feature-pipeline-preview - SKILL.md Agent Skill

name: feature-pipeline-preview description: Use when implementing a new NeoHaskell feature end-to-end — from ADR draft through PR merge — and the work needs the structured 17-phase pipeline with ADR, security and performance reviews (each grounded against feature complexity and the Jess persona), outside-in test design, build loop, and PR/CI handling. Triggers on 'run feature pipeline', 'feature pipeline', 'start pipeline', 'implement feature', 'new NeoHaskell feature', 'implement issue #N', 'add type to nhcore', 'pipeline init', 'next pipeline phase', 'approve phase N'. kind: process executor: opus model: claude-opus-4-7

NeoHaskell feature pipeline

Coordinates the 18-phase pipeline for landing a NeoHaskell feature. State machine lives in scripts/pipeline.py; each phase dispatches to a child node. Replaces the older neohaskell-feature-pipeline skill.

The pipeline has two distinguishing characteristics:

Grounding loop — after every security and performance deep-audit (phases 4, 5, 12, 13), a grounding pass filters findings through the feature's complexity tier and Jess's 15-minute pocket so simple features do not accumulate cascades of unneeded caching, SHA pins, constant-time compares, or INLINE pragmas. See references/grounding-loop.md.
Quality-rubric review-gates — phases 6 (DevEx), 7 (architecture), 8 (test spec) each run a produce step followed by an independent review-quality step that validates the artefact against a per-phase rubric (devex-rubric.md, architecture-rubric.md, test-spec-rubric.md). These review steps replace the previous human PAUSE gates — the pipeline advances automatically on RUBRIC: pass and halts with a structured rubric record on RUBRIC: fail.

Shared invariants

All state lives in .pipeline/state.json and .pipeline/findings-*.json. Drive via python3 .claude/skills/feature-pipeline-preview/scripts/pipeline.py <cmd>.
The pipeline must be initialised before any phase runs. Phase 1 calls pipeline.py init.
PAUSE-gated phases (marked 🔒 below) stop the orchestrator until the maintainer runs pipeline.py approve <N>. The PAUSE remains only on phases 3 (ADR draft — opens a draft PR for review), 16 (PR finalize — marks the draft ready), 17 (Opus PR review), and 18 (CI cycle / merge) — every other phase auto-gates on its rubric or findings.
Every step reads references/jess-persona.md and references/nhcore-context.md when delegating to a sonnet child.
The grounding pass is mandatory — it cannot be skipped to "save time", and it cannot be merged with the deep-audit step.
The review-quality steps on phases 6/7/8 are mandatory and must be run by a different agent invocation from the producer step (independence is what makes the gate trustworthy).
Trust-but-verify after every leaf. When any leaf calls pipeline.py complete <N>, the orchestrator MUST independently re-verify the leaf's claimed outputs before advancing the cursor. For build/test-gated phases (11, 15), re-read .pipeline/test.log and confirm 0 failures across all suites. For findings-gated phases (4, 5, 12, 13, 17), re-read the findings JSON and confirm the leaf's blocker-count claim. For fix-findings (14), confirm blockers == 0 across both findings files. Canonical implementation: scripts/verify-leaf.py — invoke as python3 .claude/skills/feature-pipeline-preview/scripts/verify-leaf.py <N>. If the verify fails, refuse to advance and surface the mismatch.
Diff-shape policing. After each phase, run git diff --name-only HEAD and check the changed paths against references/phase-allowed-paths.md. Any path outside the phase's allow-list refuses complete <N> and surfaces the offending paths.

