config.example.yaml

# Hone configuration example
#
# Usage:
# 1) Copy this file to config.yaml
# 2) Fill in or adjust the values as needed. Most provider tokens are config-owned;
#    public SMS/Captcha and selected operational secrets use runtime env as noted below.

# UI / CLI display language: zh or en
language: zh

# Process runtime timezone (IANA name). This is the single default for cron,
# date keys, report schedules, and user-visible local timestamps. In production
# containers, set it explicitly because host detection commonly resolves to UTC.
timezone: "Asia/Shanghai"

# iMessage config
imessage:
  enabled: false
  # Leave empty to listen to all users; otherwise only the specified user is tracked
  target_handle: ""
  db_path: "~/Library/Messages/chat.db"
  poll_interval: 2 # polling interval in seconds
  # Local HTTP listener used by iMessage-side scheduled task callbacks
  listen_addr: "127.0.0.1:8078"

# Feishu channel config (Rust business process + Go official SDK facade)
feishu:
  enabled: false
  # Feishu credentials are read from config.yaml
  app_id: ""
  app_secret: ""
  # Internal allowlist of email addresses; empty means allow all; supports "*"
  allow_emails: []
  # Internal allowlist of mobile numbers; accepts "+8613800138000" or "13800138000"
  allow_mobiles: []
  # Fallback allowlist: when the contact API cannot provide an email or mobile number, allow the open_id directly
  allow_open_ids: []
  # Chat scope:
  # - DM_ONLY: only process direct messages
  # - GROUPCHAT_ONLY: only process group chats
  # - ALL: process both
  # Legacy dm_only is still accepted for backward compatibility.
  chat_scope: DM_ONLY
  # Maximum length of a single message; longer messages are split on the Rust side
  max_message_length: 3500
  # Local JSON-RPC address from Rust to the Go facade
  facade_url: "http://127.0.0.1:18115/rpc"
  # Listener address for callback events from the Go facade to Rust
  callback_addr: "127.0.0.1:18116"
  # The Go facade's own listener address; used for local health checks
  facade_addr: "127.0.0.1:18115"
  # Timeout, in seconds, for waiting on the facade long connection at startup
  startup_timeout_seconds: 20

# Telegram Bot config
telegram:
  enabled: false
  # Telegram Bot token
  bot_token: "REPLACE_ME"
  # Allowed user ID list; empty means allow all; supports "*"
  allow_from: []
  # Chat scope:
  # - DM_ONLY: only process direct messages
  # - GROUPCHAT_ONLY: only process group chats
  # - ALL: process both
  # Legacy dm_only is still accepted for backward compatibility.
  chat_scope: DM_ONLY
  # Maximum length of a single message (Telegram sendMessage limit is about 4096)
  max_message_length: 3500

# Discord Bot config
discord:
  enabled: false
  # Discord Bot token
  bot_token: ""
  # Allowed user ID list; empty means allow all; supports "*"
  allow_from: []
  # Chat scope:
  # - DM_ONLY: only process direct messages
  # - GROUPCHAT_ONLY: only process group chats
  # - ALL: process both
  # Legacy dm_only is still accepted for backward compatibility.
  chat_scope: DM_ONLY
  # Maximum length of a single message (Discord limit is 2000)
  max_message_length: 1800
  group_reply:
    # Whether explicit-trigger group replies are enabled
    enabled: true
  watch:
    enabled: false
    # Channel IDs to watch; can also be specified on the command line with --channel-id (higher priority)
    channel_ids: []
    # Run mode: true = long-running loop; false = one-shot run (suitable for cron)
    loop: true
    # Whether to print verbose logs (also prints on ticks with no new messages)
    verbose: false

# Group-chat shared session / context compression policy
group_context:
  # Whether group-context governance is enabled
  enabled: true
  # Whether each group shares one session
  shared_session_enabled: true
  # Whether untriggered group text should be buffered for the next explicit trigger
  pretrigger_window_enabled: true
  # Maximum buffered group messages to carry into the next explicit trigger
  pretrigger_window_max_messages: 10
  # Only keep buffered group messages from the latest N seconds
  pretrigger_window_max_age_seconds: 300
  # Number of recent valid messages to restore for a shared group session
  recent_context_limit: 18
  # Trigger compression after the shared session exceeds this number of messages
  compress_threshold_messages: 24
  # Trigger compression after the shared session exceeds this many bytes
  compress_threshold_bytes: 49152
  # Keep the latest N messages after compression
  retain_recent_after_compress: 8

