config.example.yaml

# Orchestrator configuration example.
# Copy to config.yaml and adjust values.

llama_server_url: http://127.0.0.1:8080
project_dir: ./project

# Maximum retries per phase before escalation
max_retries: 3

# Total orchestrator iterations before halting
max_total_iterations: 20

# Architect self-evaluation passes
architect_max_passes: 3
architect_pass_threshold: 4

# Quiz turns (ingestion phase)
quiz_min_turns: 3
quiz_max_turns: 10

# Require human approval at gate transitions
require_human_approval: true

# Expose server to non-RFC1918 IPs (LAN-only by default)
expose_public: false

# -----------------------------------------------------------------------------
# Multi-endpoint pool (H5.2) — optional.
#
# Declare one or more OpenAI-compatible endpoints (local llama.cpp, OpenRouter,
# OpenAI, Ollama, ...) and a per-role routing table. If both blocks are
# omitted, the orchestrator falls back to the single `llama_server_url` above.
#
# Env overrides: ORCHESTRATOR_ENDPOINT_<NAME>_{BASE_URL,MODEL,API_KEY}
# (<NAME> is upper-cased, non-alphanumerics → underscore).
# -----------------------------------------------------------------------------
# endpoints:
#   - name: local-gpu
#     base_url: http://127.0.0.1:8080
#     model: qwen3-coder:30b
#   - name: openrouter
#     base_url: https://openrouter.ai/api
#     model: anthropic/claude-sonnet-4.5
#     api_key_env: OPENROUTER_API_KEY
#     headers:
#       HTTP-Referer: https://guildr.local
#       X-Title: guildr
#     extra_body: {}   # omit the Qwen thinking hint for strict providers
#
# routing:
#   architect:
#     - openrouter                   # short form: endpoint default model
#     - endpoint: local-gpu          # long form: per-role model override
#       model: qwen3-coder:30b
#   coder:
#     - local-gpu
#   reviewer:
#     - openrouter
#
# MemPalace MCP exposure for selected opencode roles. Enabled by default for
# coder/tester/reviewer/narrator. Deterministic wake-up prompt injection remains
# authoritative; MCP grants live memory lookup to tool-using roles that can
# tolerate extra context. Set `enabled: false` to remove it from generated
# opencode config.
#
# memory_mcp:
#   enabled: true
#   roles: [coder, tester, reviewer, narrator]
#   command: ["python", "-m", "mempalace.mcp_server"]
#   timeout_ms: 5000