reported in the wild), so the unknown-command error points at
+// the real command.
+func TestRunUnknownLoginSuggestsAuthLogin(t *testing.T) {
+	var stdout, stderr bytes.Buffer
+	exitCode := runWithDeps([]string{"login"}, &stdout, &stderr, appDeps{})
+	if exitCode != 2 {
+		t.Fatalf("exit = %d, want 2 (unknown command)", exitCode)
+	}
+	if !strings.Contains(stderr.String(), `"zero auth login"`) {
+		t.Fatalf("stderr = %q, want a zero auth login suggestion", stderr.String())
+	}
+	// An unrelated unknown command gets no misleading suggestion.
+	stderr.Reset()
+	runWithDeps([]string{"frobnicate"}, &stdout, &stderr, appDeps{})
+	if strings.Contains(stderr.String(), "did you mean") {
+		t.Fatalf("stderr = %q, unrelated commands must not get the auth hint", stderr.String())
+	}
+}

internal/cli/auth.go

@@ -283,6 +283,9 @@ func newAuthManager(deps appDeps, out io.Writer) (*oauth.Manager, error) {
 		// carries no token material. A real browser launch is intentionally not
 		// performed (the sandbox/headless contexts make printing the safer default).
 		OpenBrowser: func(string) error { return nil },
+		// `zero auth login <preset>` (e.g. xai) should resolve the baked-in preset
+		// without the operator exporting ZERO_OAUTH_ALLOW_PRESETS first.
+		AllowPresets: true,
 	})
 }
 

internal/cli/auth_propagation_test.go

@@ -0,0 +1,180 @@
+package cli
+
+import (
+	"bytes"
+	"errors"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// seedStoredProviderKey writes key into a credential store rooted beside the
+// returned config path, and returns deps fields wired at that path — the same
+// layout applyStoredProviderKey resolves through in production.
+func seedStoredProviderKey(t *testing.T, providerName, key string) (userConfigPath string) {
+	t.Helper()
+	t.Setenv("ZERO_CRED_STORAGE", "encrypted-file") // never touch the real OS keychain in tests
+	dir := t.TempDir()
+	store, err := config.ProviderKeyStoreAt(dir)
+	if err != nil {
+		t.Fatalf("ProviderKeyStoreAt: %v", err)
+	}
+	if err := store.Set(providerName, key); err != nil {
+		t.Fatalf("seed stored key: %v", err)
+	}
+	return filepath.Join(dir, "config.json")
+}
+
+// applyStoredProviderKeyAt must fill APIKey from the credential store for an
+// apiKeyStored profile — the single central helper the newProvider wrapper uses.
+func TestApplyStoredProviderKeyAtFillsFromCredstore(t *testing.T) {
+	configPath := seedStoredProviderKey(t, "echo", "sk-stored-test")
+	ucp := func() (string, error) { return configPath, nil }
+
+	stored := applyStoredProviderKeyAt(config.ProviderProfile{Name: "echo", APIKeyStored: true}, ucp)
+	if stored.APIKey != "sk-stored-test" {
+		t.Fatalf("APIKey = %q, want the stored key", stored.APIKey)
+	}
+
+	// A profile that did NOT opt into stored-key auth must be left unchanged —
+	// a stale store entry must not silently reactivate credentials.
+	plain := applyStoredProviderKeyAt(config.ProviderProfile{Name: "echo"}, ucp)
+	if plain.APIKey != "" {
+		t.Fatalf("non-stored profile APIKey = %q, want empty", plain.APIKey)
+	}
+}
+
+// fillAppDeps wraps newProvider so EVERY surface that builds a runtime provider
+// (buildProvider, the ACP builder — which is just deps.newProvider — and exec's
+// escalation switcher) gets the stored key applied. Passing the raw resolved
+// profile previously sent unauthenticated requests for apiKeyStored profiles,
+// the default onboarding outcome. Testing the wrap covers all those surfaces.
+func TestFillAppDepsWrapsNewProviderWithStoredKey(t *testing.T) {
+	configPath := seedStoredProviderKey(t, "echo", "sk-stored-wrap")
+
+	var captured config.ProviderProfile
+	deps := fillAppDeps(appDeps{
+		userConfigPath: func() (string, error) { return configPath, nil },
+		newProvider: func(profile config.ProviderProfile) (zeroruntime.Provider, error) {
+			captured = profile
+			return nil, nil // only the captured profile matters here
+		},
+	})
+
+	if _, err := deps.newProvider(config.ProviderProfile{Name: "echo", APIKeyStored: true}); err != nil {
+		t.Fatalf("wrapped newProvider: %v", err)
+	}
+	if captured.APIKey != "sk-stored-wrap" {
+		t.Fatalf("wrapped newProvider built with APIKey = %q, want the stored key (unauthenticated regression)", captured.APIKey)
+	}
+}
+
+// buildProvider (the TUI/exec STARTUP construction site) must export
+// ZERO_PROVIDER so children spawned at any point in the run are pinned to the
+// parent's provider from launch — not only after an in-session switch. Without
+// this, a provider switch persisted by another zero process mid-session moves
+// new children onto a different provider (and credentials) than the parent.
+func TestBuildProviderExportsActiveProviderEnv(t *testing.T) {
+	t.Setenv(config.ActiveProviderEnv, "stale-from-elsewhere")
+	// The exporter is injected (defaultAppDeps wires config.SetActiveProviderEnv;
+	// fillAppDeps deliberately leaves it nil so ordinary tests never mutate the
+	// process env). Inject the real one here to assert the actual env effect.
+	deps := appDeps{
+		exportActiveProvider: config.SetActiveProviderEnv,
+		newProvider: func(config.ProviderProfile) (zeroruntime.Provider, error) {
+			return nil, nil
+		},
+	}
+
+	resolved := config.ResolvedConfig{Provider: config.ProviderProfile{Name: "pinned", ProviderKind: config.ProviderKindAnthropic, Model: "m"}}
+	if _, err := buildProvider(resolved, deps); err != nil {
+		t.Fatalf("buildProvider: %v", err)
+	}
+	if got := os.Getenv(config.ActiveProviderEnv); got != "pinned" {
+		t.Fatalf("%s = %q after startup build, want %q (children would spawn on the stale provider)", config.ActiveProviderEnv, got, "pinned")
+	}
+
+	// A FAILED build must not move the pin: the process never committed to the
+	// new provider, so children must keep resolving whatever was in effect.
+	t.Setenv(config.ActiveProviderEnv, "still-current")
+	failing := appDeps{
+		exportActiveProvider: config.SetActiveProviderEnv,
+		newProvider: func(config.ProviderProfile) (zeroruntime.Provider, error) {
+			return nil, errors.New("boom")
+		},
+	}
+	if _, err := buildProvider(resolved, failing); err == nil {
+		t.Fatal("expected build error")
+	}
+	if got := os.Getenv(config.ActiveProviderEnv); got != "still-current" {
+		t.Fatalf("%s = %q after failed build, want it untouched", config.ActiveProviderEnv, got)
+	}
+
+	// No provider profile at all → no provider, no pin change.
+	if _, err := buildProvider(config.ResolvedConfig{}, deps); err != nil {
+		t.Fatalf("buildProvider(empty): %v", err)
+	}
+	if got := os.Getenv(config.ActiveProviderEnv); got != "still-current" {
+		t.Fatalf("%s = %q after profile-less build, want it untouched", config.ActiveProviderEnv, got)
+	}
+}
+
+// A mid-run --allow-escalation model switch previously rebuilt the provider
+// from the PURE resolved profile (the stored key lives only in buildProvider's
+// local copy), so the escalated provider was keyless and the next completion
+// call 401'd mid-run. Every profile handed to newProvider on this path must
+// carry the stored key.
+func TestRunExecEscalationSwitchKeepsStoredKey(t *testing.T) {
+	dataHome := t.TempDir()
+	t.Setenv("XDG_DATA_HOME", dataHome)
+	cwd := t.TempDir()
+
+	if _, ok := mustUpgradeTarget(t, "claude-haiku-4.5"); !ok {
+		t.Skip("registry has no upgrade target for claude-haiku-4.5")
+	}
+
+	configPath := seedStoredProviderKey(t, "echo", "sk-stored-escalation")
+
+	var builtProfiles []config.ProviderProfile
+	var stdout, stderr bytes.Buffer
+	exitCode := runWithDeps([]string{
+		"exec",
+		"--allow-escalation",
+		"--model", "claude-haiku-4.5",
+		"--init-session-id", "escalation_stored_key_run",
+		"escalate please",
+	}, &stdout, &stderr, appDeps{
+		getwd:          func() (string, error) { return cwd, nil },
+		userConfigPath: func() (string, error) { return configPath, nil },
+		resolveConfig: func(_ string, overrides config.Overrides) (config.ResolvedConfig, error) {
+			model := "claude-haiku-4.5"
+			if overrides.Provider.Model != "" {
+				model = overrides.Provider.Model
+			}
+			cfg := execResolvedConfig()
+			cfg.Provider.ProviderKind = config.ProviderKindAnthropic
+			cfg.Provider.Model = model
+			cfg.Provider.APIKeyStored = true // key lives ONLY in the credstore
+			cfg.MaxTurns = 3
+			return cfg, nil
+		},
+		newProvider: func(profile config.ProviderProfile) (zeroruntime.Provider, error) {
+			builtProfiles = append(builtProfiles, profile)
+			return &usageEmittingEscalatingProvider{escalate: len(builtProfiles) == 1}, nil
+		},
+	})
+	if exitCode != exitSuccess {
+		t.Fatalf("exit = %d, want success; stderr=%s", exitCode, stderr.String())
+	}
+	if len(builtProfiles) < 2 {
+		t.Fatalf("want at least 2 provider builds (initial + escalated), got %d", len(builtProfiles))
+	}
+	for i, profile := range builtProfiles {
+		if profile.APIKey != "sk-stored-escalation" {
+			t.Fatalf("provider build %d APIKey = %q, want the stored key (escalated build went out keyless)", i, profile.APIKey)
+		}
+	}
+}

internal/cli/exec.go

@@ -365,6 +365,9 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 	var modelSwitcher func(context.Context, string) (agent.Provider, error)
 	if options.allowEscalation {
 		modelSwitcher = func(_ context.Context, modelID string) (agent.Provider, error) {
+			// deps.newProvider is wrapped (fillAppDeps) to apply the stored key, so
+			// the escalated provider is authenticated even though resolved.Provider
+			// is the pure profile — no per-site key handling here.
 			switchedProfile := resolved.Provider
 			switchedProfile.Model = modelID
 			switchedProvider, err := deps.newProvider(switchedProfile)

internal/cli/oauth_credential_test.go

@@ -24,6 +24,7 @@ func TestProviderHasOAuthLogin(t *testing.T) {
 
 func TestSetupRequiredRecognizesOAuthLogin(t *testing.T) {
 	path := filepath.Join(t.TempDir(), "tok.json")
+	t.Setenv("ZERO_OAUTH_STORAGE", "file") // an inherited "keyring" would ignore the temp path and hit the OS keychain
 	t.Setenv("ZERO_OAUTH_TOKENS_PATH", path)
 	store, err := oauth.NewStore(oauth.StoreOptions{FilePath: path})
 	if err != nil {
@@ -45,3 +46,88 @@ func TestSetupRequiredRecognizesOAuthLogin(t *testing.T) {
 		t.Fatal("a keyless provider with no OAuth login must require onboarding")
 	}
 }
+
+// A profile renamed away from its catalog ID must still get a runtime OAuth
+// resolver when the login is stored under the catalog name — and a CASE-VARIANT
+// profile name must not drop the catalog-ID candidate (the store is
+// case-sensitive; a case-insensitive dedupe would swallow it).
+func TestOAuthResolverForProfileFallsBackToCatalogID(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "tok.json")
+	t.Setenv("ZERO_OAUTH_STORAGE", "file") // an inherited "keyring" would ignore the temp path and hit the OS keychain
+	t.Setenv("ZERO_OAUTH_TOKENS_PATH", path)
+	store, err := oauth.NewStore(oauth.StoreOptions{FilePath: path})
+	if err != nil {
+		t.Fatalf("NewStore: %v", err)
+	}
+	if err := store.Save(oauth.ProviderKey("chatgpt"), oauth.Token{AccessToken: "tok", ExpiresAt: time.Now().Add(time.Hour)}); err != nil {
+		t.Fatalf("seed token: %v", err)
+	}
+
+	// A renamed profile resolves the catalog-ID login, and the returned key MUST be
+	// that login's key — the same key the Codex account-header resolver reads from,
+	// so bearer and account can never come from different logins.
+	if resolver, key := oauthLoginForProfile(config.ProviderProfile{Name: "codex", CatalogID: "chatgpt"}); resolver == nil {
+		t.Fatal("renamed profile with a catalog-ID login must get a resolver (unauthenticated-children regression)")
+	} else if key != oauth.ProviderKey("chatgpt") {
+		t.Fatalf("bound key = %q, want the chatgpt login key (bearer/account must share it)", key)
+	}
+	// Name differs from catalog ID only in CASE: the exact-case store key is
+	// "provider:chatgpt", so "chatgpt" must survive as a distinct candidate.
+	if resolver, key := oauthLoginForProfile(config.ProviderProfile{Name: "ChatGPT", CatalogID: "chatgpt"}); resolver == nil {
+		t.Fatal("case-variant profile name must still resolve the catalog-ID login (case-insensitive dedupe regression)")
+	} else if key != oauth.ProviderKey("chatgpt") {
+		t.Fatalf("case-variant bound key = %q, want the chatgpt login key", key)
+	}
+	if resolver, _ := oauthLoginForProfile(config.ProviderProfile{Name: "chatgpt", CatalogID: "chatgpt"}); resolver == nil {
+		t.Fatal("catalog-named profile must get a resolver")
+	}
+	if resolver, key := oauthLoginForProfile(config.ProviderProfile{Name: "openai", CatalogID: "openai"}); resolver != nil || key != "" {
+		t.Fatalf("a profile with no stored login must get no resolver and an empty key, got key=%q", key)
+	}
+}
+
+// The catalog-ID fallback must NOT reach a sibling profile's OAuth login when
+// this profile has its own key — withBearer would erase the key and silently
+// bill the model call under the sibling's account (a real cross-profile
+// override the fallback introduced). A profile WITH its own resolved key gets no
+// resolver even when a catalog-shared login exists.
+func TestOAuthResolverForProfileDoesNotOverrideExplicitKey(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "tok.json")
+	t.Setenv("ZERO_OAUTH_STORAGE", "file") // an inherited "keyring" would ignore the temp path and hit the OS keychain
+	t.Setenv("ZERO_OAUTH_TOKENS_PATH", path)
+	store, err := oauth.NewStore(oauth.StoreOptions{FilePath: path})
+	if err != nil {
+		t.Fatalf("NewStore: %v", err)
+	}
+	// A personal login stored under the catalog name.
+	if err := store.Save(oauth.ProviderKey("anthropic"), oauth.Token{AccessToken: "personal-tok", ExpiresAt: time.Now().Add(time.Hour)}); err != nil {
+		t.Fatalf("seed token: %v", err)
+	}
+
+	// Work profile: same catalog, its OWN resolved API key, deliberately NOT the
+	// personal OAuth account. It must keep using its key (no resolver attached).
+	work := config.ProviderProfile{Name: "anthropic-work", CatalogID: "anthropic", APIKey: "sk-work-key"}
+	if resolver, key := oauthLoginForProfile(work); resolver != nil || key != "" {
+		t.Fatalf("a profile with its own API key must not get a catalog-shared OAuth resolver (cross-profile override), got key=%q", key)
+	}
+	// A raw auth-header credential must be protected the same way.
+	header := config.ProviderProfile{Name: "anthropic-hdr", CatalogID: "anthropic", AuthHeaderValue: "Bearer byo"}
+	if resolver, _ := oauthLoginForProfile(header); resolver != nil {
+		t.Fatal("a profile with its own auth header must not get a catalog-shared OAuth resolver")
+	}
+	// A same-NAME login must not override an explicit key either (the name
+	// candidate is gated on the profile being keyless, not just the catalog one).
+	if err := store.Save(oauth.ProviderKey("anthropic-work"), oauth.Token{AccessToken: "name-tok", ExpiresAt: time.Now().Add(time.Hour)}); err != nil {
+		t.Fatalf("seed name token: %v", err)
+	}
+	if resolver, _ := oauthLoginForProfile(work); resolver != nil {
+		t.Fatal("a profile with its own API key must not get an OAuth resolver even from a same-name login")
+	}
+	// Sanity: a keyless sibling on the same catalog DOES resolve the shared login.
+	keyless := config.ProviderProfile{Name: "anthropic-oauth", CatalogID: "anthropic"}
+	if resolver, key := oauthLoginForProfile(keyless); resolver == nil {
+		t.Fatal("a keyless profile should still resolve the catalog-shared login")
+	} else if key != oauth.ProviderKey("anthropic") {
+		t.Fatalf("keyless sibling bound key = %q, want the catalog login key", key)
+	}
+}

internal/cli/oauth_provider.go

@@ -4,7 +4,6 @@ import (
 	"context"
 	"errors"
 	"net/http"
-	"strings"
 	"time"
 
 	"github.com/Gitlawb/zero/internal/config"
@@ -12,42 +11,48 @@ import (
 	"github.com/Gitlawb/zero/internal/providers/providerio"
 )
 
-// providerOAuthName resolves the OAuth login name for a provider profile. It is
-// the name a user logs in under (`zero auth login <name>`) whose token then
-// authenticates this provider's model calls. For now it is the profile name;
-// the provider presets map well-known kinds (openai/anthropic) to canonical
-// names on top of this.
-func providerOAuthName(profile config.ProviderProfile) string {
-	return strings.TrimSpace(profile.Name)
-}
-
-// oauthResolverForProfile returns a TokenResolver that authenticates a provider's
-// model calls with the user's OAuth login, or nil when no login exists for this
-// provider. Returning nil for the no-login case keeps API-key users free of any
-// per-request store lookups: the resolver is only attached once a login is
-// present at construction time.
-func oauthResolverForProfile(profile config.ProviderProfile) providerio.TokenResolver {
-	name := providerOAuthName(profile)
-	if name == "" {
-		return nil
+// oauthLoginForProfile resolves the user's OAuth login for a provider ONCE and
+// returns both a TokenResolver that authenticates model calls with it and the
+// credential-store key it bound to. It returns (nil, "") when no login exists —
+// keeping API-key users free of any per-request store lookups, since the resolver
+// is only attached when a login is present at construction time.
+//
+// The returned key is the single source of truth for "which login is this
+// provider using": callers pass it to providers.Options.OAuthLoginKey so the
+// Codex chatgpt-account-id header reads its account from the exact login that
+// issued the bearer token, instead of doing a second, independent lookup that
+// could select a different login (a backend-rejected mismatch).
+//
+// Candidate login names (profile name, then a catalog-ID fallback, both gated on
+// the profile having no own configured credential) come from the shared
+// ProviderProfile.OAuthLoginCandidates so the runtime resolver, the Codex account
+// resolver, and the onboarding presence check never diverge.
+func oauthLoginForProfile(profile config.ProviderProfile) (providerio.TokenResolver, string) {
+	candidates := profile.OAuthLoginCandidates()
+	if len(candidates) == 0 {
+		return nil, ""
 	}
 	store, err := oauth.NewStore(oauth.StoreOptions{})
 	if err != nil {
-		return nil
+		return nil, ""
 	}
-	key := oauth.ProviderKey(name)
-	if _, ok, loadErr := store.Load(key); loadErr != nil || !ok {
-		// No login (or an unreadable/invalid key) → API-key auth, no resolver.
-		return nil
+	_, key, ok := oauth.FirstStored(store, candidates)
+	if !ok {
+		// No login under any candidate (or unreadable/invalid keys) → API-key
+		// auth, no resolver.
+		return nil, ""
 	}
 	manager, err := oauth.NewManager(oauth.ManagerOptions{
 		Store:      store,
 		HTTPClient: &http.Client{Timeout: 30 * time.Second},
+		// Refreshing a token the user logged into (possibly a preset provider like
+		// xAI) re-resolves that provider's OAuth config, which needs the preset.
+		AllowPresets: true,
 	})
 	if err != nil {
-		return nil
+		return nil, ""
 	}
-	return func(ctx context.Context, forceRefresh bool) (string, string, bool, error) {
+	resolver := func(ctx context.Context, forceRefresh bool) (string, string, bool, error) {
 		var token string
 		var rerr error
 		if forceRefresh {
@@ -64,4 +69,5 @@ func oauthResolverForProfile(profile config.ProviderProfile) providerio.TokenRes
 		}
 		return "Authorization", "Bearer " + token, true, nil
 	}
+	return resolver, key
 }

internal/cli/provider_setup.go

@@ -488,7 +488,7 @@ func validateProviderRuntimeReady(profile config.ProviderProfile) error {
 }
 
 func providerProfileHasCredential(profile config.ProviderProfile) bool {
-	return strings.TrimSpace(profile.APIKey) != "" || profile.APIKeyStored || strings.TrimSpace(profile.AuthHeaderValue) != ""
+	return profile.HasConfiguredCredential()
 }
 
 func providerKindForDescriptor(descriptor providercatalog.Descriptor) config.ProviderKind {

internal/cli/setup.go

@@ -314,10 +314,14 @@ func setupRequired(resolved config.ResolvedConfig) bool {
 }
 
 // providerHasOAuthLogin reports whether a stored OAuth login exists for the
-// provider, keyed by its profile name or catalog id.
+// provider. It uses the same ProviderProfile.OAuthLoginCandidates the runtime
+// resolver does, so the "authenticated in the UI" gate and the "authenticated at
+// runtime" resolver can never diverge. This is only consulted for a profile with
+// no usable API-key credential (setupRequired returns early otherwise), where
+// the candidate set is permissive (profile name + catalog ID).
 func providerHasOAuthLogin(profile config.ProviderProfile, oauthLogins map[string]bool) bool {
-	for _, name := range []string{strings.TrimSpace(profile.Name), strings.TrimSpace(profile.CatalogID)} {
-		if name != "" && oauthLogins[name] {
+	for _, name := range profile.OAuthLoginCandidates() {
+		if oauthLogins[name] {
 			return true
 		}
 	}

internal/cli/specialist_test.go

@@ -8,6 +8,8 @@ import (
 	"runtime"
 	"strings"
 	"testing"
+
+	"github.com/Gitlawb/zero/internal/config"
 )
 
 func TestRunSpecialistListShowAndPath(t *testing.T) {
@@ -326,7 +328,7 @@ func setSpecialistConfigRoot(t *testing.T) string {
 	default:
 		t.Setenv("XDG_CONFIG_HOME", root)
 	}
-	configRoot, err := os.UserConfigDir()
+	configRoot, err := config.UserConfigDir()
 	if err != nil {
 		t.Fatalf("UserConfigDir() error = %v", err)
 	}

internal/config/active_provider_env_test.go

@@ -0,0 +1,49 @@
+package config
+
+import (
+	"os"
+	"testing"
+)
+
+// SetActiveProviderEnv must export the provider name so a child process (whose
+// resolution reads the inherited environment via the nil-map -> os.Getenv path in
+// applyEnv) resolves the SAME active provider. A blank name CLEARS it.
+func TestSetActiveProviderEnvRoundTrips(t *testing.T) {
+	t.Setenv(ActiveProviderEnv, "") // register cleanup; restores the original after
+
+	if err := os.Unsetenv(ActiveProviderEnv); err != nil {
+		t.Fatalf("unset: %v", err)
+	}
+	SetActiveProviderEnv("") // blank with nothing set stays unset
+	if got := os.Getenv(ActiveProviderEnv); got != "" {
+		t.Fatalf("empty name set the env to %q, want unset", got)
+	}
+
+	// Regression: switching from a named provider back to an unnamed/default profile
+	// must CLEAR a previously exported value, not leave it stale for children.
+	if err := os.Setenv(ActiveProviderEnv, "stale-provider"); err != nil {
+		t.Fatalf("seed stale: %v", err)
+	}
+	SetActiveProviderEnv("") // blank must unset the stale value
+	if got := os.Getenv(ActiveProviderEnv); got != "" {
+		t.Fatalf("blank name left stale env %q, want cleared", got)
+	}
+	cfgStale := &FileConfig{}
+	applyEnv(cfgStale, nil)
+	if cfgStale.ActiveProvider != "" {
+		t.Fatalf("applyEnv still resolved stale provider %q after clear", cfgStale.ActiveProvider)
+	}
+
+	SetActiveProviderEnv("ollama-cloud")
+	if got := os.Getenv(ActiveProviderEnv); got != "ollama-cloud" {
+		t.Fatalf("env = %q, want ollama-cloud", got)
+	}
+
+	// applyEnv with a nil map is the exact path a child uses: it falls back to
+	// os.Getenv, so the inherited value selects the active provider.
+	cfg := &FileConfig{}
+	applyEnv(cfg, nil)
+	if cfg.ActiveProvider != "ollama-cloud" {
+		t.Fatalf("applyEnv ActiveProvider = %q, want ollama-cloud (inherited env not read)", cfg.ActiveProvider)
+	}
+}

internal/config/credentials.go

@@ -155,6 +155,63 @@ func MigratePlaintextProviderKeys(path string, store APIKeySetter) (int, error)
 	return migrated, nil
 }
 
+// HasConfiguredCredential reports whether the profile is set up to authenticate
+// with its own API key (inline, stored, or a raw auth header). It is the single
+// definition of "this profile is key-authed", shared by OAuthLoginCandidates and
+// the cli/tui setup surfaces so they never disagree about whether a profile is
+// keyed. APIKeyEnv is deliberately NOT included: an env var may be unset at
+// runtime while the profile actually relies on an OAuth login, so an
+// env-configured-but-empty profile must still be allowed to resolve a token.
+func (profile ProviderProfile) HasConfiguredCredential() bool {
+	return strings.TrimSpace(profile.APIKey) != "" ||
+		profile.APIKeyStored ||
+		strings.TrimSpace(profile.AuthHeaderValue) != ""
+}
+
+// OAuthLoginCandidates returns the login names to try, in order, when resolving
+// this profile's OAuth token — used by the runtime bearer resolver, the Codex
+// account-header resolver, and the onboarding presence check so all three agree.
+//
+// It returns NO candidates for a profile that carries its own configured
+// credential: attaching an OAuth resolver to such a profile makes withBearer
+// erase the key the instant a token resolves, silently routing/billing the call
+// through the OAuth account instead of the intended key (the cross-profile
+// override the review caught — and, via the profile name, a same-name login
+// overriding an explicit key). A stored-key profile that reaches this keyless
+// (e.g. a transiently unreadable keyring) is likewise treated as key-auth, so it
+// fails with a clear missing-credential error rather than silently borrowing an
+// unrelated OAuth login. APIKeyEnv is intentionally not part of that gate (an
+// env var may be unset while the profile relies on OAuth).
+//
+// For a keyless profile the profile name comes first (a login under your own
+// profile name is unambiguously yours), then the catalog ID as a FALLBACK (so a
+// profile renamed away from its catalog ID — e.g. {name:"codex",
+// catalogID:"chatgpt"} — still finds a `zero auth login chatgpt` token).
+// Candidates are trimmed, blank-skipped, and de-duplicated CASE-SENSITIVELY — the
+// OAuth token store is a case-sensitive map, so "ChatGPT" and "chatgpt" are
+// distinct keys and both must survive.
+func (profile ProviderProfile) OAuthLoginCandidates() []string {
+	if profile.HasConfiguredCredential() {
+		return nil
+	}
+	var names []string
+	add := func(s string) {
+		s = strings.TrimSpace(s)
+		if s == "" {
+			return
+		}
+		for _, existing := range names {
+			if existing == s { // case-sensitive: the store lookup is case-sensitive
+				return
+			}
+		}
+		names = append(names, s)
+	}
+	add(profile.Name)
+	add(profile.CatalogID)
+	return names
+}
+
 // ApplyStoredAPIKey fills profile.APIKey from the credential store when it is not
 // already resolved — an inline config key or a resolved APIKeyEnv always wins, so
 // this only supplies a key for providers whose secret lives in the store. A nil

internal/config/credentials_test.go

@@ -172,3 +172,69 @@ func TestProviderProfileAPIKeyStoredRoundTrips(t *testing.T) {
 		t.Fatalf("no inline key expected, got %q", p.APIKey)
 	}
 }
+
+// OAuthLoginCandidates always offers the profile name, adds the catalog ID as a
+// fallback ONLY when the profile has no effective own credential, and dedupes
+// case-sensitively (the OAuth store is a case-sensitive map).
+func TestOAuthLoginCandidates(t *testing.T) {
+	cases := []struct {
+		name    string
+		profile ProviderProfile
+		want    []string
+	}{
+		{
+			name:    "renamed keyless profile falls back to catalog id",
+			profile: ProviderProfile{Name: "codex", CatalogID: "chatgpt"},
+			want:    []string{"codex", "chatgpt"},
+		},
+		{
+			name:    "case-variant name keeps distinct catalog-id candidate",
+			profile: ProviderProfile{Name: "ChatGPT", CatalogID: "chatgpt"},
+			want:    []string{"ChatGPT", "chatgpt"},
+		},
+		{
+			name:    "exact-duplicate name and catalog id collapse",
+			profile: ProviderProfile{Name: "chatgpt", CatalogID: "chatgpt"},
+			want:    []string{"chatgpt"},
+		},
+		{
+			// A configured key must block ALL candidates (name included): a login
+			// under the profile's own name would otherwise erase the key too.
+			name:    "own inline key blocks every candidate",
+			profile: ProviderProfile{Name: "anthropic-work", CatalogID: "anthropic", APIKey: "sk-work"},
+			want:    nil,
+		},
+		{
+			name:    "own auth header blocks every candidate",
+			profile: ProviderProfile{Name: "acme", CatalogID: "anthropic", AuthHeaderValue: "Bearer x"},
+			want:    nil,
+		},
+		{
+			// APIKeyStored means key-auth even if the key isn't loaded here (e.g. a
+			// transiently unreadable keyring): don't silently borrow an OAuth login.
+			name:    "stored-key profile blocks every candidate",
+			profile: ProviderProfile{Name: "openai-stored", CatalogID: "openai", APIKeyStored: true},
+			want:    nil,
+		},
+		{
+			// APIKeyEnv is NOT a configured credential for gating: the env var may be
+			// unset while the profile relies on an OAuth login, so candidates stand.
+			name:    "env-only profile still yields candidates",
+			profile: ProviderProfile{Name: "xai", CatalogID: "xai", APIKeyEnv: "XAI_API_KEY"},
+			want:    []string{"xai"},
+		},
+		{
+			name:    "empty profile yields no candidates",
+			profile: ProviderProfile{},
+			want:    nil,
+		},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			got := c.profile.OAuthLoginCandidates()
+			if strings.Join(got, ",") != strings.Join(c.want, ",") {
+				t.Fatalf("OAuthLoginCandidates() = %#v, want %#v", got, c.want)
+			}
+		})
+	}
+}

internal/config/max_turns_env_test.go

@@ -0,0 +1,56 @@
+package config
+
+import (
+	"os"
+	"testing"
+)
+
+// SetMaxTurnsEnv must export the per-run turn budget so a child process (whose
+// resolution reads the inherited environment via the nil-map -> os.Getenv path in
+// applyEnv) runs with the SAME budget the user set via /turns. n <= 0 is a no-op,
+// and a non-numeric/zero env value must not clobber the configured budget.
+func TestSetMaxTurnsEnvRoundTrips(t *testing.T) {
+	t.Setenv(MaxTurnsEnv, "") // register cleanup; restores the original after
+
+	if err := os.Unsetenv(MaxTurnsEnv); err != nil {
+		t.Fatalf("unset: %v", err)
+	}
+	SetMaxTurnsEnv(0) // non-positive must not set anything
+	if got := os.Getenv(MaxTurnsEnv); got != "" {
+		t.Fatalf("n<=0 set the env to %q, want unset", got)
+	}
+
+	SetMaxTurnsEnv(150)
+	if got := os.Getenv(MaxTurnsEnv); got != "150" {
+		t.Fatalf("env = %q, want 150", got)
+	}
+
+	// applyEnv with a nil map is the exact path a child uses: it falls back to
+	// os.Getenv, so the inherited value sets the budget.
+	cfg := &FileConfig{MaxTurns: defaultMaxTurns}
+	applyEnv(cfg, nil)
+	if cfg.MaxTurns != 150 {
+		t.Fatalf("applyEnv MaxTurns = %d, want 150 (inherited env not read)", cfg.MaxTurns)
+	}
+
+	// A garbage env value must leave the configured budget untouched.
+	if err := os.Setenv(MaxTurnsEnv, "not-a-number"); err != nil {
+		t.Fatalf("set: %v", err)
+	}
+	cfg2 := &FileConfig{MaxTurns: defaultMaxTurns}
+	applyEnv(cfg2, nil)
+	if cfg2.MaxTurns != defaultMaxTurns {
+		t.Fatalf("garbage env clobbered MaxTurns to %d, want default %d", cfg2.MaxTurns, defaultMaxTurns)
+	}
+
+	// An over-ceiling env value (e.g. a raw shell export that bypasses the /turns
+	// clamp) must be clamped to MaxTurnsCeiling at the read site.
+	if err := os.Setenv(MaxTurnsEnv, "999999"); err != nil {
+		t.Fatalf("set: %v", err)
+	}
+	cfg3 := &FileConfig{MaxTurns: defaultMaxTurns}
+	applyEnv(cfg3, nil)
+	if cfg3.MaxTurns != MaxTurnsCeiling {
+		t.Fatalf("applyEnv MaxTurns = %d, want clamped to ceiling %d", cfg3.MaxTurns, MaxTurnsCeiling)
+	}
+}

internal/config/paths.go

@@ -4,6 +4,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"runtime"
 	"strings"
 )
 
@@ -33,13 +34,32 @@ func DefaultResolveOptions(workspaceRoot string) (ResolveOptions, error) {
 }
 
 func DefaultUserConfigPath() (string, error) {
-	userConfigDir, err := os.UserConfigDir()
+	userConfigDir, err := UserConfigDir()
 	if err != nil {
 		return "", fmt.Errorf("resolve user config directory: %w", err)
 	}
 	return filepath.Join(userConfigDir, "zero", "config.json"), nil
 }
 
+// UserConfigDir returns the base directory Zero stores per-user config under.
+// It mirrors os.UserConfigDir everywhere except macOS: there Go defaults to
+// ~/Library/Application Support, but Zero deliberately uses ~/.config (XDG-style,
+// matching Linux and Claude Code) so a single config path works cross-platform.
+// $XDG_CONFIG_HOME still wins on macOS when set.
+func UserConfigDir() (string, error) {
+	if runtime.GOOS == "darwin" {
+		if xdg := strings.TrimSpace(os.Getenv("XDG_CONFIG_HOME")); xdg != "" {
+			return xdg, nil
+		}
+		home, err := os.UserHomeDir()
+		if err != nil {
+			return "", fmt.Errorf("resolve home directory: %w", err)
+		}
+		return filepath.Join(home, ".config"), nil
+	}
+	return os.UserConfigDir()
+}
+
 func existingConfigFile(path string) (string, error) {
 	info, err := os.Stat(path)
 	if err != nil {

internal/config/paths_test.go

@@ -105,7 +105,7 @@ func setUserConfigRoot(t *testing.T) string {
 		t.Setenv("XDG_CONFIG_HOME", root)
 	}
 
-	configRoot, err := os.UserConfigDir()
+	configRoot, err := UserConfigDir()
 	if err != nil {
 		t.Fatalf("UserConfigDir() error = %v", err)
 	}
@@ -130,3 +130,21 @@ func mkdirAll(t *testing.T, path string) {
 		t.Fatalf("create directory %s: %v", path, err)
 	}
 }
+
+func TestDefaultUserConfigPathUsesXDGConfigOnMacOS(t *testing.T) {
+	if runtime.GOOS != "darwin" {
+		t.Skip("macOS-specific config path behavior")
+	}
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("XDG_CONFIG_HOME", "")
+
+	path, err := DefaultUserConfigPath()
+	if err != nil {
+		t.Fatalf("DefaultUserConfigPath() error = %v", err)
+	}
+	want := filepath.Join(home, ".config", "zero", "config.json")
+	if path != want {
+		t.Fatalf("DefaultUserConfigPath() = %q, want %q", path, want)
+	}
+}

internal/config/resolver.go

@@ -2,9 +2,11 @@ package config
 
 import (
 	"encoding/json"
+	"errors"
 	"fmt"
 	"net/url"
 	"os"
+	"strconv"
 	"strings"
 
 	"github.com/Gitlawb/zero/internal/modelregistry"
@@ -13,7 +15,40 @@ import (
 	"github.com/Gitlawb/zero/internal/sandbox"
 )
 
-const defaultMaxTurns = 30
+// ErrNoActiveProvider marks a resolve failure caused solely by a missing or
+// unresolvable active provider. The interactive TUI treats this as "needs
+// onboarding" (drop into the setup wizard) rather than a fatal config error.
+var ErrNoActiveProvider = errors.New("no active provider configured")
+
+// ErrProviderRequiresModel marks a resolve failure caused solely by the active
+// provider missing a model with no catalog default to fall back on (custom
+// openai-/anthropic-compatible endpoints — Zero cannot guess a gateway's model).
+// Like ErrNoActiveProvider, the interactive TUI treats it as "needs onboarding"
+// and drops into the setup wizard so the user can fix it; headless commands
+// (zero config, zero exec) still fail with the actionable message.
+var ErrProviderRequiresModel = errors.New("provider requires model")
+
+// setupFixableError tags an error with a sentinel for errors.Is WITHOUT
+// changing its message: the sentinel's own text must not prefix what the user
+// sees (the wrapped message is already complete and actionable).
+type setupFixableError struct {
+	err      error
+	sentinel error
+}
+
+func (e *setupFixableError) Error() string { return e.err.Error() }
+
+func (e *setupFixableError) Unwrap() []error { return []error{e.err, e.sentinel} }
+
+// defaultMaxTurns is the per-run tool-turn budget when none is configured. 30 was
+// too low for real multi-step agentic work (agents ran out mid-task before reaching
+// later steps); 50 matches the old "deep" preset. Raise per-session with /turns.
+const defaultMaxTurns = 50
+
+// MaxTurnsCeiling caps the per-run tool-turn budget so a stray env value or typo
+// can't set an absurd ceiling. Shared between applyEnv (read site) and the /turns
+// command (write site) so the bound holds even if the env is set by a raw shell.
+const MaxTurnsCeiling = 500
 
 // defaultDeferThreshold is the number of deferred-eligible (MCP) tools at which
 // Zero collapses their full JSON schemas into compact `tool_search` reminder
@@ -186,6 +221,9 @@ func mergeConfig(dst *FileConfig, src FileConfig) {
 	if src.Preferences.Recaps != nil {
 		dst.Preferences.Recaps = src.Preferences.Recaps
 	}
+	if strings.TrimSpace(src.Preferences.Theme) != "" {
+		dst.Preferences.Theme = strings.TrimSpace(src.Preferences.Theme)
+	}
 	mergeLocalControlConfig(&dst.LocalControl, src.LocalControl)
 }
 
@@ -358,6 +396,18 @@ func effectiveProviderKind(profile ProviderProfile) ProviderKind {
 	return ""
 }
 
+// catalogDefaultModel returns the default model of a catalog provider ("" when
+// the id is unknown), used to fill a missing profile.Model for the official-API
+// kinds so a hand-written profile without a model resolves instead of bricking
+// every command that resolves config up front (zero config, bare zero setup).
+func catalogDefaultModel(catalogID string) string {
+	descriptor, ok := providercatalog.Get(catalogID)
+	if !ok {
+		return ""
+	}
+	return strings.TrimSpace(descriptor.DefaultModel)
+}
+
 func catalogDescriptorForProfile(profile ProviderProfile) (providercatalog.Descriptor, bool) {
 	catalogID := providercatalog.NormalizeID(profile.CatalogID)
 	if catalogID == "" {
@@ -431,11 +481,53 @@ func mergeProfile(base ProviderProfile, next ProviderProfile) ProviderProfile {
 	return base
 }
 
+// ActiveProviderEnv selects the active provider profile by name (read in applyEnv).
+const ActiveProviderEnv = "ZERO_PROVIDER"
+
+// SetActiveProviderEnv exports the active provider name to the process environment
+// so a spawned child process (which inherits the environment) resolves the SAME
+// provider profile — and therefore the same credentials (env key / stored key /
+// OAuth) — as its parent. Without this a sub-agent re-resolves config.json's
+// default provider and can land on one whose credentials don't match the parent's
+// live selection, failing auth the instant it spawns. A blank name CLEARS the
+// variable: switching back to an unnamed/default profile must not keep exporting a
+// stale provider to children.
+func SetActiveProviderEnv(name string) {
+	name = strings.TrimSpace(name)
+	if name == "" {
+		_ = os.Unsetenv(ActiveProviderEnv)
+		return
+	}
+	_ = os.Setenv(ActiveProviderEnv, name)
+}
+
+// MaxTurnsEnv overrides the per-run tool-turn budget by name (read in applyEnv).
+const MaxTurnsEnv = "ZERO_MAX_TURNS"
+
+// SetMaxTurnsEnv exports the per-run tool-turn budget to the process environment so
+// a spawned child (sub-agent / swarm member, which inherits the environment) runs
+// with the SAME budget the user set via /turns. Without it a child re-resolves
+// config.json's default and a large delegated task can exhaust its turns mid-run
+// (exit 4 / max-turns). No-op for n <= 0.
+func SetMaxTurnsEnv(n int) {
+	if n > 0 {
+		_ = os.Setenv(MaxTurnsEnv, strconv.Itoa(n))
+	}
+}
+
 func applyEnv(cfg *FileConfig, env map[string]string) {
-	activeProvider := strings.TrimSpace(envValue(env, "ZERO_PROVIDER"))
+	activeProvider := strings.TrimSpace(envValue(env, ActiveProviderEnv))
 	if activeProvider != "" {
 		cfg.ActiveProvider = activeProvider
 	}
+	if maxTurns := strings.TrimSpace(envValue(env, MaxTurnsEnv)); maxTurns != "" {
+		if n, err := strconv.Atoi(maxTurns); err == nil && n > 0 {
+			if n > MaxTurnsCeiling {
+				n = MaxTurnsCeiling
+			}
+			cfg.MaxTurns = n
+		}
+	}
 
 	applyProviderEnv(cfg, ProviderKindOpenAI, envProfile{
 		Name:    string(ProviderKindOpenAI),
@@ -677,11 +769,16 @@ func unionStrings(base []string, extra []string) []string {
 }
 
 type normalizeOptions struct {
-	defaultOpenAIModel bool
+	// defaultModels fills a missing profile.Model for the official-API provider
+	// kinds (openai from the model registry; anthropic/google from their catalog
+	// descriptors). Off for provider-command configs (command.go), which must
+	// surface exactly what the external command returned rather than inventing
+	// a model behind its back.
+	defaultModels bool
 }
 
 func normalizeProviders(providers []ProviderProfile, activeName string, envMaps ...map[string]string) ([]ProviderProfile, ProviderProfile, error) {
-	return normalizeProvidersWithOptions(providers, activeName, normalizeOptions{defaultOpenAIModel: true}, envMaps...)
+	return normalizeProvidersWithOptions(providers, activeName, normalizeOptions{defaultModels: true}, envMaps...)
 }
 
 func normalizeProvidersWithoutModelDefaults(providers []ProviderProfile, activeName string, envMaps ...map[string]string) ([]ProviderProfile, ProviderProfile, error) {
@@ -696,7 +793,7 @@ func normalizeProvidersWithOptions(providers []ProviderProfile, activeName strin
 	}
 	if len(providers) == 0 {
 		if activeName != "" {
-			return nil, ProviderProfile{}, fmt.Errorf("active provider %q not found", activeName)
+			return nil, ProviderProfile{}, fmt.Errorf("%w: active provider %q not found", ErrNoActiveProvider, activeName)
 		}
 		return []ProviderProfile{}, ProviderProfile{}, nil
 	}
@@ -728,10 +825,13 @@ func normalizeProvidersWithOptions(providers []ProviderProfile, activeName strin
 	}
 
 	if !activeFound {
-		return nil, ProviderProfile{}, fmt.Errorf("active provider %q not found", activeName)
+		return nil, ProviderProfile{}, fmt.Errorf("%w: active provider %q not found", ErrNoActiveProvider, activeName)
 	}
 	if active.Model == "" {
-		return nil, ProviderProfile{}, providerError(active, "provider %s requires model", active.Name)
+		return nil, ProviderProfile{}, &setupFixableError{
+			err:      providerError(active, "provider %s requires model — add \"model\" to its entry in config.json, or re-run: zero setup <catalog-id> --model <model>", active.Name),
+			sentinel: ErrProviderRequiresModel,
+		}
 	}
 
 	return normalized, active, nil
@@ -778,7 +878,7 @@ func normalizeProvider(profile ProviderProfile, env map[string]string, options n
 	case ProviderKindOpenAI:
 		if profile.BaseURL == "" || isOfficialOpenAIBaseURL(profile.BaseURL) {
 			profile.BaseURL = OpenAIBaseURL
-			if options.defaultOpenAIModel && profile.Model == "" {
+			if options.defaultModels && profile.Model == "" {
 				profile.Model = modelregistry.DefaultModelID
 			}
 			return profile, nil
@@ -795,6 +895,9 @@ func normalizeProvider(profile ProviderProfile, env map[string]string, options n
 	case ProviderKindAnthropic:
 		if profile.BaseURL == "" || isOfficialAnthropicBaseURL(profile.BaseURL) {
 			profile.BaseURL = AnthropicBaseURL
+			if options.defaultModels && profile.Model == "" {
+				profile.Model = catalogDefaultModel("anthropic")
+			}
 			return profile, nil
 		}
 		return ProviderProfile{}, providerError(profile, "anthropic provider %s requires official baseURL %s", profile.Name, AnthropicBaseURL)
@@ -809,6 +912,9 @@ func normalizeProvider(profile ProviderProfile, env map[string]string, options n
 	case ProviderKindGoogle:
 		if profile.BaseURL == "" || isOfficialGoogleBaseURL(profile.BaseURL) {
 			profile.BaseURL = GoogleBaseURL
+			if options.defaultModels && profile.Model == "" {
+				profile.Model = catalogDefaultModel("google")
+			}
 			return profile, nil
 		}
 		return ProviderProfile{}, providerError(profile, "google provider %s requires official baseURL %s", profile.Name, GoogleBaseURL)

internal/config/resolver_test.go

@@ -1,6 +1,7 @@
 package config
 
 import (
+	"errors"
 	"os"
 	"path/filepath"
 	"reflect"
@@ -883,8 +884,8 @@ func TestResolveAllowsNoConfiguredProviders(t *testing.T) {
 	if HasProviderProfile(resolved.Provider) {
 		t.Fatalf("Provider = %#v, want zero value", resolved.Provider)
 	}
-	if resolved.MaxTurns != 30 {
-		t.Fatalf("MaxTurns = %d, want 30", resolved.MaxTurns)
+	if resolved.MaxTurns != defaultMaxTurns {
+		t.Fatalf("MaxTurns = %d, want default %d", resolved.MaxTurns, defaultMaxTurns)
 	}
 }
 
