Imported from zhongjis/agent-skills (
skills/AGENTS.md). Install upstream withnpx skills add zhongjis/agent-skills --skill skills. Copyright stays with the author.
skills
Purpose
- Canonical authored and adapted skill trees.
Ownership
- Each direct child directory is a category grouping (
engineering/,tools/,productivity/,misc/,in-progress/); each skill leaf is one complete directory nested one level under its category. - Root
skill-harnesses.nixowns sparse logical harness routing. - Root
skill-selection.nixowns global selection exclusions.
Local Contracts
- Preserve each whole tree, including scripts, references, assets, licenses, fixtures, and test material.
- Skills absent from
skill-harnesses.nixare logical common. - Routed skills belong only to their named logical harness, not common.
- Authored/adapted leaves have no
skills-lock.jsonentry. - Excluded authored leaves remain canonical here and may have one matching relative project projection at
.agents/skills/<name>→../../skills/<category>/<name>.
Work Guidance
- Move whole directories; keep only the explicit project projections owned by
skill-selection.nix; do not create compatibility copies or symlinks in old roots. - Keep skill names, profile membership, global exclusion, logical routing, and project projections aligned.
engineering/research/owns ordinary primary-source research and explicitly opted-in deep workflow routing; deep findings require atomic evidence and verified coverage, failures remain visible, and research or workflow execution does not authorize persistence, rendering, or publication.code-review/SKILL.mdand its context-gathering / parallel-axes references own scope-matched local reviews, shared dependency-contract evidence, and scope-grounded requirements; preserve these contracts across review passes.herdr-bulk-action/SKILL.mdowns user-invoked, explicitly confirmed generic Herdr batches: referenced task behavior with user overrides, task-needed isolation, and a bounded lifecycle; itsevals/evals.jsonowns focused scenarios. Keepherdr-bulk-reviewseparate.engineering/delegate-using-herdr/SKILL.mdowns single-task, user-invoked issue-tracker-to-Herdr handoff: tracker resolution (supplied URL, named tracker, repository convention, otherwise one question), latest-intent handoff-contract agreement (including a diff-size budget with a stop-and-report point and the issue's transitional code and seams), supplied-issue reuse or creation, parent-source worktree resolution, same-kind agent dispatch (Pi worker prompts always end with literalulw), and default immediate handoff without parent monitoring or integration. Its naming contract keeps repository branch conventions distinct from lowercase issue-slug worktree basenames (numeric identifiers prefixedgh-), friendly labels, and unique ≤32-character lowercase-hyphen live agent names without session/tab mutation; flag-specific worker contracts live inreferences/review.md(--review, fail-closed local Claude review in a sibling side pane before delivery) andreferences/auto-delivery.md(--auto, fail-closed remote delivery through merge and tracker-native closeout). The parent reads each selected reference and embeds its full contract before dispatch, with local review before publication and remote gates unchanged; itsevals/evals.jsonowns focused delegation scenarios.engineering/pr/owns evidence-backed PR-body drafting and revision adapted from Matt Pocock's upstreampr: explicit user formatting before template/body precedence, a required Summary visual, collapsed verification with merge-critical gaps visible under risks, omission of inapplicable visual evidence,<MISSING_EVIDENCE>for unobserved proof, grouped issue/document/report/companion-PR references with- N/Aonly when none exist, an issue-only[<primary item ID>]title prefix, and focused eval scenarios;tools/gh/owns generic GitHub CLI, authentication, PR creation, comment, and review operations; installedbefore-and-afterowns specialized user-visible UI media attachment, usingagent-browserfor capture.engineering/retro/is adapted from Matt Pocock's upstreamretro; it locates session logs and steering files through the global Agent Environment instruction (nix-configcommon/instructions/personal) and names an owning source for each finding.engineering/programming/owns code and test quality (testing policy, mocking ladder, test anti-patterns) and routes test-first work one-way toengineering/tdd/.engineering/tdd/is adapted from Matt Pocock's upstreamtddand owns only test-first sequencing: seams agreed once per task (revisited on scope or interface change), vertical slices, and red → green with refactoring deferred to review; it must not referenceprogrammingor restate test-quality policy.engineering/address-comments/defaults an invocation without a PR identifier to the open PR associated with the current branch; an explicit PR number or URL selects and checks out that PR first. Default mode executes its scoped plan and amendments without approval; explicit--manualrequires approval for listed local work and named remote actions while retaining the same safety and completion gates. It owns bounded required-check handling: replies follow publication, each push gets one 5-minute required-check window, every failure introduced relative to the PR base is repaired, and evidenced base/pre-existing, flaky, or infrastructure failures stay open. Reply/review writes are serialized; prefer REST root-comment replies and verify submitted comments, non-pending parent reviews, and intended reply targets. Draft publication is limited to authorized workflow-owned reviews withCOMMENT; non-thread replies exclude findings already represented inline. Threads replied to in a pass stay unresolved for the reviewer; onlysettledprior-round threads are resolved, where the PR author replied last and the same reviewer has later PR activity.
Verification
- Run
nix eval --file tests/selector.nixand full checks intests/AGENTS.md.
Child DOX Index
- Other nested AGENTS.md files under skill fixtures are test material unless indexed here.