# Nano Banana image generation config
nano_banana:
  enabled: true
  # Single implementation: use the OpenRouter image generation API (modalities=["image","text"])
  base_url: "https://openrouter.ai/api/v1"
  model: "google/gemini-2.0-flash-exp"
  default_image_count: 3
  # API key prefers the llm.providers.openrouter.api_key/api_keys pool; legacy single-key
  # OpenRouter fields remain readable during migration.
  # Request timeout, max_tokens, temperature, HTTP referrer, and title currently use fixed client defaults.
  download_dir: "gen_images"

# Financial Modeling Prep API config
# Outbound transactional email for push subscriptions (Cloudflare Email Sending).
# Leave api_token empty and export HONE_CLOUDFLARE_EMAIL_API_TOKEN instead — a credential
# that can send mail from a verified domain must never live in the repository.
# With no token resolved the feature is simply off: pushes still deliver, they
# just do not also arrive by mail.
email:
  account_id: "52dfc1420d779b403c08196b792ce926"
  from_address: "no-reply@hone-claw.com"
  from_name: "HONE"
  api_token: ""
  api_token_env: "HONE_CLOUDFLARE_EMAIL_API_TOKEN"
  api_base: "https://api.cloudflare.com/client/v4"
  timeout: 20
  # Hone's own ceiling, below Cloudflare's daily quota, so a scheduler fault
  # cannot burn the whole allowance in one morning.
  daily_limit: 150
  # Signs the login-free unsubscribe links carried by every delivered push.
  # Environment-only. Empty means no link is emitted rather than a broken one.
  unsubscribe_secret_env: "HONE_UNSUBSCRIBE_SECRET"

fmp:
  # Legacy single-key fallback; prefer api_keys for new configs
  api_key: ""
  # Financial Modeling Prep API key pool with failover support
  api_keys: []
  base_url: "https://financialmodelingprep.com/api"
  timeout: 60

# Proactive event engine: polls FMP for market events, deduplicates, and
# dispatches to subscribers (per-holding or global). High-severity events push
# immediately; Medium/Low roll up into pre/post-market digests.
event_engine:
  # Master switch. Default off — enable only when you want proactive pushes.
  enabled: false
  # poll_intervals 覆盖 news / price，以及需要分钟级发现的 earnings_surprise。
  # earnings calendar / corp_action / macro / analyst_grade / sec_filings 走 cron-aligned ——
  # 在 digest.default_slots 各 slot 前 `prefetch_offset_mins` 分钟跑一次,
  # 保证 flush 推送的数据永远是刚拉的,不受用户重启时机漂移。
  # earnings_surprise 只在美股常见盘前/盘后财报发布窗口访问网络，窗口外 no-op；
  # 冷启动时每个 poller 立即跑一次。
  poll_intervals:
    # 新闻。最小推荐 300 (5m)，典型 600-1800 (10-30m)，最大 3600 (1h)。
    # 低于 300 时 FMP 返回不会更新，只是多烧配额。
    news_secs: 1800 # 30 min
    # 盘中价格异动。最小推荐 60 (1m)，典型 180-600 (3-10m)，最大 900 (15m)。
    # 降到 60 时 N 只持仓 × 每分钟一次 quote，会显著吃 FMP 配额；
    # 非交易时段 poller 会自动 skip，不必担心夜里空烧。
    price_secs: 900 # 15 min
    # 正式财报实际值。仅 ET 工作日 04:00-10:00、16:00-23:00 请求；默认 10m，
    # 配置下限 60s。成功质量判断后同一季度不会重复调用 SEC/LLM。
    earnings_surprise_secs: 600
  digest:
    # 管理员视角默认槽位时刻 —— 在 actor 没有自定义 prefs.digest_slots 时,
    # UnifiedDigestScheduler 用这组时刻作为默认 slot 触发推送。每个 slot 可配
    # `time`(本地 HH:MM)和可选 `label`(渲染 header 用)。
    # Defaults target CN users: 08:30 卷起夜里队列成早间摘要;09:00 把美股盘后
    # 隔夜汇总延后到醒来后,避免半夜打扰。
    default_slots:
      - {time: "08:30", label: "盘前摘要"}
      - {time: "09:00", label: "晨间摘要"}
    # Max items rendered per digest flush. Events beyond this (sorted by severity
    # desc, occurred_at desc) are dropped with a "…… 另 N 条已省略" footer so a
    # Low-severity storm can't bury the user. 0 disables the cap.
    max_items_per_batch: 20
    # cron-aligned poller 在每个 default slot 时刻前多少分钟执行拉取。
    # 日频来源不再按固定 24h interval 轮询,而是在 default_slots 各 slot 前 offset 跑一次,
    # 保证推送数据永远是 flush 前刚拉的。
    # 默认 30:数值越小数据越新,但留给 EventStore/Router 的缓冲越紧;低于 5 不推荐。
    prefetch_offset_mins: 30
    # 同一用户两次 digest 的最小间隔。用户配置多个窗口时，避免 09:00/12:00/16:00
    # 反复消费同一批主题；0 表示不限制。
    min_gap_minutes: 180
  thresholds:
    # |pct| < low → no event; low ≤ |pct| < high → Low (digest); |pct| ≥ high → High (immediate push)
    # Production defaults tuned for US equities: 10% rarely triggers, 5% misses many real moves.
    price_alert_low_pct: 2.5
    price_alert_high_pct: 6.0
    volume_sigma: 2.0
    high_severity_daily_cap: 8
    # Minimum minutes between two High sink pushes for the *same* ticker.
    # Protects against one stock firing price + news + filing in quick succession.
    # Later Highs within the window get demoted to digest (status="cooled_down").
    # 0 disables the cooldown.
    same_symbol_cooldown_minutes: 60
    # Low→Medium news convergence caps. Per-symbol prevents one ticker from
    # dominating; per-tick prevents broad multi-ticker bursts crowding digest.
    news_upgrade_per_symbol_per_tick: 3
    news_upgrade_per_tick: 12
    # 用户 per-actor price_high_pct_override 低于该值时，只会让价格异动进 digest；
    # 即时推送仍要求达到该系统最小阈值，除非 payload 标明高仓位。
    price_min_direct_pct: 6.0
    # 价格异动跨档再提醒步长。high=6, step=2 时,盘中跨 +6/+8/+10
    # 或 -6/-8/-10 会形成独立事件；未跨新档只更新摘要最新态。
    price_realert_step_pct: 2.0
    # 价格 band 单一推送规则(v0.5.2 起替代旧 cap+gap 双保险):新档 pct 必须比
    # 当日已 sink-sent 的最大档高出本字段值,否则降级 digest。0 = 关闭(无脑全推)。
    # 默认 2.0 与 price_realert_step_pct 一致,意为「每跨一个新 band 必推」 ——
    # 避免大行情长尾失声(POC 实证 AAOI 6→16 序列 cap=2 仅 6/8,advance=2 全 6 档)。
    price_band_min_advance_pct: 2.0
    # 收盘 price_close 默认只进摘要,避免美股收盘在北京时间凌晨即时打扰。
    price_close_direct_enabled: false
    large_position_weight_pct: 20.0
    # High macro 只有临近发生窗口才即时推；未来 7 天日历保留为摘要/预告。
    macro_immediate_lookahead_hours: 6
    macro_immediate_grace_hours: 2
  renderer:
    # Severity list that goes through LLM polish. Empty list = pure template.
    # polish_llm names an llm.profiles entry; empty falls back to legacy llm.openrouter.sub_model/model.
    llm_polish_for: []
    polish_llm: aux
    template_dir: "./data/event_templates"