Steps

Init pipeline — read ./01-init-pipeline/SKILL.md and follow it with {feature_name, issue_number, module_path, test_path, branch_name, adr_number}. Verify: .pipeline/state.json exists and pipeline.py status lists phase 1 as completed.
Classify feature — spawn an Agent (model: haiku) and instruct it to read ./02-classify-feature/SKILL.md and follow it. Verify: .pipeline/classification.json exists with tier ∈ {trivial, simple, moderate, complex, security-critical}.
ADR draft 🔒 — spawn an Agent (model: opus) and instruct it to read ./03-adr-draft/SKILL.md and follow it. Verify: docs/decisions/NNNN-slug.md exists with Status: Proposed, the ADR is committed, and a draft PR is open (pipeline.py get pr_url returns a URL). Then pipeline.py complete 3 and stop until pipeline.py approve 3 — the maintainer reviews the ADR in the draft PR.
Security review (ADR) — spawn an Agent (model: haiku) and instruct it to read ./04-security-adr/SKILL.md and follow it. Verify: .pipeline/findings-04.json exists with blockers >= 0. If blockers > 0, stop and surface to maintainer.
Performance review (ADR) — spawn an Agent (model: haiku) and instruct it to read ./05-performance-adr/SKILL.md and follow it. Verify: .pipeline/findings-05.json exists. May run in parallel with step 4.
DevEx review — spawn an Agent (model: haiku) and instruct it to read ./06-devex-review/SKILL.md and follow it. Verify: .pipeline/devex-review-rubric.json exists with "verdict": "pass". On fail, the rubric record names the failing checks and the pipeline halts here.
Architecture design — spawn an Agent (model: haiku) and instruct it to read ./07-architecture-design/SKILL.md and follow it. Verify: .pipeline/architecture-rubric.json exists with "verdict": "pass". On fail, the rubric record names the failing checks and the pipeline halts.
Test spec design — spawn an Agent (model: haiku) and instruct it to read ./08-test-spec-design/SKILL.md and follow it. Verify: .pipeline/test-spec-rubric.json exists with "verdict": "pass". On fail, the rubric record names the failing checks and the pipeline halts.
Test writing — spawn an Agent (model: sonnet) and instruct it to read ./09-test-writing/SKILL.md and follow it. Verify: tests compile and ALL fail.
Implementation — spawn an Agent (model: sonnet) and instruct it to read ./10-implementation/SKILL.md and follow it. Verify: source files exist at the planned paths.
Build loop — spawn an Agent (model: haiku) and instruct it to read ./11-build-loop/SKILL.md and follow it. Verify: cabal build all and cabal test pass; .pipeline/hlint.log exists (warnings are captured for the PR body, not gated here).
Security review (impl) — spawn an Agent (model: haiku) and instruct it to read ./12-security-impl/SKILL.md and follow it. Verify: .pipeline/findings-12.json exists.
Performance review (impl) — spawn an Agent (model: haiku) and instruct it to read ./13-performance-impl/SKILL.md and follow it. Verify: .pipeline/findings-13.json exists. May run in parallel with step 12.
Fix findings — spawn an Agent (model: sonnet) and instruct it to read ./14-fix-findings/SKILL.md and follow it. Verify: blockers == 0 across findings-12 and findings-13 after fixes; cabal test still green.
Final verify — spawn an Agent (model: haiku) and instruct it to read ./15-final-verify/SKILL.md and follow it. Verify: clean build, all tests pass; .pipeline/hlint.log refreshed (hlint is captured for the PR body, not gated).
Finalize PR (mark ready) 🔒 — spawn an Agent (model: haiku) and instruct it to read ./16-create-pr/SKILL.md and follow it. It commits/pushes the implementation, sets the final title/body, and converts the draft PR opened in phase 3 to ready for review (gh pr ready) — it does NOT create a second PR. Verify: python3 .claude/skills/feature-pipeline-preview/scripts/pipeline.py get pr_url returns the draft PR URL and gh pr view --json isDraft -q .isDraft is false. The submit leaf calls pipeline.py complete 16 itself; then stop until the maintainer runs python3 .claude/skills/feature-pipeline-preview/scripts/pipeline.py approve 16.
Opus PR review 🔒 — spawn an Agent (model: opus) and instruct it to read ./17-opus-pr-review/SKILL.md and follow it. Verify: .pipeline/findings-17.json exists. Then PAUSE until pipeline.py approve 17.
CI cycle — spawn an Agent (model: haiku) and instruct it to read ./18-ci-cycle/SKILL.md and follow it. Verify: CI green and PR merged.

Phase completion is owned by each phase's own skill — the orchestrator never calls pipeline.py complete <N> unconditionally. Rubric-gated phases (6, 7, 8) and the script-style record leaves (2, 4, 5, 12, 13) call complete <N> only when the gate passes; the test-writing, implementation, build-loop, fix-findings, final-verify, and PR-finalize phases call it from their last step on success. PAUSE-gated phases (3, 16, 17, 18) then stop and wait for the maintainer's python3 .claude/skills/feature-pipeline-preview/scripts/pipeline.py approve <N> before continuing.

Refusals

.pipeline/state.json already exists with a different feature → refuse and ask the maintainer to run pipeline.py reset or pick the right state.
A phase's verify check fails → stop, surface the failure, do not advance the cursor.
A grounding pass is missing from a security/perf phase → refuse to mark the phase complete.
The maintainer asks to skip the grounding pass → refuse; explain that the grounding pass is the load-bearing guard against cascade reviews.