Agent Skills: Performance Optimizer

Paths: File paths (references/, ../ln-*) are relative to this skill directory.

Type: L2 Domain Coordinator Category: 8XX Optimization

Performance Optimizer

Runtime-backed multi-cycle optimization coordinator. Profiles, researches, validates, and executes optimization hypotheses until target reached, plateau detected, or budget exhausted.

Inputs

| Input | Required | Description | |-------|----------|-------------| | target | Yes | endpoint, function, or pipeline to optimize | | observed_metric | Yes | current performance problem | | target_metric | No | user or research-derived target | | max_cycles | No | default 3 |

Purpose & Scope

• Detect whether optimization is the right tool
• Run iterative cycles: profile -> gate -> research -> target -> context -> validate -> execute
• Preserve optimization artifacts under .hex-skills/optimization/{slug}/ and runtime state under .hex-skills/optimization/runtime/runs/{run_id}/
• Resume deterministically from the last checkpointed phase
• Keep cycle summaries machine-readable

Runtime Contract

MANDATORY READ: Load references/ci_tool_detection.md MANDATORY READ: Load references/coordinator_runtime_contract.md, references/optimization_runtime_contract.md, references/coordinator_summary_contract.md

Runtime CLI:

node references/scripts/optimization-runtime/cli.mjs start --slug {slug} --manifest-file .hex-skills/optimization/{slug}/manifest.json
node references/scripts/optimization-runtime/cli.mjs status --slug {slug}
node references/scripts/optimization-runtime/cli.mjs record-worker-result --payload '{...}'
node references/scripts/optimization-runtime/cli.mjs record-summary --payload '{...}'
node references/scripts/optimization-runtime/cli.mjs record-cycle --payload '{...}'
node references/scripts/optimization-runtime/cli.mjs checkpoint --phase PHASE_8_EXECUTE --payload '{...}'
node references/scripts/optimization-runtime/cli.mjs advance --to PHASE_9_CYCLE_BOUNDARY

Runtime Layout

Runtime state is run-scoped, while optimization artifacts stay slug-scoped:

.hex-skills/optimization/
  runtime/active/ln-810/{slug}.json
  runtime/runs/{run_id}/manifest.json
  runtime/runs/{run_id}/state.json
  runtime/runs/{run_id}/checkpoints.json
  runtime/runs/{run_id}/history.jsonl
  runtime-artifacts/runs/{run_id}/optimization-coordinator/ln-810--{slug}.json
  runtime-artifacts/runs/{run_id}/optimization-worker/{worker}--{child_identifier}.json
  {slug}/context.md
  {slug}/ln-814-log.tsv

Workflow

Phase 0: Preflight

1. Validate:
   • target identifiable
   • observed metric provided
   • git clean state
   • test infrastructure exists
2. Detect stack and optional service topology.
3. Derive slug.
4. Build manifest with:
   • slug
   • target
   • observed_metric
   • target_metric
   • execution_mode
   • cycle_config
5. Start runtime and checkpoint PHASE_0_PREFLIGHT.

Phase 1: Parse Input

1. Normalize the problem statement.
2. Set or defer target_metric.
3. Checkpoint PHASE_1_PARSE_INPUT.

Phase 2: Profile

1. Compute deterministic child metadata:
   • identifier=ln-811--{slug}--cycle-{current_cycle}
   • child run_id
   • exact summaryArtifactPath=.hex-skills/runtime-artifacts/runs/{parent_run_id}/optimization-worker/ln-811--{slug}--cycle-{current_cycle}.json
2. Checkpoint PHASE_2_PROFILE with child_run.
3. Invoke ln-811-performance-profiler with the child runId and exact summaryArtifactPath.
4. Read the emitted optimization-worker summary envelope from the exact artifact path.
5. Record the worker summary with record-worker-result.

Phase 3: Wrong Tool Gate

Evaluate profiler output:

| Gate | Meaning | Action | |------|---------|--------| | PROCEED | optimization work is justified | continue | | CONCERNS | measurements usable but imperfect | continue with warning | | BLOCK | wrong tool, already optimized, or infrastructure-bound | aggregate and exit | | WAIVED | user overrides BLOCK | continue with explicit waiver |

Rules:

• cycle 1 BLOCK -> finish as diagnostic result
• cycle 2+ BLOCK due to already_optimized or within_industry_norm -> finish as successful stop

Checkpoint PHASE_3_WRONG_TOOL_GATE with:

• gate_verdict
• stop_reason when blocked
• final_result when terminal

Phase 4: Research

1. Compute deterministic child metadata for ln-812.
2. Checkpoint PHASE_4_RESEARCH with child_run.
3. Invoke ln-812-optimization-researcher with the child runId and exact summaryArtifactPath.
4. Read and record the emitted optimization-worker summary envelope.
5. If no hypotheses remain, stop after aggregate/report.

Phase 5: Set Target

1. Resolve target metric:
   • user-specified target wins
   • otherwise use research target with confidence
   • otherwise default to 50% improvement
2. Checkpoint PHASE_5_SET_TARGET with target_metric.

Phase 6: Write Context

1. Build .hex-skills/optimization/{slug}/context.md.
2. Include:
   • problem statement
   • performance map
   • target metrics
   • hypotheses and conflicts
   • local codebase findings
   • previous cycles
3. Checkpoint PHASE_6_WRITE_CONTEXT with context_file.

