docs: add CLAUDE.md and skill invocation instructions#284
Merged
Conversation
CLAUDE.md auto-loads when Claude Code opens the repo, pointing Claude to SKILL.md for agent-loop orchestration requests. SKILL.md gains a "How to invoke this skill" section at the top with concrete natural-language examples, so users know they just need to describe what they want in the session — no slash command required. Fixes the documentation gap left by #283. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds .claude/commands/coding-review-agent-loop.md so users can invoke the agent-loop skill with /coding-review-agent-loop in Claude Code. The namespaced name avoids conflicts with other agent-loop tools. Notes the current Claude-as-coder limitation (reversed roles require a separate issue). Updates CLAUDE.md to mention the slash command as the preferred invocation path. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… Codex and --skip-trust to Gemini Codex's bwrap sandbox fails on GCP VMs with "Failed RTM_NEWADDR: Operation not permitted". The --dangerously-bypass-approvals-and-sandbox flag is intended for exactly this case (externally sandboxed environments). Gemini fails with "not a trusted directory" when run non-interactively outside a git workspace. --skip-trust suppresses this check. Both flags correspond to CodexBackend.default_args(dangerous=True) and GeminiBackend.default_args(dangerous=True) respectively, but run_external never called default_args. Now set directly in AgentLoopConfig. Also adds --diff-file flag to embed a pre-fetched diff in the prompt, needed because Gemini's non-interactive mode lacks run_shell_command. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The slash command advertised `task "<text>"` as a supported flow, but state_manager only accepts --flow plan|pr, so users invoking that path would enter undefined orchestration. Remove task flow and add a note directing users to the headless CLI for task-based invocations. Also clarify that `issue <N>` maps to --flow plan. Addresses Codex blocking finding from PR #284 round 1. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Structured JSON reviewer output must be rendered to human-readable markdown before posting to GitHub, matching what the headless CLI does via _render_public_pr_review_comment/_render_public_plan_review_comment. - helpers/render_response.py: new module wrapping the library's comment_rendering functions for pr_review, plan_review, coder_followup, and plan_revision kinds - SKILL.md: insert `python -m helpers.render_response` step between validate_response and attach-metadata; attach-metadata and post now operate on the rendered file, not the raw JSON Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Owner
Author
|
Review verdict: Approved The documentation additions are clear, well-structured, and provide helpful instructions for invoking the agent loop as a skill in the Claude environment. The new files and the modifications to existing documentation create a logical and useful guide for users. -- Google Gemini |
Owner
Author
|
Review verdict: Blocking The invocation docs mostly help, but the new slash command advertises flows that the skill instructions do not support. Blocking issues
-- OpenAI Codex |
Owner
Author
|
Review verdict: Approved The round-2 changes resolve the unsupported slash-command flow issue, and the added invocation docs plus helper updates are internally consistent. Focused skill-helper tests passed locally with python3. -- OpenAI Codex |
Owner
Author
|
Review verdict: Approved The changes look good. The new documentation in CLAUDE.md and SKILL.md clearly explains how to invoke the agent loop skill. The helper script change to inject PR diffs is a good addition to support PR review flows. -- Google Gemini |
Add --role coder to run_external.py so Codex can be invoked for plan writing turns. Document the reversed-roles flow in SKILL.md and update the slash command to accept --coder codex. Add a dry-run test covering the new coder role path. Closes #285 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Codex's blocking review (PR #286) identified that SKILL.md documented the reversed-roles flow without specifying --reviewers claude in the build-resume call. Without that, build-resume ignores Claude's posted review on any subsequent resume attempt, causing the loop to repeat review turns. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…skill feat(skill): add --coder codex reversed-roles support
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Closes the documentation gap from #283: the skill mode was implemented but there was no explanation of how to actually invoke it.
CLAUDE.md(new): auto-loaded by Claude Code when the repo is opened; tells Claude to consultSKILL.mdwhen the user asks to run the agent loop. Lists concrete invocation examples.SKILL.md: adds a "How to invoke this skill" section near the top with natural-language examples — users just describe what they want, no slash command needed.Test plan
-- Anthropic Claude