tui-live-log

Metadata

Description

## Description

When a codex agent is actively running and producing output (visible in `.wg/agents/<id>/output.log`), the TUI's live log pane shows empty. Confirmed today on agent-941 (create-agents-md cycle iteration 2): output.log was 28+ lines and growing while the TUI live log pane showed nothing.

User quote: 'if create-agents-md is running i don't see it. the log is completely empty' — followed by 'codex log view... it's busted'

This is distinct from:
- `tui-detail-view` (iteration-specific FLIP score labels)
- `log-raw-view` (head/tail summary feature)
- `fix-pty-output` (already-fixed PTY doubling)

This is specifically: **codex agents' stream isn't being picked up by the TUI's live log pane**, while claude agents' streams are. The wg-level log entries (`wg show <task>` Log section) DO populate correctly for codex — only the live pty/stdout stream rendering in the TUI is missing.

### Likely investigation
1. Compare how claude vs codex agent streams are tee'd / read by the TUI. Likely `src/tui/log_view` or wherever the live pane subscribes.
2. Codex CLI emits JSONL on stdout (item.completed / item.started / agent_message etc) — confirm the TUI's parser handles that format. The PTY-doubling fix (`fix-pty-output`) handled streaming_text vs finalized message races; this might be a similar handler bypass for codex's structured stream.
3. Check whether the live pane is reading `output.log` directly, or subscribing to a stream channel that the codex handler isn't writing to.

### Repro
1. Start a codex task (any `--model codex:*` add)
2. Open TUI, navigate to the running agent's log pane
3. Observe: pane is empty even while `tail -f .wg/agents/<id>/output.log` shows active writes

## Validation

- [ ] Failing test first: render the TUI log pane with a fixture that mimics a running codex agent's output stream; assert non-empty pane content
- [ ] Claude tasks' streams continue to render unchanged (regression check)
- [ ] Manual smoke: dispatch a codex task on `codex:gpt-5.5`, open TUI live log on the agent — pane shows the codex stream as it arrives
- [ ] cargo build + cargo test pass

Depends on

• done .assign-tui-live-log

Required by

• (none)

Log

• 2026-04-28T21:44:00.566277693+00:00 Task paused
• 2026-04-28T21:44:04.139262037+00:00 Task published
• 2026-04-28T21:44:32.624830596+00:00 Lightweight assignment: agent=Careful Programmer (f5143935), exec_mode=full, context_scope=task, reason=Careful Programmer (0.73 score, 182 tasks) is ideal for precision-critical TUI/stream debugging with regression requirements.
• 2026-04-28T21:44:34.176710439+00:00 Spawned by coordinator --executor claude --model opus
• 2026-04-28T21:44:44.699982878+00:00 Starting investigation: examining TUI log pane code and codex stream handling
• 2026-04-28T21:48:41.071427039+00:00 Root cause identified: parse_raw_stream_line in src/tui/viz_viewer/state.rs only handles claude (assistant/user/system/tool_call) and native (turn) JSONL formats. Codex emits thread.started/turn.started/item.completed/item.started/turn.completed which all return None — so codex agents' raw_stream.jsonl produces zero stream_events and the TUI live log pane shows empty.
• 2026-04-28T21:51:30.854701693+00:00 Task unclaimed: agent 'agent-946' (PID 612167) process exited
• 2026-04-28T21:51:40.457225879+00:00 Spawned by coordinator --executor claude --model opus
• 2026-04-28T22:07:03.258275751+00:00 Validated: 21/21 agent_stream_tests pass (5 new codex tests + 16 existing). Pre-existing config::test_global_config_path failure is unrelated (verified by reverting). cargo build + cargo test pass; updated wg binary installed via cargo install --path .
• 2026-04-28T22:07:42.438085424+00:00 Committed: f3adc29c3 — pushed to remote (wg/agent-946/tui-live-log)
• 2026-04-28T22:07:51.062860864+00:00 Task pending eval (agent reported done; awaiting `.evaluate-*` to score)
• 2026-04-28T22:10:44.761071819+00:00 PendingEval → Done (evaluator passed; downstream unblocks)