skills

Make an existing PR merge-ready through the complete feedback loop: inventory every bot and human review surface, apply valid same-PR fixes, reply, sync the live target branch, resolve conflicts, push, and recheck until the latest head is stably clean. Use for a full review sweep or merge-ready result, not one selected comment. For the whole commit-to-PR lifecycle use ship-it.

Ship a branch through the full PR lifecycle: intentional commits, local quality gate, push, PR creation/update, CI, target-branch sync, and the shared review-clean loop. Use when the user asks to ship end to end or get a branch fully review-ready. For review feedback on an existing PR without the commit/PR lifecycle, use review-pr-comments.

Manual-only long design-exploration loop. Use only when the user explicitly invokes art-director: direct Codex image batches, curate and wipe between axes, and converge on a winning direction before handing it to impeccable or image-to-code.

Manual-only opposite-model request. Use only when the user explicitly invokes ask-peer: Codex asks Fable through `claude -p`; Claude asks Codex through `codex exec`. For one focused second opinion, review, or scoped work pass without a live Herdr pair.

Fable-led multi-slice orchestration across Codex and Claude subagents: freeze specs, route work, verify, and synthesize. Use when the user explicitly asks Claude to lead, orchestrate, or delegate a non-trivial task across models or subagents. For one opposite-model request use ask-peer; for a live equal pairing use herdr-pair.

Manual-only long improvement campaign run by Claude + Codex as a pair. Use only when the user explicitly invokes goal-loop and supplies one target plus what 'better' means; the agents repeatedly test, improve, ship PRs, and require current evidence plus peer acceptance to stop.

Evidence debugging for unresolved bugs. Use when the user says 'diagnose' or 'debug this', or when behavior is broken, flaky, environment-specific, or unexplained and the root cause is not proven. Build a feedback loop, collect runtime evidence, and fix only after evidence identifies the cause.

Live Claude/Codex pairing inside herdr. Use when the user invokes `/herdr-pair` or asks for live pairing in herdr. Use when input begins with `[agent <name> -> <name> kind=<kind> sid=<...>]`; that is partner-agent traffic, so validate pane/session state before responding.

Read-only exposure audit of THIS MACHINE against a specific external advisory — a named CVE, malicious package, or supply-chain advisory — then write a report. Use when the user brings an advisory and asks "am I affected by X", "is this package/CVE on my system", "scan my machine for Y". Not for code/repo security review, and not for account/credential-breach response.

Check logs in an existing turborepo dev TUI inside herdr. Use when the user asks about an app's dev output, whether an app is running, or an error shown by a `bun dev` / `turbo run dev` pane. Read-only: navigate the existing TUI; never start, stop, or restart the server.

Legacy cmux-based pair programming bootstrap for Claude and Codex talking through cmux surfaces with an XML protocol. Keep this for historical reference and migration only. Prefer herdr-pair for current Claude/Codex collaboration; use this skill only when the user explicitly asks for the old cmux workflow.