Deep Context: Artifact Classifier and Smart Router

Execute this skill to analyze conversation context, detect relevant domains, and determine the optimal review routing strategy.

Execution Instructions

When invoked, analyze the current conversation and produce a context report.

Dependency Check

Before executing any step, verify the domain-registry is present:

[skills-root]/domain-registry/README.md

Where [skills-root] is the parent of this skill's directory. Resolve with ls ../ from this skill's location.

If missing → stop immediately and output:

⚠ Missing required skill for context:

  domain-registry
    Expected: {skills-root}/domain-registry/README.md

Install:
  git clone https://github.com/mikeng-io/agent-skills /tmp/agent-skills
  cp -r /tmp/agent-skills/skills/domain-registry {skills-root}/

Dependency present → proceed to Step 1.

Step 1: Analyze Conversation

Extract from the conversation:

conversation_signals:
  files_mentioned: []        # File paths referenced (e.g., "src/auth.go")
  artifacts_mentioned: []    # Other artifacts (Figma files, spreadsheets, docs)
  topics: []                 # Key topics discussed
  concerns: []               # What the user is worried about
  explicit_domains: []       # Domains explicitly named by user
  explicit_routing: ""       # If user said "multi-model", "debate", "thorough", etc.
  intent: ""                 # What the user wants to accomplish

Step 2: Classify Artifact Type

Determine the primary artifact type from signals:

artifact_type_rules:
  code:
    signals:
      - File extensions: .go, .py, .ts, .js, .rs, .java, .rb, .kt, .swift, .c, .cpp
      - Mentions of: function, class, module, API, endpoint, service, handler
      - Git references or diff mentions

  financial:
    signals:
      - Terms: P&L, revenue, profit, loss, budget, forecast, balance sheet, ROI, EBITDA
      - Spreadsheet files: .xlsx, .csv with financial column names
      - GAAP, IFRS, accounting mentions

  marketing:
    signals:
      - Terms: campaign, audience, conversion, CTR, CPC, funnel, brand, messaging, copywriting
      - Marketing channel names: email, social, PPC, SEO, content marketing

  research:
    signals:
      - Terms: literature, sources, evidence, citations, study, paper, methodology
      - Academic language patterns
      - Bibliography or reference mentions

  creative:
    signals:
      - Design files: .fig, .sketch, .psd, .ai, .xd
      - Terms: design, visual, layout, color, typography, UX, wireframe, mockup, copy

  mixed:
    signals:
      - Signals from 2+ categories above present simultaneously

Artifact type selection:

If signals from 2+ types → mixed
If dominant signals clearly point to one type → that type
Default → code (most common)

Step 2.5: Enrich Context with DeepWiki (Optional)

If the artifact type is code or mixed and a GitHub repository can be identified, optionally invoke the deepwiki skill to get architectural context before domain selection. This enriches domain detection beyond what file names and conversation signals alone can provide.

When to invoke:

Artifact type is code or mixed
Minimal conversation context (low-confidence signals from Step 1)
User asks about architectural intent, not just file contents
Repository is large or unfamiliar — conversation signals may not surface all relevant domains

Invocation:

Skill("deepwiki")

Pass the question: "What are the primary technical domains and architectural concerns of this codebase? What systems, components, or subsystems are involved?"

Use the returned answer to:

Supplement explicit_domains before running domain-registry matching
Confirm or expand artifact_type if the codebase is more complex than conversation signals suggested
Add technical terms from the answer to the conversation signals for better domain trigger matching

If deepwiki returns availability: "unavailable" → skip and proceed to Step 3 with signals from Step 1 only. Non-blocking.

Step 3: Select Domains from Domain-Registry

Read the domain-registry files to match detected signals:

# Read all three domain category files
Read: [skills-root]/domain-registry/domains/technical.md
Read: [skills-root]/domain-registry/domains/business.md
Read: [skills-root]/domain-registry/domains/creative.md

Where [skills-root] is the parent directory of the skill invoking this step. Resolve by navigating up from the current skill's location using ls or Glob if needed.

