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-812-optimization-researcher

Type: L3 Worker Category: 8XX Optimization

Researches competitive benchmarks, industry standards, and solution approaches for bottlenecks identified by the profiler. Generates prioritized hypotheses for the executor.

---

Overview

Aspect	Details
Input	Performance map from profiler (real measurements: baseline metrics, per-step timing, bottleneck classification, optimization hints)
Output	Industry benchmarks, solution candidates, prioritized hypotheses (H1..H7)
Pattern	Research-first: competitors → industry → local codebase → solutions → hypotheses

---

Workflow

Phases: Competitive Analysis → Bottleneck-Specific Research → Local Codebase Check → Hypothesis Generation → Research Report

---

Phase 1: Competitive Analysis

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

Goal

Establish what "good" looks like for this type of operation. Define target metric if user did not provide one.

Research Queries

Goal	Query Template	Tool
Industry benchmark	"{domain} API response time benchmark {year}"	WebSearch
Competitor performance	"{competitor_type} {operation} latency"	WebSearch
Standard expectations	"acceptable response time for {operation_type}"	WebSearch
Framework-specific guidance	"{framework} {operation} performance best practices"	Context7 / Ref

Output

Field	Description
industry_benchmark	Expected performance range for this operation type
competitor_approaches	How top systems solve this (2-3 examples)
recommended_target	Suggested target metric (if user did not specify)
target_metrics	Per-metric quantitative targets (see below)
sources	URLs with dates for all findings

Target Metric Research

For each metric present in performance_map.baseline, research a quantitative target:

Metric	Query Template	Tool
wall_time_ms	"{domain} API response time benchmark {year}"	WebSearch
cpu_time_ms	"{framework} handler CPU time benchmark"	WebSearch
memory_peak_mb	"{domain} API memory usage benchmark {year}"	WebSearch
http_round_trips	"{domain} API call count optimization best practice"	WebSearch
io_bytes	"{domain} file processing throughput benchmark"	WebSearch

Output format:

target_metrics:
  wall_time_ms:
    value: 500
    source: "industry benchmark: translation APIs p95 200-500ms"
    confidence: HIGH
  http_round_trips:
    value: 2
    source: "best practice: batch API reduces N calls to 1-2"
    confidence: HIGH
  memory_peak_mb:
    value: 128
    source: "similar workload: 64-128MB typical"
    confidence: MEDIUM

Confidence levels: HIGH = benchmark found with source, MEDIUM = derived from best practices, LOW = estimated from general guidelines. Only include metrics present in the profiler baseline.

---

Phase 2: Bottleneck-Specific Research

MANDATORY READ: Load research_query_templates.md for per-type query templates.

Research Strategy

Based on the primary bottleneck type from the profiler:

Bottleneck Type	Research Focus
Architecture	Batching, pipelining, parallelism, DataLoader pattern
I/O-Network	Connection pooling, HTTP/2, multiplexing, caching
I/O-DB	Query optimization, indexes, eager loading, bulk operations
I/O-File	Streaming, async I/O, memory-mapped files
CPU	Algorithm alternatives, vectorization, caching computation, OSS replacement
Cache	Eviction policies, cache key design, invalidation strategies, tiered caching, warm-up
External	Caching layer, circuit breaker, fallback strategies, provider alternatives

Research Protocol

FOR each top bottleneck (max 3):
  1. Select query templates from research_query_templates.md
  2. Execute research chain: Context7 → Ref → WebSearch (per research_tool_fallback.md)
  3. Collect solution approaches with expected impact
  4. Note technology prerequisites (libraries, infrastructure)

Solution Evaluation

Field	Description
solution	Name/description of the approach
source	Where found (URL, docs section)
expected_impact	Estimated improvement (e.g., "9x reduction for N=9")
complexity	Low / Medium / High
prerequisites	What's needed (library, infrastructure, API support)
feasibility	HIGH / MEDIUM / LOW — based on prerequisites availability

---

Phase 3: Local Codebase Check

Before recommending external solutions, check if the codebase already has the capability:

Check	How
Batch/bulk methods on client classes	Grep for batch, bulk, multi in client/service classes
Cache infrastructure	Grep for redis, memcache, cache, @cached, lru_cache
Connection pool configuration	Grep for pool_size, max_connections, pool in config
Async variants	Grep for async_, aio, Async prefix/suffix on methods
Unused configuration	Read client/service config for batch_size, max_connections params

Impact on Feasibility

Finding	Effect
Batch API exists, not used	Feasibility = HIGH, Complexity = LOW
Cache infra exists, not configured for this path	Feasibility = HIGH, Complexity = LOW-MEDIUM
No existing capability, requires new library	Feasibility = MEDIUM, Complexity = MEDIUM-HIGH
Requires infrastructure change	Feasibility = LOW, Complexity = HIGH

