Ln 812 Optimization Researcher

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. If shared/ is missing, fetch files via WebFetch from https://raw.githubusercontent.com/levnikolaevich/claude-code-skills/master/skills/{path}.

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

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

Output

Target Metric Research

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

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:

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

Phase 3: Local Codebase Check

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

Impact on Feasibility

Phase 4: Generate Hypotheses (3-7)

Hypothesis Sources (Priority Order)

Hypothesis Format

Anti-Bias Checks (mandatory before finalizing)

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:

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

References

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

Runtime Summary Artifact

MANDATORY READ: Load shared/references/coordinator_summary_contract.md

Emit an optimization-worker summary envelope.

Managed mode:

• ln-810 passes deterministic runId and exact summaryArtifactPath
• write the summary to the provided summaryArtifactPath

Standalone mode:

• omit runId and summaryArtifactPath
• write .hex-skills/runtime-artifacts/runs/{run_id}/optimization-worker/ln-812--{identifier}.json

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 prepared with sources
• Optimization research artifact written to the shared location

Version: 2.0.0 Last Updated: 2026-03-14