mirror of
https://gitee.com/mateos/mateclaw.git
synced 2026-09-14 03:33:43 +08:00
0476447ab6
A bundle of stability fixes that all surfaced together while running the same long-form generation task across multiple turns. Each one addresses a distinct way the previous behavior silently dropped content the user had already seen on screen. 1. Mid-turn narrative persistence (StateGraphReActAgent + SummarizingNode). Intermediate ReasoningNode rounds and SummarizingNode broadcast their content_delta directly to the SSE channel for live display, but the StreamAccumulator only received the final answer. After refresh the assistant message showed only tool_call cards with no body text. StateGraphReActAgent now also forwards STREAMED_CONTENT (already set per round) as a persistOnly StreamDelta whenever it changes, so every narrative chunk lands in the accumulator's content buffer and gets written to mate_message. SummarizingNode now writes its summary into the same key so summarize narratives persist too. 2. Follow-up message queue, not dispose (ChatController#interruptStream). Sending a new message while a turn was running called requestInterrupt, which dispose()d the active Reactor chain mid LLM call. That cancelled the in-flight generation, lost partial tokens, and left the user staring at a half-finished bubble. The endpoint now uses enqueueMessage in all paths, matching the "wait for current turn, then run" behavior. The old requestInterrupt API is kept for any future force-replace UI but no caller routes to it. 3. Queued user message ordering (ChatStreamTracker.QueuedInput + ChatController.startQueuedMessage). interruptStream used to save the queued user message immediately, before the in-flight assistant message finalized in doOnError. listMessages orders by create_time ASC, so the queued user message ended up above the assistant reply it was supposed to follow. QueuedInput now carries contentParts; persistence is delayed to startQueuedMessage, which runs only after Asst-N is on disk. 4. JVM shutdown flush (ChatStreamTracker @PreDestroy + emergencySaveAccumulator). A mvn spring-boot:run restart used to wipe in-flight turns: SSE emitter timed out, ShutdownHook fired, HikariPool closed before doOnError could save. ChatStreamTracker now exposes an emergency-save callback per RunState; ChatController registers one per stream that snapshots the accumulator and writes status="interrupted_shutdown". @PreDestroy walks active runs, invokes the callback, then disposes. Spring's reverse-order bean teardown keeps ConversationService and Hikari alive long enough for the save to complete. 5. Observation thresholds for summarize (GraphObservationProperties + application.yml). The previous total-chars threshold of 12 KB triggered summarize after one or two RFC reads, costing a 40 to 80 second compaction LLM call per loop. Tuned to: total 200 KB, single 16 KB, large-result 32 KB, rounds safety net 25. Java field defaults reverted to the conservative original values so application.yml stays the source of truth. 6. Frontend thinking segmentation (useChat.ts thinking_delta + phase). Multi-round ReAct turns merged every reasoning + summarize round's thinking into one segment, accumulating to 9 KB+ in a single bubble. thinking_delta now uses findLast(running) so a tool_call_started or phase transition closes the previous segment and the next delta opens a fresh one. phase event also closes running thinking/content segments. 7. Other small things bundled: removed a debug metadata-keys log that flooded the log file with one line per stream chunk; fixed three stale tests that didn't compile after earlier constructor changes (WikiLogServiceTest, WikiOverviewSpliceTest, WikiProcessingServiceLazyTest); added rfc-066 documenting the unified message queue + priority refactor as the next logical step on top of these stabilizations. Verified end-to-end with multiple full sessions: a four-minute generation that produced the expected docx and a follow-up enqueue that ran cleanly after the previous turn naturally completed, without the old "Disposable unavailable" interrupt path.
