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 或 kingbase 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:
    # Issue #50: explicit Hikari sizing. Default of 10 was being exhausted
    # by O(100) cron jobs firing on minute boundaries (each run holds 3-4
    # connections sequentially) plus ChannelHealthMonitor's per-minute scan.
    # 30 leaves headroom for HTTP / SSE / channel adapters even with the
    # cron concurrency limiter (CronJobService.MAX_CONCURRENT_CRON_RUNS=8)
    # at full saturation.
    hikari:
      maximum-pool-size: 30
      minimum-idle: 5
      connection-timeout: 30000
      idle-timeout: 600000
      max-lifetime: 1800000
      leak-detection-threshold: 60000

  # 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
    # Disable Flyway's built-in ${...} placeholder substitution. Some migrations
    # (e.g. V85 ckjia MCP seed) intentionally store ${ENV_VAR} literals so that
    # mateclaw's runtime header parser (McpClientManager.parseHeaders) can
    # expand them at request time. Flyway would otherwise try to resolve those
    # tokens at migration time and fail with "No value provided for placeholder".
    # Safe to disable globally because no migration in this project relies on
    # Flyway-side ${...} substitution.
    placeholder-replacement: false

  h2:
    console:
      enabled: ${H2_CONSOLE_ENABLED:false}
      path: /h2-console

  # Spring AI Alibaba (DashScope) — Spring AI Alibaba 1.1.x configuration path.
  #
  # The api-key here is only consumed by Spring AI Alibaba's auto-configured beans
  # as a *fallback*. The real source of truth for every provider/key/model is the
  # admin UI ("Settings → Models", persisted in mate_model_provider /
  # mate_model_config); AgentDashScopeChatModelBuilder resolves the key per-request
  # from the provider row first, only falling back to this property when the row
  # is incomplete. The placeholder default keeps DashScopeChatAutoConfiguration
  # happy at startup when no env var is set (Docker / fresh install) — leave it
  # alone unless you know what you're doing.
  ai:
    dashscope:
      api-key: ${DASHSCOPE_API_KEY:configure-in-admin-ui}
      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
    # SQL parameter values can contain prompts/model thinking. Keep them out of
    # logs by default; operators may explicitly opt in for short-lived diagnosis.
    log-impl: ${MATECLAW_MYBATIS_LOG_IMPL:org.apache.ibatis.logging.nologging.NoLoggingImpl}

# SpringDoc OpenAPI
# 注意：/swagger-ui*、/v3/api-docs*、/webjars/** 落在 SecurityConfig 的
# .anyRequest().permitAll()，即 Swagger UI 当前公开可访问。如生产需要收口，
# 在 SecurityConfig 加显式规则（不要在此处配）。
springdoc:
  api-docs:
    path: /v3/api-docs
  swagger-ui:
    path: /swagger-ui.html
  # 把单个嵌套 query 参数对象（如分页 wrapper）拍平成独立字段，减少 schema 噪音
  default-flat-param-object: true

# MateClaw 自定义配置
mateclaw:
  execution-evidence:
    # Observe receipts only. Enforcement requires managed verification scopes and is not available yet.
    mode: observe
    retention-days: 90
    max-summary-bytes: 2048
    max-observations: 32
    default-list-limit: 20
    max-list-limit: 100
    cleanup-interval-ms: 60000
    cleanup-max-batches: 10
  a2a:
    enabled: ${MATECLAW_A2A_ENABLED:false}
    # Public base URL for Agent Cards. Production deployments should set this
    # explicitly so peers do not depend on proxy-derived request headers.
    base-url: ${MATECLAW_A2A_BASE_URL:}
    call-timeout-ms: ${MATECLAW_A2A_CALL_TIMEOUT_MS:120000}
    max-tasks: ${MATECLAW_A2A_MAX_TASKS:1000}
    task-ttl-seconds: ${MATECLAW_A2A_TASK_TTL_SECONDS:3600}
    max-response-bytes: ${MATECLAW_A2A_MAX_RESPONSE_BYTES:1048576}
    outbound-timeout-ms: ${MATECLAW_A2A_OUTBOUND_TIMEOUT_MS:120000}
    allow-private-outbound: ${MATECLAW_A2A_ALLOW_PRIVATE_OUTBOUND:false}
    sweep-interval-ms: ${MATECLAW_A2A_SWEEP_INTERVAL_MS:60000}

