ny-auto-pipeline

ny-auto-pipeline

The autonomous orchestration pattern for shipping non-trivial features end-to-end with subagent-backed review at every phase, in a single conversation, without interactive gates.

When to use

Invoke ONLY when the user explicitly asks for an autonomous full-pipeline run. Examples:

• "auto brainstorm → spec → review → plan → subagent-driven impl"
• "ทำ pipeline ต้นจนจบไม่ต้องหยุดถาม"
• "เอา Subagent มาทำตั้งแต่ brainstorm จนถึง impl"
• "run the full skill chain autonomously"

Do NOT invoke for: small fixes, single-file edits, code review only, exploratory questions, or any task where the user hasn't granted explicit autonomous authority.

Why this exists

The default superpowers:brainstorming and superpowers:writing-plans skills have hard interactive gates: they ask one question at a time and require the user to approve the design before plan-writing. That discipline is correct for collaborative work. It's wrong when the user has explicitly said "go autonomously" — every "any feedback before I continue?" message becomes friction the user explicitly told you to skip.

This skill encodes the override semantics the user opts into, AND the discipline that has to stay even in autonomous mode: spec written and reviewed, plan written, every implementation task implemented + spec-reviewed + code-reviewed before the next task starts.

The pipeline

Seven phases. The orchestrator (you) coordinates; specialized subagents do the per-task work. Several phases compose existing skills — references inline so the orchestrator invokes the canonical source rather than reinventing it:

• Phase 1 builds on superpowers:brainstorming (with interactive gate overridden)
• Phase 4 builds on superpowers:writing-plans
• Phase 5 builds on superpowers:subagent-driven-development + superpowers:test-driven-development
• Phase 5 dispatch mechanics — which model tier each subagent gets (Haiku/Sonnet/Opus), how to stay responsive while subagents run, and parallel-vs-serialize decisions — follow ny-orchestrate-subagents
• Phase 6 composes superpowers:verification-before-completion + 9arm-skills:scrutinize
• Phase 7 composes 9arm-skills:post-mortem when warranted
• Empirical verification during impl composes 9arm-skills:debug-mantra

This skill is the autonomous full-pipeline composition (brainstorm → … → post-mortem). For the lower-level question of how to route and run subagents — capability-tiered dispatch, background multitasking, and the conflict-detection heuristic — see the sibling skill ny-orchestrate-subagents. ny-auto-pipeline decides WHAT phases run and in what order; ny-orchestrate-subagents decides HOW each subagent is tiered, dispatched, and parallelized vs serialized.

Phase 1: Brainstorm                (orchestrator, with context-gathering)
Phase 2: Spec write                (orchestrator)
Phase 3: Spec review               (subagent — fresh context)
Phase 4: Plan write                (orchestrator, after applying review fixes)
Phase 5: Subagent-driven impl      (one subagent per task + 2 review subagents per task)
Phase 6: Scrutinize                (orchestrator — read the actual diffs)
Phase 7: Post-mortem               (orchestrator — capture what happened)

Phase 1 — Brainstorm (no interactive questions)

• Invoke superpowers:brainstorming to load its checklist for reference, then override the interactive question loop. You already have the user's go-ahead.
• Gather repo context with ny-codemap / ny-codegraph (faster + structured than grep/find).
• If the task is too large for a single spec, decompose into sub-projects and pick the first to ship — but say so explicitly. Don't bury the decomposition.
• Pick approach. Don't ping-pong 2-3 options to the user; pick one with reasoning visible in your message.

Phase 2 — Spec write

• Save to docs/superpowers/specs/YYYY-MM-DD-<topic>-design.md (or the project's documented spec location).
• Sections: Goal, Non-goals, High-level architecture, CLI/API surface (if applicable), File structure, Behavior/concurrency invariants (if applicable), Tests, Acceptance criteria, Known limitations, Open questions resolved at plan-time.
• Be loud about what's out of scope. The spec is the contract; if something is deferred, name it.
• Commit before the review subagent runs.

Phase 3 — Spec review (subagent)

• Dispatch one subagent with the general-purpose or a language-specific reviewer (e.g. ecc:rust-reviewer) agent type.
• Brief includes:
  • Path to the spec
  • Paths to all referenced source files (so the reviewer can verify load-bearing claims)
  • Specific questions framed for review (placeholders, internal consistency, technical correctness against actual code, ambiguity, scope, testability)
  • Output format: Blockers / Fixes recommended / Open questions / Sign-off
• Apply fixes inline. Don't re-dispatch unless the reviewer flagged blockers and you've made non-trivial changes. Commit the fixes.

Phase 4 — Plan write

• Invoke superpowers:writing-plans semantics: bite-sized TDD-shaped tasks, exact file paths, complete code in every step, exact commands with expected output.
• Self-review: spec coverage table mapping each spec section to a task, placeholder scan, type consistency.
• Include a BLOCKED fallback for any task with non-trivial risk. Stating "if X fails, the pivot is Y" gives the implementer a path forward that doesn't require pinging you mid-run.
• Commit.