  # 财报 poller 参数。
  # - window_days:EarningsPoller 每 tick 向 FMP earning_calendar 拉 [today, today+N]。
  #   这决定 Hone 从多早开始"关注"一家公司的财报 —— 默认 14 天:启动时你会一次看到
  #   未来 14 天内所有持仓的财报预告(Medium 级,进 digest)。
  #   **v0.1.46 起**,poller 只产出稳定 id 的 `earnings:{SYM}:{DATE}` teaser 事件
  #   (Medium 入 digest);T-3/T-2/T-1 倒计时由 UnifiedDigestScheduler 在每个 slot
  #   触发时从 EventStore 现算 —— 这样 poller cron 漂移不会让倒计时 off-by-one,
  #   T-1 也由早 8:30 的 pre-market 窗口天然覆盖,不再单独升 High。整条 lifecycle
  #   共享 `earnings_upcoming` kind,把它加进 `blocked_kinds` 即全程静音。
  #   典型取值:7(仅本周)、14(两周,默认)、21(三周,但信号老化)。小于 3 会让 T-3
  #   倒计时失效(因为 poller 看不到 T-3 的财报)。
  # quality_review: EarningsSurprisePoller 原始数据只有 EPS actual vs estimate。
  #   对亏损公司或 EPS 接近 0 的公司,百分比 surprise 会非常误导。开启后,若近期
  #   SEC 8-K 财报新闻稿可抓取且 OpenRouter 可用,会用 LLM 综合收入、指引、backlog、
  #   GAAP/non-GAAP 利润、EBIT/EBITA/EBITDA、现金流与风险来判断 immediate / digest。
  #   失败、缺少上下文或低置信时直接跳过 candidate,不再产出 EPS-only 推送。
  earnings:
    window_days: 14
    quality_review:
      enabled: true
      llm: earnings_quality
      model: x-ai/grok-4.5
      max_review_tokens: 1800
      min_review_confidence: 0.65
      min_immediate_confidence: 0.9
      sec_recent_hours: 72
      context_max_chars: 9000
    # A-tier 公司在财报投递后异步执行，不阻塞 T0。它逐项核对画像里的未决问题
    # 与管理层承诺，因此必须使用独立于短财报卡的 completion budget。
    continuity_review:
      enabled: true
      llm: earnings_continuity
      model: x-ai/grok-4.5
      max_review_tokens: 3600