  # DeepSeek Harness runtime. The executable and Cordis composition are kept
  # outside the Spring classpath and can be supplied by environment variables.
  agent:
    runtime:
        dsh:
          command: ${DSH_JSONRPC_AGENT:}
          cordis-config: ${DSH_CORDIS_CONFIG:}
          working-directory: ${DSH_CWD:}
          base-url: ${DEEPSEEK_BASE_URL:}
          model-name: ${DEEPSEEK_MODEL:}
          api-key: ${DEEPSEEK_API_KEY:}
          manifest-url: ${DSH_MANIFEST_URL:}
          github-release-url: ${DSH_GITHUB_RELEASE_URL:https://api.github.com/repos/deepseek-ai/deepseek-harness/releases/latest}
          install-root: ${DSH_INSTALL_ROOT:${user.home}/.mateclaw/runtimes/deepseek-harness}
  server:
    # Public base URL used to build absolute download links for tool-generated
    # files (e.g. https://mateclaw.example.com). Leave empty to fall back to the
    # current request's host, and to a relative path when no request is bound.
    # Set this when agents deliver download links to channels/clients that cannot
    # resolve a relative URL (IM messages, copied links, external downloads).
    public-base-url: ${MATECLAW_PUBLIC_BASE_URL:}
  openapi:
    # SpringDoc OpenAPI 元信息（驱动 /swagger-ui.html 与 /v3/api-docs）。
    # server-url 留空时由 SpringDoc 从请求 host 推导，避免 Try it out 打到错误地址。
    # 生产若需固定，通过 MATECLAW_OPENAPI_SERVER_URL 覆盖（如 https://mate.example.com）。
    title: ${MATECLAW_OPENAPI_TITLE:MateClaw REST API}
    version: ${MATECLAW_OPENAPI_VERSION:1.0}
    server-url: ${MATECLAW_OPENAPI_SERVER_URL:}
    # description 留空则使用 OpenApiConfig 中的内置默认描述
    description: ${MATECLAW_OPENAPI_DESCRIPTION:}
    # 是否公开 Swagger UI / OpenAPI 文档路径（/swagger-ui*、/v3/api-docs*、/webjars/**）。
    # true  = 任何人可浏览（本地开发 / 内网默认）；
    # false = 需要全局管理员（ROLE_ADMIN）才能访问，由 SecurityConfig 强制。
    # 生产数据库 profile（mysql/kingbase/postgres）默认覆盖为 false。
    expose-ui: ${MATECLAW_OPENAPI_EXPOSE_UI:true}
  jwt:
    secret: ${JWT_SECRET:MateClaw-JWT-Secret-Key-2024-Please-Change-In-Production}
    expiration: 86400000
  sso:
    # 全局开关（默认关闭，不影响现有部署）
    enabled: ${SSO_ENABLED:false}
    # 「仅允许绑定已有账号」模式（false = 允许自动创建新用户）
    link-only: ${SSO_LINK_ONLY:false}
    # 新建 SSO 用户的默认角色
    default-role: ${SSO_DEFAULT_ROLE:user}
    feishu:
      enabled: ${SSO_FEISHU_ENABLED:false}
      app-id: ${SSO_FEISHU_APP_ID:}
      app-secret: ${SSO_FEISHU_APP_SECRET:}
      # 国际版切换: feishu (国内) / lark (国际版 Lark)
      domain: ${SSO_FEISHU_DOMAIN:feishu}
      # SSO 回调地址，通常 https://your-domain/login?sso=callback
      redirect-uri: ${SSO_FEISHU_REDIRECT_URI:}
  # 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
  tools:
    disclosure:
      # progressive: deferred schemas stay behind tool_search/tool_describe and
      #   execute through tool_call in the same action round. enable_tool remains
      #   available only for backwards compatibility with older conversations.
      # legacy: advertise every bound tool up front (pre-disclosure behavior).
      mode: ${MATECLAW_TOOLS_DISCLOSURE_MODE:progressive}
  context:
    prefix-budget:
      # Ratio remains useful for small contexts; this hard ceiling is what keeps
      # a provider-declared 1M window from advertising ~30k schema tokens forever.
      tool-schema-max-tokens: ${MATECLAW_TOOL_SCHEMA_MAX_TOKENS:12000}
  workspace:
    sandbox:
