fix(tui): explicitly request the initial window size at startup

Zero is always in alt-screen mode (useAltScreen() is unconditionally
true), and transcriptView's `if m.altScreen && m.height > 0` gate is what
pads the rendered frame out to fill the full terminal height
(scrollableTranscriptItemsView pads bodyWindow to window.height). If
m.height is ever 0 - Bubble Tea documents an initial WindowSizeMsg as
delivered automatically on program start, but that's the terminal
proactively pushing a size, and nothing else ever asks again if that
push is missed (a slow/unusual terminal, a multiplexer, a startup race)
- Zero falls back to the unpadded, non-fullscreen render path for the
rest of the session: content only occupies as many rows as it needs,
leaving the remaining terminal height as plain unfilled background
instead of the alt-screen's painted panel color.

Fixed by explicitly firing tea.RequestWindowSize in Init(), the same
defensive pattern already used for tea.RequestBackgroundColor right below
it - Zero doesn't rely solely on an unprompted push for either.

CHANGELOG.md

@@ -11,7 +11,14 @@ tagged. Until then, source builds report the version `dev`.
 ### 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 +210,72 @@ 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)
+		// 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.
+		forwardingOpts := zeroruntime.CollectOptions{OnUsage: options.OnUsage}
+		if options.OnText != nil {
+			forwardingOpts.OnText = func(s string) { mark(); options.OnText(s) }
+		}
+		if options.OnReasoning != nil {
+			forwardingOpts.OnReasoning = func(s string) { mark(); options.OnReasoning(s) }
+		}
+		if options.OnToolCallStart != nil {
+			forwardingOpts.OnToolCallStart = func(id, name string) { mark(); options.OnToolCallStart(id, name) }
+		}
+		if options.OnToolCallDelta != nil {
+			forwardingOpts.OnToolCallDelta = func(id, fragment string) { mark(); 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 +284,45 @@ 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.
+		for attempt := 1; attempt <= maxStreamStallRetries &&
+			isStreamTimeoutError(collected.Error) && !forwardedAnything &&
+			collected.Text == "" && len(collected.ToolCalls) == 0; attempt++ {
+			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 +675,20 @@ 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 +713,53 @@ 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/stall_retry_test.go

@@ -0,0 +1,122 @@
+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". 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).
+type stallProvider struct {
+	calls         int32
+	stallBefore   int32
+	partialText   string
+	reasoningText 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)
+	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}
+		}
+		ch <- zeroruntime.StreamEvent{
+			Type:  zeroruntime.StreamEventError,
+			Error: "provider stream error: no output for 10m (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 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) {
+	p := &stallProvider{stallBefore: 1, reasoningText: "thinking hard"}
+	_, 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 got := atomic.LoadInt32(&p.calls); got != 1 {
+		t.Fatalf("reasoning-then-stall must not retry, got %d calls", got)
+	}
+}
+
+// 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/cli/app.go

@@ -217,9 +217,9 @@ 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)
@@ -641,6 +641,7 @@ 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,
@@ -1033,9 +1034,10 @@ 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 +1045,7 @@ 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 +1055,10 @@ 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

@@ -1397,7 +1397,7 @@ 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())
 	}
 }

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/oauth_provider.go

@@ -43,6 +43,9 @@ func oauthResolverForProfile(profile config.ProviderProfile) providerio.TokenRes
 	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

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/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/resolver.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"net/url"
 	"os"
+	"strconv"
 	"strings"
 
 	"github.com/Gitlawb/zero/internal/modelregistry"
@@ -13,7 +14,15 @@ import (
 	"github.com/Gitlawb/zero/internal/sandbox"
 )
 
-const defaultMaxTurns = 30
+// 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 +195,9 @@ 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)
 }
 
@@ -431,11 +443,53 @@ 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),

internal/config/resolver_test.go

@@ -883,8 +883,8 @@ 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)
 	}
 }
 

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/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/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/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_test.go

@@ -264,6 +264,13 @@ 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_TOKENS_PATH", t.TempDir()+"/tokens.json")
+
 	transport := &captureTransport{
 		responseBody: "data: [DONE]\n\n",
 	}

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
 	}

internal/providers/openai/codex_responses.go

@@ -59,6 +59,7 @@ const (
 	responsesEventOutputItemAdded   = "response.output_item.added"
 	responsesEventContentPartAdded  = "response.content_part.added"
 	responsesEventOutputTextDelta   = "response.output_text.delta"
+	responsesEventReasoningDelta    = "response.reasoning_summary_text.delta"
 	responsesEventOutputTextDone    = "response.output_text.done"
 	responsesEventFunctionArgsDelta = "response.function_call_arguments.delta"
 	responsesEventContentPartDone   = "response.content_part.done"
@@ -86,6 +87,10 @@ type responsesRequest struct {
 // here; omitted entirely when the caller requests no (or an unsupported) effort.
 type responsesReasoning struct {
 	Effort string `json:"effort,omitempty"`
+	// Summary requests a streamed reasoning summary ("auto" lets the API pick a
+	// level). Without it the backend emits no reasoning events, so a long thinking
+	// phase produces zero visible output and reads as a hang in the UI.
+	Summary string `json:"summary,omitempty"`
 }
 
 // inputItem is one element of the Responses `input` array. The Type field
@@ -256,7 +261,9 @@ func (p *CodexProvider) buildResponsesRequest(request zeroruntime.CompletionRequ
 	// and an empty or unsupported effort simply omits the field — without this the
 	// caller's chosen effort was silently dropped for every Codex model.
 	if effort := openAIReasoningEffort(request.ReasoningEffort); effort != "" {
-		req.Reasoning = &responsesReasoning{Effort: effort}
+		// Summary "auto" makes the backend stream reasoning_summary_text deltas so a
+		// long thinking phase shows live progress instead of looking hung.
+		req.Reasoning = &responsesReasoning{Effort: effort, Summary: "auto"}
 	}
 	return req, nil
 }
@@ -382,10 +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
 	}
@@ -501,6 +508,17 @@ func (p *CodexProvider) emitResponsesEvent(
 			})
 		}
 		return true
+	case responsesEventReasoningDelta:
+		// Reasoning summary deltas: surface as live "thinking" so a long reasoning
+		// phase shows progress (and keeps the activity clock fresh) instead of
+		// looking like a hang. Requested via reasoning.summary="auto".
+		if event.Delta != "" {
+			providerio.SendEvent(ctx, events, zeroruntime.StreamEvent{
+				Type:    zeroruntime.StreamEventReasoning,
+				Content: event.Delta,
+			})
+		}
+		return true
 	case responsesEventFunctionArgsDelta:
 		p.handleFunctionArgsDelta(ctx, &event, state, events)
 		return true

internal/providers/openai/codex_test.go

@@ -485,6 +485,58 @@ func TestCodexProviderForwardsReasoningEffort(t *testing.T) {
 	if reasoning["effort"] != "high" {
 		t.Fatalf("body.reasoning.effort = %#v, want high", reasoning["effort"])
 	}
+	// A reasoning summary must be requested so the backend streams
+	// reasoning_summary_text deltas — otherwise a long think shows no live output.
+	if reasoning["summary"] != "auto" {
+		t.Fatalf("body.reasoning.summary = %#v, want auto", reasoning["summary"])
+	}
+}
+
+func TestCodexProviderStreamsReasoningSummaryDeltas(t *testing.T) {
+	// reasoning_summary_text deltas must surface as StreamEventReasoning (live
+	// "thinking"), in order, alongside the normal text output. Without this a long
+	// reasoning phase produces zero visible output and reads as a hang.
+	var rec codexRequest
+	srv := newCodexResponsesServer(t, &rec,
+		`{"type":"response.created","response":{"id":"resp-1","status":"in_progress"}}`,
+		`{"type":"response.reasoning_summary_text.delta","delta":"Thinking. "}`,
+		`{"type":"response.reasoning_summary_text.delta","delta":"Planning."}`,
+		`{"type":"response.output_text.delta","delta":"done"}`,
+		`{"type":"response.completed","response":{"id":"resp-1","status":"completed"}}`,
+	)
+	defer srv.Close()
+
+	provider, err := NewCodexProvider(CodexOptions{
+		Options:   Options{APIKey: "sk-test", BaseURL: srv.URL, Model: "gpt-5"},
+		AccountID: "acc-x",
+	})
+	if err != nil {
+		t.Fatalf("NewCodexProvider: %v", err)
+	}
+	stream, err := provider.StreamCompletion(context.Background(), zeroruntime.CompletionRequest{
+		Messages:        []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "solve it"}},
+		ReasoningEffort: "high",
+	})
+	if err != nil {
+		t.Fatalf("StreamCompletion: %v", err)
+	}
+	var reasoning, text []string
+	for ev := range stream {
+		switch ev.Type {
+		case zeroruntime.StreamEventReasoning:
+			reasoning = append(reasoning, ev.Content)
+		case zeroruntime.StreamEventText:
+			text = append(text, ev.Content)
+		case zeroruntime.StreamEventError:
+			t.Fatalf("unexpected error event: %q", ev.Error)
+		}
+	}
+	if got := strings.Join(reasoning, ""); got != "Thinking. Planning." {
+		t.Fatalf("reasoning deltas = %q, want %q", got, "Thinking. Planning.")
+	}
+	if got := strings.Join(text, ""); got != "done" {
+		t.Fatalf("text deltas = %q, want %q", got, "done")
+	}
 }
 
 func TestCodexProviderOmitsReasoningWhenUnset(t *testing.T) {

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

@@ -23,6 +23,26 @@ 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 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.
+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)", idleTimeout*streamContentStallFactor)
+	}
+	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 +53,14 @@ var ErrStreamIdle = errors.New("idle timeout (upstream stopped sending data)")
 // 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
+
 // 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 +105,42 @@ 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.
+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
+	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 +260,36 @@ 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 := idleTimeout * streamContentStallFactor
+		content := time.NewTimer(contentTimeout)
+		defer content.Stop()
+		contentC = content.C
+		resetContent = reset(content, contentTimeout)
 	}
 
 	for {
@@ -232,6 +305,12 @@ 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 +325,11 @@ 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 +4,7 @@ import (
 	"context"
 	"errors"
 	"io"
+	"net/http"
 	"strings"
 	"testing"
 	"time"
@@ -201,3 +202,133 @@ func TestUpstreamUnreachable(t *testing.T) {
 		})
 	}
 }
+
+// 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.
+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 200ms. 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 200ms 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 (idle × factor) 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, "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)
+	}
+	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/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/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/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/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 +1,117 @@
+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
+)
+
+// 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
+}
+
+// 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 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

@@ -40,41 +40,58 @@ 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).
@@ -280,10 +297,39 @@ 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 +348,12 @@ 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 +382,21 @@ 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).
@@ -608,56 +675,57 @@ 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 +790,18 @@ 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 {
@@ -736,6 +816,13 @@ 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 +899,14 @@ 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 +985,17 @@ 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 +1037,13 @@ 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 {
@@ -953,9 +1066,19 @@ 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
 			}
 			if m.mcpCommandCancel != nil {
@@ -1003,6 +1126,11 @@ 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
 			}
@@ -1012,11 +1140,24 @@ func (m model) updateModel(msg tea.Msg) (tea.Model, tea.Cmd) {
 			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 +1278,9 @@ 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 +1318,17 @@ 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 +1353,7 @@ 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 +1370,7 @@ 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 +1395,7 @@ 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 +1454,9 @@ 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 +1599,10 @@ 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 +1713,7 @@ 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
@@ -1856,6 +2015,14 @@ 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 +2046,27 @@ 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
 }
@@ -2030,10 +2215,14 @@ 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
@@ -3199,14 +3388,20 @@ 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 +3418,20 @@ 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 +3672,26 @@ 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 +3955,8 @@ 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

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/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

@@ -166,6 +166,10 @@ 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
@@ -1229,9 +1233,29 @@ 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/rendering.go

@@ -36,7 +36,7 @@ func pickerBusyText(name string) string {
 			Title: "Busy",
 			Lines: []string{"Can't change " + label + " while a run is in progress."},
 		}},
-		Hints: []string{"press Esc to cancel the run, then try again"},
+		Hints: []string{"press Esc twice to cancel the run, then try again"},
 	})
 }
 

internal/tui/session_controls.go

@@ -9,6 +9,7 @@ import (
 	tea "charm.land/bubbletea/v2"
 
 	"github.com/Gitlawb/zero/internal/agent"
+	"github.com/Gitlawb/zero/internal/config"
 	"github.com/Gitlawb/zero/internal/modelregistry"
 	"github.com/Gitlawb/zero/internal/sessions"
 	"github.com/Gitlawb/zero/internal/usage"
@@ -256,6 +257,42 @@ func (m model) handleSelfCorrectCommand(args string) (model, string) {
 	return m, m.selfCorrectText()
 }
 