  # SEC filings poller 配置。
  # forms: 每 tick 对每只 watchlist ticker 拉哪些 form 类型。FMP 的 /v3/sec_filings
  #   必须按 type 过滤,不支持一次取多 type;每个 form 一次 HTTP。
  #   Severity 由 form 决定(代码硬映射,不在 config 里):
  #     8-K / S-1            → High(突发披露 / 增发,稀释信号)
  #     10-Q / 10-K          → Medium(季报 / 年报)
  #     DEF 14A              → Low(委托书 / 治理薪酬)
  #   想关掉某 form 直接从这里删掉它,不需要改代码。空数组等于关闭整个 poller。
  # enrichment: 是否给 SEC filing 调 LLM 生成 ~200 字业务摘要(长期主线投资者
  #   视角:跳过 GAAP 数字,抓 backlog / 资本配置 / 风险因子新增)。
  #   当前默认使用 x-ai/grok-4.3 替代已下线的 Grok 4.1 Fast；
  #   实际质量、延迟和成本以 OpenRouter 当前模型为准。
  #   失败 fallback 走原始 form/link body,不阻塞推送。
  sec_filings:
    forms:
      - 8-K
      - 10-Q
      - 10-K
      - S-1
      - DEF 14A
    enrichment:
      enabled: true
      llm: filing_summary
      model: x-ai/grok-4.3
      max_summary_tokens: 800
      # SEC EDGAR 强制要求 User-Agent 包含联系邮箱,否则会被限流或拒绝。
      # 部署方应改成自己的真实联系邮箱(SEC 不强制可达,但需有公司 / 产品名 + 邮箱)。
      user_agent: "honeclaw event-engine ops@honeclaw.local"

  # 全局 digest LLM 子配置 —— UnifiedDigestScheduler 在每个 slot 触发时把这一节
  # 作为 global news path 的采集 / dedup / LLM 旋钮使用。
  # 触发完全由 per-actor `prefs.digest_slots` 驱动,这里不再有调度时刻。
  # 候选池流水:
  # - 候选池 = RSS news + FMP trusted news(允许 Low 进入预选);collector 再过滤
  #   legal_ad / transcript / 已广播事件。Macro floor 不从这里采集,由 personalize 分类产生。
  # - Pass 1 (pass1_llm/pass1_model) 批量打分 + cluster + 一句话 takeaway
  # - Pass 2 (pass2_llm/pass2_model) 抓原文(若 fetch_full_text=true) + 精读 + 短评 + 排序
  # 输出经 unified scheduler per-actor fan-out,与 buffer/synth 候选合流后渲染。
  # 默认 enabled=false。`enabled=false` 时 unified scheduler 不装配 global curator,
  # global news section 跳过;pipeline 仍按 buffer/synth 候选推送 actor。
  global_digest:
    enabled: false
    lookback_hours: 24 # 候选池回看窗口
    # *_llm fields name llm.profiles entries; matching *_model fields are legacy OpenRouter fallbacks.
    pass1_llm: digest_fast
    pass1_model: x-ai/grok-4.3
    pass2_llm: digest_strong
    pass2_model: x-ai/grok-4.3 # 替代已下线的 Grok 4.1 Fast；实际价格 / context / reasoning 能力以 OpenRouter 当前模型为准
    pass2_top_n: 15 # Pass 1 排序后送 Pass 2 精读的候选数上限
    final_pick_n: 8 # 单次推送最终保留的条数上限
    fetch_full_text: false # Pass 2 是否 GET 原文 → html2text;失败 fallback FMP text
    # Jina Reader API key,用于 Pass 2 抓不到原文时(reuters / wsj / barrons / nyt 等
    # 付费墙站点)走 r.jina.ai 二次抓取。免费层 1M tokens/月,jina.ai 注册即得。
    # 留空则跳过 Jina fallback,直接落到 FMP text。
    jina_api_key: ""
    # 事件级去重(POC 验证 2026-04-26):collector 之后 / Pass1 之前,用强 LLM
    # 把同事件多源报道合并成 1 条代表,避免 picks 被 5 条 Hormuz 这种重复挤满。
    # 必须用强模型,nova-lite 这种会过度归类成 theme(把不同事件合到大筐里)。
    event_dedupe_enabled: true
    event_dedupe_llm: digest_strong
    mainline_distill_llm: mainline_short
    event_dedupe_model: x-ai/grok-4.3