@@ -898,6 +899,9 @@ func TestResolveRejectsActiveProviderWithoutConfiguredProfiles(t *testing.T) {
 	if !strings.Contains(err.Error(), `active provider "ghost" not found`) {
 		t.Fatalf("error = %q, want active provider missing message", err.Error())
 	}
+	if !errors.Is(err, ErrNoActiveProvider) {
+		t.Fatalf("error = %v, want errors.Is(err, ErrNoActiveProvider)", err)
+	}
 	if resolved.ActiveProvider != "" {
 		t.Fatalf("ActiveProvider = %q, want empty", resolved.ActiveProvider)
 	}
@@ -1551,3 +1555,97 @@ func TestResolveMaxTurnsZeroFallsBackToDefault(t *testing.T) {
 		t.Fatalf("MaxTurns = %d, want default %d", resolved.MaxTurns, defaultMaxTurns)
 	}
 }
+
+// The reported brick: a hand-written google profile with an apiKey but no
+// model made EVERY resolving command fail ("provider google requires model"),
+// including zero config and bare zero setup — the only commands that could
+// have fixed it. Official-API kinds now fall back to their catalog default
+// model, exactly like the openai kind always has.
+func TestResolveDefaultsGoogleModelFromCatalog(t *testing.T) {
+	path := writeConfig(t, `{
+		"activeProvider": "google",
+		"providers": [{"name": "google", "provider_kind": "google", "apiKey": "AIza-x"}]
+	}`)
+	resolved, err := Resolve(ResolveOptions{UserConfigPath: path, Env: map[string]string{}})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v, want model defaulted from the google catalog", err)
+	}
+	if resolved.Provider.Model != "gemini-2.5-pro" {
+		t.Fatalf("Model = %q, want the google catalog default gemini-2.5-pro", resolved.Provider.Model)
+	}
+	if resolved.Provider.BaseURL != GoogleBaseURL {
+		t.Fatalf("BaseURL = %q, want %q", resolved.Provider.BaseURL, GoogleBaseURL)
+	}
+}
+
+func TestResolveDefaultsAnthropicModelFromCatalog(t *testing.T) {
+	path := writeConfig(t, `{
+		"activeProvider": "anthropic",
+		"providers": [{"name": "anthropic", "provider_kind": "anthropic", "apiKey": "sk-ant-x"}]
+	}`)
+	resolved, err := Resolve(ResolveOptions{UserConfigPath: path, Env: map[string]string{}})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v, want model defaulted from the anthropic catalog", err)
+	}
+	if resolved.Provider.Model != "claude-sonnet-4.5" {
+		t.Fatalf("Model = %q, want the anthropic catalog default claude-sonnet-4.5", resolved.Provider.Model)
+	}
+}
+
+// An explicitly configured model must always win over the catalog default.
+func TestResolveKeepsExplicitGoogleModel(t *testing.T) {
+	path := writeConfig(t, `{
+		"activeProvider": "google",
+		"providers": [{"name": "google", "provider_kind": "google", "apiKey": "AIza-x", "model": "gemini-2.5-flash"}]
+	}`)
+	resolved, err := Resolve(ResolveOptions{UserConfigPath: path, Env: map[string]string{}})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+	if resolved.Provider.Model != "gemini-2.5-flash" {
+		t.Fatalf("Model = %q, want the explicitly configured gemini-2.5-flash", resolved.Provider.Model)
+	}
+}
+
+// Provider-command configs must NOT get invented models: the without-defaults
+// path surfaces exactly what the external command returned.
+func TestNormalizeWithoutModelDefaultsStillRequiresGoogleModel(t *testing.T) {
+	_, _, err := normalizeProvidersWithoutModelDefaults(
+		[]ProviderProfile{{Name: "google", ProviderKind: ProviderKindGoogle, APIKey: "AIza-x"}},
+		"google", map[string]string{})
+	if err == nil {
+		t.Fatal("provider-command path must still require an explicit model")
+	}
+	if !strings.Contains(err.Error(), "requires model") {
+		t.Fatalf("error = %q, want requires-model", err.Error())
+	}
+	// The residual error must tell the user how to fix it.
+	if !strings.Contains(err.Error(), "config.json") {
+		t.Fatalf("error = %q, want an actionable config.json hint", err.Error())
+	}
+}
+
+// The residual requires-model failure (custom endpoint, no catalog default)
+// must be tagged setup-fixable WITHOUT changing its message: the interactive
+// TUI matches the sentinel to drop into the wizard, while headless commands
+// print the exact actionable text.
+func TestResolveRequiresModelErrorIsSetupFixable(t *testing.T) {
+	path := writeConfig(t, `{
+		"activeProvider": "gw",
+		"providers": [{"name": "gw", "provider_kind": "openai-compatible", "baseURL": "https://gw.example/v1", "apiKey": "sk-x"}]
+	}`)
+	_, err := Resolve(ResolveOptions{UserConfigPath: path, Env: map[string]string{}})
+	if err == nil {
+		t.Fatal("expected requires-model error for a custom endpoint without model")
+	}
+	if !errors.Is(err, ErrProviderRequiresModel) {
+		t.Fatalf("error = %v, want errors.Is(err, ErrProviderRequiresModel)", err)
+	}
+	// The sentinel must NOT prefix the user-facing message.
+	if !strings.HasPrefix(err.Error(), "provider gw requires model") {
+		t.Fatalf("message = %q, want it to start with the provider error, not the sentinel text", err.Error())
+	}
+	if strings.Contains(err.Error(), "sk-x") {
+		t.Fatalf("error leaked API key: %q", err.Error())
+	}
+}

internal/config/types.go

@@ -97,6 +97,10 @@ type ToolsConfig struct {
 
 type PreferencesConfig struct {
 	FavoriteModels []string `json:"favoriteModels,omitempty"`
+	// Theme is the persisted TUI palette preference — "auto" or a registered theme
+	// name (e.g. "dracula"). Applied at startup below the --theme flag and
+	// ZERO_THEME, so a /theme choice survives restart. Empty = unset (defaults auto).
+	Theme string `json:"theme,omitempty"`
 	// Recaps is a tri-state: nil (unset) defaults to ON; an explicit false means
 	// the user turned post-turn recaps off. A *bool is its own tri-state, so no
 	// custom unmarshal is needed (unlike ToolsConfig.DeferThreshold's int).

internal/config/validate_test.go

@@ -66,10 +66,13 @@ func TestValidateFileRedactsSecretInIssue(t *testing.T) {
 }
 
 func TestValidateFileMissingModelWarns(t *testing.T) {
+	// An openai-compatible CUSTOM endpoint has no catalog default to fall back
+	// on, so a missing model is still a real issue — and the message must tell
+	// the user how to fix it.
 	path := writeValidateFixture(t, `{
 		"activeProvider": "main",
 		"providers": [
-			{"name": "main", "provider_kind": "anthropic"}
+			{"name": "main", "provider_kind": "openai-compatible", "baseURL": "https://gateway.example/v1"}
 		]
 	}`)
 
@@ -80,6 +83,30 @@ func TestValidateFileMissingModelWarns(t *testing.T) {
 	if !strings.Contains(issues[0].Message, "requires model") {
 		t.Fatalf("expected requires-model issue, got %#v", issues)
 	}
+	if !strings.Contains(issues[0].Message, "config.json") {
+		t.Fatalf("requires-model issue should carry an actionable hint, got %#v", issues)
+	}
+}
+
+func TestValidateFileDefaultsOfficialKindModels(t *testing.T) {
+	// Official-API kinds (anthropic/google) fall back to their catalog default
+	// model, so a hand-written model-less profile validates clean instead of
+	// bricking zero config / bare zero setup — the only commands that could
+	// have fixed it (the reported google case).
+	path := writeValidateFixture(t, `{
+		"activeProvider": "google",
+		"providers": [
+			{"name": "google", "provider_kind": "google", "apiKey": "AIza-x"},
+			{"name": "anthropic", "provider_kind": "anthropic", "apiKey": "sk-ant-x"}
+		]
+	}`)
+
+	_, issues := ValidateFile(path)
+	for _, issue := range issues {
+		if strings.Contains(issue.Message, "requires model") {
+			t.Fatalf("official-kind profiles must default their model, got issue %#v", issue)
+		}
+	}
 }
 
 func TestValidateFileValidConfigHasNoIssues(t *testing.T) {

internal/config/writer.go

@@ -154,6 +154,28 @@ func SetRecapsEnabled(path string, enabled bool) (FileConfig, error) {
 	return cfg, nil
 }
 
+// SetTheme persists the TUI theme preference, mirroring SetFavoriteModels
+// (read-modify-atomic-write). A blank theme clears the stored preference.
+func SetTheme(path string, theme string) (FileConfig, error) {
+	path = strings.TrimSpace(path)
+	if path == "" {
+		return FileConfig{}, fmt.Errorf("config path is required")
+	}
+	cfg := FileConfig{}
+	if data, err := os.ReadFile(path); err == nil {
+		if err := json.Unmarshal(data, &cfg); err != nil {
+			return FileConfig{}, fmt.Errorf("invalid config JSON %s: %w", path, err)
+		}
+	} else if !os.IsNotExist(err) {
+		return FileConfig{}, fmt.Errorf("read config %s: %w", path, err)
+	}
+	cfg.Preferences.Theme = strings.TrimSpace(theme)
+	if err := writeConfigFile(path, cfg); err != nil {
+		return FileConfig{}, err
+	}
+	return cfg, nil
+}
+
 func normalizeFavoriteModels(models []string) []string {
 	seen := map[string]bool{}
 	favorites := make([]string, 0, len(models))

internal/config/writer_test.go

@@ -272,6 +272,39 @@ func TestSetFavoriteModelsPersistsUserPreferences(t *testing.T) {
 	}
 }
 
+func TestSetThemePersistsUserPreference(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "zero.json")
+	writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "openai",
+		Providers: []ProviderProfile{
+			{Name: "openai", ProviderKind: ProviderKindOpenAI, Model: "gpt-4.1"},
+		},
+	}, 0o600)
+
+	cfg, err := SetTheme(path, "  dracula  ")
+	if err != nil {
+		t.Fatalf("SetTheme() error = %v", err)
+	}
+	if cfg.Preferences.Theme != "dracula" {
+		t.Fatalf("Theme = %q, want dracula (trimmed)", cfg.Preferences.Theme)
+	}
+	persisted := readConfigFixture(t, path)
+	if persisted.Preferences.Theme != "dracula" {
+		t.Fatalf("persisted Theme = %q, want dracula", persisted.Preferences.Theme)
+	}
+	if persisted.ActiveProvider != "openai" || len(persisted.Providers) != 1 {
+		t.Fatalf("provider config was not preserved by SetTheme: %#v", persisted)
+	}
+
+	// A blank value clears the stored preference.
+	if cfg, err = SetTheme(path, ""); err != nil {
+		t.Fatalf("SetTheme(\"\") error = %v", err)
+	}
+	if cfg.Preferences.Theme != "" {
+		t.Fatalf("SetTheme(\"\") should clear the theme, got %q", cfg.Preferences.Theme)
+	}
+}
+
 func TestRecapsPreferenceRoundTrips(t *testing.T) {
 	// Default (unset) is ON.
 	if !(PreferencesConfig{}).RecapsEnabled() {

internal/mcp/client_test.go

@@ -1,11 +1,11 @@
 package mcp
 
 import (
-	"bytes"
 	"context"
 	"encoding/json"
 	"errors"
 	"fmt"
+	"io"
 	"net/http"
 	"net/http/httptest"
 	"os"
@@ -393,28 +393,43 @@ func TestConnectRejectsUnsupportedTransport(t *testing.T) {
 }
 
 func TestClientRequestWaitsForMatchingResponseID(t *testing.T) {
-	var incoming bytes.Buffer
-	incomingWriter := newMessageWriter(&incoming)
-	if err := incomingWriter.write(rpcMessage{Method: "notifications/progress"}); err != nil {
-		t.Fatal(err)
-	}
-	if err := incomingWriter.write(rpcMessage{
-		ID:    99,
-		Error: &rpcError{Code: -32000, Message: "wrong response"},
-	}); err != nil {
-		t.Fatal(err)
-	}
-	if err := incomingWriter.write(rpcMessage{
-		ID:     1,
-		Result: mustRaw(map[string]any{"value": "matched"}),
-	}); err != nil {
-		t.Fatal(err)
-	}
+	// Pipes, not a pre-filled bytes.Buffer: the fake server responds only AFTER
+	// reading the outgoing request, matching production ordering. With responses
+	// pre-loaded, the client's reader goroutine could consume the matching
+	// response — and hit EOF — before request() registered its pending id,
+	// flaking as "request() error = EOF" on fast runners.
+	serverReader, clientWriter := io.Pipe() // client → server
+	clientReader, serverWriter := io.Pipe() // server → client
+	t.Cleanup(func() {
+		_ = clientWriter.Close()
+		_ = serverWriter.Close()
+	})
+
+	serverErr := make(chan error, 1)
+	go func() {
+		requests := newMessageReader(serverReader)
+		request, err := requests.read()
+		if err != nil {
+			serverErr <- err
+			return
+		}
+		responses := newMessageWriter(serverWriter)
+		for _, message := range []rpcMessage{
+			{Method: "notifications/progress"},
+			{ID: 99, Error: &rpcError{Code: -32000, Message: "wrong response"}},
+			{ID: request.ID, Result: mustRaw(map[string]any{"value": "matched"})},
+		} {
+			if err := responses.write(message); err != nil {
+				serverErr <- err
+				return
+			}
+		}
+		serverErr <- nil
+	}()
 
-	var outgoing bytes.Buffer
 	client := &Client{
-		reader: newMessageReader(&incoming),
-		writer: newMessageWriter(&outgoing),
+		reader: newMessageReader(clientReader),
+		writer: newMessageWriter(clientWriter),
 		nextID: 1,
 	}
 	var result struct {
@@ -426,6 +441,9 @@ func TestClientRequestWaitsForMatchingResponseID(t *testing.T) {
 	if result.Value != "matched" {
 		t.Fatalf("result.Value = %q, want matched response", result.Value)
 	}
+	if err := <-serverErr; err != nil {
+		t.Fatalf("fake server error: %v", err)
+	}
 }
 
 type errorCloser struct {

internal/oauth/manager.go

@@ -44,10 +44,17 @@ type Manager struct {
 
 // ManagerOptions configures a Manager.
 type ManagerOptions struct {
-	Store         *Store
-	Registry      *Registry
-	HTTPClient    *http.Client
-	Env           map[string]string
+	Store      *Store
+	Registry   *Registry
+	HTTPClient *http.Client
+	Env        map[string]string
+	// AllowPresets opts this manager into the baked-in OAuth presets without the
+	// operator exporting ZERO_OAUTH_ALLOW_PRESETS — used by the interactive wizard
+	// and CLI login (and the runtime token refresh) for a provider the user chose
+	// to sign into, whose preset client identity is public (e.g. xAI). It layers the
+	// flag onto Env (or the process environment when Env is nil), preserving any
+	// ZERO_OAUTH_<NAME>_* overrides. Leave false for hermetic tests.
+	AllowPresets  bool
 	Now           func() time.Time
 	RefreshBuffer time.Duration
 	Out           io.Writer
@@ -83,9 +90,13 @@ func NewManager(opts ManagerOptions) (*Manager, error) {
 	if open == nil {
 		open = func(string) error { return nil }
 	}
+	env := opts.Env
+	if opts.AllowPresets {
+		env = envWithPresetsAllowed(env)
+	}
 	return &Manager{
 		store: opts.Store, registry: registry, client: client,
-		env: opts.Env, now: now, buffer: buffer, out: out, openBrowser: open,
+		env: env, now: now, buffer: buffer, out: out, openBrowser: open,
 	}, nil
 }
 

internal/oauth/manager_test.go

@@ -52,6 +52,30 @@ func managerFor(t *testing.T, env map[string]string, openBrowser func(string) er
 	return m
 }
 
+// AllowPresets layers the opt-in onto the manager's env so a login/refresh for a
+// preset provider resolves; without it (and no Env) the env stays nil so callers
+// and tests remain hermetic.
+func TestNewManagerAllowPresetsForcesOptIn(t *testing.T) {
+	store, err := NewStore(StoreOptions{FilePath: filepath.Join(t.TempDir(), "tok.json")})
+	if err != nil {
+		t.Fatalf("NewStore: %v", err)
+	}
+	m, err := NewManager(ManagerOptions{Store: store, AllowPresets: true})
+	if err != nil {
+		t.Fatalf("NewManager: %v", err)
+	}
+	if m.env["ZERO_OAUTH_ALLOW_PRESETS"] != "1" {
+		t.Fatalf("AllowPresets should force the opt-in; got env[%q]=%q", "ZERO_OAUTH_ALLOW_PRESETS", m.env["ZERO_OAUTH_ALLOW_PRESETS"])
+	}
+	m2, err := NewManager(ManagerOptions{Store: store})
+	if err != nil {
+		t.Fatalf("NewManager: %v", err)
+	}
+	if m2.env != nil {
+		t.Fatalf("env should stay nil without AllowPresets, got %v", m2.env)
+	}
+}
+
 func TestManagerLoginLoopback(t *testing.T) {
 	fp := newFakeProvider(t, `{"access_token":"at","refresh_token":"rt","token_type":"Bearer","expires_in":3600}`)
 	env := map[string]string{

internal/oauth/presets.go

@@ -1,6 +1,33 @@
 package oauth
 
-import "strings"
+import (
+	"os"
+	"strings"
+)
+
+// envWithPresetsAllowed returns an env map that opts into the baked-in presets
+// (as if ZERO_OAUTH_ALLOW_PRESETS=1 were exported) so an interactive wizard/CLI
+// login for a well-known public client (e.g. xAI) can use the preset without the
+// operator setting the flag themselves. It copies base — or snapshots the process
+// environment when base is nil — because envValue treats a non-nil map as hermetic
+// (a missing key does NOT fall back to os.Getenv), so a partial map would silently
+// drop the operator's ZERO_OAUTH_<NAME>_* overrides. The flag is then forced on.
+func envWithPresetsAllowed(base map[string]string) map[string]string {
+	env := make(map[string]string, len(base)+1)
+	if base == nil {
+		for _, kv := range os.Environ() {
+			if eq := strings.IndexByte(kv, '='); eq > 0 {
+				env[kv[:eq]] = kv[eq+1:]
+			}
+		}
+	} else {
+		for k, v := range base {
+			env[k] = v
+		}
+	}
+	env["ZERO_OAUTH_ALLOW_PRESETS"] = "1"
+	return env
+}
 
 // providerPreset is a baked-in default OAuth configuration for a well-known
 // provider. Every field is overridable per provider via ZERO_OAUTH_<NAME>_*

internal/oauth/presets_test.go

@@ -46,6 +46,46 @@ func TestResolveConfigUsesXAIPreset(t *testing.T) {
 	}
 }
 
+// envWithPresetsAllowed lets a caller (the wizard / CLI login / runtime refresh)
+// opt into the preset without exporting ZERO_OAUTH_ALLOW_PRESETS: an otherwise-inert
+// xAI config now resolves from the baked-in preset.
+func TestEnvWithPresetsAllowedEnablesPreset(t *testing.T) {
+	r := NewRegistry()
+	// Baseline: a hermetic empty map keeps the preset inert.
+	if _, _, err := r.ResolveConfig("xai", map[string]string{}); err == nil {
+		t.Fatal("xai should not resolve without the opt-in")
+	}
+	// The helper forces the opt-in so the preset resolves.
+	cfg, _, err := r.ResolveConfig("xai", envWithPresetsAllowed(map[string]string{}))
+	if err != nil {
+		t.Fatalf("envWithPresetsAllowed should enable the xai preset: %v", err)
+	}
+	if cfg.ClientID != "b1a00492-073a-47ea-816f-4c329264a828" {
+		t.Fatalf("client_id = %q", cfg.ClientID)
+	}
+}
+
+// The helper copies the base map (never mutating it) and keeps ZERO_OAUTH_<NAME>_*
+// overrides — critical because envValue treats a non-nil map as hermetic, so a
+// partial map would silently drop them.
+func TestEnvWithPresetsAllowedPreservesOverrides(t *testing.T) {
+	base := map[string]string{"ZERO_OAUTH_XAI_CLIENT_ID": "custom-id"}
+	env := envWithPresetsAllowed(base)
+	if env["ZERO_OAUTH_ALLOW_PRESETS"] != "1" {
+		t.Fatalf("opt-in flag not set: %v", env)
+	}
+	if _, mutated := base["ZERO_OAUTH_ALLOW_PRESETS"]; mutated {
+		t.Fatal("envWithPresetsAllowed mutated the caller's base map")
+	}
+	cfg, _, err := NewRegistry().ResolveConfig("xai", env)
+	if err != nil {
+		t.Fatalf("ResolveConfig with override env: %v", err)
+	}
+	if cfg.ClientID != "custom-id" {
+		t.Fatalf("override dropped through the helper: client_id = %q", cfg.ClientID)
+	}
+}
+
 func TestResolveConfigEnvOverridesPreset(t *testing.T) {
 	r := NewRegistry()
 	env := map[string]string{

internal/oauth/store.go

@@ -41,6 +41,26 @@ func ValidateKey(key string) error {
 // ProviderKey builds the store key for a provider login.
 func ProviderKey(name string) string { return KeyPrefixProvider + name }
 
+// FirstStored returns the token and its ProviderKey for the FIRST candidate name
+// that has a token in the store, with ok=false when none do. Callers pass
+// ProviderProfile.OAuthLoginCandidates() so that everything derived from a login
+// — the bearer token AND any header claim like chatgpt-account-id — comes from
+// the SAME login; selecting independently per consumer could otherwise pair a
+// bearer from one login with an account header from another. A load error on a
+// candidate is treated as a miss (skip to the next), never a hard failure.
+func FirstStored(store *Store, candidates []string) (Token, string, bool) {
+	if store == nil {
+		return Token{}, "", false
+	}
+	for _, name := range candidates {
+		key := ProviderKey(name)
+		if token, ok, err := store.Load(key); err == nil && ok {
+			return token, key, true
+		}
+	}
+	return Token{}, "", false
+}
+
 // Status is a redaction-safe summary of a stored token (no secret material).
 type Status struct {
 	Key             string    `json:"key"`

internal/providermodeldiscovery/discovery.go

@@ -85,6 +85,104 @@ func Discover(ctx context.Context, profile config.ProviderProfile, options Optio
 	}
 }
 
+// DiscoverOllamaContextWindow asks a local Ollama daemon for a model's context
+// length via its native /api/show endpoint. The generic /v1/models probe
+// (parseModelsResponse) only ever returns id/description — OpenAI-compatible
+// listings don't carry context-window metadata, and a custom/local model tag
+// (e.g. a user's own Ollama pull, including a ":cloud"-tagged model proxied
+// through the local daemon) has no curated-catalog entry to borrow one from
+// either, so modelContextWindow has nothing to show for it without this. Only
+// meaningful for the local Ollama provider (baseURL like
+// http://localhost:11434/v1) — Ollama Cloud's hosted API is a different
+// service and isn't assumed to expose the same endpoint.
+func DiscoverOllamaContextWindow(ctx context.Context, baseURL string, model string, options Options) (int, error) {
+	model = strings.TrimSpace(model)
+	if model == "" {
+		return 0, fmt.Errorf("model name is required")
+	}
+	endpoint, err := ollamaShowEndpoint(baseURL)
+	if err != nil {
+		return 0, err
+	}
+	payload, err := json.Marshal(struct {
+		Model string `json:"model"`
+	}{Model: model})
+	if err != nil {
+		return 0, err
+	}
+	request, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, strings.NewReader(string(payload)))
+	if err != nil {
+		return 0, err
+	}
+	request.Header.Set("Content-Type", "application/json")
+	request.Header.Set("Accept", "application/json")
+
+	client := options.HTTPClient
+	if client == nil {
+		client = &http.Client{Timeout: 10 * time.Second}
+	}
+	response, err := client.Do(request)
+	if err != nil {
+		return 0, err
+	}
+	defer response.Body.Close()
+
+	body, err := io.ReadAll(io.LimitReader(response.Body, 1<<20))
+	if err != nil {
+		return 0, err
+	}
+	if response.StatusCode < 200 || response.StatusCode >= 300 {
+		return 0, fmt.Errorf("ollama show endpoint returned %s: %s", response.Status, strings.TrimSpace(string(body)))
+	}
+	return parseOllamaShowContextWindow(body)
+}
+
+// ollamaShowEndpoint derives the native Ollama API root from the
+// OpenAI-compatible base URL Zero stores for the provider (".../v1") and
+// appends /api/show.
+func ollamaShowEndpoint(baseURL string) (string, error) {
+	baseURL = strings.TrimSpace(baseURL)
+	if baseURL == "" {
+		return "", fmt.Errorf("provider base URL is required for ollama model discovery")
+	}
+	parsed, err := url.Parse(baseURL)
+	if err != nil || parsed.Scheme == "" || parsed.Host == "" {
+		return "", fmt.Errorf("invalid provider base URL %q", baseURL)
+	}
+	path := strings.TrimRight(parsed.Path, "/")
+	path = strings.TrimSuffix(path, "/v1")
+	parsed.Path = strings.TrimRight(path, "/") + "/api/show"
+	parsed.RawQuery = ""
+	parsed.Fragment = ""
+	return parsed.String(), nil
+}
+
+// parseOllamaShowContextWindow extracts the context length from an
+// /api/show response. Ollama reports it under model_info with an
+// architecture-prefixed key (e.g. "llama.context_length",
+// "qwen2.context_length") — the prefix varies by model family, so this scans
+// for any key ending in ".context_length" rather than hardcoding one.
+func parseOllamaShowContextWindow(body []byte) (int, error) {
+	// model_info mixes value types (strings like "general.architecture", numbers
+	// like "*.context_length"), so decode values generically and only try to
+	// read a number out of the one key we actually care about.
+	var payload struct {
+		ModelInfo map[string]any `json:"model_info"`
+	}
+	if err := json.Unmarshal(body, &payload); err != nil {
+		return 0, fmt.Errorf("decode ollama show response: %w", err)
+	}
+	for key, value := range payload.ModelInfo {
+		if !strings.HasSuffix(key, ".context_length") {
+			continue
+		}
+		if window, ok := value.(float64); ok && window > 0 {
+			return int(window), nil
+		}
+	}
+	return 0, fmt.Errorf("ollama show response did not report a context length")
+}
+
 func discoverOpenAIModels(ctx context.Context, profile config.ProviderProfile, options Options) ([]Model, error) {
 	endpoint, err := modelsEndpoint(profile.BaseURL)
 	if err != nil {

internal/providermodeldiscovery/discovery_test.go

@@ -2,6 +2,7 @@ package providermodeldiscovery
 
 import (
 	"context"
+	"io"
 	"net/http"
 	"net/http/httptest"
 	"strings"
@@ -267,3 +268,61 @@ func modelIDs(models []Model) []string {
 	}
 	return ids
 }
+
+// TestDiscoverOllamaContextWindowFetchesFromNativeShowEndpoint: the generic
+// /v1/models probe never carries context-window metadata (parseModelsResponse
+// only extracts id/description), so a custom/local Ollama model tag with no
+// curated-catalog match has no other source for it. This exercises the
+// Ollama-native /api/show fallback that fills that gap.
+func TestDiscoverOllamaContextWindowFetchesFromNativeShowEndpoint(t *testing.T) {
+	var gotPath, gotMethod, gotBody string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		gotMethod = r.Method
+		body, _ := io.ReadAll(r.Body)
+		gotBody = string(body)
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{
+			"model_info": {
+				"general.architecture": "qwen2",
+				"qwen2.context_length": 131072
+			}
+		}`))
+	}))
+	defer server.Close()
+
+	window, err := DiscoverOllamaContextWindow(context.Background(), server.URL+"/v1", "kimi-k2.7-code:cloud", Options{HTTPClient: server.Client()})
+	if err != nil {
+		t.Fatalf("DiscoverOllamaContextWindow returned error: %v", err)
+	}
+	if window != 131072 {
+		t.Fatalf("context window = %d, want 131072", window)
+	}
+	if gotPath != "/api/show" {
+		t.Fatalf("requested path = %q, want /api/show (not under /v1)", gotPath)
+	}
+	if gotMethod != http.MethodPost {
+		t.Fatalf("method = %q, want POST", gotMethod)
+	}
+	if !strings.Contains(gotBody, `"kimi-k2.7-code:cloud"`) {
+		t.Fatalf("request body = %q, want it to name the model", gotBody)
+	}
+}
+
+func TestDiscoverOllamaContextWindowRequiresModelName(t *testing.T) {
+	if _, err := DiscoverOllamaContextWindow(context.Background(), "http://localhost:11434/v1", "", Options{}); err == nil {
+		t.Fatal("expected an error for an empty model name")
+	}
+}
+
+func TestDiscoverOllamaContextWindowErrorsWhenShowOmitsContextLength(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"model_info": {"general.architecture": "qwen2"}}`))
+	}))
+	defer server.Close()
+
+	if _, err := DiscoverOllamaContextWindow(context.Background(), server.URL+"/v1", "some-model", Options{HTTPClient: server.Client()}); err == nil {
+		t.Fatal("expected an error when no *.context_length key is present")
+	}
+}

internal/providers/anthropic/provider.go

@@ -211,11 +211,11 @@ func (provider *Provider) stream(ctx context.Context, body []byte, events chan<-
 	err = providerio.ScanSSEDataWithContext(streamCtx, cancelStream, response.Body, provider.streamIdleTimeout, func(data string) bool {
 		return provider.emitPayload(ctx, data, state, events)
 	})
-	if errors.Is(err, providerio.ErrStreamIdle) {
+	if errors.Is(err, providerio.ErrStreamIdle) || errors.Is(err, providerio.ErrStreamStalled) {
 		state.closeOpen(ctx, events)
 		providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{
 			Type:  zeroruntime.StreamEventError,
-			Error: provider.redact(fmt.Sprintf("provider stream error: idle timeout after %s (upstream stopped sending data)", provider.streamIdleTimeout)),
+			Error: provider.redact("provider stream error: " + providerio.StreamTimeoutMessage(err, provider.streamIdleTimeout)),
 		})
 		return
 	}

internal/providers/factory.go

@@ -26,6 +26,13 @@ type Options struct {
 	// OAuth bearer token (preferred over the API key). nil => API-key auth only.
 	// Applied to the OpenAI and Anthropic providers.
 	OAuthResolver providerio.TokenResolver
+	// OAuthLoginKey is the credential-store key OAuthResolver bound to (empty when
+	// there is no OAuth login). It is the SAME selection the bearer resolver made,
+	// resolved ONCE by the caller so the Codex chatgpt-account-id header reads its
+	// account from the exact login that issued the bearer token — never a second,
+	// independent lookup that could pick a different login (a backend-rejected
+	// mismatch).
+	OAuthLoginKey string
 }
 
 // New creates a runtime provider for a resolved provider profile.
@@ -251,14 +258,22 @@ func isCodexCatalog(profile config.ProviderProfile, _ resolvedProfile) bool {
 
 // newCodexProvider builds a Codex-flavored openai provider for the chatgpt
 // catalog. The Codex headers (`originator`, `chatgpt-account-id`) are
-// injected by the CodexProvider wrapper. The `chatgpt-account-id` is
-// resolved dynamically from the stored OAuth token's Account field on
-// every request so a refresh that rotates the token (and its account
-// claim) takes effect immediately — a static AccountID captured at
-// construction would go stale on the first refresh.
+// injected by the CodexProvider wrapper. The `chatgpt-account-id` is read from
+// the stored OAuth token's Account field on every request so a refresh that
+// rotates the token (and its account claim) takes effect immediately — a static
+// AccountID captured at construction would go stale on the first refresh.
+//
+// The *login* the header reads from is options.OAuthLoginKey — the SAME key the
+// bearer resolver bound, resolved ONCE by the caller. The bearer resolver
+// refreshes the token in place under that key; reading the account from the same
+// key keeps header and token on one login for the provider's life. Resolving the
+// key independently here (a second oauth.FirstStored) could, after a transient
+// per-candidate load error or a mid-session `zero auth login`, pick a different
+// login than the bearer — a mismatch the backend rejects.
 func newCodexProvider(profile config.ProviderProfile, resolved resolvedProfile, options Options) (zeroruntime.Provider, error) {
+	accountKey := options.OAuthLoginKey
 	resolver := openai.CodexAccountResolver(func(ctx context.Context) (string, bool, error) {
-		account := codexAccountFromStore(profile.Name)
+		account := codexAccountForKey(accountKey)
 		if account == "" {
 			return "", false, nil
 		}
@@ -288,17 +303,20 @@ func newCodexProvider(profile config.ProviderProfile, resolved resolvedProfile,
 	})
 }
 
-// codexAccountFromStore reads the chatgpt_account_id from the stored OAuth
-// token for the given provider name. It is called per-request (via the
-// resolver) so a refresh that rotates the token takes effect immediately
-// without restarting the agent. Returns "" when no token is stored (the
-// Codex provider then omits the header and the user sees a clear 401).
-func codexAccountFromStore(providerName string) string {
+// codexAccountForKey reads the chatgpt_account_id from the token currently
+// stored under key. Called per-request (via the resolver) so a refresh that
+// rotates the token — saved back in place under the same key — takes effect
+// immediately without restarting the agent. Returns "" when key is empty (no
+// OAuth login), or the token is absent or carries no account claim.
+func codexAccountForKey(key string) string {
+	if key == "" {
+		return ""
+	}
 	store, err := oauth.NewStore(oauth.StoreOptions{})
 	if err != nil {
 		return ""
 	}
-	token, ok, err := store.Load(oauth.ProviderKey(providerName))
+	token, ok, err := store.Load(key)
 	if err != nil || !ok {
 		return ""
 	}

internal/providers/factory_test.go

@@ -264,6 +264,14 @@ func TestNewRejectsUnsupportedProviderKind(t *testing.T) {
 }
 
 func TestNewRoutesChatGPTCatalogToCodexProvider(t *testing.T) {
+	// Isolate the OAuth token store to an empty temp path so the factory reads no
+	// stored login — otherwise this test picks up the developer's real chatgpt
+	// OAuth token and the "want empty chatgpt-account-id" assertion fails locally
+	// (it still passes in CI, where no login is stored). Mirrors the isolation in
+	// TestNewRoutesChatGPTCatalogWithStoredAccountID.
+	t.Setenv("ZERO_OAUTH_STORAGE", "file")
+	t.Setenv("ZERO_OAUTH_TOKENS_PATH", t.TempDir()+"/tokens.json")
+
 	transport := &captureTransport{
 		responseBody: "data: [DONE]\n\n",
 	}
@@ -315,11 +323,10 @@ func TestNewRoutesChatGPTCatalogToCodexProvider(t *testing.T) {
 
 func TestNewRoutesChatGPTCatalogWithStoredAccountID(t *testing.T) {
 	// The factory reads the stored OAuth token's Account field for the
-	// chatgpt-account-id header. Seed a token in a temp store and point
-	// ZERO_OAUTH_TOKENS_PATH at it so the factory picks it up.
-	dir := t.TempDir()
-	t.Setenv("ZERO_OAUTH_TOKENS_PATH", dir+"/tokens.json")
-	store, err := newOAuthStoreForTest()
+	// chatgpt-account-id header, from the login key the CALLER supplies in
+	// Options.OAuthLoginKey (the same key the bearer resolver bound). Seed a
+	// token in an isolated temp store, then pass that key.
+	store, err := newOAuthStoreForTest(t)
 	if err != nil {
 		t.Fatalf("NewStore: %v", err)
 	}
@@ -337,8 +344,9 @@ func TestNewRoutesChatGPTCatalogWithStoredAccountID(t *testing.T) {
 		CatalogID: "chatgpt",
 		Model:     "gpt-5",
 	}, Options{
-		HTTPClient: &http.Client{Transport: transport},
-		UserAgent:  "zero-factory-test",
+		HTTPClient:    &http.Client{Transport: transport},
+		UserAgent:     "zero-factory-test",
+		OAuthLoginKey: oauth.ProviderKey("chatgpt"),
 	})
 	if err != nil {
 		t.Fatalf("New() error = %v", err)
@@ -400,10 +408,61 @@ func (transport *captureTransport) body() io.Reader {
 	return strings.NewReader(transport.requestBody)
 }
 
-// newOAuthStoreForTest builds a Store pointed at the current
-// ZERO_OAUTH_TOKENS_PATH (set by the caller). It exists so the chatgpt
-// factory tests can seed a token without copying the path-handling dance
-// from internal/cli.
-func newOAuthStoreForTest() (*oauth.Store, error) {
+// newOAuthStoreForTest pins the OAuth token store to a plain temp FILE and
+// returns a Store on it. Pinning ZERO_OAUTH_STORAGE matters as much as the
+// path: an inherited "keyring" value would send NewStore to the OS keychain
+// and ignore ZERO_OAUTH_TOKENS_PATH entirely, making the test read/write the
+// developer's real logins. Exists so the chatgpt factory tests can seed a
+// token without copying the path-handling dance from internal/cli.
+func newOAuthStoreForTest(t *testing.T) (*oauth.Store, error) {
+	t.Helper()
+	t.Setenv("ZERO_OAUTH_STORAGE", "file")
+	t.Setenv("ZERO_OAUTH_TOKENS_PATH", t.TempDir()+"/tokens.json")
 	return oauth.NewStore(oauth.StoreOptions{})
 }
+
+// codexAccountForKey reads the chatgpt-account-id from the token stored under a
+// FIXED key — the key the caller (cli.oauthLoginForProfile) already bound for the
+// bearer token, passed through providers.Options.OAuthLoginKey. The account is
+// therefore always read from the same login that issued the bearer (no second,
+// independent selection). It re-reads per call, so an in-place token refresh (new
+// account claim under the SAME key) is picked up; an empty key (no OAuth login)
+// or a missing/account-less token yields "".
+func TestCodexAccountForKey(t *testing.T) {
+	store, err := newOAuthStoreForTest(t)
+	if err != nil {
+		t.Fatalf("NewStore: %v", err)
+	}
+	if err := store.Save(oauth.ProviderKey("chatgpt"), oauth.Token{AccessToken: "tok-1", Account: "acc-catalog-7"}); err != nil {
+		t.Fatalf("Save chatgpt: %v", err)
+	}
+	// A different login WITHOUT an account claim, to prove the read stays on the
+	// requested key and never crosses to another login's account.
+	if err := store.Save(oauth.ProviderKey("codex"), oauth.Token{AccessToken: "tok-codex"}); err != nil {
+		t.Fatalf("Save codex: %v", err)
+	}
+
+	if got := codexAccountForKey(oauth.ProviderKey("chatgpt")); got != "acc-catalog-7" {
+		t.Fatalf("account = %q, want acc-catalog-7", got)
+	}
+	// The bound key's token has no account → "", NOT the other login's account.
+	if got := codexAccountForKey(oauth.ProviderKey("codex")); got != "" {
+		t.Fatalf("account = %q, want empty (must not cross to another login's account)", got)
+	}
+	// Empty key (no OAuth login) → header omitted.
+	if got := codexAccountForKey(""); got != "" {
+		t.Fatalf("account for empty key = %q, want empty", got)
+	}
+	// Unknown key → "".
+	if got := codexAccountForKey(oauth.ProviderKey("nope")); got != "" {
+		t.Fatalf("account for unknown key = %q, want empty", got)
+	}
+
+	// An in-place refresh under the bound key IS reflected per request.
+	if err := store.Save(oauth.ProviderKey("chatgpt"), oauth.Token{AccessToken: "tok-refreshed", Account: "acc-rotated"}); err != nil {
+		t.Fatalf("refresh chatgpt: %v", err)
+	}
+	if got := codexAccountForKey(oauth.ProviderKey("chatgpt")); got != "acc-rotated" {
+		t.Fatalf("account = %q, want acc-rotated (in-place refresh must be picked up)", got)
+	}
+}

internal/providers/gemini/provider.go

@@ -158,10 +158,10 @@ func (provider *Provider) stream(ctx context.Context, body []byte, events chan<-
 	err = providerio.ScanSSEDataWithContext(streamCtx, cancelStream, response.Body, provider.streamIdleTimeout, func(data string) bool {
 		return provider.emitPayload(ctx, data, &state, events)
 	})
-	if errors.Is(err, providerio.ErrStreamIdle) {
+	if errors.Is(err, providerio.ErrStreamIdle) || errors.Is(err, providerio.ErrStreamStalled) {
 		providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{
 			Type:  zeroruntime.StreamEventError,
-			Error: provider.redact(fmt.Sprintf("provider stream error: idle timeout after %s (upstream stopped sending data)", provider.streamIdleTimeout)),
+			Error: provider.redact("provider stream error: " + providerio.StreamTimeoutMessage(err, provider.streamIdleTimeout)),
 		})
 		return
 	}
@@ -342,7 +342,7 @@ func (provider *Provider) geminiRequest(request zeroruntime.CompletionRequest) (
 			declarations = append(declarations, geminiFunctionDeclaration{
 				Name:        tool.Name,
 				Description: tool.Description,
-				Parameters:  tool.Parameters,
+				Parameters:  sanitizeGeminiSchema(tool.Parameters),
 			})
 		}
 		mapped.Tools = []geminiToolGroup{{FunctionDeclarations: declarations}}
@@ -350,6 +350,71 @@ func (provider *Provider) geminiRequest(request zeroruntime.CompletionRequest) (
 	return mapped, nil
 }
 
