deep-analysis

name	deep-analysis
description	Use when surface-level analysis is insufficient and problems require hypothesis-driven investigation across multiple system components
effort	high

Tier-aware discipline — apply at session start and before every major step:

1. Read widest → narrowest before duplicating effort:
   • T3 (cross-project): mcp__plugin_nx_nexus__nx_answer(...) for verb-shape questions; mcp__plugin_nx_nexus__search(...) for keyword lookup.
   • T2 (project): mcp__plugin_nx_nexus__memory_search(query="<topic>", project="<repo>").
   • T1 (siblings, this session): mcp__plugin_nx_nexus__scratch(action="search", query="<topic>").
2. Reuse plans before dispatching multiple agents: mcp__plugin_nx_nexus__plan_search(query="<task>", limit=3).
3. Write back at end — findings not stored are findings lost. Pick the tier that matches the audience:
   • mcp__plugin_nx_nexus__scratch(action="put", ..., tags="<topic>") for sibling agents downstream THIS session (T1, narrowest scope, cheapest write).
   • mcp__plugin_nx_nexus__memory_put(...) for project-scoped decisions, future sessions same project (T2).
   • mcp__plugin_nx_nexus__store_put(...) for permanent cross-project knowledge, future sessions everywhere (T3).
   • mcp__plugin_nx_nexus__plan_save(...) for multi-agent pipeline outcomes (so future callers hit plan-match).

Deep Analysis Skill

Delegates to the deep-analyst agent.

Model Selection

Default: sonnet. Escalate via model parameter on the Agent tool:

Task Shape	Model	When
Single-module or focused question	sonnet (default)	Most analysis tasks
>3 modules, system-level, or novel architecture	opus	Cross-cutting investigations

When This Skill Activates

• When investigating performance mysteries
• When debugging multi-component interactions
• When understanding complex system behavior
• When surface-level analysis is insufficient
• When root cause analysis requires deep investigation
• After debugger if issue is cross-cutting

Pre-Dispatch: Seed Link Context

Before dispatching the deep-analyst agent, seed T1 scratch with link targets so the auto-linker can create catalog links when the agent stores findings:

1. If the task references an RDR (pattern RDR-\d+) or a known document, resolve it: mcp__plugin_nx_nexus-catalog__search(query="RDR-NNN or document title")
2. Check T1 scratch for rdr-planning-context
3. Write link context to scratch:

   mcp__plugin_nx_nexus__scratch(action="put", content='{"targets": [{"tumbler": "<resolved-tumbler>", "link_type": "relates"}], "source_agent": "deep-analyst"}', tags="link-context")
   

4. If no RDR/document reference found, skip seeding (the auto-linker handles empty context gracefully)

Agent Invocation

Use the Agent tool to invoke deep-analyst:

## Relay: deep-analyst

**Task**: [what needs to be done]
**Bead**: [ID] or 'none'

### Input Artifacts
- Files: [relevant files]

### Deliverable
Analysis report with findings and recommendations

### Quality Criteria
- [ ] Multiple hypotheses explored
- [ ] Root cause(s) identified with confidence
- [ ] Recommendations are actionable

For full relay structure and optional fields, see RELAY_TEMPLATE.md.

Investigation Methodology

The deep-analyst uses mcp__plugin_nx_sequential-thinking__sequentialthinking:

1. Form initial hypothesis about the problem
2. Identify evidence needed to validate/refute
3. Gather evidence systematically — use query(question=..., subtree=..., follow_links="cites") for citation-aware evidence gathering, or query(question=..., content_type=...) for type-scoped retrieval
4. Evaluate hypothesis against evidence
5. If refuted, branch to new hypothesis; iterate until root cause found
6. Synthesize findings and provide actionable recommendations

Success Criteria

• Problem clearly understood and scoped
• Multiple hypotheses explored
• Root cause(s) identified with confidence
• Conclusions supported by evidence
• Recommendations are actionable
• Findings stored in nx store for future reference
• T2 memory updated with session findings (if multi-session work)

Agent-Specific PRODUCE

• Analysis Findings: Store in nx T3 via store_put tool: content="# Analysis: {topic}\n{findings}", collection="knowledge", title="analysis-{topic}-{date}", tags="analysis"
• Hypothesis Results: Document with confidence levels in nx T3
• Recommendations: Include in output as "Recommended Next Step" for caller to dispatch strategic-planner
• Analysis Chain: Use T1 scratch to track hypothesis progression during investigation:
  • scratch tool: action="put", content="Analysis step {N}: {hypothesis}\nEvidence: {evidence}\nConfidence: {level}", tags="analysis,step-{N}"
  • scratch_manage tool: action="promote", entry_id="", project="{project}", title="analysis-chain.md"