+// maxTurnsCeiling caps the per-session /turns budget so a typo (e.g. /turns 99999)
+// can't set an absurd ceiling; real multi-step tasks fit comfortably under it. Shared
+// with config so applyEnv enforces the same bound on an inherited ZERO_MAX_TURNS.
+const maxTurnsCeiling = config.MaxTurnsCeiling
+
+func (m model) handleTurnsCommand(args string) (model, string) {
+	args = strings.TrimSpace(args)
+	if args == "" || strings.EqualFold(args, "status") {
+		return m, m.turnsText()
+	}
+	n, err := strconv.Atoi(args)
+	if err != nil || n < 1 {
+		return m, "Turns\nUsage: /turns <n> — set the per-run tool-turn budget for this session (a positive number, e.g. /turns 150)."
+	}
+	if n > maxTurnsCeiling {
+		n = maxTurnsCeiling
+	}
+	m.agentOptions.MaxTurns = n
+	// Propagate the budget to spawned sub-agents / swarm members (which inherit the
+	// environment) so a delegated task gets the same budget, not config.json's default.
+	config.SetMaxTurnsEnv(n)
+	return m, m.turnsText()
+}
+
+func (m model) turnsText() string {
+	return renderCommandOutput(commandOutput{
+		Title:  "Turns",
+		Status: commandStatusOK,
+		Sections: []commandSection{{
+			Title: "State",
+			Lines: []string{fmt.Sprintf("max tool-turns per run: %d", m.agentOptions.MaxTurns)},
+		}},
+		Hints: []string{fmt.Sprintf("/turns <n> sets this session's tool-turn budget (max %d); raise it for long multi-step tasks like delegation/swarm runs", maxTurnsCeiling)},
+	})
+}
+
 func (m model) selfCorrectText() string {
 	depth := "LSP diagnostics only — fast, scoped to changed files"
 	if m.selfCorrectTests {

internal/tui/session_test.go

@@ -775,6 +775,8 @@ func TestEscCancelRecordsSessionError(t *testing.T) {
 		t.Fatal("expected prompt submit to start an agent run")
 	}
 
+	updated, _ = next.Update(testKey(tea.KeyEsc))
+	next = updated.(model)
 	updated, _ = next.Update(testKey(tea.KeyEsc))
 	next = updated.(model)
 
@@ -836,7 +838,10 @@ func TestCancelledRunFlushesCheckpointSessionEvents(t *testing.T) {
 		next = updated.(model)
 		if _, ok := runtimeMsg.(permissionRequestMsg); ok {
 			// Cancel mid-run via Esc while the permission prompt is pending: this
-			// unblocks the goroutine through ctx cancellation.
+			// unblocks the goroutine through ctx cancellation. Esc now requires a
+			// second press within the confirmation window to actually cancel.
+			updated, _ = next.Update(testKey(tea.KeyEsc))
+			next = updated.(model)
 			updated, _ = next.Update(testKey(tea.KeyEsc))
 			next = updated.(model)
 			cancelled = true

internal/tui/sidebar.go

@@ -17,6 +17,7 @@ import (
 	"time"
 
 	"charm.land/lipgloss/v2"
+	"github.com/charmbracelet/x/ansi"
 
 	"github.com/Gitlawb/zero/internal/tools"
 )
@@ -261,13 +262,11 @@ func (m model) swarmSpawnedAgents() []swarmAgent {
 			switch a.state {
 			case "done", "failed", "completed", "cancelled":
 				a.finishing = true
-				if m.pending {
-					// Run still going: keep it visible and clickable, no fade.
-					a.finishedAt = time.Time{}
-					live = append(live, a)
-					continue
-				}
-				// Turn ended: fade out over the linger window, then drop.
+				// A member drops once ITS OWN task completes — not when the whole turn
+				// ends: fade out over the linger window from when it was first seen
+				// finished (stamped each tick by stampSwarmDone), then remove. This
+				// holds whether or not the overall run is still in flight, mirroring
+				// how finished specialists drop (sidebarSpecialists).
 				doneAt, stamped := m.swarmDoneAt[a.id]
 				if stamped && m.now().Sub(doneAt) >= sidebarAgentLinger {
 					continue // past the linger window — remove
@@ -609,6 +608,15 @@ func (m model) renderContextSidebar(width, height int) []string {
 	}
 	add(tokenLine)
 
+	// Hover highlight: resolved by STABLE IDENTITY (sessionID / stepIndex), not a
+	// cached line offset — see hoveredSidebarLineOffset. A row whose identity no
+	// longer resolves (it disappeared since the hover was last set from a real
+	// mouse motion) simply doesn't highlight, rather than a coincidentally-matching
+	// unrelated row lighting up.
+	if lineOffset, ok := m.hoveredSidebarLineOffset(width); ok && lineOffset >= 0 && lineOffset < len(lines) {
+		lines[lineOffset] = zeroTheme.hover.Render(ansi.Strip(lines[lineOffset]))
+	}
+
 	// Normalize every row to exactly width cells.
 	for i := range lines {
 		lines[i] = padStyledLine(lines[i], width)
@@ -622,6 +630,31 @@ func (m model) renderContextSidebar(width, height int) []string {
 	return lines
 }
 
+// hoveredSidebarLineOffset resolves the hovered sidebar row's CURRENT line offset
+// fresh on every call, by re-matching m.hover's stable identity (sessionID for an
+// agent, stepIndex for a plan step) against a freshly computed hit list — never a
+// cached index. Returns false when the hover isn't sidebar-scoped, or when the
+// identity no longer resolves (that row disappeared since the hover was last set
+// by a real mouse motion — a linger window elapsing, a plan step completing —
+// with no intervening motion to re-target it).
+func (m model) hoveredSidebarLineOffset(width int) (int, bool) {
+	switch m.hover.kind {
+	case hoverSidebarAgent:
+		for _, hit := range m.sidebarAgentSelectables(width) {
+			if hit.sessionID == m.hover.sessionID {
+				return hit.lineOffset, true
+			}
+		}
+	case hoverPlanStep:
+		for _, hit := range m.sidebarPlanSelectables(width) {
+			if hit.stepIndex == m.hover.stepIndex {
+				return hit.lineOffset, true
+			}
+		}
+	}
+	return 0, false
+}
+
 // sidebarHeader renders a bold-muted uppercase section label. Bold muted (vs the
 // faint body items and placeholders) gives the header enough weight to read as a
 // section heading rather than more filler. The width arg is unused — kept so it

internal/tui/sidebar_test.go

@@ -391,32 +391,32 @@ func TestSidebarShowsSwarmSpawnedAgents(t *testing.T) {
 // row (e.g. a resumed transcript that dropped the call): the member is still
 // A finished member stays visible (and clickable) while the run is still in
 // flight, so the user can inspect it; only once the turn ends does it drop.
-func TestSwarmAgentsPersistWhileRunInFlight(t *testing.T) {
+func TestSwarmAgentDropsOnOwnCompletionEvenMidRun(t *testing.T) {
 	base := time.Date(2026, 6, 24, 12, 0, 0, 0, time.UTC)
 	m := sidebarTestModel()
 	m.now = func() time.Time { return base }
-	m.pending = true  // run still going
+	m.pending = true  // run STILL going — a member must still drop on its OWN completion
 	m.activeRunID = 7 // exercise the run-scoped filter with a non-zero id
 	m.transcript = append(m.transcript,
 		transcriptRow{kind: rowToolCall, tool: "swarm_spawn", detail: "build homepage", runID: 7},
 		transcriptRow{kind: rowToolResult, tool: "swarm_spawn", detail: "Spawned subagent as task subagent-1 on team default.", runID: 7},
 		transcriptRow{kind: rowToolResult, tool: "swarm_collect", detail: "Results: 1 task(s)\n- subagent-1 [done] build homepage", runID: 7},
 	)
-	// Long past the linger window — but it must still show while pending.
-	m.swarmDoneAt = map[string]time.Time{"subagent-1": base.Add(-10 * sidebarAgentLinger)}
-
+	// Just finished, still within the linger window: shows briefly with a fading ✓.
+	m.swarmDoneAt = map[string]time.Time{"subagent-1": base.Add(-sidebarAgentLinger / 2)}
 	agents := m.swarmSpawnedAgents()
 	if len(agents) != 1 {
-		t.Fatalf("a finished member must stay while the run is in flight, got %d: %+v", len(agents), agents)
+		t.Fatalf("within the linger window a finished member should still show (fading), got %d: %+v", len(agents), agents)
 	}
 	if !agents[0].finishing {
 		t.Fatalf("a finished member should render done (✓), got %+v", agents[0])
 	}
 
-	// Once the turn ends, the long-finished member fades out and drops.
-	m.pending = false
+	// Past the linger window: it drops — even though the overall run is still in
+	// flight (previously a finished member lingered until the whole turn ended).
+	m.swarmDoneAt = map[string]time.Time{"subagent-1": base.Add(-2 * sidebarAgentLinger)}
 	if got := len(m.swarmSpawnedAgents()); got != 0 {
-		t.Fatalf("after the turn ends a long-finished member should drop, got %d", got)
+		t.Fatalf("a member past its linger window must drop on its own completion mid-run, got %d", got)
 	}
 }
 

internal/tui/subchat_regression_test.go

@@ -0,0 +1,118 @@
+package tui
+
+import (
+	"strings"
+	"testing"
+
+	tea "charm.land/bubbletea/v2"
+)
+
+// The subchat drill-in (viewing a subagent/swarm child session) swaps the on-screen
+// transcript to m.subchat.childRows, but mouse hit-testing and the footer's pinned
+// plan panel were never made subchat-aware — they kept operating on/showing the
+// PARENT run's state while the child session was on screen. These tests guard the
+// fix: mouse selection must resolve against the child rows, the plan panel must not
+// show in that view, and a selection in progress must extend across a wheel-scroll
+// instead of freezing at whatever was visible when the drag started.
+
+func TestFooterHidesPlanPanelDuringSubchat(t *testing.T) {
+	m := runningPlanModel(t, 3)
+	m.altScreen = true
+	m.height = 30
+	width := m.chatColumnWidth()
+
+	withoutSubchat := m.footerView(width)
+	if !strings.Contains(withoutSubchat, "Step number 1 here") {
+		t.Fatalf("sanity check failed: pinned plan panel should render outside subchat, got:\n%s", withoutSubchat)
+	}
+
+	m.subchat.active = true
+	m.subchat.childSessionID = "child-1"
+	withSubchat := m.footerView(width)
+	if strings.Contains(withSubchat, "Step number 1 here") {
+		t.Fatalf("plan panel must NOT show while viewing a subchat child session, got:\n%s", withSubchat)
+	}
+}
+
+func TestTranscriptSelectionInSubchatUsesChildSessionRows(t *testing.T) {
+	m := mouseTestModel()
+	m.mouseCapture = true
+	// Parent transcript has DIFFERENT text than the child session: if selection
+	// hit-tests against the wrong (parent) rows, this proves it by either matching
+	// nothing or matching the wrong text.
+	m.transcript = appendRow(m.transcript, rowUser, "parent transcript text")
+	m.subchat.active = true
+	m.subchat.childSessionID = "child-1"
+	m.subchat.childRows = appendRow(nil, rowUser, "hello world")
+
+	textY := topmostVisibleTranscriptMouseY(t, m)
+	updated, _ := m.Update(testMouseClick(tea.MouseLeft, 3, textY))
+	m = updated.(model)
+	updated, _ = m.Update(testMouseMotion(tea.MouseLeft, 8, textY))
+	m = updated.(model)
+
+	if got := m.selectedTranscriptText(); got != "hello" {
+		t.Fatalf("selectedTranscriptText() in subchat = %q, want hello (from the CHILD session's rows)", got)
+	}
+}
+
+func TestTranscriptSelectionExtendsAcrossWheelScroll(t *testing.T) {
+	m := mouseTestModel()
+	m.mouseCapture = true
+	// Enough rows that the transcript overflows the viewport, so a wheel-scroll is
+	// a real (non-clamped) scroll. chatScrollOffset=0 anchors to the BOTTOM (newest
+	// content, see transcriptViewport.window: start = totalLines-height-offset), so
+	// with overflow content the topmost VISIBLE line is not the transcript's first
+	// line — topmostVisibleTranscriptMouseY (window-aware) finds it correctly.
+	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
+
+	// Wheel UP reveals OLDER content above (offset increases -> window.start
+	// decreases): re-evaluating the SAME on-screen Y afterward must land on an
+	// EARLIER (smaller bodyY) line now that the viewport has shifted, extending the
+	// selection upward instead of leaving the cursor pinned to the pre-scroll line.
+	updated, _ = m.Update(testMouseWheel(tea.MouseWheelUp, 0, textY))
+	m = updated.(model)
+
+	if m.chatScrollOffset == scrollBefore {
+		t.Fatal("sanity check failed: wheel-up should have scrolled (80 rows overflow a 30-row terminal)")
+	}
+	if !m.transcriptSelection.active {
+		t.Fatal("selection must survive a wheel-scroll, not be cleared")
+	}
+	if m.transcriptSelection.cursor.bodyY >= cursorBefore {
+		t.Fatalf("selection cursor bodyY = %d, want < %d (it must extend to follow the scroll, not freeze)", m.transcriptSelection.cursor.bodyY, cursorBefore)
+	}
+}
+
+// topmostVisibleTranscriptMouseY returns the on-screen Y of the topmost currently
+// VISIBLE selectable text line — window-aware (unlike firstTranscriptTextMouseY,
+// which walks the full unwindowed layout and only lands on-screen when everything
+// fits in one viewport). It resolves against transcriptHitTestSource, the same
+// subchat-aware source transcriptLineAtMouse uses, so it works for both the parent
+// transcript and a subchat child session.
+func topmostVisibleTranscriptMouseY(t *testing.T, m model) int {
+	t.Helper()
+	header, items, width := m.transcriptHitTestSource()
+	frame := m.scrollableTranscriptFrame(header, m.footerView(width))
+	metrics := measureTranscriptBodyItems(items, m.transcriptBodyHeights)
+	window := transcriptViewportForLayout(metrics, frame, m.chatScrollOffset).window()
+	layout := layoutVisibleTranscriptBodyItems(items, metrics, window)
+	for _, line := range layout.selectable {
+		if line.text != "" && !line.toggle {
+			return frame.bodyRect.y + (line.bodyY - window.start)
+		}
+	}
+	t.Fatalf("no selectable visible transcript text line found: %#v", layout.selectable)
+	return 0
+}

internal/tui/theme.go

@@ -12,7 +12,8 @@ import (
 // may be swapped at startup (background detection / ZERO_THEME / --theme) or live
 // via /theme. Colors are truecolor hex; lipgloss downsamples on limited displays
 // and renders plain text when there is no TTY (tests). Every renderer consumes
-// these named styles — no hex literal may appear outside this file.
+// these named styles — no hex literal may appear outside theme_palettes.go (the
+// palette tables + theme registry).
 type tuiTheme struct {
 	// Base tokens.
 	ink        lipgloss.Style // primary text
@@ -29,6 +30,7 @@ type tuiTheme struct {
 	line       lipgloss.Style // default borders, rules, status separators
 	lineStrong lipgloss.Style // emphasized borders
 	selection  lipgloss.Style // transcript selection highlight
+	hover      lipgloss.Style // hovered clickable row (specialist card, toggle, sidebar/plan row)
 