+// geminiSchemaFields is the subset of JSON Schema keywords Google's
+// functionDeclarations[].parameters accepts (the Generative AI Schema type).
+// Anything outside it — most notably OpenAI's `additionalProperties`, which Zero
+// emits on every tool's parameters and which Gemini 400s on ("Unknown name
+// \"additionalProperties\" … Cannot find field") — must be dropped before the
+// request goes out. Kept as an allowlist rather than a denylist so a schema
+// keyword Zero (or an MCP server) adds later can't silently leak an unsupported
+// field into the Gemini payload.
+var geminiSchemaFields = map[string]bool{
+	"type": true, "format": true, "title": true, "description": true,
+	"nullable": true, "enum": true, "items": true, "properties": true,
+	"required": true, "anyOf": true, "propertyOrdering": true, "default": true,
+	"minimum": true, "maximum": true, "minItems": true, "maxItems": true,
+	"minLength": true, "maxLength": true, "minProperties": true,
+	"maxProperties": true, "pattern": true, "example": true,
+}
+
+// sanitizeGeminiSchema returns a copy of a JSON-Schema map keeping only the
+// keywords Gemini supports (geminiSchemaFields), recursing through the nested
+// schemas under `properties`, `items`, and `anyOf`. Returns nil for a nil input
+// so a parameterless tool stays parameterless.
+func sanitizeGeminiSchema(schema map[string]any) map[string]any {
+	if schema == nil {
+		return nil
+	}
+	out := make(map[string]any, len(schema))
+	for key, value := range schema {
+		if !geminiSchemaFields[key] {
+			continue
+		}
+		switch key {
+		case "properties":
+			if props, ok := value.(map[string]any); ok {
+				cleaned := make(map[string]any, len(props))
+				for name, sub := range props {
+					if subMap, ok := sub.(map[string]any); ok {
+						cleaned[name] = sanitizeGeminiSchema(subMap)
+					} else {
+						cleaned[name] = sub
+					}
+				}
+				value = cleaned
+			}
+		case "items":
+			if subMap, ok := value.(map[string]any); ok {
+				value = sanitizeGeminiSchema(subMap)
+			}
+		case "anyOf":
+			if variants, ok := value.([]any); ok {
+				cleaned := make([]any, len(variants))
+				for i, variant := range variants {
+					if subMap, ok := variant.(map[string]any); ok {
+						cleaned[i] = sanitizeGeminiSchema(subMap)
+					} else {
+						cleaned[i] = variant
+					}
+				}
+				value = cleaned
+			}
+		}
+		out[key] = value
+	}
+	return out
+}
+
 func mapMessages(messages []zeroruntime.Message) (*geminiContent, []geminiContent, error) {
 	systemParts := []geminiPart{}
 	contents := []geminiContent{}

internal/providers/gemini/provider_test.go

@@ -3,6 +3,7 @@ package gemini
 import (
 	"context"
 	"encoding/json"
+	"fmt"
 	"net/http"
 	"net/http/httptest"
 	"reflect"
@@ -504,3 +505,129 @@ func eventsOfType(events []zeroruntime.StreamEvent, eventType zeroruntime.Stream
 	}
 	return matching
 }
+
+// TestSanitizeGeminiSchemaStripsUnsupportedFields: the sanitizer keeps only
+// Gemini-supported keywords and recurses into properties/items/anyOf, so no
+// OpenAI-ism (additionalProperties, $schema, patternProperties) survives at any
+// depth while legitimate fields (type/description/enum/required/default) stay.
+func TestSanitizeGeminiSchemaStripsUnsupportedFields(t *testing.T) {
+	in := map[string]any{
+		"type":                 "object",
+		"additionalProperties": false,
+		"$schema":              "https://json-schema.org/draft/2020-12/schema",
+		"required":             []string{"path"},
+		"properties": map[string]any{
+			"path": map[string]any{
+				"type":        "string",
+				"description": "a path",
+				"default":     ".",
+			},
+			"nested": map[string]any{
+				"type":                 "object",
+				"additionalProperties": false, // must be stripped at depth too
+				"properties": map[string]any{
+					"tags": map[string]any{
+						"type":  "array",
+						"items": map[string]any{"type": "string", "additionalProperties": true},
+					},
+				},
+			},
+		},
+	}
+	out := sanitizeGeminiSchema(in)
+
+	assertNoAdditionalProps(t, out, "root")
+	if out["additionalProperties"] != nil {
+		t.Fatal("top-level additionalProperties must be stripped")
+	}
+	if out["$schema"] != nil {
+		t.Fatal("$schema must be stripped")
+	}
+	// Legitimate fields survive.
+	if out["type"] != "object" {
+		t.Fatalf("type should survive, got %v", out["type"])
+	}
+	props := out["properties"].(map[string]any)
+	pathProp := props["path"].(map[string]any)
+	if pathProp["description"] != "a path" || pathProp["default"] != "." {
+		t.Fatalf("legitimate property fields must survive: %#v", pathProp)
+	}
+	// nil input stays nil (parameterless tool).
+	if sanitizeGeminiSchema(nil) != nil {
+		t.Fatal("nil schema should map to nil")
+	}
+}
+
+// assertNoAdditionalProps fails if additionalProperties appears anywhere in the
+// (recursively walked) schema.
+func assertNoAdditionalProps(t *testing.T, node any, path string) {
+	t.Helper()
+	switch v := node.(type) {
+	case map[string]any:
+		if _, ok := v["additionalProperties"]; ok {
+			t.Fatalf("additionalProperties leaked at %s", path)
+		}
+		for k, sub := range v {
+			assertNoAdditionalProps(t, sub, path+"."+k)
+		}
+	case []any:
+		for i, sub := range v {
+			assertNoAdditionalProps(t, sub, fmt.Sprintf("%s[%d]", path, i))
+		}
+	}
+}
+
+// TestGeminiRequestOmitsAdditionalPropertiesInToolSchema: end-to-end, the tool
+// parameters Zero emits (schemaToRuntimeMap always writes additionalProperties)
+// must not reach Gemini — otherwise every functionDeclaration is 400-rejected
+// ("Unknown name additionalProperties"), which broke all tool-using exec calls
+// against Google (issue #373).
+func TestGeminiRequestOmitsAdditionalPropertiesInToolSchema(t *testing.T) {
+	var gotBody map[string]any
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if err := json.NewDecoder(r.Body).Decode(&gotBody); err != nil {
+			t.Fatalf("decode request body: %v", err)
+		}
+		writeSSE(w, `{"usageMetadata":{"promptTokenCount":1}}`)
+	}))
+	defer server.Close()
+
+	provider, err := New(Options{APIKey: "k", BaseURL: server.URL + "/", Model: "models/gemini-2.5-flash"})
+	if err != nil {
+		t.Fatalf("New returned error: %v", err)
+	}
+	stream, err := provider.StreamCompletion(context.Background(), zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "hi"}},
+		Tools: []zeroruntime.ToolDefinition{{
+			Name:        "grep",
+			Description: "search",
+			Parameters: map[string]any{ // exactly what schemaToRuntimeMap produces
+				"type":                 "object",
+				"additionalProperties": false,
+				"required":             []string{"pattern"},
+				"properties": map[string]any{
+					"pattern": map[string]any{"type": "string", "description": "regex"},
+				},
+			},
+		}},
+	})
+	if err != nil {
+		t.Fatalf("StreamCompletion returned error: %v", err)
+	}
+	drain(stream)
+
+	tools := gotBody["tools"].([]any)
+	decls := tools[0].(map[string]any)["functionDeclarations"].([]any)
+	params := decls[0].(map[string]any)["parameters"].(map[string]any)
+	if _, ok := params["additionalProperties"]; ok {
+		t.Fatalf("additionalProperties must not be sent to Gemini: %#v", params)
+	}
+	// The tool is still usable: its real schema survived.
+	if params["type"] != "object" {
+		t.Fatalf("tool parameters type should survive, got %v", params["type"])
+	}
+	props := params["properties"].(map[string]any)
+	if _, ok := props["pattern"]; !ok {
+		t.Fatalf("tool property must survive sanitization: %#v", props)
+	}
+}

internal/providers/openai/codex_responses.go

@@ -389,10 +389,10 @@ func (p *CodexProvider) streamResponses(
 	err = providerio.ScanSSEDataWithContext(streamCtx, cancelStream, response.Body, inner.streamIdleTimeout, func(data string) bool {
 		return p.emitResponsesEvent(ctx, data, state, events)
 	})
-	if errors.Is(err, providerio.ErrStreamIdle) {
+	if errors.Is(err, providerio.ErrStreamIdle) || errors.Is(err, providerio.ErrStreamStalled) {
 		providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{
 			Type:  zeroruntime.StreamEventError,
-			Error: p.redact(fmt.Sprintf("provider stream error: idle timeout after %s (upstream stopped sending data)", inner.streamIdleTimeout)),
+			Error: p.redact("provider stream error: " + providerio.StreamTimeoutMessage(err, inner.streamIdleTimeout)),
 		})
 		return
 	}

internal/providers/openai/provider.go

@@ -102,10 +102,11 @@ func New(options Options) (*Provider, error) {
 		endpoint = baseURL + "/chat/completions"
 	}
 
-	httpClient := options.HTTPClient
-	if httpClient == nil {
-		httpClient = http.DefaultClient
-	}
+	// Route through providerio.HTTPClient so a missing client gets the shared,
+	// stall-hardened transport (bounded response-header wait + shorter idle-conn
+	// reuse) rather than the raw http.DefaultClient, which hangs on a reused-dead
+	// pooled connection (the macOS-only "both providers stall" bug).
+	httpClient := providerio.HTTPClient(options.HTTPClient)
 
 	maxTokens := options.MaxTokens
 	if maxTokens < 0 {
@@ -218,12 +219,12 @@ func (provider *Provider) stream(ctx context.Context, body []byte, events chan<-
 	err = providerio.ScanSSEDataWithContext(streamCtx, cancelStream, response.Body, provider.streamIdleTimeout, func(data string) bool {
 		return provider.emitPayload(ctx, data, state, events)
 	})
-	if errors.Is(err, providerio.ErrStreamIdle) {
+	if errors.Is(err, providerio.ErrStreamIdle) || errors.Is(err, providerio.ErrStreamStalled) {
 		state.flushBufferedContent(events)
 		state.closeBufferedOpen(events)
 		sendEvent(ctx, events, zeroruntime.StreamEvent{
 			Type:  zeroruntime.StreamEventError,
-			Error: provider.redact(fmt.Sprintf("provider stream error: idle timeout after %s (upstream stopped sending data)", provider.streamIdleTimeout)),
+			Error: provider.redact("provider stream error: " + providerio.StreamTimeoutMessage(err, provider.streamIdleTimeout)),
 		})
 		return
 	}

internal/providers/providerio/providerio.go

@@ -6,10 +6,12 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"math"
 	"net/http"
 	"net/url"
 	"os"
 	"regexp"
+	"runtime"
 	"strconv"
 	"strings"
 	"time"
@@ -23,6 +25,26 @@ const maxSSELineBytes = 16 * 1024 * 1024
 // closing the connection. Callers surface it as an idle-timeout error.
 var ErrStreamIdle = errors.New("idle timeout (upstream stopped sending data)")
 
+// ErrStreamStalled reports that a streaming upstream kept the connection alive
+// (SSE keep-alives reset the idle watchdog, so it never fired) but produced no
+// actual output for ContentStallTimeout(idle). Without this an upstream that
+// heartbeats-but-stalls — observed on chatgpt/gpt-5.x and ollama reasoning
+// models — would hang the agent indefinitely.
+var ErrStreamStalled = errors.New("stream stalled (upstream kept the connection alive but produced no output)")
+
+// StreamTimeoutMessage returns the human-readable detail for a stream-timeout
+// error (ErrStreamIdle or ErrStreamStalled), given the configured idle timeout.
+// Callers prepend their own "provider stream error: " prefix. A stalled stream
+// gets a distinct, actionable message — it did NOT stop sending data (keep-alives
+// kept arriving); it just produced no output, which usually means the model is
+// stuck or very slow.
+func StreamTimeoutMessage(err error, idleTimeout time.Duration) string {
+	if errors.Is(err, ErrStreamStalled) {
+		return fmt.Sprintf("no output for %s (the model kept the connection alive but produced nothing — it may be stuck; try a faster model or lower reasoning effort)", ContentStallTimeout(idleTimeout))
+	}
+	return fmt.Sprintf("idle timeout after %s (upstream stopped sending data)", idleTimeout)
+}
+
 // DefaultStreamIdleTimeout is the single source of truth for how long every
 // provider waits on a silent stream before aborting it. The watchdog only fires
 // on genuine silence — SSE keep-alive comments reset it (see
@@ -33,6 +55,33 @@ var ErrStreamIdle = errors.New("idle timeout (upstream stopped sending data)")
 // Override globally with ZERO_STREAM_IDLE_TIMEOUT.
 const DefaultStreamIdleTimeout = 5 * time.Minute
 
+// ContentStallTimeout bounds a heartbeat-but-no-output stream: keep-alives reset
+// the idle watchdog (a heartbeating upstream is not "dead"), but if NO real data
+// line arrives for this long the stream is aborted (ErrStreamStalled). Scaled
+// above idleTimeout (so a slow-but-producing request that emits data between
+// heartbeats is never killed — the watchdog only ever fires when NOTHING real
+// arrives), but only 1.2× rather than the old 2×: a genuine heartbeat-pause
+// stall on chatgpt/gpt-5.x rarely recovers, so a ~10-minute dead wait (at the 5m
+// idle default) was a terrible UX for a doomed turn. At the default idle this is
+// 6 minutes; it still scales with ZERO_STREAM_IDLE_TIMEOUT. A returned value <= 0
+// (idle watchdog disabled) leaves the content watchdog off too.
+func ContentStallTimeout(idleTimeout time.Duration) time.Duration {
+	if idleTimeout <= 0 {
+		return 0
+	}
+	// 1.2× computed as idle + idle/5 (not idle*6/5), plus a clamp: a
+	// pathologically large ZERO_STREAM_IDLE_TIMEOUT could make idle*6 overflow
+	// int64 and wrap to a NEGATIVE duration, which would arm the content timer
+	// to fire immediately and abort every stream. Clamping to the max duration
+	// on overflow just means "effectively no content watchdog" — the sane
+	// result for an absurd idle timeout.
+	extra := idleTimeout / 5
+	if idleTimeout > math.MaxInt64-extra {
+		return math.MaxInt64
+	}
+	return idleTimeout + extra
+}
+
 // streamIdleTimeoutEnv is the global override for the stream idle timeout. It
 // accepts a Go duration ("5m", "300s", "90s") or a bare number of seconds
 // ("300"). A value of "0", "off", "none", or "disabled" turns the watchdog off
@@ -77,12 +126,58 @@ func NormalizeBaseURL(baseURL string, defaultBaseURL string, label string) (stri
 	return baseURL, nil
 }
 
-// HTTPClient returns the configured client or the process default.
+// sharedHTTPClient is the process-wide client used when a provider supplies none.
+// It tunes the default transport to defeat the stale-pooled-connection hang: Go
+// keeps idle keep-alive connections in a pool, and a later request can reuse one
+// the server/NAT has silently dropped. Because the model call is a POST (non-
+// idempotent), Go will NOT auto-retry it on a fresh connection — so it blocks
+// forever waiting for a response that never arrives. This is provider-agnostic
+// (every provider shares the default transport), which is why it reproduced on
+// BOTH chatgpt and ollama, and it surfaces on macOS far more than Linux because
+// macOS keeps dead pooled connections around longer.
+//
+//   - ResponseHeaderTimeout bounds the gap between finishing the request and the
+//     first response header, so a reused-dead connection fails fast (then the
+//     reconnect path re-dials) instead of hanging. It does NOT bound streaming
+//     after the 200 header arrives, so slow first tokens / long reasoning are
+//     unaffected (the stream-idle + content watchdogs cover that).
+//   - IdleConnTimeout is shortened so a connection that went idle across a pause
+//     is closed (and re-dialed fresh) rather than reused after it has gone stale.
+//   - DisableKeepAlives on darwin only: ResponseHeaderTimeout/IdleConnTimeout
+//     above catch a reused connection that's fully dead (never responds) or
+//     idle past the timeout, but not one that's alive-but-severely-degraded —
+//     e.g. reused shortly after a prior request on the same host (well within
+//     30s, common across quick retries of the same turn), where it still
+//     delivers real bytes, just at a crippled rate, resetting Zero's stream
+//     idle/content-stall watchdogs (which only fire on true silence) without
+//     ever recovering. That degraded-not-dead case is indistinguishable from
+//     genuine backend slowness from inside the stream, so the only reliable
+//     fix is removing pooling from the equation entirely on the one platform
+//     where this class of bug has actually reproduced. A fresh TCP+TLS
+//     handshake per request costs low tens of milliseconds — negligible next
+//     to the minutes-long stalls this avoids — and this doesn't touch
+//     Linux/Windows, where the underlying OS doesn't keep dead/degraded
+//     pooled connections around as long.
+var sharedHTTPClient = func() *http.Client {
+	transport := http.DefaultTransport.(*http.Transport).Clone()
+	// 120s, not 60s: a slow cloud proxy (e.g. ollama `*:cloud`) can withhold its
+	// 200 response header until the upstream model emits a first token, so a 60s cap
+	// risked aborting a legitimately-slow-but-alive request. 120s still bounds a
+	// truly dead reused connection (which never responds) while tolerating slow
+	// header delivery; slow first tokens after the header are covered by the idle +
+	// content-stall watchdogs.
+	transport.ResponseHeaderTimeout = 120 * time.Second
+	transport.IdleConnTimeout = 30 * time.Second
+	transport.DisableKeepAlives = runtime.GOOS == "darwin"
+	return &http.Client{Transport: transport}
+}()
+
+// HTTPClient returns the configured client or the shared, stall-hardened default.
 func HTTPClient(client *http.Client) *http.Client {
 	if client != nil {
 		return client
 	}
-	return http.DefaultClient
+	return sharedHTTPClient
 }
 
 // SendEvent writes a provider event without blocking cancellation cleanup.
@@ -202,21 +297,36 @@ func ScanSSEDataWithContext(
 	// The idle watchdog is optional. When idleTimeout <= 0 it is disabled, but we
 	// STILL run the goroutine + select loop so ctx cancellation is always honored
 	// (a nil idleC channel simply never fires in the select).
-	var idleC <-chan time.Time
+	var idleC, contentC <-chan time.Time
 	resetIdle := func() {}
+	resetContent := func() {}
 	if idleTimeout > 0 {
+		// reset drains the timer's channel if it already fired before re-arming, so
+		// a stale tick can't trip the watchdog after activity resumed.
+		reset := func(t *time.Timer, d time.Duration) func() {
+			return func() {
+				if !t.Stop() {
+					select {
+					case <-t.C:
+					default:
+					}
+				}
+				t.Reset(d)
+			}
+		}
 		idle := time.NewTimer(idleTimeout)
 		defer idle.Stop()
 		idleC = idle.C
-		resetIdle = func() {
-			if !idle.Stop() {
-				select {
-				case <-idle.C:
-				default:
-				}
-			}
-			idle.Reset(idleTimeout)
-		}
+		resetIdle = reset(idle, idleTimeout)
+
+		// Content watchdog: only real data lines reset it (keep-alives do not), so a
+		// stream that heartbeats without producing output is bounded instead of
+		// hanging forever.
+		contentTimeout := ContentStallTimeout(idleTimeout)
+		content := time.NewTimer(contentTimeout)
+		defer content.Stop()
+		contentC = content.C
+		resetContent = reset(content, contentTimeout)
 	}
 
 	for {
@@ -232,6 +342,12 @@ func ScanSSEDataWithContext(
 			// a timeout instead of blocking the agent forever.
 			cancel()
 			return ErrStreamIdle
+		case <-contentC:
+			// Upstream kept heartbeating (idle watchdog never fired) but produced
+			// no output for the content window. Treat as stalled and abort rather
+			// than hanging forever.
+			cancel()
+			return ErrStreamStalled
 		case item, ok := <-payloads:
 			if !ok {
 				// Reader finished: deliver its terminal status (EOF -> nil,
@@ -246,8 +362,11 @@ func ScanSSEDataWithContext(
 			}
 			resetIdle()
 			if item.keepAlive {
+				// A heartbeat keeps the connection "alive" (idle watchdog) but is
+				// NOT output, so it must not reset the content watchdog.
 				continue
 			}
+			resetContent()
 			if !handle(item.data) {
 				// The provider asked to stop (e.g. it already emitted an error
 				// for this payload). Abort the read and end like ScanSSEData:

internal/providers/providerio/providerio_test.go

@@ -4,6 +4,9 @@ import (
 	"context"
 	"errors"
 	"io"
+	"math"
+	"net/http"
+	"runtime"
 	"strings"
 	"testing"
 	"time"
@@ -201,3 +204,170 @@ func TestUpstreamUnreachable(t *testing.T) {
 		})
 	}
 }
+
+// ContentStallTimeout is 1.2× the idle timeout, disabled (0) when idle is
+// disabled, and must never overflow to a negative duration for an absurdly
+// large idle timeout (which would arm the content timer to fire immediately
+// and abort every stream).
+func TestContentStallTimeout(t *testing.T) {
+	cases := []struct {
+		idle time.Duration
+		want time.Duration
+	}{
+		{5 * time.Minute, 6 * time.Minute},
+		{30 * time.Second, 36 * time.Second},
+		{100 * time.Millisecond, 120 * time.Millisecond},
+		{0, 0},
+		{-1, 0},
+	}
+	for _, c := range cases {
+		if got := ContentStallTimeout(c.idle); got != c.want {
+			t.Fatalf("ContentStallTimeout(%v) = %v, want %v", c.idle, got, c.want)
+		}
+	}
+	// Overflow guard: idle*6 would wrap negative; the clamp keeps it positive.
+	if got := ContentStallTimeout(math.MaxInt64); got <= 0 {
+		t.Fatalf("ContentStallTimeout(MaxInt64) = %v, must stay positive (no overflow wrap)", got)
+	}
+}
+
+// A heartbeating-but-output-less upstream must not hang forever: SSE keep-alives
+// feed the idle watchdog (so it never fires), but the content watchdog
+// (ContentStallTimeout(idle)) aborts with ErrStreamStalled when no real data
+// line arrives. This is the gpt-5.x / ollama "still generating forever" hang.
+func TestScanSSEDataWithContextAbortsOnContentStall(t *testing.T) {
+	pr, pw := io.Pipe()
+	defer func() { _ = pw.Close() }()
+
+	stop := make(chan struct{})
+	defer close(stop)
+	go func() {
+		// Heartbeat every 30ms — comfortably under the 100ms idle timeout, so idle
+		// never fires under CI/GC jitter — but never send a data line.
+		for {
+			select {
+			case <-stop:
+				return
+			default:
+			}
+			if _, err := io.WriteString(pw, ": keep-alive\n\n"); err != nil {
+				return
+			}
+			time.Sleep(30 * time.Millisecond)
+		}
+	}()
+
+	cancelled := false
+	cancel := func() { cancelled = true }
+
+	done := make(chan error, 1)
+	go func() {
+		// idle 100ms → content stall at 120ms (ContentStallTimeout = idle*1.2).
+		// Keep-alives reset idle but not content.
+		done <- ScanSSEDataWithContext(context.Background(), cancel, pr, 100*time.Millisecond, func(string) bool { return true })
+	}()
+
+	select {
+	case err := <-done:
+		if !errors.Is(err, ErrStreamStalled) {
+			t.Fatalf("err = %v, want ErrStreamStalled", err)
+		}
+	case <-time.After(3 * time.Second):
+		t.Fatal("ScanSSEDataWithContext hung on a heartbeat-but-no-output stream")
+	}
+	if !cancelled {
+		t.Fatal("content stall did not cancel the request context")
+	}
+}
+
+// Real data lines reset the content watchdog, so a slow-but-producing stream that
+// runs longer than the content window is never aborted.
+func TestScanSSEDataWithContextContentResetsOnData(t *testing.T) {
+	pr, pw := io.Pipe()
+
+	go func() {
+		// One data line every 30ms for ~300ms (well past the 120ms content window,
+		// with comfortable slack under the 100ms idle timeout), so the content
+		// watchdog keeps resetting, then close cleanly.
+		for i := 0; i < 10; i++ {
+			if _, err := io.WriteString(pw, "data: chunk\n\n"); err != nil {
+				return
+			}
+			time.Sleep(30 * time.Millisecond)
+		}
+		_ = pw.Close()
+	}()
+
+	n := 0
+	done := make(chan error, 1)
+	go func() {
+		done <- ScanSSEDataWithContext(context.Background(), func() {}, pr, 100*time.Millisecond, func(string) bool { n++; return true })
+	}()
+
+	select {
+	case err := <-done:
+		if err != nil {
+			t.Fatalf("err = %v, want nil (data kept the stream alive)", err)
+		}
+	case <-time.After(3 * time.Second):
+		t.Fatal("ScanSSEDataWithContext hung on a producing stream")
+	}
+	if n != 10 {
+		t.Fatalf("handled %d data lines, want 10", n)
+	}
+}
+
+// StreamTimeoutMessage must give a stalled stream a distinct, accurate detail —
+// it reports the content window (ContentStallTimeout) and must NOT claim the
+// upstream "stopped sending data" (keep-alives were still arriving).
+func TestStreamTimeoutMessage(t *testing.T) {
+	idle := 5 * time.Minute
+	if msg := StreamTimeoutMessage(ErrStreamIdle, idle); !strings.Contains(msg, "idle timeout after 5m") {
+		t.Fatalf("idle message = %q, want it to mention the 5m idle timeout", msg)
+	}
+	stalled := StreamTimeoutMessage(ErrStreamStalled, idle)
+	if !strings.Contains(stalled, "no output for 6m") {
+		t.Fatalf("stalled message = %q, want it to report the 6m content window (idle 5m × 1.2)", stalled)
+	}
+	if strings.Contains(stalled, "stopped sending data") {
+		t.Fatalf("stalled message must not claim the upstream stopped sending data: %q", stalled)
+	}
+}
+
+// HTTPClient(nil) must return the shared, stall-hardened transport (bounded
+// response-header wait + shorter idle-conn reuse) that defeats the macOS stale-
+// pooled-connection hang; an explicit client is returned untouched.
+func TestHTTPClientReturnsStallHardenedSharedClient(t *testing.T) {
+	got := HTTPClient(nil)
+	if got == nil {
+		t.Fatal("HTTPClient(nil) returned nil")
+	}
+	tr, ok := got.Transport.(*http.Transport)
+	if !ok {
+		t.Fatalf("transport type = %T, want *http.Transport", got.Transport)
+	}
+	if tr.ResponseHeaderTimeout != 120*time.Second {
+		t.Fatalf("ResponseHeaderTimeout = %v, want 120s", tr.ResponseHeaderTimeout)
+	}
+	if tr.IdleConnTimeout != 30*time.Second {
+		t.Fatalf("IdleConnTimeout = %v, want 30s", tr.IdleConnTimeout)
+	}
+	// DisableKeepAlives is the stronger mitigation scoped to darwin only: the
+	// two timeouts above catch a reused connection that's fully dead or idle
+	// too long, but not one that's alive-and-degraded (still delivers real
+	// bytes, just at a crippled rate) — which resets Zero's stream watchdogs
+	// without ever recovering. Removing pooling from the equation entirely is
+	// only worth its reconnect cost on the platform this class of bug has
+	// actually reproduced on.
+	wantDisableKeepAlives := runtime.GOOS == "darwin"
+	if tr.DisableKeepAlives != wantDisableKeepAlives {
+		t.Fatalf("DisableKeepAlives = %v, want %v (GOOS=%s)", tr.DisableKeepAlives, wantDisableKeepAlives, runtime.GOOS)
+	}
+	if HTTPClient(nil) != got {
+		t.Fatal("HTTPClient(nil) must return a shared instance so the conn pool is reused")
+	}
+	custom := &http.Client{}
+	if HTTPClient(custom) != custom {
+		t.Fatal("an explicit client must be returned unchanged")
+	}
+}

internal/specialist/exec.go

@@ -204,7 +204,11 @@ type ChildRunResult struct {
 	Events   []streamjson.Event
 	Stderr   string
 	ExitCode int
-	Started  bool
+	// Signal is a human-readable description (e.g. "signal: killed") when the child
+	// was terminated by a signal rather than exiting normally; empty otherwise. It
+	// turns an opaque exit -1 into an actionable reason (SIGKILL ~ out of memory).
+	Signal  string
+	Started bool
 }
 
 func (executor Executor) Run(ctx context.Context, params TaskParameters, options TaskRunOptions) (ExecResult, error) {
@@ -566,17 +570,35 @@ func (executor Executor) runBuiltArgs(ctx context.Context, built BuildArgsResult
 		exitCode := run.exitCodeOr(-1)
 		summary := SummarizeStream(run.Events, exitCode)
 		executor.recordSpecialistStop(accounting, summary, "error", summary.ExitCode, err, false)
-		return ExecResult{}, err
+		// Carry the child session id even on a post-start failure so a caller (the
+		// swarm launcher -> FailWithSession) can still make the failed member
+		// drillable; the session exists once the child has started.
+		return ExecResult{SessionID: built.SessionID}, err
 	}
 	summary := SummarizeStream(run.Events, run.ExitCode)
 	rolledUp := executor.rollUpSpecialistUsage(accounting, summary)
 	executor.recordSpecialistStop(accounting, summary, summary.Status, summary.ExitCode, nil, rolledUp)
 	return ExecResult{
-		Result:    BuildFinalResult(run.Events, run.Stderr, run.ExitCode),
+		Result:    BuildFinalResult(run.Events, run.Stderr, run.ExitCode, run.Signal),
 		SessionID: built.SessionID,
 	}, nil
 }
 
+// availableSpecialistList renders the registered specialist names for a corrective
+// "not found" error, so a model that guessed a wrong name learns the real options.
+func availableSpecialistList(result LoadResult) string {
+	names := make([]string, 0, len(result.Specialists))
+	for _, manifest := range result.Specialists {
+		if n := strings.TrimSpace(manifest.Metadata.Name); n != "" {
+			names = append(names, n)
+		}
+	}
+	if len(names) == 0 {
+		return "(none registered)"
+	}
+	return strings.Join(names, ", ")
+}
+
 func (executor Executor) loadManifest(name string) (Manifest, error) {
 	name = strings.TrimSpace(name)
 	if name == "" {
@@ -592,7 +614,12 @@ func (executor Executor) loadManifest(name string) (Manifest, error) {
 	}
 	manifest, ok := Find(result, name)
 	if !ok {
-		return Manifest{}, fmt.Errorf("specialist %q not found", name)
+		// Corrective error: a model that invents a specialist name (e.g.
+		// "validator-runner", "file-writer") otherwise gets an opaque "not found"
+		// and keeps retrying made-up names. List the ACTUALLY-registered ones so it
+		// self-corrects — no hardcoded names, since a custom registry may not have
+		// worker/explorer/code-review.
+		return Manifest{}, fmt.Errorf("specialist %q not found. Available: %s. Pick one of these whose tools fit the task, or omit the name to use the default", name, availableSpecialistList(result))
 	}
 	return manifest, nil
 }
@@ -783,15 +810,22 @@ func runChildProcess(ctx context.Context, binaryPath string, args []string, prog
 	}
 	exitCode := 0
 	started := true
+	signalDesc := ""
 	if err := command.Wait(); err != nil {
 		var exitErr *osexec.ExitError
 		if errors.As(err, &exitErr) {
 			exitCode = exitErr.ExitCode()
+			if exitCode < 0 && exitErr.ProcessState != nil {
+				// ExitCode() is -1 when the child was terminated by a signal rather
+				// than exiting. ProcessState.String() is the portable description
+				// (e.g. "signal: killed") — capture it so the failure isn't opaque.
+				signalDesc = exitErr.ProcessState.String()
+			}
 		} else {
 			return ChildRunResult{Events: events, Stderr: stderr.String(), ExitCode: -1, Started: started}, fmt.Errorf("run specialist child: %w", err)
 		}
 	}
-	return ChildRunResult{Events: events, Stderr: stderr.String(), ExitCode: exitCode, Started: started}, nil
+	return ChildRunResult{Events: events, Stderr: stderr.String(), ExitCode: exitCode, Signal: signalDesc, Started: started}, nil
 }
 
 func (run ChildRunResult) exitCodeOr(defaultExitCode int) int {

internal/specialist/load_manifest_error_test.go

@@ -0,0 +1,32 @@
+package specialist
+
+import (
+	"strings"
+	"testing"
+)
+
+// A model that invents a specialist name (e.g. "validator-runner", "file-writer")
+// must get a corrective error listing the real specialists, so it self-corrects to a
+// capable one instead of looping on made-up names that spawn doomed sub-agents.
+func TestLoadManifestUnknownNameListsAvailable(t *testing.T) {
+	executor := Executor{
+		Load: func(LoadOptions) (LoadResult, error) {
+			return LoadResult{Specialists: []Manifest{
+				{Metadata: Metadata{Name: "worker"}},
+				{Metadata: Metadata{Name: "explorer"}},
+				{Metadata: Metadata{Name: "code-review"}},
+			}}, nil
+		},
+	}
+
+	_, err := executor.loadManifest("validator-runner")
+	if err == nil {
+		t.Fatal("an unknown specialist name must error")
+	}
+	msg := err.Error()
+	for _, want := range []string{"validator-runner", "not found", "worker", "explorer", "code-review"} {
+		if !strings.Contains(msg, want) {
+			t.Fatalf("error %q missing %q (must steer the model to a real specialist)", msg, want)
+		}
+	}
+}

internal/specialist/manifest.go

@@ -11,6 +11,7 @@ import (
 	"strings"
 	"time"
 
+	"github.com/Gitlawb/zero/internal/config"
 	"github.com/Gitlawb/zero/internal/modelregistry"
 )
 
@@ -120,7 +121,7 @@ var knownToolNames = map[string]bool{
 }
 
 func DefaultPaths(workspaceRoot string) (Paths, error) {
-	userConfigDir, err := os.UserConfigDir()
+	userConfigDir, err := config.UserConfigDir()
 	if err != nil {
 		return Paths{}, fmt.Errorf("resolve user config directory: %w", err)
 	}

internal/specialist/output_tool.go

@@ -260,6 +260,12 @@ func formatTaskOutputSummary(task background.Task, summary StreamResult, rawLine
 	if !task.CompletedAt.IsZero() {
 		fmt.Fprintf(&builder, "completed_at: %s\n", task.CompletedAt.Format(time.RFC3339))
 		fmt.Fprintf(&builder, "exit_code: %d\n", task.ExitCode)
+		if task.ExitCode < 0 {
+			// A negative exit code means the background child was terminated by a
+			// signal rather than exiting (the int-only launcher can't carry the
+			// signal name). Give the same actionable hint the foreground path does.
+			builder.WriteString("note: terminated by a signal — killed before it finished. Likely causes: the OS out-of-memory killer (common when many sub-agents run in parallel — try fewer), a timeout, or cancellation.\n")
+		}
 	}
 
 	if summary.Text != "" {

internal/specialist/output_tool_test.go

@@ -124,10 +124,16 @@ func TestOutputToolBlocksUntilTaskCompletes(t *testing.T) {
 	tool := NewOutputTool(manager)
 	tool.PollInterval = time.Millisecond
 
+	// timeout is a ceiling, not a target: the test still returns as soon as the
+	// 5ms write lands (PollInterval keeps it responsive), it just tolerates a
+	// slow/contended CI runner without failing. A 100ms budget against a 5ms
+	// write only left 20x margin, which Windows CI runners (coarser default
+	// timer granularity, heavier scheduling contention than Linux/macOS) blew
+	// through and failed the test on an otherwise-unrelated PR merge.
 	result := tool.Run(context.Background(), map[string]any{
 		"task_id": "child_task",
 		"block":   true,
-		"timeout": 100,
+		"timeout": 2000,
 	})
 
 	if result.Status != tools.StatusOK || !strings.Contains(result.Output, "output:\nfinished") {

internal/specialist/streamer.go

@@ -134,12 +134,17 @@ func SummarizeStream(events []streamjson.Event, processExitCode int) StreamResul
 	return result
 }
 
-func BuildFinalResult(events []streamjson.Event, stderrOutput string, processExitCode int) tools.Result {
+func BuildFinalResult(events []streamjson.Event, stderrOutput string, processExitCode int, signalDesc string) tools.Result {
 	summary := SummarizeStream(events, processExitCode)
 	hasErrors := len(summary.Errors) > 0 || summary.ExitCode != 0
 	if summary.Status != "" && summary.Status != "success" && summary.Status != "ok" {
 		hasErrors = true
 	}
+	// A captured kill signal must always surface, even if a late run_end reported a
+	// clean exit just before the child was killed during teardown.
+	if strings.TrimSpace(signalDesc) != "" {
+		hasErrors = true
+	}
 	if !hasErrors {
 		output := summary.Text
 		if summary.SessionID != "" {
@@ -149,7 +154,15 @@ func BuildFinalResult(events []streamjson.Event, stderrOutput string, processExi
 	}
 
 	var builder strings.Builder
-	fmt.Fprintf(&builder, "Subagent failed (exit %d)\n", summary.ExitCode)
+	if signal := strings.TrimSpace(signalDesc); signal != "" {
+		// The child was terminated by a signal (exit code -1) rather than exiting.
+		// Surface the signal instead of an opaque "exit -1", and list the common
+		// causes evenhandedly — this branch also covers timeouts and cancellations,
+		// so don't assert OOM.
+		fmt.Fprintf(&builder, "Subagent terminated by a signal (%s) — it was killed before it finished. Common causes: an out-of-memory kill, a timeout, or cancellation; check the signal to tell which. If you were running many sub-agents in parallel, reduce the concurrency.\n", signal)
+	} else {
+		fmt.Fprintf(&builder, "Subagent failed (exit %d)\n", summary.ExitCode)
+	}
 	if len(summary.Errors) > 0 {
 		fmt.Fprintf(&builder, "errors: %s\n", strings.Join(summary.Errors, "; "))
 	}

internal/specialist/streamer_test.go

@@ -25,7 +25,7 @@ func TestParseStreamAndBuildFinalResultSuccess(t *testing.T) {
 	if summary.SessionID != "child" || summary.Text != "done" || summary.ExitCode != 0 || len(summary.Tools) != 1 || summary.Tools[0] != "grep" {
 		t.Fatalf("unexpected summary: %#v", summary)
 	}
-	result := BuildFinalResult(events, "", 0)
+	result := BuildFinalResult(events, "", 0, "")
 	if result.Status != tools.StatusOK || result.Output != "session_id: child\ndone" {
 		t.Fatalf("unexpected final result: %#v", result)
 	}
@@ -41,7 +41,7 @@ func TestBuildFinalResultUsesTextDeltasWhenFinalMissing(t *testing.T) {
 	if err != nil {
 		t.Fatalf("ParseStream returned error: %v", err)
 	}
-	result := BuildFinalResult(events, "", 0)
+	result := BuildFinalResult(events, "", 0, "")
 	if result.Status != tools.StatusOK || result.Output != "session_id: child\nhello world" {
 		t.Fatalf("unexpected final result: %#v", result)
 	}
@@ -82,7 +82,7 @@ func TestBuildFinalResultErrorIncludesDiagnostics(t *testing.T) {
 	if err != nil {
 		t.Fatalf("ParseStream returned error: %v", err)
 	}
-	result := BuildFinalResult(events, "stderr text", 0)
+	result := BuildFinalResult(events, "stderr text", 0, "")
 	if result.Status != tools.StatusError {
 		t.Fatalf("Status = %s, want error", result.Status)
 	}
@@ -93,6 +93,33 @@ func TestBuildFinalResultErrorIncludesDiagnostics(t *testing.T) {
 	}
 }
 
+// A child terminated by a signal (exit code -1) must surface the signal + an
+// actionable hint, not an opaque "Subagent failed (exit -1)".
+func TestBuildFinalResultSurfacesKillSignal(t *testing.T) {
+	events, err := ParseStream(strings.NewReader(strings.Join([]string{
+		`{"schemaVersion":2,"type":"run_start","runId":"run_1","sessionId":"child"}`,
+		`{"schemaVersion":2,"type":"tool_call","runId":"run_1","id":"call_1","name":"read_file"}`,
+		"",
+	}, "\n")))
+	if err != nil {
+		t.Fatalf("ParseStream returned error: %v", err)
+	}
+	result := BuildFinalResult(events, "", -1, "signal: killed")
+	if result.Status != tools.StatusError {
+		t.Fatalf("Status = %s, want error", result.Status)
+	}
+	// Surfaces the signal + lists causes (OOM/timeout/cancellation) without asserting
+	// OOM as the sole cause.
+	for _, want := range []string{"terminated by a signal", "signal: killed", "out-of-memory", "timeout", "cancellation"} {
+		if !strings.Contains(result.Output, want) {
+			t.Fatalf("kill output missing %q:\n%s", want, result.Output)
+		}
+	}
+	if strings.Contains(result.Output, "Subagent failed (exit") {
+		t.Fatalf("signaled child should not show the opaque exit line:\n%s", result.Output)
+	}
+}
+
 func TestParseStreamRejectsInvalidLines(t *testing.T) {
 	_, err := ParseStream(strings.NewReader(`{"schemaVersion":2,"runId":"run_1"}` + "\n"))
 	if err == nil || !strings.Contains(err.Error(), "type is required") {

internal/swarm/coordinator.go

@@ -166,6 +166,12 @@ func (c *Coordinator) Fail(id, errMsg string) error {
 	return c.finish(id, StatusFailed, "", errMsg, "")
 }
 
+// FailWithSession marks a task failed but keeps its child session id, so a member
+// that ran and failed (e.g. exited non-zero / hit max-turns) is still drillable.
+func (c *Coordinator) FailWithSession(id, errMsg, sessionID string) error {
+	return c.finish(id, StatusFailed, "", errMsg, sessionID)
+}
+
 func (c *Coordinator) finish(id string, status TaskStatus, result, errMsg, sessionID string) error {
 	c.mu.Lock()
 	defer c.mu.Unlock()

internal/swarm/launcher_specialist.go

@@ -2,8 +2,10 @@ package swarm
 
 import (
 	"context"
+	"errors"
 
 	"github.com/Gitlawb/zero/internal/specialist"
+	"github.com/Gitlawb/zero/internal/tools"
 )
 
 // NewSpecialistLauncher adapts internal/specialist.Executor into a
@@ -54,7 +56,18 @@ func NewSpecialistLauncher(executor specialist.Executor) MemberLauncher {
 			MemberAutonomy: true,
 		})
 		if err != nil {
-			return MemberResult{}, err
+			// Preserve the child session id on a post-start failure too (exec.go
+			// returns it on the error path) so lifecycle's FailWithSession can keep
+			// the failed member drillable, not just the StatusError case below.
+			return MemberResult{SessionID: res.SessionID}, err
+		}
+		if res.Result.Status == tools.StatusError {
+			// The child ran but its task FAILED (e.g. non-zero exit / max-turns).
+			// Surface it as a member failure so the swarm marks it [failed], not
+			// [done] — otherwise the orchestrator (which can't see the AGENTS panel)
+			// trusts incomplete work. Keep the session id so the failed member is
+			// still drillable, and carry its report as the failure message.
+			return MemberResult{SessionID: res.SessionID}, errors.New(res.Result.Output)
 		}
 		return MemberResult{Result: res.Result.Output, SessionID: res.SessionID}, nil
 	}}

internal/swarm/launcher_specialist_test.go

@@ -67,3 +67,50 @@ func TestSpecialistLauncherRunsUnregisteredSwarmAgent(t *testing.T) {
 		t.Fatalf("member args missing agent type: %#v", gotArgs)
 	}
 }
+
+// A member whose child exits non-zero (e.g. exit 4 / max-turns) must be reported as
+// a FAILURE — otherwise the swarm marks it [done] and the orchestrator trusts
+// incomplete work. The failed member keeps its session id (drill-in) and the child
+// report rides along as the failure message.
+func TestSpecialistLauncherMarksNonZeroExitAsFailed(t *testing.T) {
+	four := 4
+	executor := specialist.Executor{
+		BinaryPath:   "/usr/local/bin/zero",
+		NewSessionID: func() (string, error) { return "member_task", nil },
+		Load: func(specialist.LoadOptions) (specialist.LoadResult, error) {
+			return specialist.LoadResult{}, nil
+		},
+		RunChild: func(ctx context.Context, binaryPath string, args []string, progress func(streamjson.Event)) (specialist.ChildRunResult, error) {
+			return specialist.ChildRunResult{
+				Events: []streamjson.Event{
+					{Type: streamjson.EventRunStart, SessionID: "member_task"},
+					{Type: streamjson.EventFinal, Text: "i could not finish the objective"},
+					{Type: streamjson.EventRunEnd, Status: "error", ExitCode: &four},
+				},
+				ExitCode: 4,
+			}, nil
+		},
+	}
+
+	handle, err := NewSpecialistLauncher(executor).Launch(context.Background(), MemberSpec{
+		ID:           "m1",
+		TaskID:       "m1",
+		AgentType:    "subagent",
+		Team:         "probe",
+		Task:         "a task too big for the budget",
+		SystemPrompt: "You are a subagent spawned to complete a specific task.",
+	})
+	if err != nil {
+		t.Fatalf("launch: %v", err)
+	}
+	res, err := handle.Wait()
+	if err == nil {
+		t.Fatal("a member that exited non-zero must be reported as FAILED, got nil error")
+	}
+	if !strings.Contains(err.Error(), "exit 4") {
+		t.Fatalf("failure should carry the child report (exit 4), got %q", err.Error())
+	}
+	if res.SessionID != "member_task" {
+		t.Fatalf("a failed member must keep its session id for drill-in, got %q", res.SessionID)
+	}
+}

internal/swarm/lifecycle.go

@@ -71,7 +71,9 @@ func (s *Swarm) watch(t *Team, m *Member, spec MemberSpec) {
 			}
 			// fall through to record the original error if relaunch fails
 		}
-		_ = s.coord.Fail(m.TaskID, memberError(err))
+		// res.SessionID is preserved by the handle even on error, so a member that
+		// ran then failed stays drillable; a pure launch error carries an empty id.
+		_ = s.coord.FailWithSession(m.TaskID, memberError(err), res.SessionID)
 	} else {
 		_ = s.coord.CompleteWithSession(m.TaskID, res.Result, res.SessionID)
 	}

internal/swarm/tools.go

@@ -103,6 +103,25 @@ func errResult(format string, a ...any) tools.Result {
 	return tools.Result{Status: tools.StatusError, Output: "Error: " + fmt.Sprintf(format, a...)}
 }
 
+// agentTypeProperty builds the schema for an agent-type parameter, constraining it
+// to the roster's currently registered types (the built-ins plus any user-registered
+// agents) via an enum so a model can't spawn or hand off to an unknown type — the
+// schema teaches the valid set up front instead of the model learning it from a
+// failed call. Falls back to a free-form string if the roster is unavailable or
+// empty (so the tool never advertises an empty, un-satisfiable enum).
+func agentTypeProperty(sw *Swarm, description string) tools.PropertySchema {
+	prop := tools.PropertySchema{Type: "string", Description: description}
+	if sw == nil {
+		return prop
+	}
+	if registry := sw.Registry(); registry != nil {
+		if types := registry.AgentTypes(); len(types) > 0 {
+			prop.Enum = types
+		}
+	}
+	return prop
+}
+
 // ---- swarm_spawn -----------------------------------------------------------
 
 type spawnTool struct {
@@ -118,7 +137,7 @@ func (t *spawnTool) Parameters() tools.Schema {
 	return tools.Schema{
 		Type: "object",
 		Properties: map[string]tools.PropertySchema{
-			"agent_type": {Type: "string", Description: "Roster agent type to spawn (e.g. teammate, subagent)."},
+			"agent_type": agentTypeProperty(t.sw, "Roster agent type to spawn."),
 			"task":       {Type: "string", Description: "The focused task/briefing for the member."},
 			"team":       {Type: "string", Description: "Team to spawn into. Defaults to \"default\"."},
 		},
@@ -353,7 +372,7 @@ func (t *handoffTool) Parameters() tools.Schema {
 		Type: "object",
 		Properties: map[string]tools.PropertySchema{
 			"task_id":       {Type: "string", Description: "Task id to hand off."},
-			"to_agent_type": {Type: "string", Description: "Roster agent type to take over."},
+			"to_agent_type": agentTypeProperty(t.sw, "Roster agent type to take over."),
 			"note":          {Type: "string", Description: "Optional handoff note delivered to the new member's inbox."},
 			"team":          {Type: "string", Description: "Team the task belongs to. Defaults to \"default\"."},
 		},

internal/swarm/tools_test.go

@@ -2,6 +2,7 @@ package swarm
 
 import (
 	"context"
+	"slices"
 	"strings"
 	"testing"
 	"time"
@@ -10,6 +11,46 @@ import (
 	"github.com/Gitlawb/zero/internal/tools"
 )
 
