agents.yaml

proxy:
  url: "http://localhost:3457"
  manager_url: "http://localhost:3400"
  timeout_ms: 900000
  ssh_key: "~/.ssh/claude-proxy-agents"
  # gh_token loaded from GH_TOKEN env var at runtime

base_dir: "/Users/paultarr/Documents/Git"
orchestrator_dir: "/Users/paultarr/Documents/Git/claude-agent-orchestrator"

providers:
  global:
    # Fleet treasury — canonical receiving wallet for all revenue paths.
    # Set this env var in the host environment to override the baked-in default.
    # Baked-in default: Base (L2) EVM wallet owned by the fleet.
    FLEET_WALLET_ADDRESS: "0x468EC325f3797F5968dEcC757FA0B960Bd0f78Ef"
    FLEET_WALLET_NETWORK: "Base"
    # Alias used by agent-reviewer survival-plan.ts (and any agent that reads
    # TREASURY_WALLET_ADDRESS rather than FLEET_WALLET_ADDRESS).
    TREASURY_WALLET_ADDRESS: "0x468EC325f3797F5968dEcC757FA0B960Bd0f78Ef"

  claude:
    model: "claude-opus-4-6"
    limits:
      hourly: 1000000       # ~1M tokens/hour
      daily: 5000000        # ~5M tokens/day
      weekly: 25000000      # ~25M tokens/week
  openai:
    model: "gpt-5.4-mini"
    limits:
      hourly: 2000000
      daily: 10000000
      weekly: 50000000
  gemini:
    model: "gemini-2.5-pro"
    api_key_env: "GEMINI_API_KEY"
  grok:
    model: "grok-4"
    api_key_env: "XAI_API_KEY"
    base_url: "https://api.x.ai/v1"
    limits:
      hourly: 500000
      daily: 2000000
      weekly: 10000000
  deepseek:
    model: "deepseek-chat"
    api_key_env: "DEEPSEEK_API_KEY"
    base_url: "https://api.deepseek.com/v1"
    limits:
      hourly: 2000000
      daily: 10000000
      weekly: 50000000

verification:
  enabled: true
  sources: ["github", "linear"]
  min_score: 0.7
  max_revisions: 1

# ── Escalation config ──────────────────────────────────────────────────────────
# Controls automatic escalation thresholds.
# max_followup_depth: maximum length of a follow-up task chain before the next
#   task is routed to the escalation queue instead of auto-dispatched.
#   Depth 0 = root task, 1 = first follow-up, 2 = second follow-up, etc.
#   Default is 3.  Set to 0 to disable the depth cap entirely.
escalation:
  retry_limit: 3
  max_followup_depth: 3

# ── Task type definitions ──────────────────────────────────────────────────────
# Built-in types (implementation, research, facilitation) have hardcoded
# defaults in the verifier and do not need entries here — but entries here
# take precedence over the built-ins if you need to customise them.
#
# Add new non-coding agent task types here without touching reviewer-client.ts:
#
#   task_types:
#     my-type:
#       verification_prompt: |
#         You are verifying ... Respond with ONLY a JSON object ...
#       prompt_header: My Task Header    # default: "Task"
#       result_header: Agent Output      # default: "Agent Response"
#       dimensions: [dim1, dim2]         # documentation only
#
task_types:
  planning:
    verification_prompt: |
      You are a quality reviewer for plans and roadmaps produced by an AI agent. Given a planning request and the agent's plan, assess the quality of the plan.

      Respond with ONLY a JSON object (no markdown, no code fences):
      {
        "approved": true/false,
        "score": 0.0-1.0,
        "notes": "Brief assessment of plan quality",
        "revision": "If not approved, specific guidance for improvement (omit if approved)",
        "dimensions": {
          "feasibility": 0.0-1.0 (can this plan realistically be executed given constraints?),
          "completeness": 0.0-1.0 (does the plan cover all required steps and dependencies?),
          "risk_assessment": 0.0-1.0 (are risks identified with mitigations?),
          "clarity": 0.0-1.0 (are steps concrete and unambiguous enough to act on?)
        }
      }

      Evaluate plan quality on:
      - **Feasibility**: Is the plan realistic given time, resources, and technical constraints?
      - **Completeness**: Are all required steps, dependencies, and success criteria included?
      - **Risk assessment**: Are blockers and risks identified? Are mitigations proposed?
      - **Clarity**: Are steps specific enough that an agent can act on them without further clarification?
      - **Sequencing**: Are steps in a logical order with dependencies respected?

