109 KiB

Raw Blame History

Stalwart JMAP ↔ Telegram Bot — Rust solution design

中文版本 / Chinese version → design.zh-CN.md

Status: Design and implementation have been implemented (The code can be found in the warehouses src/, web/, cloudflare-worker/; this article also retains the original records of design decisions and selection evaluations. The description of stages 1/2/4 belongs to the historical plan, and the current progress can be found in §10.5) Author: Cowork (team: MessageWeave) Date: 2026-09-21 Target readers: Codex CLI (lead), follow-up AI coding agent, end-user review

Related documents (separation of responsibilities to avoid duplication):

docs/charter.md — Project charter: project goals, technology selection, security invariants, implementation phases, prohibited matters, test acceptance and stable ID registry

AGENTS.md — Universal, language-independent code writing and environment building specifications (excluding project-specific content)

docs/deployment.md — Deployment, operation, maintenance and release: general HTTPS-only Docker container, Secrets, Redis (the only carrier of status), short request Webhook/Push/reconciliation routing, health check, CI release and items that still need to be confirmed

This document only retains content directly related to product behavior, code architecture, module interfaces, state machines, data flow, error handling, testing and implementation phases. Deployment operations and Docker/universal container platform details have been moved to deployment.md; project constraints and security invariants are in docs/charter.md; general, language-agnostic code writing and environment building specifications are in ../AGENTS.md.

[Task and Scope](#1-Task and Scope)
[jmap-client capability analysis](#2-jmap-client-capability analysis)
[Adaptation of authentication and email operations] (#3-Adaptation of authentication and email operations)
[Telegram channel implementation selection](#4-telegram-channel implementation selection)
[Overall architecture and data flow](#5-Overall architecture and data flow)
[Module Division](#6-Module Division)
[Configuration and Security](#7-Configuration and Security)
[Error Handling and Observability](#8-Error Handling and Observability)
[Test Strategy](#9-Test Strategy)
[Phase-based implementation plan](#10-Phase-based implementation plan)
[Historical issues and decision-making archives (product/architecture category, both have been concluded)] (#11-Historical issues and decision-making archives, product architecture category have been concluded)
[AI auxiliary capabilities: architecture, confirmation threshold, failure fallback] (#12-ai-auxiliary capability architecture confirmation threshold failure fallback)

Deployment/platform decisions confirmed (Single Account, App Password+Basic, Redis Hosting+AOF, Platform HTTPS URL, External Cron) and all converged archive (including Q-DEP-A/Q-DEP-B, see the Confirmed Decisions section of docs/deployment.md); Unfinished code gaps can be found in docs/opengaps.md.

1. Task and Scope

1.1 Goals

Build a Telegram robot written in Rust as a personal email assistant for Stalwart JMAP mailbox:

Query/read emails, view folders, send emails, manage keywords, etc. through Telegram commands.
Leverage JMAP's Push HTTPS callback (+ external Cron reconciliation) to proactively push to Telegram when new emails arrive. (EventSource/SSE/long polling is non-target, see deployment.md NG-POLLING-SSE/NG-LONG-POLLING/C-NO-LONG-CONN.)
Single-user or multi-account deployment (the default is for single-account self-hosting scenarios).

1.2 Scope (this stage)

✅ Research + Plan Design (this document)
❌ Do not create/modify project code or initialize cargo project
✅ Output detailed plans for lead and user review, marking key uncertain items

1.3 Research basis

stalwartlabs/jmap-client (main branch, as of this research)
Project directory: /home/okabe/Repo/messageweave/ (see §10.0 and ../AGENTS.md §3.3 for tool chain requirements)

2. jmap-client capability analysis

Source: Read directly from the stalwartlabs/jmap-client warehouse stalwartlabs/jmap-client/src/lib.rs, stalwartlabs/jmap-client/src/client.rs, stalwartlabs/jmap-client/src/email/, stalwartlabs/jmap-client/src/email_submissi on/, stalwartlabs/jmap-client/src/event_source/, stalwartlabs/jmap-client/Cargo.toml, stalwartlabs/jmap-client/README.md, stalwartlabs/jmap-client/examples/.

