• NEVER skip Level 1 unless user explicitly requests starting at Level 2/3 (via /research [topic] --level 2 or "Start with Level 2 investigation of [topic]")
• NEVER use serial investigation at Level 3 (must be parallel)
• NEVER use Sonnet at Level 3 (must be Opus)
• NEVER forget extended thinking at Level 3
• NEVER implement solutions (always delegate to plugin-improve)

Level 1: Quick Check (5-10 min, Sonnet, no extended thinking)

Goal: Find quick answer from local knowledge base or JUCE API docs

Sources: Local troubleshooting docs, Context7 JUCE documentation

Exit criteria: HIGH confidence solution → present decision menu, otherwise escalate to Level 2

See references/research-protocol.md#level-1-quick-check for detailed process.

Level 2: Moderate Investigation (15-30 min, Sonnet, no extended thinking)

Goal: Deep-dive JUCE docs, forums, GitHub for authoritative answers

Sources: Context7 deep-dive, JUCE forum search, GitHub issue search

Exit criteria: MEDIUM-HIGH confidence solution → present decision menu, otherwise escalate to Level 3

See references/research-protocol.md#level-2-moderate-investigation for detailed process.

Level 3: Deep Research (30-60 min, Opus, extended thinking 15k budget)

Goal: Parallel subagent investigation for novel/complex problems

Model requirements: claude-opus-4-1-20250805 with extended-thinking (15k budget)

Process: Spawn 2-3 parallel research subagents via Task tool, synthesize findings

Exit criteria: ALWAYS present decision menu (no further escalation)

See references/research-protocol.md#level-3-deep-research for detailed process.

Report Generation

Each level generates a structured report using templates in assets/:

• Level 1: assets/level1-report-template.md
• Level 2: assets/level2-report-template.md
• Level 3: assets/level3-report-template.md

Reports include: findings summary, confidence assessment, recommended solution, and source references.

Progress tracking: Use assets/research-progress.md template to track investigation progress across levels.

Decision Menus

After each level, present decision menu using checkpoint protocol format:

Example (Level 1 - HIGH confidence):

✓ Level 1 complete (found solution in local docs)

Solution: [Brief description of solution]
Source: troubleshooting/[category]/[file].md
Confidence: HIGH (exact match)

What's next?

1. Apply solution (recommended)
2. Review full findings
3. Continue deeper - Escalate to Level 2
4. Other

Choose (1-4): _

Example (Level 2 - MEDIUM confidence):

✓ Level 2 complete (found 2 potential solutions)

Recommended: [Solution 1 name]
Alternative: [Solution 2 name]
Confidence: MEDIUM (verified by JUCE forum + docs)

What's next?

1. Apply recommended solution
2. Review all findings
3. Try alternative approach
4. Continue deeper - Escalate to Level 3
5. Other

Choose (1-5): _

Example (Level 3 - comprehensive investigation):

✓ Level 3 complete (parallel investigation)

Investigated 3 approaches:
- Approach A: [Brief description] (recommended)
- Approach B: [Brief description] (viable alternative)
- Approach C: [Brief description] (not recommended)

Confidence: HIGH after synthesis

What's next?

1. Apply recommended solution (recommended)
2. Review detailed comparison
3. Try alternative approach B
4. Document findings
5. Other

Choose (1-5): _

NEVER use AskUserQuestion tool for decision menus. Always use inline numbered lists with "Choose (1-N): _" format matching checkpoint protocol.

See <state_requirement name="checkpoint_protocol"> below for response handling.

Integration Points

See references/integrations.md for troubleshooter and troubleshooting-docs integration details.

Error Handling

See references/error-handling.md for timeout, failure, and fallback patterns.