229 lines
8.5 KiB
YAML
229 lines
8.5 KiB
YAML
server:
|
||
port: 18088
|
||
servlet:
|
||
context-path: /
|
||
|
||
spring:
|
||
application:
|
||
name: mateclaw-server
|
||
messages:
|
||
basename: messages
|
||
encoding: UTF-8
|
||
servlet:
|
||
multipart:
|
||
max-file-size: 100MB
|
||
max-request-size: 200MB
|
||
threads:
|
||
virtual:
|
||
enabled: true
|
||
profiles:
|
||
active: dev
|
||
|
||
# 数据源（默认 H2，生产切换为 mysql profile）
|
||
datasource:
|
||
url: jdbc:h2:file:./data/mateclaw;MODE=MySQL;DATABASE_TO_LOWER=TRUE;CASE_INSENSITIVE_IDENTIFIERS=TRUE
|
||
driver-class-name: org.h2.Driver
|
||
username: sa
|
||
password:
|
||
|
||
# SQL 初始化由 DatabaseBootstrapRunner 接管，关闭 Spring 自动执行
|
||
sql:
|
||
init:
|
||
mode: never
|
||
|
||
# Flyway 数据库版本迁移
|
||
flyway:
|
||
enabled: true
|
||
baseline-on-migrate: true
|
||
baseline-version: "1"
|
||
locations:
|
||
- classpath:db/migration/h2
|
||
validate-on-migrate: true
|
||
clean-disabled: true
|
||
|
||
h2:
|
||
console:
|
||
enabled: ${H2_CONSOLE_ENABLED:false}
|
||
path: /h2-console
|
||
|
||
# Spring AI Alibaba (DashScope) - Spring AI Alibaba 1.1.x 配置路径
|
||
ai:
|
||
dashscope:
|
||
api-key: ${DASHSCOPE_API_KEY:your-dashscope-api-key-here}
|
||
chat:
|
||
options:
|
||
model: qwen-max
|
||
temperature: 0.7
|
||
max-tokens: 4096
|
||
# Spring AI 1.1.x 会话记忆配置（使用内嵌 H2 时无需额外配置）
|
||
chat:
|
||
observations:
|
||
log-prompt: false
|
||
log-completion: false
|
||
memory:
|
||
repository:
|
||
jdbc:
|
||
initialize-schema: embedded
|
||
# Spring AI Retry 配置：429/503/529 归为可重试（TransientAiException），启用指数退避
|
||
# RFC-012 M1：max-attempts 从 5 调到 2，避免与 wiki 层 callLlmWithResilientRetry 的 5×N 嵌套放大；
|
||
# 真正的重试主控权交给业务层（wiki / agent），Spring AI 仅负责"一次性瞬时抖动"的快速重试
|
||
retry:
|
||
max-attempts: 2
|
||
on-http-codes: 429, 503, 529
|
||
backoff:
|
||
initial-interval: 3000
|
||
multiplier: 3
|
||
max-interval: 60000
|
||
# 禁用 Spring AI MCP Client 自动配置（由 McpClientManager 自行管理生命周期）
|
||
mcp:
|
||
client:
|
||
enabled: false
|
||
|
||
# MyBatis Plus
|
||
mybatis-plus:
|
||
configuration:
|
||
map-underscore-to-camel-case: true
|
||
log-impl: org.apache.ibatis.logging.slf4j.Slf4jImpl
|
||
|
||
# SpringDoc OpenAPI
|
||
springdoc:
|
||
api-docs:
|
||
path: /v3/api-docs
|
||
swagger-ui:
|
||
path: /swagger-ui.html
|
||
|
||
# MateClaw 自定义配置
|
||
mateclaw:
|
||
jwt:
|
||
secret: ${JWT_SECRET:MateClaw-JWT-Secret-Key-2024-Please-Change-In-Production}
|
||
expiration: 86400000
|
||
# 搜索配置已迁移至数据库（mate_system_setting 表），通过 UI 系统设置管理
|
||
# MCP server 配置已迁移至数据库（mate_mcp_server 表），通过 UI 管理
|
||
mcp:
|
||
enabled: true
|
||
skill:
|
||
workspace:
|
||
root: ${user.home}/.mateclaw/skills
|
||
auto-init: true
|
||
delete-policy: archive
|
||
hub:
|
||
base-url: https://clawhub.ai
|
||
search-path: /api/v1/search
|
||
http-timeout: 15
|
||
http-retries: 3
|
||
plugin:
|
||
enabled: true
|
||
user-dir: ${user.home}/.mateclaw/plugins
|
||
# RFC-017: 声明式 Hook 系统
|
||
hooks:
|
||
enabled: true
|
||
global-rate-limit: 200 # 全局每秒最大派发数
|
||
global-concurrency: 32 # 同时活跃的派发任务上限
|
||
dispatch-deadline: 5s # 单事件派发硬预算
|
||
trusted-domains: [] # HttpAction 允许调用的域名（精确或后缀匹配），留空禁用 HTTP
|
||
http:
|
||
connect-timeout: 2s
|
||
read-timeout: 3s
|
||
audit:
|
||
enabled: true # 每次派发写 mate_hook_run
|
||
retain-days: 7
|
||
# RFC-014: Anthropic prompt cache 标记
|
||
llm:
|
||
cache:
|
||
enabled: true # 总开关；false 时全部走 NoOp
|
||
min-prompt-tokens: 1024 # 累计 prompt token 低于此值则跳过缓存（避免 cache write 倒亏）
|
||
max-breakpoints: 4 # 单请求最多打几个 cache_control 断点（Anthropic 上限 4）
|
||
ttl: DEFAULT # DEFAULT(5min) | EXTENDED_1H（需要 anthropic-beta header）
|
||
include-tools-block: true # 工具 schema 段是否独立打断点（CONVERSATION_HISTORY 策略）
|
||
adaptive:
|
||
enabled: true # 自适应降级（连续 miss 后短路到 NoOp，冷却后恢复）
|
||
miss-threshold: 5
|
||
cool-down-ms: 60000
|
||
# RFC-009 P3.3: per-provider health tracker for the multi-model failover chain.
|
||
# When a provider hits failure-threshold consecutive failures it enters a cooldown
|
||
# window during which the chain walker skips it, avoiding repeated 5-retry stalls
|
||
# against a known-broken provider on every conversation turn.
|
||
failover:
|
||
health:
|
||
enabled: true
|
||
failure-threshold: 3
|
||
cooldown-ms: 300000
|
||
|
||
# MateClaw Agent 配置
|
||
mate:
|
||
agent:
|
||
graph:
|
||
observation:
|
||
# 与 GraphObservationProperties.java 默认值对齐，参考 openclaw token-budget 设计
|
||
max-single-observation-chars: 16000
|
||
max-total-observation-chars: 200000
|
||
large-result-threshold: 32000
|
||
min-rounds-for-summarize: 25
|
||
head-ratio: 0.4
|
||
truncation-marker: "\n\n... [内容已截断，共 %d 字符，保留前后关键片段] ...\n\n"
|
||
tool:
|
||
timeout:
|
||
default-timeout-seconds: 300
|
||
per-category:
|
||
shell: 120
|
||
web: 30
|
||
# RFC-008 Phase 3: tool-result three-layer budget (per-result spill + per-turn aggregate budget).
|
||
# Layer 1 (per-tool cap) lives inside individual tools; Layer 2 spills oversized
|
||
# single results to disk; Layer 3 enforces an aggregate cap on the combined
|
||
# response size of one tool turn. The full output is preserved on disk and
|
||
# the in-context preview points the agent at the spill file (read_file tool).
|
||
tool-result:
|
||
enabled: true
|
||
per-result-threshold-chars: 16000 # was 4000 — prevents WebSearch spill-to-disk
|
||
per-turn-budget-chars: 32000 # was 16000 — headroom for multi-tool turns
|
||
preview-head-chars: 800
|
||
storage-base-dir: ""
|
||
# Retrieval-style tools that must NEVER be spilled. Spilling read_file's
|
||
# output causes a recursion: the agent reads the spill path, that read also
|
||
# exceeds the threshold, gets spilled to a new path, agent reads that one,
|
||
# ad infinitum until MAX_TOOL_CALLS_PER_STEP is hit. Add MCP-provided
|
||
# readers here if they have the same role.
|
||
excluded-tools:
|
||
- read_file
|
||
- read_workspace_memory_file
|
||
conversation:
|
||
window:
|
||
# 测试时临时调低：2000 token ≈ 2000 中文字，3 轮对话即可触发压缩
|
||
# 生产环境应改回 128000
|
||
default-max-input-tokens: 128000
|
||
compact-trigger-ratio: 0.75
|
||
preserve-recent-pairs: 2
|
||
summary-max-tokens: 300
|
||
# Wiki 知识库配置
|
||
wiki:
|
||
enabled: true # 是否启用 Wiki 知识库功能
|
||
max-chunk-size: 30000 # LLM 单次处理最大字符数（超过则分块）
|
||
max-context-chars: 10000 # 注入 Agent prompt 的 Wiki 摘要最大字符数
|
||
max-pages-per-raw: 15 # 单个原始材料最多生成的 Wiki 页面数
|
||
max-parallel-phase-b-pages: 3 # RFC-012 follow-up #3：单 chunk 内 phase B 阶段并行处理的 page 数
|
||
auto-process-on-upload: true # 上传原始材料后是否自动触发 AI 消化
|
||
upload-dir: ./data/wiki-uploads # 上传文件存储目录
|
||
max-scan-files: 500 # 目录扫描最大文件数
|
||
max-scan-file-size: 52428800 # 扫描时跳过大于此大小的文件（字节，默认 50MB）
|
||
# Dream v2 feature flags — production defaults after GA validation
|
||
memory:
|
||
# Phase 1: lifecycle mediator wiring
|
||
lifecycle-mediator-enabled: true
|
||
dream:
|
||
focused-enabled: true
|
||
archive-enabled: true
|
||
archive-keep-days: 30
|
||
max-candidates-per-dream: 100
|
||
# Phase 2: SOUL auto-evolution and provider decorators
|
||
soul-update-interval: 20 # 20 writes trigger one SOUL.md LLM update (0 = off)
|
||
provider-retry-attempts: 1 # 1 = no retry (enable when external providers added)
|
||
provider-metrics-enabled: false # actuator dependency now present; enable when external providers added
|
||
# Phase 3: fact projection
|
||
fact:
|
||
projection-enabled: true
|
||
projection-rebuild-cron: "0 */30 * * * ?"
|
||
llm-extraction-enabled: false # placeholder impl, enable after Phase 3 L4+
|
||
contradiction-check-enabled: false # experimental simple detection, enable after LLM batch impl
|
||
trust-half-life-days: 60
|
||
forget-enabled: true
|