      # Global fallback filesystem boundary for file/shell tools. When a
      # conversation has no per-workspace base path configured, operations are
      # confined to this root instead of running unconstrained against the whole
      # filesystem (fail-closed default). Set enabled=false to restore the legacy
      # unconstrained behaviour for unconfigured conversations.
      enabled: ${MATECLAW_WORKSPACE_SANDBOX_ENABLED:true}
      root: ${MATECLAW_WORKSPACE_SANDBOX_ROOT:${user.dir}/data/workspace}
  chat:
    upload:
      # Root directory for conversation chat attachments when neither the active
      # agent nor its workspace configures a base path. Attachments resolve to
      # {baseDir}/{conversationId}/{storedName}. When a workspace/agent base path
      # IS configured, attachments land under {basePath}/chat-uploads/{convId}/
      # instead; reads and cleanup still check this default dir so legacy uploads
      # remain resolvable. Defaults to the legacy location for zero-config parity.
      base-dir: ${MATECLAW_CHAT_UPLOAD_BASE_DIR:data/chat-uploads}
      # Organize new attachments/generated media into per-day sub-directories:
      # {convDir}/yyyy-MM-dd/{storedName}. Serving URLs stay flat and reads
      # probe both layouts, so this can be toggled at any time; files written
      # under the previous layout remain resolvable either way. The day is the
      # server's local date (a UTC container groups by UTC days).
      date-folders: ${MATECLAW_CHAT_UPLOAD_DATE_FOLDERS:true}
  skill:
    reflection:
      # Out-of-band skill reflection: after a conversation reaches the cadence
      # below, an async reviewer reads the recent window and creates or
      # improves skills through the same skill_manage pipeline the agent uses.
      # Runs off the request thread, so it never consumes the live turn's
      # context window.
      # Explicit opt-in: transcript/catalog content may be sent to the selected
      # model provider. auto-apply is a second, independent mutation gate.
      enabled: ${MATECLAW_SKILL_REFLECTION_ENABLED:false}
      auto-apply: ${MATECLAW_SKILL_REFLECTION_AUTO_APPLY:false}
      # Review once this many new messages have accumulated since the last
      # attempt. 0 disables the cadence gate entirely.
      review-turn-interval: ${MATECLAW_SKILL_REFLECTION_TURN_INTERVAL:8}
      # Substance floor — a window with fewer assistant turns than this rarely
      # contains a reusable workflow, so the review is skipped before any LLM
      # call is made.
      min-assistant-turns: ${MATECLAW_SKILL_REFLECTION_MIN_ASSISTANT_TURNS:2}
      # Most recent messages handed to the reviewer.
      max-messages: ${MATECLAW_SKILL_REFLECTION_MAX_MESSAGES:24}
      # Per-conversation cooldown between reviews, in minutes. Applies on top
      # of the cadence gate, so a busy conversation reviews at most this often.
      cooldown-minutes: ${MATECLAW_SKILL_REFLECTION_COOLDOWN_MINUTES:30}
      # Hard cap on create/edit/patch actions applied by a single review.
      max-actions-per-run: ${MATECLAW_SKILL_REFLECTION_MAX_ACTIONS:3}
      # Character budget for the existing-skill catalog shown to the reviewer.
      # Skill bodies are truncated to fit; the reviewer is told not to target
      # truncated text with a patch.
      catalog-char-budget: ${MATECLAW_SKILL_REFLECTION_CATALOG_BUDGET:8000}
      # Reviewer model id. Empty follows the system default model.
      model-id: ${MATECLAW_SKILL_REFLECTION_MODEL_ID:}
    routine:
      # Routine mining: a nightly cross-session pass that clusters the opening
      # request of recent conversations and promotes the ones the user makes
      # habitually into class-level skills. Recurrence is invisible to the
      # per-conversation reflection reviewer above — inside one window a weekly
      # request is indistinguishable from a one-off — so this pass supplies the
      # cross-session evidence that reviewer structurally cannot see.
      # Explicit opt-in because mining persists conversation-derived patterns
      # and promotion sends evidence to the selected model provider.
      enabled: ${MATECLAW_SKILL_ROUTINE_ENABLED:false}
      cron: ${MATECLAW_SKILL_ROUTINE_CRON:0 0 3 * * *}
      # How far back each sweep looks. A routine the user stops doing decays
      # out of this window on its own.
      lookback-days: ${MATECLAW_SKILL_ROUTINE_LOOKBACK_DAYS:30}
      # Shingle-similarity above which two openers count as the same request.
      # Tuned toward precision — a false merge invents a routine the user does
      # not have, which is worse than missing one until the next sweep.
      similarity-threshold: ${MATECLAW_SKILL_ROUTINE_SIMILARITY:0.62}
      # Promotion gate. Both must hold: occurrences proves repetition, distinct
      # days proves habit rather than one afternoon of retries.
      min-occurrences: ${MATECLAW_SKILL_ROUTINE_MIN_OCCURRENCES:3}
      min-distinct-days: ${MATECLAW_SKILL_ROUTINE_MIN_DISTINCT_DAYS:3}
      # Shortest opener worth clustering, and the prefix length fed to the
      # shingler.
      min-opener-chars: ${MATECLAW_SKILL_ROUTINE_MIN_OPENER_CHARS:8}
      max-opener-chars: ${MATECLAW_SKILL_ROUTINE_MAX_OPENER_CHARS:400}
      # Conversation ids retained per candidate as promotion evidence.
      max-samples-per-candidate: ${MATECLAW_SKILL_ROUTINE_MAX_SAMPLES:8}
      # Candidates promoted per sweep, bounding LLM cost per run.
      max-promotions-per-run: ${MATECLAW_SKILL_ROUTINE_MAX_PROMOTIONS:2}
      # Conversations scanned per sweep, bounding query and memory cost.
      max-conversations-per-run: ${MATECLAW_SKILL_ROUTINE_MAX_CONVERSATIONS:1000}
      # Transcript shaping for the synthesis prompt.
      transcript-messages-per-sample: ${MATECLAW_SKILL_ROUTINE_TRANSCRIPT_MESSAGES:12}
      transcript-truncate-chars: ${MATECLAW_SKILL_ROUTINE_TRANSCRIPT_TRUNCATE:800}
      # Synthesis model id. Empty follows the system default model.
      model-id: ${MATECLAW_SKILL_ROUTINE_MODEL_ID:}
    upload:
      # Size caps for skill bundle ZIPs (upload endpoint and marketplace
      # install). The archive is buffered in memory during extraction, so
      # max-total-size-mb also bounds peak heap usage per install. Uploads
      # additionally pass through spring.servlet.multipart limits above.
      max-entry-size-mb: ${MATECLAW_SKILL_UPLOAD_MAX_ENTRY_SIZE_MB:1}
      max-total-size-mb: ${MATECLAW_SKILL_UPLOAD_MAX_TOTAL_SIZE_MB:50}
    workspace:
      # Skill workspace root. Override with MATECLAW_SKILL_WORKSPACE_ROOT to
      # relocate it onto a persistent volume — in Docker this is pointed at
      # /app/data/skills so the existing server_data volume persists installed
      # skills, accumulated LESSONS.md, and skill runtime files across restarts.
      root: ${MATECLAW_SKILL_WORKSPACE_ROOT:${user.home}/.mateclaw/skills}
      auto-init: true
      delete-policy: archive
    disclosure:
      load-skill-tool:
        # When false, the load_skill meta tool is not advertised to agents and
        # the skill catalog guidance falls back to readSkillFile. Escape hatch
        # for operators who don't want the explicit skill-load entry point.
        enabled: ${MATECLAW_SKILL_LOAD_SKILL_TOOL_ENABLED:true}
    curator:
      enabled: true
      cron: "0 0 2 * * *"         # daily 02:00 — staggered away from wiki / backup jobs
      stale-after-days: 30
      archive-after-days: 90
      # AGENT_CREATED scopes the sweep to skills written autonomously
      # (origin=agent|routine). Skills a user asked for in a conversation are
      # stamped origin=user and are never aged out under this scope.
      scope: AGENT_CREATED         # AGENT_CREATED | ALL_DYNAMIC | OFF
      protect-prefixes:
        - "sys-"
        - "ops-"
      # Restore point captured before every mutating sweep. The sweep archives
      # skills and, with consolidation on, rewrites their bodies — unattended
      # and overnight, so a bad pass is usually noticed long after it ran.
      # Disabling this makes those changes one-way.
      backup-enabled: ${MATECLAW_SKILL_CURATOR_BACKUP_ENABLED:true}
      backup-keep: ${MATECLAW_SKILL_CURATOR_BACKUP_KEEP:5}
    hub:
      base-url: https://clawhub.ai
      search-path: /api/v1/search
      skills-path: /api/v1/skills
      download-path: /api/v1/download
      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
  security:
    # Hosts/IPs/CIDR blocks allowed through the SSRF guards (browser, hooks, image
    # download) even though they are loopback/private/link-local/metadata. Each
    # entry is a literal hostname, a literal IP, or an IPv4 CIDR block. Keep narrow.
    # Example: [192.168.100.100, 192.168.100.0/24, internal.corp]
    ssrf-allowlist: []
  # 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
  # Multi-agent delegation (DelegateAgentTool).
  delegation:
    # Wall-clock budget for one delegateParallel batch (shared across all
    # children — they run concurrently on virtual threads, so this is total
    # latency, not per-child). 300 s headroom is needed because thinking
    # models (Kimi / GLM / MiniMax) routinely take 90–290 s per LLM turn
    # when the child must produce multi-section structured output.
    parallel-timeout-seconds: 300
    # Default wall-clock budget for detached delegateAsync children. A caller
    # may request a different positive timeout up to 86400 seconds; keeping a
    # default bound prevents abandoned background children from running forever.
    async-timeout-seconds: 3600

# MateClaw Agent 配置
mate:
  channel:
    # 入站消息去重：IM 平台在 ack 迟到 / 丢失 / 非 200 时会重投同一条消息，
    # 不去重则每次重投都会跑一轮完整 Agent 回合，用户看到重复答复。
    dedup:
      enabled: true
      # 需明显长于各平台重投窗口（秒级到分钟级）
      ttl: 5m
      max-size: 2000
  agent:
    # Deterministic Markdown cleanup of the final answer (heading spaces, glued
    # ---, table pipe alignment) before persistence / channel delivery. Set to
    # false to pass model output through verbatim.
    markdown-normalize-enabled: true
    reasoning:
      # How much of a turn's reasoning reaches the message record.
      #   all      — every iteration's reasoning, kept where it happened. The
      #              reasoning behind each tool call is what a replay needs.
      #   terminal — only the iteration that produced the final answer. Smaller
      #              rows, but a long tool loop persists as a bare conclusion.
      retention: all
    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...[TRUNCATED: %d chars total, middle omitted. Do NOT infer or fabricate omitted content; retrieve the full data (e.g. read_file) or tell the user the result is incomplete.]...\n\n"
    tool:
      timeout:
        default-timeout-seconds: 300
        per-category:
          shell: 120
          web: 30
    # Tool-result budget (per-result spill + per-turn aggregate budget).
    # The executor tries to spill the RAW result first so the full output is
    # preserved on disk; the in-context preview points the agent at the spill
    # file via read_file. When spill is disabled, the tool is on the exclusion
    # list, the body is at or below the threshold, or the disk write fails,
    # the executor falls back to inline hard-truncation to the same character
    # cap. Per-turn aggregate caps the combined size across one tool turn.
    tool-result:
      enabled: true
      per-result-threshold-chars: 8000    # aligned with executor hard cap; > this size → spill, ≤ → inline verbatim
      per-turn-budget-chars: 32000        # headroom for multi-tool turns
      preview-head-chars: 800
      excluded-tool-inline-chars: 2500
      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.
      # readSkillFile / load_skill deliberately return the full SKILL.md so the
      # model never misses mandatory sections (API parameter tables etc.);
      # spilling them defeats that and leaves the model an 800-char preview,
      # causing wrong-parameter tool calls. Their references/scripts reads are
      # already self-paginated to 8000 chars, so excluding them stays bounded.
      excluded-tools:
        - read_file
        - read_workspace_memory_file
        - readSkillFile
        - load_skill
      # Spill files are deleted after this many days. Default 0 disables the
      # scheduled sweep entirely so a summary/preview that points at a spill
      # path stays valid for the whole life of the conversation. Files are
      # still purged when the conversation is deleted explicitly via
      # ConversationService.deleteConversation. Raise to a positive value if
      # disk pressure outweighs recoverability for your deployment.
      retention-days: 0
      cleanup-cron: "0 0 3 * * ?"
    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
    # Always-on injection budget — bounds the per-turn size of the user/feedback structured block
    system-block-max-chars: 4000               # char cap on the always-on block; over budget drops oldest by Updated date (LRU); 0 = unlimited
    system-block-max-entries-per-type: 40       # max entries injected per type (user/feedback); 0 = unlimited
    # Structured-memory consolidation — separate maintenance task; LLM merges duplicate/stale user/feedback entries (shared + per-owner) to curb storage growth
    structured-consolidation-enabled: true            # off = injection cap only, no storage-side merge
    structured-consolidation-min-entries: 8            # buckets with fewer entries skip the LLM call to save cost
    structured-consolidation-cron: "0 30 3 * * ?"     # own schedule, decoupled from dreaming-enabled / dreaming-cron
    structured-consolidation-max-owners-per-run: 50    # cap LLM cost per agent per run; remaining owners picked up next run; 0 = unlimited
    # Always-on file ceilings — deterministic backstop so PROFILE.md / MEMORY.md (LLM-rewritten) cannot grow per-turn context without bound
    profile-max-chars: 4000                            # PROFILE.md hard cap; truncates at a section boundary if the rewrite overruns; 0 = unlimited
    memory-md-max-chars: 8000                          # MEMORY.md hard cap; 0 = unlimited