      Scoring guide:
      - 0.9-1.0: Excellent — comprehensive, feasible plan with clear steps, dependencies, and risk mitigations
      - 0.7-0.89: Good — solid plan, minor gaps in risk coverage or step specificity
      - 0.5-0.69: Acceptable — plan is understandable but lacks depth or has unclear steps
      - Below 0.5: Needs revision — incomplete, infeasible, or too vague to execute
    prompt_header: Planning Request
    result_header: Agent Plan
    dimensions: [feasibility, completeness, risk_assessment, clarity]

  coordination:
    verification_prompt: |
      You are a quality reviewer for coordination and orchestration decisions produced by an AI agent. The agent evaluates cross-team dependencies, assigns work, and tracks progress across multiple parties.

      Respond with ONLY a JSON object (no markdown, no code fences):
      {
        "approved": true/false,
        "score": 0.0-1.0,
        "notes": "Brief assessment of coordination quality",
        "revision": "If not approved, specific guidance for improvement (omit if approved)",
        "dimensions": {
          "assignment_quality": 0.0-1.0 (were tasks assigned to the right agents/teams?),
          "dependency_mapping": 0.0-1.0 (are cross-party dependencies identified and sequenced?),
          "communication": 0.0-1.0 (are handoffs and status updates clearly defined?),
          "completeness": 0.0-1.0 (does the coordination plan cover all parties and tasks?)
        }
      }

      Evaluate coordination quality on:
      - **Assignment quality**: Were tasks given to the right agents or teams based on capability?
      - **Dependency mapping**: Are cross-party dependencies explicitly identified and ordered?
      - **Communication**: Are handoff points, status updates, and escalation paths defined?
      - **Completeness**: Does the plan account for all parties involved in the work?
      - **Conflict avoidance**: Are parallel workstreams structured to avoid conflicts?

      Scoring guide:
      - 0.9-1.0: Excellent — clear assignments, full dependency graph, defined handoffs
      - 0.7-0.89: Good — solid coordination with minor gaps in dependency or communication coverage
      - 0.5-0.69: Acceptable — workable plan but some assignments are vague or dependencies implicit
      - Below 0.5: Needs revision — key parties missing, dependencies unresolved, or assignments unclear
    prompt_header: Coordination Request
    result_header: Coordination Plan
    dimensions: [assignment_quality, dependency_mapping, communication, completeness]