 	// Title bar.
 	badge lipgloss.Style // ` 0 ` brand chip: onAccent on accent, bold
@@ -91,9 +93,10 @@ type tuiTheme struct {
 }
 
 // palette is the raw color-token table for one theme. buildTheme turns it into a
-// resolved tuiTheme. darkPalette and lightPalette are the ONLY place hex literals
-// live. All dark tints stay darker than ink so every pairing survives 256-color
-// downsampling; the light set is dark-on-light with the same intent inverted.
+// resolved tuiTheme. The palette literals and the ordered theme registry live in
+// theme_palettes.go (the only place hex literals live). Dark palettes keep tints
+// darker than ink so every pairing survives 256-color downsampling; light palettes
+// are dark-on-light with the same intent inverted.
 type palette struct {
 	panel     string // card backgrounds (the terminal canvas itself is never painted full-bleed)
 	promptBg  string // submitted user prompt background
@@ -124,73 +127,6 @@ type palette struct {
 	cardPerm  string // permission card border (amber mixed into line)
 }
 
-// darkPalette is the original Lime palette: a near-black chat surface with one
-// lime accent. bg (#070708) is the terminal's own canvas — deliberately never
-// painted — so no token references it.
-var darkPalette = palette{
-	panel:     "#0e0e10",
-	promptBg:  "#262626",
-	line:      "#242429",
-	line2:     "#414147",
-	ink:       "#ececee",
-	muted:     "#9a9aa2", // secondary text — lifted so it clearly out-ranks faint
-	faint:     "#8a8a92", // hints/metadata — nudged up to separate from faintest
-	faintest:  "#7c7c82", // line numbers/separators — pinned at the WCAG-AA floor on the dark panel
-	accent:    "#caff3f",
-	green:     "#5dd1a4",
-	red:       "#ff7a7a",
-	amber:     "#ffc25c",
-	blue:      "#7db4ff",
-	gitAdd:    "#7db87a",
-	gitDel:    "#b87a7a",
-	addBg:     "#18352c",
-	delBg:     "#241819",
-	addBgWord: "#2e654d", // changed span within an added line — brighter green (sep 1.83 vs addBg, addInk 7.1:1)
-	delBgWord: "#502d30", // changed span within a deleted line — brighter red (sep 1.44 vs delBg, delInk 7.7:1)
-	permBg:    "#1c1915",
-	selBg:     "#32401b", // selected row bg — brightened from #1d2114 so the highlighted row separates from the panel (sep 1.18→1.73) while ink label contrast stays ~9.4:1
-	addInk:    "#bdeed7",
-	delInk:    "#f2c4c4",
-	onAccent:  "#000000",
-	cardRun:   "#5a6b2e",
-	cardErr:   "#6b3434",
-	cardPerm:  "#6b5a2e",
-}
-
-// lightPalette is dark-on-light: light-gray surfaces with near-black ink. The
-// muted/faint/faintest grays get progressively LIGHTER (toward the surface) so the
-// same text hierarchy reads on white; the lime accent is darkened to keep contrast
-// against a light background, and the diff/permission tints become light surfaces.
-var lightPalette = palette{
-	panel:     "#ececed", // card backgrounds (slightly off-white)
-	promptBg:  "#dcdce0", // submitted prompt background
-	line:      "#cfcfd5", // default borders
-	line2:     "#b0b0b8", // emphasized borders
-	ink:       "#1b1b1d", // primary text (near-black)
-	muted:     "#4a4a52", // secondary text — darkened so it clearly out-ranks faint on light
-	faint:     "#57575f", // hints, metadata — nudged down to separate from faintest
-	faintest:  "#646469", // line numbers, separators — pinned at the WCAG-AA floor on the light panel
-	accent:    "#477006", // brand lime, darkened to AA contrast on light
-	green:     "#1c7a4a", // success / diff add
-	red:       "#c0322c", // errors / diff del
-	amber:     "#8f6200", // permission / warnings
-	blue:      "#1d5fd6", // grep locations
-	gitAdd:    "#2f7a3a", // footer PR diff additions
-	gitDel:    "#a83a3a", // footer PR diff deletions
-	addBg:     "#e2f3e6", // diff added-line bg (light green)
-	delBg:     "#fbe6e6", // diff deleted-line bg (light red)
-	addBgWord: "#97d3ab", // changed span within an added line — deeper green (sep 1.49 vs addBg, addInk 4.7:1)
-	delBgWord: "#f1b3b3", // changed span within a deleted line — deeper red (sep 1.48 vs delBg, delInk 4.6:1)
-	permBg:    "#fbf0d8", // permission card bg (light amber)
-	selBg:     "#cfe78f", // selected row bg (light accent) — deepened from #e7f2cd so it separates from the near-white panel (sep 1.01→1.15); ink label stays >12:1, faint ~5:1
-	addInk:    "#1d5b37", // added-line text
-	delInk:    "#8e2d2d", // deleted-line text
-	onAccent:  "#ffffff", // text on the (dark) accent/amber fills
-	cardRun:   "#94ad60", // running card border
-	cardErr:   "#cf9a9a", // errored card border
-	cardPerm:  "#c9ad7d", // permission card border
-}
-
 // buildTheme resolves a palette into the styles every renderer uses.
 func buildTheme(p palette) tuiTheme {
 	col := func(s string) color.Color { return lipgloss.Color(s) }
@@ -210,6 +146,13 @@ func buildTheme(p palette) tuiTheme {
 		line:       fg(p.line),
 		lineStrong: fg(p.line2),
 		selection:  lipgloss.NewStyle().Background(col(p.accent)).Foreground(col(p.onAccent)),
+		// A full block (like selection) would be too heavy for something that
+		// repaints on every mouse movement; a soft foreground reads as
+		// "interactive" at a glance without competing with real content colors
+		// (e.g. a specialist card's red/green status glyph). Amber (not the brand
+		// lime accent) — lime read as too glaring for something that repaints
+		// continuously as the cursor moves.
+		hover: fg(p.amber),
 
 		badge: lipgloss.NewStyle().Background(col(p.accent)).Foreground(col(p.onAccent)).Bold(true),
 

