fix: sub-agent hang/crash stranding runs mid-phase with no save
Some checks failed
port-to-omp / port (push) Failing after 2s

A stalled sub-agent (provider stream never settling after the final tool
call, or a throwing session-event listener recursing through the SDK's
run-failure path to stack overflow) left the phase stuck in_progress with
no state save, no error, and no completion message — observed twice in
freno-dev, both times after comments wrote findings.md.

- agent-runner: 60-min settle watchdog on every agent phase
  (PYGIENIUM_AGENT_TIMEOUT_MS, env-tunable); timeout disposes the session
  and fails the phase loudly with a /pygienium-resume hint instead of
  hanging the check-runner's await forever.
- agent-runner: applySessionEvent — the session.subscribe listener can no
  longer throw into the SDK event pipeline (guards for partial/malformed
  events, optional-chained message_update, throwing chat forwarder).
- comments: scan no longer regenerates the full report as its final
  message (findings.md is the source of truth); fix phase reads
  findings.md with embedded fallback.
- check-runner: failing state-save inside the catch path can't double-
  fault or escape as an unhandled rejection; completion posting is
  best-effort.
- tests: watchdog timeout test, comments task-text regression tests,
  applySessionEvent malformed-shape unit tests. 146 pass, tsc clean.
- README: document PYGIENIUM_AGENT_TIMEOUT_MS.
This commit is contained in:
2026-08-11 08:47:12 -04:00
parent 6d23b04ef6
commit 8768f9de97
7 changed files with 435 additions and 48 deletions

View File

@@ -36,6 +36,12 @@ Pygienium reads its chat-rendering style from pi's `settings.json` (`~/.pi/agent
No entry means `"verbose"` (the default). An unreadable or missing `settings.json` also falls back to `"verbose"`.
Environment variable (read at run time):
| Env var | Default | Description |
| --- | --- | --- |
| `PYGIENIUM_AGENT_TIMEOUT_MS` | `3600000` (60 min) | Per-agent-phase settle deadline. A sub-agent session that never settles (stalled provider stream, hung retry/auto-compaction after its last tool call) is aborted and the phase fails with a visible error instead of hanging the run mid-transition with no state update and no completion message. Lower it (e.g. `1200000`) to fail fast on flaky providers; raise it for very large repos.
## Commands
Every command accepts a `[path]` target (default: the current directory) and is