agents:
  # ── Orchestrator ────────────────────────────────────────────────
  claude-agent-orchestrator:
    dir: "claude-agent-orchestrator"
    repo: "git@github.com:rapartlu/agent-orchestrator.git"
    provider: "claude"
    model: "claude-opus-4-6"
    pool: "orchestrator"
    description: "Core orchestration infrastructure (Claude). Daemon, state store, dispatching, triggers."
    capabilities: ["typescript", "orchestration", "sqlite"]
    github: "rapartlu/agent-orchestrator"
    owns_topics: ["orchestrator", "daemon", "routing", "dispatcher", "planner", "deployer", "state", "triggers", "sync", "waste", "guard", "metric"]
    director: true                 # Fleet Director — Q3 2026 (decided at kickoff #1209)
    director_duties: ["weekly-retro", "goals-yaml", "quarterly-retro-issue"]
    auto_reroute_rejection_threshold: 4
    # Housekeeping fires at cycle modulo 0 → first slot in each 5h window (~:00)
    housekeeping_offset_cycles: 0
    docker:
      port: 3472
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  codex-agent-orchestrator:
    dir: "claude-agent-orchestrator"
    repo: "git@github.com:rapartlu/agent-orchestrator.git"
    provider: "openai"
    model: "gpt-5.5"
    pool: "orchestrator"
    description: "Core orchestration infrastructure (Codex). Parallel throughput for orchestrator pool."
    capabilities: ["typescript", "orchestration", "sqlite"]
    github: "rapartlu/agent-orchestrator"
    owns_topics: ["orchestrator", "daemon", "routing", "dispatcher", "planner", "deployer", "state", "triggers", "sync"]
    docker:
      port: 3480
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  # ── Reviewer ────────────────────────────────────────────────────
  claude-orchestrator-reviewer:
    dir: "claude-orchestrator-reviewer"
    repo: "git@github.com:rapartlu/agent-reviewer.git"
    provider: "claude"
    model: "claude-sonnet-4-6"
    pool: "reviewer"
    description: "PR reviewer, verifier, supervisor (Claude). Also serves as LLM backend."
    capabilities: ["typescript", "llm", "review", "verification"]
    capability_tags: ["review-only"]
    github: "rapartlu/agent-reviewer"
    owns_topics: ["review", "verification", "supervisor", "improvement", "escalation", "telegram", "quality"]
    # Housekeeping fires at cycle modulo 10 → second slot in each 5h window (~:50min offset)
    housekeeping_offset_cycles: 10
    docker:
      port: 3474
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  codex-orchestrator-reviewer:
    dir: "claude-orchestrator-reviewer"
    repo: "git@github.com:rapartlu/agent-reviewer.git"
    provider: "openai"
    model: "gpt-5.4-mini"
    pool: "reviewer"
    description: "PR reviewer, verifier, supervisor (Codex). Parallel throughput for reviewer pool."
    capabilities: ["typescript", "llm", "review", "verification"]
    github: "rapartlu/agent-reviewer"
    owns_topics: ["review", "verification", "supervisor", "improvement", "escalation", "telegram", "quality"]
    docker:
      port: 3481
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  # ── Dashboard ───────────────────────────────────────────────────
  claude-orchestrator-dashboard:
    dir: "claude-orchestrator-dashboard"
    repo: "git@github.com:rapartlu/agent-dashboard.git"
    provider: "claude"
    model: "claude-sonnet-4-6"
    pool: "dashboard"
    description: "Dashboard, CLI commands, observability UI (Claude)."
    capabilities: ["typescript", "html", "css", "express", "dashboard"]
    github: "rapartlu/agent-dashboard"
    owns_topics: ["dashboard", "cli", "status", "health", "metrics", "activity", "ui", "escalation-ui"]
    # Housekeeping fires at cycle modulo 20 → third slot in each 5h window (~:100min offset)
    housekeeping_offset_cycles: 20
    docker:
      port: 3473
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  codex-orchestrator-dashboard:
    dir: "claude-orchestrator-dashboard"
    repo: "git@github.com:rapartlu/agent-dashboard.git"
    provider: "openai"
    model: "gpt-5.4-mini"
    pool: "dashboard"
    description: "Dashboard, CLI commands, observability UI (Codex). Parallel throughput for dashboard pool."
    capabilities: ["typescript", "html", "css", "express", "dashboard"]
    github: "rapartlu/agent-dashboard"
    owns_topics: ["dashboard", "cli", "status", "health", "metrics", "activity", "ui", "escalation-ui"]
    docker:
      port: 3482
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  # ── Telegram ────────────────────────────────────────────────────
  claude-orchestrator-telegram:
    dir: "claude-agent-orchestrator"
    provider: "claude"
    model: "claude-haiku-4-5"
    description: "Dedicated Telegram command handler. Quick queries only."
    capabilities: ["telegram", "status"]
    owns_topics: []
    docker:
      port: 3477
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  # ── Research pool (Claude + Codex) ──────────────────────────────
  claude-research-agent:
    dir: "claude-research-agent"
    repo: "git@github.com:rapartlu/research-agent.git"
    provider: "claude"
    model: "claude-opus-4-6"
    pool: "research"
    description: "Research agent: web search, doc analysis, technology evaluation (Claude)."
    capabilities: ["research", "web-search", "documentation", "analysis"]
    capability_tags: ["research-only"]
