dev-invoke-claude-code

name	dev_invoke_claude-code
description	Codex-native adapter for invoking Anthropic Claude Code CLI as an independent read-only QAQC reviewer using markdown file handoff. Use when the user explicitly asks Codex to ask Claude, Claude Code, Opus, or Anthropic for a QAQC/code review, second-model review, or independent review. Do not trigger for generic QAQC unless the user asks for Claude specifically. Triggers: claude review, claude code review, ask claude, invoke claude, claude qaqc, opus review, anthropic review, second model with claude Prerequisites: Claude Code CLI authenticated and available as `claude`. Default model: opus with xhigh effort for deep reviews; use sonnet/high for faster or lower-cost review passes.
shared_corpus	false
harness_scope	codex_only
source_owner	gpt-cmdr
security_review	internal

Invoking Claude Code For QAQC

Codex-native provider handoff skill. This is not a shared domain workflow and should not be mirrored into .claude/skills/. It lets Codex request an independent Claude Code review while keeping shared repo policy in AGENTS.md.

Use this skill when the user explicitly asks Codex to have Claude Code review code, documentation, notebooks, or agent-framework changes.

Safety Rules

• Use Claude as a read-only reviewer by default.
• Do not let Claude edit files, commit, push, install tools, or mutate the worktree.
• Do not overwrite root-level TASK.md or OUTPUT.md; those may belong to another handoff.
• Put each review in an ignored workspace such as working/agent-reviews/claude-code/{timestamp}-{slug}/.
• Give Claude the same shared repo contract Codex used: nearest AGENTS.md, relevant local AGENTS.md files, and task-specific files.
• Do not use --dangerously-skip-permissions unless the user explicitly requests it and the workspace is externally sandboxed.

Handoff Pattern

Create a review workspace:

$reviewDir = "working/agent-reviews/claude-code/20260426-claude-qaqc"
New-Item -ItemType Directory -Force $reviewDir | Out-Null

Write TASK.md in that folder. Include:

• the review target files or directories
• the exact concern or review objective
• the shared instruction files Claude should read first
• a strict read-only constraint
• the expected finding format

Minimum TASK.md shape:

# Claude Code QAQC Review

You are performing an independent read-only QAQC review for ras-commander.

## Read First

1. AGENTS.md
2. The nearest local AGENTS.md files for the target paths
3. docs/development/multi-harness-agent-contract.md when reviewing agent infrastructure

## Target

- path/to/file_or_directory

## Review Focus

- correctness bugs
- behavioral regressions
- missing tests or validation gaps
- security or supply-chain risks, if relevant

## Constraints

- Do not edit files.
- Do not run destructive commands.
- Do not commit or push.
- Write findings first, ordered by severity.
- Include file and line references when possible.

Invoke Claude Code non-interactively and capture its report:

claude --print `
  --model opus `
  --effort xhigh `
  --permission-mode auto `
  --disallowedTools "Edit,Write,MultiEdit,NotebookEdit" `
  --output-format text `
  --append-system-prompt "Read-only QAQC review. Do not edit files. Return only the review report." `
  "Read $reviewDir/TASK.md, perform the review, and write the final report to stdout." `
  > "$reviewDir/OUTPUT.md"

Use sonnet with --effort high when the user wants a faster review:

claude --print `
  --model sonnet `
  --effort high `
  --permission-mode auto `
  --disallowedTools "Edit,Write,MultiEdit,NotebookEdit" `
  --output-format text `
  --append-system-prompt "Read-only QAQC review. Do not edit files. Return only the review report." `
  "Read $reviewDir/TASK.md, perform the review, and write the final report to stdout." `
  > "$reviewDir/OUTPUT.md"

Output Handling

After Claude returns:

1. Read OUTPUT.md.
2. Validate whether each finding is actionable and grounded in the repo.
3. Do not blindly apply Claude's recommendations.
4. Summarize Claude's findings to the user and distinguish:
   • confirmed issues
   • plausible issues needing local verification
   • items you reject after inspection

Optional Session Capture

If session metadata is needed, run with JSON output and save the raw result:

claude --print `
  --model opus `
  --effort xhigh `
  --permission-mode auto `
  --disallowedTools "Edit,Write,MultiEdit,NotebookEdit" `
  --output-format json `
  "Read $reviewDir/TASK.md and perform the read-only QAQC review." `
  > "$reviewDir/RUN.json"

If using JSON output, extract the review text into OUTPUT.md before presenting the result.

Explicit Opt-In: Claude Ultrareview

Claude Code also exposes claude ultrareview. Treat that as a separate, explicit opt-in path because it is cloud-hosted and branch/PR oriented.

Use it only when the user asks for Claude Ultrareview or a cloud-hosted Claude review of the current branch/PR. Do not substitute it for the default local markdown handoff above.

Failure Handling

• If claude is not found, report that Claude Code CLI is not installed or not on PATH.
• If authentication fails, ask the user to authenticate Claude Code.
• If Claude requests edit permissions, stop and rerun with a stricter read-only prompt rather than granting broad permissions.
• If Claude's review is empty or generic, tighten TASK.md with explicit files, line ranges, and review questions.