Repository navigation
Conversation
Implementer agents (Tier 2) take an `accepted` proposal and open a PR in the matching sibling repo under mctlhq/. Triggered manually only — no auto- trigger on .status.yaml change (would require Argo Events Sensor, out of scope for v1). What's new: - orchestrator/run_implementer.py — discovers proposals by glob, marks .status.yaml in-progress, gh-clones the target repo, runs the SDK with the per-service implementer sub-agent, pushes the branch, opens a PR, marks .status.yaml implemented + PR URL. - orchestrator/options.py — IMPLEMENTER_BUDGET_USD ($3 soft cap) and build_implementer_agent_options() with cwd=cloned-repo and add_dirs=[]. - agents/<svc>/.claude/agents/implementer.md for all 7 services. Shared body, per-service notes (sanity-check command, repo idiosyncrasies). - Dockerfile installs `gh` CLI from cli.github.com — required by the orchestrator for clone + PR open. Bumps the implied image tag to 1.2.0 (referenced from the new CWFT in mctl-gitops). - requirements.txt: PyYAML for .status.yaml IO. Idempotency: a proposal stuck in `in-progress` is skipped on the next run unless --force is passed. .status.yaml schema follows the meta-plan (status / updated_at / updated_by / pr / notes). File absent = proposed.
…ter scaffold Address codex review on PR #2. P1 (run_implementer.py:188): discovery hard-filtered status == "accepted", so a wedged in-progress proposal could never be retried via --force; the operator had to hand-edit .status.yaml. Thread `include_in_progress` from the CLI --force flag through `find_accepted_proposals` so a force-run returns both accepted and in-progress proposals; `implement_one` already short-circuits in-progress unless force is set, so the two flags are naturally paired. P2 (run_implementer.py:232): the orchestrator copies implementer.md into the cloned repo's .claude/agents/ before running the SDK. If the sub-agent uses broad staging (git add -A), the runtime scaffold leaks into the service-repo PR. After copying the file, write the path to .git/info/exclude (per-clone, untracked-by-design ignore list). This keeps the repo's own .gitignore untouched — modifying it would itself be an out-of-scope change. Idempotent: re-runs don't duplicate the entry.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 09792ecc21
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| print(f"Unknown service '{args.service}'. Available: {', '.join(SERVICES)}", file=sys.stderr) | ||
| sys.exit(2) | ||
|
|
||
| ensure_auth_for_sdk() |
There was a problem hiding this comment.
The CLI currently calls ensure_auth_for_sdk() unconditionally, so python -m orchestrator.run_implementer --dry-run still fails if CLAUDE_CODE_OAUTH_TOKEN/ANTHROPIC_API_KEY are absent. That breaks the advertised dry-run behavior (discovery/spec parsing without executing the SDK) and makes non-auth validation in CI or local troubleshooting impossible. Gate this auth check behind not args.dry_run so dry-run can execute proposal discovery without Claude credentials.
Useful? React with 👍 / 👎.
mashkovd
left a comment
mashkovd
left a comment
There was a problem hiding this comment.
The cleanup behavior contradicts the stated post-mortem guarantee. The finally block removes target unconditionally, including CalledProcessError and other failures, although the comment says “Keep target dir for post-mortem on failure; clean only on success.” Track success and delete only then, or update the contract if retaining the failed clone is intentionally no longer desired.
|
Закрываю: функциональность уже на Проверено: Tier 2-пайплайн работает в проде: CWFT Открываю заново, если найдётся расхождение. |
Adds Tier 2 implementer agents to the orchestrator.
Fixes addressed from codex review on PR #2: