Agent skill catalog

See the small state transitions that pui-dev and pui-maintain compose without preloading the full library.

The machine registry at internal/agent-operations/skills.yaml is the source for paths, capability bands, task classes, inputs, and outputs. This page explains the same library for people.

How loading works

The entrypoint chooses one leaf ID. The resolver returns its one registered file:

Terminal

pnpm agent:skill -- pui-trace --mode human-assisted --mode-source current-user

The leaf returns a structured handoff. The entrypoint validates it before resolving the next leaf:

Terminal

pnpm agent:skill -- --handoff <handoff.json>

A handoff carries typed artifacts and at most one next skill. A terminal handoff ends the chain. Leaf skills cannot load other leaves themselves.

Enter and bound the work

Skill	One transition
`pui-assess`	Unassessed context to an unsigned U0-C4 local task-fit result
`pui-agent-identify`	Non-exempt request context to an unsigned closed-set ModelTrace record
`pui-orient`	Unknown context to a mode-aware contribution envelope
`pui-select`	Unbounded request to one read-only work proposal or no-work result
`pui-claim`	Authorized proposal to one posted claim or exact re-selection result
`pui-unclaim`	One owned claim to a recorded release
`pui-trace`	Bounded subject to an authority and evidence map

Shape governed artifacts

Skill	One transition
`pui-brainstorm`	Unresolved product direction to the smallest decision packet
`pui-spec`	Traced semantic scope to a governed spec graph change
`pui-contract`	Governed fact to a readable contract projection

Implement one owning slice

Skill	One transition
`pui-module`	Bounded Module slice to portable implementation and evidence
`pui-host`	Traced host responsibility to a Host Capability realization
`pui-adapter-assess`	Bounded Adapter question to a read-only assessment packet
`pui-adapter`	Bounded Adapter scope to one implementation slice
`pui-prototype`	Governed component slice to one coherent Prototype delivery
`pui-regression`	Reproduced governed failure to a bounded repair and regression evidence

Build evidence and reader projections

Skill	One transition
`pui-test`	Governed behavior to executable evidence
`pui-docs`	Governed repository fact to human documentation
`pui-validate`	Candidate change to a proportional evidence report
`pui-review`	Candidate exact head to an independent revision-bound packet at one declared review class and an authorized review receipt
`pui-evidence-publish`	Prepared, separately authorized Issue evidence to one additive comment receipt; no uploads or metadata changes
`pui-integrate`	Independently approved exact head to a rule-compliant merge receipt
`pui-collaborate`	Governed live collaboration state to one verified reversible mutation receipt

Inspect and continue repository operations

The inspection leaves are read-only intake and diagnosis transitions. Their credential boundary protects observation; it does not cap the follow-on skill chain. pui-collaborate then applies one exact-target reversible metadata, update-branch, ready-for-review, thread, review-request, status-comment, or CI-recheck action under current-user or standing authorization. Its registered runtime verifies a purpose-bound request digest and fresh live state, performs zero or one mutation, reconciles an unknown outcome once without retrying, and returns a schema-validated receipt. Review disposition and merge retain their separate exact-head leaves.

Skill	One transition
`pui-issue`	Bounded Issue portfolio to an operations report
`pui-pr`	Bounded pull-request portfolio to an integration-state report
`pui-ci`	One workflow failure to an owning evidence map
`pui-govern`	One collaboration question to a governance drift report
`pui-deploy`	One delivery surface to a revision-bound evidence report
`pui-deps`	One dependency question to an impact and risk report
`pui-dependency-update`	Governed dependency report to a bounded manifest and lockfile update
`pui-package-budget`	Accepted capability cost evidence to a bounded numeric package-budget transaction; preserves the blocking gate and independent review

Prepare and audit a release

Skill	One transition
`pui-release-prep`	Explicit release intent to a reviewable candidate state
`pui-release-audit`	Completed publication to reconciled immutable evidence

Run autonomous maintenance

pui-maintain routes one stage at a time. Observation, verification, and review use fresh contexts where the protocol requires independence.

Skill	One transition
`pui-mission`	Selected candidate to a frozen bounded mission and lease
`pui-observe`	Frozen mission to a finding or no-finding report
`pui-verify`	Candidate finding to an independent classification
`pui-remediate`	Accepted finding to the bounded repair and remediation packet
`pui-maintenance-review`	Actual remediation to an independent technical verdict
`pui-maintenance-close`	Reviewed remediation to synchronized closure
`pui-record`	Supported non-remediation terminal outcome to a synchronized run record

dot’s owner-authorized exception

The owner’s 2026-10-06 dot exception replaces only ModelTrace-specific sampling and record prerequisites with the exact visible not-measured dot declaration. dot uses the explicit publisher flags or a real connected-service path with equivalent live checks; it must not fabricate a modeltrace-record, authentication or a passed handoff validator. All other authorization, exact-state, DCO, CI and independent-review conditions remain. Contributor Agents gives the declaration, command syntax and limits. The measured skill description below applies to non-exempt Agents.

Capability is a ceiling when the Agent works alone

pui-agent-identify is U0, public-read, and disposable-output-only: request-context in, modeltrace-record out. It uses genuine active-model literal sampling and pinned offline scoring, not a system/harness label fallback. Carry its exact private reference and canonical public-receipt digest through handoffs. Every non-exempt Agent-originated commit or Issue/PR creation, material update, comment, review or current collaboration write requires a current record in both modes; local edits, read-only review and verification do not. Humans and deterministic non-LLM automation need no model test. An independent model context measures itself. ModelTrace is unsigned closed-set attribution, not authenticated identity, permission, independent review, acceptance or activation of pending scheduled scopes. See Contributor Agents for disclosure and freshness.

C1 covers bounded facts plus factual or documentation review. C2 adds test and bounded-regression review plus exact-target collaboration mutations such as review submission and exact-head integration. C3 adds bounded semantic implementation and governed-slice review. C4 adds cross-domain semantics, governance judgment, release-evidence review, and release preparation. pui-review declares one content review class from C1 through C4; the review and merge write primitives themselves use the C2 exact-target mutation floor. The registry calls a leaf’s threshold autonomousMinimumBand because it applies when an Agent chooses or advances work alone.

During human-assisted work, the same result is advice. It changes scope, validation, review depth, and stated limitations, but it does not block a task the current user requested. In autonomous work, an eligible independent Agent can continue through recheck, finding-backed REQUEST_CHANGES, clean exact-head APPROVE, and exact-head merge when active standing authorization, trusted evidence, live platform permission, review state, and repository rules all agree. Human escalation is reserved for unresolved product direction and privileged or irreversible operations; Discord or Poppy trust still applies only to the surfaces they control.