In Herdr, initiate an independent, read-only Codex reviewer, with the caller verifying locatable review conclusions.
Router Rules
- Use only when Herdr + Codex review is explicitly requested, or when this skill invocation already has that task context. Ordinary correctness review goes to the existing review process; maintainability issues follow
code-quality-review; Codex collaboration outside Herdr must not go through this skill. - If already assigned as a leaf reviewer, complete the review directly, without invoking this skill, orchestra, or more Agents. Discussing/researching these skills does not authorize starting processes.
- Read its
herdr-orchestravia the loadedreferences/delegation.mdname/actual path; it is the sole process control owner. This package does not allocate panes, does not repeat environment probing, and does not start Codex directly. When dependencies are missing, report the source pathskills/developer-tools-integrations/herdr-orchestra/and the incomplete items; do not bypass Herdr. - Verify the identity of the currently loaded package; a third-party entry with the same name is not this source. If an old entry containing an automatic CHANGELOG is still loaded, report the mismatch and do not modify the installation, global configuration, or hooks. For host loading boundaries, see the repository
docs/harnesses.md。
Compact Workflow
- Per the review contract , specify the cwd, user standards, and one review scope, and capture HEAD/ref, diff, and relevant input content. By default, all current changes include staged, unstaged, and untracked; when the user specifies staged-only, preserve that boundary.
- Prepare the six-field text handoff and the complete leaf prompt. Using the Codex CLI boundary 's interactive root parameter
--sandbox read-only --ask-for-approval never --no-alt-screen, request a newkind=codexworker; verify the current help at actual execution time. If the user specifies a model, preserve it; otherwise inherit the configuration. - Hand it to orchestra for one dispatch and collection. The prompt is task text, not a shell command or
/reviewUI operation; a startup failure must not switch toexec reviewor another reviewer. - Verify identity, target, lifecycle, and complete response against the receipt. When truncated, only orchestra requests the missing small segment from the same ready worker; the read-only reviewer does not write temp files or expand the sandbox. If recovery is incomplete, keep it incomplete.
- Before acceptance, recheck the input and finding anchors. If the target changes, mark scope drift; only verify conclusions when complete and scope-consistent. Output accepted findings, important excluded false positives, and uncertainties; do not auto-fix or recursively re-review.
Output Contract
- Return the actual reviewed cwd/scope/input identity, worker identity/lifecycle, response coverage and completeness, and the caller's scope recheck result.
- List findings by priority with path/line, trigger scenario, impact, evidence, and verification gaps. Insufficient evidence goes under uncertainty; no actionable findings must have complete review basis.
- Separate passed, failed, not-run checks, and missing evidence. blocked, timeout, unknown, empty response, missing segments, and scope drift must not be written as passed; leftover worker resource status comes from the orchestra receipt.
- Save captured text only within the caller's existing artifact write permissions; do not extend review authorization to fixes, CHANGELOG, Git write operations, installation, trust/config/hooks changes, or resource cleanup.
In-package resources are resolved by the actual directory published when this skill is loaded;<skill-dir> denotes that directory, and relative positions between installation directories are not assumed. For behavioral fixtures see evals, for independent lexical input see trigger cases。Create handoff summarizes sources, actual checks, and unverified boundaries; these materials do not prove that a new session has loaded this package.