2.1 crate overview

item	value
crate name	`jmap-client` (crates.io)
Protocol coverage	JMAP Core (RFC 8620), Mail (RFC 8621), WebSocket (RFC 8887), Sieve (draft-12)
Asynchronous runtime	tokio + reqwest
License	Apache-2.0 OR MIT
`forbid(unsafe_code)`	Yes (stated at top of lib.rs)✅
Default features	0.4.2 Tested `default = ["async", "websockets", "aws_lc_rs"]` (including WebSocket stack `tokio-tungstenite`). ⚠️ `default-features = true` will implicitly enable WebSocket, violating `C-NO-LONG-CONN`. This project actually uses `default-features = false, features = ["async", "rustls"]` (without `websockets`), subject to `Cargo.toml`.

2.2 Module structure (`src/`)

lib.rs            顶层：URI / Method / DataType / PushObject / Error 枚举
client.rs         Client：认证、connect、session、send_request、event_source、ws
core/             request / response / get / set / query / query_changes / changes / error / session
email/            mod + helpers（email_get/email_query/email_import/email_copy/email_set/email_parse…）
email_submission/ helpers（email_submission_get/query/set → 发送邮件 + 状态查询）
mailbox/          helpers（mailbox_create/query/get/set/destroy…）
thread/           thread/get
identity/         Identity（发件身份）get/set
blob/             blob 上传/复制（附件）
sieve/            SieveScript（服务端过滤脚本）
vacation_response/
principal/        Principal（账户/共享）
push_subscription/ PushSubscription（HTTP push 回调注册）
event_source/     SSE 流：mod / parser / stream
client_ws/        WebSocket 客户端（feature = "websockets"）

2.3 Key capabilities mapped to this Bot

Bot requirements	jmap-client API	Remarks
Login/Session	`Client::new().credentials(...).connect(url)`	Supports Basic and Bearer; connect resolves session URL, capabilities
Column folder	`mailbox_query` / `mailbox_get`	With `role` to identify INBOX/Important/Draft, etc.
List emails	`email_query` (Filter + Comparator + anchor paging)	Filter: `subject`/`from`/`to`/`in_mailbox`/`has_keyword`/`after`/`before`…
Read email text	`email_get` + `Property` selection	`BodyStructure`/`BodyValues`/`Preview`/`TextBody`/`HtmlBody`
Read attachments	`Blob/get` (blobId) or `email_parse`	Large attachments need to be downloaded in fragments/streaming
Send email (2 steps)	① `email_set`/`email_import` create draft ② `email_submission_set` send	submission associate identityId
Delete/Archive	`email_set` (keywords `$seen`/`$flagged`), `mailbox_destroy`	JMAP has no real "delete", relying on keyword/move
Search	`email_query` (`Filter::text`) + `SearchSnippet/get` highlighting	`Filter` is a serde single-label enumeration, no comparator syntax; `SearchSnippet/get` only returns `emailId`/`subject`/`preview`, no `bodyProperties`/`parts`, body-level highlighting in locked version `0.4.2` can't do it (downgraded to a pure list of IDs)
Real-time notification (Push + reconciliation)	Push HTTPS callback → `StateChange`; external Cron calls `/reconcile` and uses `Email/changes` to make up the difference	Requires public HTTPS entrance (deployment.md `C-HTTPS-INBOUND`/`FLOW-NEW-MAIL`); Push is not the only reliable source
SSE / WebSocket (non-target)	`event_source` / `client_ws`	This deployment does not use (deployment.md `NG-POLLING-SSE`/`NG-LONG-POLLING`/`C-NO-LONG-CONN`); only crate capabilities are listed for investigation

2.4 Authentication mechanism (`stalwartlabs/jmap-client/src/client.rs`)

Credentials::Basic { username, secret } — username/password, natively supported by Stalwart.
Credentials::Bearer { token, .. } — OAuth2 access_token; optional refresh_token + refresh_url + refresh_grace, the client will automatically refresh before expiration.
connect() Function: GET session URL → Parse accounts/capabilities/download_url/upload_url/event_source_url, cache account_id.
Support custom reqwest::Client (Client::new().client(reqwest_client)): can inject proxy, TLS configuration, timeout, UA.
Support accept_long_responses / event_source(ping, ..) for long connection keep-alive.

