TUI Rework — feature/tui-rework

Autonomous fix branch for rendering/streaming/tooling issues found via session-log analysis. Work is sequential: each P lands as its own commit with a regression test.

Status: ALL DONE. P0–P4 committed. Run the full gate before merging: go vet ./... && go test -count=1 -race ./tui/ ./tools/ ./internal/agentic/ ./internal/app/

Commits on this branch (newest first)

P0 solution (committed b4a5d4c + 7ba5ddc)

Symptom: while a tool runs, its widget body was empty and the footer showed only wall-clock "elapsed Ns" — no bytes/lines growth. Bash output appeared only at exit.

Findings from the trace:

• Bash stdout streaming already worked end-to-end (progressWriter → EventToolProgress → tracker.OnProgress → SetOutput); tests TestToolProgress_ShowsPartialOutputWhileRunning / TestToolStreamingRepro_* pass.
• Arg streaming exists for OpenAI (EventToolCallDelta + Partial snapshot) and Anthropic (input_json_delta + ContentIndex). For providers emitting neither, no widget can exist before the call arrives — nothing to show "preparing…" in. The visible "frozen" symptom was the missing growth indicator, not missing widgets.

What landed:

1. Live-progress footer (tui/tool_execution.go): ToolExecutionComponent tracks outputBytes/outputLines (updated in SetOutput, which serves both partial progress and final result). While Pending/Running the duration line renders elapsed 12.3s · 1.2 KB · 84 lines; suffix omitted with no output yet (fast tools stay clean) and dropped on completion (plain Took Xs). Helpers: progressSuffix, formatByteSize, formatLineCount. Test: TestToolExecution_LiveProgressFooter (+ formatter unit tests).
2. Provider delta diagnostic (internal/agentic): toolCallDeltasThisRound counted in handleStreamToolCallDelta, reset per round, logged at handleStreamDone as stream round done: tool_call deltas=N. A zero count in session logs proves the active provider/model can't stream args, settling the "verify provider" question with data instead of code reading.

P3 solution (committed 1c5b39b)

Bug: context_lines was advertised in the schema and parsed into searchParams.ContextLines but NEVER used — searchWithPattern/formatResults never received it. The old test only asserted non-empty output (passed vacuously). Models learned "search gives no context" and fell back to bash grep (321 bash-grep vs 100 search calls across 15 sessions).

Fix:

1. searchResult gained Before/After; searchFile captures ±N lines per match via the matchContext helper (kept gocognit under budget), threaded through searchFiles → searchWithPattern → Execute/searchPipePattern.
2. Rendered grep-style in formatFileContentLines: context rows N- content vs match rows N: content (shared sanitizeSearchLine helper).
3. Default context_lines changed 1 → 0 (opt-in; plain searches stay compact).
4. New max_lines param (schema + params + doc): truncateOutputLines caps output content lines like | head -N, emitting a … (K lines truncated by max_lines) marker after the header.
5. Tests: TestSearchTool_Execute_WithContextLines now asserts context rows actually appear; TestSearchTool_Execute_ContextLinesDefaultOff; TestSearchTool_Execute_MaxLinesTruncates; TestTruncateOutputLines table test. Documented in tools/search.long.md.

Not done (nice-to-have, low value): invert_match, count_only, files_only modes (60 bash greps used -v; 4 -c; 1 -l).

P4 solution (committed bfa61b2)

Rewrote the compositor around an explicit-scrollback model instead of the old accreted six-strategy scroll dispatch:

• scrollTop watermark — transcript rows emitted into terminal scrollback exactly once, in order, monotonically advancing, clamped to the chrome band start so chrome can never leak (structural, not heuristic).
• One scroll mechanism (emitScrollbackAdvance) with two sub-cases: first-frame top-down write (blank screen) and steady-state bottom scroll-then-write. The first/large/bare/deleted/shrink/resize strategy dispatch is gone.
• prevWindowFull discriminator — a previous window counts as "full" only if every transcript region row is non-blank; partial windows (blank padding from bottom-aligned short content) take the full-range re-emit path. This was the subtle case behind the last content-loss bug (the start frame).
• DECSTBM scroll region [1, height-chromeH] confines native terminal scroll to the transcript, so pinned chrome (status/editor/footer/bubbles) never moves and never enters scrollback.
• Chrome classification is role-based: the single HeightAllocated child is the scrollable transcript; every child after it is pinned chrome (Scene.ChromeHeight computed in buildScene, tui/tui.go).

Net −208 lines in compositor.go. Both test emulators (TermEmulator, screenEmulator) gained DECSTBM support. Refactored into small helpers to stay within gocognit 15 / gocyclo 12 (transcriptWindow, emitFirstFrameScroll, emitSteadyScroll, steadyWriteRange, prevWindowFull, scrollTarget, advanceScrollback, repaintWindow).

P1 solution (committed 3c92af0)

ToolExecutionComponent.expandedSet flag: an explicit per-widget toggle (Enter/Ctrl+O on a focused block, HandleInput → setExpandedExplicit) wins over the global view policy in BOTH directions and persists across streaming re-renders. effectiveExpanded() checks expandedSet first (tui/tool_execution.go). The global toggle-all clears overrides via ClearExplicitExpand() in ChatViewport.invalidateAllToolWidgets (tui/chat_viewport.go) so Ctrl+O-all still flips everything uniformly.

P2 solution (committed 8269910)

Steering recall already worked: Alt+E → App.handleEditSteering (internal/app/submithandler.go:166) flushes SteeringQueue back into the input line and clears the bubble/footer. The gap was discoverability:

• bubble footer now shows "Alt+E to edit" (tui/chat_viewport_components.go);
• KbEditSteering registered (tui/keybindings.go);
• listed in /hotkeys help (core/commands/hotkeys.go).

Key lessons / gotchas

• Don't restructure compose's canvas layout (capping/shifting rows broke 15 tests). The virtual-buffer invariant: scrolled-off transcript rows must stay addressable so the scroll path can emit them to native scrollback.
• The tui/ test suite cancellation: go test ./tui/ -count=1 occasionally gets canceled in this environment; build+vet are reliable. Re-run if canceled.
• Reference implementations: ../pi renders one flat buffer and lets chrome scroll into history (accepts the leak). ../opencode uses OpenTUI (Zig-backed cell-buffer compositor, internal scroll, no native scrollback). Goa's DECSTBM-region approach pins chrome without adopting a full cell-buffer model.
• Test emulators (tui/term_emulator.go, tui/streaming_scroll_test.go) both model DECSTBM scroll regions; replay tests through emu.Scrollback() to assert what lands in history.
• Gate before committing: go vet ./tui/, go test -count=1 -race ./tui/, gocognit -over 15 ./tui/, gocyclo -over 12 ./tui/ (must be empty output).