AGENT SKILL

pmo-delivery

Use for delivery planning, handoff, task source of truth, engineering sequencing, and semantic handoff review.

License 许可
MIT
最近更新
2026-10-14
Stars
★ 3
Forks
1

01 前言 Front Matter

name
pmo-delivery
description
Use for delivery planning, handoff, task source of truth, engineering sequencing, and semantic handoff review.

03 SKILL.md 正文

pmo-delivery

Purpose

Translate approved scope/design into executable work items and handoff notes, and review whether the resulting contract is actually sufficient for a developer to start, integrate, and demonstrate.

Trigger

Use when creating or reviewing DELIVERY.md, sequencing work, assigning owners, checking task-source consistency, or preparing/reviewing a handoff.

Intents

Intent Use when Read set
delivery_planning (default) Creating or changing work items PROJECT.md, DELIVERY.md, design context
handoff_review Preparing for, or checking, the Handoff gate see below

Required Inputs

PROJECT.md, task source of truth, relevant design context, and the context/artifact contract in pmo-config/context-map.json and pmo-config/artifact-policy.json.

For handoff_review, additionally: HANDOFF.md, DESIGN/BUILD-SPEC.md, RAID-log.md and decision-log.md if present, and pmo-config/handoff-policy.json. When DESIGN/VISUAL-DIRECTION.md, DESIGN/DESIGN-SYSTEM.md, and DESIGN/DESIGN-SYSTEM.html all exist, also read DESIGN/VISUAL-REVIEW.json and the declared capture paths. Read only these. Do not read release artifacts, and do not read source/** beyond the rows already cited in PROJECT.md unless a specific finding requires it.

Allowed Context

Use the context router handoff set. Do not read release artifacts unless the user asks for release readiness.

Mode Behavior

Use pmo-config/policy.json for mode, status, review-stage, task-source, strict-trigger, and sentinel values. Use pmo-config/artifact-policy.json to decide whether DELIVERY.md is required, conditional, or replaceable by GitHub Issues for the active mode and gate. Use pmo-config/handoff-policy.json for handoff artifacts, review lenses, blocking points, and owner policy.

Execution Steps - delivery_planning

  1. Confirm task source of truth: file or GitHub.
  2. Create work items with mode, strict trigger, requirement ref, design ref, status, review stage, evidence, and labels.
  3. Check enum values against runtime policy.
  4. Flag missing blockers or client decisions.
  5. Route implementation deviations through CHANGE-REQUESTS.json; do not silently edit the execution contract or acceptance criteria to absorb them.

Execution Steps - handoff_review

The deterministic validator proves the contract is complete. It cannot prove it is sensible. This review is the part that reads for sense, and its output is structured so the validator can then check that the reading happened, covered everything, and is still current.

  1. Run the gate first: node cli/axiom.mjs validate --project <project> --mode <mode> --gate Handoff. Do not re-report anything it already caught. Structural gaps are its job; this review starts where it stops.

  2. Read the artifacts listed above. Read them together, not one at a time — most real findings are contradictions between documents, each of which is defensible alone.

  3. Walk every lens in pmo-config/handoff-policy.json semantic_review.lenses. Record each one in the output even when it found nothing; a lens with no entry is an unfinished review, and HANDOFF-010 will say so.

  4. Record findings in HANDOFF-REVIEW.json (see templates/HANDOFF-REVIEW.json).

  5. Record both freshness digests. The handoffDigest tool (src/tools/digest-tools.ts) prints them:

    • source_snapshot.digest — the material the requirements came from
    • review_inputs.digest — the governed artifacts you actually read

    A review that records only the first keeps reporting as current after someone rewrites the build sequence or waives a build-spec section.

  6. If the visual-artifact triple exists, treat Visual Proof as separate conditional candidate evidence: confirm the manifest has two committed captures, an input digest, complete rubric, named human reviewer, and direction decision references. Do not judge aesthetics yourself, replace the human declaration, or describe this check as a new approval.

  7. Re-run the gate, then node cli/axiom.mjs handoff for stage verdicts.

  8. When the Claude Design provider track is active, merge only current accepted output into the Final Handoff — the output set whose digest DESIGN/CLAUDE-DESIGN/REVIEW.json records. Unaccepted or stale output is not handoff content (DPROV-005/DPROV-006).

