fix: drop npm@latest from the OIDC publish path

Rely on the npm bundled with the pinned Node 24 toolchain (>= 11.5.1,
the trusted-publishing minimum) instead of pulling an unpinned
npm@latest from the registry inside the publish job, and fail fast
with a clear error if a future toolchain change regresses it.

.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: ${{ github.token }}
@@ -92,3 +96,83 @@ jobs:
           tag="${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
+    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 @@
+{}

internal/agent/loop.go

@@ -210,27 +210,29 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			}
 		}
 
-		// Wrapped callbacks flag ANY output forwarded this turn, so a stall retry
-		// never re-streams on top of output the user already saw — even across the
-		// reactive-compaction retry below, which overwrites `collected`.
-		forwardedAnything := false
-		mark := func() { forwardedAnything = true }
-		// OnUsage is telemetry, not visible output, so it does NOT mark
-		// forwardedAnything: a provider that emits an early usage event (e.g. prompt
-		// tokens) then stalls with no content must still be safely retried. The gate
-		// only guards against re-streaming output the user actually SAW.
+		// 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) { mark(); options.OnText(s) }
+			forwardingOpts.OnText = func(s string) { forwardedVisibleText = true; options.OnText(s) }
 		}
 		if options.OnReasoning != nil {
-			forwardingOpts.OnReasoning = func(s string) { mark(); options.OnReasoning(s) }
+			forwardingOpts.OnReasoning = func(s string) { options.OnReasoning(s) }
 		}
 		if options.OnToolCallStart != nil {
-			forwardingOpts.OnToolCallStart = func(id, name string) { mark(); options.OnToolCallStart(id, name) }
+			forwardingOpts.OnToolCallStart = func(id, name string) { options.OnToolCallStart(id, name) }
 		}
 		if options.OnToolCallDelta != nil {
-			forwardingOpts.OnToolCallDelta = func(id, fragment string) { mark(); options.OnToolCallDelta(id, fragment) }
+			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)
@@ -284,15 +286,28 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			result.Messages = copyMessages(messages)
 			return result, ctx.Err()
 		}
-		// A stream idle/stall timeout with NO output forwarded yet this turn can be
-		// safely re-issued on a fresh connection — nothing was shown, so there's no
-		// duplication (this recovers the macOS stale-pooled-connection hang when it
-		// slips past the transport's response-header timeout). A turn that already
-		// streamed partial text/tool-calls is NOT retried (it would duplicate) and
-		// falls through to the error return below. Capped + exponential backoff.
+		// 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) && !forwardedAnything &&
-			collected.Text == "" && len(collected.ToolCalls) == 0; attempt++ {
+			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

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

@@ -10,19 +10,25 @@ import (
 )
 
 // stallProvider connects successfully (HTTP 200) but the stream emits a stall/idle
-// timeout error for the first stallBefore calls, then succeeds with "done". When
-// partialText is set it emits that text BEFORE the stall, simulating a stall after
-// partial output (which must NOT be retried — re-issuing would duplicate it).
+// 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
+	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, 3)
+	ch := make(chan zeroruntime.StreamEvent, 5)
 	if n <= p.stallBefore {
 		if p.partialText != "" {
 			ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventText, Content: p.partialText}
@@ -30,9 +36,16 @@ func (p *stallProvider) StreamCompletion(_ context.Context, _ zeroruntime.Comple
 		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 10m (the model produced nothing)",
+			Error: "provider stream error: no output for 6m (the model produced nothing)",
 		}
 		close(ch)
 		return ch, nil
@@ -72,21 +85,54 @@ func TestRunDoesNotRetryStallAfterPartialOutput(t *testing.T) {
 	}
 }
 