+func TestSpawnAndHandoffAgentTypeEnumReflectsRoster(t *testing.T) {
+	sw := newSwarmFor(t, newLauncher(okFor))
+	want := sw.Registry().AgentTypes() // sorted built-ins: subagent, teammate
+
+	spawnEnum := (&spawnTool{sw: sw}).Parameters().Properties["agent_type"].Enum
+	if !slices.Equal(spawnEnum, want) {
+		t.Fatalf("spawn agent_type enum = %v, want roster %v", spawnEnum, want)
+	}
+	if !slices.Contains(spawnEnum, "subagent") || !slices.Contains(spawnEnum, "teammate") {
+		t.Fatalf("enum missing built-in roster types: %v", spawnEnum)
+	}
+	handoffEnum := (&handoffTool{sw: sw}).Parameters().Properties["to_agent_type"].Enum
+	if !slices.Equal(handoffEnum, want) {
+		t.Fatalf("handoff to_agent_type enum = %v, want roster %v", handoffEnum, want)
+	}
+
+	// A user-registered custom type must extend the enum (not a hardcoded list).
+	if err := sw.Registry().Register(Definition{AgentType: "researcher"}); err != nil {
+		t.Fatalf("register custom agent: %v", err)
+	}
+	spawnEnum = (&spawnTool{sw: sw}).Parameters().Properties["agent_type"].Enum
+	if !slices.Contains(spawnEnum, "researcher") {
+		t.Fatalf("custom agent type not reflected in spawn enum: %v", spawnEnum)
+	}
+	handoffEnum = (&handoffTool{sw: sw}).Parameters().Properties["to_agent_type"].Enum
+	if !slices.Contains(handoffEnum, "researcher") {
+		t.Fatalf("custom agent type not reflected in handoff enum: %v", handoffEnum)
+	}
+}
+
+func TestAgentTypePropertyFallsBackWithoutRoster(t *testing.T) {
+	prop := agentTypeProperty(nil, "desc")
+	if prop.Type != "string" || prop.Description != "desc" {
+		t.Fatalf("unexpected property: %+v", prop)
+	}
+	if prop.Enum != nil {
+		t.Fatalf("nil swarm must not advertise an (empty) enum, got %v", prop.Enum)
+	}
+}
+
 func TestCollapseRuneSafeTruncation(t *testing.T) {
 	// 300 multi-byte runes: truncation must not split a rune (no invalid UTF-8).
 	got := collapse(strings.Repeat("🚀", 300))

internal/tools/exec_command.go

@@ -28,8 +28,17 @@ const (
 	completedSessionRetention = 30 * time.Second
 	maxExecSessions           = 64
 	recentExecOutputBytes     = 4096
-	execSessionStopTimeout    = 3 * time.Second
-	execSessionEvictedMessage = "[zero] session evicted: too many background terminals\n"
+	// maxExecOutputBufferBytes caps the undrained output an unpolled session can
+	// accumulate. Without a cap, a long-lived background session nobody polls
+	// again (e.g. a dev server left running after its initiating run was
+	// cancelled) grows this buffer forever as long as the process keeps writing
+	// output, with no ceiling — this previously ran a session's memory into the
+	// tens of gigabytes over several hours and got the whole zero process
+	// OOM-killed by the OS.
+	maxExecOutputBufferBytes         = 2 * 1024 * 1024
+	execSessionStopTimeout           = 3 * time.Second
+	execSessionEvictedMessage        = "[zero] session evicted: too many background terminals\n"
+	execOutputBufferTruncatedMessage = "[zero] output buffer truncated: undrained output exceeded 2MiB, oldest output dropped"
 )
 
 type execSessionManager struct {
@@ -233,6 +242,10 @@ type ExecSessionSnapshot struct {
 	Status       string
 	ExitCode     *int
 	RecentOutput string
+	// OutputTruncated reflects the buffer's last-known truncation state (see
+	// execOutputBuffer.peekTruncated), so a session listing (/ps) still shows
+	// this even if the session is reaped before ever being polled again.
+	OutputTruncated bool
 }
 
 type ExecSessionController interface {
@@ -317,24 +330,26 @@ func (session *execSession) snapshot() ExecSessionSnapshot {
 		status = "exited"
 	}
 	return ExecSessionSnapshot{
-		ID:           session.id,
-		Command:      session.commandText,
-		Cwd:          session.cwd,
-		RelativeCwd:  session.relativeCwd,
-		StartedAt:    startedAt,
-		LastUsedAt:   lastUsedAt,
-		TTY:          session.tty,
-		Status:       status,
-		ExitCode:     copiedExit,
-		RecentOutput: session.output.recentString(),
+		ID:              session.id,
+		Command:         session.commandText,
+		Cwd:             session.cwd,
+		RelativeCwd:     session.relativeCwd,
+		StartedAt:       startedAt,
+		LastUsedAt:      lastUsedAt,
+		TTY:             session.tty,
+		Status:          status,
+		ExitCode:        copiedExit,
+		RecentOutput:    session.output.recentString(),
+		OutputTruncated: session.output.peekTruncated(),
 	}
 }
 
 type execOutputBuffer struct {
-	mu     sync.Mutex
-	data   []byte
-	recent []byte
-	notify chan struct{}
+	mu        sync.Mutex
+	data      []byte
+	recent    []byte
+	truncated bool
+	notify    chan struct{}
 }
 
 func newExecOutputBuffer() *execOutputBuffer {
@@ -344,6 +359,10 @@ func newExecOutputBuffer() *execOutputBuffer {
 func (buffer *execOutputBuffer) Write(p []byte) (int, error) {
 	buffer.mu.Lock()
 	buffer.data = append(buffer.data, p...)
+	if len(buffer.data) > maxExecOutputBufferBytes {
+		buffer.data = buffer.data[len(buffer.data)-maxExecOutputBufferBytes:]
+		buffer.truncated = true
+	}
 	buffer.recent = append(buffer.recent, p...)
 	if len(buffer.recent) > recentExecOutputBytes {
 		buffer.recent = buffer.recent[len(buffer.recent)-recentExecOutputBytes:]
@@ -373,6 +392,34 @@ func (buffer *execOutputBuffer) drainString() string {
 	return out
 }
 
+// consumeTruncated reports whether the buffer has dropped output to stay
+// within maxExecOutputBufferBytes since the last call, resetting the flag.
+// Kept as an out-of-band signal rather than text embedded in drainString's
+// result: an earlier version prefixed a notice directly into the drained
+// string, but that notice always sat ~maxExecOutputBufferBytes (2MiB) before
+// the end of the combined output, far outside the byte-budget head/tail
+// window truncateExecOutput keeps afterward (at most 400KB even at the
+// tool's maximum max_output_tokens) — so the notice was reliably swallowed
+// or chopped by that second, unrelated truncation pass. Surfacing it as a
+// separate bool lets the caller report it regardless of where in the byte
+// stream the overflow happened.
+func (buffer *execOutputBuffer) consumeTruncated() bool {
+	buffer.mu.Lock()
+	defer buffer.mu.Unlock()
+	truncated := buffer.truncated
+	buffer.truncated = false
+	return truncated
+}
+
+// peekTruncated reports the buffer's last-known truncation state without
+// clearing it, for status views (e.g. /ps) that shouldn't consume the signal
+// a subsequent poll is still meant to report via consumeTruncated.
+func (buffer *execOutputBuffer) peekTruncated() bool {
+	buffer.mu.Lock()
+	defer buffer.mu.Unlock()
+	return buffer.truncated
+}
+
 type execCommandTool struct {
 	baseTool
 	workspaceRoot string
@@ -485,25 +532,28 @@ func (tool execCommandTool) run(ctx context.Context, args map[string]any, engine
 	if err != nil {
 		return errorResult("Error starting exec_command: " + err.Error())
 	}
-	output := session.collect(ctx, time.Duration(yieldTimeMS)*time.Millisecond)
+	output, outputTruncated := session.collect(ctx, time.Duration(yieldTimeMS)*time.Millisecond)
 	if ctx != nil && ctx.Err() != nil && !session.doneClosed() {
 		session.terminate()
-		output += session.collect(context.Background(), time.Second)
+		more, moreTruncated := session.collect(context.Background(), time.Second)
+		output += more
+		outputTruncated = outputTruncated || moreTruncated
 	}
 	exitCode, exited := session.exitStatus()
 	if exited {
 		tool.manager.remove(session.id)
 	}
 	return execToolResult(execToolResultInput{
-		commandText:     commandText,
-		output:          output,
-		sessionID:       session.id,
-		exitCode:        exitCode,
-		exited:          exited,
-		relativeCwd:     relativeCwd,
-		tty:             session.tty,
-		plan:            session.plan,
-		maxOutputTokens: maxOutputTokens,
+		commandText:           commandText,
+		output:                output,
+		outputBufferTruncated: outputTruncated,
+		sessionID:             session.id,
+		exitCode:              exitCode,
+		exited:                exited,
+		relativeCwd:           relativeCwd,
+		tty:                   session.tty,
+		plan:                  session.plan,
+		maxOutputTokens:       maxOutputTokens,
 	})
 }
 
@@ -558,26 +608,49 @@ func (tool execCommandTool) startSession(commandText string, absoluteCwd string,
 	return session, nil
 }
 
-func (session *execSession) collect(ctx context.Context, wait time.Duration) string {
+// collect drains the session's output buffer, returning the accumulated text
+// and whether the buffer ever had to drop output to stay within
+// maxExecOutputBufferBytes since the last collect call.
+func (session *execSession) collect(ctx context.Context, wait time.Duration) (string, bool) {
 	if ctx == nil {
 		ctx = context.Background()
 	}
 	deadline := time.Now().Add(wait)
 	var builder strings.Builder
+	truncated := false
+	finish := func() (string, bool) {
+		if session.output.consumeTruncated() {
+			truncated = true
+		}
+		return builder.String(), truncated
+	}
 	for {
 		if chunk := session.output.drainString(); chunk != "" {
 			builder.WriteString(chunk)
+			if session.output.consumeTruncated() {
+				truncated = true
+			}
+			// A background process that keeps writing output faster than this
+			// loop can drain it (a chatty dev server, a watch-mode rebuild loop,
+			// a crash-restart spew) would otherwise never let drainString return
+			// empty, so the deadline check below — only reached on an empty
+			// drain — would never fire. That turns `wait` into an unbounded
+			// block regardless of what the caller asked for. Check it here too,
+			// so continuous output can never starve the timeout.
+			if time.Now().After(deadline) {
+				return finish()
+			}
 			continue
 		}
 		if session.doneClosed() {
 			if chunk := session.output.drainString(); chunk != "" {
 				builder.WriteString(chunk)
 			}
-			return builder.String()
+			return finish()
 		}
 		remaining := time.Until(deadline)
 		if remaining <= 0 {
-			return builder.String()
+			return finish()
 		}
 		timer := time.NewTimer(remaining)
 		select {
@@ -602,9 +675,9 @@ func (session *execSession) collect(ctx context.Context, wait time.Duration) str
 				default:
 				}
 			}
-			return builder.String()
+			return finish()
 		case <-timer.C:
-			return builder.String()
+			return finish()
 		}
 	}
 }
@@ -701,22 +774,23 @@ func (tool writeStdinTool) RunWithOptions(ctx context.Context, args map[string]a
 			}
 		}
 	}
-	output := session.collect(ctx, time.Duration(yieldTimeMS)*time.Millisecond)
+	output, outputTruncated := session.collect(ctx, time.Duration(yieldTimeMS)*time.Millisecond)
 	exitCode, exited := session.exitStatus()
 	if exited {
 		tool.manager.remove(session.id)
 	}
 	return execToolResult(execToolResultInput{
-		commandText:     session.commandText,
-		output:          output,
-		sessionID:       session.id,
-		exitCode:        exitCode,
-		exited:          exited,
-		relativeCwd:     session.relativeCwd,
-		tty:             session.tty,
-		interrupted:     interrupted,
-		plan:            session.plan,
-		maxOutputTokens: maxOutputTokens,
+		commandText:           session.commandText,
+		output:                output,
+		outputBufferTruncated: outputTruncated,
+		sessionID:             session.id,
+		exitCode:              exitCode,
+		exited:                exited,
+		relativeCwd:           session.relativeCwd,
+		tty:                   session.tty,
+		interrupted:           interrupted,
+		plan:                  session.plan,
+		maxOutputTokens:       maxOutputTokens,
 	})
 }
 
@@ -741,16 +815,22 @@ func shouldInterruptExecSession(chars string, tty bool) bool {
 }
 
 type execToolResultInput struct {
-	commandText     string
-	output          string
-	sessionID       int
-	exitCode        int
-	exited          bool
-	relativeCwd     string
-	tty             bool
-	interrupted     bool
-	plan            zeroSandbox.CommandPlan
-	maxOutputTokens int
+	commandText string
+	output      string
+	// outputBufferTruncated is true when the session's undrained output buffer
+	// had to drop bytes to stay within maxExecOutputBufferBytes since it was
+	// last collected — data that is gone for good, unlike the head/tail
+	// truncation below (which the caller can recover by polling again or
+	// raising max_output_tokens).
+	outputBufferTruncated bool
+	sessionID             int
+	exitCode              int
+	exited                bool
+	relativeCwd           string
+	tty                   bool
+	interrupted           bool
+	plan                  zeroSandbox.CommandPlan
+	maxOutputTokens       int
 }
 
 func execToolResult(input execToolResultInput) Result {
@@ -771,16 +851,26 @@ func execToolResult(input execToolResultInput) Result {
 	} else {
 		meta["session_id"] = strconv.Itoa(input.sessionID)
 	}
+	if input.outputBufferTruncated {
+		meta["output_buffer_truncated"] = "true"
+	}
 
 	status := StatusOK
 	if input.exited && ((input.exitCode != 0 && !input.interrupted) || meta[SandboxLikelyDeniedMeta] == "true") {
 		status = StatusError
 	}
 	body := formatExecCommandOutput(output, input.sessionID, input.exited, input.exitCode, input.interrupted)
+	if input.outputBufferTruncated {
+		// Appended after truncateExecOutput's own head/tail slicing, not
+		// embedded in the text that goes through it — a marker inside that
+		// text can land in the discarded middle or get chopped at a small
+		// max_output_tokens budget. Appending here guarantees it survives.
+		body += "\n" + execOutputBufferTruncatedMessage
+	}
 	return Result{
 		Status:    status,
 		Output:    body,
-		Truncated: truncated,
+		Truncated: truncated || input.outputBufferTruncated,
 		Meta:      meta,
 		Display: Display{
 			Summary: execDisplaySummary(input.commandText, input.sessionID, input.exited, input.exitCode),

internal/tools/exec_command_test.go

@@ -363,6 +363,152 @@ func TestStartExecProcessFallsBackAfterPTYStartMutation(t *testing.T) {
 	}
 }
 
+// TestExecOutputBufferCapsUndrainedData: a session nobody polls must not grow
+// its undrained buffer without bound — a long-lived background process that
+// keeps writing while unpolled previously ran a session's memory into the
+// tens of gigabytes and got the whole zero process OOM-killed by the OS.
+func TestExecOutputBufferCapsUndrainedData(t *testing.T) {
+	buffer := newExecOutputBuffer()
+
+	chunk := strings.Repeat("x", 1024)
+	writes := (maxExecOutputBufferBytes / len(chunk)) + 10 // comfortably over the cap
+	for i := 0; i < writes; i++ {
+		if _, err := buffer.Write([]byte(chunk)); err != nil {
+			t.Fatalf("Write: %v", err)
+		}
+	}
+
+	buffer.mu.Lock()
+	dataLen := len(buffer.data)
+	buffer.mu.Unlock()
+	if dataLen > maxExecOutputBufferBytes {
+		t.Fatalf("undrained buffer grew to %d bytes, want <= %d", dataLen, maxExecOutputBufferBytes)
+	}
+
+	out := buffer.drainString()
+	// drainString itself carries no truncation marker: a marker embedded in
+	// the drained text would always sit ~maxExecOutputBufferBytes before the
+	// end of the string, past any realistic head/tail truncation window a
+	// caller applies afterward (see execToolResult), so it's reliably lost.
+	// The signal lives out of band on consumeTruncated/peekTruncated instead.
+	if !strings.HasSuffix(out, chunk) {
+		t.Fatal("drained output should keep the most recent bytes, not the oldest")
+	}
+	if len(out) > maxExecOutputBufferBytes {
+		t.Fatalf("drained output = %d bytes, want <= %d (no marker text mixed in)", len(out), maxExecOutputBufferBytes)
+	}
+
+	if !buffer.peekTruncated() {
+		t.Fatal("peekTruncated should report the overflow without clearing it")
+	}
+	if !buffer.peekTruncated() {
+		t.Fatal("a second peekTruncated call should still report it — peek must not consume")
+	}
+	if !buffer.consumeTruncated() {
+		t.Fatal("consumeTruncated should report the overflow")
+	}
+	if buffer.consumeTruncated() {
+		t.Fatal("consumeTruncated should reset after being read once")
+	}
+	if buffer.peekTruncated() {
+		t.Fatal("peekTruncated should reflect the reset state after consumeTruncated")
+	}
+
+	// A second drain with no intervening writes must be empty.
+	if got := buffer.drainString(); got != "" {
+		t.Fatalf("drainString after a full drain = %q, want empty", got)
+	}
+}
+
+// TestExecToolResultSurfacesBufferTruncationOutsideByteBudget: an earlier
+// version embedded the truncation notice directly in the drained text, which
+// a review found sits ~maxExecOutputBufferBytes before the end of the
+// string — far past the head/tail window truncateExecOutput keeps even at
+// the tool's smallest allowed max_output_tokens, so the notice was reliably
+// swallowed or chopped. The notice must survive regardless of how small the
+// byte budget is, and it must not count against that budget.
+func TestExecToolResultSurfacesBufferTruncationOutsideByteBudget(t *testing.T) {
+	hugeOutput := strings.Repeat("y", maxExecOutputBufferBytes+1024)
+
+	result := execToolResult(execToolResultInput{
+		commandText:           "cmd",
+		output:                hugeOutput,
+		outputBufferTruncated: true,
+		sessionID:             1,
+		exited:                false,
+		maxOutputTokens:       1, // the schema's own declared minimum
+	})
+
+	if !strings.Contains(result.Output, execOutputBufferTruncatedMessage) {
+		t.Fatalf("result output should contain the buffer-truncation notice even at the smallest max_output_tokens, got %q", result.Output[:min(len(result.Output), 200)])
+	}
+	if result.Meta["output_buffer_truncated"] != "true" {
+		t.Fatalf("result meta should flag output_buffer_truncated, got %#v", result.Meta)
+	}
+	if !result.Truncated {
+		t.Fatal("result should report Truncated when the buffer dropped output")
+	}
+}
+
+// TestCollectRespectsDeadlineUnderContinuousOutput asserts collect() returns
+// close to its requested deadline even while output keeps arriving. Before
+// the corresponding fix, the deadline was only checked in the branch reached
+// when drainString returns empty — a background process producing output
+// fast/continuously enough to keep the buffer perpetually non-empty could
+// theoretically starve that check indefinitely. This synthetic writer
+// (8 goroutines, tight loop) did not reliably reproduce that starvation under
+// Go's scheduler in practice — the reader still won the race often enough —
+// so this test doesn't prove the old code could hang; it just pins down the
+// intended behavior (bounded by wait) going forward.
+func TestCollectRespectsDeadlineUnderContinuousOutput(t *testing.T) {
+	session := &execSession{
+		id:     1000,
+		output: newExecOutputBuffer(),
+		done:   make(chan struct{}),
+	}
+
+	stop := make(chan struct{})
+	var wg sync.WaitGroup
+	// Several concurrent writers, each pushing a decent-sized chunk in a tight
+	// loop: a single slow writer lets the reader win the race often enough
+	// (the runtime schedules a gap between writes) to mask the bug. Enough
+	// parallel writers keep the buffer non-empty essentially continuously,
+	// closer to a real chatty process whose PTY reads land in bursts.
+	const writers = 8
+	chunk := []byte(strings.Repeat("x", 256))
+	wg.Add(writers)
+	for i := 0; i < writers; i++ {
+		go func() {
+			defer wg.Done()
+			for {
+				select {
+				case <-stop:
+					return
+				default:
+					session.output.Write(chunk)
+				}
+			}
+		}()
+	}
+	t.Cleanup(func() {
+		close(stop)
+		wg.Wait()
+	})
+
+	const wait = 200 * time.Millisecond
+	start := time.Now()
+	_, _ = session.collect(context.Background(), wait)
+	elapsed := time.Since(start)
+
+	// Generous slack over `wait` for scheduling jitter under a continuously
+	// writing goroutine — this must stay a small multiple of wait, not
+	// "however long the writer keeps going" (which is what the bug produced:
+	// this test would hang past the 30s test timeout without the fix).
+	if elapsed > 3*wait {
+		t.Fatalf("collect took %v under continuous output, want close to the %v deadline", elapsed, wait)
+	}
+}
+
 // resilientTempDir is like t.TempDir() but tolerates the Windows handle-release
 // lag: a SIGKILL'd child process that had the dir as its cwd may not have
 // released it the instant it is reaped, so the immediate RemoveAll t.TempDir()

internal/tui/background_terminals.go

@@ -146,6 +146,9 @@ func formatBackgroundTerminalRow(session tools.ExecSessionSnapshot, now time.Tim
 	}
 	preview := compactCommandOutputText(session.RecentOutput)
 	prefix := fmt.Sprintf("%d · %s · %s · %s", session.ID, session.Status, age, cwd)
+	if session.OutputTruncated {
+		prefix += " · output truncated"
+	}
 	if preview != "" {
 		return prefix + " · " + command + " · " + preview
 	}

internal/tui/clipboard.go

@@ -56,6 +56,13 @@ func pasteFromClipboardCmd() tea.Cmd {
 // spec prompt, the MCP manager, an open picker, the detailed transcript) swallow
 // the paste; empty content is a no-op.
 func (m model) routePaste(content string) (tea.Model, tea.Cmd) {
+	// A paste is a deliberate action, same as a keypress or click — it means
+	// the user moved on to something else, so it disarms a stale Esc
+	// cancel-confirmation. tea.PasteMsg/clipboardReadMsg aren't
+	// tea.KeyPressMsg, so they don't go through that generic keypress hook;
+	// do it once here, before any early return, so both the bracketed-paste
+	// and right-click-paste paths are covered uniformly.
+	m = m.disarmCancelConfirmation()
 	if content == "" {
 		// Empty text clipboard — the user may have pasted a screenshot.
 		// Probe the OS clipboard for image content asynchronously.

internal/tui/command_center.go

@@ -441,6 +441,8 @@ func (m model) handleModelCommand(args string) (model, string) {
 	m.providerProfile = nextProfile
 	m.provider = nextProvider
 	m.providerName = displayValue(nextProfile.Name, string(metadata.ProviderKind))
+	// Keep sub-agent child processes on the same provider we just switched to.
+	config.SetActiveProviderEnv(nextProfile.Name)
 	m.modelName = target.modelID
 	resetEffort := false
 	if m.reasoningEffort != "" && !reasoningEffortAllowed(target.reasoningEfforts, m.reasoningEffort) {
@@ -488,16 +490,16 @@ func (m model) handleModelCommand(args string) (model, string) {
 // picker calls this when a model from a non-active provider is chosen, so the
 // picker can list every saved provider and switch across them (like a unified
 // provider+model selector). The key is loaded from the encrypted store / env.
-func (m model) switchProviderModel(providerName, modelID string) (model, string) {
+func (m model) switchProviderModel(providerName, modelID string) (model, string, tea.Cmd) {
 	if m.pending {
-		return m, "Model\nCannot switch providers while a run is active."
+		return m, "Model\nCannot switch providers while a run is active.", nil
 	}
 	if m.newProvider == nil {
-		return m, "Model\nProvider rebuild is not available for this TUI session."
+		return m, "Model\nProvider rebuild is not available for this TUI session.", nil
 	}
 	target, ok := m.savedProviderByName(providerName)
 	if !ok {
-		return m, "Model\nunknown provider " + strconv.Quote(providerName)
+		return m, "Model\nunknown provider " + strconv.Quote(providerName), nil
 	}
 	target = m.profileWithCredential(target)
 	target.Model = strings.TrimSpace(modelID)
@@ -506,21 +508,35 @@ func (m model) switchProviderModel(providerName, modelID string) (model, string)
 	// was deleted/unreadable the marker can still be set, and building a keyless
 	// provider would only fail later with a 401. Local/no-auth providers need no key.
 	if strings.TrimSpace(target.APIKey) == "" && strings.TrimSpace(target.AuthHeaderValue) == "" && !(hasDescriptor && descriptor.Local) {
-		return m, "Model\nprovider " + strconv.Quote(providerName) + " has no usable credential — run setup or `zero auth login " + providerName + "`."
+		return m, "Model\nprovider " + strconv.Quote(providerName) + " has no usable credential — run setup or `zero auth login " + providerName + "`.", nil
 	}
 	next, err := m.newProvider(target)
 	if err != nil {
-		return m, "Model\n" + redaction.RedactString(err.Error(), redaction.Options{ExtraSecretValues: []string{target.APIKey}})
+		return m, "Model\n" + redaction.RedactString(err.Error(), redaction.Options{ExtraSecretValues: []string{target.APIKey}}), nil
 	}
 	m.provider = next
 	m.providerProfile = target
 	m.providerName = target.Name
 	m.modelName = target.Model
+	// Keep sub-agent child processes on the same provider we just switched to.
+	config.SetActiveProviderEnv(target.Name)
 	if strings.TrimSpace(m.userConfigPath) != "" {
 		_, _ = config.SetActiveProvider(m.userConfigPath, target.Name)
 		_, _ = config.SetProviderModel(m.userConfigPath, target.Name, target.Model)
 	}
-	return m, fmt.Sprintf("Model\nSwitched to %s · %s", target.Name, target.Model)
+	// Warm discovery for the provider we just switched to, same as Init() does
+	// for the provider active at launch — otherwise the context-usage gauge has
+	// no window for this provider until /model happens to be opened separately.
+	var cmds []tea.Cmd
+	if hasDescriptor {
+		if cmd := m.modelPickerProviderDiscoveryCmd(descriptor, target); cmd != nil {
+			cmds = append(cmds, cmd)
+		}
+		if cmd := m.ollamaContextWindowDiscoveryCmd(descriptor, target.BaseURL, target.Model); cmd != nil {
+			cmds = append(cmds, cmd)
+		}
+	}
+	return m, fmt.Sprintf("Model\nSwitched to %s · %s", target.Name, target.Model), tea.Batch(cmds...)
 }
 
 // profileWithCredential fills a profile's APIKey for provider construction the same

internal/tui/command_views.go

@@ -467,7 +467,7 @@ func providerNextActionLines(profile config.ProviderProfile, snapshot zerocomman
 }
 
 func providerProfileHasCredential(profile config.ProviderProfile) bool {
-	return strings.TrimSpace(profile.APIKey) != "" || profile.APIKeyStored || strings.TrimSpace(profile.AuthHeaderValue) != ""
+	return profile.HasConfiguredCredential()
 }
 
 func providerCredentialRequired(profile config.ProviderProfile, providerKind string) bool {

internal/tui/commands.go

@@ -40,6 +40,7 @@ const (
 	commandImage
 	commandAddDir
 	commandSelfCorrect
+	commandTurns
 	commandUnknown
 )
 
@@ -241,6 +242,13 @@ var commandDefinitions = []commandDefinition{
 		description: "Show or set post-edit self-correction depth (LSP-only default; on/tests/full add the project test plan; off/lsp disable tests, LSP-only).",
 		kind:        commandSelfCorrect,
 	},
+	{
+		name:        "/turns",
+		usage:       "/turns [n]",
+		group:       commandGroupSession,
+		description: "Show or set the per-run tool-turn budget for this session (raise it for long multi-step tasks).",
+		kind:        commandTurns,
+	},
 	{
 		name:        "/help",
 		usage:       "/help",
@@ -273,9 +281,9 @@ var commandDefinitions = []commandDefinition{
 	},
 	{
 		name:        "/theme",
-		usage:       "/theme [auto|dark|light]",
+		usage:       "/theme [list|auto|name]",
 		group:       commandGroupSession,
-		description: "Switch or show the color theme (auto-detects terminal background).",
+		description: "Pick a color theme (no arg opens the picker; auto detects the terminal background).",
 		kind:        commandTheme,
 	},
 	{

internal/tui/context_gauge_test.go

@@ -27,3 +27,29 @@ func TestContextWindowSegment(t *testing.T) {
 		t.Fatalf("gauge = %q, want 161K/200K · 81%%", got)
 	}
 }
+
+// TestStatusLineDoesNotDuplicateTokenFigureNextToGauge: the gauge's "used"
+// figure and the plain usage segment's token count both read the exact same
+// latestUsageTokens value — once the gauge is showing it, the usage segment
+// must fall back to cost-only, or the same number renders twice side by
+// side in the footer (e.g. "◔ 161K/200K · 81% │ 161K tok").
+func TestStatusLineDoesNotDuplicateTokenFigureNextToGauge(t *testing.T) {
+	m := newModel(t.Context(), Options{ModelName: "claude-sonnet-4.5"})
+	if _, err := m.usageTracker.Record(usage.RecordInput{
+		ModelID: "claude-sonnet-4.5",
+		Usage:   zeroruntime.Usage{InputTokens: 160_000, OutputTokens: 1000},
+	}); err != nil {
+		t.Fatalf("Record: %v", err)
+	}
+
+	status := plainRender(t, m.statusLine(110))
+	if !strings.Contains(status, "161K/200K") {
+		t.Fatalf("status line = %q, expected the context gauge to render", status)
+	}
+	if strings.Count(status, "161K") != 1 {
+		t.Fatalf("status line = %q, the 161K token figure must appear exactly once, not duplicated next to the gauge", status)
+	}
+	if strings.Contains(status, "tok") {
+		t.Fatalf("status line = %q, the plain token segment should be suppressed once the gauge shows the same figure", status)
+	}
+}

internal/tui/file_view.go

@@ -0,0 +1,251 @@
+// file_view.go is the drill-in view for a touched file, opened from the FILES
+// sidebar section (files_panel.go). It reuses the subchat pattern: while
+// active, the chat column's body swaps to this file's content — the sidebar,
+// composer, and scroll engine keep working unchanged (transcriptBodyItems is
+// the single source the viewport, renderer, and hit-testers all read, so
+// swapping there keeps every consumer consistent). Two modes:
+//
+//	diff (default) — the file's edit cards from this session, full-depth
+//	full           — the file as it stands on disk, syntax highlighted, with
+//	                 gutter markers on the lines this session's diffs added
+//
+// d/f switch modes (with an empty composer), Esc returns to the chat at the
+// scroll position it was left at.
+package tui
+
+import (
+	"bufio"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+)
+
+// fileViewMaxLines caps the full-file mode so a giant generated file can't
+// freeze a render; the tail collapses into a "… N more lines" trailer.
+const fileViewMaxLines = 4000
+
+const (
+	fileViewDiff = iota
+	fileViewFull
+)
+
+// fileViewState manages the drill-in view for a touched file. When active, the
+// transcript body swaps to the file's diff/content instead of the chat rows.
+type fileViewState struct {
+	active bool
+	path   string // workspace-relative, as carried by changedFiles
+	mode   int    // fileViewDiff | fileViewFull
+	// parentScrollOffset preserves the chat scroll position so closing the view
+	// returns to the same spot (mirrors subchatState).
+	parentScrollOffset int
+}
+
+// openFileView activates the drill-in for path in diff mode. Opening from an
+// already-open view (clicking another FILES row) keeps the original saved chat
+// scroll position rather than saving the file view's own offset as "parent".
+// Re-opening the file that is ALREADY being viewed is a no-op: a stray
+// re-click must not bounce the user from full mode back to diff or reset
+// their scroll position.
+func (m model) openFileView(path string) model {
+	if m.fileView.active && m.fileView.path == path {
+		return m
+	}
+	if !m.fileView.active {
+		m.fileView.parentScrollOffset = m.chatScrollOffset
+	}
+	m.fileView.active = true
+	m.fileView.path = path
+	m.fileView.mode = fileViewDiff
+	// A file only the git sweep knows about (bash/subagent mutation) has no edit
+	// cards to stack — open straight on the full file instead of a placeholder.
+	if len(m.fileViewResultRows()) == 0 {
+		m.fileView.mode = fileViewFull
+	}
+	m.chatScrollOffset = 0
+	m = m.clearHover() // bodyY numbering differs between the file body and the chat
+	return m
+}
+
+// exitFileView deactivates the view and restores the chat scroll position.
+func (m model) exitFileView() model {
+	if !m.fileView.active {
+		return m
+	}
+	m.chatScrollOffset = m.fileView.parentScrollOffset
+	m.fileView = fileViewState{}
+	m = m.clearHover()
+	return m
+}
+
+// setFileViewMode switches diff/full, resetting the scroll to the bottom-anchored
+// start since the two bodies have unrelated heights.
+func (m model) setFileViewMode(mode int) model {
+	if !m.fileView.active || m.fileView.mode == mode {
+		return m
+	}
+	m.fileView.mode = mode
+	m.chatScrollOffset = 0
+	return m
+}
+
+// fileViewNavBar renders the single-line header shown in place of the pinned
+// title bar while the view is active: the path plus the key hints. One line
+// exactly, so every scrollableTranscriptFrame computed against it agrees with
+// the title bar's geometry.
+func (m model) fileViewNavBar(width int) string {
+	mode := "diff"
+	other := "f full"
+	if m.fileView.mode == fileViewFull {
+		mode = "full"
+		other = "d diff"
+	}
+	left := zeroTheme.accent.Render("← "+truncatePathLeft(m.fileView.path, maxInt(8, width/2))) +
+		zeroTheme.faint.Render("  ·  "+mode)
+	right := zeroTheme.faint.Render(other + " · esc back")
+	return fitStyledLine(joinHeaderLine(left, right, width), width)
+}
+
+// fileViewBodyItems builds the body items the transcript machinery renders
+// while the view is active — one pre-rendered block, so scrolling and height
+// accounting flow through the exact same path as chat rows.
+func (m model) fileViewBodyItems(width int) []transcriptBodyItem {
+	var block string
+	if m.fileView.mode == fileViewFull {
+		block = m.renderFileViewFull(width)
+	} else {
+		block = m.renderFileViewDiff(width)
+	}
+	return []transcriptBodyItem{transcriptBlockBodyItem(transcriptBodyItemRow, -1, block)}
+}
+
+// fileViewResultRows returns the transcript's tool-result rows that touched the
+// viewed file, in chronological order.
+func (m model) fileViewResultRows() []transcriptRow {
+	var rows []transcriptRow
+	for _, row := range m.transcript {
+		if row.kind != rowToolResult {
+			continue
+		}
+		for _, p := range row.changedFiles {
+			if p == m.fileView.path {
+				rows = append(rows, row)
+				break
+			}
+		}
+	}
+	return rows
+}
+
+// renderFileViewDiff renders the session's edits to the file as its full-depth
+// tool cards (bodyCap 0, the detailed-view depth), stacked chronologically —
+// the same cards the chat shows, so the diffs read identically in both places.
+func (m model) renderFileViewDiff(width int) string {
+	rows := m.fileViewResultRows()
+	if len(rows) == 0 {
+		return zeroTheme.faint.Render("No recorded edits for this file in this session.")
+	}
+	rc := buildRowContext(m.transcript)
+	opts := cardRenderOptions{bodyCap: 0, cwd: m.cwd}
+	var b strings.Builder
+	for i, row := range rows {
+		if i > 0 {
+			b.WriteString("\n\n")
+		}
+		if len(rows) > 1 {
+			b.WriteString(zeroTheme.faint.Render(fmt.Sprintf("edit %d of %d", i+1, len(rows))))
+			b.WriteString("\n")
+		}
+		b.WriteString(m.renderRowModeUncached(row, width, rc, opts))
+	}
+	return b.String()
+}
+
+// renderFileViewFull renders the file as it currently stands on disk, syntax
+// highlighted, with a line-number gutter and an accent ▎ marker on the lines
+// this session's diffs added (matched by exact text — an approximation that
+// tolerates later drift; a stale marker just doesn't highlight).
+func (m model) renderFileViewFull(width int) string {
+	target := m.fileView.path
+	if !filepath.IsAbs(target) {
+		target = filepath.Join(m.cwd, target)
+	}
+	// Stream the read and stop at the cap: os.ReadFile would load a multi-GB
+	// file wholesale before any truncation, which is the exact render freeze
+	// fileViewMaxLines exists to prevent.
+	file, err := os.Open(target)
+	if err != nil {
+		return zeroTheme.faint.Render("Could not read file: " + err.Error())
+	}
+	defer file.Close()
+	var lines []string
+	truncated := false
+	scanner := bufio.NewScanner(file)
+	scanner.Buffer(make([]byte, 0, 64*1024), 16*1024*1024)
+	for scanner.Scan() {
+		if len(lines) == fileViewMaxLines {
+			truncated = true
+			break
+		}
+		lines = append(lines, scanner.Text())
+	}
+	if err := scanner.Err(); err != nil {
+		if len(lines) == 0 {
+			return zeroTheme.faint.Render("Could not read file: " + err.Error())
+		}
+		truncated = true // e.g. a single over-long line mid-file: show what we have
+	}
+
+	changed := m.fileViewChangedLines()
+	gutterW := len(fmt.Sprintf("%d", len(lines)))
+	textBudget := maxInt(8, width-gutterW-3) // gutter + space + marker column
+	// Highlight with an effectively-infinite measure so the highlighter never
+	// wraps — output lines stay 1:1 with file lines and the gutter numbering
+	// can't desync. Each line is then truncated to the column budget below.
+	display, ok := highlightCodeForPath(lines, m.fileView.path, 1<<20, nil)
+	if !ok || len(display) != len(lines) {
+		display = lines // no lexer for this path: render plain
+	}
+
+	var b strings.Builder
+	for i, line := range display {
+		line = fitStyledLine(line, textBudget)
+		if i > 0 {
+			b.WriteString("\n")
+		}
+		marker := " "
+		if changed[strings.TrimSpace(lines[i])] {
+			marker = zeroTheme.accent.Render("▎")
+		}
+		b.WriteString(zeroTheme.faintest.Render(fmt.Sprintf("%*d ", gutterW, i+1)))
+		b.WriteString(marker)
+		b.WriteString(line)
+	}
+	if truncated {
+		// No exact remaining-line count: computing one would require reading the
+		// rest of the file, defeating the bounded read above.
+		b.WriteString("\n")
+		b.WriteString(zeroTheme.faint.Render(fmt.Sprintf("… more lines (file truncated at %d for display)", len(lines))))
+	}
+	return b.String()
+}
+
+// fileViewChangedLines collects the trimmed text of every line the session's
+// diffs ADDED to the viewed file, for the full-mode gutter markers. Very short
+// lines ("}", "return", ")") are skipped: text-matching them would mark every
+// unrelated occurrence across the file, which misleads far more than a missing
+// marker on a brace line ever could.
+func (m model) fileViewChangedLines() map[string]bool {
+	changed := map[string]bool{}
+	for _, row := range m.fileViewResultRows() {
+		for _, line := range strings.Split(row.detail, "\n") {
+			if !strings.HasPrefix(line, "+") || strings.HasPrefix(line, "+++") {
+				continue
+			}
+			if text := strings.TrimSpace(strings.TrimPrefix(line, "+")); len(text) >= 4 {
+				changed[text] = true
+			}
+		}
+	}
+	return changed
+}

