| name | pip-design |
| description | Design or refine an Apache Pulsar PIP (Pulsar Improvement Proposal) from a draft into a votable, minimal, 5-year-durable proposal. Runs multi-model multi-viewpoint reviews, an iterative minimalism-refinement loop to a fixpoint, and (when the PIP has a reference implementation) the full design-doc + code + CI pipeline. Use when asked to design a PIP, review/harden a PIP draft, or drive a proposal to "nothing left to take away." |
| argument-hint | <path-to-pip-draft.md> [--impl <impl-branch-or-worktree>] [--reviewers fable,opus,codex] |
| allowed-tools | Bash(*), Read, Write, Edit, Glob, Grep, Agent, Skill, SendMessage |
PIP Design
Take a Pulsar Improvement Proposal from a rough draft to a proposal that is votable, minimal, and built to hold for 5+ years — through disciplined multi-model review and an iterative minimalism refinement loop.
"Perfection is achieved, not when there is nothing more to add, but when there is nothing left to take away." — Antoine de Saint-Exupéry
"Simple can be harder than complex. You have to work hard to make your thinking clean to make it simple." — Steve Jobs
These two quotes are the skill's north star. Minimalism is the goal; the work of getting there is making the design's thinking clean.
When to use
- A PIP draft exists (even rough) and needs to be hardened into something the community can vote on.
- A design has grown across revisions and needs a subtractive pass ("what can we remove?").
- You want independent, adversarial review of interface decisions that must live for years.
The core loop
Design refinement is iterative to a fixpoint, not one-shot. Repeat REVIEW → DECIDE → APPLY → VERIFY until a full review round surfaces nothing new.
1. Understand the draft against the code
Read the PIP end to end. If it has (or claims) a reference implementation, verify every claim against the actual code — grep the shipped types and their production call sites, not just declarations. Do not trust prose that says "X is used" or "Y is a well-known extension point"; confirm a real consumer exists. Fabricated or aspirational claims are the first thing to fix.
2. Multi-model, multi-viewpoint review
Run independent reviews from different models, each with a distinct lens. Convergence across models = highest confidence; a lone finding is a hypothesis to verify.
Suggested panel (adapt to what's available):
- Fable 5 — design coherence + subtractive minimalism (smallest sufficient type set; does each element pull its weight in the whole shape?).
- Opus 4.8 (xhigh / think hard) — durability + implementability (which decisions are safe 5-year bets vs churn-risk; premature generalization; missing invariants that force a later break; third-party-implementer traps).
- Codex (GPT-5.5, high reasoning) — adversarial ("argue the design is bigger than it needs to be"; security; backward-compat).
Give each reviewer the SAME lens questions + a shared list of seed candidates, but let each bring its angle. Have them return structured verdicts: CUT / RESHAPE / KEEP per element, with rationale, an additive add-back path, and the tradeoff. (Run reviewers as parallel subagents; Codex via codex exec -c model_reasoning_effort="high".)
Run the reviewers READ-ONLY. They analyze; you decide.
3. Decide — the minimalism judgment
For every public type / method / field / enum value / resolved decision, ask:
- Is there a real in-tree consumer today, or is it speculative generality?
- If removed, what concretely breaks — and could an existing element meet the need?
- Is this exact shape a safe 5-year bet, or a guess likely to churn (forcing a compat break)?
- Can a future revision add it back additively when a real need appears? (If yes, removal now is safe — a shipped public API is forever; add-back is cheap.)
The deciding principle (earned across a real PIP):
Cut what carries contract weight — an interaction rule, a merge precedence, a duplicate path — when it has no consumer. KEEP zero-rule projections that serve the audience the SPI exists for (third-party implementers), safety guards together with their escape valves, and service seams that have a consuming twin.
Corollaries:
- A lying accessor (always returns a constant / fabricated value because the surrounding seam can't populate it) is worse than absent — cut it.
- A duplicate path ("a second way to express what the design already expresses once") — cut one.
- "Wired + functional but no in-tree producer" does NOT by itself clear the cut bar for an SPI meant for third parties. What clears it is no-consumer AND contract-weight.
- Guard against manufactured minimalism. Removing a safety guard, a zero-rule projection serving the SPI's audience, fixed RFC/standard vocabulary, or a shape whose removal forces a future break is not simplification — it manufactures traps and gaps. Do not invent findings to seem thorough.
- Grep production call sites, not declarations — the sharpest cuts come from seeing what values callers actually pass (e.g. a port field that every caller sets to 0).
4. Apply — the full pipeline
Each accepted change runs the whole pipeline so the proposal and its implementation never drift:
- Design doc: make the edit AND write a decision record — for a cut: what / why / additive add-back path / tradeoff; for a contested keep: the rationale. Kept-and-why is as valuable as cut-and-why. Maintain an "honesty list" of deliberate keeps with zero/low consumers so a later pass doesn't re-litigate them.
- Code (if there's a reference impl): make the change minimal and surgical; STOP and report if a "dead" element turns out load-bearing.
- Verify: compile + the affected tests; for anything wire-observable (auth bytes, TLS handshake, HTTP encoding) prove wire-equivalence before "exercising" a speculative element — if it can't be preserved, cut instead.
5. Verify the whole & re-review (fixpoint)
Re-run the multi-model review on the refined design. Keep iterating until a full round returns no agreed CUT/RESHAPE that survives scrutiny. That is the fixpoint — "nothing left to take away." Each round typically finds less; when the reviewers who drove the earlier cuts now find nothing, you're there. Confirm the major interface decisions hold in shipped code (not just prose), and that no remaining shape forces a future backward-incompatible break.
If the PIP has a reference implementation + CI
Keep the implementation green throughout, and decompose it for review:
- Split into a stacked PR series where each chunk builds on the last and is individually CI-triggerable. Defer deletions to the last chunk so every earlier chunk is forward-reference-free (a clean wholesale checkout from the green monolith).
- Prove equivalence: the last chunk's tree must equal the monolith (minus any excluded log), byte-for-byte. This is the correctness gate for the decomposition.
- CI discipline: verify the full local gate before pushing (not a subset); triage failures real-vs-flake by the actual failing test + its stack, not the group name; rerun known flakes; a sweeping default change (e.g. secure-by-default) breaks tests only CI's full matrix surfaces — expect a few principled iterations.
- Orchestration: one worktree owner at a time; use the raw git binary for ref/SHA-critical ops if a proxy tool rewrites them; keep a running implementation log of every decision + tradeoff.
Deliverable
A PIP where the surface is small enough to defend, every remaining element earns its place with a real consumer (or is a documented safety-guard / projection / seam), every deferred capability has a clean additive path, the major decisions are verified 5-year-durable, and the decision records make the "why" auditable. Simple achieved — nothing to add, nothing to take away.
