rdr-gate

RDR Gate Skill

Optional validation for high-stakes decisions. Most RDRs don't need a formal gate — use this when the decision is expensive to reverse and you want to confront what you don't actually know before committing.

Delegates Layer 3 to the substantive-critic agent (sonnet). See registry.yaml.

When This Skill Activates

• User says "gate this RDR", "finalization check", "is this RDR ready?"
• User invokes /nx:rdr-gate
• User wants to validate an RDR before locking it as Final

Input

• RDR ID (required) — e.g., 003

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.

Three Validation Layers (run in sequence)

Layer 1 — Structural Validation (no AI)

Read the RDR markdown file. Check that these sections are present AND non-empty (not just the heading with placeholder text):

• Problem Statement
• Context (with Background and Technical Environment subsections)
• Research Findings (with Investigation and Key Discoveries subsections)
• Proposed Solution (with Approach and Technical Design subsections)
• Alternatives Considered (at least one alternative with Pros/Cons/Rejection reason)
• Trade-offs (with Consequences and Risks subsections)
• Implementation Plan / Approach / Steps (with at least one numbered Phase/Step/Stage)
• Finalization Gate (must have written responses, not just template placeholders)

If any section is missing or contains only placeholder text (e.g., [What is the specific challenge]):

• Report which sections are incomplete
• STOP — do not proceed to Layer 2 or 3
• Status remains Draft

Layer 2 — Assumption Audit (from T2, no AI)

mcp__plugin_nx_nexus__memory_get(project="{repo}_rdr", title=""

Filter entries matching NNN-research-*. Analyze:

1. Count by classification: verified, documented, assumed
2. Count by verification method: source_search, spike, docs_only
3. Flag high-risk items: classification=assumed AND verification_method=docs_only

Display:

Assumption Audit for RDR NNN:
- 3 verified (2 source search, 1 spike)
- 1 documented (docs only)
- 2 assumed — ⚠ UNRESOLVED
  [seq 4] "Library X supports feature Y" (docs only) ← HIGH RISK
  [seq 6] "Latency under 100ms" (docs only) ← HIGH RISK

If assumed findings remain:

• Ask: "Proceed with 2 unverified assumptions? (recorded as acknowledged)"
• If yes: update T2 records with acknowledged: true
• If no: STOP — user should verify or remove assumptions first

Layer 3 — AI Critique (substantive-critic agent)

Dispatch the substantive-critic agent via Agent tool with this relay:

## Relay: substantive-critic

**Task**: Critique RDR NNN for internal consistency, missing failure modes, scope creep, and proportionality.
**Bead**: none

### Input Artifacts
- nx store: none
- nx memory: {repo}_rdr/NNN (status and research records)
- Files: docs/rdr/NNN-*.md

### Deliverable
Structured critique with pass/warn/fail per finalization gate criterion:
1. Contradiction Check — pass/warn/fail
2. Assumption Verification — pass/warn/fail
3. Scope Verification — pass/warn/fail
4. Cross-Cutting Concerns — pass/warn/fail
5. Proportionality — pass/warn/fail

### Quality Criteria
- [ ] Every fail has a specific section reference and fix suggestion
- [ ] Warns are actionable but non-blocking
- [ ] Prior RDR search attempted (may return empty on cold-start)

Prior-art search (within the agent): Use catalog if available, fall back to raw search:

• First, try catalog (structured metadata): mcp__plugin_nx_nexus__catalog_search(query="relevant terms from problem statement", content_type="rdr")
  • If results found, use mcp__plugin_nx_nexus__catalog_links(tumbler="<result>", direction="both") to discover related RDRs in the graph
• If catalog empty or not initialized, fall back to T3 semantic search:
  • Use store_list tool to enumerate collections, filter for rdr__
  • mcp__plugin_nx_nexus__search(query="relevant query terms from RDR problem statement", corpus="{each_collection}", limit=5 If no collections found: "No prior RDRs indexed. Cross-project prior-art search will improve as RDRs are indexed and closed."