internal/tui/theme_palettes.go

@@ -0,0 +1,457 @@
+package tui
+
+import "strings"
+
+// theme_palettes.go is the sole home of raw color hex in the TUI: every palette
+// literal and the ordered theme registry live here, so theme.go stays hex-free
+// (just the palette struct, buildTheme, and the resolved tuiTheme). Adding a theme
+// is a new palette{...} literal plus one themeRegistry entry — nothing else.
+//
+// Contrast discipline: every palette must satisfy the same WCAG invariants the
+// built-ins do (asserted across the whole registry in theme_select_test.go): ink
+// and faintest ≥ AA on panel; onAccent ≥ AA on accent; addInk/delInk readable on
+// their diff bands and word spans; the muted>faint>faintest ramp monotonic toward
+// the surface; and selBg visibly separated from panel while its label stays legible.
+
+// darkPalette is the original Lime palette: a near-black chat surface with one
+// lime accent. bg (#070708) is the terminal's own canvas — deliberately never
+// painted — so no token references it.
+var darkPalette = palette{
+	panel:     "#0e0e10",
+	promptBg:  "#262626",
+	line:      "#242429",
+	line2:     "#414147",
+	ink:       "#ececee",
+	muted:     "#9a9aa2", // secondary text — lifted so it clearly out-ranks faint
+	faint:     "#8a8a92", // hints/metadata — nudged up to separate from faintest
+	faintest:  "#7c7c82", // line numbers/separators — pinned at the WCAG-AA floor on the dark panel
+	accent:    "#caff3f", // original bright lime (the brand accent)
+	green:     "#5dd1a4",
+	red:       "#ff7a7a",
+	amber:     "#ffc25c",
+	blue:      "#7db4ff",
+	gitAdd:    "#7db87a",
+	gitDel:    "#b87a7a",
+	addBg:     "#18352c",
+	delBg:     "#241819",
+	addBgWord: "#2e654d", // changed span within an added line — brighter green (sep 1.83 vs addBg, addInk 7.1:1)
+	delBgWord: "#502d30", // changed span within a deleted line — brighter red (sep 1.44 vs delBg, delInk 7.7:1)
+	permBg:    "#1c1915",
+	selBg:     "#32401b", // selected row bg — brightened from #1d2114 so the highlighted row separates from the panel (sep 1.18→1.73) while ink label contrast stays ~9.4:1
+	addInk:    "#bdeed7",
+	delInk:    "#f2c4c4",
+	onAccent:  "#000000",
+	cardRun:   "#5a6b2e",
+	cardErr:   "#6b3434",
+	cardPerm:  "#6b5a2e",
+}
+
+// draculaPalette — the Dracula scheme (dracula.com): muted-violet surface, purple
+// accent, high-chroma pink/green/cyan signals.
+var draculaPalette = palette{
+	panel:     "#282a36",
+	promptBg:  "#383c4d",
+	line:      "#363a4b",
+	line2:     "#484c62",
+	ink:       "#f8f8f2",
+	muted:     "#b9bccb",
+	faint:     "#a2a5b8",
+	faintest:  "#9195ac",
+	accent:    "#bd93f9",
+	green:     "#50fa7b",
+	red:       "#ff5555",
+	amber:     "#ffb86c",
+	blue:      "#8be9fd",
+	gitAdd:    "#77c58c",
+	gitDel:    "#d98d8d",
+	addBg:     "#1c3b2a",
+	delBg:     "#3a2026",
+	addBgWord: "#235035",
+	delBgWord: "#5e333b",
+	permBg:    "#322a1e",
+	selBg:     "#504482",
+	addInk:    "#cbf2dd",
+	delInk:    "#f4c9c9",
+	onAccent:  "#000000",
+	cardRun:   "#7c6aa6",
+	cardErr:   "#98505a",
+	cardPerm:  "#9a7c62",
+}
+
+// nordPalette — the Nord scheme (nordtheme.com): cool polar-night slate with a
+// frost-blue accent and desaturated aurora signals.
+var nordPalette = palette{
+	panel:     "#3b4252",
+	promptBg:  "#464f62",
+	line:      "#434c5e",
+	line2:     "#4c566a",
+	ink:       "#eceff4",
+	muted:     "#c8cfda",
+	faint:     "#b4bdcb",
+	faintest:  "#a5afc1",
+	accent:    "#88c0d0",
+	green:     "#a3be8c",
+	red:       "#bf616a",
+	amber:     "#d08770",
+	blue:      "#81a1c1",
+	gitAdd:    "#8ba077",
+	gitDel:    "#b0757d",
+	addBg:     "#37433a",
+	delBg:     "#45383d",
+	addBgWord: "#456d46",
+	delBgWord: "#6d4650",
+	permBg:    "#47413a",
+	selBg:     "#40688a",
+	addInk:    "#d6ecca",
+	delInk:    "#f0c6cb",
+	onAccent:  "#000000",
+	cardRun:   "#4c6672",
+	cardErr:   "#6b4a51",
+	cardPerm:  "#6a564d",
+}
+
+// gruvboxPalette — Gruvbox dark (medium): warm retro browns with an olive-green
+// accent and high-contrast cream ink.
+var gruvboxPalette = palette{
+	panel:     "#32302f",
+	promptBg:  "#3c3836",
+	line:      "#504945",
+	line2:     "#665c54",
+	ink:       "#ebdbb2",
+	muted:     "#c9b99a",
+	faint:     "#b7a78d",
+	faintest:  "#a89984",
+	accent:    "#8ec07c",
+	green:     "#b8bb26",
+	red:       "#fb4934",
+	amber:     "#fabd2f",
+	blue:      "#83a598",
+	gitAdd:    "#98971a",
+	gitDel:    "#cc241d",
+	addBg:     "#2f3a29",
+	delBg:     "#3b2b29",
+	addBgWord: "#3e5236",
+	delBgWord: "#593733",
+	permBg:    "#38331e",
+	selBg:     "#3d4e30",
+	addInk:    "#c5e6b0",
+	delInk:    "#f0c0bb",
+	onAccent:  "#000000",
+	cardRun:   "#6f8460",
+	cardErr:   "#a5493d",
+	cardPerm:  "#a5833a",
+}
+
+// tokyoNightPalette — Tokyo Night (storm): deep indigo surface, soft blue accent,
+// cool high-contrast ink.
+var tokyoNightPalette = palette{
+	panel:     "#1e2030",
+	promptBg:  "#2c3149",
+	line:      "#262a3d",
+	line2:     "#3b4261",
+	ink:       "#c8d3f5",
+	muted:     "#a9b1d0",
+	faint:     "#9099b2",
+	faintest:  "#838ba8",
+	accent:    "#82aaff",
+	green:     "#c3e88d",
+	red:       "#ff757f",
+	amber:     "#ffc777",
+	blue:      "#86e1fc",
+	gitAdd:    "#96bf7d",
+	gitDel:    "#c77e85",
+	addBg:     "#20303b",
+	delBg:     "#37222c",
+	addBgWord: "#2b5a4a",
+	delBgWord: "#5c2e3a",
+	permBg:    "#2a2419",
+	selBg:     "#2a385b",
+	addInk:    "#b8e4d3",
+	delInk:    "#f3c4cb",
+	onAccent:  "#000000",
+	cardRun:   "#4b5d8b",
+	cardErr:   "#7d4857",
+	cardPerm:  "#7d6954",
+}
+
+// catppuccinPalette — Catppuccin Mocha: soft lavender surface with a mauve accent
+// and pastel signals.
+var catppuccinPalette = palette{
+	panel:     "#1e1e2e",
+	promptBg:  "#34364b",
+	line:      "#313244",
+	line2:     "#45475a",
+	ink:       "#cdd6f4",
+	muted:     "#a6adc8",
+	faint:     "#9399b2",
+	faintest:  "#83889f",
+	accent:    "#cba6f7",
+	green:     "#a6e3a1",
+	red:       "#f38ba8",
+	amber:     "#f9e2af",
+	blue:      "#89b4fa",
+	gitAdd:    "#8cbf8a",
+	gitDel:    "#cc8a9b",
+	addBg:     "#24312b",
+	delBg:     "#3c2a32",
+	addBgWord: "#2f5140",
+	delBgWord: "#56333f",
+	permBg:    "#29261b",
+	selBg:     "#322e46",
+	addInk:    "#c4ecd6",
+	delInk:    "#f4cdd6",
+	onAccent:  "#000000",
+	cardRun:   "#7a6a99",
+	cardErr:   "#8a5b72",
+	cardPerm:  "#8d8274",
+}
+
+// oneDarkPalette — Atom One Dark: slate-gray surface with a blue accent.
+var oneDarkPalette = palette{
+	panel:     "#2e323b",
+	promptBg:  "#3a3f4a",
+	line:      "#393f4a",
+	line2:     "#4b525f",
+	ink:       "#abb2bf",
+	muted:     "#a2a9b6",
+	faint:     "#9aa1af",
+	faintest:  "#969cab",
+	accent:    "#61afef",
+	green:     "#98c379",
+	red:       "#e06c75",
+	amber:     "#e5c07b",
+	blue:      "#56b6c2",
+	gitAdd:    "#82a06a",
+	gitDel:    "#bd8087",
+	addBg:     "#2c382b",
+	delBg:     "#3a2d2f",
+	addBgWord: "#3d5a3a",
+	delBgWord: "#5c3e40",
+	permBg:    "#3c3826",
+	selBg:     "#354256",
+	addInk:    "#cdeab3",
+	delInk:    "#f0c3c7",
+	onAccent:  "#000000",
+	cardRun:   "#496c8c",
+	cardErr:   "#7c515b",
+	cardPerm:  "#7e735e",
+}
+
+// solarizedDarkPalette — Solarized Dark (Ethan Schoonover): the signature teal
+// base03/base02 surface with a cyan accent and the fixed accent wheel.
+var solarizedDarkPalette = palette{
+	panel:     "#073642",
+	promptBg:  "#0b3b46",
+	line:      "#123f48",
+	line2:     "#4b636c",
+	ink:       "#cdd6d6",
+	muted:     "#a9b3b3",
+	faint:     "#9ba5a5",
+	faintest:  "#929c9c",
+	accent:    "#3bb3a6", // solarized cyan, brightened to clear AA on the lifted base02 panel
+	green:     "#859900",
+	red:       "#dc322f",
+	amber:     "#b58900",
+	blue:      "#268bd2",
+	gitAdd:    "#93a05e",
+	gitDel:    "#c67b71",
+	addBg:     "#123f31",
+	delBg:     "#45302e",
+	addBgWord: "#1e5c44",
+	delBgWord: "#6a3f3a",
+	permBg:    "#2e2a18",
+	selBg:     "#17505a",
+	addInk:    "#c3ecd6",
+	delInk:    "#f2cbc6",
+	onAccent:  "#000000",
+	cardRun:   "#1d6b6c",
+	cardErr:   "#6d393d",
+	cardPerm:  "#5b6028",
+}
+
+// rosePinePalette — Rosé Pine (main): muted rose-quartz base with a soft-rose
+// accent and cool pine/foam signals.
+var rosePinePalette = palette{
+	panel:     "#1f1d2e",
+	promptBg:  "#2f2b47",
+	line:      "#2b2840",
+	line2:     "#403d52",
+	ink:       "#e0def4",
+	muted:     "#a8a3c0",
+	faint:     "#928ea9",
+	faintest:  "#8985a0",
+	accent:    "#ebbcba",
+	green:     "#31748f",
+	red:       "#eb6f92",
+	amber:     "#f6c177",
+	blue:      "#9ccfd8",
+	gitAdd:    "#5e88a2",
+	gitDel:    "#d589a7",
+	addBg:     "#1f2d3a",
+	delBg:     "#3a1f2d",
+	addBgWord: "#274b5e",
+	delBgWord: "#763a4f",
+	permBg:    "#4a3e3d",
+	selBg:     "#44415a",
+	addInk:    "#cfe8ef",
+	delInk:    "#f6c9cd",
+	onAccent:  "#000000",
+	cardRun:   "#7c6673",
+	cardErr:   "#7c4662",
+	cardPerm:  "#806857",
+}
+
+// everforestPalette — Everforest dark (medium): warm forest-gray surface with a
+// sage-green accent.
+var everforestPalette = palette{
+	panel:     "#333c43",
+	promptBg:  "#3d484d",
+	line:      "#414b52",
+	line2:     "#55636b",
+	ink:       "#d3c6aa",
+	muted:     "#b0bab0",
+	faint:     "#a4aea3",
+	faintest:  "#9ca99b",
+	accent:    "#a7c080",
+	green:     "#83c092",
+	red:       "#e67e80",
+	amber:     "#dbbc7f",
+	blue:      "#7fbbb3",
+	gitAdd:    "#8faa78",
+	gitDel:    "#c08888",
+	addBg:     "#2c3f37",
+	delBg:     "#3e3234",
+	addBgWord: "#324e3b",
+	delBgWord: "#573a3c",
+	permBg:    "#3c382d",
+	selBg:     "#3b482e",
+	addInk:    "#cfead0",
+	delInk:    "#f4cfcd",
+	onAccent:  "#000000",
+	cardRun:   "#798b6b",
+	cardErr:   "#9c676b",
+	cardPerm:  "#96896b",
+}
+
+// lightPalette is dark-on-light: a warm cream surface (so cards lift off the
+// terminal page, which Zero never paints) with near-black ink and an olive-lime
+// accent that keeps the brand identity while clearing AA on the light panel. The
+// muted/faint/faintest grays get progressively LIGHTER toward the surface; diff and
+// permission tints are light surfaces. Replaces the old flat cool-gray light set
+// whose muddy panel, sub-AA diff spans, and near-invisible selBg read as broken.
+var lightPalette = palette{
+	panel:     "#efebd4",
+	promptBg:  "#e3ddc2",
+	line:      "#d8d2bd",
+	line2:     "#b7b199",
+	ink:       "#22201a",
+	muted:     "#4b5149",
+	faint:     "#575e55",
+	faintest:  "#636a61",
+	accent:    "#54700a",
+	green:     "#1e725c",
+	red:       "#c02434",
+	amber:     "#8a5f00",
+	blue:      "#1f66c0",
+	gitAdd:    "#3d6f46",
+	gitDel:    "#a34a4a",
+	addBg:     "#ddf0df",
+	delBg:     "#f8dcdc",
+	addBgWord: "#a2daae",
+	delBgWord: "#f2b6b6",
+	permBg:    "#f7ebc6",
+	selBg:     "#d4e08f",
+	addInk:    "#0c4026",
+	delInk:    "#641a1d",
+	onAccent:  "#ffffff",
+	cardRun:   "#b0be7e",
+	cardErr:   "#d8b0a8",
+	cardPerm:  "#d6c496",
+}
+
+// solarizedLightPalette — Solarized Light: the base3/base2 cream surface with the
+// same fixed accent wheel as Solarized Dark, dark-on-light.
+var solarizedLightPalette = palette{
+	panel:     "#eee8d5",
+	promptBg:  "#e1d9be",
+	line:      "#d8d1bc",
+	line2:     "#c0b89e",
+	ink:       "#304049",
+	muted:     "#495b61",
+	faint:     "#506469",
+	faintest:  "#576b72",
+	accent:    "#0c665c", // solarized cyan, darkened for AA on the cream panel (with white onAccent)
+	green:     "#859900",
+	red:       "#dc322f",
+	amber:     "#7a5c00", // solarized yellow, darkened so white onAccent works on amber fills
+	blue:      "#268bd2",
+	gitAdd:    "#788d34",
+	gitDel:    "#ac4f50",
+	addBg:     "#dde8c6",
+	delBg:     "#f1ddd2",
+	addBgWord: "#b9d488", // deepened so the changed span separates from addBg (sep 1.28)
+	delBgWord: "#edc4b4",
+	permBg:    "#f0e6bd",
+	selBg:     "#a6d6c4",
+	addInk:    "#38480a",
+	delInk:    "#6f1614",
+	onAccent:  "#ffffff", // white — accent and amber are both dark on this light theme
+	cardRun:   "#7fbaaf",
+	cardErr:   "#d8837a",
+	cardPerm:  "#c4ae63",
+}
+
+// themeEntry is one registered theme: Name is the /theme value + ZERO_THEME/--theme
+// token (lowercase, kebab), Label is the picker display text, and IsDark groups the
+// picker (Dark/Light sections) and drives which built-in `auto` resolves to.
+type themeEntry struct {
+	Name    string
+	Label   string
+	Palette palette
+	IsDark  bool
+}
+
+// themeRegistry is the ordered source of truth for every selectable theme. Order is
+// the picker order: all Dark themes first, then all Light, with the brand
+// dark/light built-ins leading their groups. themeModes (theme_select.go) prepends
+// `auto` to this. Append here to add a theme — nothing else needs editing.
+var themeRegistry = []themeEntry{
+	{Name: "dark", Label: "dark", Palette: darkPalette, IsDark: true},
+	{Name: "dracula", Label: "Dracula", Palette: draculaPalette, IsDark: true},
+	{Name: "nord", Label: "Nord", Palette: nordPalette, IsDark: true},
+	{Name: "gruvbox", Label: "Gruvbox", Palette: gruvboxPalette, IsDark: true},
+	{Name: "tokyo-night", Label: "Tokyo Night", Palette: tokyoNightPalette, IsDark: true},
+	{Name: "catppuccin", Label: "Catppuccin", Palette: catppuccinPalette, IsDark: true},
+	{Name: "one-dark", Label: "One Dark", Palette: oneDarkPalette, IsDark: true},
+	{Name: "solarized-dark", Label: "Solarized Dark", Palette: solarizedDarkPalette, IsDark: true},
+	{Name: "rose-pine", Label: "Rosé Pine", Palette: rosePinePalette, IsDark: true},
+	{Name: "everforest", Label: "Everforest", Palette: everforestPalette, IsDark: true},
+	{Name: "light", Label: "light", Palette: lightPalette, IsDark: false},
+	{Name: "solarized-light", Label: "Solarized Light", Palette: solarizedLightPalette, IsDark: false},
+}
+
+// themeByName indexes the registry by lowercased name for O(1) lookup. Built as a
+// var initializer (not init()) so it is ready before themeModes' package-var
+// initializer calls themeNames().
+var themeByName = func() map[string]themeEntry {
+	byName := make(map[string]themeEntry, len(themeRegistry))
+	for _, entry := range themeRegistry {
+		byName[entry.Name] = entry
+	}
+	return byName
+}()
+
+// lookupTheme resolves a theme name (case/space-insensitive) to its entry.
+func lookupTheme(name string) (themeEntry, bool) {
+	entry, ok := themeByName[strings.ToLower(strings.TrimSpace(name))]
+	return entry, ok
+}
+
+// themeNames returns every registered theme name in registry (picker) order.
+func themeNames() []string {
+	names := make([]string, len(themeRegistry))
+	for index, entry := range themeRegistry {
+		names[index] = entry.Name
+	}
+	return names
+}