Phase 5 — Subagent-driven implementation

Per task, in strict sequence (NEVER parallel — implementation subagents conflict on file state). This is the ny-orchestrate-subagents conflict-detection heuristic applied: every implementer shares one git working tree, so their write-sets overlap → serialize. The escape hatch is isolation — give each its own git worktree — but the default for in-tree edits is sequential. (Read-only reviewers operate on a captured diff, so the two reviewers per task can run concurrently.)

1. Implementer subagent (general-purpose, model per the ny-orchestrate-subagents capability ladder: Haiku for mechanical, Sonnet for integration, Opus for architecture)
   • Brief: full task text pasted verbatim, scene-setting context, parent commit SHA, ESCALATE list, reporting format
   • Do NOT make the subagent read the plan file — paste the task text into the prompt
2. Handle the report:
   • DONE → proceed to review
   • DONE_WITH_CONCERNS → assess concerns; correctness concerns must be addressed before review
   • NEEDS_CONTEXT → provide context, re-dispatch
   • BLOCKED → see "Escalation" below
3. Spec-compliance reviewer subagent — verify the implementation matches the task by reading the actual diff, not the implementer's claims
4. Code-quality reviewer subagent (ecc:<lang>-reviewer if available, else ecc:code-reviewer) — run AFTER spec compliance passes
5. Fix any review findings before the next task

Track progress via TaskCreate/TaskUpdate at the pipeline-phase granularity (7 tasks), NOT per implementation task — the plan's checkboxes are the implementation tracker.

Phase 6 — Scrutinize (delegate to the scrutinize skill)

This phase composes two existing skills — invoke them; do not reinvent.

1. superpowers:verification-before-completion — Iron Law: "NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE." Re-run the baselines yourself (don't trust subagent-reported counts). If you haven't run the verification command in this message, you cannot claim it passes.
2. 9arm-skills:scrutinize (from thananon/9arm-skills — skills/engineering/scrutinize/SKILL.md) — outsider-perspective end-to-end review with the workflow: Intent → Trace → Verify → Report. Specifically:
   • Intent. State the goal in one sentence. Ask whether a simpler / smaller alternative achieves the same goal. If a better alternative exists, surface it.
   • Trace. Walk the actual code path end-to-end through the real code (not just the diff). Include unchanged code at the seams — bugs hide there.
   • Verify. For each claim the change makes, walk the trace and confirm/refute it. Consider edge cases, concurrent callers, error paths, silent contract changes.
   • Report. One tight section per finding, ordered by severity. Cite file:line. Close with a one-line verdict: ship / fix-then-ship / rework / reject.

Concretely run, in this order:

1. git log --oneline <base-sha>..HEAD — commit count + messages
2. git diff --stat <base-sha>..HEAD — files changed
3. git diff <base-sha>..HEAD on load-bearing files (concurrency primitives, public API, cross-module callers). Look for the patterns scrutinize calls out: silent scope creep, overpromised work, unwrap() / expect() / as any / // TODO, weaker test invariants than spec required.
4. Re-run baselines yourself (verification-before-completion's gate).
5. Manual smoke against a real input the test suite doesn't cover — subagent "I ran it" is not equivalent to you seeing it pass.

If scrutinize finds something the reviewers missed: fix inline if mechanical, escalate via advisor if structural. Don't silently shrug.

Phase 7 — Post-mortem (delegate to the post-mortem skill when warranted)

If the pipeline involved real debugging or a non-trivial blocker, invoke 9arm-skills:post-mortem (from thananon/9arm-skills — skills/engineering/post-mortem/SKILL.md). Its required inputs match exactly what a real blocker generates: reliable repro, root cause identified, fix identified, fix validated. Its structure (Summary / Symptom / Root cause / Why it produced the symptom / Fix / How it was found / Why it slipped through / Validation) is the canonical engineering writeup.

When to invoke the full post-mortem skill (separate doc):

• A blocker required an advisor call and a real pivot
• A diagnostic claim from a subagent turned out to be wrong on inspection, and the eventual cause is worth recording
• A test was misdesigned in a way that hid a real bug
• The work shipped with a documented limitation that future engineers will want context on

When NOT to invoke post-mortem (per its own NOT-WHEN list):

• Fix not validated yet
• Trivial work (the PR description is the record)
• A clean run with no blockers — a short wrap-up message is sufficient

For clean runs (no advisor calls, no blockers, no deferred acceptance criteria), produce a compact user-facing wrap-up only:

• Commits that landed (SHAs + messages, compact table)
• Baselines green (test counts, re-run by you in Phase 6)
• What's PARTIAL or DEFERRED — say so plainly, don't paper over with the older healthier number
• Suggested next-session pickup if obvious

For runs with real friction, the post-mortem doc is the deliverable; the user-facing message can be a one-line pointer to it.

Do NOT write a post-mortem on top of an unvalidated fix. post-mortem requires validated repro and validated fix — if you don't have both, you're documenting a hypothesis.

