feat(terminal): Log view — transcript live tail (main + subagent transcripts)

[thanhsmind]

What

A second, additive observation surface for Claude Code panes: a Log toggle in the terminal view that tails the agent's own on-disk session transcript instead of screenshotting the terminal grid.

The existing screen view polls pane.read — an 80-line ANSI snapshot every 1.5s — so anything that scrolls past between ticks is unrecoverable, and there is no concept of a tool call, command output, or subagent. The transcript (~/.claude/projects/<encoded-cwd>/…) is append-only and written by Claude Code itself, so a byte cursor over it is gap-free by construction: every command with its full output, thinking, edits, and subagent activity between two polls arrives, regardless of poll interval. This is the transcript-as-history-source candidate from the distillery (collie, docs/distillery/porting-log.md:104), finally built.

How

• src/transcript/ (new): resolve pane → transcript via the snapshot's foreground_cwd ?? cwd (D5 precedence) and Claude Code's project-dir encoding; gap-free tail reading complete lines only (a mid-append partial line waits for the next poll); defensive per-record rendering — unknown record shapes pass through raw, nothing is silently dropped; cursor validated against path escape.
• Subagents: Claude Code ≥2.1.x writes Task/teammate activity to <session-id>/subagents/agent-*.jsonl, not as isSidechain records in the main file (verified live against 2.1.222). The watch set is therefore multi-file: main session + every subagent transcript, merged by record timestamp, subagent lines prefixed ⑂, files born mid-watch tailed from byte 0.
• GET /api/panes/:pane/activity (new route): stateless per request like every existing route — the (opaque, multi-entry) cursor round-trips through the client; the gateway stores nothing, keeping the never-store rule (src/store/mod.rs) intact. Panes without a transcript answer available:false (data, not an error).
• Web UI: a "Log" footer toggle swaps the xterm screen for an append-only <pre> with a client-owned 200-line ring and stick-to-bottom (release on scroll-up). Opening backfills the tail so the ring lands full. Screen view, PageUp history, keys/reply sheets are untouched.
• Drive-by fix: demo_config() now builds its JSON via serde_json — the Windows temp dir's backslashes made --demo panic on invalid JSON escapes.

Scope & limits

• Open backfill window 512 KiB, per-record caps (tool output 40 lines, text 80 lines, 400 chars/line), 400 lines per poll, 200-line client ring. Truncation is always marked (… (+K lines)), never silent.
• Claude Code panes only for now — other panes show "No transcript" and keep the screen view as before. No WebSocket (consistent with decision 675fc93a), no persistence, no process ownership.
• Design notes and resolved decisions: plans/260805-1500-transcript-as-history-source/plan.md.

Testing

• 16 unit tests on the transcript module (encoding, EOF-open backfill, incremental append, partial-line handling, torn-window start, subagent merge, mid-watch subagent pickup, session-switch divider, cursor path-escape rejection, render shapes, ANSI strip/clip).
• 3 handler tests (auth-opaque 404, unknown pane 404, available:false).
• 5 new vitest tests on the Log toggle (viewport swap, cursorless open, cursor round-trip across toggles, 200-line ring, screen-poll resume). 112 web tests / 338 lib tests green (the one failing lib test, doctor::checks::agent_presets_fix_seeds…, fails on a clean checkout of main too — pre-existing, unrelated).
• Verified end-to-end against a live deployment: real Claude Code 2.1.222 session with an active teammate; open lands with the ring full, subagent lines stream live.

🤖 Generated with Claude Code