Gate Aggregation

• Any fail → gate fails. Status remains Draft.
• Warns only → gate passes. Warns surfaced to user but do not block.
• All pass → gate passes.

Important: The AI critique supplements but does not replace the author completing the Finalization Gate section with written responses. The gate should verify that the Finalization Gate section contains substantive written responses, not just "N/A" or placeholder text.

On Pass

1. Write gate result to T2: mcp__plugin_nx_nexus__memory_put(content="outcome: PASSED\ndate: YYYY-MM-DD\ncritical_count: 0\nsignificant_count: N\nobservation_count: N\nsummary: One-sentence summary", project="{repo}_rdr", title="{id}-gate-latest", ttl="permanent", tags="rdr,gate"
2. Append gate findings to the RDR's Revision History section
3. Print: > Run '/nx:rdr-accept <id>' to accept this RDR.

Status remains Draft until the author explicitly accepts via /nx:rdr-accept.

On Fail

1. Write gate result to T2 (same format, outcome: "BLOCKED")
2. Display the critique with specific sections to address
3. Status remains Draft

Relay Template (Use This Format)

When dispatching the substantive-critic agent via Agent tool for Layer 3 critique, use this exact structure:

## Relay: substantive-critic

**Task**: Critique RDR NNN for internal consistency, missing failure modes, scope creep, and proportionality.
**Bead**: [ID] (status: [status]) or 'none'

### Input Artifacts
- nx store: [prior RDR collections or "none"]
- nx memory: {repo}_rdr/NNN (status and research records)
- nx scratch: [scratch IDs from Layer 1/2 or "none"]
- Files: docs/rdr/NNN-*.md

### Deliverable
Structured critique with pass/warn/fail per finalization gate criterion:
1. Contradiction Check
2. Assumption Verification
3. Scope Verification
4. Cross-Cutting Concerns
5. Proportionality

### Quality Criteria
- [ ] Every fail has a specific section reference and fix suggestion
- [ ] Warns are actionable but non-blocking
- [ ] Prior RDR search attempted (may return empty on cold-start)

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

For additional optional fields, see RELAY_TEMPLATE.md.

Success Criteria

• RDR directory resolved from .nexus.yml indexing.rdr_paths[0] (default docs/rdr)
• Layer 1 structural validation completed (all required sections present and non-empty)
• Layer 2 assumption audit completed (findings counted by classification and method)
• High-risk items flagged (classification=assumed AND verification_method=docs_only)
• Layer 3 AI critique dispatched and results aggregated
• Gate result determined: pass (all pass or warns only) or fail (any fail)
• Gate result written to T2 as {id}-gate-latest (both pass and fail)
• On pass: gate findings appended to Revision History, accept prompt displayed
• On fail: specific sections to address displayed to user

Agent-Specific PRODUCE

Outputs generated by the substantive-critic agent (Layer 3):

• T3 knowledge: Gate results via store_put tool: content="# Gate: RDR NNN\n{critique}", collection="knowledge", title="gate-rdr-NNN-{date}", tags="rdr,gate,critique"
• T2 memory: Gate result record via memory_put tool: project="{repo}_rdr", title="{id}-gate-latest", ttl="permanent", tags="rdr,gate" (outcome: PASSED or BLOCKED)
• T1 scratch: Layer 1/2 validation notes via scratch tool: action="put", content="Gate RDR NNN: Layer 1 structural check", tags="rdr,gate" (promoted to T2 on completion)

Session Scratch (T1): Use scratch tool for ephemeral notes during multi-layer validation. Flagged items auto-promote to T2 at session end.

Known Limitations

T2 retrieval is O(N): Layer 2's memory_get tool with project="{repo}_rdr", title="" returns all records. Client-side filtering by title pattern (NNN-research-*) is required. Validate that parsed records have rdr_id and seq fields before using them.