iop/configs/edge.yaml

302 lines
11 KiB
YAML

edge:
id: "edge-example"
name: "Example Edge"
server:
listen: "0.0.0.0:9090"
# advertise_host is the host other components (Node bootstrap, OpenAI client)
# use to reach this edge. Empty means resolve automatically from interfaces.
advertise_host: ""
# bootstrap is the artifact/bootstrap channel iop-edge can render in
# bootstrap commands and `env` output. Leave artifact_base_url empty until a
# field channel is provisioned.
bootstrap:
artifact_base_url: ""
tls:
enabled: false
logging:
level: "info"
pretty: true
# path is the edge log file. Empty falls back to <binary-dir>/logs/edge.log
# for the bundled dev binary; explicit values win.
path: ""
metrics:
port: 19092
# control_plane connects this Edge to a Control Plane instance.
# Set enabled: true and wire_addr to the Control Plane TCP endpoint to
# activate the outbound connector. Leave enabled: false (default) for
# standalone / local-only Edge deployments.
control_plane:
enabled: false
wire_addr: ""
reconnect_interval_sec: 5
# refresh enables the Edge-local admin API used by `iop-edge config refresh`.
# Keep it loopback-only unless an operator-specific access control layer is added.
refresh:
enabled: false
listen: "127.0.0.1:19093"
# Requests with an estimated input token count at or above this value are
# classified as long-context for admission policy.
long_context_threshold_tokens: 100000
a2a:
enabled: false
listen: "0.0.0.0:8081"
path: "/a2a"
node: ""
adapter: "cli"
session_id: "a2a"
timeout_sec: 120
bearer_token: ""
openai:
enabled: false
listen: "0.0.0.0:18081"
bearer_token: ""
# principal_tokens maps IOP bearer-token-operation identities to hashed
# OpenAI-compatible bearer tokens. Raw tokens are never stored here; only
# a SHA-256 hash (lowercase hex, 64 chars) of each issued token is kept.
# When set, callers authenticate via hash match against this list; the
# legacy single bearer_token above still works as an unmapped fallback.
# Example (values below are illustrative hashes, not real tokens):
# principal_tokens:
# - token_ref: "iop-tok-alice"
# token_hash_sha256: "<sha256 hex of the issued token>"
# principal_ref: "user:alice"
# principal_alias: "alice"
principal_tokens: []
# Chat Completions provider routes default to metadata.iop_response_mode=passthrough.
# Callers can opt into passthrough+sideband or transformed per request metadata;
# this config has no global response-mode switch.
node: ""
adapter: "ollama"
target: ""
models: []
# Legacy/compatibility model_routes (discouraged for new deploys).
# model_routes was the adapter/target route catalog that mapped external model
# ids to adapter + target. It is superseded by the provider-pool approach:
# 1. Define canonical routing keys with models[].
# 2. Declare provider candidates with nodes[].providers[].
# 3. Link them via models[].providers[provider_id] = served_model.
# For details see the Provider-pool section below.
# This block is a backward-compat fallback only. New deploys should use
# top-level models[] + nodes[].providers[].
session_id: "openai"
