Plan Spec

/plan-spec

You are the orchestrator for the fourth and final skill in the System-tier planning chain. The deliverable is a populated draft spec at ai-docs/specs/<spec-id>.md plus the same plan at ai-docs/plans/<plan-name>.md updated to status: promoted with successor_spec: <spec-id> set.

This skill is the gate between forward-looking design (the plan) and behavior-preserving execution (the spec). The promotion is explicit — you call scripts/plans.py promote, which itself shells out to scripts/specs.py init. No silent promotion.

How success is judged

• The Stage 1 gate held: any unresolved P0 fork in §6 produced an explicit ABORT naming each fork and its /decide <slug> — refusal is the correct outcome, not a failure.
• The spec at ai-docs/specs/<spec-id>.md is transcribed from plan §3-§5 (P1 forks land in §Exceptions as deferred), not invented.
• The link is bidirectional: plan status: promoted + successor_spec: set, spec pointing back via Provenance.
• Promotion went through scripts/plans.py promote — never a hand-rolled scaffold. Write toward these gates from Stage 0.

Core beliefs

1. The promotion gate is real. A plan with unresolved P0 forks in §6 is NOT promotable. Refuse and recommend /decide for each blocking fork. The cost of a premature promotion is a spec that forces the implementer to re-derive the architecture.
2. Spec sections come from plan sections. The spec's Architecture reflects plan §3 + §5; Implementation reflects plan §4 + §3; Exceptions reflects plan §5 (smells accepted) + §6 (P1 deferred forks). Don't invent — transcribe.
3. Plan ↔ spec link is bidirectional. The plan's successor_spec points at the spec; the spec's motivating_decision (if set) plus a # Provenance section point back at the plan. This is the audit trail.
4. One plan, one spec. If a plan should yield multiple specs (different code roots, different timelines), it was actually two plans — abandon the merge, re-run /scope-feature for each.

Scope (this skill itself)

• Project root: this worktree's root.
• Python: .venv/bin/python for scripts/plans.py and scripts/specs.py (they use the shared PyYAML-backed frontmatter parser from scripts/_lib/).
• Read: ai-docs/plans/<name>.md.
• Write (via plans.py promote): ai-docs/specs/<spec-id>.md, ai-docs/plans/<name>.md (status + successor_spec).
• Write (this skill): spec body sections after promote scaffold.

Pipeline

Stage 0 — Setup

PLAN_NAME="<arg>"
PLAN_PATH="ai-docs/plans/${PLAN_NAME}.md"
SPEC_ID="<plan-name unless --spec-id was supplied>"

Optional --spec-id <id> overrides the default successor spec id (the plan name). Use the override only for a real collision or when one plan intentionally promotes to a differently named spec.

Verify plan exists and status: architected. If status is anything else, abort and recommend the matching stage skill (or /decide if P0 forks remain).

Stage 1 — Validate promotion gate

Read ${PLAN_PATH} §6 (Open Decisions). Count entries under each priority heading:

• P0 forks > 0 → ABORT. Print each P0 fork name and the recommended /decide <slug> invocation. Do NOT proceed.
• P1 forks > 0 → continue, but include them in spec §Exceptions with (deferred from plan §6 — resolve during /refactor-subsystem Phase 2).

Read §5 (Architecture Fit). Verify:

• At least one ADR conformance OR an explicit "no constraining priors" note. Empty §5 → ABORT, recommend re-running /architecture-fit.

Stage 2 — Determine code roots

Ask the user for the spec's code roots (the file paths the spec will track for # spec:<id>::IM-N annotations). Default candidates:

• Subsystems named in plan frontmatter subsystems: field (translate to file/dir paths via the subsystem doc's first "Files" section).
• Files named in §3 "Files touched (estimate)".

Present the candidate list and wait for user approval / amendment. Approval token: approved / approve / go / lgtm / proceed / yes.

Stage 3 — Promote

.venv/bin/python scripts/plans.py promote "${PLAN_NAME}" \
  --spec-id "${SPEC_ID}" \
  --code-roots "<root1>" \
  [--code-roots "<root2>" ...] \
  --allow-missing

--allow-missing is appropriate for System-tier specs because new subsystems may not yet have files on disk. The promote subcommand:

1. Verifies plan is not already promoted.
2. Shells out to scripts/specs.py init <spec-id> --code-roots ... with title and motivating_decision propagated from the plan.
3. Writes the spec scaffold under ai-docs/specs/<spec-id>.md.
4. Updates the plan's status: to promoted and sets successor_spec: to the spec id.

If promote fails (non-zero exit), surface the error and stop. Do not re-run silently — the plan and spec must end in a consistent state.

Stage 4 — Fill spec body from plan

Open the new spec at ai-docs/specs/${SPEC_ID}.md. The scaffold has five standard narrative sections (Goals / Architecture / Implementation / Learnings / Exceptions). Fill them by transcribing from the plan:

• Goals ← plan §1 problem statement + §2 success criteria.
• Architecture ← plan §3 cross-subsystem call graph + plan §5 decision conformance + pattern alignment.
• Implementation ← plan §4 behaviors-to-preserve become characterization-test items (AR-N); files to touch become IM-N items, ordered test-first.
• Learnings ← empty placeholder (/refactor-subsystem Phase 2b fills post-execution).
• Exceptions ← plan §5 smells-accepted + plan §6 P1 deferred forks.

Replace the scaffold's status: STUB with status: draft once the sections are filled.

Add a # Provenance block immediately after the frontmatter:

# Provenance

Promoted from plan `<plan-name>` (`ai-docs/plans/<plan-name>.md`).
- Plan §1-2 → Goals
- Plan §3, §5 → Architecture
- Plan §4, §3 → Implementation (test-first)
- Plan §5, §6 → Exceptions

This is the audit trail — when the spec drifts during implementation, the plan is the original-intent record.

Stage 5 — Verify links

.venv/bin/python scripts/plans.py audit
.venv/bin/python scripts/specs.py inventory-check "${SPEC_ID}"

The plan audit confirms successor_spec resolves; the spec inventory-check confirms the scaffold parses and the code roots are addressable.

Stage 6 — Summarize

Report to the user in ≤10 lines:

• Plan path: ai-docs/plans/<name>.md (now status: promoted).
• Spec path: ai-docs/specs/<spec-id>.md (now status: draft).
• Code roots tracked in spec.
• ADR ids carried over from plan §5.
• P1 forks deferred to spec §Exceptions (count + names).
• Recommended next command:
  • /refactor-subsystem <spec-id> to begin behavior-preserving execution.
  • Or, if the spec is greenfield (no code-roots exist yet): manual implementation against the spec, with /refactor-subsystem joining once code lands.

Non-goals

• Authoring decisions (that's /decide — must happen before this skill if there are P0 forks).
• Implementing the feature (that's /refactor-subsystem).
• Re-running impact analysis (that's /impact-feature; if the impact map is stale, the plan should be reverted to impacted and re- walked, not patched here).
• Editing the plan's §1-§4 (those are frozen once the plan is architected; only status and successor_spec change).

When things go sideways