← graph

verify-agents-md

Verify AGENTS.md actually landed on main; diagnose + retry if not (cycle with create-agents-md)

Metadata

Statusdone
Assignedagent-943
Agent identity3184716484e6f0ea08bb13539daf07686ee79d440505f1fdf2de0357707034c3
Modelclaude:opus
Created2026-04-28T21:23:18.817713258+00:00
Started2026-04-28T21:38:26.424230699+00:00
Completed2026-04-28T21:38:51.706250731+00:00
Tagseval-scheduled, converged
Tokens205627 in / 1778 out
Eval score0.75
└ blocking impact0.88
└ completeness0.82
└ constraint fidelity0.25
└ coordination overhead0.88
└ correctness0.85
└ downstream usability0.75
└ efficiency0.85
└ intent fidelity0.79
└ style adherence0.88

Description

Description

Cycle partner for create-agents-md. Today (2026-04-28) we hit a confusing state: agent-933 ran the codex task, reported success, ran wg done, status went to Done — but AGENTS.md does not exist on main and there is no feat: create-agents-md commit. The codex handler appears to register task completion without going through the same git commit + merge lifecycle as the claude handler.

This task is the verification + iteration step. Each cycle iteration:

  1. Read git: does AGENTS.md exist on main? Is it byte-identical to CLAUDE.md? Is there a recent feat commit?
  2. If yes (everything landed): call wg done --converged <THIS_TASK_ID> to terminate the cycle cleanly.
  3. If no:
    • Diagnose the gap: read /home/erik/workgraph/.wg/agents/<latest-create-agents-md-agent>/output.log and look for what step of the wg done pipeline didn't fire.
    • Append findings to docs/codex-handler-merge-bug.md (create if absent — this becomes the artifact for the eventual real fix).
    • If the diagnosis points to a wg-side bug (codex handler doesn't commit/merge like claude does): file a fix task with wg add titled 'Codex handler: missing git commit + merge in wg done lifecycle', --model claude:opus, with a clear repro pulled from THIS task's findings, then exit this iteration normally — the cycle continues but the actual fix happens out-of-band.
    • If the diagnosis points to a fixable task-description issue with create-agents-md (e.g. codex misunderstood instructions): edit the task body via whatever wg surface allows, then wg retry create-agents-md for the next iteration.
    • If neither: log puzzlement, do not converge, let max-iterations end the cycle.
  4. Cycle config: max iterations = 3. After 3, the cycle stops on its own and the user investigates manually.

Important

  • This task itself runs on claude:opus (NOT codex) — it is the diagnostic agent, must not be subject to the broken codex pipeline.
  • It is allowed to file new tasks via wg add and to call wg retry create-agents-md — these are normal chat-agent operations.
  • Eval scores from .flip-* / .evaluate-* on this task are advisory and do not block the cycle. Convergence is purely via wg done --converged.
  • DO NOT manually create AGENTS.md from this task. The whole point is to verify codex can do it; circumventing defeats the smoke test.

Validation

  • Task converges (--converged) only when AGENTS.md is on main + byte-identical to CLAUDE.md + a feat commit exists referencing create-agents-md
  • If a wg-side bug is diagnosed, a fix task is filed referencing this task as source
  • docs/codex-handler-merge-bug.md exists with concrete findings if the cycle hits the failure mode
  • Cycle does NOT exceed 3 iterations (wg cycles should show max-iterations=3 on this loop)
  • No regressions: claude-handler tasks still commit + merge normally

Depends on

Required by

Log