Escalation

The concrete advisor mechanism is the sibling skill ny-principal-advisor — a fresh-context Principal SWE/Architect subagent dispatched on Fable (fallback Opus). It defines the brief template, the dispatch rules, and how to weigh the answer.

Call the advisor (ny-principal-advisor) when:

1. An implementer subagent reports BLOCKED for the second time on the same root cause. Two attempts on the same error means the approach itself is wrong — not the implementation.
2. Spec acceptance criteria conflict with empirical reality. (Example from a past run: spec said "scaffolded project should boot via bun run dev", reality was Bun's file: install symlinks individual files back to the source repo, causing a dual-React copy. Advisor said "stop digging, this is a separate sub-project, pivot the test.")
3. You're about to silently change the spec to make a test pass. Don't. Surface the conflict to the advisor.

Don't call the advisor for:

• Compile errors with obvious fixes
• Tests that fail with clear messages
• Subagent reports of "DONE" — those go to the reviewer
• Single false positives from loop-detection hooks (they're cumulative session counters, not actionable signals)

Empirical-first verification (delegate to debug-mantra)

When a subagent reports BLOCKED or DONE_WITH_CONCERNS with a diagnostic claim ("X is broken because Y"), invoke 9arm-skills:debug-mantra (from thananon/9arm-skills — skills/engineering/debug-mantra/SKILL.md) before pivoting. Its four-step discipline is:

1. Reproduce reliably. Build a runnable repro before anything else. No repro → stop, get one.
2. Know the fail path. Debugger first; then source trace + knob enumeration; then in-code instrumentation. Don't skip ahead.
3. Falsify the hypothesis. Run the disproof first. If the hypothesis dies, you saved yourself from chasing a phantom.
4. Every run is a breadcrumb. Cross-reference all attempts.

This is the right discipline when a subagent's diagnostic claim needs verification before you commit to a pivot. The orchestrator-side reproductions of "dual-React surfaces in build mode too" or "the linker error persists with Option pivot" are both step-1 work — confirm the failure mode at your own command line, with eyes on the actual output, before re-dispatching with a different approach.

Coupled with the advisor tool at real decision points, debug-mantra step 3 (falsify) is what discriminates between "we need a small fix" and "the spec premise is broken."

Test design discipline

The orchestrator writes the plan; the implementer writes the test code per the plan. If the plan's test is wrong, the implementer will execute it faithfully and then report a real failure on a correct implementation — wasting a cycle.

Lessons:

• Sequential reuse can inflate concurrent assertions. A race test that asserts "exactly M claims out of N attempts" without holding the claims open lets early claimers drop before late contenders attempt — the late ones then succeed too, totaling > M. Use a two-barrier design: one barrier to force simultaneous start, a second barrier to force every contender to finish reporting before any claim drops.
• Debug-only assertions don't catch release-mode bugs. If your fix replaces a debug_assert!, the regression test MUST pass --release to prove the invariant survives optimization.
• Run flaky-suspect tests 5× before declaring done. A test that passes once might pass by accident.

Honest reporting

When the pipeline ends:

• Name the commit SHAs that landed
• Name the baselines that are green (and the counts)
• Name what's PARTIAL (worked at code level, deferred at runtime) — don't paper over with the older healthier number
• Name the next-session pickup if the work has obvious follow-ups

DON'T:

• Pretend deferred acceptance criteria are met
• Hide regressions behind older benchmark numbers
• Claim "shipped" when an E2E test was trimmed for the wrong reason

Compact at phase boundaries, not mid-task

If a StrategicCompact or similar hook fires:

• Mid-task (between TDD steps in an implementer dispatch): ignore, finish the task, compact between phases
• Between phases (after a phase ships and commits): natural break, compact is fine
• Cumulative file-modified warnings: usually false signal of "scattered" work in long autonomous sessions — actual per-task scope is bounded by the spec

Output format expectations

The user invoked ny-auto-pipeline because they want results, not narration. Match it:

• One sentence per phase transition: "Spec written. Dispatching reviewer."
• After each task ships: short status line with commit SHA and test counts
• Final wrap-up: a compact summary table (commits + baselines), known limitations, suggested next pickup

Match the user's language. If they wrote Thai, reply Thai; if English, English. Technical terms stay English (commit SHA, baseline counts, file paths, code blocks).

Anti-patterns to avoid

• Pinging the user with "Should I continue?" between phases. The user already said go.
• Spawning implementation subagents in parallel. They conflict on file state.
• Skipping the code-quality review on "simple" tasks. Static file emission is a tax you pay for the next bug being caught early.
• Trying to fix BLOCKED tasks by re-dispatching the same agent with the same approach. Dispatch FRESH with a NEW approach, or call advisor.
• Padding the wrap-up with adjectives. The git log + test counts speak for themselves.
• Skipping Phase 6 because "the reviewers already looked." The reviewers are subagents with bounded context; you have the full conversation. They miss things you'll catch.
• Writing the post-mortem as a separate doc/commit. It's a section in the wrap-up message, not a deliverable in its own right.