internal/tui/file_view_test.go

@@ -0,0 +1,288 @@
+package tui
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	tea "charm.land/bubbletea/v2"
+
+	"github.com/Gitlawb/zero/internal/agent"
+	"github.com/Gitlawb/zero/internal/sessions"
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+// TestFileViewOpenExitRestoresScroll: opening saves the chat scroll position,
+// resets it for the file body, and Esc restores it; switching files while open
+// keeps the ORIGINAL saved position (not the file view's own).
+func TestFileViewOpenExitRestoresScroll(t *testing.T) {
+	m := filesPanelTestModel()
+	m.chatScrollOffset = 12
+
+	m = m.openFileView("web/app.js")
+	if !m.fileView.active || m.fileView.mode != fileViewDiff {
+		t.Fatalf("open should activate in diff mode: %+v", m.fileView)
+	}
+	if m.chatScrollOffset != 0 || m.fileView.parentScrollOffset != 12 {
+		t.Fatalf("open should reset scroll and save the parent offset: offset=%d saved=%d", m.chatScrollOffset, m.fileView.parentScrollOffset)
+	}
+
+	m.chatScrollOffset = 5 // scrolled within the file body
+	m = m.openFileView("internal/tui/sidebar.go")
+	if m.fileView.parentScrollOffset != 12 {
+		t.Fatalf("switching files must keep the original parent offset, got %d", m.fileView.parentScrollOffset)
+	}
+
+	m = m.exitFileView()
+	if m.fileView.active || m.chatScrollOffset != 12 {
+		t.Fatalf("exit should restore the chat scroll: active=%v offset=%d", m.fileView.active, m.chatScrollOffset)
+	}
+}
+
+// TestFileViewEscAndModeKeys: Esc exits the view via the model's key handler;
+// d/f switch modes while the composer is empty and never while typing.
+func TestFileViewEscAndModeKeys(t *testing.T) {
+	m := filesPanelTestModel()
+	m = m.openFileView("web/app.js")
+
+	updated, _ := m.Update(tea.KeyPressMsg{Code: 'f', Text: "f"})
+	m = updated.(model)
+	if m.fileView.mode != fileViewFull {
+		t.Fatal("f should switch to full mode")
+	}
+	updated, _ = m.Update(tea.KeyPressMsg{Code: 'd', Text: "d"})
+	m = updated.(model)
+	if m.fileView.mode != fileViewDiff {
+		t.Fatal("d should switch back to diff mode")
+	}
+
+	// With text in the composer, d/f type as normal characters.
+	m.input.SetValue("say")
+	updated, _ = m.Update(tea.KeyPressMsg{Code: 'f', Text: "f"})
+	m = updated.(model)
+	if m.fileView.mode != fileViewDiff {
+		t.Fatal("f while typing must not hijack the composer")
+	}
+	m.input.SetValue("")
+
+	updated, _ = m.Update(tea.KeyPressMsg{Code: tea.KeyEsc})
+	m = updated.(model)
+	if m.fileView.active {
+		t.Fatal("Esc should exit the file view")
+	}
+}
+
+// TestFileViewDiffBody: diff mode stacks the file's edit cards chronologically
+// with "edit N of M" labels; a file with no recorded edits shows the quiet
+// placeholder.
+func TestFileViewDiffBody(t *testing.T) {
+	m := filesPanelTestModel()
+	m = m.openFileView("internal/tui/sidebar.go")
+	body := plainRender(t, m.renderFileViewDiff(78))
+	if !strings.Contains(body, "edit 1 of 2") || !strings.Contains(body, "edit 2 of 2") {
+		t.Fatalf("expected chronological edit labels:\n%s", body)
+	}
+	if !strings.Contains(body, "added one") || !strings.Contains(body, "three") {
+		t.Errorf("expected both diffs' content:\n%s", body)
+	}
+
+	m.fileView.path = "never/touched.go"
+	if got := plainRender(t, m.renderFileViewDiff(78)); !strings.Contains(got, "No recorded edits") {
+		t.Errorf("untouched file should show the placeholder, got:\n%s", got)
+	}
+}
+
+// TestFileViewFullBody: full mode shows the on-disk content with line numbers
+// and marks session-added lines with the gutter marker; a missing file degrades
+// to a readable error line.
+func TestFileViewFullBody(t *testing.T) {
+	dir := t.TempDir()
+	if err := os.WriteFile(filepath.Join(dir, "app.js"), []byte("let a = 1\nlet untouched = 0\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	m := filesPanelTestModel()
+	m.cwd = dir
+	m.transcript = append(m.transcript, transcriptRow{
+		kind: rowToolResult, tool: "write_file", id: "w9", status: tools.StatusOK,
+		detail:       "+let a = 1",
+		changedFiles: []string{"app.js"},
+	})
+	m = m.openFileView("app.js")
+	m = m.setFileViewMode(fileViewFull)
+
+	body := m.renderFileViewFull(78)
+	plain := plainRender(t, body)
+	if !strings.Contains(plain, "1 ") || !strings.Contains(plain, "let untouched = 0") {
+		t.Fatalf("full view should show numbered file content:\n%s", plain)
+	}
+	lines := strings.Split(plain, "\n")
+	if len(lines) != 2 {
+		t.Fatalf("one rendered line per file line, got %d:\n%s", len(lines), plain)
+	}
+	if !strings.Contains(lines[0], "▎") {
+		t.Errorf("session-added line should carry the gutter marker: %q", lines[0])
+	}
+	if strings.Contains(lines[1], "▎") {
+		t.Errorf("untouched line must not carry the marker: %q", lines[1])
+	}
+
+	m.fileView.path = "gone.js"
+	if got := plainRender(t, m.renderFileViewFull(78)); !strings.Contains(got, "Could not read file") {
+		t.Errorf("missing file should degrade to an error line, got:\n%s", got)
+	}
+}
+
+// TestFileViewSwapsTranscriptBody: while active, transcriptBodyItems returns
+// the file body (a single block) instead of the chat rows, and the pinned
+// title bar swaps to the one-line nav bar — the geometry every frame consumer
+// relies on.
+func TestFileViewSwapsTranscriptBody(t *testing.T) {
+	m := filesPanelTestModel()
+	m = m.openFileView("internal/tui/sidebar.go")
+
+	items := m.transcriptBodyItems(m.chatColumnWidth(), "")
+	if len(items) != 1 {
+		t.Fatalf("file view should swap the body to a single block item, got %d items", len(items))
+	}
+	nav := plainRender(t, m.pinnedTitleBar(m.chatColumnWidth()))
+	if !strings.Contains(nav, "sidebar.go") || !strings.Contains(nav, "esc back") {
+		t.Fatalf("nav bar should show the path and key hints: %q", nav)
+	}
+	if lines := len(viewLines(m.fileViewNavBar(m.chatColumnWidth()))); lines != 1 {
+		t.Fatalf("nav bar must be exactly one line (title-bar geometry), got %d", lines)
+	}
+
+	// The whole view renders without panicking in both modes and shows the nav.
+	if view := plainRender(t, m.transcriptView()); !strings.Contains(view, "esc back") {
+		t.Fatal("transcript view should carry the file nav bar")
+	}
+}
+
+// TestSubchatEntryClosesFileView: drilling into an AGENTS row while a file view
+// is open closes the file view first (the subchat owns the single-column view).
+func TestSubchatEntryClosesFileView(t *testing.T) {
+	store := sessions.NewStore(sessions.StoreOptions{RootDir: t.TempDir()})
+	if _, err := store.Create(sessions.CreateInput{SessionID: "sess-1"}); err != nil {
+		t.Fatal(err)
+	}
+	m := filesPanelTestModel()
+	m.sessionStore = store
+	m.swarmSessionMap = map[string]string{"subagent-1": "sess-1"}
+	m.transcript = append(m.transcript,
+		transcriptRow{kind: rowToolCall, tool: "swarm_spawn", detail: "build it", runID: 1},
+		transcriptRow{kind: rowToolResult, tool: "swarm_spawn", detail: "Spawned subagent as task subagent-1 on team default.", runID: 1},
+	)
+	m.activeRunID = 1
+	m = m.openFileView("web/app.js")
+
+	width := sidebarWidth(m.width)
+	agents := m.sidebarAgentSelectables(width)
+	if len(agents) == 0 {
+		t.Fatal("expected a clickable agent row")
+	}
+	click := testMouseClick(tea.MouseLeft, m.chatColumnWidth()+3, agents[0].lineOffset)
+	updated, _, handled := m.handleTranscriptSelectionMouse(click)
+	if !handled {
+		t.Fatal("agent row click should be handled")
+	}
+	if updated.fileView.active {
+		t.Fatal("entering the subchat should close the file view")
+	}
+	if !updated.subchat.active {
+		t.Fatal("subchat should be active")
+	}
+}
+
+// TestChangedFilesRehydration: a persisted tool-result payload's changedFiles
+// restores onto the rehydrated transcript row, so the FILES panel survives
+// /resume.
+func TestChangedFilesRehydration(t *testing.T) {
+	events := []sessions.Event{{
+		Type:    sessions.EventToolResult,
+		Payload: json.RawMessage(`{"toolCallId":"t1","name":"edit_file","status":"ok","output":"+x","changedFiles":["pkg/a.go","pkg/b.go"]}`),
+	}}
+	rows := transcriptRowsFromSessionEvents(events)
+	if len(rows) != 1 {
+		t.Fatalf("expected 1 row, got %d", len(rows))
+	}
+	got := rows[0].changedFiles
+	if len(got) != 2 || got[0] != "pkg/a.go" || got[1] != "pkg/b.go" {
+		t.Fatalf("changedFiles not rehydrated: %v", got)
+	}
+}
+
+// TestOpenFileViewSamePathIsNoOp: re-clicking the FILES row of the file already
+// being viewed must not clobber the user's mode or scroll — the old
+// unconditional openFileView bounced full mode back to diff and reset scroll.
+func TestOpenFileViewSamePathIsNoOp(t *testing.T) {
+	m := filesPanelTestModel()
+	m = m.openFileView("web/app.js")
+	m = m.setFileViewMode(fileViewFull)
+	m.chatScrollOffset = 7 // scrolled within the file body
+
+	m = m.openFileView("web/app.js")
+	if m.fileView.mode != fileViewFull {
+		t.Fatal("re-opening the same file must keep full mode")
+	}
+	if m.chatScrollOffset != 7 {
+		t.Fatalf("re-opening the same file must keep the scroll, got %d", m.chatScrollOffset)
+	}
+	// A DIFFERENT file still switches (and resets to diff mode as documented).
+	m = m.openFileView("internal/tui/sidebar.go")
+	if m.fileView.path != "internal/tui/sidebar.go" || m.fileView.mode != fileViewDiff {
+		t.Fatalf("opening another file should switch views: %+v", m.fileView)
+	}
+}
+
+// TestFileViewFullBodyTruncatesLongFile: the full view stops reading at
+// fileViewMaxLines (bounded read — a giant file must not be loaded wholesale)
+// and appends the truncation trailer.
+func TestFileViewFullBodyTruncatesLongFile(t *testing.T) {
+	dir := t.TempDir()
+	var b strings.Builder
+	for i := 0; i < fileViewMaxLines+50; i++ {
+		fmt.Fprintf(&b, "line %d\n", i)
+	}
+	if err := os.WriteFile(filepath.Join(dir, "big.txt"), []byte(b.String()), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	m := filesPanelTestModel()
+	m.cwd = dir
+	m.gitTouched = []gitSweepFile{{path: "big.txt"}}
+	m = m.openFileView("big.txt")
+
+	plain := plainRender(t, m.renderFileViewFull(80))
+	lines := strings.Split(plain, "\n")
+	if len(lines) != fileViewMaxLines+1 { // capped content + the trailer line
+		t.Fatalf("rendered %d lines, want %d content lines + 1 trailer", len(lines), fileViewMaxLines)
+	}
+	if !strings.Contains(lines[len(lines)-1], "truncated") {
+		t.Fatalf("expected the truncation trailer, got %q", lines[len(lines)-1])
+	}
+}
+
+// TestFileViewKeysDeferToBlockingModal: with a permission prompt up, Esc and
+// the d/f shortcuts belong to the prompt — the drill-in must not swallow them
+// (Esc exiting the view instead of reaching the prompt's deny handling).
+func TestFileViewKeysDeferToBlockingModal(t *testing.T) {
+	m := filesPanelTestModel()
+	m = m.openFileView("web/app.js")
+	m.pendingPermission = &pendingPermissionPrompt{
+		request: agent.PermissionRequest{ToolName: "write_file"},
+		decide:  func(agent.PermissionDecision) {},
+	}
+
+	updated, _ := m.Update(tea.KeyPressMsg{Code: 'f', Text: "f"})
+	m = updated.(model)
+	if m.fileView.mode != fileViewDiff {
+		t.Fatal("f with a permission prompt up must not switch file-view modes")
+	}
+	updated, _ = m.Update(tea.KeyPressMsg{Code: tea.KeyEsc})
+	m = updated.(model)
+	if !m.fileView.active {
+		t.Fatal("Esc with a permission prompt up must not exit the file view")
+	}
+}

internal/tui/files_git_sweep.go

@@ -0,0 +1,237 @@
+// files_git_sweep.go fills the FILES sidebar's blind spot: mutations that
+// bypass the file tools entirely — bash/exec_command scaffolding (npm create,
+// go generate, heredoc writes) and subagents editing the shared workspace. None
+// of those produce a changedFiles-carrying tool result, so the transcript-derived
+// roster (files_panel.go) never sees them. The sweep asks git instead: a
+// baseline `git status --porcelain` snapshot is taken at startup (Init), and a
+// re-run after each command tool result / turn end reports any NEWLY dirty
+// paths, which merge into the roster with a diffstat from `git diff --numstat`.
+// Pre-existing dirty state stays in the baseline and never shows; a non-git
+// workspace fails the first command and the sweep silently stays off.
+package tui
+
+import (
+	"context"
+	"strconv"
+	"strings"
+	"time"
+
+	tea "charm.land/bubbletea/v2"
+)
+
+// gitSweepTimeout bounds each background git invocation so a hung index lock
+// can never stall message delivery (the command runs off the update goroutine).
+const gitSweepTimeout = 3 * time.Second
+
+// gitSweepFile is one workspace file git reports as dirty: workspace-relative
+// path (git's own output convention, matching changedFiles), whether it is new
+// (untracked/added), and the tracked diffstat when known (0/0 for untracked).
+type gitSweepFile struct {
+	path    string
+	created bool
+	adds    int
+	dels    int
+}
+
+// gitSweepMsg carries one sweep's result. baseline marks the startup snapshot
+// (recorded as "pre-existing, never show") vs a live re-check (merged into the
+// roster). ok is false when git failed (not a repo, no git binary) — the
+// handler then marks the sweep unavailable rather than retrying every turn.
+type gitSweepMsg struct {
+	baseline bool
+	ok       bool
+	files    []gitSweepFile
+}
+
+// gitSweepCmd runs the status+numstat pair off the update goroutine. cwd is the
+// workspace root the TUI resolves paths against.
+func gitSweepCmd(parent context.Context, cwd string, baseline bool) tea.Cmd {
+	return func() tea.Msg {
+		if parent == nil {
+			parent = context.Background()
+		}
+		ctx, cancel := context.WithTimeout(parent, gitSweepTimeout)
+		defer cancel()
+		// --untracked-files=all enumerates files inside new directories; plain
+		// porcelain collapses them to "dir/", which is useless as a FILES row.
+		status, err := defaultPRCommandRunner(ctx, cwd, "git", "status", "--porcelain", "--untracked-files=all")
+		if err != nil {
+			return gitSweepMsg{baseline: baseline, ok: false}
+		}
+		files := parseGitPorcelain(status)
+		// Diffstat for tracked modifications; untracked files have no diff to
+		// stat. -z makes the output machine-parseable: NUL-terminated records, no
+		// path quoting, and renames as explicit preimage/postimage fields instead
+		// of the brace-mangled "src/{old => new}" a plain --numstat emits (which
+		// would never match a porcelain path). Best-effort: a failure (e.g. unborn
+		// HEAD in a fresh repo) keeps the file list with zero counts rather than
+		// dropping the sweep.
+		if numstat, err := defaultPRCommandRunner(ctx, cwd, "git", "diff", "HEAD", "--numstat", "-z"); err == nil {
+			stats := parseGitNumstat(numstat)
+			for i := range files {
+				if counts, ok := stats[files[i].path]; ok {
+					files[i].adds, files[i].dels = counts[0], counts[1]
+				}
+			}
+		}
+		return gitSweepMsg{baseline: baseline, ok: true, files: files}
+	}
+}
+
+// parseGitPorcelain parses `git status --porcelain` v1 output into sweep files.
+// Format: two status columns, a space, then the path ("XY path"); renames show
+// "old -> new" (keep the new side); untracked entries are "?? path".
+func parseGitPorcelain(out string) []gitSweepFile {
+	var files []gitSweepFile
+	for _, line := range strings.Split(out, "\n") {
+		if len(line) < 4 {
+			continue
+		}
+		code, path := line[:2], strings.TrimSpace(line[3:])
+		if to, _, found := cutRename(path); found {
+			path = to
+		}
+		path = unquoteGitPath(path)
+		if path == "" {
+			continue
+		}
+		files = append(files, gitSweepFile{
+			path:    path,
+			created: code == "??" || strings.Contains(code, "A"),
+		})
+	}
+	return files
+}
+
+// cutRename splits a porcelain rename value "old -> new", returning the new
+// path. found is false for ordinary (non-rename) paths.
+func cutRename(path string) (string, string, bool) {
+	if idx := strings.Index(path, " -> "); idx >= 0 {
+		return path[idx+4:], path[:idx], true
+	}
+	return "", "", false
+}
+
+// unquoteGitPath strips the quotes git wraps around paths with special
+// characters ("web/my file.js"). Escapes inside are left as-is — such a path
+// still identifies the file well enough for a sidebar row.
+func unquoteGitPath(path string) string {
+	if len(path) >= 2 && strings.HasPrefix(path, `"`) && strings.HasSuffix(path, `"`) {
+		return path[1 : len(path)-1]
+	}
+	return path
+}
+
+// parseGitNumstat parses `git diff --numstat -z` output into path → [added,
+// deleted]. -z records are NUL-terminated "added\tdeleted\tpath"; a rename/copy
+// instead ends the counts record with an EMPTY path ("added\tdeleted\t") and
+// appends two more NUL-terminated fields, preimage then postimage — the stats
+// key is the postimage (the path porcelain/changedFiles report). Paths are
+// never quoted or brace-mangled in -z mode. Binary files report "-" counts and
+// are skipped.
+func parseGitNumstat(out string) map[string][2]int {
+	stats := map[string][2]int{}
+	fields := strings.Split(out, "\x00")
+	for i := 0; i < len(fields); i++ {
+		parts := strings.SplitN(fields[i], "\t", 3)
+		if len(parts) != 3 {
+			continue
+		}
+		adds, errA := strconv.Atoi(parts[0])
+		dels, errD := strconv.Atoi(parts[1])
+		if errA != nil || errD != nil {
+			continue
+		}
+		path := parts[2]
+		if path == "" {
+			// Rename/copy record: the next two fields are preimage, postimage.
+			if i+2 >= len(fields) {
+				break // truncated output — drop the partial record
+			}
+			path = fields[i+2]
+			i += 2
+		}
+		if path != "" {
+			stats[path] = [2]int{adds, dels}
+		}
+	}
+	return stats
+}
+
+// handleGitSweepMsg folds a sweep result into the model: the baseline snapshot
+// records what was already dirty before this TUI session (those paths never
+// show), a live sweep upserts newly dirty paths into gitTouched in first-seen
+// order. A failed sweep marks git unavailable so no further sweeps are issued.
+func (m model) handleGitSweepMsg(msg gitSweepMsg) model {
+	m.gitSweepInFlight = false
+	if !msg.ok {
+		m.gitSweepUnavailable = true
+		if m.gitFileBaseline == nil {
+			m.gitFileBaseline = map[string]bool{}
+		}
+		return m
+	}
+	if msg.baseline {
+		baseline := make(map[string]bool, len(msg.files))
+		for _, f := range msg.files {
+			baseline[f.path] = true
+		}
+		m.gitFileBaseline = baseline
+		return m
+	}
+	for _, f := range msg.files {
+		if m.gitFileBaseline[f.path] {
+			continue
+		}
+		found := false
+		for i := range m.gitTouched {
+			if m.gitTouched[i].path == f.path {
+				m.gitTouched[i] = f
+				found = true
+				break
+			}
+		}
+		if !found {
+			// Copy-on-append: model copies share the backing array, so an in-place
+			// append from two update branches could alias. Rebuilding is cheap at
+			// sidebar scale.
+			m.gitTouched = append(append([]gitSweepFile(nil), m.gitTouched...), f)
+		}
+	}
+	return m
+}
+
+// maybeGitSweep issues a live sweep when one is useful and none is running:
+// the baseline exists (Init's snapshot answered), git works here, and the
+// workspace is known. Returns the (possibly nil) command to batch.
+func (m model) maybeGitSweep() (model, tea.Cmd) {
+	if m.gitSweepInFlight || m.gitSweepUnavailable || m.gitFileBaseline == nil || strings.TrimSpace(m.cwd) == "" {
+		return m, nil
+	}
+	m.gitSweepInFlight = true
+	return m, gitSweepCmd(m.ctx, m.cwd, false)
+}
+
+// gitTouchedFiles adapts the sweep results to the roster's touchedFile shape,
+// for merging under the transcript-derived entries (files_panel.go), newest
+// detection first — gitTouched upserts in first-seen order, so iterating it in
+// reverse keeps the merged FILES tail consistent with the roster's
+// newest-first ordering. No transcript row backs them (lastRowIndex -1), so
+// selecting one skips the scroll/tint and the drill-in opens on the full file.
+func (m model) gitTouchedFiles() []touchedFile {
+	if len(m.gitTouched) == 0 {
+		return nil
+	}
+	files := make([]touchedFile, 0, len(m.gitTouched))
+	for i := len(m.gitTouched) - 1; i >= 0; i-- {
+		f := m.gitTouched[i]
+		files = append(files, touchedFile{
+			path:         f.path,
+			created:      f.created,
+			adds:         f.adds,
+			dels:         f.dels,
+			lastRowIndex: -1,
+		})
+	}
+	return files
+}

internal/tui/files_git_sweep_test.go

@@ -0,0 +1,243 @@
+package tui
+
+import (
+	"os"
+	"os/exec"
+	"path/filepath"
+	"testing"
+)
+
+func TestParseGitPorcelain(t *testing.T) {
+	out := " M web/app.js\n?? web/new.css\nA  docs/added.md\nR  old.go -> pkg/new.go\n?? \"web/my file.js\"\n\n"
+	files := parseGitPorcelain(out)
+	if len(files) != 5 {
+		t.Fatalf("expected 5 entries, got %d: %+v", len(files), files)
+	}
+	byPath := map[string]gitSweepFile{}
+	for _, f := range files {
+		byPath[f.path] = f
+	}
+	if f := byPath["web/app.js"]; f.created {
+		t.Error("modified file must not read as created")
+	}
+	if f := byPath["web/new.css"]; !f.created {
+		t.Error("untracked file should read as created")
+	}
+	if f := byPath["docs/added.md"]; !f.created {
+		t.Error("index-added file should read as created")
+	}
+	if _, ok := byPath["pkg/new.go"]; !ok {
+		t.Errorf("rename should keep the new path: %+v", files)
+	}
+	if _, ok := byPath["web/my file.js"]; !ok {
+		t.Errorf("quoted path should unquote: %+v", files)
+	}
+}
+
+// TestParseGitNumstat feeds the exact `git diff --numstat -z` record shapes:
+// plain "added\tdeleted\tpath\0", binary "-\t-\tpath\0", and the rename form
+// "added\tdeleted\t\0preimage\0postimage\0" — the stats must key the postimage
+// (the path porcelain/changedFiles report), which the old newline parser lost
+// (plain --numstat brace-mangles renames, so they never matched a roster path).
+func TestParseGitNumstat(t *testing.T) {
+	out := "12\t3\tweb/app.js\x00" +
+		"-\t-\tassets/logo.png\x00" +
+		"4\t0\t\x00old.go\x00pkg/new.go\x00" +
+		"1\t1\tweb/my file.js\x00" // -z paths are never quoted
+	stats := parseGitNumstat(out)
+	if got := stats["web/app.js"]; got != [2]int{12, 3} {
+		t.Errorf("web/app.js = %v, want [12 3]", got)
+	}
+	if _, ok := stats["assets/logo.png"]; ok {
+		t.Error("binary '-' entries must be skipped")
+	}
+	if got := stats["pkg/new.go"]; got != [2]int{4, 0} {
+		t.Errorf("renamed file must key the postimage path = %v, want [4 0]", got)
+	}
+	if _, ok := stats["old.go"]; ok {
+		t.Error("the rename preimage must not appear as its own entry")
+	}
+	if got := stats["web/my file.js"]; got != [2]int{1, 1} {
+		t.Errorf("path with spaces = %v, want [1 1] (no quoting in -z mode)", got)
+	}
+	// Truncated rename record (missing postimage) must not panic or mis-key.
+	if got := parseGitNumstat("4\t0\t\x00old.go"); len(got) != 0 {
+		t.Errorf("truncated rename record should be dropped, got %v", got)
+	}
+}
+
+// TestGitSweepMergeAndBaseline: the baseline snapshot hides pre-existing dirty
+// paths; live sweeps upsert only new ones; a failed sweep latches unavailable.
+func TestGitSweepMergeAndBaseline(t *testing.T) {
+	m := model{}
+	m = m.handleGitSweepMsg(gitSweepMsg{baseline: true, ok: true, files: []gitSweepFile{{path: "dirty-before.go"}}})
+	if !m.gitFileBaseline["dirty-before.go"] {
+		t.Fatal("baseline should record pre-existing dirty paths")
+	}
+
+	m = m.handleGitSweepMsg(gitSweepMsg{ok: true, files: []gitSweepFile{
+		{path: "dirty-before.go", adds: 9},
+		{path: "kanban/board.tsx", created: true, adds: 120},
+	}})
+	if len(m.gitTouched) != 1 || m.gitTouched[0].path != "kanban/board.tsx" {
+		t.Fatalf("only newly dirty paths should merge: %+v", m.gitTouched)
+	}
+
+	// Re-sweep updates stats in place, no duplicate.
+	m = m.handleGitSweepMsg(gitSweepMsg{ok: true, files: []gitSweepFile{{path: "kanban/board.tsx", created: true, adds: 150, dels: 2}}})
+	if len(m.gitTouched) != 1 || m.gitTouched[0].adds != 150 {
+		t.Fatalf("re-sweep should upsert stats: %+v", m.gitTouched)
+	}
+
+	failed := model{}
+	failed = failed.handleGitSweepMsg(gitSweepMsg{baseline: true, ok: false})
+	if !failed.gitSweepUnavailable {
+		t.Fatal("a failed sweep should latch unavailable")
+	}
+	if _, cmd := failed.maybeGitSweep(); cmd != nil {
+		t.Fatal("no further sweeps once unavailable")
+	}
+}
+
+// TestMaybeGitSweepGating: no sweep before the baseline answers, none while one
+// is in flight, and the in-flight flag sets when one is issued.
+func TestMaybeGitSweepGating(t *testing.T) {
+	m := model{cwd: "/tmp"}
+	if _, cmd := m.maybeGitSweep(); cmd != nil {
+		t.Fatal("no baseline yet: sweep must not run")
+	}
+	m.gitFileBaseline = map[string]bool{}
+	next, cmd := m.maybeGitSweep()
+	if cmd == nil || !next.gitSweepInFlight {
+		t.Fatal("with a baseline, a sweep should be issued and marked in flight")
+	}
+	if _, again := next.maybeGitSweep(); again != nil {
+		t.Fatal("single-flight: no second sweep while one runs")
+	}
+}
+
+// TestTouchedFilesMergesGitSweep: git-discovered files append below the
+// transcript-derived entries, deduped by path (a tool-result entry wins).
+func TestTouchedFilesMergesGitSweep(t *testing.T) {
+	m := filesPanelTestModel()
+	m.gitTouched = []gitSweepFile{
+		{path: "kanban/board.tsx", created: true, adds: 120},
+		{path: "web/app.js", adds: 999}, // duplicate of a transcript entry
+	}
+	files := m.touchedFiles()
+	var kanban *touchedFile
+	appCount := 0
+	for i := range files {
+		if files[i].path == "kanban/board.tsx" {
+			kanban = &files[i]
+		}
+		if files[i].path == "web/app.js" {
+			appCount++
+			if files[i].adds == 999 {
+				t.Error("transcript-derived entry should win over the git duplicate")
+			}
+		}
+	}
+	if kanban == nil || !kanban.created || kanban.lastRowIndex != -1 {
+		t.Fatalf("git-only file should merge with created badge and no row anchor: %+v", files)
+	}
+	if appCount != 1 {
+		t.Fatalf("web/app.js should appear exactly once, got %d", appCount)
+	}
+}
+
+// TestOpenFileViewGitOnlyFallsBackToFull: a file with no edit cards (git sweep
+// discovery) opens straight on the full-file view instead of an empty diff.
+func TestOpenFileViewGitOnlyFallsBackToFull(t *testing.T) {
+	m := filesPanelTestModel()
+	m.gitTouched = []gitSweepFile{{path: "kanban/board.tsx", created: true}}
+	m = m.openFileView("kanban/board.tsx")
+	if m.fileView.mode != fileViewFull {
+		t.Fatal("git-only file should open in full mode")
+	}
+	m = m.exitFileView()
+	m = m.openFileView("web/app.js")
+	if m.fileView.mode != fileViewDiff {
+		t.Fatal("a file with edit cards still opens in diff mode")
+	}
+}
+
+// TestGitSweepCmdAgainstRealRepo: end-to-end against a real git repo — the
+// baseline sees pre-existing dirt, a post-mutation sweep reports the new file
+// as created and the modified file with its numstat.
+func TestGitSweepCmdAgainstRealRepo(t *testing.T) {
+	if _, err := exec.LookPath("git"); err != nil {
+		t.Skip("git not installed")
+	}
+	dir := t.TempDir()
+	run := func(args ...string) {
+		t.Helper()
+		cmd := exec.Command("git", args...)
+		cmd.Dir = dir
+		cmd.Env = append(os.Environ(),
+			"GIT_AUTHOR_NAME=t", "GIT_AUTHOR_EMAIL=t@t", "GIT_COMMITTER_NAME=t", "GIT_COMMITTER_EMAIL=t@t")
+		if out, err := cmd.CombinedOutput(); err != nil {
+			t.Fatalf("git %v: %v\n%s", args, err, out)
+		}
+	}
+	write := func(name, content string) {
+		t.Helper()
+		if err := os.WriteFile(filepath.Join(dir, name), []byte(content), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	run("init", "-q")
+	write("tracked.txt", "one\ntwo\n")
+	run("add", ".")
+	run("commit", "-q", "-m", "seed")
+	write("pre-existing.txt", "dirt\n") // dirty BEFORE the TUI "opens"
+
+	baseline := gitSweepCmd(nil, dir, true)().(gitSweepMsg)
+	if !baseline.ok || len(baseline.files) != 1 || baseline.files[0].path != "pre-existing.txt" {
+		t.Fatalf("baseline should see only the pre-existing dirt: %+v", baseline)
+	}
+
+	// The "agent" now scaffolds via the shell.
+	write("scaffolded.txt", "hello\n")
+	write("tracked.txt", "one\ntwo\nthree\nfour\n")
+
+	sweep := gitSweepCmd(nil, dir, false)().(gitSweepMsg)
+	if !sweep.ok {
+		t.Fatal("sweep failed")
+	}
+	byPath := map[string]gitSweepFile{}
+	for _, f := range sweep.files {
+		byPath[f.path] = f
+	}
+	if f, ok := byPath["scaffolded.txt"]; !ok || !f.created {
+		t.Fatalf("scaffolded file should report created: %+v", sweep.files)
+	}
+	if f := byPath["tracked.txt"]; f.adds != 2 || f.dels != 0 {
+		t.Fatalf("tracked.txt numstat = +%d −%d, want +2 −0", f.adds, f.dels)
+	}
+
+	// A rename WITH edits: porcelain reports the new path, and the -z numstat
+	// rename record must resolve to the same (postimage) path so the diffstat
+	// actually attaches — the plain --numstat form brace-mangles the path and
+	// the counts were silently lost.
+	run("add", ".")
+	run("commit", "-q", "-m", "second")
+	run("mv", "tracked.txt", "renamed.txt")
+	write("renamed.txt", "one\ntwo\nthree\nfour\nfive\n")
+	renameSweep := gitSweepCmd(nil, dir, false)().(gitSweepMsg)
+	if !renameSweep.ok {
+		t.Fatal("rename sweep failed")
+	}
+	renamed := map[string]gitSweepFile{}
+	for _, f := range renameSweep.files {
+		renamed[f.path] = f
+	}
+	if f, ok := renamed["renamed.txt"]; !ok || f.adds != 1 {
+		t.Fatalf("renamed file should carry its numstat under the new path, got %+v", renameSweep.files)
+	}
+
+	// Not-a-repo → ok=false (the sweep latches off).
+	if msg := gitSweepCmd(nil, t.TempDir(), false)().(gitSweepMsg); msg.ok {
+		t.Fatal("a non-repo should report ok=false")
+	}
+}

internal/tui/files_panel.go

@@ -0,0 +1,416 @@
+// files_panel.go renders the FILES section of the right context sidebar: the
+// workspace files this session has touched, newest first, with an A/M badge and
+// a +added/−removed diffstat per file, plus a pulsing row for the file whose
+// write is streaming right now. Like the swarm roster (sidebar.go), the touched
+// set is not separate model state — it is recovered on demand from the
+// transcript's tool-result rows (their changedFiles), so it survives resume for
+// free and can never drift from what the chat shows.
+//
+// Interaction (see handleTranscriptSelectionMouse): the first click on a row
+// SELECTS the file — its edit cards tint in the chat and the transcript scrolls
+// to the most recent one; a second click (or a click while the drill-in is
+// already open) opens the file view (file_view.go). Esc clears the selection.
+package tui
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+
+	tea "charm.land/bubbletea/v2"
+
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+// maxSidebarFiles caps the FILES rows so the section stays a glanceable set,
+// not a scrolling log; older files collapse into a "+N more" trailer.
+const maxSidebarFiles = 6
+
+// touchedFile is one workspace file the session has mutated, aggregated across
+// every tool-result row that touched it.
+type touchedFile struct {
+	path         string
+	adds         int  // total added lines across its diffs
+	dels         int  // total removed lines across its diffs
+	edits        int  // number of mutations
+	created      bool // first touch created the file (write_file "Created …")
+	failed       bool // the latest touch errored
+	lastRowIndex int  // transcript index of the most recent result touching it
+}
+
+// touchedFiles recovers the session's touched-file roster from the transcript,
+// most recently touched first. Recomputed on demand (the value-receiver model
+// can't persist a registry from View), mirroring swarmSpawnedAgents; the scan is
+// a single pass over the transcript per render, same order as the sidebar's
+// other sections.
+func (m model) touchedFiles() []touchedFile {
+	var files []touchedFile
+	index := map[string]int{}
+	for i, row := range m.transcript {
+		if row.kind != rowToolResult || len(row.changedFiles) == 0 {
+			continue
+		}
+		adds, dels := planDiffStat(row.detail)
+		// A multi-file result (apply_patch spanning several files) must not charge
+		// the WHOLE patch's totals to every file it touched — split the diff by its
+		// per-file headers and attribute each file its own counts.
+		var perFile map[string][2]int
+		if len(row.changedFiles) > 1 {
+			perFile = perFileDiffStats(row.detail)
+		}
+		for _, path := range row.changedFiles {
+			if path == "" {
+				continue
+			}
+			at, seen := index[path]
+			if !seen {
+				index[path] = len(files)
+				files = append(files, touchedFile{
+					path:         path,
+					created:      resultRowCreatedFile(row),
+					lastRowIndex: i,
+				})
+				at = len(files) - 1
+			}
+			fileAdds, fileDels := adds, dels
+			if perFile != nil {
+				// Headers were found: use this file's own counts. A path the split
+				// couldn't attribute shows 0/0 rather than the inflated patch totals.
+				counts := perFile[path]
+				fileAdds, fileDels = counts[0], counts[1]
+			}
+			files[at].adds += fileAdds
+			files[at].dels += fileDels
+			files[at].edits++
+			files[at].failed = row.status == tools.StatusError
+			files[at].lastRowIndex = i
+		}
+	}
+	// Most recently touched first: the file being worked on now belongs at the
+	// top, like the sidebar's ACTIVITY feed. Sorted by last touch (not reversed
+	// first-seen order): a file touched early and touched again last must list
+	// first. Stable, so same-row files keep their changedFiles order.
+	sort.SliceStable(files, func(i, j int) bool {
+		return files[i].lastRowIndex > files[j].lastRowIndex
+	})
+	// Merge the git sweep's discoveries (bash/subagent mutations that carry no
+	// changedFiles — files_git_sweep.go) below the transcript-derived entries,
+	// skipping paths a tool result already reported.
+	for _, f := range m.gitTouchedFiles() {
+		if _, seen := index[f.path]; seen {
+			continue
+		}
+		files = append(files, f)
+	}
+	return files
+}
+
+// perFileDiffStats splits a (possibly multi-file) unified diff into per-path
+// +added/−removed counts, keyed by the workspace-relative path from each
+// section's "+++ b/…" header (the "--- a/…" old side when the new side is
+// /dev/null, i.e. a deletion). Returns nil when the detail carries no file
+// headers at all (write_file/edit_file previews may be bare hunks) — the caller
+// then falls back to whole-detail totals, which are correct for a single file.
+func perFileDiffStats(detail string) map[string][2]int {
+	stats := map[string][2]int{}
+	current, oldPath := "", ""
+	normalize := func(line string, prefix string) string {
+		path := strings.TrimSpace(line[4:])
+		if tab := strings.IndexByte(path, '\t'); tab >= 0 {
+			path = path[:tab] // unified-diff convention: optional timestamp after a tab
+		}
+		path = unquoteGitPath(path)
+		if path == "/dev/null" {
+			return ""
+		}
+		return strings.TrimPrefix(path, prefix)
+	}
+	for _, line := range strings.Split(detail, "\n") {
+		switch {
+		case strings.HasPrefix(line, "--- "):
+			oldPath = normalize(line, "a/")
+		case strings.HasPrefix(line, "+++ "):
+			current = normalize(line, "b/")
+			if current == "" {
+				current = oldPath // deletion: charge the removed lines to the old path
+			}
+		case current == "":
+			// Lines before the first header (or in an unattributable section).
+		case strings.HasPrefix(line, "+"):
+			counts := stats[current]
+			counts[0]++
+			stats[current] = counts
+		case strings.HasPrefix(line, "-"):
+			counts := stats[current]
+			counts[1]++
+			stats[current] = counts
+		}
+	}
+	if len(stats) == 0 {
+		return nil
+	}
+	return stats
+}
+
+// resultRowCreatedFile reports whether a tool-result row reads as a file
+// creation ("tool result: write_file ok Created x (…)"), so the FILES row can
+// badge it A rather than M. Heuristic on the confirmation text write_file
+// already emits; a miss just shows M, never breaks anything.
+func resultRowCreatedFile(row transcriptRow) bool {
+	return strings.Contains(row.text, " Created ")
+}
+
+// liveEditingPath is the path of the file whose mutating tool call is streaming
+// its arguments RIGHT NOW (the same live write streamingToolCallView previews),
+// or "" when no mutation is in flight. It renders as a pulsing row pinned atop
+// the FILES section.
+func (m model) liveEditingPath() string {
+	if m.streamCallDecoder == nil {
+		return ""
+	}
+	switch m.streamCallName {
+	case "write_file", "edit_file", "apply_patch":
+		return m.streamCallDecoder.path
+	}
+	return ""
+}
+
+// fileHit marks a rendered FILES row (by its ABSOLUTE index inside the rendered
+// sidebar) that is clickable, carrying the file it refers to.
+type fileHit struct {
+	lineOffset int
+	path       string
+}
+
+// sidebarFilesHeader renders the FILES section header with the touched count.
+func (m model) sidebarFilesHeader(width int) string {
+	n := len(m.touchedFiles())
+	if n == 0 {
+		return sidebarHeader("FILES", width)
+	}
+	return sidebarHeaderWithCount("FILES", fmt.Sprintf("%d", n), zeroTheme.muted, width)
+}
+
+// sidebarFileLines renders the FILES section body: the live "writing" pulse row
+// first, then the touched files newest-first, capped at maxSidebarFiles with a
+// "+N more" trailer. Returns nil when the session has touched nothing (the
+// caller shows a placeholder). The paired hits slice records each clickable
+// row's index WITHIN the returned lines.
+func (m model) sidebarFileLines(width int) ([]string, []fileHit) {
+	files := m.touchedFiles()
+	live := m.liveEditingPath()
+	if len(files) == 0 && live == "" {
+		return nil, nil
+	}
+	room := maxInt(4, width-3)
+	var lines []string
+	var hits []fileHit
+
+	// The file being written this instant: an accent pulse, not yet clickable
+	// (its result row — the diff — doesn't exist until the call completes).
+	if live != "" {
+		lines = append(lines, " "+zeroTheme.accent.Render("●")+" "+
+			zeroTheme.ink.Render(truncatePathLeft(live, room)))
+	}
+
+	// Filter the live row out FIRST so the "+N more" trailer counts only rows
+	// that could actually render — counting the skipped live entry would inflate
+	// the overflow by one.
+	visible := files[:0:0]
+	for _, f := range files {
+		if f.path != live {
+			visible = append(visible, f)
+		}
+	}
+	shown := 0
+	for _, f := range visible {
+		if shown >= maxSidebarFiles {
+			lines = append(lines, "   "+zeroTheme.faint.Render(fmt.Sprintf("+%d more", len(visible)-shown)))
+			break
+		}
+		shown++
+		hits = append(hits, fileHit{lineOffset: len(lines), path: f.path})
+		lines = append(lines, m.renderFileRow(f, room))
+	}
+	return lines, hits
+}
+
+// renderFileRow renders one touched file: status badge, left-truncated path,
+// and the +/− diffstat, with the selected file carrying an accent marker so the
+// selection reads in the sidebar as well as in the chat tint.
+func (m model) renderFileRow(f touchedFile, room int) string {
+	badge := zeroTheme.muted.Render("M")
+	switch {
+	case f.failed:
+		badge = zeroTheme.red.Render("✗")
+	case f.created:
+		badge = zeroTheme.green.Render("A")
+	}
+	// Reserve the diffstat's width so the path truncates around it.
+	stat := ""
+	if f.adds > 0 || f.dels > 0 {
+		stat = fmt.Sprintf(" +%d −%d", f.adds, f.dels)
+	}
+	pathRoom := maxInt(4, room-len(stat))
+	pathStyle := zeroTheme.muted
+	lead := " "
+	if m.selectedFile == f.path {
+		lead = zeroTheme.accent.Render("▸")
+		pathStyle = zeroTheme.ink
+	}
+	line := lead + badge + " " + pathStyle.Render(truncatePathLeft(f.path, pathRoom))
+	if stat != "" {
+		line += zeroTheme.faintest.Render(stat)
+	}
+	return line
+}
+
+// sidebarFileSelectables returns the clickable FILES rows with their ABSOLUTE
+// index inside the rendered sidebar. The FILES section renders after PLAN in
+// renderContextSidebar: AGENTS header + body, blank + PLAN header + body, then
+// blank + FILES header, then the file rows. The offset accounting mirrors
+// sidebarPlanSelectables exactly, extended one section down.
+func (m model) sidebarFileSelectables(width int) []fileHit {
+	lines, hits := m.sidebarFileLines(width)
+	if len(lines) == 0 {
+		return nil
+	}
+	agentBody := len(m.sidebarAgentLines(width))
+	if agentBody == 0 {
+		agentBody = 1 // the "no agents spawned" placeholder occupies one line
+	}
+	planBody := len(m.sidebarPlanLines(width))
+	if planBody == 0 {
+		planBody = 1 // the "no active plan" placeholder occupies one line
+	}
+	base := 1 + agentBody + 2 + planBody + 2 // sections above + (blank + FILES header)
+	for i := range hits {
+		hits[i].lineOffset += base
+	}
+	return hits
+}
+
+// fileRowAtMouse maps a left-click in the context sidebar to a touched file,
+// mirroring planStepAtMouse's column/x gate. Rows truncated away by the
+// sidebar's height budget (or colliding with the token floor line) never match.
+func (m model) fileRowAtMouse(msg tea.MouseMsg) (string, bool) {
+	if !m.sidebarActive() {
+		return "", false
+	}
+	if m.setup.visible || m.providerWizard != nil || m.mcpAddWizard != nil || m.mcpManager != nil || m.picker != nil || m.suggestionsActive() {
+		return "", false
+	}
+	sidebarW := sidebarWidth(m.width)
+	if sidebarW <= 0 {
+		return "", false
+	}
+	x0 := m.chatColumnWidth() + 3 // " │ " divider between the columns
+	x, y := mouseX(msg), mouseY(msg)
+	if x < x0 || x >= x0+sidebarW {
+		return "", false
+	}
+	for _, hit := range m.sidebarFileSelectables(sidebarW) {
+		if hit.lineOffset == y && hit.lineOffset < m.height-1 && hit.path != "" {
+			return hit.path, true
+		}
+	}
+	return "", false
+}
+
+// rowTouchesSelectedFile reports whether a transcript row's mutation touched
+// the currently selected file, so its card renders with the selection tint.
+func (m model) rowTouchesSelectedFile(row transcriptRow) bool {
+	if m.selectedFile == "" || row.kind != rowToolResult {
+		return false
+	}
+	for _, path := range row.changedFiles {
+		if path == m.selectedFile {
+			return true
+		}
+	}
+	return false
+}
+
+// lastRowIndexForFile returns the transcript index of the most recent
+// tool-result row that touched path, or -1.
+func (m model) lastRowIndexForFile(path string) int {
+	for i := len(m.transcript) - 1; i >= 0; i-- {
+		row := m.transcript[i]
+		if row.kind != rowToolResult {
+			continue
+		}
+		for _, p := range row.changedFiles {
+			if p == path {
+				return i
+			}
+		}
+	}
+	return -1
+}
+
+// selectFile marks path as the selected file and scrolls the transcript so its
+// most recent edit card is in view; the card tint comes from the renderers
+// reading selectedFile (rowTouchesSelectedFile).
+func (m model) selectFile(path string) model {
+	m.selectedFile = path
+	if offset, ok := m.scrollOffsetForTranscriptRow(m.lastRowIndexForFile(path)); ok {
+		m.chatScrollOffset = offset
+		if offset == 0 {
+			m.chatBodyLines = 0
+		}
+	}
+	return m
+}
+
+// scrollOffsetForTranscriptRow computes the chatScrollOffset that places the
+// given transcript row's first rendered line at the top of the viewport (with
+// one line of breathing room), using the same layout metrics the scroll engine
+// itself uses. ok is false outside alt-screen or when the row isn't rendered
+// (e.g. skipped/collapsed).
+func (m model) scrollOffsetForTranscriptRow(rowIndex int) (int, bool) {
+	if rowIndex < 0 || !m.altScreen || m.height <= 0 {
+		return 0, false
+	}
+	width := m.chatColumnWidth()
+	items := m.transcriptBodyItems(width, "")
+	metrics := measureTranscriptBodyItems(items, m.transcriptBodyHeights)
+	startY := -1
+	for _, span := range metrics.spans {
+		if span.kind == transcriptBodyItemRow && span.rowIndex == rowIndex {
+			startY = span.startY
+			break
+		}
+	}
+	if startY < 0 {
+		return 0, false
+	}
+	frame := m.scrollableTranscriptFrame(m.pinnedTitleBar(width), m.footerView(width))
+	viewport := transcriptViewportForLayout(metrics, frame, m.chatScrollOffset)
+	// The offset counts lines below the fold: window.start == total-height-offset,
+	// so placing startY at the window top (minus one context line) means
+	// offset = total - height - (startY - 1), clamped to the scroll range.
+	offset := metrics.totalLines() - frame.bodyRect.height - (startY - 1)
+	return clampInt(offset, 0, viewport.maxOffset()), true
+}
+
+// truncatePathLeft fits a path into room cells by dropping leading components
+// ("internal/tui/sidebar.go" → "…/tui/sidebar.go"), keeping the filename — the
+// part that identifies the file — intact for as long as possible.
+func truncatePathLeft(path string, room int) string {
+	if room <= 0 {
+		return ""
+	}
+	if len(path) <= room {
+		return path
+	}
+	const ellipsis = "…/"
+	parts := strings.Split(path, "/")
+	for start := 1; start < len(parts); start++ {
+		candidate := ellipsis + strings.Join(parts[start:], "/")
+		if len(candidate) <= room {
+			return candidate
+		}
+	}
+	// Even "…/name" overflows: fall back to a right-truncated basename.
+	return truncateRunes(ellipsis+parts[len(parts)-1], room)
+}