internal/tui/theme_picker_test.go

@@ -0,0 +1,340 @@
+package tui
+
+import (
+	"context"
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"reflect"
+	"testing"
+
+	tea "charm.land/bubbletea/v2"
+)
+
+// A committed theme is written to user config and reloaded at startup (via
+// Options.SavedTheme -> resolveThemeMode), so a /theme choice survives restart.
+func TestThemeChoicePersistsAcrossRestart(t *testing.T) {
+	defer applyTheme(themeDark, true)
+	cfgPath := filepath.Join(t.TempDir(), "config.json")
+
+	// First session: pick a color theme via the /theme command (same commit path the
+	// picker uses).
+	m := newModel(context.Background(), Options{UserConfigPath: cfgPath})
+	m, _ = m.handleThemeCommand("dracula")
+	if m.themeMode != themeMode("dracula") {
+		t.Fatalf("themeMode = %q, want dracula", m.themeMode)
+	}
+	data, err := os.ReadFile(cfgPath)
+	if err != nil {
+		t.Fatalf("theme commit should have written config: %v", err)
+	}
+	var cfg struct {
+		Preferences struct {
+			Theme string `json:"theme"`
+		} `json:"preferences"`
+	}
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		t.Fatalf("config is not valid JSON: %v", err)
+	}
+	if cfg.Preferences.Theme != "dracula" {
+		t.Fatalf("preferences.theme = %q, want dracula", cfg.Preferences.Theme)
+	}
+
+	// Second session: the persisted theme seeds startup (no flag / env).
+	restarted := newModel(context.Background(), Options{UserConfigPath: cfgPath, SavedTheme: "dracula"})
+	if restarted.themeMode != themeMode("dracula") {
+		t.Fatalf("restarted themeMode = %q, want dracula (from saved config)", restarted.themeMode)
+	}
+}
+
+// resolveThemeMode precedence: flag > ZERO_THEME > persisted config > auto.
+func TestResolveThemeModeConfigFallback(t *testing.T) {
+	if got := resolveThemeMode("", "", "nord"); got != themeMode("nord") {
+		t.Errorf("saved-only = %q, want nord", got)
+	}
+	if got := resolveThemeMode("dracula", "", "nord"); got != themeMode("dracula") {
+		t.Errorf("flag should beat saved: got %q, want dracula", got)
+	}
+	if got := resolveThemeMode("", "gruvbox", "nord"); got != themeMode("gruvbox") {
+		t.Errorf("env should beat saved: got %q, want gruvbox", got)
+	}
+	if got := resolveThemeMode("", "", "bogus-theme"); got != themeAuto {
+		t.Errorf("unknown saved theme should fall back to auto, got %q", got)
+	}
+}
+
+// themePickerRowIndex finds the picker row whose Value is name (index-independent,
+// so tests survive registry reordering / additions).
+func themePickerRowIndex(t *testing.T, p *commandPicker, name string) int {
+	t.Helper()
+	for i, item := range p.items {
+		if item.Value == name {
+			return i
+		}
+	}
+	t.Fatalf("theme picker has no row for %q", name)
+	return -1
+}
+
+// assertPreviewMatchesSelection checks the live palette equals the palette of the
+// currently highlighted theme row (the auto row has no palette of its own).
+func assertPreviewMatchesSelection(t *testing.T, m model) {
+	t.Helper()
+	sel := m.picker.items[m.picker.selected]
+	entry, ok := lookupTheme(sel.Value)
+	if !ok {
+		return
+	}
+	if r, _, _, _ := zeroTheme.inkColor.RGBA(); r != mustR(t, entry.Palette.ink) {
+		t.Errorf("preview of %q did not apply its palette (ink mismatch)", sel.Value)
+	}
+}
+
+// Bare `/theme` opens the popup picker (like /model and /effort) with one row per
+// theme mode, preselecting the active preference. An explicit `/theme <mode>` must
+// keep taking the text path so scripts and the existing state view still work.
+func TestThemePickerOpensOnBareTheme(t *testing.T) {
+	defer applyTheme(themeDark, true)
+	m := newModel(context.Background(), Options{})
+	m.themeMode = themeLight
+	m.input.SetValue("/theme")
+
+	updated, cmd := m.Update(testKey(tea.KeyEnter))
+	m = updated.(model)
+	if cmd != nil {
+		t.Fatalf("opening the theme picker should not emit a cmd, got %T", cmd)
+	}
+	if m.picker == nil || m.picker.kind != pickerTheme {
+		t.Fatalf("expected the theme picker to open, got %#v", m.picker)
+	}
+	if len(m.picker.items) != len(themeModes) {
+		t.Fatalf("picker has %d items, want %d", len(m.picker.items), len(themeModes))
+	}
+	for i, want := range themeModes {
+		if m.picker.items[i].Value != want {
+			t.Errorf("item %d value = %q, want %q", i, m.picker.items[i].Value, want)
+		}
+	}
+	if got := m.picker.items[m.picker.selected].Value; got != string(themeLight) {
+		t.Errorf("preselected value = %q, want the active mode %q", got, themeLight)
+	}
+}
+
+// An explicit mode argument runs the text handler directly and never opens the
+// popup — guards the dispatch branch and keeps /theme <mode> scriptable.
+func TestThemeArgSkipsPicker(t *testing.T) {
+	defer applyTheme(themeDark, true)
+	m := newModel(context.Background(), Options{})
+	m.input.SetValue("/theme dark")
+
+	updated, _ := m.Update(testKey(tea.KeyEnter))
+	m = updated.(model)
+	if m.picker != nil {
+		t.Fatalf("explicit /theme dark must not open a picker, got %#v", m.picker)
+	}
+	if m.themeMode != themeDark {
+		t.Fatalf("after /theme dark, mode = %q", m.themeMode)
+	}
+}
+
+// Moving the cursor live-previews each palette: the global zeroTheme swaps as the
+// selection changes, so the whole UI (and the overlay) repaints in the hovered
+// theme without committing the preference.
+func TestThemePickerPreviewsOnMove(t *testing.T) {
+	defer applyTheme(themeDark, true)
+	m := newModel(context.Background(), Options{})
+	m.themeMode = themeDark
+	m.hasDarkBg = true
+	applyTheme(themeDark, true)
+	m.input.SetValue("/theme")
+	updated, _ := m.Update(testKey(tea.KeyEnter))
+	m = updated.(model)
+
+	// Each Down moves to the next theme and live-applies its palette; Up follows back.
+	for i := 0; i < 3; i++ {
+		updated, _ = m.Update(testKey(tea.KeyDown))
+		m = updated.(model)
+		assertPreviewMatchesSelection(t, m)
+	}
+	updated, _ = m.Update(testKey(tea.KeyUp))
+	m = updated.(model)
+	assertPreviewMatchesSelection(t, m)
+
+	// Preview must not have written the committed preference.
+	if m.themeMode != themeDark {
+		t.Errorf("preview mutated the committed mode to %q", m.themeMode)
+	}
+}
+
+// Enter on a fixed mode commits it: the palette stays applied, the preference is
+// recorded, the popup closes, and the switch is noted in the transcript.
+func TestThemePickerCommitAppliesAndRecords(t *testing.T) {
+	defer applyTheme(themeDark, true)
+	m := newModel(context.Background(), Options{})
+	m.themeMode = themeDark
+	m.hasDarkBg = true
+	applyTheme(themeDark, true)
+	m.input.SetValue("/theme")
+	updated, _ := m.Update(testKey(tea.KeyEnter))
+	m = updated.(model)
+
+	// Choose a color theme by locating its row (index-independent), then confirm.
+	m.picker.selected = themePickerRowIndex(t, m.picker, "dracula")
+	updated, cmd := m.Update(testKey(tea.KeyEnter))
+	m = updated.(model)
+	if cmd != nil {
+		t.Fatalf("committing a fixed theme should not emit a cmd, got %T", cmd)
+	}
+	if m.picker != nil {
+		t.Fatal("committing should close the picker")
+	}
+	if m.themeMode != themeMode("dracula") {
+		t.Fatalf("committed mode = %q, want dracula", m.themeMode)
+	}
+	if r, _, _, _ := zeroTheme.inkColor.RGBA(); r != mustR(t, draculaPalette.ink) {
+		t.Error("commit did not leave the dracula palette applied")
+	}
+	if !transcriptContains(m.transcript, "active theme") {
+		t.Errorf("commit should record the switch in the transcript:\n%s", transcriptText(m.transcript))
+	}
+}
+
+// Committing `auto` re-probes the terminal background (like the text dispatch) so
+// the palette re-detects light/dark instead of reusing the preview's reading.
+func TestThemePickerAutoCommitReprobesBackground(t *testing.T) {
+	defer applyTheme(themeDark, true)
+	m := newModel(context.Background(), Options{})
+	m.themeMode = themeDark
+	m.input.SetValue("/theme")
+	updated, _ := m.Update(testKey(tea.KeyEnter))
+	m = updated.(model)
+
+	m.picker.selected = 0 // "auto"
+	updated, cmd := m.Update(testKey(tea.KeyEnter))
+	m = updated.(model)
+	if m.themeMode != themeAuto {
+		t.Fatalf("committed mode = %q, want auto", m.themeMode)
+	}
+	if cmd == nil {
+		t.Fatal("committing auto must return a background re-probe cmd")
+	}
+	if reflect.ValueOf(cmd).Pointer() != reflect.ValueOf(tea.RequestBackgroundColor).Pointer() {
+		t.Error("committing auto must return tea.RequestBackgroundColor")
+	}
+}
+
+// Typing to filter the list re-previews the newly-highlighted mode, so the applied
+// palette never diverges from the row the popup points at. Backspacing back to an
+// empty query re-previews the (restored) selection too.
+func TestThemePickerFilterRePreviews(t *testing.T) {
+	defer applyTheme(themeDark, true)
+	m := newModel(context.Background(), Options{})
+	m.themeMode = themeDark
+	m.hasDarkBg = true
+	applyTheme(themeDark, true)
+	m.input.SetValue("/theme")
+	updated, _ := m.Update(testKey(tea.KeyEnter))
+	m = updated.(model)
+
+	// Preview light by moving, then filter to "light" — the applied palette must
+	// track the filtered highlight (still light here), not lag behind.
+	updated, _ = m.Update(testKey(tea.KeyDown)) // dark -> light preview
+	m = updated.(model)
+	updated, _ = m.Update(testKeyText("l"))
+	m = updated.(model)
+	updated, _ = m.Update(testKeyText("i"))
+	m = updated.(model)
+	if got := m.picker.items[m.picker.selected].Value; got != string(themeLight) {
+		t.Fatalf("filter 'li' should highlight light, got %q", got)
+	}
+	if r, _, _, _ := zeroTheme.inkColor.RGBA(); r != mustR(t, lightPalette.ink) {
+		t.Error("filtering to light did not keep the light palette previewed")
+	}
+
+	// Re-filter to "dark": highlight and preview must both switch to dark.
+	updated, _ = m.Update(testKey(tea.KeyBackspace))
+	m = updated.(model)
+	updated, _ = m.Update(testKey(tea.KeyBackspace))
+	m = updated.(model)
+	updated, _ = m.Update(testKeyText("d"))
+	m = updated.(model)
+	if got := m.picker.items[m.picker.selected].Value; got != string(themeDark) {
+		t.Fatalf("filter 'd' should highlight dark, got %q", got)
+	}
+	if r, _, _, _ := zeroTheme.inkColor.RGBA(); r != mustR(t, darkPalette.ink) {
+		t.Error("filtering to dark did not re-preview the dark palette")
+	}
+}
+
+// A filter that matches nothing must not strand the last preview: the palette
+// falls back to the committed theme, and committing (Enter on the empty list) is a
+// clean no-op that leaves the committed mode and palette in agreement.
+func TestThemePickerEmptyFilterRestoresCommitted(t *testing.T) {
+	defer applyTheme(themeDark, true)
+	m := newModel(context.Background(), Options{})
+	m.themeMode = themeDark
+	m.hasDarkBg = true
+	applyTheme(themeDark, true)
+	m.input.SetValue("/theme")
+	updated, _ := m.Update(testKey(tea.KeyEnter))
+	m = updated.(model)
+
+	// Preview another theme by moving, then type a char that matches no theme name
+	// ('q' appears in none) -> the list empties.
+	updated, _ = m.Update(testKey(tea.KeyDown))
+	m = updated.(model)
+	updated, _ = m.Update(testKeyText("q"))
+	m = updated.(model)
+	if len(m.picker.items) != 0 {
+		t.Fatalf("filter 'q' should match no theme, got %d items", len(m.picker.items))
+	}
+	if r, _, _, _ := zeroTheme.inkColor.RGBA(); r != mustR(t, darkPalette.ink) {
+		t.Error("an empty filter should restore the committed dark palette, not keep the preview")
+	}
+
+	// Enter on the empty list closes the picker without changing the committed
+	// preference, and leaves the palette matching it.
+	updated, _ = m.Update(testKey(tea.KeyEnter))
+	m = updated.(model)
+	if m.picker != nil {
+		t.Fatal("Enter on an empty filter should close the picker")
+	}
+	if m.themeMode != themeDark {
+		t.Fatalf("empty-list commit must not change the mode, got %q", m.themeMode)
+	}
+	if r, _, _, _ := zeroTheme.inkColor.RGBA(); r != mustR(t, darkPalette.ink) {
+		t.Error("after an empty-list commit the palette must match the committed dark mode")
+	}
+}
+
+// Esc dismisses the picker without choosing and restores the committed palette,
+// undoing whatever the live preview applied.
+func TestThemePickerEscRestoresCommittedTheme(t *testing.T) {
+	defer applyTheme(themeDark, true)
+	m := newModel(context.Background(), Options{})
+	m.themeMode = themeDark
+	m.hasDarkBg = true
+	applyTheme(themeDark, true)
+	m.input.SetValue("/theme")
+	updated, _ := m.Update(testKey(tea.KeyEnter))
+	m = updated.(model)
+
+	// Preview another theme by moving, then bail out.
+	updated, _ = m.Update(testKey(tea.KeyDown))
+	m = updated.(model)
+	assertPreviewMatchesSelection(t, m)
+	if r, _, _, _ := zeroTheme.inkColor.RGBA(); r == mustR(t, darkPalette.ink) {
+		t.Fatal("expected a non-dark preview to be applied before Esc")
+	}
+	updated, _ = m.Update(testKey(tea.KeyEsc))
+	m = updated.(model)
+	if m.picker != nil {
+		t.Fatal("Esc should close the picker")
+	}
+	if m.themeMode != themeDark {
+		t.Fatalf("Esc must not change the committed mode, got %q", m.themeMode)
+	}
+	if r, _, _, _ := zeroTheme.inkColor.RGBA(); r != mustR(t, darkPalette.ink) {
+		t.Error("Esc did not restore the committed dark palette")
+	}
+}

internal/tui/theme_select.go

@@ -1,6 +1,10 @@
 package tui
 
-import "strings"
+import (
+	"strings"
+
+	"github.com/Gitlawb/zero/internal/config"
+)
 
