docketSAFE

From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: docket
description: Use when adopting, repairing, auditing or comparing a repository's agent workflow (queues, worktrees, paired review, gates, review ledger); when coordinating a multi-agent fan-out or an unattended run; when running a full or incremental codebase review; when triaging the queues; or when cleaning up after PRs land. Not needed for ordinary single-task work in a repository whose AGENTS.md and guides already encode the workflow; those documents are sufficient there.
---

# Docket

A project run this way answers four questions from its own files at any moment: what is being worked on, what is ready to work on, what was decided, and what is known to be broken. Everything below exists to keep those answers true while many agents work at once.

**Core principle:** every unit of work is a branch, a worktree, a draft PR with an exact claim, an independent review, and a cleanup. Every idea that is not that unit of work goes into a queue, immediately, and the current work continues. Two things may go straight to `main` without a branch: documentation that records work to be done (queue entries, triage, plans, design specs; these may also ride along in whatever PR they came up in, which does not then do that work), and housekeeping metadata files when the user says so for that case; see [pr-and-landing](references/pr-and-landing.md#direct-to-main).

The repository's own `AGENTS.md` and guides win wherever they exist and are what an agent reads for everyday work; this skill is for the operations below, and for installing those guides where they are missing (see `references/setup.md`). Loading it for an ordinary implementation task in an already-set-up repository spends context on rules the repository already states.

## Quick start

The user needs only these phrases; this skill works out the rest by inspecting the repository first.

| The user says | What it means |
| --- | --- |
| "Use docket to set up or audit this project" | Inspect first. A project with a contract gets an audit and repairs on a branch; one without gets the workflow installed; an empty one gets initialised before any product code. Propose hosting settings, never apply them. If the repository records a policy that conflicts with the default model (for example one queue, no markers), report the conflict and ask which way to go before changing it. |
| "Use docket to do a full / incremental review" | The codebase review procedure in [reviews](references/reviews.md), landed as its own PR. |
| "Use docket to fan out the ready TODOs" / "run overnight" | The coordinator procedure in [parallel-work](references/parallel-work.md), with the autonomy mode stated up front. |
| "Grab a TODO" / any single task | Not this skill. The repository's `AGENTS.md` and guides. |

## Choose the operation

| Operation | Read | Tools |
| --- | --- | --- |
| Adopt, repair, audit or compare the workflow in a repository | [setup](references/setup.md), then the pillar references it points to | `install.sh [--check]`, `check-queues`, `check-links`, `review-due`, `gh api` protection inspection |
| Coordinate a fan-out of writers and reviewers, or an unattended run | [parallel-work](references/parallel-work.md), [pr-and-landing](references/pr-and-landing.md) | `claim-check`, `start-work`, `check-pr-markers`, `pr-labels` |
| Triage the queues, split a batch, or reconcile claims | [queues](references/queues.md) | `check-queues`, `claim-check` |
| Run a full or incremental codebase review, or reconcile finding closure | [reviews](references/reviews.md) | `review-due` |
| Define or repair gates, respond to a red main, change protection | [quality-gates](references/quality-gates.md) | `ci.sh`, templates under `templates/.github/` |
| Land a stack, clean up after merges | [pr-and-landing](references/pr-and-landing.md) | `cleanup-landed` |
| Implement one queued or requested task in a set-up repository | the repository's `AGENTS.md`, `docket/CONTRACT.md` and guides; not this skill | the repository's `docket/scripts/` |

## Five pillars

| Pillar | Non-negotiable | Reference |
| --- | --- | --- |
| Parallel work | One owner, one branch, one worktree per task. Shared interfaces land before dependents start. The coordinator integrates and serialises queue edits. | [parallel-work](references/parallel-work.md) |
| Paired review | Every code-writing agent gets a separate reviewer agent before the PR is called ready. The reviewer verifies claims, the author fixes, a PR comment records who reviewed what. | [parallel-work](references/parallel-work.md#writer-and-reviewer-pairs) |
| Queues, not tangents | Discoveries become queue entries with a source, in the repository that owns them, then the current task resumes. Ready means unblocked. Claims are exact slugs in draft PRs. | [queues](references/queues.md) |
| Gates everywhere | Fast hooks, full CI on PRs and on pushes to main, scheduled extended runs on main, protection that makes the user's merge safe rather than ceremonial. Empty suites never pass. | [quality-gates](references/quality-gates.md) |
| Review ledger | Full reviews record a revision and findings; incremental reviews start from the last revision and verify claimed fixes. Findings are promoted deliberately into a review backlog. | [reviews](references/reviews.md) |

PRs, stacking, landing and cleanup are in [pr-and-landing](references/pr-and-landing.md). Installing or adapting the system is in [setup](references/setup.md).

## The unit of work, start to finish

Script paths below are the installed paths (`docket/scripts/`), the only ones anyone runs; the skill's `templates/docket/scripts/` copies reach a repository only through `install.sh` (see Bundled tools).

1. **Select.** Follow the user's selection, or take the highest-priority unclaimed entry from the named queue in its Ready section. Never switch queues or stages because the requested one is empty; report that instead.
2. **Check claims.** `bash docket/scripts/claim-check.sh <slug>`

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (1)

.aiignore

Gates applied: no_behavioural_pass.