mikeng-io/agent-skills/bridge-claude

bridge: claude
model_family: anthropic/claude
availability: conditional # Depends on executor — not always available
connection_preference:
  1: native-dispatch # Executor is Claude Code — Task tool / Agent Teams
  2: cli # Any other executor — invoke `claude -p` CLI
  3: api # Last resort — Anthropic HTTP API via ANTHROPIC_API_KEY
  4: skip # None available — return SKIPPED (non-blocking)
 
native_dispatch:
  detection: "Task tool is accessible with subagent_type parameter"
  reliability: "HIGH — actual capability check, not env var"
  subagent_types: ["explore", "librarian", "oracle", "metis", "momus"]

# Check if running INSIDE Claude Code (native dispatch)
claude_native:
  signal: "Task tool (uppercase) is accessible with subagent_type parameter"
  check: "Can invoke Task() with subagent_type='explore' or 'oracle'"
  reliability: HIGH
 
# Environment variables (backup signals, less reliable)
claude_env:
  signal: "CLAUDE_CODE_SESSION is set"
  check: "echo ${CLAUDE_CODE_SESSION:-}"
  reliability: LOW (may not be set in all contexts)

# Step 1: Check if Task tool is accessible
# If you're reading this from inside a Claude Code session, you have access to Task tool
# The key signal is: can you invoke Task with subagent_type parameter?
 
# Step 2: Determine native dispatch type
native_dispatch_type:
  agent_teams:
    condition: "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 AND (domains >= 3 OR intensity = thorough)"
    description: "Teammates share a task list and communicate directly"
 
  task_tool:
    condition: "Default — always available as fallback"
    description: "Sub-agents report results to parent only"

{
  "executor_type": "claude-code",
  "native_dispatch": {
    "available": true,
    "tool_name": "Task",
    "subagent_types": ["explore", "librarian", "oracle", "metis", "momus"],
    "session_continuity": true
  },
  "recommended_dispatch": "native"
}

which claude

echo ${ANTHROPIC_API_KEY:+found}

{
  "bridge": "claude",
  "status": "SKIPPED",
  "skip_reason": "No Anthropic access — Task tool unavailable, claude CLI not found, ANTHROPIC_API_KEY not set",
  "outputs": []
}

dispatch_mode:
  agent_teams:
    condition: "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 AND (domains >= 3 OR intensity = thorough)"
    description: "Teammates share a task list and communicate directly — best for complex multi-domain work"
    preferred: true
    caveat:
      "Env var is necessary but not sufficient — TeamCreate may still fail if this bridge
      is executing as a nested sub-agent (Task → Task). Always guard with a fallback."
 
  task_tool:
    condition: "Default — always available as fallback"
    description: "Sub-agents report results to parent — best for focused, independent tasks"
    preferred: false

0. ATTEMPT TeamCreate → "bridge-claude-{session_id}"
   IF TeamCreate fails → go to Step 3B (Task Tool fallback)
 
1. TeamCreate succeeded → continue
2. Spawn teammates:
   - One per domain from bridge_input.domains
   - Devil's Advocate (always)
   - Integration Checker (always)
3. Create tasks in shared task list — one per teammate
4. Teammates self-coordinate: domain experts complete their analysis,
   Devil's Advocate challenges findings via direct messages,
   Integration Checker surfaces cross-component issues
5. Wait for all tasks complete
6. Synthesize via TaskList
7. TeamDelete → clean up
   IF any step 2–6 fails → call TeamDelete before returning SKIPPED

You are a {expert_role}. Your task: {task_description}
Scope: {scope} | Context: {context_summary} | Intensity: {intensity}
Focus: {focus_areas from domain-registry}
Return your output as the JSON structure defined in bridge-commons Agent Prompt Template.

You are a Devil's Advocate for this analysis session.
Scope: {scope} | Context: {context_summary} | Intensity: {intensity}
 
Your obligations (read debate-protocol/experts/devils-advocate.md for full protocol):
- MUST challenge every CRITICAL and HIGH finding not originated by you, via direct teammate messages
- SHOULD challenge MEDIUM findings when you detect a pattern across multiple findings
- Cross-domain synthesis: actively look for findings whose combination implies a new, higher-severity issue not caught by any single domain expert
- Pre-mortem focus: for each component in scope, ask "what would cause this to fail in production?"
 
Challenge quality standard: a valid challenge must either (a) identify a missing assumption, (b) propose an alternative explanation that lowers severity, or (c) surface a scenario where the finding does not apply.
 
Message domain expert teammates directly to challenge their findings. Do not wait for them to send to you first.
 
Return outputs JSON with domain: "cross-domain". Include both challenge outcomes (findings you successfully challenged/withdrew) and new findings you discovered.

You are an Integration Checker for this analysis session.
Scope: {scope} | Context: {context_summary} | Intensity: {intensity}
 
Focus areas (read debate-protocol/experts/integration-checker.md for full protocol):
- Interface mismatches: where does component A assume something about component B that isn't guaranteed?
- Undocumented contracts: implicit dependencies that work by accident, not by design
- Error propagation gaps: errors that one component produces but callers don't handle
- Timing and ordering dependencies: race conditions, initialization ordering, cascading failures
- Cross-cutting assumptions: things that must be true globally but are only enforced locally
 
For each finding from domain experts: does it have cross-component implications beyond its stated scope?
If yes, surface those as integration findings even if the original finding is withdrawn.
 
Return outputs JSON with domain: "integration".

Task 1: {domain_1} expert — focus: {focus_areas}, scope: {scope}
Task 2: {domain_2} expert — focus: {focus_areas}, scope: {scope}
...
Task N:   Devil's Advocate — challenge assumptions, find failure modes (domain: "cross-domain")
Task N+1: Integration Checker — cross-component impacts, implicit contracts (domain: "integration")

CAPABILITY_FLAGS={resolved from bridge-commons capability_profile}
 
timeout {final_timeout} claude -p "{constructed_prompt}" \
  --output-format json \
  $CAPABILITY_FLAGS