The twelve lenses

Lens The question it asks
value_and_scope_slice Does the slice deliver the value the target milestone has to show?
capability_lifecycle Is each capability complete across its lifecycle, or only the happy path? A count that can go down but never up cannot be demonstrated twice.
data_cardinality_and_units Do entities, cardinality, quantities, and units support the declared use cases? A "stock" entity with no quantity and no unit is a name, not a model.
state_transitions_and_rollback Does every state machine declare guards, terminal states, and how a transition is reversed?
concurrency_and_idempotency Two users at once, a retried request, two callers allocating an id. Is any of that specified?
dependencies_and_build_order Can the declared sequence actually be executed in that order? Work-item numbering is not build order.
ownership_and_capacity Does every stream have a named owner, an integrator, and stated capacity that fits the horizon?
acceptance_seed_reachability Can each acceptance case be reached from the declared seed data?
automated_manual_test_split Is each case classified, and does the classification have a runner or a person behind it?
privacy_and_data_classification Do declared data elements, free text, files, and metadata have classification decisions? Does a privacy commitment in one document contradict a feature in another?
environment_and_device_constraints Does the serving model satisfy the declared device and runtime capabilities?
demo_startup_reset_and_recovery Is there a declared startup, reset, degraded, and recovery path?

Rules for findings

  • Every finding cites evidence. evidence_refs names the artifact rows the finding is derived from. A finding with no evidence is an opinion.
  • Every finding has a suggestion and an owner. Naming a problem without naming the next action or who takes it is not a review outcome.
  • Never invent a requirement. If a capability looks incomplete relative to what the source already asked for, that is a completeness finding against the existing requirement — not new scope. If it is genuinely new scope, it is an open question for a human, not a finding.
  • Separate blocking points. before_build is a different claim from before_demo. Getting this wrong is the most damaging thing this review can do: mislabelling a demo blocker as a build blocker stalls a team that could be working, and the reverse produces a demo-day surprise.
  • Do not restate deterministic findings. If HANDOFF-004 already caught the inverted sequence, do not raise it again here.
  • Do not echo source content. Findings are read in CI logs and dashboards. Cite the location; do not paste the row.

What an AI reviewer may and may not close

May close a finding (status: resolved) when the artifacts now show it was fixed, and cite the change as evidence. Every non-open status needs a decision_ref that resolves.

These limits are enforced by HANDOFF-010, not merely requested here. An AI-authored review that sets accepted_risk, or that closes a privacy finding, fails the gate. A closure under a human-only lens additionally has to cite a DEC-### that exists in decision-log.md with a named decider.

May not close, regardless of how obvious the answer looks — these need a named human and a decision record:

  • anything under privacy_and_data_classification or environment_and_device_constraints (configured in semantic_review.human_only_close_lenses)
  • anything requiring a business, legal, security, commercial, or budget decision
  • anything that would change scope, a deadline, or an accepted risk

Set status: open and name the human in owner.

Output Contract

For delivery_planning: work-item changes, dependency notes, owner gaps, and validation result.

For handoff_review: a HANDOFF-REVIEW.json conforming to templates/HANDOFF-REVIEW.json, plus a short summary naming what blocks build, what blocks demo, and what needs a human decision. When Visual Proof applies, also identify missing or stale VISUAL-REVIEW.json evidence without making an aesthetic verdict.

Approval Rules

Mode escalation and high-risk work require human confirmation before delivery starts.

A handoff review is not an approval. HANDOFF-REVIEW.json is candidate evidence produced by a reviewer. It never substitutes for the Design Ready approval row in PROJECT.md, and the Handoff gate introduces no new approval of its own — it reuses the one that already exists. An AI must never move an approval from pending to approved, and must never present a readiness score as a decision. See docs/concepts/human-authority.md.

Validation Command

node cli/axiom.mjs validate --project <project> --mode <mode> --gate Scope
node cli/axiom.mjs validate --project <project> --mode <mode> --gate Handoff
node cli/axiom.mjs handoff --project <project> --mode <mode>

Prohibited Actions

Do not create hidden task systems, duplicate source of truth, hardcode mode/gate matrices, or add features outside scope. Do not hardcode review lenses, blocking points, or owner tokens — they come from pmo-config/handoff-policy.json. Do not record a review whose digests you did not recompute — both of them, after every other artifact is final.

Completion Criteria

Every work item references an existing requirement/business rule and has valid mode, status, and review stage.

For handoff_review: every lens is recorded, every finding has evidence, a suggestion, an owner, and a blocking point, the digest is current, and every finding that needs a human decision is still open with that human named. If the visual-artifact triple exists, conditional Visual Proof evidence must also be present and current, while Design Ready remains human-owned.

Visual Proof has no .agents/skills mirror. Follow the active .claude/skills/ workflow and use the generated skills/ package only through its repository build command.

04 同仓库更多 Skills