  # Per-source 开关。每个字段对应 hone-event-engine 内一个 EventSource,
  # 设成 false 即不 spawn 该 source(最省 FMP 配额)。
  # 想要"poller 仍跑、只是 router 丢弃某 kind"的兜底关法,改用下方 disabled_kinds。
  sources:
    news: true # /v3/stock_news → NewsCritical
    price: true # /v3/quote 按 watch pool 拉 → PriceAlert / Weekly52High / Weekly52Low
    extended_hours: true # /v3/historical-chart/1min?extended=true → pre/post-market PriceAlert
    earnings_calendar: true # /v3/earning_calendar → EarningsUpcoming
    corp_action: true # 拆股 / 分红日历 → Split / Dividend
    sec_filings: true # SEC filings per-ticker → SecFiling(forms come from sec_filings.forms above)
    macro_calendar: true # /v3/economic_calendar → MacroEvent
    analyst_grade: true # 按 watch pool 拉 → AnalystGrade
    earnings_surprise: true # 按 watch pool 拉 → EarningsReleased
    # 通用 RSS 新闻源 —— 给 global_digest 候选池补 FMP 漏掉的料。
    # POC 验证 FMP 漏掉 Bloomberg 93%、SpaceNews 100%、STAT News 100% 的关键料。
    # 入 events 表 source = "rss:{handle}",collector 与 FMP news 一同拉。
    rss_feeds:
      - handle: bloomberg_markets
        url: https://feeds.bloomberg.com/markets/news.rss
        interval_secs: 1800
      - handle: spacenews
        url: https://spacenews.com/feed/
        interval_secs: 3600
      - handle: stat_news
        url: https://www.statnews.com/feed/
        interval_secs: 3600
      # 大V速报：白毛使用用户确认的 aichainmap 公共整理 feed，条目仍以 X 原链
      # 为内容来源；SemiAnalysis 使用官方 RSS。Jukan 仍需部署方配置合法 bridge。
      - handle: influencer_serenity
        url: https://serenity-webhook.pages.dev/feed
        interval_secs: 900
      - handle: influencer_semianalysis
        url: https://semianalysis.com/feed
        interval_secs: 3600
      # 关键事件链一手来源：只有注册的 key_event_* handle 且 feed 与条目均在
      # 对应官方域名时，才可标为“一手确认”。没有稳定官方 RSS 的公司应通过
      # 管理员研究资料库导入官方原文，不能用搜索摘要或自媒体替代。
      - handle: key_event_openai
        url: https://openai.com/news/rss.xml
        interval_secs: 3600
    # Telegram 公开频道监听(web preview `t.me/s/<handle>`),产出 social_post。
    # 空数组表示不启用;开启时每条配置会启动一个独立 poller loop。
    telegram_channels: []
  # 全局 event kind 黑名单。router 在 per-user NotificationPrefs 之前先过一遍;
  # 入库仍然发生(便于日报 / 统计),只是不再分发给任何 actor。即使用户把某个
  # kind 放进 allow_kinds,在这里禁用后仍然不会推。合法 tag:
  #   earnings_upcoming, earnings_released, earnings_call_transcript,
  #   news_critical,
  #   price_alert, weekly52_high, weekly52_low,
  #   dividend, split, sec_filing, analyst_grade,
  #   macro_event, social_post
  disabled_kinds: []
  # news_classifier_llm names an llm.profiles entry; news_classifier_model is the
  # legacy OpenRouter fallback when that profile is empty. Keep CI/tests mocked;
  # use real API only for manual live-quality checks.
  news_classifier_llm: news_classifier
  news_classifier_model: "x-ai/grok-4.3"
  # Global default phrase for judging whether uncertain-source news is important.
  # Per-actor notif_prefs news_importance_prompt overrides this value.
  news_importance_prompt: "公司或潜在影响公司长期逻辑和宏观叙事的重大事件"

llm:
  provider: openrouter
  default_profile: main
  auxiliary_profile: aux
  openrouter:
    # Legacy config-only fallback. Prefer the llm.providers.openrouter.api_key/api_keys pool.
    model: "moonshotai/kimi-k2.5"
    # Legacy auxiliary fallback model for heartbeats / session compression / other background checks.
    # This example uses llm.auxiliary_profile instead.
    sub_model: "moonshotai/kimi-k2.5"
    timeout: 120
    max_retries: 3
    max_tokens: 32768
  providers:
    openrouter:
      # OpenRouter-specialized OpenAI-compatible provider kind.
      # Use openai_compatible for other OpenAI-compatible endpoints.
      kind: openrouter
      base_url: "https://openrouter.ai/api/v1"
      api_key: ""
      api_keys: []
      timeout: 120
      max_retries: 3
  profiles:
    main:
      provider: openrouter
      model: "moonshotai/kimi-k2.5"
      params:
        max_tokens: 32768
    aux:
      provider: openrouter
      model: "moonshotai/kimi-k2.5"
      params:
        max_tokens: 4096
    news_classifier:
      provider: openrouter
      model: "x-ai/grok-4.3"
      params:
        max_tokens: 64
        temperature: 0
    filing_summary:
      provider: openrouter
      model: "x-ai/grok-4.3"
      params:
        max_tokens: 800
        temperature: 0.2
        # SEC enrichment 的 prompt/renderer 契约是纯文本；不要在这里强制 json_object。
    earnings_quality:
      provider: openrouter
      model: "x-ai/grok-4.5"
      params:
        max_tokens: 1800
        temperature: 0.2
        reasoning:
          effort: low
        response_format:
          type: json_object
    earnings_continuity:
      provider: openrouter
      model: "x-ai/grok-4.5"
      params:
        max_tokens: 3600
        temperature: 0.2
        reasoning:
          effort: low
        response_format:
          type: json_object
    digest_fast:
      provider: openrouter
      model: "x-ai/grok-4.3"
      params:
        max_tokens: 1200
        temperature: 0.2
    digest_strong:
      provider: openrouter
      model: "x-ai/grok-4.3"
      params:
        max_tokens: 1600
        temperature: 0.2
        reasoning:
          effort: low
    mainline_short:
      provider: openrouter
      model: "x-ai/grok-4.3"
      params:
        max_tokens: 1200
        temperature: 0.2
  auxiliary:
    # Optional legacy OpenAI-compatible auxiliary path for heartbeats / session compression.
    # Used only when llm.auxiliary_profile is empty; then it overrides openrouter.sub_model.
    base_url: ""
    api_key: ""
    model: ""
    timeout: 120
    max_retries: 3
    max_tokens: 32768

storage:
  # Anchors file-backed state paths; sessions themselves live in PostgreSQL.
  sessions_dir: "./data/sessions"
  # Per-actor daily conversation quota files. Legacy conversation_quota_db_path
  # is still accepted as an alias, but new configs should use this directory key.
  conversation_quota_dir: "./data/conversation_quota"
  llm_audit_retention_days: 30
  llm_audit_enabled: true
  portfolio_dir: "./data/portfolio"
  cron_jobs_dir: "./data/cron_jobs"
  gen_images_dir: "./data/gen_images"
  notif_prefs_dir: "./data/notif_prefs"

cloud:
  # PostgreSQL is the mandatory authoritative store in every deployment mode.
  # OSS remains optional; without it, generated files use their local fallback path.
  # Keep real PG / OSS credentials in the process environment or local .env only.
  # Set cloud.strict_no_local_storage=true to reject those optional local file fallbacks.
  # HONE_CLOUD_STRICT_NO_LOCAL_STORAGE=true has the same startup-fail effect.
  strict_no_local_storage: false
  # Private-R2 public-community delivery is opt-in. Keep `off` until the edge
  # Worker, private bucket binding, and shared signing secret are ready.
  community_delivery:
    # off | shadow | prefer. Both shadow/prefer may issue the short-lived edge
    # cookie; clients continue using the legacy API unless separately enabled.
    mode: "off"
    # Runtime clamps this value to 60..3600 seconds.
    token_ttl_secs: 900
    # Name of the environment variable containing the HMAC secret. After
    # trimming whitespace it must be 32..1024 UTF-8 bytes. Never put the
    # secret bytes themselves in this YAML file.
    secret_env: "HONE_COMMUNITY_EDGE_HMAC_SECRET"
  postgres:
    database_url: ""
    database_url_env: "DATABASE_URL"
    host: ""
    host_env: "HONE_POSTGRES_HOST"
    port: 0
    port_env: "HONE_POSTGRES_PORT"
    user: ""
    user_env: "HONE_POSTGRES_USER"
    password: ""
    password_env: "HONE_POSTGRES_PASSWORD"
    database: ""
    database_env: "HONE_POSTGRES_DATABASE"
    sslmode: "disable"
    # Optional local proxy, for example socks5://127.0.0.1:1082.
    proxy: ""
    proxy_env: "HONE_POSTGRES_PROXY"
    # Force direct TCP for PG even if proxy is configured in env/config.
    no_proxy: false
    no_proxy_env: "HONE_POSTGRES_NO_PROXY"
  oss:
    # aliyun_oss | r2 | s3
    provider: "aliyun_oss"
    provider_env: "HONE_OSS_PROVIDER"
    access_key_id: ""
    access_key_id_env: "HONE_OSS_ACCESS_KEY_ID"
    access_key_secret: ""
    access_key_secret_env: "HONE_OSS_ACCESS_KEY_SECRET"
    bucket: ""
    bucket_env: "HONE_OSS_BUCKET"
    endpoint: ""
    endpoint_env: "HONE_OSS_ENDPOINT"
    region: ""
    region_env: "HONE_OSS_REGION"
    public_upload_prefix: "public-uploads"
    # Optional reqwest proxy, for example socks5://127.0.0.1:1082.
    proxy: ""
    proxy_env: "HONE_OSS_PROXY"

