Claude Subagent

Claude Subagent

Use Claude Code as a one-shot, read-only reviewer. Claude runs in the command's working directory, so it can inspect that workspace directly.

Run a review

1. Define one bounded, read-only review assignment. Include the goal, relevant paths, known constraints, expected deliverable, and non-file conversation context Claude cannot infer.
2. Explicitly prohibit file changes and external-system changes. Ask Claude to inspect primary artifacts and return findings, evidence, disagreements, or a recommendation for Codex to verify and synthesize.
3. Put the prompt in a scratch file outside the repository. Invoke the scripts/claude-subagent helper beside this SKILL.md from the target working directory, piping the prompt on stdin. The helper runs the user's existing c function in non-interactive print mode. Always pass an effort: use --effort high by default, and use --effort xhigh for difficult problems. Do not use other effort levels.
4. Start one call and wait on that same process. If the command tool yields a running-session identifier, resume that session; do not submit a duplicate merely because Claude has not produced output yet.
5. Treat stdout as Claude's response and judge success by the exit code, not by anything on stderr. A zero exit with non-empty stdout is a completed call. On a non-zero exit, report the failure and inspect the workspace before deciding whether a retry is safe. A failed delegation may be retried at most twice, for no more than three attempts total. Each retry must follow that inspection; never retry blindly after a task that may have changed files or external state. Count an attempt as failed only after the process exits non-zero or exits without a response on stdout. The unrecognized_model stderr diagnostic described below is not a failed attempt and must not trigger a retry; keep waiting on the same process.

The unrecognized_model stderr line is NOT a failure

The relay drives a private model id (model_hub/es1_orange_o48) that the Claude CLI doesn't have in its built-in table, so it prints one diagnostic line to stderr:

[claude-code:unrecognized_model] {"model":"model_hub/es1_orange_o48","query_source":"generate_session_title"}

This is cosmetic and expected. The real request still goes out and returns normally — query_source is usually generate_session_title (the title side-channel, not your task). It cannot be silenced via modelOverrides (that mapping only accepts real Anthropic model ids and ignores unknown keys), so do not chase it. Never treat this line, or any stderr text, as proof the request failed, and never conclude "no review body" from it — read stdout.

To avoid parsing stderr entirely, capture the two streams separately, or ask for structured output and read the result field only when is_error is false:

claude-subagent < /tmp/claude-subagent-prompt.md >/tmp/out.txt 2>/tmp/err.txt   # body in out.txt
claude-subagent --output-format json < /tmp/claude-subagent-prompt.md           # {"is_error":false,...,"result":"..."}

The helper already pins a 1M context window, so long prompts and large repository context are fine.

Example command shape:

"${CODEX_HOME:-$HOME/.codex}/skills/claude-subagent/scripts/claude-subagent" --effort high < /tmp/claude-subagent-prompt.md

Pass ordinary Claude CLI flags after the helper name when useful, for example --model opus, --tools "", or --output-format json. Always pass either --effort high or --effort xhigh as described above.

Safety and verification

The helper deliberately inherits c's --dangerously-skip-permissions. This does not broaden the user's authorization: keep Claude inside the current task, do not pass secrets unnecessarily, and always state the read-only boundary in the prompt. Never use Claude as the writer, implementer, or external-system operator through this skill, even though the underlying helper has those technical capabilities.

After Claude returns, verify consequential claims and inspect any changes or tests yourself. Present Claude's result as a delegated contribution, not as independently confirmed fact.