    github: "rapartlu/research-agent"
    owns_topics: ["research", "investigation", "comparison", "evaluation", "exploration", "temporal", "architecture"]
    # Housekeeping fires at cycle modulo 30 → fourth slot in each 5h window (~:150min offset)
    housekeeping_offset_cycles: 30
    docker:
      port: 3478
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  codex-research-agent:
    dir: "claude-research-agent"
    repo: "git@github.com:rapartlu/research-agent.git"
    provider: "openai"
    model: "gpt-5.5"
    pool: "research"
    description: "Research agent (Codex). Parallel throughput for research pool."
    capabilities: ["research", "web-search", "documentation", "analysis"]
    capability_tags: ["research-only"]
    github: "rapartlu/research-agent"
    owns_topics: ["research", "investigation", "comparison", "evaluation", "exploration", "temporal", "architecture"]
    docker:
      port: 3483
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  # ── Proxy ───────────────────────────────────────────────────────
  claude-proxy:
    dir: "claude-proxy"
    repo: "git@github.com:rapartlu/agent-proxy.git"
    provider: "claude"
    model: "claude-opus-4-6"
    pool: "proxy"
    description: "Proxy server: CLI sessions, container management (Claude)."
    capabilities: ["typescript", "docker", "proxy", "api", "infrastructure"]
    github: "rapartlu/agent-proxy"
    owns_topics: ["proxy", "docker", "container", "infrastructure", "auth", "ssh"]
    # Housekeeping fires at cycle modulo 40 → fifth slot in each 5h window (~:200min offset)
    housekeeping_offset_cycles: 40
    docker:
      port: 3471
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  codex-proxy:
    dir: "claude-proxy"
    repo: "git@github.com:rapartlu/agent-proxy.git"
    provider: "openai"
    model: "gpt-5.4-mini"
    pool: "proxy"
    description: "Proxy server (Codex). Parallel throughput for proxy pool."
    capabilities: ["typescript", "docker", "proxy", "api", "infrastructure"]
    github: "rapartlu/agent-proxy"
    owns_topics: ["proxy", "docker", "container", "infrastructure", "auth", "ssh"]
    docker:
      port: 3484
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  # ── Meeting Facilitator ────────────────────────────────────────
  meeting-facilitator-agent:
    dir: "meeting-facilitator-agent"
    repo: "git@github.com:rapartlu/meeting-facilitator-agent.git"
    provider: "claude"
    model: "claude-sonnet-4-6"
    description: "Meeting facilitator. Evaluates requests, selects format and participants, runs structured discussions, synthesises outcomes."
    capabilities: ["facilitation", "synthesis", "coordination"]
    github: "rapartlu/meeting-facilitator-agent"
    owns_topics: ["facilitation", "meetings", "meeting", "standup", "coordination", "suppression", "surge"]
    # Housekeeping fires at cycle modulo 50 → sixth slot in each 5h window (~:250min offset)
    housekeeping_offset_cycles: 50
    docker:
      port: 3485
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  # ── Grok meeting voice ─────────────────────────────────────────
  grok-meeting-voice:
    dir: "grok-meeting-voice"
    provider: "grok"
    model: "grok-4"
    description: "Non-voting standup participant (Grok). Provides viewpoints from a materially different model lineage. Article VI cognitive diversity."
    capabilities: ["facilitation", "standup", "synthesis"]
    owns_topics: ["standup", "meeting-voice"]
    allowed_types: ["facilitation"]
    docker:
      port: 3486
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  # ── Deepseek background ────────────────────────────────────────
  deepseek-background:
    dir: "deepseek-background"
    provider: "deepseek"
    model: "deepseek-chat"
    description: "Cheap background workloads (Deepseek V3): embeddings, semantic memory reindex, fuzzing. Volume tier."
    capabilities: ["embeddings", "semantic-memory", "background", "fuzzing", "analysis"]
    owns_topics: ["embeddings", "semantic-memory", "reindex", "fuzzing", "background"]
    docker:
      port: 3487
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  # ── Deepseek reasoning reviewer ────────────────────────────────
  deepseek-reasoning:
    dir: "claude-orchestrator-reviewer"
    repo: "git@github.com:rapartlu/agent-reviewer.git"
    provider: "deepseek"
    model: "deepseek-reasoner"
    pool: "reviewer"
    description: "Third voice in the reviewer pool (Deepseek R1). Independent reasoning for PR review and verification."
    capabilities: ["typescript", "llm", "review", "verification", "reasoning"]
    github: "rapartlu/agent-reviewer"
    owns_topics: ["review", "verification", "reasoning"]
    docker:
      port: 3488
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"

  # ── Gemini synthesis ───────────────────────────────────────────
  gemini-synth:
    dir: "gemini-synth"
    provider: "gemini"
    model: "gemini-2.5-pro"
    description: "Long-context whole-repo review and standup synthesis (Gemini 2.5 Pro). Article VI cognitive diversity."
    capabilities: ["synthesis", "long-context", "review", "standup", "analysis"]
    owns_topics: ["synthesis", "standup-synthesis", "long-context-review", "whole-repo"]
    docker:
      port: 3489
      api_key: "cheese"
      permissions: "bypassPermissions"
      session: "fresh"