Termination Gatekeeper

Purpose

Externalize completion judgment so work is called done only after required evidence exists.

Inspect First

• task spec, feature list, OpenSpec tasks, changed files, validation scripts, tests, runtime docs, release checklist, and current command output

Procedure

1. Define done for the task.

   • Create or update docs/development/definition-of-done.md.
   • Create or update task-specific docs/development/termination-evidence.md when useful.
   • Include static, runtime, e2e or integration, docs, and clean-state evidence as applicable.

2. Classify changed surface.

   • UI/browser changes usually require accessible e2e or browser workflow evidence.
   • API/runtime/data changes usually require runtime and integration evidence.
   • Docs-only changes can use docs/link/placeholder checks.
   • Risky changes should include pre-edit investigation evidence from $pre-edit-investigation-gate.
   • Browser-facing lifecycle changes should include click-path evidence from $click-path-and-browser-qa.

3. Add advisory validation.

   • Create or recommend scripts/termination-check.
   • Use scripts/check_termination_evidence.py from this plugin.
   • Do not force heavy e2e on repos where the changed surface does not need it.

4. Block premature completion in instructions.

   • Completion is not "code looks done."
   • Completion requires evidence from commands, runtime checks, or documented blockers.

5. Route evidence.

   • Update feature-state-machine evidence for passing features.
   • Include evidence in PR output via task-planning-and-pr-protocol.
   • Add final validation or blocked outcome events to docs/observability/task-trace.jsonl when the repo uses $trajectory-risk-supervisor.
   • Use project-local-skill-generator to create project-definition-of-done when this repo should expose completion rules as a local Codex skill.

Validation

• Run ./scripts/termination-check when present.
• Confirm evidence names exact commands or runtime checks.
• Confirm any skipped layer has a documented reason.

Completion Criteria

• The task has objective completion evidence or explicit unresolved blockers.
• The agent cannot honestly claim done based only on code inspection.