2.5 Error model (`Error` enumeration)

Transport(reqwest::Error)   网络/TLS/超时
Parse(serde_json::Error)    序列化
Internal(String)            客户端内部
Problem(Box<ProblemDetails>) JMAP problem-details
Server(String)
Method(MethodError)         JMAP method-level 错误（NotJSON/Forbidden/RateLimit/StateMismatch…）
Set(SetError<String>)       /set 级别错误（每条记录的 creation/update/destroy 失败）
WebSocket(...)              ws 错误（feature 开启时）

MethodErrorType fine-grained: ServerUnavailable/ServerFail/RateLimit/InvalidArguments/Forbidden/StateMismatch/TooManyChanges… → Can directly drive Bot’s retry/current limiting/state reset strategy.

2.6 Real-time channel decision-making (determined, see deployment.md)

Deployment target = Generic HTTPS-only Docker, no long connections (deployment.md C-NO-LONG-CONN/C-HTTPS-INBOUND). Current channel = JMAP Push HTTPS callback (first explicitly registered by the administrator via POST /api/push/register, see §7.3) + external Cron /reconcile reconciliation. Push subscriptions are not created automatically, and Push cannot be regarded as the only reliable source. EventSource/SSE and WebSocket are both marked as non-target (NG-POLLING-SSE). Only the following table is retained for reference for crate capability research.

Channel	Whether to use	Remarks
Push Subscription (HTTP callback)	Accessed, explicit registration required	Short request model, suitable for stateless + no long connection; `POST /api/push/register` creates subscription, `/push/jmap` automatically completes Stalwart verification writeback
EventSource/SSE	❌ non-target	long connection, conflicts with `C-NO-LONG-CONN` (`NG-POLLING-SSE`)
WebSocket (RFC 8887)	❌ Non-target	Long connection, conflicts with `C-NO-LONG-CONN`; and needs server support

Reconciliation: External HTTPS Cron periodically adjusts /reconcile (deployment.md FLOW-RECONCILE), uses Email/changes + Redis sinceState to make up for the difference, and also serves as cursor reconstruction after Redis is lost.

3. Authentication and email operation adaptation

3.1 Authentication Strategy (Confirmed: App Password + Basic)

Confirm adoption (C-AUTH-APP-BASIC): Credentials::Basic (Stalwart account email + App Password). App Password can be revoked/expired independently without using a master password. Configuration injection, no code required.
Not used: OAuth/OIDC Bearer (increases complexity when there is no OIDC requirement, do not select); master password Basic.
Security: The password is only saved in the configuration or key manager; wrapped with secrecy::SecretString during runtime, the log will never print plain text (see §7.2).
For runtime injection methods, see deployment.md (C-NO-SECRET-IN-IMAGE).
Single Account (REQ-SINGLE-ACCOUNT): This instance only connects to one Stalwart account; multiple accounts = deploy multiple bot instances.

3.2 Mailbox operation adaptation layer (JMAP ↔ Bot semantics)

It is recommended to encapsulate a layer of domain semantics in the domain::jmap module and only expose business verbs to the upper layer:

Bot verb	Wrapper function	Internal JMAP
`list_folders()`	→ `Vec<Folder>`	`mailbox_query` + `mailbox_get` (cache role→id)
`list_emails(folder, page)`	→ `Vec<EmailMetadata>`	`email_query`(anchor paging) + `email_get` does not pass `properties` (take the server's default full value, including `From`; the list page must also display the sender)
`read_email(id, want_body)`	→ `EmailBody { text, html, attachments }`	`email_get`([BodyStructure, BodyValues, BlobIds]); Multiple part original text (`REQ-JMAP-RAW-MULTIPART`): Filter "parts with `part_id` and `bodyValue`" in order of `text_body` The last splicing is `text`; if there is no available part → return clear error (do not silently return an empty string). Attachments use `Blob/get`
`send_email(to, subject, body, attachments)`	→ `EmailId`	① `email_import`/`email_set` Create draft ② `email_submission_set`(onSend)
`set_flag(id, keyword)`	→ ()	`email_set` keywords
`search_emails(account_id, query, limit)`	→ `Result<Vec<SearchResult>, JmapError>`	`email_query` (`Filter::text`, `limit` capped at 100) gets the ID + `SearchSnippet/get` gets the highlight; when the latter does not support (`unknownMethod`/timeout) downgrade returns empty snippets Instead of reporting an error, the caller renders a pure ID list

Design points: anchor+position paging of email_query is JMAP standard practice and is more stable than traditional offset; sinceState + changes is used for incremental synchronization to avoid repeated pulling of the entire volume.

4. Telegram channel implementation selection

Telegram channel is self-developed and implemented using reqwest in src/channel.rs (ARCH-DEPS-STAGE4), without introducing third-party Bot framework. Evaluated teloxide but not adopted: This project only uses a few calls of the Bot API, and it is not worth introducing the two layers of abstraction of dptree scheduling and built-in session middleware; throttle and Redis session capabilities are partially implemented using src/channel.rs retries and external Redis TTL. Candidate comparison (teloxide / grammers / old telegram-bot), 6 reasons why we originally preferred teloxide, feature set plan and teloxide-core downgrade plan, all recorded in docs/retired.md.

5. Overall architecture and data flow

5.1 High-level topology

            ┌───────────────────────────────────────────────────┐
            │                    Telegram                       │
            └───────────────────────┬───────────────────────────┘
                     Webhook(HTTPS) │
            ┌───────────────────────▼───────────────────────────┐
            │   HTTP 入口 (axum，单端口 PORT)   │
            │  POST /webhook/tg  ·  POST /push/jmap  ·          │
            │  POST /reconcile   ·  GET /healthz /ready          │
            │  GET / + admin session + config APIs (管理 SPA)     │
            └───────────────────────┬───────────────────────────┘
                                命令 │            通知 ↓(Redis Streams worker)
            ┌───────────────────────▼───────────────────────────┐
            │            bot 层 (handlers)                     │
            │  /list /read /send /search /folders /flag ...     │
            └───────────────────────┬───────────────────────────┘
                                领域 │
            ┌───────────────────────▼───────────────────────────┐
            │        jmap 适配层 (JmapService)                   │
            │  list_folders/list_emails/read_email/send_email…  │
            │  + sinceState 缓存 → 外部 Redis (MOD-SINCESTATE)  │
            └───────────────────────┬───────────────────────────┘
                                JMAP │ HTTPS（短请求）
            ┌───────────────────────▼───────────────────────────┐
            │     Stalwart JMAP Server                          │
            │  session_url / email/* / mailbox/* / submission/* │
            └───────────────────────▲───────────────────────────┘
                     Push 回调(HTTPS)│  + 外部 Cron 对账(FLOW-RECONCILE)
            ┌───────────────────────┴───────────────────────────┐
            │  push handler → Redis Streams(MOD-STREAMS)        │
            │  worker: Email/changes → 通知 → 发往 TG → 推进游标│
            └───────────────────────────────────────────────────┘

Deployment form = General HTTPS-only Docker, no long connection (deployment.md C-NO-LONG-CONN/C-HTTPS-INBOUND): Telegram is Webhook only; JMAP real-time is Push callback only; slow tasks are asynchronously stepped into Redis Streams; reconciliation is triggered by External HTTPS Cron. EventSource/SSE and long polling are non-targets (NG-POLLING-SSE/NG-LONG-POLLING). Channel decoupling (see §5.2): The domain layer interacts with the channel layer through domain Command/Notification, and no specific channel SDK type appears in the domain layer; DingTalk/Feishu only retains extension bits and does not implement them in advance.