timeout_sec: 120
strict_output: true
strict_stream_buffer: false
# === Provider-pool (models[] / nodes[].providers[]) — recommended ===
# Top-level models[] defines canonical routing keys and their provider-pool mapping.
# Each entry id is the external model id; providers map maps provider id → served model.
#
# Example:
# models:
# - id: "qwen3.6:35b"
# display_name: "Qwen 3.6 35B"
# context_window_tokens: 262144
# providers:
# vllm-gpu: "nvidia/Qwen3.6-35B-A3B-NVFP4"
# ollama-local: "qwen3.6:35b"
#
# nodes[].providers[] defines each provider candidate with catalog and execution fields (Provider-First).
# nodes[].providers[].id is the stable provider identity referenced by models[].providers keys.
# nodes[].providers[].type — runtime type (vllm, ollama, lemonade, sglang, openai_api, cli).
# nodes[].providers[].category — "api", "cli", or "local_inference".
# nodes[].providers[].endpoint / base_url / command — type-specific execution fields.
# nodes[].providers[].models — served model names this provider can serve.
# nodes[].providers[].health — observed health state string.
# nodes[].providers[].capacity — provider-pool max concurrent execution slots; 0 is not dispatchable.
# nodes[].providers[].total_context_tokens — provider runtime total KV/context budget used for
# long-context admission. Must satisfy total_context_tokens >= context_window_tokens *
# long_context_capacity for every model group referencing this provider. Live-apply on refresh.
# nodes[].providers[].long_context_capacity — concurrent long-context slots (context_window-sized
# requests). Long requests occupy a normal slot and a long slot; 0 means no dedicated long limit.
# Live-apply on refresh.
# nodes[].providers[].max_queue — max queue depth (per-provider).
# nodes[].providers[].queue_timeout_ms — queue timeout in milliseconds.
# nodes[].providers[].lifecycle_capabilities — coarse lifecycle capabilities list.
# nodes[].providers[].enabled — on/off dispatch switch; omit or true = enabled, false = excluded from
# dispatch pool. Disabled providers appear in status with status=disabled, capacity=0. Does not
# stop adapter processes. Classified as live-apply (no restart required) on config refresh.
console:
adapter: "cli"
target: "codex"
session_id: "default"
background: false
timeout_sec: 240
# Top-level models[] defines canonical routing keys and their provider-pool mapping.
# models[].id is the external model id; providers maps provider id → served model.
models:
- id: "qwen3.6:35b"
display_name: "Qwen 3.6 35B"
context_window_tokens: 262144
default_max_tokens: 32768
min_max_tokens: 32768
default_thinking_token_budget: 1024
providers:
mac-mlx-vllm: "mlx-community/Qwen3.6-35B-A3B-4bit"
# Example: CLI providers (codex, codex-exec) do not serve models.
# They are declared in nodes[].providers[] for dispatch, not in models[].
# - id: "codex-task"
# providers:
# codex: "app-server"
# codex-exec: "exec"
nodes:
# id is the stable node identity; omitting it falls back to an auto UUID (dev only).
# agent_kind selects the registration kind; omitting it defaults to "generic-node".
# Allowed values: "generic-node" (default).
#
# Provider-First example (recommended):
- id: "node-example-01"
alias: "example-node"
token: "<node-token>"
agent_kind: "generic-node"
providers:
- id: "claude-tui"
type: "cli"
category: "cli"
command: "claude"
args:
- "--dangerously-skip-permissions"
env:
- "TERM=xterm-256color"
mode: "persistent-lazy"
capacity: 1
- id: "codex"
type: "cli"
category: "cli"
command: "codex"
args:
- "app-server"
mode: "codex-app-server"
capacity: 1
- id: "codex-exec"
type: "cli"
category: "cli"
command: "codex"
args:
- "exec"
- "--json"
resume_args:
- "exec"
- "resume"
- "--json"
output_format: "codex-json"
mode: "codex-exec"
capacity: 1
# Mac MLX vLLM provider (local GPU inference).
# Tracked secrets (API key, etc.) are not stored in this file.
- id: "mac-mlx-vllm"
type: "openai_api"
category: "local_inference"
endpoint: "http://127.0.0.1:8002/v1"
models:
- "mlx-community/Qwen3.6-35B-A3B-4bit"
health: "healthy"
capacity: 2
# Long-context admission budget: total_context_tokens >= 262144 * 1.
# mac-mlx-vllm runtime KV budget is 262144 (max_kv_size / max_request_tokens),
# so only one full 262144-window long request fits at a time. long_context_capacity
# is the long-slot count, not the normal capacity (2); do not conflate them.
total_context_tokens: 262144
long_context_capacity: 1
# Legacy adapters configuration (compat override example):
# adapters:
# cli:
# enabled: true
# profiles:
# claude-tui:
# command: "claude"
# args:
# - "--dangerously-skip-permissions"
# env:
# - "TERM=xterm-256color"
# persistent: true
# terminal: true
# response_idle_timeout_ms: 5000
# startup_idle_timeout_ms: 5000
# mode: "persistent-lazy"
# codex:
# command: "codex"
# args:
# - "app-server"
# mode: "codex-app-server"
# codex-exec:
# command: "codex"
# args:
# - "exec"
# - "--json"
# resume_args:
# - "exec"
# - "resume"
# - "--json"
# output_format: "codex-json"
# mode: "codex-exec"
runtime:
concurrency: 1
# === Provider-pool example (recommended) ===
# Full example node with provider pool (Provider-First).
# nodes[].providers[].id is the stable provider identity referenced by models[].providers keys.
# nodes:
# - id: "node-gpu-01"
# alias: "gpu-node"
# providers:
# - id: "vllm-gpu"
# type: "vllm"
# category: "api"
# endpoint: "http://127.0.0.1:8000/v1"
# models:
# - "nvidia/Qwen3.6-35B-A3B-NVFP4"
# health: "healthy"
# capacity: 4
# max_queue: 16
# queue_timeout_ms: 30000
# priority: 1
# request_timeout_ms: 120000
# lifecycle_capabilities: ["scale_up", "scale_down"]
# # enabled: false # exclude from dispatch pool (live-apply)
# - id: "ollama-local"
# type: "ollama"
# category: "local_inference"
# base_url: "http://127.0.0.1:11434"
# models:
# - "qwen3.6:35b"
# health: "healthy"
# capacity: 2
# max_queue: 8
# queue_timeout_ms: 30000
# priority: 2
#
# === Legacy adapters (compat override example — discouraged for new deploys) ===
# Only use when explicit adapter instance override or compatibility before compilation is required.
# - id: "node-vllm-01"
# alias: "vllm-gpu-node"
# token: "<node-token>"
# adapters:
# openai_compat_instances:
# - name: "vllm-gpu"
# enabled: true
# provider: "vllm"
# endpoint: "http://127.0.0.1:8000/v1"
# capacity: 4
# max_queue: 16
# queue_timeout_ms: 30000
# providers:
# - id: "vllm-gpu-legacy"
# type: "vllm"
# category: "api"
# adapter: "vllm-gpu"
# models:
# - "nvidia/Qwen3.6-35B-A3B-NVFP4"