Participant: codex (AI agent)
Understanding
The ecosystem already has the intended Doctor → Repair → Validator split, but
the paths are not currently one working chain. Standalone repair-agent is
scheduled and healthy in dry-run mode, validator-agent is actively approving
exact PR heads, while skills-agent discovery is completely blocked by the
invalid 0014_publication-freeze package. Adding more skills before restoring
discovery would increase inventory without increasing capability.
Todo2code can provide a useful deterministic evidence plane: its offline
pipeline produced a successful immutable run and graph over skills-agent.
Its proposed edits must remain advisory because stale TODO declarations can
produce grounded-looking plans for paths that no longer exist.
Execution plan after approval
- Create a skills-agent repair PR for
0014, schema validation isolation and
GLM 5.2 configuration alignment.
- Require its normal CI and independent validator-agent exact-head approval.
- Add the read-only
0015_todo2code-governance-health pilot as a separate
skills-agent PR and validate the todo2code manifest/graph contract.
- Prove one correlation-bound Doctor → Repair PR → Validator rejection or
approval cycle with auto-merge disabled.
- Only then add dependency and branch-lifecycle repair skills in separate PRs.
- Record results in this ticket and close it after all cross-repository
evidence is immutable and reproducible.
Actual changes
- Audited local code, current GitHub workflow runs and protected variables.
- Ran todo2code deterministically against skills-agent with all LLM stages
disabled; the run succeeded and wrote artifacts only under
/tmp.
- Created this coordination plan only. No implementation or external state was
changed.
Blockers
- Explicit approval of the proposed skill IDs and phased ordering.
- The OpenRouter Repair key total limit must be resolved before live Repair can
prove a paid patch/PR hand-off.
- repair-agent’s existing local changes belong to another workstream and must
not be included.