5.2 Channel abstraction and multi-channel expansion strategy

Status quo: The first (and currently only) channel is Telegram. The goal is to retain the ability to expand DingTalk/Feishu in parallel, but not to achieve it ahead of time.
Principle: Decoupling the domain layer and channel layer - Mail/JMAP/AI/intention state machine ** must not rely on Telegram type**.
Abstract (not over-engineered): No empty traits are reserved. src/channel.rs had three zero implementations #[expect(dead_code)] Placeholder trait (Channel / Notifier / MessageAdapter, comment readme "Stable ID + Phase 0 placeholder"), Deleted (see docs/retired.md for registration) - Actual Telegram outbound channel::telegram::TelegramClient::send_text (self-developed by reqwest), by worker.rs MetadataWorker and notify.rs are held directly and are never passed through these three traits. The command is parsed in parse_intent (src/worker.rs), rendering logic is scattered in src/worker.rs / src/notify.rs. What is passed between the domain and the channel is the domain Notification (src/domain.rs), and the channel only adapts at the edge.
Constraints:
- The public interface of jmap / llm / intent state machine / notify::core only accepts/returns domain types;
- Specific channel SDK types must not appear in the domain layer (the channel layer does not currently introduce third-party Bot frameworks, see §4);
- Add new channel = new adapter + assembly, do not change the domain layer. When abstraction is really needed, add it with the implementation and leave no empty traits.
What not to do: Do not define a polymorphic configuration registry, do not abstract the "channel capability matrix" in advance, do not build a plugins mechanism; evolve on demand (YAGNI).

5.3 Running model (short request, no long connection)

Single binary message-weave, #[tokio::main].
On startup:
1. Load configuration (config::Config).
2. Construct JmapService (Client::connect completes session parsing, account_id caching, mailbox role→id mapping warm-up); sinceState is restored from external Redis (MOD-SINCESTATE).
3. Start the HTTP portal (axum, single port PORT): /webhook/tg, /push/jmap, /reconcile, /healthz, /ready; three of the write paths first pass the SAF-AUTH-* portal authentication (fail-closed, §7.3), /healthz, /ready For the public probe (SAF-PROBE-PUBLIC); /ready has done end-to-end detection (configuration + Redis + outbound read-only detection JMAP session GET and TG getMe, each 3s, parallel, about 3s at worst).
4. Start Redis Streams worker (background task) to consume Push events → Email/changes → Notification → Send to TG → Push sinceState → XACK.
5. Does not hold any long connections and does not build self-timers (C-NO-LONG-CONN); reconciliation is triggered by external HTTPS Cron /reconcile (FLOW-RECONCILE).
Notification sending and command processing share Arc<JmapService>, and internal tokio::sync::RwLock protects the variable cache; cross-request status always falls to external Redis (C-REDIS-ONLY-STATE).
Multiple instances and load balancing (deployment form, ARCH-LB-WORKER): Since the state is all external Redis (C-REDIS-ONLY-STATE) and the delivery is idempotent (MOD-DEDUP), the same image can be instantiated across multiple serverless platforms, free Cloudflare Worker is used as the only entrance and failover (C-LB-SINGLE-REG-URL); /reconcile Use Redis to lock a single instance (SAF-RECONCILE-LOCK), and Streams to automatically amortize using the same consumer group (MOD-STREAMS-GROUP). See deployment.md §10 for details. **Field/channel logic does not need to be changed. **
Production redline (C-NO-DB / C-NO-LOCAL-WRITE / C-LOG-STDOUT-ONLY / SAF-LOG-PURITY / C-NO-STATEFUL-RECOVERY, see deployment.md §0 / §8.2):
- Production uses no database (C-NO-DB): No SQLite/Postgres/MySQL/embedded database; Redis is the only production state store.
- Disable local file/directory writing (C-NO-LOCAL-WRITE): no log files, no data files, no temporary cache, no local volumes mounted.
- Log only writes to stdout/stderr (C-LOG-STDOUT-ONLY): The container/platform is responsible for collecting and placing disk; disable the file log backend.
- Log and Redis write content constraints (SAF-LOG-PURITY): Only structured events, counts, timestamps, desensitized summaries; Prohibited key original text, JMAP email body, AI request/response, attachment content.
- It is prohibited to rely on in-process status for production recovery (C-NO-STATEFUL-RECOVERY): Restart and continuation (deduplication, sinceState, Streams breakpoints, circuit breaker counts, sessions) are all implemented by external Redis + JMAP reconciliation; the in-process cache is only for performance optimization, and the loss must be safe and reentrant.
Graceful exit: signal + CancellationToken, flush Redis Streams pending entries before exiting (Do not write local files, C-NO-LOCAL-WRITE).
Current status: Configuration guidance, HTTP entry and write entry authentication completed (R1/SAF-AUTH-*, fail-closed); JMAP session, Email/changes, PushSubscription create/update, Redis Streams worker and /reconcile have been connected. /reconcile and the Push closed loop still need to complete end-to-end acceptance in the real Stalwart, Redis, and Telegram environments. You cannot claim that the production link has been verified based on local access control alone.