-// A stall after only REASONING was forwarded (no text/tool-calls) must NOT be
-// retried: the user already saw the reasoning, so re-issuing with callbacks
-// re-enabled would duplicate it. collected.Text is empty here, so this exercises
-// the turn-level forwarded-output gate specifically, not the text check.
-func TestRunDoesNotRetryStallAfterForwardedReasoning(t *testing.T) {
+// 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"}
-	_, err := Run(context.Background(), "go", p, Options{
+	result, err := Run(context.Background(), "go", p, Options{
 		Registry:    tools.NewRegistry(),
 		OnReasoning: func(string) {},
 	})
-	if err == nil {
-		t.Fatal("a stall after forwarded reasoning must NOT be retried; want an error")
+	if err != nil {
+		t.Fatalf("a stall after only reasoning should retry to success, got %v", err)
 	}
-	if got := atomic.LoadInt32(&p.calls); got != 1 {
-		t.Fatalf("reasoning-then-stall must not retry, got %d calls", got)
+	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)
 	}
 }
 

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/providers/providerio/providerio.go

@@ -6,10 +6,12 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"math"
 	"net/http"
 	"net/url"
 	"os"
 	"regexp"
+	"runtime"
 	"strconv"
 	"strings"
 	"time"
@@ -25,9 +27,9 @@ 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 streamContentStallFactor × the idle timeout. Without this an
-// upstream that heartbeats-but-stalls — observed on chatgpt/gpt-5.x and ollama
-// reasoning models — would hang the agent indefinitely.
+// 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
@@ -38,7 +40,7 @@ var ErrStreamStalled = errors.New("stream stalled (upstream kept the connection
 // 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)", idleTimeout*streamContentStallFactor)
+		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)
 }
@@ -53,13 +55,32 @@ func StreamTimeoutMessage(err error, idleTimeout time.Duration) string {
 // Override globally with ZERO_STREAM_IDLE_TIMEOUT.
 const DefaultStreamIdleTimeout = 5 * time.Minute
 
-// streamContentStallFactor 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 streamContentStallFactor × the idle timeout, the stream is
-// treated as stalled and aborted (ErrStreamStalled). It is > 1 so a slow-but-
-// producing request — one that emits data between heartbeats — is never killed; the
-// content watchdog only catches a stream that heartbeats while producing nothing.
-const streamContentStallFactor = 2
+// 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
@@ -122,6 +143,21 @@ func NormalizeBaseURL(baseURL string, defaultBaseURL string, label string) (stri
 //     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
@@ -132,6 +168,7 @@ var sharedHTTPClient = func() *http.Client {
 	// content-stall watchdogs.
 	transport.ResponseHeaderTimeout = 120 * time.Second
 	transport.IdleConnTimeout = 30 * time.Second
+	transport.DisableKeepAlives = runtime.GOOS == "darwin"
 	return &http.Client{Transport: transport}
 }()
 
@@ -285,7 +322,7 @@ func ScanSSEDataWithContext(
 		// 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 := idleTimeout * streamContentStallFactor
+		contentTimeout := ContentStallTimeout(idleTimeout)
 		content := time.NewTimer(contentTimeout)
 		defer content.Stop()
 		contentC = content.C

internal/providers/providerio/providerio_test.go

@@ -4,7 +4,9 @@ import (
 	"context"
 	"errors"
 	"io"
+	"math"
 	"net/http"
+	"runtime"
 	"strings"
 	"testing"
 	"time"
@@ -203,11 +205,36 @@ 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
-// (idleTimeout × streamContentStallFactor) aborts with ErrStreamStalled when no
-// real data line arrives. This is the gpt-5.x / ollama "still generating forever"
-// hang.
+// (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() }()
@@ -235,7 +262,8 @@ func TestScanSSEDataWithContextAbortsOnContentStall(t *testing.T) {
 
 	done := make(chan error, 1)
 	go func() {
-		// idle 100ms → content stall at 200ms. Keep-alives reset idle but not content.
+		// 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 })
 	}()
 
@@ -258,7 +286,7 @@ func TestScanSSEDataWithContextContentResetsOnData(t *testing.T) {
 	pr, pw := io.Pipe()
 
 	go func() {
-		// One data line every 30ms for ~300ms (well past the 200ms content window,
+		// 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++ {
@@ -290,16 +318,16 @@ func TestScanSSEDataWithContextContentResetsOnData(t *testing.T) {
 }
 
