[gering]

Why

Local python3 scripts/check-structure.py went red while CI stayed green: per-test timeouts (F7) and test_codex_models / test_kimi_acp failures under load (F8).

What the measurements showed

Per-test wall time measured alone, with 3 full runs in parallel, and with 12 parallel runs of a single test:

• macOS assesses every new executable on its first exec (syspolicyd, ~150 ms; a symlink to an already-run file costs ~4 ms). Parallel runs pile up on it. Tests that wrote a fresh stub per case or per Env spent most of their time waiting on it: test_kimi_acp took 67 s wall for 12 s CPU, and 419–590 s in parallel. test_agent_registry took 444 s in parallel. Linux CI has no such check. The mechanism is confirmed by research; the pile-up under load is inferred from the measurements.
• test_codex_models gave the helper deadlines of 0.4–2 s, and a working fake peer had to start inside them. It failed in 12 of 12 parallel runs.
• test_sessions waited a fixed 1 s per pane poll in sessions.py.
• test_insights_handoff looked for leftover temp files in the shared $TMPDIR and saw files from parallel runs.

Changes

• swarm tests: one shared fake per class or module (kimi, grok-compat). The codex test now separates the deadline under test from the time a working peer gets, and asserts that rejected cases were not merely timeouts. The kimi fake now waits up to 10 s, not 1 s, for the client's kill.
• work-system: SESSIONS_POLL_MS test seam; content-addressed stub links in test_agent_registry and test_herdr_shell; a private TMPDIR for the leak check.
• check-structure.py: a TEST_TIMEOUTS table. Only test_mandate (75 s) gets its own 300 s; the 120 s default stays. A stale entry is an error.
• Knowledge note: deployment/ci-structure-checks.md.

Proof

• A full run with 3 in parallel: 1360 s → 391 s. kimi 419 → 13.5 s, agent_registry 444 → 22 s, sessions 80 → 29 s, herdr_shell 67 → 31 s.
• test_codex_models, 12 runs in parallel: 0/12 → 12/12 green.
• Final head: 3 parallel and 3 sequential check-structure.py runs, all OK.

Review

One local /swarm:review round (Claude lenses, gpt-6.1-sol, grok-4.7; 15 of 15 voices) found 9 issues. 6 are fixed in this PR:

• the shared fake directories now clean up after themselves (grok-compat, insights-handoff TMPDIR);
• the stty and raw_grok stubs read their paths from env, so the stub cache now actually hits;
• HANG_BOUND is 8 s, below the helper's 10 s default, so it still catches a helper that ignores --timeout;
• SESSIONS_POLL_MS can only shorten the poll, never lengthen it;
• the knowledge page's pluginVersion stamp is updated.

3 were not acted on:

• one shared stub helper across test files: tests stay self-contained by convention;
• the 5 s descendants deadline: 12 of 12 runs passed under 12x load, so this is an accepted residual;
• the leak check only sees the private TMPDIR: the old check was no stronger.

check-structure.py passes on the fixed head.

Not done

• The test_kimi_acp 1 s kill race never reproduced in 12 parallel runs. The 10 s bound comes from reading the code and costs nothing on the passing path.
• test_herdr_launch is out of scope (its own lane).

🤖 Generated with Claude Code

https://claude.ai/code/session_01WFDBrBNHxBX8MUG6oAC8n6