5.4 Data flow: new email push (critical path, FLOW-NEW-MAIL)

Stalwart JMAP Push → POST /push/jmap (StateChange{Email/EmailDelivery: new_state})
  → 校验 pushSubscriptionId + verificationCode → 幂等去重(MOD-DEDUP) → 入 Redis Streams(MOD-STREAMS) → 立即 2xx ACK
  → worker: XREADGROUP → 用 sinceState(存 Redis, MOD-SINCESTATE) 调 Email/changes → 取 created[] 的 id
  → read_email 封装 email_get（Subject, Preview, From, TextBody, BodyValues, Size, ReceivedAt, HasAttachment）→ EmailMetadata
  → notify::core 组装领域 Notification（仅元数据 + 行内按钮意图，绝不含正文）
  → channel::telegram::TelegramClient::send_notification(chat, notification)（内部渲染文本后经 send_text 发出）
  → 更新 sinceState（写外部 Redis）→ XACK
  → 若 Redis 丢失：由外部 Cron 对账(FLOW-RECONCILE) 重建游标并补发

Key: Email/changes lists the new email id in created to avoid full email_query; sinceState is stored in external Redis (C-REDIS-ONLY-STATE). When Redis is lost, it will be reconstructed from JMAP by reconciliation, and the fact source is in JMAP. Notifications only contain sender/subject/time (+number of attachments), the text never appears in the notification (see docs/charter.md §3 Security Boundary SAF-NOTIFY-META). No long connection: Push is a short request callback, reconciliation is triggered by external HTTPS Cron (C-NO-LONG-CONN), and is not a self-sustaining timer in the container.

5.5 Data flow: command `/read <seq>` (long mail branch)

TG /read 3 → handler 取会话里的 folder+page 游标
  → JmapService.list_emails(folder, page=3)
  → 取第 3 封 id → read_email(id, body=true)   # JMAP 直取，绝不经 LLM
  → 格式化：
      正文 ≤ long_email_char_limit(默认 4000) → 正常发送（HTML escape）
      正文 > 阈值 → 不发全文，发 Preview(120) + 发件人/主题/附件概览 + 选项：
        "[AI 总结](约300字，需确认) / [继续查看原文(截断，标注‘完整请电脑查看’)]"
  → reply
（"AI 总结"分支必须用户确认后才把正文交 LLM 生成 ~300 字摘要；见 §12.3）

5.6 Session State Machine (Conversation FSM)

There is currently no FSM. The command route is parsed into Intent by parse_intent in src/worker.rs; the AI authorization state is a Boolean expiration time in Redis (see the AI authorization state section of docs/reference.md for the key and TTL), not a multi-step confirmation state.