 // StreamTimeoutMessage must give a stalled stream a distinct, accurate detail —
-// it reports the content window (idle × factor) and must NOT claim the upstream
-// "stopped sending data" (keep-alives were still arriving).
+// 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 10m") {
-		t.Fatalf("stalled message = %q, want it to report the 10m content window", stalled)
+	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)
@@ -324,6 +352,17 @@ func TestHTTPClientReturnsStallHardenedSharedClient(t *testing.T) {
 	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")
 	}

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/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/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"
@@ -790,6 +791,18 @@ 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)
 	// 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 {
@@ -2222,15 +2235,18 @@ 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()
@@ -2806,6 +2822,15 @@ 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 ""
@@ -2821,6 +2846,10 @@ 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)
 }
 

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"},

internal/tui/working_status_test.go

@@ -106,6 +106,46 @@ func TestQuietGenerationHint(t *testing.T) {
 	}
 }
 
+// TestQuietGenerationHintEscalatesPastHalfIdleTimeout: a heartbeating-but-
+// silent stream (chatgpt/gpt-5.x, ollama reasoning models — see
+// providerio.ErrStreamStalled) and a genuinely healthy-but-slow one look
+// identical under the plain "still generating… Xs" cue — the ticking number
+// is the only signal, whether real content is still coming or nothing ever
+// will. Past half the provider's idle timeout the cue must say so explicitly
+// and name when Zero's own content-stall watchdog will act, rather than
+// leaving the user to guess whether this is a hang.
+func TestQuietGenerationHintEscalatesPastHalfIdleTimeout(t *testing.T) {
+	// 30s idle timeout: half (15s) sits comfortably above quietWorkingHint (8s),
+	// leaving a clean window to observe the plain cue before it escalates.
+	t.Setenv("ZERO_STREAM_IDLE_TIMEOUT", "30s")
+	base := time.Date(2026, 6, 25, 12, 0, 0, 0, time.UTC)
+	m := model{now: func() time.Time { return base }}
+	m.activeRunID = 7
+	m.turnStartedAt = base
+
+	// Quiet past quietWorkingHint but under half the (30s) idle timeout -> the
+	// plain cue, same as TestQuietGenerationHint's normal case.
+	m.lastStreamActivity = base.Add(-10 * time.Second)
+	got := m.quietGenerationHint()
+	if !strings.Contains(got, "still generating") {
+		t.Fatalf("under half the idle timeout: want the plain cue, got %q", got)
+	}
+	if strings.Contains(got, "auto-recover") {
+		t.Fatalf("under half the idle timeout: should not have escalated yet, got %q", got)
+	}
+
+	// Quiet for >= half the idle timeout -> escalates to name the watchdog and
+	// its ceiling (providerio.ContentStallTimeout(30s) = 30s*1.2 = 36s here).
+	m.lastStreamActivity = base.Add(-15 * time.Second)
+	got = m.quietGenerationHint()
+	if !strings.Contains(got, "unusually quiet") || !strings.Contains(got, "auto-recover") {
+		t.Fatalf("past half the idle timeout: want an escalated cue naming the watchdog, got %q", got)
+	}
+	if !strings.Contains(got, "36s") {
+		t.Fatalf("escalated cue should name the content-stall ceiling (ContentStallTimeout(30s) = 36s), got %q", got)
+	}
+}
+
 // TestQuietHintHiddenWhenSidebarShowsIt: when the context sidebar is up (it
 // carries the "generating…" pulse in ACTIVITY), the working line must NOT also
 // show the hint — it appears in exactly one place.

release-please-config.json

@@ -0,0 +1,12 @@
+{
+  "$schema": "https://raw.githubusercontent.com/googleapis/release-please/main/schemas/config.json",
+  "release-type": "node",
+  "include-component-in-tag": false,
+  "bootstrap-sha": "02f0c0929bec65f5fc41bdcab3eb518d2b0b329a",
+  "packages": {
+    ".": {
+      "package-name": "@gitlawb/zero",
+      "release-as": "0.1.0"
+    }
+  }
+}