Repository navigation
fail-safe CLI: unknown verb exits 0; no exit-10 for agent logical failure #22
Description
Two fail-safety gaps found while wiring the Argo lane (adlc-argo-wf !3 review). Both are small, both have silent-success failure modes.
M1 — unknown top-level verb exits 0 (must)
src/dispatch.mjs:154-157 — the default: case prints help and returns 0. Consequence, verified: adlc-cli factory-mission --issue … --run-id … prints help and exits 0 — a run that never happened reports success. (The Argo seed template shipped exactly this invocation.) Fix: return 1 for an unrecognized command string; keep bare help/no-args at 0.
M2 — exit-10 contract unimplemented (must, retry-gating)
The Argo lane retry gate (expression: asInt(lastRetry.exitCode) != 10) needs logical failure to exit 10. Today:
src/commands/agent.mjs:62-64returns the raw agent code — logical failure and usage error both yield 1, indistinguishable.src/commands/workflow.mjs:174,215return 0/1 (PAUSED→ 1, which a runner retries as infra).
Fix: classify agent logical failure → return 10; workflow terminal logical failure → 10; decide a non-retry paused code with the lane template. Related: workflow run/resume: distinguish FAILED, PAUSED and ABORTED in exit code (all collapse to 1) #19 / Unknown verb exits 0 — silent no-op reports success #20.
Acceptance
- Unknown verb → non-zero exit with usage hint.
- Simulated agent logical failure → exit 10, zero Argo retries; OOM-kill/infra → retried with backoff.
- Paused workflow → documented code the lane expression excludes.
Lane context: adlc-argo-wf !3 comment (2026-09-30) carries the full five-item review; agentic-container side (exit-10 propagation) tracked separately.
Activity
kanfil commentedon Sep 30, 2026
Mission started on
tikalk/adlc-cli#22(fail-safe CLI: unknown-verb exit + exit-10). Runmission-adlc-cli-22-20260930, worktree-isolated (fix/issue-22-cli-fail-safe), autonomous. Drafts stay local; step findings publish here as markers.kanfil commentedon Sep 30, 2026
Test step done — RED gate passed (executor re-ran the suite: 12 tests, 3 pass, 9 fail).
tests/cli-failsafe.test.mjs. Genuine behavior failure: unknown verb exits 0, expected 1 (0 !== 1). Guards green: bare→0,help→0,run-without-prompt→1.src/exit-codes.mjsabsent). Implementer contract: provideclassifyAgentExit({code, signal})(0→0, non-zero child→10, signal→130) andclassifyWorkflowStatus(status)(COMPLETED→0, PAUSED→3, FAILED→10, ABORTED→1); wire intodispatch.mjsdefault (unknown→1),agent.mjspost-spawn return,workflow.mjsrun/resume terminal mapping. Lane expression (applied in adlc-argo-wf, not here):asInt(lastRetry.exitCode) != 10 && asInt(lastRetry.exitCode) != 3.fix: machine-readable CLI exit codes (unknown verb, logical failure, …
kanfil commentedon Sep 30, 2026
Mission complete — audit trail:
fix/issue-22-cli-fail-safeoff main; remote lease acquired; issue OPEN unlabeled → autonomous.tests/cli-failsafe.test.mjs(12 tests), RED verified by executor re-run (9 fail: 1 behavior + 8 seam-missing).src/exit-codes.mjs+ wiring indispatch.mjs/agent.mjs/workflow.mjs; GREEN verified (12/12 new, 318/318 full, 34 suites)..adlc/workflows/runs/mission-adlc-cli-22-20260930/(brief.md, iterations.md, mission.yml).agent-authoredlabels do not exist in this repo → no label transitions; stamp carried in PR body. No engine-managed run → completion recorded in local state, notworkflow state advance.Remaining for the lane: container propagation (tikalk/engineering/agentic-container#1) + lane expression application (adlc-argo-wf values MR).