Improve performance end to end

Perf Max

The "fast" pillar of the triad (its siblings: bug-zero = correct, ux-maximizer = beautiful). Point it at working code and drive the latency down by measurement, not guesswork. It conducts the profiling + optimization skills you already have and adds the discipline they need: profile first, optimize the real hotspot, prove behavior is unchanged, measure the win, and re-profile because the bottleneck moves.

Core Insight: Performance work fails two ways: optimizing by vibe (you speed up code that wasn't the bottleneck) and optimizing too early (you lock in structure the project hasn't earned). This vertical refuses both: every change targets a profiler-scored hotspot, and the whole pass is gated to settled code. Like its siblings, it's a successive-prompter with a green ratchet: a "faster" change that alters behavior or reds a test is reverted, not kept.

What perf-max adds (the only new logic)

1. Profile-first discipline: no optimization without a measured target. Vibe- tuning is forbidden; the profiler picks the work.
2. Behavior proof per change: output identical + tests green, or the "speedup" reverts. Fast and wrong is just wrong.
3. Measure the win: before/after numbers per change; an optimization with no measurable gain gets reverted (don't keep complexity that didn't pay).
4. Re-profile loop: the hotspot moves after you fix the top one; re-profile and stop on saturation (diminishing speedup per effort), not guesswork.

Preconditions / guards

1. Late-stage only. Optimization is a maturity activity. If features are still churning (>10 modified files OR >3 in-flight beads, or core goals not yet WORKING), STOP. Premature optimization wastes passes on code about to be rewritten and can lock in the wrong structure. Profile a thing you're about to ship.
2. A representative workload exists. You can't optimize what you can't measure. If there's no realistic benchmark/usage to profile against, establish one FIRST; that's the precondition, not optional.
3. Green baseline + checkpoint before any change.

The Pipeline (staged: issue each step only when the prior finishes)

Phase 0  ORIENT    Read code + intent; establish a representative workload + a green
                   baseline; checkpoint. (Re-read AGENTS.md/README, deep-investigate.)
Phase 1  PROFILE   hotspot-scout → a SCORED hotspot list (CPU / mem
                   / I/O / contention) under real-world usage. No targets → no work.
Phase 2  OPTIMIZE  optimize-proof on the top hotspot(s), ONE at a time,
                   under the green ratchet. deepen to grind the list
                   systematically with progressive deepening.
Phase 3  PROVE     Behavior proof (output identical, tests green; golden/metamorphic
                   tests where apt) AND measure the win (before/after p95/throughput).
                   No measurable gain or any behavior change → revert that change.
Phase 4  RE-PROFILE Re-profile: the bottleneck has moved. Stop when K=2 passes find
                   no material win left (saturation). Else → Phase 2 on the new top.

Skills it orchestrates

100% your installed arsenal: perf-max adds sequencing + proof, not profilers. Full map + staged prompts in references/ROUTING.md. In brief: profile with hotspot-scout; optimize with optimize-proof (driven by deepen); prove with optimize-proof's behavior proofs, plus verify / testing-golden-artifacts / testing-metamorphic if installed (behavior unchanged); escalate to system-performance-remediation (the machine is the bottleneck, not the code), deadlock-finder-and-fixer (lock contention), deps-update (a slow dep), each if installed (otherwise handle it directly), or simplify-proof (the profile keeps pointing at tangled code: the real cost is complexity, not compute; it's optimize-proof's sibling, same proof spine, objective inverted to a measured complexity drop).

Shared mechanisms (inherited from project-autopilot)

Green ratchet, decision inbox, saturation convergence, capability-aware execution, as in project-autopilot; see its CONVERGENCE.md / SAFETY-RAILS.md / EXECUTION.md rather than duplicating. Integration: project-autopilot's performance lens / late-stage optimization pass routes here for the exhaustive treatment, exactly as the UX lens routes to ux-maximizer and the correctness lens to bug-zero.

THE EXACT PROMPT

/perf-max
Target: <project path or hot subsystem>
Workload: <the benchmark / realistic usage to profile against>
Mode: solo | swarm

Anti-Patterns

• Optimizing by vibe instead of a profiler-scored target: the cardinal sin.
• Running it early, on a still-churning codebase (premature optimization).
• Keeping a "faster" change that altered behavior or reds a test.
• Keeping an optimization with no measurable win (complexity that didn't pay).
• Profiling once. The hotspot moves after each fix; re-profile every round.
• Optimizing without a representative workload (you're tuning to a fiction).
• Grinding past saturation instead of stopping when the wins dry up.

References

• ROUTING.md: profile/optimize/prove/escalate map + staged prompts