Phase 7: Validate Plan

1. Compute deterministic child metadata for ln-813.
2. Checkpoint PHASE_7_VALIDATE_PLAN with validation_verdict and child_run.
3. Invoke ln-813-optimization-plan-validator with the child runId and exact summaryArtifactPath.
4. Read and record the emitted evaluation-coordinator summary envelope.
5. If verdict is NO_GO, pause runtime until user resolves or waives.

Phase 8: Execute

execution_mode=execute:

1. Compute deterministic child metadata for ln-814.
2. Checkpoint PHASE_8_EXECUTE with child_run.
3. Invoke ln-814-optimization-executor with the child runId and exact summaryArtifactPath.
4. Read and record the emitted optimization-worker summary envelope.

execution_mode=plan_only:

1. Do not run ln-814.
2. Checkpoint PHASE_8_EXECUTE as skipped_by_mode.

Phase 9: Cycle Boundary

1. Record the cycle summary with record-cycle.
2. Evaluate stop conditions:
   • target met
   • plateau
   • max cycles reached
   • no new hypotheses
3. If continuing:
   • merge previous branch when needed
   • increment current_cycle
   • checkpoint PHASE_9_CYCLE_BOUNDARY
   • advance back to PHASE_2_PROFILE
4. If stopping:
   • checkpoint PHASE_9_CYCLE_BOUNDARY with stop_reason
   • advance to PHASE_10_AGGREGATE

Phase 10: Aggregate

1. Aggregate all cycle summaries from runtime state.
2. Compute cumulative improvement.
3. Checkpoint PHASE_10_AGGREGATE.

Phase 11: Report

1. Produce final report with:
   • per-cycle summary
   • cumulative improvement
   • final result
   • gap analysis when target not met
2. Checkpoint PHASE_11_REPORT with:
   • report_ready=true
   • final_result
3. Record the optimization-coordinator summary envelope with record-summary.
4. Complete runtime only after the report checkpoint and coordinator summary exist.

Worker Invocation (MANDATORY)

Host Skill Invocation: Skill(skill: "...", args: "...") is mandatory delegation.

• Claude: call the Skill tool exactly as shown.
• Codex: if no Skill tool exists, locate the named skill in available skills, read its SKILL.md, treat args as $ARGUMENTS, execute that skill workflow, then return here with its result/artifact.
• Do not inline worker logic or mark the worker complete without executing the target skill.

| Phase | Skill | Purpose | |-------|--------|---------| | 2 | ln-811-performance-profiler | Build measured performance map | | 4 | ln-812-optimization-researcher | Research hypotheses and targets | | 7 | ln-813-optimization-plan-validator | Validate feasibility via evaluation-platform review (L2 Coordinator) | | 8 | ln-814-optimization-executor | Execute optimization strike and bisect |

Skill(skill: "ln-811-performance-profiler")
Skill(skill: "ln-812-optimization-researcher")
Agent(... Skill(skill: "ln-813-optimization-plan-validator"))
Agent(... Skill(skill: "ln-814-optimization-executor"))

TodoWrite format (mandatory)

- Start ln-810 runtime (pending)
- Run profiler and record summary (pending)
- Apply Wrong Tool Gate (pending)
- Run researcher and record summary (pending)
- Set target metric (pending)
- Write optimization context (pending)
- Validate plan and record summary (pending)
- Execute or skip by mode (pending)
- Record cycle boundary (pending)
- Aggregate results and write final report (pending)

Critical Rules

• Runtime state is the optimization orchestration SSOT.
• Worker outputs are consumed only through summary JSON artifacts.
• plan_only is a first-class execution mode, not an informal branch.
• NO_GO from ln-813 must pause runtime until explicitly resolved.
• Cycle history lives in runtime state, not in chat memory.
• A terminal diagnostic result is still a valid DONE orchestration outcome.

Definition of Done

• [ ] Runtime started and preflight/input checkpoints recorded
• [ ] Profiler and researcher summaries recorded deterministically
• [ ] Wrong Tool Gate result checkpointed
• [ ] Target metric and context file checkpointed
• [ ] Validator summary recorded; NO_GO handled via PAUSED when needed
• [ ] Executor summary recorded or skipped_by_mode checkpointed
• [ ] Cycle boundary recorded for every completed cycle
• [ ] Aggregate and final report checkpoints recorded
• [ ] Runtime completed with final result and resume-free terminal state

Phase 12: Meta-Analysis

Optional reference: load references/meta_analysis_protocol.md only when the user asks for post-run meta-analysis or protocol-formatted run reflection.

Skill type: optimization-coordinator. When requested, run after phases complete. Output to chat using the optimization-coordinator format.

Reference Files

• references/coordinator_runtime_contract.md
• references/optimization_runtime_contract.md
• references/coordinator_summary_contract.md
• ../ln-811-performance-profiler/SKILL.md
• ../ln-812-optimization-researcher/SKILL.md
• ../ln-813-optimization-plan-validator/SKILL.md
• ../ln-814-optimization-executor/SKILL.md

Version: 3.0.0 Last Updated: 2026-03-15