For each domain in the registry files:

Check if any of the domain's trigger_signals appear in the conversation
Match against: file names, topics, concerns, explicit mentions, technical terms
Select ALL matching domains (minimum 1, no maximum)

If no domains match clearly:

For code artifact type → default to api + testing
For financial → default to finance
For marketing → default to marketing
For research → default to content
For creative → default to design + ux

Step 4: Determine Routing

Apply routing decision rules:

routing_rules:
  parallel-workflow:
    conditions:
      - domains_count < 3
      - no high-stakes signals (production incident, financial risk, security breach)
      - no explicit "thorough" or "critical" in user request
    description: "Parallel independent sub-agents, no debate needed"

  debate-protocol:
    conditions_any:
      - domains_count >= 3
      - high_stakes_signals: [production, incident, breach, compliance, audit, lawsuit, financial risk]
      - explicit_signals: ["thorough", "critical", "high-stakes", "careful", "deep"]
    intensity_selection:
      thorough: explicit "thorough" or "critical" in request
      standard: default for debate-protocol routing
      quick: explicit "quick" or "fast" in request
    description: "Structured 5-phase adversarial debate across domains"

  deep-council:
    conditions_any:
      - explicit_signals: ["multi-model", "multiple models", "council", "cross-model"]
      - user explicitly requests multiple AI perspectives
      - intensity = thorough → deep-council (recommended — thorough analysis benefits from multi-model perspectives)
    description: "Multi-model review across Claude, Gemini, Codex, OpenCode bridges"

Routing priority: deep-council > debate-protocol > parallel-workflow

Step 5: Determine Confidence and Surface Missing Signals

Check each signal explicitly and record what could not be resolved:

signal_resolution:
  artifact_identified: true | false   # specific files/paths/topics clearly present?
  intent_clear: true | false          # review? audit? verify? research? explore?
  domains_detectable: true | false    # can domains be matched from domain-registry?
  scope_bounded: true | false         # is scope narrow enough to proceed without clarification?

Build the missing_signals list from any signal that resolved to false:

missing_signal_map:
  artifact_identified: false  →  "artifact"
  intent_clear: false         →  "intent"
  domains_detectable: false   →  "domains"
  scope_bounded: false        →  "scope"

Derive overall confidence from the resolved signals:

confidence_levels:
  high:    all four signals true  →  preflight_needed: false
  medium:  1 signal false         →  preflight_needed: true  (scope or domains)
  low:     2+ signals false       →  preflight_needed: true  (artifact or intent unresolved)

Step 6: Produce Context Report

Output the context report:

context_report:
  artifact_type: code | financial | marketing | creative | research | mixed
  topics: []                    # Key topics identified
  concerns: []                  # User concerns identified
  domains: []                   # Selected domain names from domain-registry
  domain_experts: []            # Corresponding expert_role values
  routing: parallel-workflow | debate-protocol | deep-council
  debate_intensity: quick | standard | thorough   # Only if routing = debate-protocol
  confidence: high | medium | low
  preflight_needed: true | false   # true if any signal resolved to false in Step 5
  missing_signals: []              # subset of ["artifact", "intent", "domains", "scope"]
                                   # empty when preflight_needed: false
  rationale: ""                    # Brief explanation of routing decision
  signals_detected:
    files: []
    topics: []
    explicit_mentions: []

missing_signals is the direct handoff to preflight — it tells preflight exactly which questions to ask, so preflight skips questions for signals that context already resolved.

Display this report to the user and/or return it to the calling skill.

Step 7: Save Artifact

Save context report to .outputs/context/{YYYYMMDD-HHMMSS}-context.md with YAML frontmatter:

---
skill: context
timestamp: {ISO-8601}
artifact_type: {artifact_type}
domains: [{domain1}, {domain2}]
routing: {routing}
confidence: {confidence}
context_summary: "{brief description of what was analyzed}"
session_id: "{unique id}"
---

Also save JSON companion: {YYYYMMDD-HHMMSS}-context.json

No symlinks. To find the latest artifact:

ls -t .outputs/context/ | head -1

QMD Integration (optional):

qmd collection add .outputs/context/ --name "context-reports" --mask "**/*.md" 2>/dev/null || true
qmd update 2>/dev/null || true

Notes

Pre-flight pattern: Other skills can call this at the start to determine routing before spawning agents
Progressive enhancement: If domain-registry is not found, fall back to basic artifact type detection
Non-blocking: Always produces a result, even with low confidence
Model-agnostic: Works in any Claude, Gemini, Codex, or OpenCode context