logging:
  level: "INFO"
  # Parsed for compatibility, but setup_logging currently always installs the
  # console formatter and does not create a file appender from logging.file.
  console: true
  file: "./data/logs/hone.log"
  # Local UDP log sink port. null uses the runtime default port 18118; there is
  # currently no config-level disable switch.
  udp_port: null

admins:
  # iMessage admin handle list (phone number or Apple ID)
  # These users have admin privileges in the iMessage channel (they can modify code and restart Hone)
  imessage_handles:
    - "" # example: "+13234567890" or "admin@example.com"
  # Telegram admin user ID list (Telegram numeric IDs)
  telegram_user_ids:
    - "" # example: "8039067465"
  # Feishu admin email list
  feishu_emails:
    - "" # example: "admin@example.com"
  # Feishu admin mobile list
  feishu_mobiles:
    - "" # example: "+8613800138000"
  # Feishu admin open_id list
  feishu_open_ids:
    - "" # example: "ou_xxx"
  # Discord admin user ID list (Discord numeric IDs)
  discord_user_ids:
    - "" # example: "123456789012345678"
  # Optional runtime admin registration passphrase.
  # Prefer setting this through the env var below instead of committing plaintext secrets.
  runtime_admin_registration_passphrase: ""
  # Env var name used when runtime_admin_registration_passphrase is empty
  runtime_admin_registration_passphrase_env: "HONE_ADMIN_REGISTER_PASSPHRASE"

web:
  # Bearer token for the Web console (recommended for remote deployments; can be left empty locally)
  auth_token: ""
  # Research report API
  research_api_base: "https://research.example.com"
  research_api_key: ""
  # Local private workflow runner base URL used by the cross-channel /report shortcut
  local_workflow_api_base: "http://127.0.0.1:3213"
  # Validate code required by the local private workflow runner /report shortcut
  local_workflow_validate_code: ""
  # Env var name used when local_workflow_validate_code is empty
  local_workflow_validate_code_env: "HONE_REPORT_VALIDATE_CODE"
  # Public SMS login uses Aliyun PNVS VerifyCode APIs and is configured by env:
  # - ALIBABA_CLOUD_ACCESS_KEY_ID / ALIBABA_CLOUD_ACCESS_KEY_SECRET (required)
  #   Compat aliases: ALIYUN_ACCESS_KEY_* or HONE_ALIYUN_ACCESS_KEY_*
  # - HONE_ALIYUN_SMS_ENDPOINT (default: dypnsapi.aliyuncs.com)
  # - HONE_ALIYUN_SMS_COUNTRY_CODE (default: 86; backend sends local CN numbers)
  # - HONE_ALIYUN_SMS_SIGN_NAME (default: 速通互联验证码)
  # - HONE_ALIYUN_SMS_TEMPLATE_CODE (default: 100001)
  # - HONE_ALIYUN_SMS_TEMPLATE_PARAM (default: {"code":"##code##","min":"5"})
  # - HONE_PUBLIC_SECURE_COOKIE=true/1/yes or false/0/no may force the public
  #   session cookie Secure flag; invalid values keep Secure=true.
  # Existing admin-created Web invite users are treated as the public-login invite list.
  # Optional Aliyun Captcha 2.0 guard for public SMS sends is configured by env:
  # - ALIBABA_CLOUD_ACCESS_KEY_ID / ALIBABA_CLOUD_ACCESS_KEY_SECRET (same AccessKey env as SMS)
  # - HONE_ALIYUN_CAPTCHA_PREFIX / HONE_ALIYUN_CAPTCHA_SCENE_ID (required to enable)
  # - HONE_ALIYUN_CAPTCHA_REGION (default: cn; use sgp for Singapore)
  # - HONE_ALIYUN_CAPTCHA_ENDPOINT (optional override)
  # - HONE_ALIYUN_CAPTCHA_ENABLED=false disables the guard even when prefix/scene are set.

