113 lines
5.5 KiB
YAML
113 lines
5.5 KiB
YAML
server:
|
||
host: 0.0.0.0
|
||
port: 18001
|
||
workers: 1
|
||
rate_limit: 60
|
||
llm:
|
||
providers:
|
||
bailian-coding:
|
||
type: openai
|
||
base_url: https://coding.dashscope.aliyuncs.com/v1
|
||
api_key: "${DASHSCOPE_API_KEY}"
|
||
models:
|
||
qwen3.7-plus: {alias: default}
|
||
qwen3-coder-plus: {alias: coding}
|
||
qwen3-max-2026-01-23: {alias: powerful}
|
||
qwen-turbo: {alias: fast}
|
||
deepseek:
|
||
type: openai
|
||
base_url: https://api.deepseek.com/v1
|
||
api_key: "${DEEPSEEK_API_KEY}"
|
||
models:
|
||
deepseek-chat: {alias: chat}
|
||
deepseek-reasoner: {alias: reasoning}
|
||
model_aliases:
|
||
default: bailian-coding/qwen3.7-plus
|
||
fast: bailian-coding/qwen-turbo
|
||
powerful: bailian-coding/qwen3-max-2026-01-23
|
||
coding: bailian-coding/qwen3-coder-plus
|
||
chat: deepseek/deepseek-chat
|
||
reasoning: deepseek/deepseek-reasoner
|
||
# U2/R3 — 模型 fallback 链：主模型失败时按序尝试 fallback。
|
||
# 等保合规场景：fallback provider 在客户内网（DashScope/DeepSeek OpenAI-compatible API），
|
||
# 不依赖 Anthropic 境外服务。fallback providers 默认 prompt_cache_enable=false
|
||
#（非 Anthropic provider 无 cache_control 透传，走自动前缀缓存）。
|
||
fallbacks:
|
||
default: [fast] # default 失败 → fast（同 provider 内降级）
|
||
# 等保合规 fallback（注释，按需启用）：将主模型 fallback 到本地推理 provider
|
||
# default: [chat, fast] # default 失败 → deepseek-chat → qwen-turbo
|
||
# U2/R3 — PII 脱敏 hook（fail-closed）。
|
||
# 启用后，所有发送至 LLM 的 prompt 先经过 pii_filter 脱敏：
|
||
# - 正则识别手机号/邮箱/身份证/银行卡，替换为 [REDACTED_TYPE:hash8]
|
||
# - 脱敏前后记录 hash 用于审计（原文不落盘）
|
||
# - 任何异常时 fail-closed（拒绝发送至 LLM）
|
||
# - 留 ner_hook 接入客户内网 NER 模型（LAC/HanLP），未配置时用正则版
|
||
pii_filter:
|
||
enabled: false # 默认关闭；等保合规场景设 true
|
||
ner_hook: null # 可选：自定义 NER 函数路径（module:func）
|
||
fail_closed: true # 脱敏失败时拒绝发送（true）或降级发送（false）
|
||
# G4/U1: Auxiliary model for cost-sensitive tasks (summarization).
|
||
# When set, ContextCompressor tries this alias first, falling back to
|
||
# the main model on failure or empty content. Commented to preserve
|
||
# default behavior — uncomment to enable.
|
||
# auxiliary_model: fast
|
||
# G9/U4: Rollback configuration. Drives RollbackExecutor subprocess timeout
|
||
# for PlanPhase.validation_command / PlanPhase.rollback_command. Per-phase
|
||
# opt-in (KTD6) — when PlanPhase.rollback_command is unset, no rollback runs.
|
||
# Canonical rollback pattern: `git checkout <specific_files>` (file-scoped,
|
||
# not `git checkout .` which would wipe unrelated changes).
|
||
rollback:
|
||
default_timeout: 30.0
|
||
# G7/U3: Three-tier fallback chain at chat REST send_message.
|
||
# main → Recovery (ReflexionEngine retry) → Emergency (rule-based classifier).
|
||
# Wired only at chat REST path (KTD5); CLI / ReWOO / Reflexion internal
|
||
# ReAct calls bypass the chain (no recursive loop).
|
||
fallback_chain:
|
||
enabled: true
|
||
recovery:
|
||
enabled: true
|
||
max_retries: 1 # ReflexionEngine max_reflections override
|
||
emergency:
|
||
enabled: true
|
||
# G6/U2: PLAN_EXEC phase policy — SOLO four-stage state machine.
|
||
# When `enabled: true`, chat WebSocket PLAN_EXEC requests build a PhasePolicy
|
||
# (Planning → Building → Verification → Delivery) and enforce per-step tool
|
||
# whitelists (R24). Transitions are LLM-driven via AdvancePhaseTool; set
|
||
# `auto_advance_after_steps` to auto-advance as a safety net (KTD6).
|
||
# Commented to preserve default behavior — uncomment to enable.
|
||
# plan_exec:
|
||
# enabled: true
|
||
# auto_advance_after_steps: 5 # optional, default = manual (LLM calls advance_phase)
|
||
# start_phase: planning # optional, default = planning
|
||
# whitelist_override: # optional, merges with default (override wins)
|
||
# planning: [search, read_file, shell]
|
||
# building: [write_file, shell, read_file]
|
||
# G10/U5: Use Redis for bus / session / task_store when REDIS_URL is set.
|
||
# Falls back to in-memory when REDIS_URL is unset (development fallback).
|
||
# The local dev environment runs pms-redis on 127.0.0.1:6379 (see .env.dev).
|
||
# Tests run in-memory for isolation; production / staging should use Redis.
|
||
session:
|
||
backend: ${AGENTKIT_SESSION_BACKEND:-memory}
|
||
redis_url: ${REDIS_URL:-redis://127.0.0.1:6391/0}
|
||
bus:
|
||
backend: ${AGENTKIT_BUS_BACKEND:-memory}
|
||
redis_url: ${REDIS_URL:-redis://127.0.0.1:6391/0}
|
||
task_store:
|
||
backend: ${AGENTKIT_TASK_STORE_BACKEND:-memory}
|
||
redis_url: ${REDIS_URL:-redis://127.0.0.1:6391/0}
|
||
skills: {auto_discover: true, paths: ["./configs/skills"]}
|
||
experts: {paths: ["./configs/experts"]}
|
||
board: {max_rounds: 5, default_template: private_board, parallel_speech: true, history_compression_threshold: 20}
|
||
logging: {level: INFO, format: text}
|
||
router: {classifier: heuristic, auction_enabled: false}
|
||
# OTel 可观测性（U1 — 默认启用）。
|
||
# OTel 依赖已升级为 required（pyproject.toml [project] dependencies），
|
||
# ImportError 静默回退已移除 — 缺包将硬失败。设 enabled=false 可显式关闭。
|
||
# otlp_endpoint 指向 OTel collector（容器名 otel-collector / jaeger）。
|
||
telemetry:
|
||
enabled: true
|
||
otlp_endpoint: http://localhost:4317 # OTLP gRPC 端点（开发环境；生产指向集群 collector）
|
||
service_name: fischer-agentkit
|
||
export_traces: true
|
||
export_metrics: true
|