 // themeMode is the operator's palette preference.
 type themeMode string
@@ -11,34 +15,49 @@ const (
 	themeLight themeMode = "light"
 )
 
-// themeModes lists the values /theme accepts.
-var themeModes = []string{string(themeAuto), string(themeDark), string(themeLight)}
+// themeModes lists the values /theme accepts, in picker order: `auto` first, then
+// every registered theme (theme_palettes.go). It is the single ordered source
+// feeding both the picker and the /theme state list — adding a registry entry
+// extends it automatically.
+var themeModes = append([]string{string(themeAuto)}, themeNames()...)
 
-// resolveThemeMode picks the preference in precedence order: an explicit value
-// (the --theme flag, threaded via Options.Theme), then ZERO_THEME, then auto.
-func resolveThemeMode(flagValue, envValue string) themeMode {
-	for _, v := range []string{flagValue, envValue} {
-		switch strings.ToLower(strings.TrimSpace(v)) {
-		case "dark":
-			return themeDark
-		case "light":
-			return themeLight
-		case "auto":
+// resolveThemeMode picks the first accepted preference from candidates in
+// precedence order — the caller passes them highest-first: the --theme flag, then
+// ZERO_THEME, then the persisted config theme. A value is accepted if it is `auto`
+// or names a registered theme; unrecognized/blank values are skipped, and an empty
+// list (or all-unrecognized) falls back to auto.
+func resolveThemeMode(candidates ...string) themeMode {
+	for _, v := range candidates {
+		s := strings.ToLower(strings.TrimSpace(v))
+		if s == "" {
+			continue
+		}
+		if s == string(themeAuto) {
 			return themeAuto
 		}
+		if _, ok := lookupTheme(s); ok {
+			return themeMode(s)
+		}
 	}
 	return themeAuto
 }
 
-// validThemeMode reports whether s names a theme mode (for /theme validation).
+// validThemeMode reports whether s names a theme mode (for /theme validation):
+// `auto` or any registered theme.
 func validThemeMode(s string) bool {
-	switch themeMode(strings.ToLower(strings.TrimSpace(s))) {
-	case themeAuto, themeDark, themeLight:
+	s = strings.ToLower(strings.TrimSpace(s))
+	if s == string(themeAuto) {
 		return true
 	}
-	return false
+	_, ok := lookupTheme(s)
+	return ok
 }
 
+// ValidThemeArg reports whether s is an acceptable --theme / ZERO_THEME value
+// (`auto` or a registered theme name). Exported so the CLI flag validator shares
+// this one source of truth instead of hardcoding the theme list.
+func ValidThemeArg(s string) bool { return validThemeMode(s) }
+
 // applyTheme swaps the active palette (zeroTheme) and the globals derived from it
 // — the streaming-fade ramp and the static render cache — so a switch repaints
 // every subsequent render. For themeAuto it resolves to dark/light from
@@ -53,12 +72,13 @@ func applyTheme(mode themeMode, hasDarkBackground bool) themeMode {
 			resolved = themeLight
 		}
 	}
-	switch resolved {
-	case themeLight:
-		zeroTheme = buildTheme(lightPalette)
-	default:
-		zeroTheme = buildTheme(darkPalette)
+	// Resolve the (now concrete) mode to its registered palette; an unknown name
+	// falls back to the dark built-in so a bad value can never leave zeroTheme unset.
+	entry, ok := lookupTheme(string(resolved))
+	if !ok {
+		entry, _ = lookupTheme(string(themeDark))
 	}
+	zeroTheme = buildTheme(entry.Palette)
 	rebuildStreamingFadePalette()
 	if defaultRenderCache != nil {
 		defaultRenderCache.clear() // old-palette entries must not be reused
@@ -66,15 +86,45 @@ func applyTheme(mode themeMode, hasDarkBackground bool) themeMode {
 	return resolved
 }
 
-// handleThemeCommand implements /theme [auto|dark|light]: no arg shows state, a
-// valid mode switches the active palette live. Mirrors handleStyleCommand.
+// previewSelectedTheme makes the live palette match the theme picker's current
+// state: the highlighted mode when a row is selectable, or the committed
+// m.themeMode when the filter matches nothing. Called on every change to the
+// picker's selection or filter (arrow/wheel moves, mouse, and query typing) so the
+// whole UI — and the overlay itself — always renders the mode the popup points at,
+// and never strands on a stale preview. It only swaps the global zeroTheme via
+// applyTheme; m.themeMode keeps the committed preference so Esc can restore it. A
+// no-op unless a theme picker is open. Runs on the Update goroutine, like every
+// zeroTheme access.
+func (m model) previewSelectedTheme() {
+	if m.picker == nil || m.picker.kind != pickerTheme {
+		return
+	}
+	if item, ok := m.picker.current(); ok {
+		applyTheme(themeMode(item.Value), m.hasDarkBg)
+		return
+	}
+	// No selectable row (the filter matched nothing): fall back to the committed
+	// theme rather than leaving the previous preview applied.
+	m.restoreCommittedTheme()
+}
+
+// restoreCommittedTheme re-applies m.themeMode after a /theme preview is dismissed
+// without choosing (Esc). Preview never wrote m.themeMode, so it still holds the
+// real preference; re-applying repaints back to it.
+func (m model) restoreCommittedTheme() {
+	applyTheme(m.themeMode, m.hasDarkBg)
+}
+
+// handleThemeCommand implements /theme [name]: `list` shows state, a registered
+// theme name (or `auto`) switches the active palette live. Bare `/theme` opens the
+// picker at the dispatch layer, so it never reaches here empty. Mirrors handleStyleCommand.
 func (m model) handleThemeCommand(args string) (model, string) {
 	arg := strings.ToLower(strings.TrimSpace(args))
 	if arg == "" || arg == "list" {
 		return m, m.themeStateText()
 	}
 	if !validThemeMode(arg) {
-		return m, "Theme\nUnknown theme: " + arg + " (use auto, dark, or light)"
+		return m, "Theme\nUnknown theme: " + arg + " (use /theme with no argument to pick from the list)"
 	}
 	m.themeMode = themeMode(arg)
 	resolved := applyTheme(m.themeMode, m.hasDarkBg)
@@ -82,11 +132,31 @@ func (m model) handleThemeCommand(args string) (model, string) {
 	if m.themeMode == themeAuto {
 		active = "auto (" + string(resolved) + ")"
 	}
-	return m, strings.Join([]string{
+	lines := []string{
 		"Theme",
 		"active theme: " + active,
 		"Already-printed scrollback keeps its previous colors; new output uses the new theme.",
-	}, "\n")
+	}
+	// Commit path (both /theme <name> and the picker route through here): persist the
+	// choice so it survives restart. Previews call applyTheme directly and never reach here.
+	if note := m.persistThemePreference(); note != "" {
+		lines = append(lines, note)
+	}
+	return m, strings.Join(lines, "\n")
+}
+
+// persistThemePreference writes the committed theme to user config so it is applied
+// again at startup (via Options.SavedTheme -> resolveThemeMode). Best-effort: returns
+// a short note to surface on failure, or "" on success / when there is no config
+// path (e.g. tests). Never called from the live-preview path.
+func (m model) persistThemePreference() string {
+	if strings.TrimSpace(m.userConfigPath) == "" {
+		return ""
+	}
+	if _, err := config.SetTheme(m.userConfigPath, string(m.themeMode)); err != nil {
+		return "note: could not save theme preference (" + err.Error() + ")"
+	}
+	return ""
 }
 
 // themeStateText renders the /theme state view.
@@ -109,6 +179,6 @@ func (m model) themeStateText() string {
 				"available: " + strings.Join(themeModes, ", "),
 			},
 		}},
-		Hints: []string{"use /theme <auto|dark|light> to switch this TUI session"},
+		Hints: []string{"run /theme with no argument to open the picker, or /theme <name> to switch directly"},
 	})
 }

internal/tui/theme_select_test.go

@@ -52,24 +52,19 @@ func wcagRatio(t *testing.T, fg, bg string) float64 {
 // The word-level diff's brighter changed-span band must keep its text AA-readable
 // and stay clearly distinct from the base add/del band, on both themes.
 func TestDiffWordSpanContrast(t *testing.T) {
-	for _, c := range []struct {
-		name string
-		pal  palette
-	}{
-		{"dark", darkPalette},
-		{"light", lightPalette},
-	} {
-		if r := wcagRatio(t, c.pal.addInk, c.pal.addBgWord); r < 4.5 {
-			t.Errorf("%s: addInk on addBgWord %.2f < 4.5 (AA)", c.name, r)
+	for _, entry := range themeRegistry {
+		name, pal := entry.Name, entry.Palette
+		if r := wcagRatio(t, pal.addInk, pal.addBgWord); r < 4.5 {
+			t.Errorf("%s: addInk on addBgWord %.2f < 4.5 (AA)", name, r)
 		}
-		if r := wcagRatio(t, c.pal.delInk, c.pal.delBgWord); r < 4.5 {
-			t.Errorf("%s: delInk on delBgWord %.2f < 4.5 (AA)", c.name, r)
+		if r := wcagRatio(t, pal.delInk, pal.delBgWord); r < 4.5 {
+			t.Errorf("%s: delInk on delBgWord %.2f < 4.5 (AA)", name, r)
 		}
-		if sep := wcagRatio(t, c.pal.addBgWord, c.pal.addBg); sep < 1.2 {
-			t.Errorf("%s: addBgWord vs addBg separation %.2f < 1.2 (span not distinct)", c.name, sep)
+		if sep := wcagRatio(t, pal.addBgWord, pal.addBg); sep < 1.2 {
+			t.Errorf("%s: addBgWord vs addBg separation %.2f < 1.2 (span not distinct)", name, sep)
 		}
-		if sep := wcagRatio(t, c.pal.delBgWord, c.pal.delBg); sep < 1.2 {
-			t.Errorf("%s: delBgWord vs delBg separation %.2f < 1.2 (span not distinct)", c.name, sep)
+		if sep := wcagRatio(t, pal.delBgWord, pal.delBg); sep < 1.2 {
+			t.Errorf("%s: delBgWord vs delBg separation %.2f < 1.2 (span not distinct)", name, sep)
 		}
 	}
 }
@@ -78,18 +73,52 @@ func TestDiffWordSpanContrast(t *testing.T) {
 // AND keep its label readable. Guards the regression this fixes: the light
 // selBg (#e7f2cd) sat at 1.01 vs the panel (#ececed) — effectively invisible.
 func TestSelectedRowBandIsVisibleAndReadable(t *testing.T) {
-	for _, c := range []struct {
-		name string
-		pal  palette
-	}{
-		{"dark", darkPalette},
-		{"light", lightPalette},
-	} {
-		if r := wcagRatio(t, c.pal.ink, c.pal.selBg); r < 4.5 {
-			t.Errorf("%s: ink on selBg contrast %.2f < 4.5 — selected-row label unreadable", c.name, r)
+	for _, entry := range themeRegistry {
+		name, pal := entry.Name, entry.Palette
+		if r := wcagRatio(t, pal.ink, pal.selBg); r < 4.5 {
+			t.Errorf("%s: ink on selBg contrast %.2f < 4.5 — selected-row label unreadable", name, r)
 		}
-		if sep := wcagRatio(t, c.pal.selBg, c.pal.panel); sep < 1.10 {
-			t.Errorf("%s: selBg vs panel separation %.2f < 1.10 — selected row does not stand out", c.name, sep)
+		if sep := wcagRatio(t, pal.selBg, pal.panel); sep < 1.10 {
+			t.Errorf("%s: selBg vs panel separation %.2f < 1.10 — selected row does not stand out", name, sep)
+		}
+	}
+}
+
+// Every registered theme — not just the two built-ins — must clear WCAG AA on its
+// text-bearing tokens against the panel and keep the muted>faint>faintest>panel
+// gray ramp monotonic in its polarity (light-on-dark for dark themes, the inverse
+// for light themes). Guards that a newly-added color palette can't ship illegible.
+func TestAllThemesContrastAndHierarchy(t *testing.T) {
+	for _, entry := range themeRegistry {
+		pal := entry.Palette
+		for _, tok := range []struct {
+			name string
+			fg   string
+		}{
+			{"ink", pal.ink}, {"muted", pal.muted}, {"faint", pal.faint},
+			{"faintest", pal.faintest}, {"accent", pal.accent},
+		} {
+			if r := wcagRatio(t, tok.fg, pal.panel); r < 4.5 {
+				t.Errorf("%s %s on panel %.2f < 4.5 (WCAG AA)", entry.Name, tok.name, r)
+			}
+		}
+		if r := wcagRatio(t, pal.onAccent, pal.accent); r < 4.5 {
+			t.Errorf("%s onAccent on accent %.2f < 4.5 (WCAG AA)", entry.Name, r)
+		}
+		// Gray ramp ordered ink -> muted -> faint -> faintest -> panel; luminance
+		// rises toward the surface on light themes, falls on dark themes.
+		chain := []float64{
+			relLum(t, pal.ink), relLum(t, pal.muted), relLum(t, pal.faint),
+			relLum(t, pal.faintest), relLum(t, pal.panel),
+		}
+		for i := 1; i < len(chain); i++ {
+			ok := chain[i] > chain[i-1]
+			if entry.IsDark {
+				ok = chain[i] < chain[i-1]
+			}
+			if !ok {
+				t.Errorf("%s hierarchy not monotonic toward surface at %d: %v", entry.Name, i, chain)
+			}
 		}
 	}
 }