internal/tui/files_panel_test.go

@@ -0,0 +1,317 @@
+package tui
+
+import (
+	"strings"
+	"testing"
+
+	tea "charm.land/bubbletea/v2"
+
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+// filesPanelTestModel is sidebarTestModel plus a couple of touched files: one
+// created, one edited twice (the second touch most recent).
+func filesPanelTestModel() model {
+	m := sidebarTestModel()
+	m.transcript = append(m.transcript,
+		transcriptRow{kind: rowToolResult, tool: "write_file", id: "f1", status: tools.StatusOK,
+			text:         "tool result: write_file ok Created web/app.js (10 lines).",
+			detail:       "+let a = 1\n+let b = 2",
+			changedFiles: []string{"web/app.js"}},
+		transcriptRow{kind: rowToolResult, tool: "edit_file", id: "f2", status: tools.StatusOK,
+			text:         "tool result: edit_file ok Edited internal/tui/sidebar.go.",
+			detail:       "+added one\n-removed one",
+			changedFiles: []string{"internal/tui/sidebar.go"}},
+		transcriptRow{kind: rowToolResult, tool: "edit_file", id: "f3", status: tools.StatusOK,
+			text:         "tool result: edit_file ok Edited internal/tui/sidebar.go.",
+			detail:       "+two\n+three\n-gone",
+			changedFiles: []string{"internal/tui/sidebar.go"}},
+	)
+	return m
+}
+
+// TestTouchedFilesAggregates: the roster is recovered from tool-result rows,
+// newest-touch first, with per-file diffstat totals, edit counts, and the
+// created badge from write_file's confirmation text.
+func TestTouchedFilesAggregates(t *testing.T) {
+	m := filesPanelTestModel()
+	files := m.touchedFiles()
+	if len(files) != 2 {
+		t.Fatalf("expected 2 touched files, got %d: %+v", len(files), files)
+	}
+	// sidebar.go was touched last, so it lists first.
+	if files[0].path != "internal/tui/sidebar.go" {
+		t.Fatalf("most recently touched file should list first, got %q", files[0].path)
+	}
+	if files[0].adds != 3 || files[0].dels != 2 || files[0].edits != 2 {
+		t.Errorf("sidebar.go diffstat = +%d −%d over %d edits, want +3 −2 over 2", files[0].adds, files[0].dels, files[0].edits)
+	}
+	if files[0].created {
+		t.Error("edited file must not read as created")
+	}
+	if !files[1].created {
+		t.Error("write_file 'Created …' result should mark the file created")
+	}
+	if files[0].lastRowIndex != len(m.transcript)-1 {
+		t.Errorf("lastRowIndex = %d, want the final transcript row", files[0].lastRowIndex)
+	}
+}
+
+// TestSidebarFileLinesRenderAndOverflow: rows carry the badge + left-truncated
+// path + diffstat; beyond maxSidebarFiles the tail collapses into "+N more";
+// hits index only real file rows.
+func TestSidebarFileLinesRenderAndOverflow(t *testing.T) {
+	m := filesPanelTestModel()
+	lines, hits := m.sidebarFileLines(34)
+	joined := plainRender(t, strings.Join(lines, "\n"))
+	if !strings.Contains(joined, "sidebar.go") || !strings.Contains(joined, "app.js") {
+		t.Fatalf("file rows missing paths:\n%s", joined)
+	}
+	if !strings.Contains(joined, "+3 −2") {
+		t.Errorf("expected aggregated diffstat +3 −2:\n%s", joined)
+	}
+	if !strings.Contains(joined, "A") || !strings.Contains(joined, "M") {
+		t.Errorf("expected A and M badges:\n%s", joined)
+	}
+	if len(hits) != 2 {
+		t.Fatalf("expected 2 clickable hits, got %d", len(hits))
+	}
+	if hits[0].path != "internal/tui/sidebar.go" {
+		t.Errorf("first hit should be the newest file, got %q", hits[0].path)
+	}
+
+	// Overflow: 8 distinct files -> maxSidebarFiles rows + a "+2 more" trailer.
+	over := sidebarTestModel()
+	for _, name := range []string{"a", "b", "c", "d", "e", "f", "g", "h"} {
+		over.transcript = append(over.transcript, transcriptRow{
+			kind: rowToolResult, tool: "edit_file", id: name, status: tools.StatusOK,
+			changedFiles: []string{name + ".go"},
+		})
+	}
+	overLines, overHits := over.sidebarFileLines(34)
+	if len(overHits) != maxSidebarFiles {
+		t.Fatalf("expected %d clickable rows, got %d", maxSidebarFiles, len(overHits))
+	}
+	if got := plainRender(t, overLines[len(overLines)-1]); !strings.Contains(got, "+2 more") {
+		t.Errorf("expected '+2 more' trailer, got %q", got)
+	}
+}
+
+// TestSidebarFilesLivePulseRow: while a mutating tool call streams its args,
+// the file it is writing shows as a pinned (unclickable) pulse row.
+func TestSidebarFilesLivePulseRow(t *testing.T) {
+	m := filesPanelTestModel()
+	m.streamCallName = "write_file"
+	m.streamCallDecoder = newStreamingDecoder()
+	m.streamCallDecoder.feed(`{"path":"web/new.css","content":"body{}`)
+
+	lines, hits := m.sidebarFileLines(34)
+	if got := plainRender(t, lines[0]); !strings.Contains(got, "web/new.css") {
+		t.Fatalf("live write should pin its path atop FILES, got %q", got)
+	}
+	for _, hit := range hits {
+		if hit.path == "web/new.css" {
+			t.Error("the live row must not be clickable (no result exists yet)")
+		}
+	}
+	// A non-mutating streaming call shows no pulse row.
+	m.streamCallName = "read_file"
+	if m.liveEditingPath() != "" {
+		t.Error("read_file must not read as a live edit")
+	}
+}
+
+// TestSidebarFileSelectablesMatchRenderedRows: the hit offsets computed by
+// sidebarFileSelectables must land exactly on the rendered FILES rows in
+// renderContextSidebar's output — the same invariant the mouse hit-test relies
+// on. Asserted against the real render, not a re-derivation of the math.
+func TestSidebarFileSelectablesMatchRenderedRows(t *testing.T) {
+	m := filesPanelTestModel()
+	width := sidebarWidth(m.width)
+	rendered := m.renderContextSidebar(width, m.height)
+	hits := m.sidebarFileSelectables(width)
+	if len(hits) == 0 {
+		t.Fatal("expected clickable FILES rows")
+	}
+	for _, hit := range hits {
+		if hit.lineOffset >= len(rendered) {
+			t.Fatalf("hit offset %d beyond rendered sidebar (%d lines)", hit.lineOffset, len(rendered))
+		}
+		line := plainRender(t, rendered[hit.lineOffset])
+		base := hit.path[strings.LastIndex(hit.path, "/")+1:]
+		if !strings.Contains(line, base) {
+			t.Errorf("hit for %q points at line %d which reads %q", hit.path, hit.lineOffset, line)
+		}
+	}
+}
+
+// TestFileRowClickSelectsThenOpens: the first sidebar click on a file selects
+// it (tint + scroll state, no view swap); the second click on the same file
+// opens the drill-in; clicking another file while the view is open switches it.
+func TestFileRowClickSelectsThenOpens(t *testing.T) {
+	m := filesPanelTestModel()
+	width := sidebarWidth(m.width)
+	hits := m.sidebarFileSelectables(width)
+	if len(hits) < 2 {
+		t.Fatal("expected two clickable FILES rows")
+	}
+	x := m.chatColumnWidth() + 3
+	click := func(hit fileHit) tea.MouseMsg { return testMouseClick(tea.MouseLeft, x, hit.lineOffset) }
+
+	m1, _, handled := m.handleTranscriptSelectionMouse(click(hits[0]))
+	if !handled {
+		t.Fatal("FILES row click should be handled")
+	}
+	if m1.selectedFile != hits[0].path || m1.fileView.active {
+		t.Fatalf("first click should select (not open): selected=%q active=%v", m1.selectedFile, m1.fileView.active)
+	}
+
+	m2, _, _ := m1.handleTranscriptSelectionMouse(click(hits[0]))
+	if !m2.fileView.active || m2.fileView.path != hits[0].path {
+		t.Fatalf("second click should open the file view: %+v", m2.fileView)
+	}
+
+	m3, _, _ := m2.handleTranscriptSelectionMouse(click(hits[1]))
+	if !m3.fileView.active || m3.fileView.path != hits[1].path {
+		t.Fatalf("clicking another file with the view open should switch it: %+v", m3.fileView)
+	}
+	if m3.selectedFile != hits[1].path {
+		t.Errorf("switching the view should move the selection too, got %q", m3.selectedFile)
+	}
+}
+
+// TestRowTouchesSelectedFileTint: only tool-result rows whose changedFiles
+// carry the selected path tint, and the render cache key differs across
+// selection states so the tint can't be served stale.
+func TestRowTouchesSelectedFileTint(t *testing.T) {
+	m := filesPanelTestModel()
+	row := m.transcript[len(m.transcript)-1]
+	if m.rowTouchesSelectedFile(row) {
+		t.Fatal("nothing selected: no row should tint")
+	}
+	m.selectedFile = "internal/tui/sidebar.go"
+	if !m.rowTouchesSelectedFile(row) {
+		t.Fatal("selected file's edit row should tint")
+	}
+	other := m.transcript[len(m.transcript)-3] // web/app.js
+	if m.rowTouchesSelectedFile(other) {
+		t.Fatal("an unrelated file's row must not tint")
+	}
+
+	rc := buildRowContext(m.transcript)
+	opts := cardRenderOptions{bodyCap: cardBodyMaxLines, cwd: m.cwd}
+	selectedKey, _ := m.renderRowCacheKey(row, 80, rc, opts, false)
+	m.selectedFile = ""
+	plainKey, _ := m.renderRowCacheKey(row, 80, rc, opts, false)
+	if selectedKey == plainKey {
+		t.Error("render cache key must change with the selection state")
+	}
+}
+
+// TestEscClearsFileSelectionThenNothing: Esc drops an active file selection
+// (before any run-level action) and is consumed doing so.
+func TestEscClearsFileSelectionThenNothing(t *testing.T) {
+	m := filesPanelTestModel()
+	m.selectedFile = "web/app.js"
+	updated, _ := m.Update(tea.KeyPressMsg{Code: tea.KeyEsc})
+	if got := updated.(model).selectedFile; got != "" {
+		t.Fatalf("Esc should clear the file selection, still %q", got)
+	}
+}
+
+func TestTruncatePathLeft(t *testing.T) {
+	if got := truncatePathLeft("internal/tui/sidebar.go", 40); got != "internal/tui/sidebar.go" {
+		t.Errorf("short path should pass through, got %q", got)
+	}
+	if got := truncatePathLeft("internal/tui/sidebar.go", 18); got != "…/tui/sidebar.go" {
+		t.Errorf("want …/tui/sidebar.go, got %q", got)
+	}
+	if got := truncatePathLeft("internal/tui/sidebar.go", 14); got != "…/sidebar.go" {
+		t.Errorf("want …/sidebar.go, got %q", got)
+	}
+}
+
+// TestTouchedFilesPerFileDiffStat: a multi-file result (apply_patch spanning
+// several files) must charge each file its OWN +/- counts from the per-file
+// diff sections — not the whole patch's totals to every file it touched.
+func TestTouchedFilesPerFileDiffStat(t *testing.T) {
+	m := sidebarTestModel()
+	patch := "diff --git a/web/one.js b/web/one.js\n" +
+		"--- a/web/one.js\n+++ b/web/one.js\n@@ -1,1 +1,2 @@\n+added one\n+added two\n" +
+		"diff --git a/web/two.js b/web/two.js\n" +
+		"--- a/web/two.js\n+++ b/web/two.js\n@@ -1,2 +1,1 @@\n-gone one\n-gone two\n-gone three\n"
+	m.transcript = append(m.transcript, transcriptRow{
+		kind: rowToolResult, tool: "apply_patch", id: "p1", status: tools.StatusOK,
+		detail:       patch,
+		changedFiles: []string{"web/one.js", "web/two.js"},
+	})
+	byPath := map[string]touchedFile{}
+	for _, f := range m.touchedFiles() {
+		byPath[f.path] = f
+	}
+	if f := byPath["web/one.js"]; f.adds != 2 || f.dels != 0 {
+		t.Errorf("web/one.js = +%d −%d, want its own +2 −0 (not the patch totals)", f.adds, f.dels)
+	}
+	if f := byPath["web/two.js"]; f.adds != 0 || f.dels != 3 {
+		t.Errorf("web/two.js = +%d −%d, want its own +0 −3 (not the patch totals)", f.adds, f.dels)
+	}
+}
+
+// TestTouchedFilesRetouchOrdersByLastTouch: touch a, then b, then a again — a
+// must list first. The old "reverse first-seen order" put b first because a's
+// slot kept its original position.
+func TestTouchedFilesRetouchOrdersByLastTouch(t *testing.T) {
+	m := sidebarTestModel()
+	for _, step := range []struct{ id, path string }{
+		{"r1", "a.go"}, {"r2", "b.go"}, {"r3", "a.go"},
+	} {
+		m.transcript = append(m.transcript, transcriptRow{
+			kind: rowToolResult, tool: "edit_file", id: step.id, status: tools.StatusOK,
+			changedFiles: []string{step.path},
+		})
+	}
+	files := m.touchedFiles()
+	if len(files) != 2 || files[0].path != "a.go" {
+		t.Fatalf("the re-touched file must list first, got %+v", files)
+	}
+}
+
+// TestSidebarFileLinesOverflowExcludesLiveRow: when the live-writing file is
+// also a touched file, it renders as the pulse row and is skipped from the
+// list — the "+N more" trailer must not count that skipped entry.
+func TestSidebarFileLinesOverflowExcludesLiveRow(t *testing.T) {
+	m := sidebarTestModel()
+	for _, name := range []string{"a", "b", "c", "d", "e", "f", "g", "h"} {
+		m.transcript = append(m.transcript, transcriptRow{
+			kind: rowToolResult, tool: "edit_file", id: name, status: tools.StatusOK,
+			changedFiles: []string{name + ".go"},
+		})
+	}
+	// h.go (a touched file) is ALSO the live write: 7 listable files remain,
+	// 6 render, so the trailer is +1 — not +2 (which counted the skipped live row).
+	m.streamCallName = "edit_file"
+	m.streamCallDecoder = newStreamingDecoder()
+	m.streamCallDecoder.feed(`{"path":"h.go","content":"x`)
+
+	lines, _ := m.sidebarFileLines(34)
+	trailer := plainRender(t, lines[len(lines)-1])
+	if !strings.Contains(trailer, "+1 more") {
+		t.Fatalf("overflow must exclude the live row from the count, got %q", trailer)
+	}
+}
+
+// TestSidebarHasContentForLiveWrite: the sidebar counts an in-flight first
+// write as content, so the FILES pulse is visible before any result row lands.
+func TestSidebarHasContentForLiveWrite(t *testing.T) {
+	m := sidebarTestModel()
+	m.plan.steps = nil // drop the helper's seeded plan: no agents/plan/files now
+	if m.sidebarHasContent() {
+		t.Fatal("session without agents/plan/files should have no sidebar content")
+	}
+	m.streamCallName = "write_file"
+	m.streamCallDecoder = newStreamingDecoder()
+	m.streamCallDecoder.feed(`{"path":"web/new.css","content":"body{}`)
+	if !m.sidebarHasContent() {
+		t.Fatal("a live in-flight write must count as sidebar content")
+	}
+}

internal/tui/flush.go

@@ -139,6 +139,10 @@ func (m model) drainFlushQueue() (model, tea.Cmd) {
 // surface ended.
 func (m *model) resetFlushFrontier(divider string) {
 	m.flushed = 0
+	// Renumbers every row's bodyY from the top (used by /clear, /resume, /rewind,
+	// /compact); a transcript-hover target's bodyY would otherwise risk
+	// coincidentally matching an unrelated row in the rebuilt transcript.
+	m.hover = hoverTarget{}
 	if divider != "" && m.flushedAny {
 		m.flushQueue = append(m.flushQueue, zeroTheme.faint.Render(divider))
 	}

internal/tui/flush_test.go

@@ -122,6 +122,8 @@ func TestEscCancellationLeavesVisibleMarker(t *testing.T) {
 
 	updated, _ := m.Update(testKey(tea.KeyEsc))
 	next := updated.(model)
+	updated, _ = next.Update(testKey(tea.KeyEsc))
+	next = updated.(model)
 	if !transcriptContains(next.transcript, "Run cancelled.") {
 		t.Fatalf("expected visible cancellation marker, got %#v", next.transcript)
 	}

internal/tui/hover.go

@@ -0,0 +1,124 @@
+package tui
+
+import tea "charm.land/bubbletea/v2"
+
+// hoverKind discriminates which clickable surface a hoverTarget refers to.
+type hoverKind int
+
+const (
+	hoverNone hoverKind = iota
+	// hoverTranscript: a specialist card or collapse/expand toggle header in the
+	// main transcript body (or the subchat child view), identified by bodyY.
+	hoverTranscript
+	// hoverSidebarAgent: a clickable AGENTS sidebar row, identified by sessionID.
+	hoverSidebarAgent
+	// hoverPlanStep: a plan step row in the sidebar's PLAN section, identified by
+	// step index.
+	hoverPlanStep
+	// hoverFileRow: a touched-file row in the sidebar's FILES section, identified
+	// by path.
+	hoverFileRow
+)
+
+// hoverTarget identifies the single clickable row (if any) currently under the
+// mouse cursor with no button pressed. Exactly one of bodyY/sessionID/stepIndex is
+// meaningful, discriminated by kind.
+//
+// The sidebar fields deliberately store a STABLE IDENTITY (sessionID, stepIndex),
+// not a raw line offset: the AGENTS/PLAN section sizes can change between when a
+// hover is detected and when the sidebar next renders (a swarm member's linger
+// window elapsing, a plan step completing) with no mouse motion in between to
+// re-resolve it. A cached raw offset would then silently point at whatever
+// unrelated row now occupies that slot. hoveredSidebarLineOffset (sidebar.go)
+// re-resolves the CURRENT line offset for this identity fresh on every render, so
+// a row that's gone simply doesn't highlight rather than a wrong one lighting up.
+type hoverTarget struct {
+	kind      hoverKind
+	bodyY     int    // hoverTranscript
+	sessionID string // hoverSidebarAgent
+	stepIndex int    // hoverPlanStep
+	filePath  string // hoverFileRow
+}
+
+// mouseHover reports whether msg is a plain cursor-movement event with NO button
+// pressed — a hover, not a drag. Requires AllMotion mouse reporting (see View's
+// wantsMouseCapture branch); CellMotion only reports motion while a button is held,
+// so this predicate would never match under the old mouse mode.
+func mouseHover(msg tea.MouseMsg) bool {
+	if _, ok := msg.(tea.MouseMotionMsg); !ok {
+		return false
+	}
+	return mouseEvent(msg).Button == tea.MouseNone
+}
+
+// updateHoverTarget resolves what's under the cursor for a hover motion, using
+// the SAME hit-testers and priority order as the click handlers in
+// handleTranscriptSelectionMouse's press case: a sidebar row (agent, then plan
+// step) takes priority since it's outside the chat column, then a clickable
+// transcript line. Clears the hover when nothing clickable is under the cursor.
+func (m model) updateHoverTarget(msg tea.MouseMsg) model {
+	if hit, ok := m.sidebarLineAtMouse(msg); ok {
+		return m.withHover(hoverTarget{kind: hoverSidebarAgent, sessionID: hit.sessionID})
+	}
+	if stepIndex, ok := m.planStepAtMouse(msg); ok {
+		return m.withHover(hoverTarget{kind: hoverPlanStep, stepIndex: stepIndex})
+	}
+	if path, ok := m.fileRowAtMouse(msg); ok {
+		return m.withHover(hoverTarget{kind: hoverFileRow, filePath: path})
+	}
+	if line, ok := m.transcriptLineAtMouse(msg); ok {
+		// A permission option reuses its OWN existing keyboard-cursor highlight
+		// (see hoverPermissionOption) rather than the m.hover mechanism, so there's
+		// nothing further to set here.
+		if line.permOption {
+			return m.hoverPermissionOption(line.permChoice).withHover(hoverTarget{})
+		}
+		// Only a card or a collapse/expand toggle header is "clickable" here; plain
+		// selectable text (e.g. a user/assistant message) isn't, so hovering over
+		// it must not light up as if it were.
+		if line.specialistCard || line.toggle {
+			return m.withHover(hoverTarget{kind: hoverTranscript, bodyY: line.bodyY})
+		}
+	}
+	return m.withHover(hoverTarget{})
+}
+
+// hoverPermissionOption moves the pending permission prompt's cursor to the
+// hovered option, reusing its EXISTING keyboard-navigation highlight (the same
+// one Tab/Shift+Tab/arrow keys move) instead of new styling — the popup already
+// renders whichever option .cursor points at differently. A no-op when no
+// permission prompt is pending or the choice isn't found (defensive; every
+// permOption line's choice always comes from permissionOptions(request) in the
+// first place).
+func (m model) hoverPermissionOption(choice permissionDecision) model {
+	if m.pendingPermission == nil {
+		return m
+	}
+	for index, option := range permissionOptions(m.pendingPermission.request) {
+		if option.choice == choice {
+			m.pendingPermission.cursor = index
+			return m
+		}
+	}
+	return m
+}
+
+func (m model) withHover(t hoverTarget) model {
+	m.hover = t
+	return m
+}
+
+// clearHover drops any hover highlight. Called wherever the underlying content a
+// hover target refers to may no longer mean the same thing on the next render — a
+// wheel-scroll (the cursor's mapping to a bodyY is only recomputed on the NEXT
+// motion event; without clearing, a stale bodyY could coincidentally match a
+// DIFFERENT row after the viewport shifts, highlighting the wrong thing) or a
+// subchat toggle (bodyY numbering is entirely different between the parent
+// transcript and a child session).
+func (m model) clearHover() model {
+	if m.hover.kind == hoverNone {
+		return m
+	}
+	m.hover = hoverTarget{}
+	return m
+}

internal/tui/hover_test.go

@@ -0,0 +1,284 @@
+package tui
+
+import (
+	"strings"
+	"testing"
+
+	tea "charm.land/bubbletea/v2"
+
+	"github.com/Gitlawb/zero/internal/agent"
+)
+
+func TestMouseHoverPredicate(t *testing.T) {
+	if !mouseHover(testMouseMotion(tea.MouseNone, 5, 5)) {
+		t.Error("a motion event with no button pressed must be a hover")
+	}
+	if mouseHover(testMouseMotion(tea.MouseLeft, 5, 5)) {
+		t.Error("a motion event WITH a button pressed is a drag, not a hover")
+	}
+	if mouseHover(testMouseClick(tea.MouseLeft, 5, 5)) {
+		t.Error("a click is not a hover")
+	}
+	if mouseHover(testMouseWheel(tea.MouseWheelDown, 5, 5)) {
+		t.Error("a wheel event is not a hover")
+	}
+}
+
+func TestUpdateHoverTargetOnSpecialistCard(t *testing.T) {
+	m := specialistClickTestModel(t, "child-sess-1")
+	x, y, ok := specialistCardClickPoint(m, 2)
+	if !ok {
+		t.Fatal("expected a specialist card selectable line in the layout")
+	}
+
+	updated, _ := m.Update(testMouseMotion(tea.MouseNone, x, y))
+	m = updated.(model)
+	if m.hover.kind != hoverTranscript {
+		t.Fatalf("hover.kind = %v, want hoverTranscript over a specialist card", m.hover.kind)
+	}
+}
+
+func TestUpdateHoverTargetOnToggleRow(t *testing.T) {
+	m := mouseTestModel()
+	m.transcript = appendTranscriptRow(m.transcript, transcriptRow{kind: rowReasoning, text: "private thought"})
+
+	width := m.chatColumnWidth()
+	body, selectable := m.transcriptBody(width, "")
+	start, _, top := m.transcriptViewportStart(body, width)
+	var target transcriptSelectableLine
+	for _, line := range selectable {
+		if line.toggle {
+			target = line
+			break
+		}
+	}
+	if !target.toggle {
+		t.Fatalf("expected reasoning header to be clickable, selectable=%#v", selectable)
+	}
+
+	updated, _ := m.Update(testMouseMotion(tea.MouseNone, target.textStart, top+target.bodyY-start))
+	m = updated.(model)
+	if m.hover.kind != hoverTranscript {
+		t.Fatalf("hover.kind = %v, want hoverTranscript over a toggle header", m.hover.kind)
+	}
+	if m.hover.bodyY != target.bodyY {
+		t.Fatalf("hover.bodyY = %d, want %d", m.hover.bodyY, target.bodyY)
+	}
+}
+
+func TestUpdateHoverTargetOnPlainTextIsNone(t *testing.T) {
+	m := mouseTestModel()
+	m.transcript = appendRow(m.transcript, rowUser, "hello world")
+	textY := firstTranscriptTextMouseY(t, m)
+
+	updated, _ := m.Update(testMouseMotion(tea.MouseNone, 3, textY))
+	m = updated.(model)
+	if m.hover.kind != hoverNone {
+		t.Fatalf("hover.kind = %v, want hoverNone over plain (non-clickable) text", m.hover.kind)
+	}
+}
+
+func TestUpdateHoverTargetOnSidebarAgentRow(t *testing.T) {
+	// Only a SWARM member row (a session mapped via swarmSessionMap) is clickable
+	// in the sidebar — a Task-delegation specialist row is not (sidebarAgentRows
+	// only records hits from the swarm loop). swarmSidebarTestModel builds exactly
+	// that: real conversation + a mapped swarm member session.
+	m := swarmSidebarTestModel(t, map[string]string{"subagent-1": "sess-1"})
+	if !m.sidebarActive() {
+		t.Fatal("sanity check failed: sidebar should be active with a swarm member present on a 100-col terminal")
+	}
+	hits := m.sidebarAgentSelectables(sidebarWidth(m.width))
+	if len(hits) == 0 {
+		t.Fatal("expected at least one clickable sidebar agent row")
+	}
+	x0 := m.chatColumnWidth() + 3
+
+	updated, _ := m.Update(testMouseMotion(tea.MouseNone, x0+1, hits[0].lineOffset))
+	m = updated.(model)
+	if m.hover.kind != hoverSidebarAgent {
+		t.Fatalf("hover.kind = %v, want hoverSidebarAgent", m.hover.kind)
+	}
+	if m.hover.sessionID != hits[0].sessionID {
+		t.Fatalf("hover.sessionID = %q, want %q", m.hover.sessionID, hits[0].sessionID)
+	}
+	// The bug this guards: hoverTarget must store the STABLE identity (sessionID),
+	// not a raw line offset — resolve it back to a line offset the same way the
+	// render path does (hoveredSidebarLineOffset) and confirm it lands on the
+	// SAME row that was actually hovered, not some other row at that raw index.
+	resolved, ok := m.hoveredSidebarLineOffset(sidebarWidth(m.width))
+	if !ok || resolved != hits[0].lineOffset {
+		t.Fatalf("hoveredSidebarLineOffset() = (%d, %v), want (%d, true)", resolved, ok, hits[0].lineOffset)
+	}
+}
+
+func TestUpdateHoverTargetOnPlanStep(t *testing.T) {
+	m := runningPlanModel(t, 3)
+	m.altScreen = true
+	m.height = 40
+	// sidebarActive requires real conversation; runningPlanModel only sets m.plan,
+	// leaving the transcript at its default (welcome-only, i.e. "empty").
+	m.transcript = appendRow(m.transcript, rowUser, "do something")
+	if !m.sidebarActive() {
+		t.Fatal("sanity check failed: sidebar should be active with a plan present on a 100-col terminal")
+	}
+	hits := m.sidebarPlanSelectables(sidebarWidth(m.width))
+	if len(hits) == 0 {
+		t.Fatal("expected at least one clickable plan step row")
+	}
+	x0 := m.chatColumnWidth() + 3
+
+	updated, _ := m.Update(testMouseMotion(tea.MouseNone, x0+1, hits[0].lineOffset))
+	m = updated.(model)
+	if m.hover.kind != hoverPlanStep {
+		t.Fatalf("hover.kind = %v, want hoverPlanStep", m.hover.kind)
+	}
+	if m.hover.stepIndex != hits[0].stepIndex {
+		t.Fatalf("hover.stepIndex = %d, want %d", m.hover.stepIndex, hits[0].stepIndex)
+	}
+	// The bug this guards: the render path must resolve stepIndex back to the
+	// CORRECT current lineOffset (hits[0].lineOffset, which is >= base=4, never a
+	// raw small index like 0/1/2) — not treat stepIndex itself as a line offset
+	// into the rendered sidebar (which would land on the AGENTS header/body).
+	resolved, ok := m.hoveredSidebarLineOffset(sidebarWidth(m.width))
+	if !ok || resolved != hits[0].lineOffset {
+		t.Fatalf("hoveredSidebarLineOffset() = (%d, %v), want (%d, true)", resolved, ok, hits[0].lineOffset)
+	}
+}
+
+func TestUpdateHoverTargetOnPermissionOptionMovesCursor(t *testing.T) {
+	m := mouseTestModel()
+	m.pending = true // the permission prompt only renders into the transcript body while m.pending
+	m.pendingPermission = &pendingPermissionPrompt{
+		request: agent.PermissionRequest{ToolName: "write_file"},
+		cursor:  0,
+	}
+	width := m.chatColumnWidth()
+	body, selectable := m.transcriptBody(width, "")
+	start, _, top := m.transcriptViewportStart(body, width)
+	var target transcriptSelectableLine
+	found := false
+	for _, line := range selectable {
+		if line.permOption && line.permChoice != permissionOptions(m.pendingPermission.request)[0].choice {
+			target = line
+			found = true
+			break
+		}
+	}
+	if !found {
+		t.Fatal("expected at least 2 permission options to test a non-default hover target")
+	}
+
+	updated, _ := m.Update(testMouseMotion(tea.MouseNone, target.textStart, top+target.bodyY-start))
+	m = updated.(model)
+	// Permission-option hover reuses the existing cursor highlight, not m.hover.
+	if m.hover.kind != hoverNone {
+		t.Fatalf("hover.kind = %v, want hoverNone (permission options use pendingPermission.cursor)", m.hover.kind)
+	}
+	wantIndex := -1
+	for index, option := range permissionOptions(m.pendingPermission.request) {
+		if option.choice == target.permChoice {
+			wantIndex = index
+			break
+		}
+	}
+	if m.pendingPermission.cursor != wantIndex {
+		t.Fatalf("pendingPermission.cursor = %d, want %d (the hovered option)", m.pendingPermission.cursor, wantIndex)
+	}
+}
+
+func TestHoverChangesTranscriptRenderOutput(t *testing.T) {
+	m := specialistClickTestModel(t, "child-sess-1")
+	width := m.chatColumnWidth()
+	without := m.transcriptBodyLayout(width, "").String()
+
+	x, y, ok := specialistCardClickPoint(m, 2)
+	if !ok {
+		t.Fatal("expected a specialist card selectable line in the layout")
+	}
+	updated, _ := m.Update(testMouseMotion(tea.MouseNone, x, y))
+	m = updated.(model)
+	if m.hover.kind != hoverTranscript {
+		t.Fatal("sanity check failed: hover should be set over the specialist card")
+	}
+	with := m.transcriptBodyLayout(width, "").String()
+
+	if without == with {
+		t.Fatal("rendering with a hover target set should differ from without (the highlight should show)")
+	}
+}
+
+func TestHoverChangesSidebarRenderOutput(t *testing.T) {
+	m := swarmSidebarTestModel(t, map[string]string{"subagent-1": "sess-1"})
+	width := m.chatColumnWidth()
+	height := m.height
+	without := strings.Join(m.renderContextSidebar(sidebarWidth(m.width), height), "\n")
+
+	hits := m.sidebarAgentSelectables(sidebarWidth(m.width))
+	if len(hits) == 0 {
+		t.Fatal("expected at least one clickable sidebar agent row")
+	}
+	updated, _ := m.Update(testMouseMotion(tea.MouseNone, width+3+1, hits[0].lineOffset))
+	m = updated.(model)
+	if m.hover.kind != hoverSidebarAgent {
+		t.Fatal("sanity check failed: hover should be set over the sidebar agent row")
+	}
+	with := strings.Join(m.renderContextSidebar(sidebarWidth(m.width), height), "\n")
+
+	if without == with {
+		t.Fatal("sidebar rendering with a hover target set should differ from without (the highlight should show)")
+	}
+}
+
+func TestHoverClearsOnWheelScroll(t *testing.T) {
+	m := mouseTestModel()
+	for i := 0; i < 80; i++ {
+		m.transcript = appendRow(m.transcript, rowUser, "line content")
+	}
+	m.hover = hoverTarget{kind: hoverTranscript, bodyY: 5}
+
+	updated, _ := m.Update(testMouseWheel(tea.MouseWheelUp, 0, 3))
+	m = updated.(model)
+	if m.hover.kind != hoverNone {
+		t.Fatalf("hover should clear on a wheel scroll, got %#v", m.hover)
+	}
+}
+
+func TestHoverClearsOnSubchatExit(t *testing.T) {
+	m := mouseTestModel()
+	m.subchat.active = true
+	m.subchat.childSessionID = "child-1"
+	m.subchat.childRows = appendRow(nil, rowUser, "child text")
+	m.hover = hoverTarget{kind: hoverTranscript, bodyY: 5}
+
+	updated, _ := m.Update(tea.KeyPressMsg(tea.Key{Code: tea.KeyEsc}))
+	m = updated.(model)
+	if m.subchat.active {
+		t.Fatal("Esc should exit the subchat view")
+	}
+	if m.hover.kind != hoverNone {
+		t.Fatalf("hover should clear on subchat exit (bodyY numbering differs), got %#v", m.hover)
+	}
+}
+
+// A sidebar row can disappear between when it was hovered and the next render
+// (a swarm member's linger window elapsing) with no mouse motion in between to
+// re-target the hover. hoveredSidebarLineOffset must not paint a highlight on
+// whatever unrelated row now occupies that identity's old slot — it must find
+// nothing and skip painting entirely.
+func TestHoveredSidebarLineOffsetSelfHealsWhenRowDisappears(t *testing.T) {
+	m := swarmSidebarTestModel(t, map[string]string{"subagent-1": "sess-1"})
+	hits := m.sidebarAgentSelectables(sidebarWidth(m.width))
+	if len(hits) == 0 {
+		t.Fatal("expected at least one clickable sidebar agent row")
+	}
+	m.hover = hoverTarget{kind: hoverSidebarAgent, sessionID: hits[0].sessionID}
+
+	// Simulate the member disappearing (a fresh model with no swarm member at all,
+	// same identity no longer present) without any mouse motion clearing m.hover.
+	m2 := swarmSidebarTestModel(t, nil)
+	m2.hover = m.hover
+
+	if _, ok := m2.hoveredSidebarLineOffset(sidebarWidth(m2.width)); ok {
+		t.Fatal("hoveredSidebarLineOffset should find nothing once the identity no longer resolves, not fall back to a stale/coincidental row")
+	}
+}

internal/tui/init_window_size_test.go

@@ -0,0 +1,59 @@
+package tui
+
+import (
+	"context"
+	"reflect"
+	"testing"
+
+	tea "charm.land/bubbletea/v2"
+)
+
+// TestInitRequestsWindowSize: Bubble Tea documents an initial WindowSizeMsg as
+// delivered automatically on program start, but that's the terminal
+// proactively pushing a size — if it's ever missed (a slow/unusual terminal,
+// a multiplexer, a startup race), nothing else asks again, m.height stays 0
+// forever, and transcriptView's `if m.altScreen && m.height > 0` gate falls
+// back to the unpadded, non-fullscreen render path for the whole session (the
+// alt-screen viewport never gets filled below the actual content). Init must
+// explicitly request it too, the same way it already explicitly requests the
+// background color rather than relying solely on an unprompted push.
+func TestInitRequestsWindowSize(t *testing.T) {
+	m := newModel(context.Background(), Options{ModelName: "gpt-4"})
+	cmd := m.Init()
+	if cmd == nil {
+		t.Fatal("Init() must return a non-nil command")
+	}
+
+	want := reflect.ValueOf(tea.RequestWindowSize).Pointer()
+	if !batchContainsCmd(t, cmd, want) {
+		t.Error("Init() must request the window size (tea.RequestWindowSize), not rely solely on the terminal's unprompted initial WindowSizeMsg")
+	}
+}
+
+// batchContainsCmd reports whether cmd (Init()'s single top-level
+// tea.Batch(...)) contains a sub-command matching want, compared by pointer
+// (the same technique TestThemeAutoReProbesBackground uses for
+// tea.RequestBackgroundColor — reliable for named top-level functions like
+// tea.RequestWindowSize, not closures). Invokes cmd itself once to unpack the
+// tea.BatchMsg, but only pointer-compares the sub-commands — never calls
+// them, since Init()'s batch also carries real closures (e.g. provider model
+// discovery) that make live network calls and must not fire in a unit test.
+func batchContainsCmd(t *testing.T, cmd tea.Cmd, want uintptr) bool {
+	t.Helper()
+	if cmd == nil {
+		return false
+	}
+	if reflect.ValueOf(cmd).Pointer() == want {
+		return true
+	}
+	batch, ok := cmd().(tea.BatchMsg)
+	if !ok {
+		return false
+	}
+	for _, sub := range batch {
+		if sub != nil && reflect.ValueOf(sub).Pointer() == want {
+			return true
+		}
+	}
+	return false
+}

internal/tui/keybinding_help.go

@@ -31,7 +31,7 @@ var keybindingGroups = []keybindingGroup{
 		bindings: []keybinding{
 			{"Enter", "send the message"},
 			{"Alt+Enter", "insert a newline (multi-line compose)"},
-			{"Esc", "cancel the run / dismiss a popup / clear the input"},
+			{"Esc (×2)", "cancel the run / dismiss a popup / clear the input"},
 			{"Ctrl+C", "cancel the run, then quit"},
 			{"?", "show this help (on an empty input)"},
 		},

internal/tui/model.go

@@ -25,6 +25,7 @@ import (
 	"github.com/Gitlawb/zero/internal/notify"
 	"github.com/Gitlawb/zero/internal/providerhealth"
 	"github.com/Gitlawb/zero/internal/providermodeldiscovery"
+	"github.com/Gitlawb/zero/internal/providers/providerio"
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/sessions"
 	"github.com/Gitlawb/zero/internal/streamjson"
@@ -40,41 +41,58 @@ const chatWheelScrollLines = 5
 const ctrlCExitConfirmDuration = 3 * time.Second
 const ctrlCExitConfirmText = "Press Ctrl+C again to exit"
 
+// escCancelConfirmDuration/escCancelConfirmText guard Esc cancelling a running
+// turn, mirroring ctrlCExitConfirmDuration/ctrlCExitConfirmText's pattern (same
+// window, same amber footer treatment) so the two "stop it" keybinds feel
+// consistent. Without this, a single stray Esc — e.g. meant to dismiss a
+// suggestion overlay that had already closed — silently threw away an
+// in-progress run with no chance to reconsider.
+const escCancelConfirmDuration = 3 * time.Second
+const escCancelConfirmText = "Press Esc again to cancel"
+
+// dragEdgeScrollInterval/dragEdgeScrollStep drive the smooth-glide auto-scroll
+// while a drag holds past the transcript edge (see edgeScrollDelta). A small step
+// on a short, steady cadence reads as a smooth continuous scroll; the wheel-scroll
+// tick size (chatWheelScrollLines) would jump too far per step for that.
+const dragEdgeScrollInterval = 70 * time.Millisecond
+const dragEdgeScrollStep = 1
+
 type model struct {
-	ctx                    context.Context
-	cwd                    string
-	userCommands           []usercommands.Command // file-sourced /commands (.zero/commands)
-	userConfigPath         string
-	doctorUserConfigPath   string
-	projectConfigPath      string
-	gitBranch              string
-	providerName           string
-	modelName              string
-	providerProfile        config.ProviderProfile
-	savedProviders         []config.ProviderProfile
-	provider               zeroruntime.Provider
-	newProvider            func(config.ProviderProfile) (zeroruntime.Provider, error)
-	probeProviderHealth    func(context.Context, providerhealth.Options) providerhealth.Result
-	discoverProviderModels func(context.Context, config.ProviderProfile) ([]providermodeldiscovery.Model, error)
-	registry               *tools.Registry
-	sessionStore           *sessions.Store
-	sandboxStore           *sandbox.GrantStore
-	mcpConfig              config.MCPConfig
-	mcpPermissionStore     *internalmcp.PermissionStore
-	mcpTokenStore          *internalmcp.TokenStore
-	mcpCommand             func(context.Context, []string) MCPCommandResult
-	sandboxSetupCommand    func(context.Context) SandboxSetupCommandResult
-	mcpViewStateCache      MCPViewState
-	mcpViewStateReady      bool
-	mcpCommandSeq          int
-	mcpCommandCancel       context.CancelFunc
-	sandboxSetupSeq        int
-	sandboxSetupInFlight   bool
-	doctorCommandSeq       int
-	doctorInFlight         bool
-	doctorFrame            int
-	activeSession          sessions.Metadata
-	sessionEvents          []sessions.Event
+	ctx                         context.Context
+	cwd                         string
+	userCommands                []usercommands.Command // file-sourced /commands (.zero/commands)
+	userConfigPath              string
+	doctorUserConfigPath        string
+	projectConfigPath           string
+	gitBranch                   string
+	providerName                string
+	modelName                   string
+	providerProfile             config.ProviderProfile
+	savedProviders              []config.ProviderProfile
+	provider                    zeroruntime.Provider
+	newProvider                 func(config.ProviderProfile) (zeroruntime.Provider, error)
+	probeProviderHealth         func(context.Context, providerhealth.Options) providerhealth.Result
+	discoverProviderModels      func(context.Context, config.ProviderProfile) ([]providermodeldiscovery.Model, error)
+	discoverOllamaContextWindow func(ctx context.Context, baseURL string, model string) (int, error)
+	registry                    *tools.Registry
+	sessionStore                *sessions.Store
+	sandboxStore                *sandbox.GrantStore
+	mcpConfig                   config.MCPConfig
+	mcpPermissionStore          *internalmcp.PermissionStore
+	mcpTokenStore               *internalmcp.TokenStore
+	mcpCommand                  func(context.Context, []string) MCPCommandResult
+	sandboxSetupCommand         func(context.Context) SandboxSetupCommandResult
+	mcpViewStateCache           MCPViewState
+	mcpViewStateReady           bool
+	mcpCommandSeq               int
+	mcpCommandCancel            context.CancelFunc
+	sandboxSetupSeq             int
+	sandboxSetupInFlight        bool
+	doctorCommandSeq            int
+	doctorInFlight              bool
+	doctorFrame                 int
+	activeSession               sessions.Metadata
+	sessionEvents               []sessions.Event
 	// titledSessions records session ids for which a model-generated title has
 	// already been attempted this process, so a finished turn re-fires the title
 	// generator at most once per session (even before its async result lands).
@@ -193,6 +211,22 @@ type model struct {
 	// sidebar; when set, the chat reflows to full width. Distinct from the
 	// availability conditions in sidebarAvailable (geometry / mode / overlays).
 	sidebarHidden bool
+	// selectedFile is the touched file selected by clicking its FILES sidebar
+	// row: its edit cards tint in the chat (rowTouchesSelectedFile) and a second
+	// click opens the drill-in file view. "" when nothing is selected; Esc clears.
+	selectedFile string
+	// fileView is the drill-in view for a touched file (file_view.go): while
+	// active the chat column's body shows the file's diff/content instead of the
+	// transcript, mirroring the subchat drill-in.
+	fileView fileViewState
+	// Git-sweep state (files_git_sweep.go): the startup snapshot of already-dirty
+	// paths (nil until Init's sweep answers), the newly dirty files discovered by
+	// live sweeps (bash/subagent mutations that carry no changedFiles), the
+	// single-flight guard, and the "not a git repo / no git" latch.
+	gitFileBaseline     map[string]bool
+	gitTouched          []gitSweepFile
+	gitSweepInFlight    bool
+	gitSweepUnavailable bool
 	// swarmDoneAt records when each swarm member was first seen finished (done/
 	// failed) in a swarm_status report, so the sidebar can linger it briefly with a
 	// fading ✓ before dropping it (a smooth exit, not an abrupt pop). Stamped in the
@@ -280,10 +314,39 @@ type model struct {
 	// (clickable suggestions, right-click paste, transcript select) pause while on.
 	mouseReleased       bool
 	transcriptSelection transcriptSelectionState
-	copyStatus          string
-	copyStatusSeq       int
-	exitConfirmActive   bool
-	exitConfirmSeq      int
+	// hover identifies the single clickable row (if any) currently under the
+	// mouse cursor with no button pressed, so it renders in a distinct style —
+	// the visual cue that it's clickable. Requires AllMotion mouse reporting
+	// (see wantsMouseCapture) since idle cursor movement carries no button.
+	hover             hoverTarget
+	copyStatus        string
+	copyStatusSeq     int
+	exitConfirmActive bool
+	exitConfirmSeq    int
+	// cancelConfirmActive/cancelConfirmSeq mirror exitConfirmActive/exitConfirmSeq
+	// (same seq-gated tea.Tick pattern) but guard a DIFFERENT action: Esc
+	// cancelling a running turn. The two are deliberately separate state (not a
+	// shared flag) since they're different actions with different consequences
+	// (quit the app vs. cancel the current run) that are armed by different
+	// keys — Ctrl+C and Esc respectively.
+	cancelConfirmActive bool
+	cancelConfirmSeq    int
+	// edgeScrollDelta drives the smooth-glide auto-scroll while a drag holds past
+	// the transcript's top/bottom edge: 0 when idle, else the signed per-tick step
+	// (matches transcriptEdgeScrollDelta's sign convention). A self-scheduling
+	// tea.Tick chain (see dragEdgeScrollTickCmd) keeps stepping it at a fixed small
+	// increment regardless of whether new raw mouse-motion events arrive — a
+	// terminal only reports motion on actual cursor movement, so without a timer
+	// the scroll would stop dead the instant the physical mouse holds still, even
+	// while parked past the edge. edgeScrollSeq invalidates a stale in-flight tick
+	// (mirroring exitConfirmSeq/copyStatusSeq) whenever the drag moves back into
+	// the body, releases, or the chain is otherwise stopped.
+	edgeScrollDelta int
+	edgeScrollSeq   int
+	// edgeScrollMouseX is the column the tick chain keeps extending the selection
+	// at — captured from the triggering drag since a timer tick carries no mouse
+	// position of its own.
+	edgeScrollMouseX int
 
 	// picker, when non-nil, is an open interactive selector overlay (/model,
 	// /effort with no argument). It captures ↑/↓/Enter/Esc and applies
@@ -302,6 +365,12 @@ type model struct {
 	// catalog descriptor ID), so /model shows each provider's real current models —
 	// the same list the provider-setup wizard discovers — not the static catalog.
 	modelPickerLiveByProvider map[string][]providermodeldiscovery.Model
+	// ollamaContextWindowByModel holds context-window sizes fetched from a local
+	// Ollama daemon's native /api/show endpoint (keyed by model name), for
+	// custom/local models that have no curated-catalog entry and whose
+	// OpenAI-compatible /v1/models listing doesn't carry that metadata at all —
+	// see modelContextWindow.
+	ollamaContextWindowByModel map[string]int
 
 	// pendingImages holds image attachments staged by /image for the next user
 	// turn; pendingImageLabels are their display names (base(path)) for the chip
@@ -330,6 +399,21 @@ type exitConfirmExpiredMsg struct {
 	seq int
 }
 
