fix(operator): bound autonomous task convergence

[urbanplanningpros]

Issue for this PR

Tracks the long-running task non-convergence failure reported in upstream openai/codex#35892 on July 29, 2026.

The report describes a finite coding task repeatedly expanding into new tasks, subagent lanes, review cycles, and verification gates over approximately three days instead of reducing one evidence-backed critical path. Local repository issues are disabled, so the upstream report is the compatibility reference.

Type of change

• Bug fix / automation continuity guard
• New feature
• Refactor / code improvement
• Documentation through executable policy

What does this PR do?

Adds a provider-neutral convergence admission guard that:

• binds execution to an immutable completion-criteria SHA-256;
• enforces maximum elapsed time, correction-cycle, and consecutive non-reduction limits;
• detects unjustified work expansion;
• preserves the existing single-agent boundary on authoritative routes;
• rejects completion-criteria drift without a new authorized task manifest;
• requires reconciliation of uncertain external writes before replay or publication;
• requires a terminal snapshot when a limit or operator stop is reached;
• verifies the task-owned diff manifest, clean snapshot state, immutable commit SHA and draft PR when repository writes are authorized;
• requires continuation state outside model memory;
• ends the current task after the terminal snapshot instead of starting another review or QA cycle;
• preserves direct OpenAI and explicitly authorized local continuity routes without adding a provider, gateway, automatic selector, Bedrock, Vertex, or Copilot dependency.

Run:

node scripts/operator/codex-task-convergence-guard.mjs \
  --input /approved/task/convergence-evidence.json \
  --json

Exit behavior:

0   task remains bounded, or a terminal snapshot is verified
75  convergence limit reached; terminal snapshot required
64  policy, routing, state, provenance or write-reconciliation failure
2   malformed evidence

How did you verify the code?

The deterministic self-test covers:

• a healthy reducing task;
• correction-cycle exhaustion;
• an operator-requested stop with a verified commit and draft PR snapshot;
• completion-criteria drift;
• forbidden subagent enablement;
• unreconciled uncertain writes;
• prohibited gateway or automatic-selection identifiers.

Remote repository CI and one disposable bounded long-running task canary remain required before promotion.

Screenshots / recordings

Not applicable; this is a non-UI automation guard.

Checklist

• No excluded provider, gateway, automatic selector, Bedrock, Vertex, or Copilot route was introduced
• No production secret, connector credential, or customer data was added
• Business-critical work continues through direct OpenAI or the explicitly authorized local route

Sanitized-Manifest: codex-task-convergence-2026-07-29
Allowed-Paths: scripts/operator/codex-task-convergence-guard.mjs, scripts/operator/codex-task-convergence-guard-selftest.mjs, docs/operator/codex-task-convergence.md

[github-actions]

This PR doesn't fully meet our contributing guidelines and PR template.

What needs to be fixed:

• PR description is missing required template sections. Please use the PR template.

Please edit this PR description to address the above within 2 hours, or it will be automatically closed.

If you believe this was flagged incorrectly, please let a maintainer know.

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

This pull request has been automatically closed because it was not updated to meet our contributing guidelines within the 2-hour window.

Feel free to open a new pull request that follows our guidelines.