evolver run writes pending review state but does not exit cleanly on Codex

[rendigua]

Summary

On Codex, evolver run can generate pending review state and local artifacts successfully, but the process does not exit naturally and remains alive as a residual node process.

This appears to be separate from:

• the old _memoryFiltering.js copy bug (#547)
• the older Codex context/session-log discovery issues (#540, #543)

Those areas improved, but this process-lifecycle issue still affects actual usability.

Environment

• OS: Windows
• Host: Codex
• node: v22.15.0
• npm: 11.6.2
• package:
  • @evomap/evolver@1.87.2
  • also observed on @evomap/evolver@1.86.1

What I did

1. Installed the latest package:

   npm install -g @evomap/evolver@1.87.2 --prefer-online --no-audit

2. In a fresh git repo, ran:

   evolver setup-hooks --platform=codex

3. Verified the generated Codex hooks:

   • session-start runs and returns {}
   • session-end runs and returns {}

4. Ran:

   evolver run

5. Waited for local artifacts/state, then checked:

   • assets/gep/candidates.jsonl
   • memory/evolution/memory_graph.jsonl
   • memory/evolution/evolution_solidify_state.json
   • residual node / evolver processes

6. Ran:

   evolver review

Actual behavior

evolver run does not exit naturally within the observation window, but by that time it has already produced:

• candidates.jsonl
• memory_graph.jsonl
• evolution_solidify_state.json with last_run

evolver review can then read the pending run correctly.

So the sequence is effectively:

1. pending run state is written
2. reviewable artifacts are written
3. but the evolver run process stays alive instead of terminating

This leaves a residual node process that may need to be killed manually before continuing.

Why this matters

This is not the same as "run failed."

The core pipeline seems to reach a reviewable state, but the process lifecycle does not finish cleanly. In practice this hurts:

• CLI usability
• automation expectations
• host integration reliability
• human review flow, because a still-running process can keep mutating or rewriting state while review is already possible

Expected behavior

After the pending state and reviewable local artifacts are fully written, evolver run should terminate cleanly instead of remaining as a residual process.

If some long-running background mode is intended, that should be explicit and documented separately from the normal evolver run behavior.

[imxyanua]

Thanks for the precise repro, @rendigua — splitting “artifacts written” from “process still alive” pointed the diagnosis at exactly the right surface (process lifecycle, not the write path), and ruling out #547 / #540 / #543 up front saved a good chunk of triage time.

Root cause confirmed: src/atp/autoDeliver.js creates its poll handle with setInterval(...) but never calls .unref() on the returned Timeout, so the refed timer pins the Node event loop after evolver run finishes draining its foreground cycle.

--loop masks the issue because the daemon’s foreground evolve loop already keeps the process alive on its own schedule; only the single-shot run path exposes the residual node process.

Internal patch is already in flight:

• add a one-line _pollInterval.unref() matching the existing pattern used in:

  • src/proxy/lifecycle/manager.js
  • src/proxy/sync/engine.js

• add a hasRef() regression assertion in test/atpAutoDeliver.test.js

That assertion was verified by surgically reverting just the .unref() call and confirming the test fails with the expected regression message.

We’ll ship the fix through the normal patch-release pipeline shortly and close this issue once the public mirror lands.