ln-811-algorithm-optimizer（ln 811 算法优化器）

Paths: File paths (shared/, references/, ../ln-*) are relative to skills repo root. If not found at CWD, locate this SKILL.md directory and go up one level for repo root.

ln-811-algorithm-optimizer

Type: L3 Worker Category: 8XX Optimization Parent: ln-810-performance-optimization-coordinator

Optimizes target function performance via autoresearch loop: establish baseline benchmark, research best practices, generate 3-7 hypotheses, apply each with keep/discard verification.

Overview

Workflow

Phases: Pre-flight → Baseline → Research → Hypothesize → Optimize Loop → Report

Phase 0: Pre-flight Checks

MANDATORY READ: Load shared/references/ci_tool_detection.md — use Benchmarks + Test Frameworks sections.

Coverage Verification

Before starting optimization, verify target function has test coverage:

Without test coverage, benchmark improvements are meaningless — the optimized function may produce wrong results faster.

Worktree & Branch Isolation

MANDATORY READ: Load shared/references/git_worktree_fallback.md — use ln-811 row.

All work (edits, benchmarks, KEEP commits) in worktree. Never modify main worktree.

Phase 1: Establish Baseline

Step 1.1: Detect or Generate Benchmark

Step 1.2: Run Baseline

Save baseline result — all improvements measured against this.

Phase 2: Research Best Practices

MANDATORY READ: Load shared/references/research_tool_fallback.md for MCP tool chain.

Research Strategy

Research Output

Collect optimization techniques applicable to the target function. For each technique note:

• Name and description
• Expected improvement category (algorithmic complexity, memory, cache, parallelism)
• Applicability conditions (data size, structure, language features)

Phase 3: Generate Hypotheses (3-7)

Hypothesis Sources

MANDATORY READ: Load optimization_categories.md for category checklist.

Hypothesis Format

Ordering

Sort by: expected_impact DESC, risk ASC. Independent hypotheses first (no dependencies).

Phase 4: Optimize Loop (Keep/Discard)

Per-Hypothesis Cycle

FOR each hypothesis (H1..H7):
  1. APPLY: Edit target function (surgical change, function body only)
  2. VERIFY: Run tests
     IF tests FAIL (assertion) → DISCARD (revert) → next hypothesis
     IF tests CRASH (runtime error, OOM, import error):
       IF fixable (typo, missing import) → fix & re-run ONCE
       IF fundamental (design flaw, incompatible API) → DISCARD + log "crash: {reason}"
  3. BENCHMARK: Run 5 times, take median
  4. COMPARE: improvement = (baseline - new) / baseline * 100
     IF improvement >= 10% → KEEP:
       git add {target_file}
       git commit -m "perf(H{N}): {description} (+{improvement}%)"
       new baseline = new median
     IF improvement < 10%  → DISCARD (revert edit)
  5. LOG: Record result to experiment log + report

Safety Rules

Keep/Discard Decision

Simplicity criterion (per autoresearch): If improvement is marginal (10-20%) and change significantly increases code complexity (>50% more lines, deeply nested logic, hard-to-read constructs), prefer DISCARD. Simpler code at near-equal performance wins.

Phase 5: Report Results

Report Schema

Experiment Log

Write to {project_root}/.optimization/ln-811-log.tsv:

Append to existing file if present (enables tracking across multiple runs).

Cleanup

Configuration

Options:
  # Target
  target_file: ""
  target_function: ""

  # Benchmark
  benchmark_runs: 5
  improvement_threshold: 10    # percent
  warmup_runs: 1

  # Hypotheses
  max_hypotheses: 7
  min_hypotheses: 3

  # Safety
  time_budget_minutes: 30
  allow_new_deps: false
  scope: "function_body"       # function_body | module

  # Verification
  run_tests: true
  run_lint: false

Error Handling

References

• benchmark_generation.md
• optimization_categories.md
• shared/references/ci_tool_detection.md (Benchmarks section)
• shared/references/research_tool_fallback.md

Definition of Done

• Test coverage for target function verified before optimization
• Target function identified and baseline benchmark established (5 runs, median)
• Research completed via MCP tool chain (Context7/Ref/WebSearch)
• 3-7 hypotheses generated, ordered by expected impact
• Each hypothesis tested: apply → tests → benchmark → keep/discard
• Each kept optimization = separate git commit with hypothesis ID
• Kept optimizations compound (each becomes new baseline)
• Marginal gains (10-20%) with complexity increase reviewed via simplicity criterion
• Tests pass after all kept optimizations
• Experiment log written to .optimization/ln-811-log.tsv
• Report returned with baseline, final, improvement%, per-hypothesis results
• Generated benchmark cleaned up if no optimizations kept
• All changes on isolated branch, pushed to remote

Version: 1.0.0 Last Updated: 2026-03-08