Skills

What to ask for next

After harness-init, the usual path is:

1. code-spec — agree on the change and how to verify it.
2. code-plan — approve the implementation steps.
3. code-execute — make the change and run the checks.
4. code-commit — commit the result when you are ready.

Commits stay manual by default. code-execute can commit when you explicitly enable automatic commits; see the config and requirements.

For a one-file change, you can ask for code-execute directly and skip the spec and plan. Validation still applies.

Full inventory

Eighteen skills are available under .agents/skills/. setup always installs the nine core skills (harness-init, adr, grilling, pragmatic-guard, qmd, research, code-spec, code-plan, code-execute). The ship set (code-commit, code-pr, code-review, code-ci) and the discovery set (create-prd, prototype, handoff, the-fool, the-jury) are optional. A fresh interactive setup preselects ship; a fresh non-interactive setup defaults to ship. Select --skills discovery or --skills ship,discovery to add sets, or --skills none for core skills only. Re-runs never remove installed files or sets. The four marked ask-only require an explicit request. PR creation, review, and CI work never merge the PR.

grilling (core) interviews toward a settled plan. the-fool challenges without forcing a decision. the-jury spawns five isolated juror agents, then deliberates with shared first-round opinions, and returns a dissent-preserving advisory verdict with confidence and a concrete test. Keep the three distinct.

Skill	Description
adr	Create, promote, and supersede decision records under docs/adrs/; repoint claims when one supersedes another.
code-ci (ask-only)	Drive a PR to merge-ready: conflicts, then comments, then failing checks. Explicit ask; never merge.
code-commit	Create a Conventional Commit when asked, or when code-execute meets the automatic commit requirements above.
code-execute	Implement an approved plan and run checks. Commit behavior follows the config described above.
code-plan	Turn an approved spec into ordered tasks, with dependencies, check points, and an explicit choice to work inline or use a subagent.
code-pr (ask-only)	Push the branch and open or amend a PR from the body template. Ask-only; never merges.
code-review (ask-only)	Review an open PR against the refs it cites; post blocking or nit findings. Never merges.
code-spec	Write a spec from a product requirements document (PRD) or confirmed request, pairing each requirement with a way to verify it.
create-prd	Clarify the problem, draft a lean PRD, and mark it stable only on approval.
grilling	Interview one question at a time until every open branch of a plan, decision, or idea is settled.
handoff (ask-only)	Save a handoff document to the OS temp directory for a fresh session. Ask-only; never automatic.
harness-init	Fold WOLVEN.md into AGENTS.md, migrate legacy ADRs, discover, stub skills, and write a session note.
pragmatic-guard	Challenge unnecessary scope and record deferred work with a trigger for revisiting it under docs/deferrals/.
prototype	Build a throwaway prototype that answers one question, then discard or promote it deliberately.
qmd	Search local markdown notes, docs, and wikis with QMD, the local search tool the agent uses; retrieve documents or set up access.
research	Investigate against primary sources, cite every claim, and land the answer as an in-repo note.
the-fool	Challenge an idea or plan with structured critique and synthesis after the user responds; does not force a decision.
the-jury	Spawn five distinct juror agents for a parallel isolated first round, deliberate with shared opinions, preserve dissent, and deliver an advisory verdict with confidence and one concrete test.

Rules applied each session

Three rules live under .agents/rules/:

• comments defines the style checked by wolven-harness comments.
• qmd-first directs the agent to search the repo’s documents before answering from memory or the web.
• yagni-strict directs the agent to challenge unnecessary scope and record deferred work under docs/deferrals/.