rdr-research

RDR Research Skill

Optionally delegates to deep-research-synthesizer (sonnet) for evidence gathering, or codebase-deep-analyzer (sonnet) for code-specific questions.

When This Skill Activates

• User says "add research finding", "update RDR research", "verify assumption"
• User invokes /nx:rdr-research
• User wants to record or classify a discovery during RDR planning

Path Detection

Resolve RDR directory from .nexus.yml indexing.rdr_paths[0]; default docs/rdr. Use the Step 0 snippet from the rdr-create skill, stored as RDR_DIR. All file paths below use $RDR_DIR in place of docs/rdr.

Subcommands

/nx:rdr-research add <id>

Inputs (prompt the user):

1. Finding text: What was discovered
2. Classification: Verified (✅) | Documented (⚠️) | Assumed (❓)
3. Verification method: Source Search | Spike | Docs Only
4. Source: Code path, URL, experiment description

Behavior:

1. Determine next sequence number: mcp__plugin_nx_nexus__memory_get(project="{repo}_rdr", title="" Filter entries where title matches NNN-research-*. Parse sequence numbers. Next seq = max + 1. If none exist, start at 1.

2. Write T2 record: mcp__plugin_nx_nexus__memory_put(content="rdr_id: NNN\nseq: {seq}\nfinding: Finding text here\nclassification: verified\nverification_method: source_search\nsource: Source description here\nacknowledged: false", project="{repo}_rdr", title="NNN-research-{seq}", ttl="permanent", tags="rdr,research,{classification}"

3. Append to RDR markdown: Add a formatted entry to the Research Findings > Key Discoveries section:

   - **✅ Verified** (source search) — Finding text here
     *Source: source description*
   

4. Catalog citation link (optional, if catalog initialized and source is a paper/URL): If the source references a paper title or URL that may be indexed in the knowledge store:

   mcp__plugin_nx_nexus__catalog_search(query="<source title or keywords>")
   

   If a matching catalog entry is found, and the current RDR has a catalog entry:

   mcp__plugin_nx_nexus__catalog_link(from_tumbler="<rdr-title>", to_tumbler="<paper-tumbler>", link_type="cites", created_by="rdr-research")
   

   Skip silently if catalog not initialized, RDR not indexed, or no match found. This enriches the citation graph incrementally as research findings are added.

/nx:rdr-research status <id>

1. List T2 entries: mcp__plugin_nx_nexus__memory_get(project="{repo}_rdr", title=""
2. Filter titles matching NNN-research-*
3. Parse and display summary:

   RDR NNN Research Status:
   - Verified (✅): 3 (2 source search, 1 spike)
   - Documented (⚠️): 1 (1 docs only)
   - Assumed (❓): 2 — ⚠ unresolved risks
     - [seq 4] "Library X supports feature Y" (docs only)
     - [seq 6] "Latency under 100ms" (docs only)
   

/nx:rdr-research verify <id> <finding-seq>

1. Read T2 record: mcp__plugin_nx_nexus__memory_get(project="{repo}_rdr", title="NNN-research-{seq}"
2. Prompt for new classification (verified or documented) and updated verification method
3. Update T2 record (overwrite with updated content)
4. Update the emoji marker in the RDR markdown file (e.g., ❓ → ✅)

Agent Dispatch

When the user asks to investigate something (not just record a finding):

• Code questions ("how does auth work in our codebase?"): Dispatch codebase-deep-analyzer agent, then record the finding
• External research ("what embedding models support CCE?"): Dispatch deep-research-synthesizer agent, then record the finding
• Simple recording: No agent needed — just write the T2 record and update markdown

Relay Template (Use This Format)

When dispatching agents (deep-research-synthesizer or codebase-deep-analyzer) via Agent tool, use this exact structure:

## Relay: deep-research-synthesizer

**Task**: [1-2 sentence summary: e.g., "Investigate embedding model support for CCE to verify RDR 003 assumption"]
**Bead**: [ID] (status: [status]) or 'none'

### Input Artifacts
- nx store: [document titles or "none"]
- nx memory: {repo}_rdr/NNN (RDR metadata and existing research records)
- nx scratch: [scratch IDs or "none"]
- Files: $RDR_DIR/NNN-*.md

### Deliverable
Research finding with classification, verification method, and source reference ready for T2 storage.

### Quality Criteria
- [ ] Finding has clear classification (Verified/Documented/Assumed)
- [ ] Verification method specified (Source Search/Spike/Docs Only)
- [ ] Source traceable to specific code path, URL, or experiment

Required: All fields must be present. Agent will validate relay before starting.

For additional optional fields, see RELAY_TEMPLATE.md.

Success Criteria

• Research finding recorded with correct classification and verification method
• T2 record written with sequential ID (NNN-research-{seq})
• RDR markdown updated with formatted finding in Key Discoveries section
• Source traceable (code path, URL, or experiment description)
• Agent dispatched correctly for investigation tasks (code vs. external research)
• Existing findings not overwritten when adding new ones

Agent-Specific PRODUCE

Outputs generated by dispatched agents (deep-research-synthesizer, codebase-deep-analyzer):

• T3 knowledge: Research findings via store_put tool: content="# RDR NNN Research: {topic}\n{findings}", collection="knowledge", title="research-rdr-NNN-{topic}", tags="rdr,research"
• T2 memory: Finding records via memory_put tool: content="...", project="{repo}_rdr", title="NNN-research-{seq}", ttl="permanent", tags="rdr,research,{classification}"
• T1 scratch: Working notes during investigation via scratch tool: action="put", content="RDR NNN research: {hypothesis}", tags="rdr,research" (promoted to T2 on completion)

Session Scratch (T1): Agents use scratch tool for ephemeral working notes during the session. Flagged items auto-promote to T2 at session end.

Notes

• The markdown document is the authoritative narrative; T2 is the queryable index
• Verification method (Source Search / Spike / Docs Only) is the most load-bearing field — it distinguishes high-confidence from low-confidence findings
• Findings marked "Docs Only" on load-bearing assumptions are the highest risk items

Known Limitations

T2 retrieval is O(N): memory_get tool with project="{repo}_rdr", title="" returns all records (RDR metadata + all research findings). Client-side filtering by title pattern (NNN-research-*) and individual memory_get calls for content are required. For projects with many RDRs (30+ at 10 findings each), this may be slow. Validate that parsed records have rdr_id and seq fields before using them — do not rely solely on title pattern matching.