internal/tui/transcript_selection.go

@@ -22,8 +22,20 @@ type transcriptSelectionPoint struct {
 
 type transcriptSelectionState struct {
 	active bool
-	anchor transcriptSelectionPoint
-	cursor transcriptSelectionPoint
+	// dragging is true from press until release — narrower than active, which
+	// stays true through the async copy-command grace window after release (a
+	// non-empty selection isn't reset to {} until transcriptCopiedMsg lands).
+	// mouseMotion gates on THIS, not active, so a genuine hover motion arriving
+	// in that post-release window doesn't get misrouted as a drag continuation.
+	// It also deliberately does NOT require msg's own Button field to be
+	// non-None on each motion event: some terminals don't restate the button on
+	// every motion report during a real drag (see
+	// TestTranscriptSelectionUpdatesOnGenericMotion), so trusting the
+	// per-event button field would break those terminals — dragging tracks the
+	// app's own press/release bracket instead.
+	dragging bool
+	anchor   transcriptSelectionPoint
+	cursor   transcriptSelectionPoint
 }
 
 type transcriptSelectableLine struct {
@@ -159,10 +171,45 @@ func (m model) finalizeTranscriptBodyRow(rendered string, selectable []transcrip
 	shifted := shiftSelectableX(selectable, gutter)
 	if m.transcriptSelection.active {
 		lines = viewLines(m.renderRenderedSelection(strings.Join(lines, "\n"), shifted, startBodyY))
+	} else if m.hover.kind == hoverTranscript {
+		// Mutually exclusive with the selection highlight above: transcriptSelection
+		// stays active for the lifetime of a persisted (copied) selection, not just
+		// during the drag — so hover correctly stays suppressed over old selected
+		// text too, not only mid-drag.
+		lines = viewLines(m.renderHoverHighlight(strings.Join(lines, "\n"), shifted, startBodyY))
 	}
 	return transcriptBodyRenderedItem{lines: lines, selectable: shifted}
 }
 
+// renderHoverHighlight re-styles the WHOLE line at m.hover.bodyY in the hover
+// accent (see renderRenderedSelection for the sibling selection-highlight logic,
+// which this mirrors). Only a clickable line (a specialist card or a
+// collapse/expand toggle header) is eligible — if the hovered bodyY no longer
+// matches one (content shifted beneath a stale hover target), this is a no-op,
+// which self-heals a hover left over from before a transcript update.
+func (m model) renderHoverHighlight(rendered string, selectable []transcriptSelectableLine, startBodyY int) string {
+	index := m.hover.bodyY - startBodyY
+	if index < 0 {
+		return rendered
+	}
+	lines := viewLines(rendered)
+	if index >= len(lines) {
+		return rendered
+	}
+	matched := false
+	for _, line := range selectable {
+		if line.bodyY == m.hover.bodyY && (line.specialistCard || line.toggle) {
+			matched = true
+			break
+		}
+	}
+	if !matched {
+		return rendered
+	}
+	lines[index] = zeroTheme.hover.Render(ansi.Strip(lines[index]))
+	return strings.Join(lines, "\n")
+}
+
 func (m model) transcriptBodyItems(width int, emptyOverlay string) []transcriptBodyItem {
 	items := []transcriptBodyItem{}
 	// Transcript ROWS render at the full chat width; row/status glyphs provide
@@ -968,16 +1015,46 @@ func splitPlainAtDisplayWidth(text string, width int) (string, string) {
 	return text, ""
 }
 
+// transcriptHitTestSource returns the header/body-items/width mouse hit-testing
+// must use to match what's actually on screen. It mirrors transcriptView's own
+// branching exactly: the subchat drill-in swaps BOTH the header (nav bar instead
+// of the pinned title bar) and the row source (the child session's rows instead
+// of the parent transcript) — and drops the sidebar-reserved width, since subchat
+// is always single-column. Hit-testing against the wrong source is why mouse
+// selection previously resolved against transcript rows that weren't even
+// visible while viewing a subagent/swarm child session.
+func (m model) transcriptHitTestSource() (header string, items []transcriptBodyItem, width int) {
+	if m.subchat.active {
+		width = chatWidth(m.width)
+		return renderSubchatNavBar(m.subchat.childSessionTitle, width), m.transcriptBodyItemsFromRows(m.subchat.childRows, width), width
+	}
+	width = m.chatColumnWidth()
+	return m.pinnedTitleBar(width), m.transcriptBodyItems(width, ""), width
+}
+
+// transcriptHitTestBlocked reports whether mouse hit-testing must be skipped
+// outright — a modal/overlay is up, or there's no alt-screen viewport at all.
+func (m model) transcriptHitTestBlocked() bool {
+	return !m.altScreen || m.height <= 0 || m.setup.visible || m.providerWizard != nil || m.mcpAddWizard != nil || m.mcpManager != nil || m.picker != nil || m.suggestionsActive()
+}
+
+// transcriptHitTestLayout computes the frame/window/layout mouse hit-testing needs,
+// shared by transcriptLineAtMouse (exact match) and nearestTranscriptLineAtMouse
+// (nearest-line fallback for scroll-driven selection extension).
+func (m model) transcriptHitTestLayout() (frame transcriptFrameLayout, window transcriptViewportWindow, layout transcriptBodyLayout) {
+	header, items, width := m.transcriptHitTestSource()
+	frame = m.scrollableTranscriptFrame(header, m.footerView(width))
+	metrics := measureTranscriptBodyItems(items, m.transcriptBodyHeights)
+	window = transcriptViewportForLayout(metrics, frame, m.chatScrollOffset).window()
+	layout = layoutVisibleTranscriptBodyItems(items, metrics, window)
+	return frame, window, layout
+}
+
 func (m model) transcriptLineAtMouse(msg tea.MouseMsg) (transcriptSelectableLine, bool) {
-	if !m.altScreen || m.height <= 0 || m.setup.visible || m.providerWizard != nil || m.mcpAddWizard != nil || m.mcpManager != nil || m.picker != nil || m.suggestionsActive() {
+	if m.transcriptHitTestBlocked() {
 		return transcriptSelectableLine{}, false
 	}
-	width := m.chatColumnWidth()
-	frame := m.scrollableTranscriptFrame(m.pinnedTitleBar(width), m.footerView(width))
-	items := m.transcriptBodyItems(width, "")
-	metrics := measureTranscriptBodyItems(items, m.transcriptBodyHeights)
-	window := transcriptViewportForLayout(metrics, frame, m.chatScrollOffset).window()
-	layout := layoutVisibleTranscriptBodyItems(items, metrics, window)
+	frame, window, layout := m.transcriptHitTestLayout()
 	_, localY, ok := frame.bodyRect.local(mouseX(msg), mouseY(msg))
 	if !ok {
 		return transcriptSelectableLine{}, false
@@ -995,6 +1072,146 @@ func (m model) transcriptLineAtMouse(msg tea.MouseMsg) (transcriptSelectableLine
 	return transcriptSelectableLine{}, false
 }
 