Invariants that still hold true: AI digests must first be explicitly authorized by the user, and LLM must not be called without authorization. Once the authorization takes effect, the digest will pull the full text of the email body and send it to LLM (read_email requests TextBody + BodyValues and fetch_text_body_values(true), see src/domain/jmap/client.rs), but the body will not be sent back to Telegram** - only the summary will be sent outbound, or it will fall back to the first 300 if it fails. characters (worker.rs:227 Note: body text never reaches Telegram).

The designed 5-state FSM (Idle / AwaitClarify / AwaitConfirm / Analyzing / AwaitFallback), together with the state transition table and channel neutrality description, can be found in docs/retired.md.

5.7 Email search (`/search`, `bfe0fd8` implemented)

/search <keyword> and the Chinese prefix search/find/retrieval (only prefix matching, to avoid "Help me search..." being hijacked into a search) triggers Intent::Search, through JmapService::search_emails(account_id, query, limit) and email_query(Filter::text) to get the ID + SearchSnippet/get Highlight subject/preview.

Key boundaries:

Highlight downgrade: SearchSnippet/get does not support (unknownMethod) or returns empty snippets instead of an error when it times out; rendering is downgraded to a pure ID list (highlighted snippets are temporarily unavailable, the following are matching email IDs), and snippets are not created.
Plain text outbound: <mark> highlighted tags are stripped by strip_mark_tags (case-insensitive, unclosed tags are discarded), and then decoded in a single pass by unescape_html_entities; tags are stripped first and then entities are decoded to prevent the literal <mark> (escaped by the server as <mark>) in the mailbox body from being restored to a real tag and accidentally deleted. subject/preview have 120/160 character truncation upper limit respectively (truncate_chars, the same tool as text truncation).
Failed mapping: JMAP side failed → SearchReply::Retry → process_telegram returns Err(()) → coordinator 503 + Retry-After retry (consistent with /reconcile, other JMAP paths), no panic, no error stack returned to user.
Empty query: /search (no keywords) → The guidance prompt "Please provide search keywords, for example: /search invoice", no request is sent. No match → No messages matching "..." were found. .
Each hit contains email_id, you can directly access /summary <email_id> to enter the AI summary process.
Body-level highlighting is not possible in locked versions: jmap-client 0.4.2’s SearchSnippet only models emailId/subject/preview, without bodyProperties/parts. According to RFC 8621 §5, the body-level body: String[Id] is not modeled by this crate and (none) deny_unknown_fields) are silently discarded by serde. The text highlighting of each part needs to bypass the crate and be sent directly to the original JMAP, which is not currently implemented.
No new environment variable; /search will give a friendly error when the configuration is missing.

6. Module division

Actual project structure (all code anchors in this document refer to this structure; see docs/retired.md for the historical target module splitting plan):

message-weave/
├── Cargo.toml                    # jmap-client 0.4.2 / redis 0.27 / reqwest 0.13；不含 teloxide
├── .env.example                  # 仅 2 个必填项的占位样例（REDIS_URL / CONFIG_ENCRYPTION_KEY），不含业务配置
├── .gitignore
├── Dockerfile                    # 仅本地开发用；生产不执行（C-DEBIAN-SLIM，边界见 deployment.md §3）
├── hoststack.yaml                # 生产部署真源（runtime/build/start/healthCheck），非网关清单，见 §5.1
├── src/
│   ├── main.rs                   # tokio main：读 PORT/REDIS_URL/CONFIG_ENCRYPTION_KEY 后启动 webhook HTTP 入口（单入口，无 CLI 子命令）
│   ├── config.rs                 # Redis 业务配置反序列化（serde → BusinessConfigWire → Config）；CONFIG_ENCRYPTION_KEY 解析
│   ├── error.rs                  # BotError 统一错误（见 §8.1）
│   ├── domain.rs                 # 领域层根（渠道中立，见 §5.2）
│   ├── domain/jmap.rs            # JmapBackend trait + MockBackend（契约测试不联网）
│   ├── domain/jmap/client.rs     # 包装 jmap_client::Client（真实只读 adapter，MOD-JMAP-CLIENT）
│   ├── state.rs                  # Redis 读写封装：配置/开关/会话/TTL 键（C-REDIS-ONLY-STATE）
│   ├── channel.rs                # 渠道层：TelegramClient（reqwest 自研出站）；领域类型在 src/domain.rs
│   ├── notify.rs                 # /push/jmap 校验入队 + Redis Streams worker + /reconcile 对账
│   ├── worker.rs                 # parse_intent → Intent，命令路由与 AI 授权判定
│   ├── ai.rs                     # LlmClient / summarize：LLM 摘要调用
│   └── web.rs                    # /config 静态前端路由 + include_str! 嵌入 + CSP
├── web/                          # 前端资源，构建期 include_str! 嵌入二进制（C-NO-LOCAL-WRITE）
│   ├── index.html
│   ├── styles.css
│   ├── config.js
│   └── config.test.mjs           # Node 原生断言测试（非 cargo 测试）
└── cloudflare-worker/            # 边缘反向代理，JS，不在 Rust 工程中