+// cancelConfirmExpiredMsg mirrors exitConfirmExpiredMsg for the Esc
+// cancel-a-run confirmation (see cancelConfirmActive).
+type cancelConfirmExpiredMsg struct {
+	seq int
+}
+
+// dragEdgeScrollTickMsg advances the smooth-glide auto-scroll one step (see
+// edgeScrollDelta). seq must match m.edgeScrollSeq or the tick is stale (the
+// drag moved back into the body, released, or was otherwise stopped since this
+// tick was scheduled) and is silently dropped — the self-scheduling chain simply
+// doesn't reschedule itself, so it terminates rather than ticking forever.
+type dragEdgeScrollTickMsg struct {
+	seq int
+}
+
 // toolCallStreamStartMsg / toolCallStreamDeltaMsg carry a tool call's live
 // argument stream from the agent goroutine to the update loop, so a file being
 // written renders as it streams (see streamingToolCallView).
@@ -551,7 +635,7 @@ func newModel(ctx context.Context, options Options) model {
 		}
 	}
 
-	userConfigDir, _ := os.UserConfigDir()
+	userConfigDir, _ := config.UserConfigDir()
 	loadedUserCommands := usercommands.Load(usercommands.DefaultPaths(cwd, userConfigDir))
 
 	registry := options.Registry
@@ -608,56 +692,57 @@ func newModel(ctx context.Context, options Options) model {
 	notifier.SetFocused(true)
 
 	m := model{
-		ctx:                    ctx,
-		cwd:                    cwd,
-		swarmDoneAt:            map[string]time.Time{},
-		userCommands:           loadedUserCommands,
-		composerCursorVisible:  true,
-		userConfigPath:         options.UserConfigPath,
-		doctorUserConfigPath:   doctorUserConfigPath,
-		projectConfigPath:      options.ProjectConfigPath,
-		savedProviders:         options.SavedProviders,
-		gitBranch:              gitBranch(cwd),
-		providerName:           options.ProviderName,
-		modelName:              options.ModelName,
-		providerProfile:        options.ProviderProfile,
-		favoriteModels:         favoriteModelSet(options.FavoriteModels),
-		recapsEnabled:          options.RecapsEnabled,
-		provider:               options.Provider,
-		newProvider:            options.NewProvider,
-		probeProviderHealth:    options.ProbeProviderHealth,
-		discoverProviderModels: options.DiscoverProviderModels,
-		registry:               registry,
-		sessionStore:           sessionStore,
-		sandboxStore:           sandboxStore,
-		mcpConfig:              options.MCPConfig,
-		mcpPermissionStore:     options.MCPPermissionStore,
-		mcpTokenStore:          options.MCPTokenStore,
-		mcpCommand:             options.MCPCommand,
-		sandboxSetupCommand:    options.SandboxSetupCommand,
-		agentOptions:           options.AgentOptions,
-		sessionCompactor:       options.SessionCompactor,
-		runtimeMessageSink:     options.RuntimeMessageSink,
-		permissionMode:         permissionMode,
-		reasoningEffort:        options.ReasoningEffort,
-		responseStyle:          defaultedResponseStyle(options.ResponseStyle),
-		themeMode:              resolveThemeMode(options.Theme, os.Getenv("ZERO_THEME")),
-		hasDarkBg:              true,
-		userAgent:              options.UserAgent,
-		usageTracker:           usageTracker,
-		transcript:             initialTranscript(),
-		transcriptBodyHeights:  newTranscriptBodyHeightCache(defaultTranscriptBodyHeightCacheMaxEntries),
-		prService:              prService,
-		prState:                prService.GetState(),
-		input:                  input,
-		spinner:                runSpinner,
-		now:                    time.Now,
-		notifier:               notifier,
-		altScreen:              options.AltScreen,
-		liveUsageCounts:        map[int]int{},
-		swarmSessionMap:        map[string]string{},
-		setup:                  newSetupState(options.Setup),
-		setupSave:              options.Setup.Save,
+		ctx:                         ctx,
+		cwd:                         cwd,
+		swarmDoneAt:                 map[string]time.Time{},
+		userCommands:                loadedUserCommands,
+		composerCursorVisible:       true,
+		userConfigPath:              options.UserConfigPath,
+		doctorUserConfigPath:        doctorUserConfigPath,
+		projectConfigPath:           options.ProjectConfigPath,
+		savedProviders:              options.SavedProviders,
+		gitBranch:                   gitBranch(cwd),
+		providerName:                options.ProviderName,
+		modelName:                   options.ModelName,
+		providerProfile:             options.ProviderProfile,
+		favoriteModels:              favoriteModelSet(options.FavoriteModels),
+		recapsEnabled:               options.RecapsEnabled,
+		provider:                    options.Provider,
+		newProvider:                 options.NewProvider,
+		probeProviderHealth:         options.ProbeProviderHealth,
+		discoverProviderModels:      options.DiscoverProviderModels,
+		discoverOllamaContextWindow: options.DiscoverOllamaContextWindow,
+		registry:                    registry,
+		sessionStore:                sessionStore,
+		sandboxStore:                sandboxStore,
+		mcpConfig:                   options.MCPConfig,
+		mcpPermissionStore:          options.MCPPermissionStore,
+		mcpTokenStore:               options.MCPTokenStore,
+		mcpCommand:                  options.MCPCommand,
+		sandboxSetupCommand:         options.SandboxSetupCommand,
+		agentOptions:                options.AgentOptions,
+		sessionCompactor:            options.SessionCompactor,
+		runtimeMessageSink:          options.RuntimeMessageSink,
+		permissionMode:              permissionMode,
+		reasoningEffort:             options.ReasoningEffort,
+		responseStyle:               defaultedResponseStyle(options.ResponseStyle),
+		themeMode:                   resolveThemeMode(options.Theme, os.Getenv("ZERO_THEME"), options.SavedTheme),
+		hasDarkBg:                   true,
+		userAgent:                   options.UserAgent,
+		usageTracker:                usageTracker,
+		transcript:                  initialTranscript(),
+		transcriptBodyHeights:       newTranscriptBodyHeightCache(defaultTranscriptBodyHeightCacheMaxEntries),
+		prService:                   prService,
+		prState:                     prService.GetState(),
+		input:                       input,
+		spinner:                     runSpinner,
+		now:                         time.Now,
+		notifier:                    notifier,
+		altScreen:                   options.AltScreen,
+		liveUsageCounts:             map[int]int{},
+		swarmSessionMap:             map[string]string{},
+		setup:                       newSetupState(options.Setup),
+		setupSave:                   options.Setup.Save,
 	}
 	// Apply an explicit theme immediately; auto stays on the dark default until
 	// Init's terminal background probe resolves it (see Init / BackgroundColorMsg).
@@ -722,6 +807,24 @@ func composerBlinkCmd() tea.Cmd {
 
 func (m model) Init() tea.Cmd {
 	cmds := []tea.Cmd{textinput.Blink, composerBlinkCmd()}
+	// Bubble Tea documents an initial WindowSizeMsg as delivered automatically
+	// on program start, so m.height/m.width are normally set before the first
+	// render. But that's the terminal proactively pushing a size — if it's
+	// ever missed (a slow/unusual terminal, a multiplexer, a startup race),
+	// nothing else ever asks again: m.height stays its zero value forever,
+	// `if m.altScreen && m.height > 0` (transcriptView) falls back to the
+	// unpadded, non-fullscreen render path for the rest of the session, and
+	// the alt-screen viewport never gets filled below the actual content.
+	// Explicitly requesting it here means Zero doesn't depend solely on the
+	// terminal's unprompted push — mirrors the RequestBackgroundColor request
+	// below for the same reason.
+	cmds = append(cmds, tea.RequestWindowSize)
+	// Baseline git snapshot for the FILES sidebar sweep: whatever is already
+	// dirty when the TUI opens is pre-existing state, not this session's work
+	// (files_git_sweep.go). Async; a non-git workspace just disables the sweep.
+	if strings.TrimSpace(m.cwd) != "" {
+		cmds = append(cmds, gitSweepCmd(m.ctx, m.cwd, true))
+	}
 	// In auto mode, ask the terminal for its background color; the reply arrives
 	// as tea.BackgroundColorMsg and selects light vs dark (see updateModel).
 	if m.themeMode == themeAuto {
@@ -736,6 +839,13 @@ func (m model) Init() tea.Cmd {
 		if cmd := m.modelPickerProviderDiscoveryCmd(descriptor, m.providerProfile); cmd != nil {
 			cmds = append(cmds, cmd)
 		}
+		// The generic discovery above has no source for a local Ollama model's
+		// context window (see ollamaContextWindowDiscoveryCmd); probe its
+		// native /api/show separately so the gauge works for custom/local
+		// Ollama models too, not just ones in the curated catalog.
+		if cmd := m.ollamaContextWindowDiscoveryCmd(descriptor, m.providerProfile.BaseURL, m.modelName); cmd != nil {
+			cmds = append(cmds, cmd)
+		}
 	}
 	if m.prService != nil && m.runtimeMessageSink != nil {
 		service := m.prService
@@ -812,6 +922,14 @@ func (m model) disarmExitConfirmation() model {
 	return m
 }
 
+func (m model) disarmCancelConfirmation() model {
+	if m.cancelConfirmActive {
+		m.cancelConfirmActive = false
+		m.cancelConfirmSeq++
+	}
+	return m
+}
+
 // Update routes every message through updateModel, then advances the flush
 // frontier for inline rendering. Alt-screen runs keep rows in the managed view
 // instead of printing into terminal scrollback (see flush.go).
@@ -890,6 +1008,17 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 			m.exitConfirmActive = false
 		}
 		return m, nil
+	case cancelConfirmExpiredMsg:
+		if msg.seq == m.cancelConfirmSeq {
+			m.cancelConfirmActive = false
+		}
+		return m, nil
+	case dragEdgeScrollTickMsg:
+		if msg.seq != m.edgeScrollSeq || m.edgeScrollDelta == 0 || !m.transcriptSelection.active {
+			return m, nil // stale, or the chain was stopped since this tick was scheduled
+		}
+		m = m.dragToEdgeScroll(m.edgeScrollDelta, m.edgeScrollMouseX)
+		return m, dragEdgeScrollTickCmd(m.edgeScrollSeq)
 	case providerWizardOAuthMsg:
 		return m.applyProviderWizardOAuth(msg)
 	case providerWizardDeviceCodeMsg:
@@ -931,6 +1060,13 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 		if !keyCtrl(msg, 'c') {
 			m = m.disarmExitConfirmation()
 		}
+		// Mirrors the exit-confirmation reset above: any key that isn't itself
+		// the confirming Esc means the user moved on to something else, so a
+		// later, unrelated Esc must arm a fresh confirmation rather than
+		// silently cancel off a stale press from seconds ago.
+		if !keyIs(msg, tea.KeyEsc) {
+			m = m.disarmCancelConfirmation()
+		}
 		// The `?` help overlay is modal: `?`, Esc, q, or Enter close it; every
 		// other key is swallowed so nothing types into the hidden composer.
 		if m.helpOverlay {
@@ -944,6 +1080,14 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 			return m.handleCtrlC()
 		case keyCtrl(msg, 'o'):
 			return m.toggleDetailedTranscript(), nil
+		case m.fileView.active && m.noBlockingModal() && m.composerValue() == "" && (keyText(msg) == "d" || keyText(msg) == "f"):
+			// Mode toggle for the file drill-in, only while the composer is empty
+			// (so mid-sentence typing is never hijacked) and no modal is up (so a
+			// permission prompt / ask-user / wizard keeps its own key handling).
+			if keyText(msg) == "f" {
+				return m.setFileViewMode(fileViewFull), nil
+			}
+			return m.setFileViewMode(fileViewDiff), nil
 		case keyCtrl(msg, 'e'):
 			// Release/recapture the mouse so the user can drag-select and copy text
 			// natively (mouse capture otherwise intercepts terminal selection).
@@ -953,11 +1097,29 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 			}
 			return m.appendSystemNotice("Mouse interaction re-enabled."), nil
 		case keyIs(msg, tea.KeyEsc):
+			// Esc is heavily overloaded below (subchat exit, MCP cancel, ask-user,
+			// permission deny, wizard/picker/suggestions dismiss, ...) before ever
+			// reaching the run-cancel fallback. Capture whether this press really
+			// is the confirming second Esc BEFORE any of those branches can fire,
+			// then disarm unconditionally: an Esc that gets consumed by one of
+			// them wasn't a confirm, so it must not leave cancelConfirmActive
+			// armed for some later, unrelated Esc to silently act on.
+			wasConfirmingCancel := m.pending && m.cancelConfirmActive
+			m = m.disarmCancelConfirmation()
 			// Subchat view exits on Esc (returns to main chat).
 			if m.subchat.active {
 				m.chatScrollOffset = m.subchat.exit()
+				m = m.clearHover() // bodyY numbering differs between subchat and the parent transcript
 				return m, nil
 			}
+			// File drill-in exits on Esc (returns to the chat at its saved scroll
+			// position); the file stays selected so a second Esc clears that. Only
+			// with no blocking modal up: Esc on a permission prompt / ask-user /
+			// wizard must reach THAT surface's deny/cancel handling below, not
+			// silently close the drill-in behind it.
+			if m.fileView.active && m.noBlockingModal() {
+				return m.exitFileView(), nil
+			}
 			if m.mcpCommandCancel != nil {
 				m.cancelMCPCommand()
 				if m.mcpAddWizard != nil {
@@ -1003,20 +1165,45 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 				if m.picker.kind == pickerModel {
 					m.clearModelPickerLoadState()
 				}
+				if m.picker.kind == pickerTheme {
+					// A live theme preview was applied while navigating; restore the
+					// committed palette since Esc dismisses without choosing.
+					m.restoreCommittedTheme()
+				}
 				m.picker = nil
 				return m, nil
 			}
 			if m.suggestionsActive() {
 				return m.dismissSuggestions(), nil
 			}
+			// A selected FILES row clears before anything run-related: the
+			// selection is a passive highlight, so Esc dropping it is cheap and
+			// expected (mirrors how editors clear selection on Esc).
+			if m.selectedFile != "" {
+				m.selectedFile = ""
+				return m, nil
+			}
 			if m.hasQueuedMessage() {
 				return m.clearQueuedMessage(), nil
 			}
-			m.clearComposer()
 			m.clearSuggestions()
 			if m.pending {
-				m.cancelRun()
+				if wasConfirmingCancel {
+					m.clearComposer()
+					m.cancelRun()
+					return m, nil
+				}
+				// First Esc only arms the confirmation — preserve whatever
+				// draft the user has typed, since nothing has actually been
+				// cancelled yet and they may not press Esc again.
+				m.cancelConfirmActive = true
+				m.cancelConfirmSeq++
+				seq := m.cancelConfirmSeq
+				return m, tea.Tick(escCancelConfirmDuration, func(time.Time) tea.Msg {
+					return cancelConfirmExpiredMsg{seq: seq}
+				})
 			}
+			m.clearComposer()
 			return m, nil
 		case keyIs(msg, tea.KeyEnter):
 			if m.transcriptDetailed {
@@ -1137,6 +1324,9 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 					return m, nil
 				}
 				m.picker.deleteQueryRune()
+				// Editing the filter changes which row is highlighted; keep the
+				// theme preview in sync with it (no-op for other pickers).
+				m.previewSelectedTheme()
 				return m, nil
 			}
 			// On an empty composer, Backspace removes the last attachment chip
@@ -1174,11 +1364,17 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 			if m.transcriptDetailed {
 				return m, nil
 			}
+			// A stationary mouse over a bodyY-keyed transcript hover target would
+			// otherwise stay lit at the scrolled-away bodyY until the next real
+			// motion event (see clearHover) — same reasoning as the wheel-scroll
+			// cases in mouse.go.
+			m = m.clearHover()
 			return m.scrollChat(m.chatPageScrollLines()), nil
 		case keyIs(msg, tea.KeyPgDown):
 			if m.transcriptDetailed {
 				return m, nil
 			}
+			m = m.clearHover()
 			return m.scrollChat(-m.chatPageScrollLines()), nil
 		case keyIs(msg, tea.KeyDown):
 			if m.transcriptDetailed {
@@ -1203,7 +1399,7 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 				if m.modelPickerIsLoading() {
 					return m, nil
 				}
-				m.picker.move(1)
+				m.pickerMoved(1)
 				return m, nil
 			}
 			if m.suggestionsActive() {
@@ -1220,6 +1416,7 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 			// ArrowUp exits subchat view (returns to main chat).
 			if m.subchat.active {
 				m.chatScrollOffset = m.subchat.exit()
+				m = m.clearHover() // bodyY numbering differs between subchat and the parent transcript
 				return m, nil
 			}
 			if m.transcriptDetailed {
@@ -1244,7 +1441,7 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 				if m.modelPickerIsLoading() {
 					return m, nil
 				}
-				m.picker.move(-1)
+				m.pickerMoved(-1)
 				return m, nil
 			}
 			if m.suggestionsActive() {
@@ -1303,6 +1500,9 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 			}
 			if keyPrintable(msg) {
 				m.picker.appendQuery(keyRunes(msg))
+				// Filtering changes the highlighted row; keep the theme preview in
+				// sync with it (no-op for other pickers).
+				m.previewSelectedTheme()
 			}
 			return m, nil
 		}
@@ -1445,6 +1645,10 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 	case tea.WindowSizeMsg:
 		m.width = msg.Width
 		m.height = msg.Height
+		// A resize re-wraps content at a new width, shifting every row's bodyY;
+		// a stale transcript-hover target could coincidentally land on an
+		// unrelated clickable row (same reasoning as clearHover's other callers).
+		m = m.clearHover()
 		// Reset the streaming-text fade state. A width change can re-wrap
 		// the in-progress text into a different number of visual lines,
 		// which invalidates the per-line age mapping. The next delta
@@ -1555,6 +1759,7 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 		}
 		m.clearStreamingToolCall() // active run finished — drop any lingering "writing" block
 		m.pending = false
+		m = m.disarmCancelConfirmation() // the run finished on its own — nothing left to confirm cancelling
 		// Fully reset the fade state at stream end. The next render
 		// emits the final row in solid ink (no settling animation), and
 		// the pending streamingFadeTickMsg that lands after this point
@@ -1652,8 +1857,13 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 			}
 			m, recapCmd = m.maybeRecapTurn(msg.runID, finalAnswer)
 		}
+		// End-of-turn git sweep: catch file mutations the tool stream couldn't
+		// report (bash scaffolding, subagent edits) so the FILES sidebar is
+		// complete once the turn settles.
+		var sweepCmd tea.Cmd
+		m, sweepCmd = m.maybeGitSweep()
 		next, queuedCmd := m.launchQueuedMessageIfReady()
-		return next, tea.Batch(titleCmd, recapCmd, queuedCmd)
+		return next, tea.Batch(titleCmd, recapCmd, sweepCmd, queuedCmd)
 	case sessionTitleGeneratedMsg:
 		return m.handleSessionTitleGenerated(msg)
 	case recapGeneratedMsg:
@@ -1804,6 +2014,14 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 		m.transcript = collapseRepeatedStatusCard(m.transcript, msg.row)
 		m.transcript = appendTranscriptRow(m.transcript, msg.row)
 		m = m.captureStepWork(msg.row)
+		// A finished command tool may have mutated files git can see but no
+		// changedFiles reports (npm create, heredoc writes, subagent edits) —
+		// re-sweep so the FILES sidebar picks them up mid-turn.
+		if msg.row.kind == rowToolResult && isPlanCommandTool(msg.row.tool) {
+			var sweep tea.Cmd
+			m, sweep = m.maybeGitSweep()
+			return m, sweep
+		}
 		return m, nil
 	case swarmSessionsMsg:
 		// Merge completed swarm members' session ids so their AGENTS sidebar rows
@@ -1834,6 +2052,8 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 	case prStateMsg:
 		m.prState = msg.state
 		return m, nil
+	case gitSweepMsg:
+		return m.handleGitSweepMsg(msg), nil
 	case prWatcherStartedMsg:
 		if msg.stop == nil {
 			return m, nil
@@ -1856,6 +2076,14 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 		return m.applySetupOAuthDeviceCode(msg)
 	case modelPickerModelsDiscoveredMsg:
 		return m.applyModelPickerModelsDiscovered(msg), nil
+	case ollamaContextWindowDiscoveredMsg:
+		if msg.err == nil && msg.contextWindow > 0 {
+			if m.ollamaContextWindowByModel == nil {
+				m.ollamaContextWindowByModel = map[string]int{}
+			}
+			m.ollamaContextWindowByModel[msg.modelName] = msg.contextWindow
+		}
+		return m, nil
 	case mcpCommandResultMsg:
 		return m.applyMCPCommandResultMessage(msg), nil
 	}
@@ -1879,9 +2107,27 @@ func (m model) View() tea.View {
 
 	view := tea.NewView(content)
 	view.AltScreen = m.altScreen
+	// Paint the whole frame with the active theme's surface. Zero never paints the
+	// terminal's own canvas, so without this a theme's text falls on the terminal
+	// background — fine when they share polarity, but a light theme's dark text on a
+	// dark terminal (or vice versa) is invisible, and a color theme never shows its
+	// real surface. Painting the panel makes every theme self-contained and legible
+	// on any terminal, and fills the transparent popup interiors (e.g. the /theme
+	// picker) too. Alt-screen only, so inline output never leaves a painted
+	// background behind in the user's scrollback after exit.
+	if m.altScreen {
+		view.BackgroundColor = zeroTheme.bgPanel
+	}
 	view.ReportFocus = m.notifier != nil
 	if m.wantsMouseCapture() {
-		view.MouseMode = tea.MouseModeCellMotion
+		// AllMotion (not CellMotion) is required for hover highlighting: it
+		// reports cursor movement even with no button pressed. CellMotion only
+		// reports motion while a button is held (drag) — see bubbletea's
+		// MouseMode docs. AllMotion has marginally worse terminal compatibility
+		// but is well supported by the terminals this app targets; the existing
+		// 15ms mouse-event throttle (mouseEventThrottleInterval) already bounds
+		// the redraw rate from the extra motion events.
+		view.MouseMode = tea.MouseModeAllMotion
 	}
 	return view
 }
@@ -2012,6 +2258,12 @@ func (m model) pinnedTitleBar(width int) string {
 	if !m.altScreen || m.height <= 0 {
 		return ""
 	}
+	// The file drill-in replaces the title bar with its nav line (path + key
+	// hints). Both are exactly one line, and every frame computation routes
+	// through here, so the swap never desyncs the viewport geometry.
+	if m.fileView.active {
+		return m.fileViewNavBar(width)
+	}
 	return m.titleBar(width)
 }
 
@@ -2030,10 +2282,14 @@ func (m model) footerView(width int) string {
 	// while the transcript scrolls underneath (a streaming turn no longer pushes
 	// the plan off-screen). Budgeted to at most a third of the screen height; a
 	// taller plan collapses to a one-line summary so the composer always stays
-	// on screen.
-	if plan := m.renderPinnedPlanPanel(width, m.pinnedPlanMaxHeight()); plan != "" {
-		footer.WriteString(plan)
-		footer.WriteString("\n")
+	// on screen. Skipped in the subchat drill-in: m.plan belongs to the PARENT
+	// run, not the subagent/swarm child session being viewed there, so pinning it
+	// above that composer would show unrelated state.
+	if !m.subchat.active {
+		if plan := m.renderPinnedPlanPanel(width, m.pinnedPlanMaxHeight()); plan != "" {
+			footer.WriteString(plan)
+			footer.WriteString("\n")
+		}
 	}
 	// The row above the composer: transient copy feedback takes priority; otherwise
 	// a faint idle affordance — discoverable key hints on the left, a jump-to-bottom
@@ -2045,15 +2301,18 @@ func (m model) footerView(width int) string {
 		footer.WriteString(fitStyledLine(joinHeaderLine("  "+left, right, width), width))
 	}
 	footer.WriteString("\n")
+	// A message typed while a run is active is queued for the next turn; show its
+	// preview directly ABOVE the input box (not below), so it reads as "waiting to
+	// send" sitting on top of what you're currently typing.
+	if queued := renderQueuedMessagePreview(m.queuedMessage, width); queued != "" {
+		footer.WriteString(queued)
+		footer.WriteString("\n")
+	}
 	footer.WriteString(m.composerBox(width))
 	if hint := m.composerDescriptionHint(width); hint != "" {
 		footer.WriteString("\n")
 		footer.WriteString(hint)
 	}
-	if queued := renderQueuedMessagePreview(m.queuedMessage, width); queued != "" {
-		footer.WriteString("\n")
-		footer.WriteString(queued)
-	}
 	footer.WriteString("\n")
 	footer.WriteString(m.statusLine(width))
 	return footer.String()
@@ -2629,6 +2888,15 @@ const quietWorkingHint = 8 * time.Second
 // quietGenerationHint returns a "still generating…" cue with an advancing
 // quiet-timer when the active turn has produced no streamed output for a while,
 // else "". The advancing number is itself the liveness signal.
+//
+// Past half the provider's idle timeout, the cue escalates to name what's
+// actually happening and when Zero will act on its own: a heartbeating-but-
+// silent stream (observed on chatgpt/gpt-5.x and ollama reasoning models,
+// see providerio.ErrStreamStalled) is bounded by the content-stall watchdog at
+// providerio.ContentStallTimeout(idle), but until it fires this exact same
+// plain "still generating… Xs" text is indistinguishable from a genuine hang —
+// the ticking number was the only signal, and it looks identical whether real
+// (if slow) content is coming or nothing ever will.
 func (m model) quietGenerationHint() string {
 	if m.activeRunID == 0 {
 		return ""
@@ -2644,6 +2912,10 @@ func (m model) quietGenerationHint() string {
 	if quiet < quietWorkingHint {
 		return ""
 	}
+	if idleTimeout := providerio.ResolveStreamIdleTimeout(0); idleTimeout > 0 && quiet >= idleTimeout/2 {
+		ceiling := providerio.ContentStallTimeout(idleTimeout)
+		return fmt.Sprintf("still generating… %s — unusually quiet, Zero will auto-recover by ~%s if it doesn't resume", formatWorkingElapsed(quiet), formatWorkingElapsed(ceiling))
+	}
 	return "still generating… " + formatWorkingElapsed(quiet)
 }
 
@@ -3199,14 +3471,20 @@ func (m model) choosePicker() (tea.Model, tea.Cmd) {
 	}
 	item, ok := picker.current()
 	if !ok {
+		if picker.kind == pickerTheme {
+			// No selectable row (e.g. the filter matched nothing): undo any live
+			// preview so the palette matches the committed m.themeMode.
+			m.restoreCommittedTheme()
+		}
 		return m, nil
 	}
+	var cmd tea.Cmd
 	switch picker.kind {
 	case pickerModel:
 		text := ""
 		if owner := strings.TrimSpace(item.OwnerProvider); owner != "" && !strings.EqualFold(owner, strings.TrimSpace(m.providerName)) {
 			// A model from another saved provider: switch provider + model together.
-			m, text = m.switchProviderModel(owner, item.Value)
+			m, text, cmd = m.switchProviderModel(owner, item.Value)
 		} else {
 			m, text = m.handleModelCommand(item.Value)
 		}
@@ -3223,8 +3501,20 @@ func (m model) choosePicker() (tea.Model, tea.Cmd) {
 		if text != "" {
 			m.transcript = reduceTranscript(m.transcript, transcriptAction{kind: actionAppendSystem, text: text})
 		}
+	case pickerTheme:
+		// The hovered palette is already live from the preview; handleThemeCommand
+		// records the choice (m.themeMode) and re-applies it, and reports the switch.
+		text := ""
+		m, text = m.handleThemeCommand(item.Value)
+		m.transcript = reduceTranscript(m.transcript, transcriptAction{kind: actionAppendSystem, text: text})
+		if m.themeMode == themeAuto {
+			// Re-probe the terminal background so a committed `auto` re-detects
+			// light/dark instead of reusing the preview's reading — mirrors the
+			// text /theme dispatch (M17).
+			return m, tea.RequestBackgroundColor
+		}
 	}
-	return m, nil
+	return m, cmd
 }
 
 func (m model) chooseSuggestion() (tea.Model, tea.Cmd) {
@@ -3465,7 +3755,26 @@ func (m model) handleSubmit() (tea.Model, tea.Cmd) {
 		m, text = m.handleSelfCorrectCommand(command.text)
 		m.transcript = reduceTranscript(m.transcript, transcriptAction{kind: actionAppendSystem, text: text})
 		return m, nil
+	case commandTurns:
+		// Changing the budget mid-run would mutate the inherited ZERO_MAX_TURNS env
+		// that sub-agents spawned later in THIS run read, making the run's budget
+		// inconsistent. Require an idle session (the new budget applies next run).
+		if m.pending && strings.TrimSpace(command.text) != "" {
+			m.transcript = reduceTranscript(m.transcript, transcriptAction{kind: actionAppendSystem, text: "Turns\nFinish or stop the current run before changing the tool-turn budget."})
+			return m, nil
+		}
+		text := ""
+		m, text = m.handleTurnsCommand(command.text)
+		m.transcript = reduceTranscript(m.transcript, transcriptAction{kind: actionAppendSystem, text: text})
+		return m, nil
 	case commandTheme:
+		// Bare `/theme` opens the popup picker (live preview on move, apply on
+		// Enter), matching /model and /effort. An explicit `/theme auto|dark|light`
+		// (or `/theme list`) still runs the text handler directly.
+		if strings.TrimSpace(command.text) == "" {
+			m.picker = m.newThemePicker()
+			return m, nil
+		}
 		text := ""
 		m, text = m.handleThemeCommand(command.text)
 		m.transcript = reduceTranscript(m.transcript, transcriptAction{kind: actionAppendSystem, text: text})
@@ -3729,7 +4038,8 @@ func (m *model) cancelRun() {
 	m.pending = false
 	m.runCancel = nil
 	m.activeRunID = 0
-	m.plan.frozenAt = m.now() // freeze the plan clock while idle (no run in flight)
+	m.cancelConfirmActive = false // whatever path got here, there's nothing left to confirm cancelling
+	m.plan.frozenAt = m.now()     // freeze the plan clock while idle (no run in flight)
 	m.pendingPermission = nil
 	m.pendingAskUser = nil
 	// The interim block renders streamingText live; a cancelled run's partial
@@ -4054,13 +4364,14 @@ func (m model) runAgentWithOptions(runID int, runCtx context.Context, prompt str
 				}
 			}
 			row := transcriptRow{
-				kind:   rowToolResult,
-				id:     effectiveToolRowID(result.ToolCallID, callSeq[result.ToolCallID]),
-				text:   toolResultRowText(result),
-				tool:   result.Name,
-				status: result.Status,
-				detail: toolResultDetail(result),
-				runID:  runID,
+				kind:         rowToolResult,
+				id:           effectiveToolRowID(result.ToolCallID, callSeq[result.ToolCallID]),
+				text:         toolResultRowText(result),
+				tool:         result.Name,
+				status:       result.Status,
+				detail:       toolResultDetail(result),
+				runID:        runID,
+				changedFiles: result.ChangedFiles,
 			}
 			// A Task result is shown by the specialist card, and update_plan by the
 			// plan panel/sidebar, so skip both redundant transcript rows.

internal/tui/model_catalog.go

@@ -84,6 +84,13 @@ func (m model) modelContextWindow(modelName string) int {
 			return window
 		}
 	}
+	// A custom/local Ollama model tag has no curated-catalog entry, and its
+	// OpenAI-compatible /v1/models listing (the discoveredContextWindow source
+	// above) doesn't carry context-window metadata at all — see
+	// ollamaContextWindowDiscoveryCmd, the only other source for this.
+	if window := m.ollamaContextWindowByModel[trimmed]; window > 0 {
+		return window
+	}
 	// Unknown: report 0 so display gauges show no denominator. Compaction enablement
 	// applies its own positive fallback via modelregistry.AgentContextWindow.
 	return 0

internal/tui/model_test.go

@@ -1108,7 +1108,87 @@ func TestPromptSubmitDoesNotStartAnotherRunWhilePending(t *testing.T) {
 	}
 }
 
-func TestEscCancelsPendingRun(t *testing.T) {
+func TestEscRequiresSecondPressToCancelPendingRun(t *testing.T) {
+	m := newModel(context.Background(), Options{})
+	cancelled := false
+	m.pending = true
+	m.activeRunID = 1
+	m.runCancel = func() { cancelled = true }
+
+	updated, cmd := m.Update(testKey(tea.KeyEsc))
+	next := updated.(model)
+
+	if cancelled {
+		t.Fatal("first Esc should not cancel the pending run")
+	}
+	if !next.pending {
+		t.Fatal("first Esc should leave the run pending")
+	}
+	if !next.cancelConfirmActive {
+		t.Fatal("first Esc should arm cancel confirmation")
+	}
+	if cmd == nil {
+		t.Fatal("first Esc should schedule confirmation expiry")
+	}
+	status := plainRender(t, next.statusLine(80))
+	if !strings.Contains(status, escCancelConfirmText) {
+		t.Fatalf("status line = %q, want cancel confirmation", status)
+	}
+
+	updated, _ = next.Update(testKey(tea.KeyEsc))
+	next = updated.(model)
+
+	if !cancelled {
+		t.Fatal("second Esc should cancel pending run")
+	}
+	if next.pending {
+		t.Fatal("expected second Esc to clear pending state")
+	}
+	if next.activeRunID != 0 || next.runCancel != nil {
+		t.Fatalf("expected active run state to clear, got id=%d cancel=%v", next.activeRunID, next.runCancel)
+	}
+	if next.cancelConfirmActive {
+		t.Fatal("cancelling should clear cancel confirmation")
+	}
+}
+
+// TestEscArmingCancelConfirmationPreservesComposerDraft: the first Esc only
+// arms the confirmation — nothing has actually been cancelled yet, so it
+// must not destroy a draft the user is still typing. Only the confirming
+// second Esc (the one that actually cancels the run) clears the composer.
+func TestEscArmingCancelConfirmationPreservesComposerDraft(t *testing.T) {
+	m := newModel(context.Background(), Options{})
+	m.pending = true
+	m.activeRunID = 1
+	m.runCancel = func() {}
+	m.input.SetValue("draft prompt")
+
+	updated, _ := m.Update(testKey(tea.KeyEsc))
+	next := updated.(model)
+
+	if !next.cancelConfirmActive {
+		t.Fatal("first Esc should arm cancel confirmation")
+	}
+	if next.composerValue() != "draft prompt" {
+		t.Fatalf("first Esc should preserve the draft, got %q", next.composerValue())
+	}
+
+	updated, _ = next.Update(testKey(tea.KeyEsc))
+	next = updated.(model)
+
+	if next.pending {
+		t.Fatal("second Esc should cancel the pending run")
+	}
+	if next.composerValue() != "" {
+		t.Fatalf("the confirming second Esc should clear the draft, got %q", next.composerValue())
+	}
+}
+
+// TestPasteDisarmsCancelConfirmation: a paste isn't a keypress, so it never
+// went through the generic "any non-Esc key disarms it" hook. Pasting is a
+// deliberate action just like typing or clicking — it must disarm a stale
+// confirmation too, or a later, unrelated Esc could silently cancel the run.
+func TestPasteDisarmsCancelConfirmation(t *testing.T) {
 	m := newModel(context.Background(), Options{})
 	cancelled := false
 	m.pending = true
@@ -1117,15 +1197,131 @@ func TestEscCancelsPendingRun(t *testing.T) {
 
 	updated, _ := m.Update(testKey(tea.KeyEsc))
 	next := updated.(model)
+	if !next.cancelConfirmActive {
+		t.Fatal("first Esc should arm cancel confirmation")
+	}
 
-	if !cancelled {
-		t.Fatal("expected Esc to cancel pending run")
+	updated, _ = next.Update(testPaste("pasted text"))
+	next = updated.(model)
+	if next.cancelConfirmActive {
+		t.Fatal("a paste should disarm the stale cancel confirmation")
 	}
-	if next.pending {
-		t.Fatal("expected Esc to clear pending state")
+
+	updated, _ = next.Update(testKey(tea.KeyEsc))
+	next = updated.(model)
+	if cancelled {
+		t.Fatal("Esc after a paste should re-arm, not immediately cancel")
 	}
-	if next.activeRunID != 0 || next.runCancel != nil {
-		t.Fatalf("expected active run state to clear, got id=%d cancel=%v", next.activeRunID, next.runCancel)
+	if !next.cancelConfirmActive {
+		t.Fatal("Esc after a paste should arm a fresh confirmation")
+	}
+}
+
+func TestEscCancelConfirmationExpires(t *testing.T) {
+	m := newModel(context.Background(), Options{})
+	m.pending = true
+	m.activeRunID = 1
+	m.runCancel = func() {}
+
+	updated, _ := m.Update(testKey(tea.KeyEsc))
+	next := updated.(model)
+	seq := next.cancelConfirmSeq
+
+	updated, _ = next.Update(cancelConfirmExpiredMsg{seq: seq - 1})
+	next = updated.(model)
+	if !next.cancelConfirmActive {
+		t.Fatal("stale expiry should not clear active cancel confirmation")
+	}
+
+	updated, _ = next.Update(cancelConfirmExpiredMsg{seq: seq})
+	next = updated.(model)
+	if next.cancelConfirmActive {
+		t.Fatal("matching expiry should clear cancel confirmation")
+	}
+	if !next.pending {
+		t.Fatal("expiring the confirmation should not cancel the run")
+	}
+}
+
+// TestEscCancelConfirmationDisarmsOnInterveningKey mirrors
+// TestCtrlCExitConfirmationDisarmsOnInterveningKey: an intervening key other
+// than Esc (typing, Ctrl+O, etc.) means the user moved on to something else,
+// so a later, unrelated Esc must arm a fresh confirmation instead of
+// silently cancelling off a stale press from seconds ago.
+func TestEscCancelConfirmationDisarmsOnInterveningKey(t *testing.T) {
+	m := newModel(context.Background(), Options{})
+	cancelled := false
+	m.pending = true
+	m.activeRunID = 1
+	m.runCancel = func() { cancelled = true }
+
+	updated, _ := m.Update(testKey(tea.KeyEsc))
+	next := updated.(model)
+	if !next.cancelConfirmActive {
+		t.Fatal("first Esc should arm cancel confirmation")
+	}
+	seq := next.cancelConfirmSeq
+
+	updated, _ = next.Update(testKey('h'))
+	next = updated.(model)
+	if next.cancelConfirmActive {
+		t.Fatal("an intervening non-Esc key should disarm cancel confirmation")
+	}
+	if next.cancelConfirmSeq == seq {
+		t.Fatal("disarming should advance the sequence so a stale expiry tick is ignored")
+	}
+
+	updated, _ = next.Update(testKey(tea.KeyEsc))
+	next = updated.(model)
+	if cancelled {
+		t.Fatal("Esc after an intervening key should re-arm, not cancel")
+	}
+	if !next.cancelConfirmActive {
+		t.Fatal("Esc after an intervening key should arm a fresh confirmation")
+	}
+}
+
+// TestEscCancelConfirmationDisarmsWhenStolenByAskUser: a mid-turn ask_user
+// prompt lands between the two Esc presses. The user's second Esc denies the
+// questionnaire (an earlier branch in the Esc handler), not a confirm, so it
+// must not leave cancelConfirmActive armed for a later, unrelated Esc to
+// silently cancel the run.
+func TestEscCancelConfirmationDisarmsWhenStolenByAskUser(t *testing.T) {
+	m := newModel(context.Background(), Options{})
+	cancelled := false
+	m.pending = true
+	m.activeRunID = 1
+	m.runCancel = func() { cancelled = true }
+
+	updated, _ := m.Update(testKey(tea.KeyEsc))
+	next := updated.(model)
+	if !next.cancelConfirmActive {
+		t.Fatal("first Esc should arm cancel confirmation")
+	}
+
+	request := agent.AskUserRequest{Questions: []agent.AskUserQuestion{{Question: "name?"}}}
+	next.pendingAskUser = &pendingAskUserPrompt{
+		request: request,
+		answer:  func([]string) {},
+		states:  newAskUserStates(request.Questions),
+	}
+
+	updated, _ = next.Update(testKey(tea.KeyEsc))
+	next = updated.(model)
+	if cancelled {
+		t.Fatal("an Esc consumed by the ask-user prompt must not cancel the run")
+	}
+	if next.cancelConfirmActive {
+		t.Fatal("an Esc consumed by the ask-user prompt must disarm the stale cancel confirmation")
+	}
+
+	updated, _ = next.Update(testKey(tea.KeyEsc))
+	next = updated.(model)
+	if cancelled {
+		t.Fatal("the next Esc should arm a fresh confirmation, not immediately cancel")
+	}
+	if !next.cancelConfirmActive {
+		t.Fatal("the next Esc should arm a fresh confirmation")
 	}
 }
 

internal/tui/mouse.go

@@ -14,6 +14,25 @@ type mouseSelectionTarget struct {
 	Index int
 }
 
+// scrollChatExtendingSelection scrolls the chat and, if a drag-selection is in
+// progress, extends its cursor to the mouse's current position against the
+// POST-scroll viewport — mirroring what a mouseMotion event does. Without this, a
+// selection's cursor is only ever updated by actual mouse movement (see
+// handleTranscriptSelectionMouse's mouseMotion case): a wheel-scroll event is a
+// distinct message type that never reaches that code, so scrolling while
+// selecting left the selection frozen at whatever line the mouse last physically
+// moved over — capped at the viewport that was visible before the scroll.
+func (m model) scrollChatExtendingSelection(delta int, msg tea.MouseMsg) model {
+	m = m.scrollChat(delta)
+	if !m.transcriptSelection.active {
+		return m
+	}
+	if line, ok := m.nearestTranscriptLineAtMouse(msg); ok {
+		m.transcriptSelection.cursor = transcriptSelectionPointForMouse(line, mouseX(msg))
+	}
+	return m
+}
+
 func (m model) handleMouse(msg tea.MouseMsg) (tea.Model, tea.Cmd) {
 	// While a provider-wizard OAuth login is in flight, ignore all mouse input
 	// (clicks and wheel) so a stray scroll can't change the selected provider
@@ -21,6 +40,15 @@ func (m model) handleMouse(msg tea.MouseMsg) (tea.Model, tea.Cmd) {
 	if m.providerWizard != nil && m.providerWizard.oauthPending {
 		return m, nil
 	}
+	// A click is a deliberate action, same as a keypress — it means the user
+	// moved on to something else, so it disarms a stale Esc cancel-confirmation
+	// rather than leaving it armed for some later, unrelated Esc to act on.
+	// Scoped to actual clicks (not hover/motion/wheel), since motion events fire
+	// continuously while the mouse merely sits over the terminal and would
+	// otherwise defeat the confirmation window on every render tick.
+	if mouseLeftPress(msg) || mouseRightPress(msg) {
+		m = m.disarmCancelConfirmation()
+	}
 	// A right-click pastes the clipboard straight into the focused field — no
 	// menu. pasteFromClipboardCmd reads the clipboard off the Update goroutine; the
 	// clipboardReadMsg result is routed by routePaste to wherever input is focused
@@ -84,10 +112,17 @@ func (m model) handleMouse(msg tea.MouseMsg) (tea.Model, tea.Cmd) {
 	if next, cmd, ok := m.handleTranscriptSelectionMouse(msg); ok {
 		return next, cmd
 	}
+	// A plain hover (cursor moved, no button pressed) never matches the
+	// press/drag/release cases above, so it falls through here — resolve what's
+	// under the cursor so it can render with the hover highlight.
+	if mouseHover(msg) {
+		return m.updateHoverTarget(msg), nil
+	}
 
 	switch {
 	case mouseWheelUp(msg):
 		m.clearMouseSelection()
+		m = m.clearHover()
 		if m.providerWizard != nil {
 			m.providerWizard.move(-1)
 			return m, nil
@@ -104,7 +139,7 @@ func (m model) handleMouse(msg tea.MouseMsg) (tea.Model, tea.Cmd) {
 			if m.modelPickerIsLoading() {
 				return m, nil
 			}
-			m.picker.move(-1)
+			m.pickerMoved(-1)
 			return m, nil
 		}
 		if m.suggestionsActive() {
@@ -116,9 +151,10 @@ func (m model) handleMouse(msg tea.MouseMsg) (tea.Model, tea.Cmd) {
 				return next, nil
 			}
 		}
-		return m.scrollChat(chatWheelScrollLines), nil
+		return m.scrollChatExtendingSelection(chatWheelScrollLines, msg), nil
 	case mouseWheelDown(msg):
 		m.clearMouseSelection()
+		m = m.clearHover()
 		if m.providerWizard != nil {
 			m.providerWizard.move(1)
 			return m, nil
@@ -135,7 +171,7 @@ func (m model) handleMouse(msg tea.MouseMsg) (tea.Model, tea.Cmd) {
 			if m.modelPickerIsLoading() {
 				return m, nil
 			}
-			m.picker.move(1)
+			m.pickerMoved(1)
 			return m, nil
 		}
 		if m.suggestionsActive() {
@@ -147,7 +183,7 @@ func (m model) handleMouse(msg tea.MouseMsg) (tea.Model, tea.Cmd) {
 				return next, nil
 			}
 		}
-		return m.scrollChat(-chatWheelScrollLines), nil
+		return m.scrollChatExtendingSelection(-chatWheelScrollLines, msg), nil
 	default:
 		return m, nil
 	}
@@ -414,6 +450,9 @@ func (m *model) selectGenericPickerAtMouse(msg tea.MouseMsg) (mouseSelectionTarg
 		if hit.y == line {
 			index := start + offset
 			m.picker.selected = index
+			// Clicking (or hovering onto) a row live-previews it, like arrow keys,
+			// so the theme picker repaints before the confirming second click.
+			m.previewSelectedTheme()
 			return mouseSelectionTarget{Scope: "picker", Kind: int(m.picker.kind), Value: item.Value, Index: index}, true
 		}
 		line++

internal/tui/mouse_test.go

@@ -407,6 +407,245 @@ func TestTranscriptSelectionReleaseExtendsRangeWithoutMotion(t *testing.T) {
 	}
 }
 