---

Phase 4: Generate Hypotheses (3-7)

Hypothesis Sources (Priority Order)

Priority	Source
1	Local codebase check (unused existing capabilities — lowest risk)
2	Research findings (proven patterns from industry)
3	Optimization hints from profiler

Hypothesis Format

Field	Description
id	H1, H2, ... H7
description	What to change and how
bottleneck_addressed	Which bottleneck from time map (step reference)
expected_impact	Estimated improvement % or multiplier
complexity	Low / Medium / High
risk	Low / Medium / High
files_to_modify	List of files that need changes
dependencies	Other hypotheses this depends on (e.g., "H2 requires H1")
conflicts_with	Hypotheses that become unnecessary if this one works

Anti-Bias Checks (mandatory before finalizing)

Bias	Check	Example
Removal bias	For each "remove X" hypothesis: generate paired "optimize X" alternative	"remove alignment" → also "optimize alignment config"
Industry bias	"Industry doesn't use X" ≠ "X not needed for us". Check: does OUR product need it?	"CAT tools skip alignment" but our users need it for quality
Premature conclusion	"X is slow" ≠ "X is wrong". Slow may mean bad implementation, not wrong approach	5.9s alignment → maybe wrong algorithm, not wrong feature

Rule: Every "remove feature" hypothesis MUST have a paired "optimize feature" hypothesis.

Fix Hierarchy (mandatory ordering)

Order hypotheses by fix level. Higher levels ALWAYS tried first:

Level	Example	Priority
1. Configuration	matching_methods="i", pool_size=10	Highest — try first
2. Infrastructure	Add cache layer, scale service
3. Framework	Use framework feature (batch API, built-in cache)
4. Application code	Refactor algorithm, add optimization
5. Feature removal	Remove functionality	Lowest — last resort only

Red flag: If highest-priority hypothesis is at level 4-5, re-examine: was a level 1-3 solution missed? Apply 5 Whys from root cause to verify.

Ordering Rules

Sort by: fix_level ASC, expected_impact DESC, complexity ASC, risk ASC.

Conflict detection: If H1 (batch API) solves the N+1 problem, H3 (parallel calls) becomes unnecessary. Mark H3.conflicts_with = ["H1"].

Dependency detection: If H2 (cache prefetch) builds on H1 (batch API), mark H2.dependencies = ["H1"].

---

Phase 5: Research Report

Report Structure

research_result:
  industry_benchmark:
    metric: "response_time"
    expected_range: "200-500ms"
    source: "..."
  recommended_target: 500          # ms — alias for target_metrics.wall_time_ms.value
  target_metrics:                  # per-metric quantitative targets
    wall_time_ms: { value: 500, source: "...", confidence: HIGH }
    http_round_trips: { value: 2, source: "...", confidence: HIGH }
    memory_peak_mb: { value: 128, source: "...", confidence: MEDIUM }
  competitor_analysis:
    - name, approach, metric, source
  solution_candidates:
    - solution, source, expected_impact, complexity, feasibility
  hypotheses:
    - id, description, bottleneck_addressed, expected_impact, complexity, risk,
      files_to_modify, dependencies, conflicts_with
  local_codebase_findings:
    - "Batch API exists: AlignmentClient.batch_align() — accepts up to 50 pairs"
    - "Redis configured but not used for alignment cache"
  research_sources:
    - url, date, relevance

---

Plan Mode

Read-only worker — all phases use MCP research tools (Ref, Context7, WebSearch) and code analysis only. Executes normally in Plan Mode via Skill().

---

Error Handling

Error	Recovery
All research tools fail	Use built-in knowledge with disclaimer: "no external sources verified"
No competitive benchmarks found	Skip industry benchmark, note "no baseline found — using general guidelines"
Cannot generate hypotheses	Return empty list — coordinator decides next step
Local codebase check finds nothing	Proceed with external research results only

---

References

• research_query_templates.md — query templates per bottleneck type
• shared/references/research_tool_fallback.md — MCP research tool priority chain

---

Definition of Done

• Competitive analysis completed (industry benchmarks, competitor approaches)
• Target metrics researched per baseline metric (value, source, confidence)
• Target metric defined (user-provided or derived from research)
• Bottleneck-specific solutions researched via MCP chain
• Local codebase checked for existing unused capabilities
• 3-7 hypotheses generated, ordered by expected impact
• Dependencies and conflicts between hypotheses identified
• Research report returned to coordinator with sources

---

Version: 2.0.0 Last Updated: 2026-03-14