+// nearestTranscriptLineAtMouse is transcriptLineAtMouse with a fallback for
+// scroll-driven selection extension: the recomputed on-screen position can land
+// exactly on a non-selectable spacer row between messages (blank lines aren't in
+// layout.selectable; a scroll can shift which physical row is text vs spacer), where
+// an EXACT match finds nothing even though real text sits one row away. Falls back
+// to the closest visible selectable line by bodyY, so the selection still extends
+// instead of silently freezing for that scroll tick. Only used for scroll
+// extension — a plain click keeps using transcriptLineAtMouse's exact match, since
+// "nothing there" is a meaningful, correct result for an intentional click below
+// the last line or in dead space.
+func (m model) nearestTranscriptLineAtMouse(msg tea.MouseMsg) (transcriptSelectableLine, bool) {
+	if line, ok := m.transcriptLineAtMouse(msg); ok {
+		return line, true
+	}
+	if m.transcriptHitTestBlocked() {
+		return transcriptSelectableLine{}, false
+	}
+	frame, window, layout := m.transcriptHitTestLayout()
+	_, localY, ok := frame.bodyRect.local(mouseX(msg), mouseY(msg))
+	if !ok || mouseX(msg) < 0 {
+		return transcriptSelectableLine{}, false
+	}
+	return nearestTranscriptSelectableAt(layout, window.start+localY)
+}
+
+// nearestTranscriptSelectableAt returns the selectable line in layout closest to
+// targetBodyY (ties go to whichever is found first) — the fallback core shared by
+// nearestTranscriptLineAtMouse (target derived from a mouse position that may
+// have shifted onto a non-selectable spacer row) and the edge-auto-scroll drag
+// case (target derived directly from the new viewport's top/bottom bound, which
+// has no "mouse position" to speak of at all).
+func nearestTranscriptSelectableAt(layout transcriptBodyLayout, targetBodyY int) (transcriptSelectableLine, bool) {
+	best := transcriptSelectableLine{}
+	bestDist := -1
+	found := false
+	for _, line := range layout.selectable {
+		dist := line.bodyY - targetBodyY
+		if dist < 0 {
+			dist = -dist
+		}
+		if !found || dist < bestDist {
+			best, bestDist, found = line, dist, true
+		}
+	}
+	return best, found
+}
+
+// transcriptEdgeScrollDelta reports the scroll delta to apply when a drag has
+// moved past the top or bottom edge of the visible transcript body while
+// staying within its horizontal span — the classic "drag past the viewport
+// edge auto-scrolls" affordance from browsers/editors. Positive reveals OLDER
+// content (dragged above the top edge, same sign convention as wheel-up);
+// negative reveals NEWER content (dragged below the bottom edge). ok is false
+// when the point isn't a vertical-edge case at all (e.g. off to the side, over
+// the sidebar) — that's not an edge-scroll, just outside the column.
+func (m model) transcriptEdgeScrollDelta(msg tea.MouseMsg) (int, bool) {
+	if m.transcriptHitTestBlocked() {
+		return 0, false
+	}
+	frame, _, _ := m.transcriptHitTestLayout()
+	x, y := mouseX(msg), mouseY(msg)
+	if x < frame.bodyRect.x || x >= frame.bodyRect.x+frame.bodyRect.width {
+		return 0, false
+	}
+	if y < frame.bodyRect.y {
+		return chatWheelScrollLines, true
+	}
+	if y >= frame.bodyRect.y+frame.bodyRect.height {
+		return -chatWheelScrollLines, true
+	}
+	return 0, false
+}
+
+// dragToEdgeScroll scrolls one step toward the edge the drag moved past, then
+// extends the selection cursor to whichever line now sits at THAT edge of the
+// new viewport, at column x. It deliberately does NOT try to re-resolve the
+// drag's (still off-screen) physical mouse position — after scrolling, that
+// position is still outside frame.bodyRect (scrolling changes what content
+// occupies a screen row, not the mouse's screen position), so
+// nearestTranscriptLineAtMouse would keep finding nothing. Anchoring to the
+// viewport edge instead is what makes the selection visibly keep pace with the
+// drag while the mouse holds past the edge. Takes a plain column rather than a
+// tea.MouseMsg because the smooth-glide tick chain (dragEdgeScrollTickMsg) calls
+// this repeatedly with no real mouse event of its own — see edgeScrollMouseX.
+func (m model) dragToEdgeScroll(delta int, x int) model {
+	m = m.scrollChat(delta)
+	_, window, layout := m.transcriptHitTestLayout()
+	target := window.start
+	if delta < 0 {
+		target = window.start + window.height - 1
+	}
+	if line, ok := nearestTranscriptSelectableAt(layout, target); ok {
+		m.transcriptSelection.cursor = transcriptSelectionPointForMouse(line, x)
+	}
+	return m
+}
+
+// dragEdgeScrollTickCmd schedules the next step of the smooth-glide edge-scroll,
+// gated by seq (see dragEdgeScrollTickMsg / edgeScrollSeq).
+func dragEdgeScrollTickCmd(seq int) tea.Cmd {
+	return tea.Tick(dragEdgeScrollInterval, func(time.Time) tea.Msg {
+		return dragEdgeScrollTickMsg{seq: seq}
+	})
+}
+
+// startEdgeScroll (re)starts the smooth-glide tick chain in the given direction —
+// a no-op if it's already running the SAME direction (a fresh raw motion event
+// arriving mid-glide must not reset the cadence, or repeated events faster than
+// dragEdgeScrollInterval would make it jerky again, defeating the point). x is
+// remembered for the ticks, which carry no mouse position of their own. Applies
+// one immediate step so there's no perceptible delay between crossing the edge
+// and the first visible movement.
+func (m model) startEdgeScroll(direction int, x int) (model, tea.Cmd) {
+	step := dragEdgeScrollStep
+	if direction < 0 {
+		step = -step
+	}
+	if m.edgeScrollDelta == step {
+		m.edgeScrollMouseX = x
+		return m, nil
+	}
+	m = m.dragToEdgeScroll(step, x)
+	m.edgeScrollDelta = step
+	m.edgeScrollMouseX = x
+	m.edgeScrollSeq++
+	return m, dragEdgeScrollTickCmd(m.edgeScrollSeq)
+}
+
+// stopEdgeScroll ends the smooth-glide tick chain (the drag moved back into the
+// body, off to the side, or released). Bumping edgeScrollSeq invalidates any
+// tick already in flight, so it lands as a no-op and doesn't reschedule itself.
+func (m model) stopEdgeScroll() model {
+	if m.edgeScrollDelta == 0 {
+		return m
+	}
+	m.edgeScrollDelta = 0
+	m.edgeScrollSeq++
+	return m
+}
+
 func (m model) transcriptViewportStart(body string, width int) (int, int, int) {
 	frame := m.scrollableTranscriptFrame(m.pinnedTitleBar(width), m.footerView(width))
 	return transcriptViewportStartForFrame(body, frame, m.chatScrollOffset)
@@ -1029,6 +1246,7 @@ func (m model) handleTranscriptSelectionMouse(msg tea.MouseMsg) (model, tea.Cmd,
 				m = m.appendSystemNotice(errMsg)
 			}
 			m.chatScrollOffset = 0
+			m = m.clearHover() // bodyY numbering differs between subchat and the parent transcript
 			return m, nil, true
 		}
 		// A click on a PLAN step row drops a transcript card listing the file
@@ -1058,6 +1276,7 @@ func (m model) handleTranscriptSelectionMouse(msg tea.MouseMsg) (model, tea.Cmd,
 				m = m.appendSystemNotice(errMsg)
 			}
 			m.chatScrollOffset = 0
+			m = m.clearHover() // bodyY numbering differs between subchat and the parent transcript
 			return m, nil, true
 		}
 		if line.toggle {
@@ -1070,21 +1289,57 @@ func (m model) handleTranscriptSelectionMouse(msg tea.MouseMsg) (model, tea.Cmd,
 		}
 		point := transcriptSelectionPointForMouse(line, mouseX(msg))
 		m.copyStatus = ""
-		m.transcriptSelection = transcriptSelectionState{active: true, anchor: point, cursor: point}
+		m.transcriptSelection = transcriptSelectionState{active: true, dragging: true, anchor: point, cursor: point}
+		// A fresh press always starts a brand-new drag: reset any glide state left
+		// over from a PREVIOUS selection that ended without going through
+		// stopEdgeScroll (e.g. a keypress mid-drag clears transcriptSelection.active
+		// directly, model.go's KeyPressMsg handler). Without this, a stale
+		// edgeScrollDelta can coincidentally match this new drag's direction and
+		// fool startEdgeScroll's "already running" fast path into silently doing
+		// nothing for the whole hold.
+		m = m.stopEdgeScroll()
 		return m, nil, true
 	case mouseMotion(msg):
-		if !m.transcriptSelection.active {
+		// Gate on dragging (the app's own press/release bracket), not active
+		// alone: active stays true through the async copy-command grace window
+		// after release, so without this a genuine hover motion arriving in
+		// that window would be misrouted here and silently move the
+		// just-released selection instead of falling through to hover
+		// handling. Deliberately NOT gated on msg's own Button field — see
+		// transcriptSelectionState.dragging's doc comment.
+		if !m.transcriptSelection.dragging {
 			return m, nil, false
 		}
-		line, ok := m.transcriptLineAtMouse(msg)
-		if ok {
+		if line, ok := m.transcriptLineAtMouse(msg); ok {
 			m.transcriptSelection.cursor = transcriptSelectionPointForMouse(line, mouseX(msg))
+			m = m.stopEdgeScroll() // back inside the body: any running glide ends
+			return m, nil, true
 		}
-		return m, nil, true
+		// The drag moved past the top/bottom edge of the visible transcript:
+		// auto-scroll toward that side and extend the selection to follow.
+		delta, ok := m.transcriptEdgeScrollDelta(msg)
+		if !ok {
+			m = m.stopEdgeScroll() // off to the side (e.g. over the sidebar), not an edge case
+			return m, nil, true
+		}
+		if m.reducedMotion {
+			// No animated glide: a single, larger step per raw motion event,
+			// matching the app's reduced-motion convention elsewhere.
+			m = m.dragToEdgeScroll(delta, mouseX(msg))
+			return m, nil, true
+		}
+		direction := 1
+		if delta < 0 {
+			direction = -1
+		}
+		var cmd tea.Cmd
+		m, cmd = m.startEdgeScroll(direction, mouseX(msg))
+		return m, cmd, true
 	case mouseRelease(msg):
 		if !m.transcriptSelection.active {
 			return m, nil, false
 		}
+		m = m.stopEdgeScroll()
 		if line, ok := m.transcriptLineAtMouse(msg); ok {
 			m.transcriptSelection.cursor = transcriptSelectionPointForMouse(line, mouseX(msg))
 		}
@@ -1093,6 +1348,10 @@ func (m model) handleTranscriptSelectionMouse(msg tea.MouseMsg) (model, tea.Cmd,
 			m.transcriptSelection = transcriptSelectionState{}
 			return m, nil, true
 		}
+		// The drag itself is over even though the selection stays active for
+		// the async copy-command grace window below — see dragging's doc
+		// comment on why mouseMotion must stop treating events as a drag now.
+		m.transcriptSelection.dragging = false
 		return m, copyTranscriptSelectionCmd(text), true
 	default:
 		return m, nil, false
@@ -1113,8 +1372,11 @@ func (m model) toggleTranscriptRow(rowIndex int) model {
 }
 
 func (m model) selectedTranscriptText() string {
-	width := m.chatColumnWidth()
-	layout := m.transcriptBodyLayout(width, "")
+	// Must read from the SAME row source transcriptLineAtMouse hit-tested against
+	// (subchat vs parent transcript) — otherwise the selection's bodyY range is
+	// matched against the wrong layout here and copy silently returns nothing.
+	_, items, _ := m.transcriptHitTestSource()
+	layout := layoutTranscriptBodyItems(items)
 	parts := []string{}
 	for _, line := range layout.selectable {
 		start, end, ok := m.selectedColumnsForTranscriptLine(line)

internal/tui/transcript_view.go

@@ -5,6 +5,11 @@ import "strings"
 func (m model) toggleDetailedTranscript() model {
 	m.transcriptDetailed = !m.transcriptDetailed
 	m.clearSuggestions()
+	if m.picker != nil && m.picker.kind == pickerTheme {
+		// Dropping a theme picker mid-preview must undo the live palette, or the
+		// previewed theme would stick while m.themeMode still holds the old one.
+		m.restoreCommittedTheme()
+	}
 	m.picker = nil
 	return m
 }

internal/tui/turns_command_test.go

@@ -0,0 +1,62 @@
+package tui
+
+import (
+	"os"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+func TestHandleTurnsCommand(t *testing.T) {
+	// Isolate the inherited-by-children budget env this command exports.
+	t.Setenv("ZERO_MAX_TURNS", "")
+
+	t.Run("sets a valid budget and exports it for children", func(t *testing.T) {
+		got, out := model{}.handleTurnsCommand("150")
+		if got.agentOptions.MaxTurns != 150 {
+			t.Fatalf("MaxTurns = %d, want 150", got.agentOptions.MaxTurns)
+		}
+		if !strings.Contains(out, "150") {
+			t.Fatalf("output should report 150, got %q", out)
+		}
+		if env := os.Getenv("ZERO_MAX_TURNS"); env != "150" {
+			t.Fatalf("ZERO_MAX_TURNS = %q, want 150 (so sub-agents inherit it)", env)
+		}
+	})
+
+	t.Run("clamps above the ceiling (env too)", func(t *testing.T) {
+		got, _ := model{}.handleTurnsCommand("99999")
+		if got.agentOptions.MaxTurns != maxTurnsCeiling {
+			t.Fatalf("MaxTurns = %d, want clamp to %d", got.agentOptions.MaxTurns, maxTurnsCeiling)
+		}
+		if env := os.Getenv("ZERO_MAX_TURNS"); env != strconv.Itoa(maxTurnsCeiling) {
+			t.Fatalf("ZERO_MAX_TURNS = %q, want clamped %d", env, maxTurnsCeiling)
+		}
+	})
+
+	t.Run("invalid / non-positive input is rejected without changing the budget", func(t *testing.T) {
+		for _, in := range []string{"abc", "0", "-5", "3.5"} {
+			m := model{}
+			m.agentOptions.MaxTurns = 42
+			got, out := m.handleTurnsCommand(in)
+			if got.agentOptions.MaxTurns != 42 {
+				t.Fatalf("input %q changed MaxTurns to %d, want unchanged (42)", in, got.agentOptions.MaxTurns)
+			}
+			if !strings.Contains(out, "Usage") {
+				t.Fatalf("input %q should show usage, got %q", in, out)
+			}
+		}
+	})
+
+	t.Run("empty / status shows the current budget without changing it", func(t *testing.T) {
+		m := model{}
+		m.agentOptions.MaxTurns = 42
+		got, out := m.handleTurnsCommand("")
+		if got.agentOptions.MaxTurns != 42 {
+			t.Fatalf("status changed MaxTurns to %d, want 42", got.agentOptions.MaxTurns)
+		}
+		if !strings.Contains(out, "42") {
+			t.Fatalf("status should report current 42, got %q", out)
+		}
+	})
+}

internal/tui/view.go

@@ -201,6 +201,9 @@ func (m model) statusLine(width int) string {
 		if m.exitConfirmActive {
 			return fitStyledLine(prefix+zeroTheme.amber.Render("●")+" "+zeroTheme.amber.Render(ctrlCExitConfirmText), width)
 		}
+		if m.cancelConfirmActive {
+			return fitStyledLine(prefix+zeroTheme.amber.Render("●")+" "+zeroTheme.amber.Render(escCancelConfirmText), width)
+		}
 		return fitStyledLine(left, width)
 	}
 
@@ -210,6 +213,8 @@ func (m model) statusLine(width int) string {
 	}
 	if m.exitConfirmActive {
 		left = prefix + zeroTheme.amber.Render("●") + " " + zeroTheme.amber.Render(ctrlCExitConfirmText)
+	} else if m.cancelConfirmActive {
+		left = prefix + zeroTheme.amber.Render("●") + " " + zeroTheme.amber.Render(escCancelConfirmText)
 	} else if summary := m.backgroundTerminalSummary(); summary != "" {
 		left += separator + zeroTheme.muted.Render(summary)
 	}
@@ -218,15 +223,19 @@ func (m model) statusLine(width int) string {
 	// Context-fill gauge: surface it down to the narrow tier (where it matters
 	// most), but skip it when the context sidebar is already showing the % so the
 	// figure isn't duplicated.
+	gaugeShown := false
 	if tier >= tierNarrow && !m.sidebarActive() {
 		if gauge := m.contextWindowSegment(); gauge != "" {
 			rightGroups = append(rightGroups, gauge)
+			gaugeShown = true
 		}
 	}
-	// The sidebar pins the token readout at its floor, so when it's open keep
-	// only the cost here — otherwise the token figure shows on both sides.
+	// The sidebar pins the token readout at its floor, and the gauge's "used"
+	// figure is the exact same number (both read latestUsageTokens) — either
+	// one showing means the plain usage segment must drop its own token count
+	// to just the cost, or the count renders twice side by side.
 	usage := m.usageStatusSegment()
-	if m.sidebarActive() {
+	if m.sidebarActive() || gaugeShown {
 		usage = m.usageCostSegment()
 	}
 	if usage != "" {