security:
  # Whether KB queries are isolated by actor/session (recommended to avoid cross-user retrieval)
  kb_actor_isolation: true
  tool_guard:
    enabled: true
    # block | audit (audit only records the warning and does not block)
    mode: "block"
    # Tool matching rules to apply (supports "*" and "!<pattern>" exclusions)
    apply_tools:
      - "*"
      - "!web_search"
      - "!data_fetch"
    # High-risk fragments that trigger blocking (matched in lowercase)
    deny_patterns:
      - "rm -rf"
      - "rm -fr"
      - "rm -r /"
      - "rm -rf /"
      - "rm -fr /"
      - "mkfs"
      - "dd if=/dev/zero"
      - "dd if=/dev/random"
      - "shutdown -h"
      - "shutdown -r"
      - "reboot"
      - "poweroff"
      - "halt"
      - ":(){ :|:& };:"
      - "del /s /q"
      - "rd /s /q"
      - "format c:"
      - "curl | sh"
      - "wget | sh"

agent:
  # Recommended: hone_cloud | codex_acp | opencode_acp.
  # Other supported runners: gemini_cli | codex_cli.
  # This example uses Codex ACP as the default local runner.
  # gemini_acp is disabled at runtime and kept below only for migration/reference.
  runner: "codex_acp"
  # Only used when runner=codex_cli; empty means use the codex-cli default model
  codex_model: "gpt-5.6-sol"
  # Daily successful-reply quota for user conversations. Set to 0 to disable the limit.
  daily_conversation_limit: 100
  # Single-step timeout used by internal runner phases such as initialize/load and stream idle gaps
  step_timeout_seconds: 180
  # Full-run timeout used for one complete answer round
  overall_timeout_seconds: 1200
  # Verbose runner debug logging is controlled by HONE_AGENT_DEBUG, not by a config field.

  system_prompt_path: "./soul.md"
  gemini_acp:
    # Deserializable legacy config only; do not set agent.runner=gemini_acp.
    command: "gemini"
    args: [ "--experimental-acp", "--sandbox", "--approval-mode", "plan" ]
    # Optional; if unset, reuse the local gemini-cli login state
    api_key: ""
    # Optional; empty uses the gemini CLI default model
    model: "gemini-2.5-pro"
  codex_acp:
    command: "codex-acp"
    args: []
    # Startup checks: @openai/codex >= 0.146.0 and @agentclientprotocol/codex-acp >= 1.1.7.
    codex_command: "codex"
    # GPT-5.6 Sol is the default high-capability Codex ACP model.
    # Hone applies model, reasoning effort, safety settings, and native developer
    # instructions through the codex-acp CODEX_CONFIG boundary before startup.
    # One deterministic Hone conversation maps to one persistent native Codex session;
    # later turns resume it and let the Codex harness own history and compaction.
    model: "gpt-5.6-sol"
    # Reasoning effort passed to Codex as model_reasoning_effort.
    variant: "xhigh"
    # Recommended: keep workspace-write + never, and let Hone run inside the actor sandbox
    sandbox_mode: "workspace-write"
    approval_policy: "never"
    dangerously_bypass_approvals_and_sandbox: false
    # Open network inside the workspace-write sandbox; without this, curl/git inside
    # the actor sandbox will hit DNS timeouts (codex defaults network_access=false).
    # Keep this convenience list for common workspace-write permissions. Advanced
    # settings may also use `extra_config_overrides` and are merged into CODEX_CONFIG.
    sandbox_permissions: ["network-full-access"]
    extra_config_overrides: []
  opencode:
    command: "opencode"
    args: [ "acp" ]
    # Leave empty to inherit your local opencode global config
    # (~/.config/opencode/opencode.json or opencode.jsonc)
    model: ""
    # Optional; only used together with a Hone-side model override
    variant: ""
    # Optional Hone-side provider/base URL override; leave empty to inherit local opencode config
    api_base_url: ""
    # Optional Hone-side API key override; leave empty to inherit local opencode auth/provider config
    api_key: ""
  earnings_workflow:
    # Administrator-only earnings preview/analysis route. This does not change
    # the runner or model used by ordinary conversations.
    runner: "opencode_acp"
    model: "google/gemini-3.1-pro-preview"
  hone_cloud:
    # Hosted service route seeded by this sample; runtime appends /api/public/v1/chat/completions when needed.
    base_url: "https://hone-claw.com"
    # Per-user Hone Cloud API key. Contact bm@hone-claw.com for an invite and key.
    api_key: ""
    # OpenAI-compatible model id placeholder used by the client request.
    model: "hone-cloud"
search:
  # Tavily is the only runtime search provider today. search.provider /
  # search.search_depth / search.topic stay in the config schema for compatibility
  # but are not wired into requests yet; the current web_search tool reads
  # search.api_keys and search.max_results only. Runtime caps max_results to 3
  # to keep Tavily response size and credits predictable for heartbeat workloads.
  provider: tavily
  # Tavily API key pool with failover support
  api_keys:
    - ""
  search_depth: "basic"
  topic: "general"
  max_results: 3