+// A completed selection stays "active" through the async copy-command grace
+// window (transcriptCopiedMsg hasn't landed yet), but the drag itself is over.
+// A genuine hover motion (no button held) arriving in that window must NOT be
+// treated as a drag continuation — it must fall through to hover handling
+// instead of silently moving the selection the user just released.
+func TestTranscriptHoverAfterReleaseDoesNotMoveSelection(t *testing.T) {
+	m := withSidebarContent(mouseTestModel())
+	m.mouseCapture = true
+	m.transcript = appendRow(m.transcript, rowUser, "hello world")
+	textY := firstTranscriptTextMouseY(t, m)
+
+	updated, _ := m.Update(testMouseClick(tea.MouseLeft, 3, textY))
+	m = updated.(model)
+	updated, cmd := m.Update(testMouseRelease(tea.MouseNone, 8, textY))
+	m = updated.(model)
+	if cmd == nil {
+		t.Fatal("sanity check failed: release should return the copy command")
+	}
+	if m.transcriptSelection.dragging {
+		t.Fatal("dragging must be false immediately after release")
+	}
+	before := m.selectedTranscriptText()
+
+	// A hover (no button) moves further along the same line, in the window
+	// before transcriptCopiedMsg has landed (m.transcriptSelection.active is
+	// still true here).
+	updated, _ = m.Update(testMouseMotion(tea.MouseNone, 10, textY))
+	m = updated.(model)
+
+	if got := m.selectedTranscriptText(); got != before {
+		t.Fatalf("a post-release hover changed the selection: got %q, want unchanged %q", got, before)
+	}
+}
+
+// Dragging a selection past the top edge of the visible transcript must
+// auto-scroll toward older content and extend the selection to follow — the
+// classic "drag past the viewport edge keeps scrolling" affordance.
+func TestTranscriptSelectionDragPastTopEdgeAutoScrolls(t *testing.T) {
+	m := mouseTestModel()
+	m.mouseCapture = true
+	// This test exercises the animated glide specifically; reducedMotion
+	// defaults from TTY detection (see defaultReducedMotion), which differs
+	// between a local dev shell and CI's non-TTY runners — set it explicitly
+	// so the test is deterministic regardless of environment.
+	m.reducedMotion = false
+	for i := 0; i < 80; i++ {
+		m.transcript = appendRow(m.transcript, rowUser, "line content")
+	}
+	textY := topmostVisibleTranscriptMouseY(t, m)
+
+	updated, _ := m.Update(testMouseClick(tea.MouseLeft, 0, textY))
+	m = updated.(model)
+	if !m.transcriptSelection.active {
+		t.Fatal("selection should be active after a left click on transcript text")
+	}
+	cursorBefore := m.transcriptSelection.cursor.bodyY
+	scrollBefore := m.chatScrollOffset
+
+	frame, _, _ := m.transcriptHitTestLayout()
+	aboveBody := frame.bodyRect.y - 1 // one row above the visible transcript body
+
+	// A left-button drag (not a hover) moved past the top edge: the smooth-glide
+	// chain must start immediately (first step + a scheduled tick).
+	updated, cmd := m.Update(testMouseMotion(tea.MouseLeft, 0, aboveBody))
+	m = updated.(model)
+	if cmd == nil {
+		t.Fatal("crossing the top edge should schedule the glide tick chain")
+	}
+	if m.edgeScrollDelta <= 0 {
+		t.Fatalf("edgeScrollDelta = %d, want > 0 (scrolling toward older content)", m.edgeScrollDelta)
+	}
+	if m.chatScrollOffset == scrollBefore {
+		t.Fatal("dragging past the top edge should auto-scroll toward older content")
+	}
+
+	// A single small step may land on a blank spacer row between messages (no
+	// text there yet) — drive the tick chain forward like a sustained real hold
+	// would, and the cursor must eventually reach a genuinely earlier line.
+	for i := 0; i < 10 && m.transcriptSelection.cursor.bodyY >= cursorBefore; i++ {
+		updated, _ = m.Update(dragEdgeScrollTickMsg{seq: m.edgeScrollSeq})
+		m = updated.(model)
+	}
+
+	if !m.transcriptSelection.active {
+		t.Fatal("selection must survive an edge auto-scroll, not be cleared")
+	}
+	if m.transcriptSelection.cursor.bodyY >= cursorBefore {
+		t.Fatalf("selection cursor bodyY = %d, want < %d (it must extend upward to follow the drag)", m.transcriptSelection.cursor.bodyY, cursorBefore)
+	}
+}
+
+// Symmetric to the top-edge case: dragging past the BOTTOM edge scrolls toward
+// newer content. Requires scrolling up first so there's somewhere to scroll back
+// down to (chatScrollOffset=0 already sits at the bottom).
+func TestTranscriptSelectionDragPastBottomEdgeAutoScrolls(t *testing.T) {
+	m := mouseTestModel()
+	m.mouseCapture = true
+	m.reducedMotion = false // deterministic across local/CI TTY-detection differences
+	for i := 0; i < 80; i++ {
+		m.transcript = appendRow(m.transcript, rowUser, "line content")
+	}
+	m.chatScrollOffset = chatWheelScrollLines * 3 // scrolled up, away from the bottom
+
+	textY := topmostVisibleTranscriptMouseY(t, m)
+	updated, _ := m.Update(testMouseClick(tea.MouseLeft, 0, textY))
+	m = updated.(model)
+	if !m.transcriptSelection.active {
+		t.Fatal("selection should be active after a left click on transcript text")
+	}
+	cursorBefore := m.transcriptSelection.cursor.bodyY
+	scrollBefore := m.chatScrollOffset
+
+	frame, _, _ := m.transcriptHitTestLayout()
+	belowBody := frame.bodyRect.y + frame.bodyRect.height // one row below the visible transcript body
+
+	updated, cmd := m.Update(testMouseMotion(tea.MouseLeft, 0, belowBody))
+	m = updated.(model)
+	if cmd == nil {
+		t.Fatal("crossing the bottom edge should schedule the glide tick chain")
+	}
+	if m.edgeScrollDelta >= 0 {
+		t.Fatalf("edgeScrollDelta = %d, want < 0 (scrolling toward newer content)", m.edgeScrollDelta)
+	}
+	if m.chatScrollOffset >= scrollBefore {
+		t.Fatal("dragging past the bottom edge should auto-scroll toward newer content (offset decreases)")
+	}
+
+	for i := 0; i < 10 && m.transcriptSelection.cursor.bodyY <= cursorBefore; i++ {
+		updated, _ = m.Update(dragEdgeScrollTickMsg{seq: m.edgeScrollSeq})
+		m = updated.(model)
+	}
+
+	if !m.transcriptSelection.active {
+		t.Fatal("selection must survive an edge auto-scroll, not be cleared")
+	}
+	if m.transcriptSelection.cursor.bodyY <= cursorBefore {
+		t.Fatalf("selection cursor bodyY = %d, want > %d (it must extend downward to follow the drag)", m.transcriptSelection.cursor.bodyY, cursorBefore)
+	}
+}
+
+// The glide tick chain must self-terminate the instant the drag moves back into
+// the visible body — otherwise a stray scheduled tick could keep scrolling after
+// the user has already dragged back to the content they wanted selected.
+func TestTranscriptSelectionEdgeGlideStopsWhenDragReturnsToBody(t *testing.T) {
+	m := mouseTestModel()
+	m.mouseCapture = true
+	m.reducedMotion = false // deterministic across local/CI TTY-detection differences
+	for i := 0; i < 80; i++ {
+		m.transcript = appendRow(m.transcript, rowUser, "line content")
+	}
+	textY := topmostVisibleTranscriptMouseY(t, m)
+
+	updated, _ := m.Update(testMouseClick(tea.MouseLeft, 0, textY))
+	m = updated.(model)
+	frame, _, _ := m.transcriptHitTestLayout()
+	updated, _ = m.Update(testMouseMotion(tea.MouseLeft, 0, frame.bodyRect.y-1))
+	m = updated.(model)
+	if m.edgeScrollDelta == 0 {
+		t.Fatal("sanity check failed: the glide chain should be running")
+	}
+	seqWhileGliding := m.edgeScrollSeq
+
+	updated, _ = m.Update(testMouseMotion(tea.MouseLeft, 0, textY))
+	m = updated.(model)
+	if m.edgeScrollDelta != 0 {
+		t.Fatalf("edgeScrollDelta = %d, want 0 (dragging back into the body must stop the glide)", m.edgeScrollDelta)
+	}
+
+	// A tick already in flight from before the drag returned must be a no-op —
+	// it carries the OLD seq, which no longer matches.
+	scrollBefore := m.chatScrollOffset
+	updated, _ = m.Update(dragEdgeScrollTickMsg{seq: seqWhileGliding})
+	m = updated.(model)
+	if m.chatScrollOffset != scrollBefore {
+		t.Fatal("a stale in-flight tick must not scroll after the glide was stopped")
+	}
+}
+
+// A keypress mid-glide clears m.transcriptSelection.active directly (the
+// KeyPressMsg handler in model.go), NOT through stopEdgeScroll — so
+// edgeScrollDelta is left stale. If a fresh drag past the SAME edge afterward
+// doesn't reset it, startEdgeScroll's "already running" fast path is fooled into
+// thinking a chain is already active and silently does nothing: no immediate
+// step, no scheduled tick. The press handler must reset glide state so every new
+// drag starts clean regardless of how the previous one ended.
+func TestTranscriptSelectionEdgeGlideRecoversAfterKeypressInterrupt(t *testing.T) {
+	m := mouseTestModel()
+	m.mouseCapture = true
+	m.reducedMotion = false // deterministic across local/CI TTY-detection differences
+	for i := 0; i < 80; i++ {
+		m.transcript = appendRow(m.transcript, rowUser, "line content")
+	}
+	textY := topmostVisibleTranscriptMouseY(t, m)
+
+	updated, _ := m.Update(testMouseClick(tea.MouseLeft, 0, textY))
+	m = updated.(model)
+	frame, _, _ := m.transcriptHitTestLayout()
+	aboveBody := frame.bodyRect.y - 1
+
+	updated, _ = m.Update(testMouseMotion(tea.MouseLeft, 0, aboveBody))
+	m = updated.(model)
+	if m.edgeScrollDelta == 0 {
+		t.Fatal("sanity check failed: the glide chain should be running")
+	}
+
+	// A keypress interrupts the drag WITHOUT releasing the mouse — clears
+	// transcriptSelection.active directly, bypassing stopEdgeScroll.
+	updated, _ = m.Update(testKey('a'))
+	m = updated.(model)
+	if m.transcriptSelection.active {
+		t.Fatal("sanity check failed: the keypress should have cleared the selection")
+	}
+	if m.edgeScrollDelta == 0 {
+		t.Fatal("sanity check failed: edgeScrollDelta should still be stale here (that's the bug this test guards)")
+	}
+
+	// A brand-new press+drag past the SAME edge must glide normally, not be
+	// silently swallowed by the stale state from the interrupted drag. The
+	// viewport shifted from the first glide's step, so the topmost visible text
+	// line's on-screen Y must be recomputed — reusing the ORIGINAL textY would
+	// fail for an unrelated reason (it may now land on a blank spacer row).
+	textY = topmostVisibleTranscriptMouseY(t, m)
+	updated, _ = m.Update(testMouseClick(tea.MouseLeft, 0, textY))
+	m = updated.(model)
+	if !m.transcriptSelection.active {
+		t.Fatal("sanity check failed: the new press should start a fresh selection")
+	}
+	scrollBefore := m.chatScrollOffset
+
+	updated, cmd := m.Update(testMouseMotion(tea.MouseLeft, 0, aboveBody))
+	m = updated.(model)
+	if cmd == nil {
+		t.Fatal("the new drag should schedule a fresh glide tick chain, not be fooled by stale state")
+	}
+	if m.chatScrollOffset == scrollBefore {
+		t.Fatal("the new drag should have applied its immediate first scroll step")
+	}
+}
+
 func TestTranscriptSelectionClearsAfterCopy(t *testing.T) {
 	m := mouseTestModel()
 	m.transcriptSelection = transcriptSelectionState{active: true}

internal/tui/oauth_device.go

@@ -44,6 +44,9 @@ func oauthDevicePrepare(name string) (oauth.DeviceAuth, oauth.Config, error) {
 		Store:       store,
 		HTTPClient:  &http.Client{Timeout: 60 * time.Second},
 		OpenBrowser: browser.OpenURL,
+		// Device-flow providers (e.g. Hugging Face) rely on the baked-in preset for
+		// their endpoints; opt in so the device code can be requested.
+		AllowPresets: true,
 	})
 	if err != nil {
 		return oauth.DeviceAuth{}, oauth.Config{}, err
@@ -62,8 +65,9 @@ func oauthDeviceComplete(name string, cfg oauth.Config, auth oauth.DeviceAuth) e
 		return err
 	}
 	manager, err := oauth.NewManager(oauth.ManagerOptions{
-		Store:      store,
-		HTTPClient: &http.Client{Timeout: 60 * time.Second},
+		Store:        store,
+		HTTPClient:   &http.Client{Timeout: 60 * time.Second},
+		AllowPresets: true, // preset config is needed to poll/exchange the device token
 	})
 	if err != nil {
 		return err
@@ -84,8 +88,9 @@ func oauthStoredToken(ctx context.Context, providerID string) string {
 		return ""
 	}
 	manager, err := oauth.NewManager(oauth.ManagerOptions{
-		Store:      store,
-		HTTPClient: &http.Client{Timeout: 30 * time.Second},
+		Store:        store,
+		HTTPClient:   &http.Client{Timeout: 30 * time.Second},
+		AllowPresets: true, // refreshing a preset-provider token re-resolves its config
 	})
 	if err != nil {
 		return ""

internal/tui/onboarding.go

@@ -755,6 +755,15 @@ func (m model) completeSetup() (tea.Model, tea.Cmd) {
 		m.providerProfile = result.Provider
 		m.providerName = result.Provider.Name
 		m.modelName = result.Provider.Model
+		// Export ZERO_PROVIDER alongside the committed profile fields (and the
+		// config setupSave already persisted as active). Unlike command_center's
+		// switch — which commits everything only after a successful build — setup
+		// commits config + profile unconditionally here, so the env must match
+		// them unconditionally too: applyEnv makes ZERO_PROVIDER WIN over config,
+		// so a stale value from an earlier /model switch left in place would send
+		// spawned children to the OLD provider even though config now names the
+		// new one. That is the exact D3 gap this closes.
+		config.SetActiveProviderEnv(result.Provider.Name)
 		if m.newProvider != nil {
 			if provider, providerErr := m.newProvider(result.Provider); providerErr == nil {
 				m.provider = provider

internal/tui/onboarding_test.go

@@ -3,6 +3,7 @@ package tui
 import (
 	"context"
 	"errors"
+	"os"
 	"strings"
 	"testing"
 
@@ -1535,3 +1536,42 @@ func absInt(value int) int {
 	}
 	return value
 }
+
+// Completing setup switches the live provider, so it must export ZERO_PROVIDER
+// exactly like the /model, /provider, and wizard switch paths — a stale value
+// from an earlier switch would otherwise win over config in every spawned
+// child (applyEnv) and pin specialists/swarm members to the OLD provider's
+// credentials.
+func TestCompleteSetupExportsActiveProviderEnv(t *testing.T) {
+	t.Setenv(config.ActiveProviderEnv, "stale-previous-provider")
+	m := newModel(context.Background(), Options{
+		Setup: SetupOptions{
+			Visible: true,
+			Providers: []SetupProviderOption{
+				{ID: "openai", Name: "OpenAI", DefaultModel: "gpt-4.1", EnvVar: "OPENAI_API_KEY", RequiresAuth: true},
+			},
+			Save: func(selection SetupSelection) (SetupResult, error) {
+				return SetupResult{
+					Provider: config.ProviderProfile{
+						Name:      selection.CatalogID,
+						CatalogID: selection.CatalogID,
+						Model:     selection.Model,
+					},
+				}, nil
+			},
+		},
+	})
+	m.width = 100
+	m.height = 30
+	m.setup.stage = setupStageReady
+
+	updated, _ := m.completeSetup()
+	next := updated.(model)
+
+	if next.providerName == "" {
+		t.Fatal("setup completion should have set a provider name")
+	}
+	if got := os.Getenv(config.ActiveProviderEnv); got != next.providerName {
+		t.Fatalf("%s = %q after setup save, want %q (children would spawn on the stale provider)", config.ActiveProviderEnv, got, next.providerName)
+	}
+}

internal/tui/options.go

@@ -20,41 +20,46 @@ import (
 
 // Options configures the reusable Zero terminal UI shell.
 type Options struct {
-	Cwd                    string
-	UserConfigPath         string
-	DoctorUserConfigPath   string
-	ProjectConfigPath      string
-	ProviderName           string
-	ModelName              string
-	ProviderProfile        config.ProviderProfile
-	SavedProviders         []config.ProviderProfile // all configured providers, for the /model multi-provider list
-	FavoriteModels         []string
-	RecapsEnabled          bool
-	Provider               zeroruntime.Provider
-	NewProvider            func(config.ProviderProfile) (zeroruntime.Provider, error)
-	ProbeProviderHealth    func(context.Context, providerhealth.Options) providerhealth.Result
-	DiscoverProviderModels func(context.Context, config.ProviderProfile) ([]providermodeldiscovery.Model, error)
-	RuntimeMessageSink     func(tea.Msg)
-	Registry               *tools.Registry
-	SessionStore           *sessions.Store
-	SandboxStore           *sandbox.GrantStore
-	MCPConfig              config.MCPConfig
-	MCPPermissionStore     *mcp.PermissionStore
-	MCPTokenStore          *mcp.TokenStore
-	MCPCommand             func(context.Context, []string) MCPCommandResult
-	SandboxSetupCommand    func(context.Context) SandboxSetupCommandResult
-	UsageTracker           *usage.Tracker
-	SessionCompactor       SessionCompactor
-	PrService              *PrService
+	Cwd                         string
+	UserConfigPath              string
+	DoctorUserConfigPath        string
+	ProjectConfigPath           string
+	ProviderName                string
+	ModelName                   string
+	ProviderProfile             config.ProviderProfile
+	SavedProviders              []config.ProviderProfile // all configured providers, for the /model multi-provider list
+	FavoriteModels              []string
+	RecapsEnabled               bool
+	Provider                    zeroruntime.Provider
+	NewProvider                 func(config.ProviderProfile) (zeroruntime.Provider, error)
+	ProbeProviderHealth         func(context.Context, providerhealth.Options) providerhealth.Result
+	DiscoverProviderModels      func(context.Context, config.ProviderProfile) ([]providermodeldiscovery.Model, error)
+	DiscoverOllamaContextWindow func(ctx context.Context, baseURL string, model string) (int, error)
+	RuntimeMessageSink          func(tea.Msg)
+	Registry                    *tools.Registry
+	SessionStore                *sessions.Store
+	SandboxStore                *sandbox.GrantStore
+	MCPConfig                   config.MCPConfig
+	MCPPermissionStore          *mcp.PermissionStore
+	MCPTokenStore               *mcp.TokenStore
+	MCPCommand                  func(context.Context, []string) MCPCommandResult
+	SandboxSetupCommand         func(context.Context) SandboxSetupCommandResult
+	UsageTracker                *usage.Tracker
+	SessionCompactor            SessionCompactor
+	PrService                   *PrService
 
 	AgentOptions    agent.Options
 	PermissionMode  agent.PermissionMode
 	ReasoningEffort modelregistry.ReasoningEffort
 	ResponseStyle   string
-	// Theme is the operator's palette preference: "auto" (default), "dark", or
-	// "light". Set from the --theme flag; falls back to ZERO_THEME then auto.
-	Theme     string
-	UserAgent string
+	// Theme is the operator's palette preference: "auto" (default), a built-in
+	// ("dark"/"light"), or a registered color theme. Set from the --theme flag;
+	// falls back to ZERO_THEME, then the persisted SavedTheme, then auto.
+	Theme string
+	// SavedTheme is the theme persisted in user config (Preferences.Theme). Applied
+	// at startup below --theme and ZERO_THEME, so a /theme choice survives restart.
+	SavedTheme string
+	UserAgent  string
 
 	// Notify configures completion / awaiting-input notifications.
 	Notify config.NotifyConfig

internal/tui/picker.go

@@ -23,6 +23,7 @@ const (
 	pickerModel pickerKind = iota
 	pickerEffort
 	pickerSession
+	pickerTheme
 )
 
 // pickerItem is one selectable row: Label is shown, Value is passed to the
@@ -564,6 +565,46 @@ type modelPickerModelsDiscoveredMsg struct {
 	err        error
 }
 
+// ollamaContextWindowDiscoveredMsg carries the result of an async /api/show
+// probe against a local Ollama daemon (see ollamaContextWindowDiscoveryCmd).
+// A zero contextWindow (or a non-nil err) means the probe found nothing
+// usable and the map is left untouched, not zeroed out.
+type ollamaContextWindowDiscoveredMsg struct {
+	modelName     string
+	contextWindow int
+	err           error
+}
+
+// ollamaContextWindowDiscoveryCmd probes a local Ollama daemon's native
+// /api/show endpoint for modelName's context length. Scoped to the local
+// Ollama provider only (catalog ID "ollama", not "ollama-cloud" — a
+// different hosted service this endpoint isn't assumed to exist on): the
+// generic /v1/models discovery has no source for this metadata for
+// custom/local model tags, so without this the context gauge just never
+// shows for them (see modelContextWindow).
+func (m model) ollamaContextWindowDiscoveryCmd(descriptor providercatalog.Descriptor, baseURL string, modelName string) tea.Cmd {
+	if descriptor.ID != "ollama" {
+		return nil
+	}
+	modelName = strings.TrimSpace(modelName)
+	baseURL = strings.TrimSpace(baseURL)
+	if modelName == "" || baseURL == "" {
+		return nil
+	}
+	discover := m.discoverOllamaContextWindow
+	if discover == nil {
+		discover = func(ctx context.Context, baseURL string, model string) (int, error) {
+			return providermodeldiscovery.DiscoverOllamaContextWindow(ctx, baseURL, model, providermodeldiscovery.Options{})
+		}
+	}
+	return func() tea.Msg {
+		ctx, cancel := context.WithTimeout(m.ctx, 8*time.Second)
+		defer cancel()
+		window, err := discover(ctx, baseURL, modelName)
+		return ollamaContextWindowDiscoveredMsg{modelName: modelName, contextWindow: window, err: err}
+	}
+}
+
 func (m model) normalizeProfileForProvider(provider providercatalog.Descriptor) config.ProviderProfile {
 	profile := m.providerProfile
 	normalizeIdentity := profileMatchesProviderBaseURL(profile, provider) ||
@@ -726,3 +767,44 @@ func (m model) newEffortPicker() *commandPicker {
 	}
 	return &commandPicker{kind: pickerEffort, title: "select reasoning effort", items: items, selected: selected}
 }
+
+// newThemePicker lists `auto` plus every registered theme as a popup, grouped into
+// Dark/Light sections (the registry is ordered dark-then-light so the group header
+// changes exactly once), preselecting the active preference. Bare `/theme` opens the
+// same overlay /model and /effort use. Moving the cursor live-previews each palette
+// (previewSelectedTheme); Enter commits the highlighted theme (choosePicker) and Esc
+// restores the previous one (Update's picker-cancel path). Items stay 1:1 with
+// themeModes and in the same order, so the popup and /theme state list agree.
+func (m model) newThemePicker() *commandPicker {
+	items := make([]pickerItem, 0, len(themeModes))
+	selected := 0
+	// `auto` sits at the top with an empty Group, so it renders header-less above
+	// the Dark/Light sections.
+	items = append(items, pickerItem{Label: string(themeAuto), Value: string(themeAuto), Meta: "match terminal"})
+	for _, entry := range themeRegistry {
+		group := "Light"
+		if entry.IsDark {
+			group = "Dark"
+		}
+		items = append(items, pickerItem{Group: group, Label: entry.Label, Value: entry.Name})
+		if entry.Name == string(m.themeMode) {
+			selected = len(items) - 1
+		}
+	}
+	// allItems lets the query filter restore rows on Backspace (one-way narrowing
+	// otherwise, since applyQuery falls back to the current items without it).
+	return &commandPicker{kind: pickerTheme, title: "select theme", items: items, allItems: append([]pickerItem{}, items...), selected: selected}
+}
+
+// pickerMoved advances the open picker's cursor by delta and live-previews the new
+// selection where the picker supports it — stepping through the /theme popup
+// repaints the UI in the hovered palette. Safe to call with no picker open. Callers
+// mutate through m.picker (a pointer) and the global theme, so the value receiver
+// is fine.
+func (m model) pickerMoved(delta int) {
+	if m.picker == nil {
+		return
+	}
+	m.picker.move(delta)
+	m.previewSelectedTheme()
+}

internal/tui/picker_test.go

@@ -14,6 +14,7 @@ import (
 
 	"github.com/Gitlawb/zero/internal/config"
 	"github.com/Gitlawb/zero/internal/modelregistry"
+	"github.com/Gitlawb/zero/internal/providercatalog"
 	"github.com/Gitlawb/zero/internal/providermodeldiscovery"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
@@ -639,14 +640,17 @@ func TestEffortPickerOpensForSupportedModel(t *testing.T) {
 	}
 }
 
-func TestThemeCommandOpensNoPicker(t *testing.T) {
-	// /theme keeps the existing shell-only message; no picker opens.
+func TestThemeCommandOpensPicker(t *testing.T) {
+	// Bare /theme opens the theme popup (live preview on move, apply on Enter),
+	// like /model and /effort. Full preview/commit/cancel behavior is covered in
+	// theme_picker_test.go; here we just pin that the no-arg command opens it.
+	defer applyTheme(themeDark, true)
 	m := newModel(context.Background(), Options{})
 	m.input.SetValue("/theme")
 	updated, _ := m.Update(testKey(tea.KeyEnter))
 	m = updated.(model)
-	if m.picker != nil {
-		t.Fatal("/theme should not open a picker")
+	if m.picker == nil || m.picker.kind != pickerTheme {
+		t.Fatalf("expected the theme picker to open, got %#v", m.picker)
 	}
 }
 
@@ -875,4 +879,100 @@ func TestModelContextWindowResolution(t *testing.T) {
 	if got := m.modelContextWindow(""); got != 0 {
 		t.Fatalf("empty model name should yield 0, got %d", got)
 	}
+	// A custom/local Ollama model tag has no curated-catalog entry and its
+	// generic /v1/models listing carries no window either — the only source is
+	// the Ollama-specific /api/show probe, landed via ollamaContextWindowByModel.
+	m.ollamaContextWindowByModel = map[string]int{"kimi-k2.7-code:cloud": 131072}
+	if got := m.modelContextWindow("kimi-k2.7-code:cloud"); got != 131072 {
+		t.Fatalf("ollama-discovered window should be used for an unregistered model, got %d", got)
+	}
+}
+
+// TestOllamaContextWindowDiscoveryCmdScopedToLocalOllama: the /api/show probe
+// only makes sense against a local Ollama daemon (catalog ID "ollama") — a
+// different provider like "ollama-cloud" is a distinct hosted service this
+// endpoint isn't assumed to exist on, and every other provider already has
+// its context window covered by the curated catalog or /v1/models discovery.
+func TestOllamaContextWindowDiscoveryCmdScopedToLocalOllama(t *testing.T) {
+	m := model{ctx: context.Background()}
+
+	ollama, ok := providercatalog.Get("ollama")
+	if !ok {
+		t.Fatal("expected \"ollama\" to be a registered catalog descriptor")
+	}
+	if cmd := m.ollamaContextWindowDiscoveryCmd(ollama, "http://localhost:11434/v1", "kimi-k2.7-code:cloud"); cmd == nil {
+		t.Fatal("expected a discovery command for the local ollama provider")
+	}
+
+	ollamaCloud, ok := providercatalog.Get("ollama-cloud")
+	if !ok {
+		t.Fatal("expected \"ollama-cloud\" to be a registered catalog descriptor")
+	}
+	if cmd := m.ollamaContextWindowDiscoveryCmd(ollamaCloud, "https://ollama.com/v1", "qwen3-coder:480b"); cmd != nil {
+		t.Fatal("ollama-cloud is a different hosted service; must not probe /api/show against it")
+	}
+
+	if cmd := m.ollamaContextWindowDiscoveryCmd(ollama, "", "kimi-k2.7-code:cloud"); cmd != nil {
+		t.Fatal("an empty base URL should yield no command")
+	}
+	if cmd := m.ollamaContextWindowDiscoveryCmd(ollama, "http://localhost:11434/v1", ""); cmd != nil {
+		t.Fatal("an empty model name should yield no command")
+	}
+}
+
+// TestOllamaContextWindowDiscoveredMsgPopulatesMap verifies the Update()
+// handler for the async probe's result actually reaches modelContextWindow's
+// fallback source, and that a failed/zero probe leaves the map untouched
+// rather than caching a bogus zero.
+func TestOllamaContextWindowDiscoveredMsgPopulatesMap(t *testing.T) {
+	m := newModel(context.Background(), Options{})
+
+	updated, cmd := m.Update(ollamaContextWindowDiscoveredMsg{modelName: "kimi-k2.7-code:cloud", contextWindow: 131072})
+	next := updated.(model)
+	if cmd != nil {
+		t.Fatal("applying a discovered window should not schedule further work")
+	}
+	if got := next.modelContextWindow("kimi-k2.7-code:cloud"); got != 131072 {
+		t.Fatalf("modelContextWindow after discovery = %d, want 131072", got)
+	}
+
+	updated, _ = next.Update(ollamaContextWindowDiscoveredMsg{modelName: "other-model", contextWindow: 0, err: errors.New("boom")})
+	next = updated.(model)
+	if got := next.modelContextWindow("other-model"); got != 0 {
+		t.Fatalf("a failed probe must not populate a window, got %d", got)
+	}
+}
+
+// TestSwitchProviderModelWarmsDiscoveryForTheNewProvider: switching providers
+// mid-session (e.g. via the /model picker) previously fired no discovery at
+// all for the newly active provider — Init() only warms the provider active
+// at launch — so the context gauge stayed blank for the rest of the session
+// unless /model happened to be reopened separately. Switching to Ollama
+// specifically must also warm the /api/show probe, the only source of a
+// context window for custom/local Ollama models.
+func TestSwitchProviderModelWarmsDiscoveryForTheNewProvider(t *testing.T) {
+	m := newModel(context.Background(), Options{
+		ProviderName:    "openai",
+		ModelName:       "gpt-5.1",
+		Provider:        &fakeProvider{},
+		ProviderProfile: config.ProviderProfile{Name: "openai", CatalogID: "openai", Model: "gpt-5.1"},
+		SavedProviders: []config.ProviderProfile{
+			{Name: "openai", CatalogID: "openai", Model: "gpt-5.1"},
+			{Name: "ollama", CatalogID: "ollama", ProviderKind: config.ProviderKindOpenAICompatible, BaseURL: "http://localhost:11434/v1", Model: "kimi-k2.7-code:cloud"},
+		},
+		NewProvider: func(config.ProviderProfile) (zeroruntime.Provider, error) {
+			return &fakeProvider{}, nil
+		},
+	})
+
+	next, text, cmd := m.switchProviderModel("ollama", "kimi-k2.7-code:cloud")
+	if !strings.Contains(text, "Switched to ollama") {
+		t.Fatalf("switch notice = %q, want it to confirm the switch", text)
+	}
+	if next.modelName != "kimi-k2.7-code:cloud" || next.providerName != "ollama" {
+		t.Fatalf("model/provider not switched: modelName=%q providerName=%q", next.modelName, next.providerName)
+	}
+	if cmd == nil {
+		t.Fatal("switching to ollama should warm both the generic and ollama-specific discovery commands")
+	}
 }

internal/tui/provider_wizard.go

@@ -20,6 +20,7 @@ import (
 	"github.com/Gitlawb/zero/internal/providercatalog"
 	"github.com/Gitlawb/zero/internal/provideroauth"
 	"github.com/Gitlawb/zero/internal/redaction"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
 // providerWizardOAuthMsg carries the result of an in-wizard browser OAuth login.
@@ -166,6 +167,10 @@ func runProviderTokenLogin(name string) error {
 		Store:       store,
 		HTTPClient:  &http.Client{Timeout: 60 * time.Second},
 		OpenBrowser: browser.OpenURL,
+		// The user explicitly chose to sign in with this provider's OAuth, so opt
+		// into its baked-in preset (e.g. xAI's public client_id); without this the
+		// config never resolves and the browser never opens.
+		AllowPresets: true,
 	})
 	if err != nil {
 		return err
@@ -883,13 +888,20 @@ func (m model) applyProviderWizard() (model, tea.Cmd) {
 	modelChoice := wizard.currentModel()
 	profile := providerWizardProfile(provider, modelChoice.ID, wizard.apiKey, wizard.baseURL, wizard.profileName)
 	runtimeProfile := providerWizardRuntimeProfile(profile)
+
+	// Build and persist into LOCALS first, committing live state only once BOTH
+	// succeed. A persist failure (read-only config, disk full) must not leave the
+	// chat running on the new provider while the status line, m.providerProfile,
+	// and the ZERO_PROVIDER export (which pins spawned children) still point at
+	// the old one — parent and children would silently run on different providers.
+	var nextProvider zeroruntime.Provider
 	if m.newProvider != nil {
-		nextProvider, err := m.newProvider(runtimeProfile)
+		built, err := m.newProvider(runtimeProfile)
 		if err != nil {
 			wizard.err = redaction.RedactString(err.Error(), redaction.Options{ExtraSecretValues: []string{profile.APIKey, runtimeProfile.APIKey}})
 			return m, nil
 		}
-		m.provider = nextProvider
+		nextProvider = built
 	}
 	if strings.TrimSpace(m.userConfigPath) != "" {
 		// Capture flip: move the freshly entered key into the encrypted credential
@@ -899,12 +911,24 @@ func (m model) applyProviderWizard() (model, tea.Cmd) {
 		profile = config.SecureProviderProfile(profile, m.userConfigPath)
 		if _, err := config.UpsertProvider(m.userConfigPath, profile, true); err != nil {
 			wizard.err = redaction.RedactString(err.Error(), redaction.Options{ExtraSecretValues: []string{secret, profile.APIKey}})
-			return m, nil
+			return m, nil // nothing committed to live state yet
 		}
 	}
+
+	// Both succeeded — commit the live provider, profile, model, and the child
+	// env export together, so they can never disagree.
+	if nextProvider != nil {
+		m.provider = nextProvider
+	}
 	m.providerProfile = profile
 	m.providerName = profile.Name
 	m.modelName = profile.Model
+	// Keep sub-agent child processes on the same provider we just switched to —
+	// same as the /model and /provider switch paths (command_center.go). Without
+	// this, a ZERO_PROVIDER exported by an earlier switch stays pointing at the
+	// OLD provider and wins over config in every spawned child (applyEnv), so
+	// specialists/swarm members run on the wrong provider's credentials.
+	config.SetActiveProviderEnv(profile.Name)
 	m.providerWizard = nil
 	return m, nil
 }
@@ -1229,9 +1253,29 @@ func (wizard *providerWizardState) renderProviderStep(width int) []string {
 	for offset, provider := range wizard.providers[start : start+maxVisible] {
 		lines = append(lines, wizard.renderSelectableProvider(width, start+offset, provider))
 	}
+	// A failed OAuth attempt leaves the wizard on this list (it does not advance),
+	// so the error must render here or the click looks like a silent no-op. The
+	// credential step renders its own copy for the ctrl+o path.
+	if wizard.oauthMode && wizard.oauthErr != "" {
+		lines = append(lines, "", zeroTheme.red.Render("OAuth login failed: "+wizard.oauthErr))
+		if hint := providerWizardOAuthErrHint(wizard.currentProvider()); hint != "" {
+			lines = append(lines, zeroTheme.faint.Render(hint))
+		}
+	}
 	return lines
 }
 
+// providerWizardOAuthErrHint returns a provider-specific next step for a failed
+// OAuth login. Hugging Face ships no public client_id (unlike xAI), so a login
+// can only work once the operator registers an app and supplies its client_id.
+func providerWizardOAuthErrHint(provider providercatalog.Descriptor) string {
+	if strings.EqualFold(provider.ID, "huggingface") {
+		return "Hugging Face needs your own app: create one at " +
+			"https://huggingface.co/settings/applications/new, then set ZERO_OAUTH_HUGGINGFACE_CLIENT_ID."
+	}
+	return ""
+}
+
 func (wizard *providerWizardState) renderSelectableProvider(width int, index int, provider providercatalog.Descriptor) string {
 	selected := index == wizard.selectedProvider
 	surface := transparentSurface

internal/tui/provider_wizard_oauth_test.go

@@ -125,6 +125,36 @@ func TestProviderWizardDeviceCodeMsgShowsCodeAndPolls(t *testing.T) {
 	}
 }
 
+// A failed OAuth attempt leaves the wizard on the provider list; the error must be
+// rendered there (not just on the credential step) so a click isn't a silent
+// no-op, and Hugging Face gets an actionable client_id hint.
+func TestProviderStepSurfacesOAuthError(t *testing.T) {
+	m := mouseTestModel()
+	m.providerWizard = m.newProviderWizard()
+	wizard := m.providerWizard
+	wizard.oauthMode = true
+	wizard.providers = providerWizardOAuthDescriptors()
+	found := false
+	for i, p := range wizard.providers {
+		if strings.EqualFold(p.ID, "huggingface") {
+			wizard.selectedProvider = i
+			found = true
+		}
+	}
+	if !found {
+		t.Fatal("huggingface should be an OAuth-capable provider")
+	}
+	wizard.oauthErr = `oauth: provider "huggingface" is not configured; set ZERO_OAUTH_HUGGINGFACE_CLIENT_ID`
+
+	view := strings.Join(wizard.renderProviderStep(72), "\n")
+	if !strings.Contains(view, "OAuth login failed") {
+		t.Fatalf("provider step should surface the OAuth error:\n%s", view)
+	}
+	if !strings.Contains(view, "huggingface.co/settings/applications/new") {
+		t.Fatalf("Hugging Face hint should point at app registration:\n%s", view)
+	}
+}
+
 func TestProviderWizardDeviceErrorSurfaced(t *testing.T) {
 	m := mouseTestModel()
 	m.providerWizard = m.newProviderWizard()

internal/tui/provider_wizard_test.go

@@ -1189,3 +1189,104 @@ func containsString(values []string, want string) bool {
 	}
 	return false
 }
+
+// Applying the wizard switches the live provider, so it must export
+// ZERO_PROVIDER exactly like the /model and /provider switch paths — a stale
+// value from an earlier switch would otherwise win over config in every
+// spawned child (applyEnv) and pin specialists/swarm members to the OLD
+// provider's credentials.
+func TestApplyProviderWizardExportsActiveProviderEnv(t *testing.T) {
+	t.Setenv(config.ActiveProviderEnv, "stale-previous-provider")
+	m := newModel(context.Background(), Options{})
+	// Isolate the SUCCESS path's persist to a temp config (the default empty
+	// path would skip persist; a future default must never reach the real user
+	// config), and stub the build so the full commit sequence runs.
+	m.userConfigPath = filepath.Join(t.TempDir(), "config.json")
+	m.newProvider = func(config.ProviderProfile) (zeroruntime.Provider, error) {
+		return &fakeProvider{}, nil
+	}
+	m.providerWizard = &providerWizardState{
+		step:        providerWizardStepModel,
+		profileName: "acme-wizard",
+		providers: []providercatalog.Descriptor{{
+			ID:                  "openrouter",
+			Name:                "OpenRouter",
+			Transport:           providercatalog.TransportOpenAICompatible,
+			DefaultBaseURL:      "https://openrouter.ai/api/v1",
+			DefaultModel:        "openai/gpt-4.1",
+			AuthEnvVars:         []string{"OPENROUTER_API_KEY"},
+			RequiresAuth:        true,
+			SupportedAPIFormats: []providercatalog.APIFormat{providercatalog.APIFormatOpenAIChatCompletions},
+		}},
+		models: []providerWizardModel{{ID: "openai/gpt-4.1", Description: "GPT-4.1"}},
+	}
+
+	updated, _ := m.applyProviderWizard()
+	next := updated
+
+	if next.providerName == "" {
+		t.Fatal("wizard apply should have set a provider name")
+	}
+	if got := os.Getenv(config.ActiveProviderEnv); got != next.providerName {
+		t.Fatalf("%s = %q after wizard apply, want %q (children would spawn on the stale provider)", config.ActiveProviderEnv, got, next.providerName)
+	}
+}
+
+// On a config PERSIST failure, applyProviderWizard must leave live state fully
+// unchanged — the chat must NOT already be running on the new provider while the
+// status line and the ZERO_PROVIDER export (which pins spawned children) still
+// point at the old one. Build and persist are staged into locals; nothing is
+// committed unless both succeed.
+func TestApplyProviderWizardPersistFailureLeavesLiveStateUnchanged(t *testing.T) {
+	t.Setenv("ZERO_CRED_STORAGE", "encrypted-file") // never touch the real OS keychain: apiKey is secured before the persist fails
+	t.Setenv(config.ActiveProviderEnv, "old-provider")
+
+	// A config path whose parent is a regular FILE, so writeConfigFile's MkdirAll
+	// fails and UpsertProvider returns an error.
+	blocker := filepath.Join(t.TempDir(), "not-a-dir")
+	if err := os.WriteFile(blocker, []byte("x"), 0o600); err != nil {
+		t.Fatalf("seed blocker file: %v", err)
+	}
+	brokenConfigPath := filepath.Join(blocker, "config.json")
+
+	oldProvider := &fakeProvider{}
+	newProvider := &fakeProvider{}
+	m := newModel(context.Background(), Options{})
+	m.provider = oldProvider
+	m.providerProfile = config.ProviderProfile{Name: "old-provider"}
+	m.providerName = "old-provider"
+	m.userConfigPath = brokenConfigPath
+	m.newProvider = func(config.ProviderProfile) (zeroruntime.Provider, error) { return newProvider, nil }
+	m.providerWizard = &providerWizardState{
+		step:        providerWizardStepModel,
+		profileName: "acme-new",
+		providers: []providercatalog.Descriptor{{
+			ID:                  "openrouter",
+			Name:                "OpenRouter",
+			Transport:           providercatalog.TransportOpenAICompatible,
+			DefaultBaseURL:      "https://openrouter.ai/api/v1",
+			DefaultModel:        "openai/gpt-4.1",
+			AuthEnvVars:         []string{"OPENROUTER_API_KEY"},
+			RequiresAuth:        true,
+			SupportedAPIFormats: []providercatalog.APIFormat{providercatalog.APIFormatOpenAIChatCompletions},
+		}},
+		models: []providerWizardModel{{ID: "openai/gpt-4.1", Description: "GPT-4.1"}},
+		apiKey: "sk-new",
+	}
+
+	updated, _ := m.applyProviderWizard()
+	next := updated
+
+	if next.providerWizard == nil || next.providerWizard.err == "" {
+		t.Fatal("a persist failure must surface a wizard error and keep the wizard open")
+	}
+	if next.provider != oldProvider {
+		t.Fatal("live provider must NOT be swapped when the persist fails")
+	}
+	if next.providerName != "old-provider" {
+		t.Fatalf("providerName = %q, want it unchanged on persist failure", next.providerName)
+	}
+	if got := os.Getenv(config.ActiveProviderEnv); got != "old-provider" {
+		t.Fatalf("%s = %q, want it unchanged on persist failure (children would diverge from parent)", config.ActiveProviderEnv, got)
+	}
+}

internal/tui/queued_message_test.go

@@ -55,6 +55,38 @@ func TestQueuedPromptPreviewAppearsInView(t *testing.T) {
 	}
 }
 
+// TestQueuedPromptPreviewSitsAboveComposer: the queued-message preview must
+// render ABOVE the input box, not below it — a message waiting to send should
+// sit on top of what you're currently typing.
+func TestQueuedPromptPreviewSitsAboveComposer(t *testing.T) {
+	m := newQueuedMessageTestModel(t)
+	m.pending = true
+	m.activeRunID = 1
+	m.runID = 1
+	m.width = 96
+	m.input.SetValue("summarize the failing test output")
+	updated, _ := m.Update(testKey(tea.KeyEnter))
+	next := updated.(model)
+
+	footer := plainRender(t, next.footerView(96))
+	lines := strings.Split(footer, "\n")
+	queuedLine, composerTopLine := -1, -1
+	for i, ln := range lines {
+		if queuedLine < 0 && strings.Contains(ln, "queued") {
+			queuedLine = i
+		}
+		if composerTopLine < 0 && strings.Contains(ln, "╭") { // composer box top border
+			composerTopLine = i
+		}
+	}
+	if queuedLine < 0 || composerTopLine < 0 {
+		t.Fatalf("expected both a queued line and a composer box; queued=%d composer=%d\n%s", queuedLine, composerTopLine, footer)
+	}
+	if queuedLine >= composerTopLine {
+		t.Fatalf("queued preview (line %d) must sit ABOVE the composer box top border (line %d):\n%s", queuedLine, composerTopLine, footer)
+	}
+}
+
 func TestAgentResponseLaunchesQueuedPrompt(t *testing.T) {
 	provider := &scriptedProvider{scripts: [][]zeroruntime.StreamEvent{{
 		{Type: zeroruntime.StreamEventText, Content: "queued answer"},