Add @claude-triggered multi-agent issue-fix pipeline - #122
Conversation
![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: b8970c89c1
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".
Route @claude fix / feature requests on issues (trusted maintainers only, gated by claude.yml's existing if:) through a gated multi-agent pipeline. - .claude/commands/fix-issue.md: shared orchestrator command. Stage 0 runs a single-agent bug-verifier gate; the multi-agent pipeline (context-builder -> fix-planner -> orchestrated fix-implementer/fix-verifier loop) runs only on a confirmed bug / actionable feature, otherwise it comments and stops. - .claude/agents/: bug-verifier (reproduction gate, with python/pytest/pip), context-builder, fix-planner, fix-implementer, fix-verifier. - .github/workflows/claude.yml: the @claude path now allows the Task tool + read tools and routes issue fix/feature requests through the pipeline, preserving the existing auto-review handling. The bug-label auto path (claude-verify-bug.yml / claude-fix-bug.yml) is untouched. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
b8970c8 to
ec9f3dd
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ec9f3dd628
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".
The @claude verification-gate + multi-agent pipeline (claude.yml + .claude/commands/fix-issue.md) is now the sole way fixes/features are implemented. Remove the older single-agent auto-fix mechanics: - delete claude-verify-bug.yml and claude-fix-bug.yml (the bug-label read-only-verify -> dispatch-fix workflows). - drop the dispatch-verify job from claude-issue-triage.yml so triage only applies labels and never auto-triggers a fix. - update claude.yml comment: @claude on an issue is the only fix/feature path. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ore tests - bug-verifier: drop raw git/gh; restrict GitHub access to the read-only ./scripts/gh.sh wrapper and add untrusted-input (prompt-injection) hardening, so the pre-gate verifier cannot push/comment with the job's write token. - fix-issue.md: add a Reporting convention so GIT: harness never calls gh/git (including the gate-failure and non-convergence stops) and returns the verdict as its final message; only GIT: self posts comments / opens PRs. - fix-verifier / fix-implementer: add Bash(pip:*) and a build-first step so a fresh/unbuilt checkout (notably the feature path) is built per CLAUDE.md instead of being reported as a test failure. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ddbf513df3
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".
The bug-verifier stays read-only and never comments; the orchestrator owns the user-facing report. Make that explicit and have the verifier emit a self-contained, comment-ready summary (for not_reproducible / needs_info / out_of_scope, stating what would unblock the issue) that the orchestrator posts verbatim — an issue comment in GIT: self, or its final message in GIT: harness. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8e4602b02f
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".
…er, unique branch - context-builder: drop raw gh/git; restrict to ./scripts/gh.sh + non-mutating git verbs (log/show/diff/blame/status), with untrusted-input hardening. - fix-planner: drop raw gh; read-only via ./scripts/gh.sh only, plus hardening. - fix-verifier: make the original-reproduction check conditional — bugs rerun the repro, features run the plan's acceptance exercise instead (no repro exists for a feature). - fix-issue.md: before opening a PR, reuse/skip an existing open bot PR for the issue, and append a unique suffix when the branch ref already exists so a re-trigger does not fail with a non-fast-forward push. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
1 participant
Summary
Makes the manual
@claudemention the single path for resolving issues (bug fixesand feature implementations), running a gated multi-agent pipeline. Removes the older
label-triggered single-agent auto-fix mechanics entirely.
Flow when a trusted maintainer (
OWNER/MEMBER/COLLABORATOR)@claudes on an issue:Added
.claude/commands/fix-issue.md— shared orchestrator command. Stage 0 is asingle-agent verification gate; the multi-agent pipeline runs only on a
confirmedbug /
actionablefeature, otherwise it comments the verdict and stops. No green-washing;bounded iteration cap.
.claude/agents/—bug-verifier(reproduction gate, withpython/pytest/pip),context-builder,fix-planner,fix-implementer,fix-verifier(all Opus; effort viarole prompts).
.github/workflows/claude.yml— the@claudepath now allows theTasktool +read tools and routes issue fix/feature requests through the verification-gate +
multi-agent pipeline. Existing auto-review handling (review threads, workflow-permission
guard, web tools) preserved unchanged.
Removed (old auto-fix mechanics)
.github/workflows/claude-verify-bug.ymland.github/workflows/claude-fix-bug.yml— the label-triggered read-only-verify → dispatch-fix workflows (from Auto-fix bug-labeled issues: verify, fix, open PR #121).
dispatch-verifyjob inclaude-issue-triage.yml— triage now only applieslabels and never auto-triggers a fix.
There is no longer any label-triggered auto-fix;
@claudeis the only trigger.Notes
main; the standalone webhook bot from an earlier revision of thisbranch was removed from the PR.
🤖 Generated with Claude Code