goldmask

Goldmask Skill — Cross-Layer Impact Analysis

Load skills: rune-orchestration, context-weaving, team-sdk, polling-guard, zsh-compat

Three-layer investigation that answers: WHAT must change (Impact), WHY it was built that way (Wisdom), and HOW RISKY the area is (Lore). Includes Collateral Damage Detection (CDD) to predict blast radius.

ATE-1 ENFORCEMENT

CRITICAL: Every Agent call in this skill MUST include team_name. Bare Agent calls without team_name cause context explosion and are blocked by the enforce-teams.sh hook.

Architecture — Goldmask's Three Eyes

                    +---------------------------------------+
                    |       GOLDMASK COORDINATOR            |
                    |       (Sonnet)                        |
                    |                                       |
                    |  Synthesize all three layers:          |
                    |  1. WHAT must change (Impact)          |
                    |  2. WHY it was built (Wisdom)          |
                    |  3. HOW RISKY is this area (Lore)      |
                    |                                       |
                    |  Produce: GOLDMASK.md + findings.json  |
                    +-------------------+-------------------+
                                        | reads all layer outputs
            +--------------------------++--------------------------+
            |                           |                           |
    +-------v--------+   +-------------v---------+   +-------------v------+
    |  IMPACT LAYER   |   |   WISDOM LAYER         |   |  LORE LAYER        |
    |  5 Tracers      |   |   1 Sage (Sonnet)       |   |  1 Analyst (Haiku) |
    |  (Haiku each)   |   |   - git blame            |   |  - git log          |
    |  - grep/glob    |   |   - commit context       |   |  - risk scoring     |
    |  - dependency   |   |   - intent classify      |   |  - co-change graph  |
    |    tracing      |   |   - caution scoring      |   |  - hotspot detect   |
    +----------------+   +------------------------+   +-------------------+

Modes

Mode	Trigger	Agents	Output
Full investigation	/rune:goldmask <diff-spec>	8 (5+1+1+1)	GOLDMASK.md + findings.json + risk-map.json
Quick check	/rune:goldmask --quick <files>	0 (deterministic)	Warnings only — compares predicted vs actual
Intelligence	/rune:goldmask --lore <diff-spec>	1 (Lore only)	risk-map.json for file sorting

MCP-First Tracer Discovery (v1.170.0+)

Investigation agents (tracers) can be discovered via MCP search:

# Phase 1: Tracer Selection
tracers = []

if mcp_available:
  candidates = agent_search({
    query: "impact analysis tracing investigation risk",
    phase: "goldmask",
    category: "investigation",
    limit: 10
  })
  tracers = candidates
  Bash("mkdir -p tmp/.rune-signals && touch tmp/.rune-signals/.agent-search-called")

if not tracers:
  # Fallback: hardcoded tracer list (current behavior)
  tracers = DEFAULT_GOLDMASK_TRACERS

This enables user-defined investigation agents (e.g., "compliance-tracer" for audit trails) to participate in Goldmask analysis alongside the 8 built-in tracers.

Phase Sequencing

Phase 1: LORE ANALYSIS (parallel with Phase 2)
    |  Lore Analyst --> risk-map.json
    |  (Haiku, ~15-30s)
    |
Phase 2: IMPACT TRACING (parallel with Phase 1, 5 tracers)
    |  Data, API, Business, Event, Config Tracers --> 5 reports
    |  (Haiku x 5, ~30-60s)
    |
    +-- Lore + Impact complete --> Phase 3 starts
    |
Phase 3: WISDOM INVESTIGATION (sequential — needs Impact output)
    |  Wisdom Sage receives:
    |    - All MUST-CHANGE + SHOULD-CHECK findings from Impact
    |    - risk-map.json from Lore (for risk tier context)
    |  Wisdom Sage --> intent classifications + caution scores
    |  (Sonnet, ~60-120s)
    |
Phase 3.5: CODEX RISK AMPLIFICATION (parallel with Phase 3, v1.51.0+)
    |  Codex traces 2nd/3rd-order risk chains
    |  Reads Impact outputs + risk-map.json
    |  --> risk-amplification.md (CDX-RISK prefix)
    |  (codex exec, opt-in, ~600s)
    |
Phase 4: COORDINATION + CDD
    |  Goldmask Coordinator merges all three layers
    |  --> GOLDMASK.md + findings.json
    |  (Sonnet, ~60-90s)
    |
Total estimated time: 3-5 minutes
Total agents: 8 (5 Haiku tracers + 1 Haiku lore-analyst + 1 Sonnet wisdom-sage + 1 Sonnet coordinator)

Orchestration Protocol

Phases 0-3: Parse input (diff-spec or file list, SEC-10 flag injection + path traversal guards), resolve changed files via git diff --name-only, generate session ID + output dir (SEC-5 validated), write inscription.json, acquire workflow lock (reader), pre-create guard + TeamCreate, write state file with session isolation.

See orchestration-protocol.md for full pseudocode.

Phases 4-6: Create 8 tasks (1 Lore + 5 Impact tracers parallel → 1 Wisdom Sage sequential → 1 Coordinator sequential). Optional Phase 3.5 Codex Risk Amplification. POLL-001 compliant monitoring. Each layer degrades independently — any combination is better than a single layer.

See phase-4-5-6-tasks-monitor-degradation.md for task creation, spawn contracts, polling config, and degradation rules. See codex-risk-amplification.md for Phase 3.5.

Phase 7: Standard 5-component team cleanup. See phase7-cleanup.md.

Phase 8: Read and present GOLDMASK.md summary.

Quick Check (--quick) / Intelligence (--lore) / Output Paths

--quick: No agents — deterministic comparison of predicted vs committed MUST-CHANGE files. --lore: Lore Analyst only → risk-map.json. Output: tmp/goldmask/{session_id}/ with 10 files.

See modes-and-output.md for mode protocols and output directory structure.

Reference Files

• trace-patterns.md — Grep/Glob patterns per language per layer
• confidence-scoring.md — Noisy-OR formula + caution scoring
• intent-signals.md — Design intent classification patterns
• output-format.md — GOLDMASK.md + findings.json + risk-map.json schemas
• investigation-protocol.md — 5-step protocol for Impact tracers
• wisdom-protocol.md — 6-step protocol for Wisdom Sage
• lore-protocol.md — Risk scoring formula for Lore Analyst

Shared Cross-Skill References

These files are consumed by other skills (forge, mend, inspect, devise) via cross-skill links:

• goldmask-quick-check.md — Quick check protocol (used by forge, mend)
• lore-layer-integration.md — Lore Layer integration patterns (used by forge, inspect, devise, mend)
• risk-tier-sorting.md — Risk tier sorting utilities (used by forge, mend)