Testing: No integrated test directory, and no stand-alone util tool module. All Rust tests are #[cfg(test)] unit tests within each module; the only test on the front end is web/config.test.mjs.

6.1 Module Responsibility Matrix

Current dependencies: jmap-client 0.4.2, redis 0.27, reqwest 0.13 have been introduced (ARCH-DEPS-STAGE1 / ARCH-DEPS-STAGE4, the version is subject to Cargo.toml). teloxide has been evaluated to be not introduced, and the Telegram channel is self-developed using reqwest (§4, see docs/retired.md for details). Each row in the table below is a real module.

Modules	Dependencies	Output	Testability	Current Status
`config`	secrecy; ring (`session_digest` SHA-256)	`Config` structure (7 fields, all with real readers)	Pure function, easy to test	Implemented (Redis business configuration serde deserialization + `validate_nonblank` fail-closed; startup env is read directly by `main.rs`, no figment/TOML)
`error`	—	`BotError` (§8.1, 4 variants)	Single test	Implemented
`domain` + `domain::jmap::client`	jmap-client 0.4.2	JMAP read-only semantics; `client` = real read-only adapter (`MOD-JMAP-CLIENT`)	mock JMAP response + `#[ignore]` real machine test	Code has been implemented, waiting for real Stalwart end-to-end verification (`cargo test -- --ignored jmap::`)
`state`	redis 0.27	Redis read and write: config/switch/session/TTL key (`C-REDIS-ONLY-STATE`)	Redis mock	implemented
`channel`	reqwest 0.13; jmap-client	`TelegramClient` (`send_text` / `send_notification`), dual channels of email and TG (event→field `Notification`→rendering)	mock HTTP	Implemented (`src/channel.rs`, self-developed by `reqwest`, no third-party Bot framework)
`worker`	—	`parse_intent` → `Intent` (6 types, including `Search(query)`: `/search` + `SearchSnippet/get` highlight rendering), command routing and AI authorization determination	Table-driven pure single test	Implemented
`ai`	reqwest 0.13	`LlmClient` / `summarize`	mock OpenAI compatible endpoint	implemented
`notify`	axum; redis	HTTP authentication, global switch, `/push/jmap` enqueue, `Email/changes` reconciliation and cursor submission	webhook verification failure path coverage 403	The real reconciliation path has been covered; the real Stalwart environment is still required for end-to-end acceptance
`web`	axum	`/config` static page + `include_str!` embedding + CSP	Front-end covered by `web/config.test.mjs`	Implemented
`debug`	reqwest 0.13; redis 0.27	Remote diagnostic surface `/debug/*` (`src/debug.rs`): 6 read-only probes `/debug/ping`, `/debug/config`, `/debug/redis`, `/debug/jmap`, `/debug/telegram`, `/debug/worker` + `POST /debug/notify` (sends a test message through the production outbound path, no independent implementation)	Only `DEBUG_ENABLED` (or `--debug`) + `DEBUG_TOKEN` mounts the route when both factors are present; the module itself is compiled unconditionally	Implemented, does not enter the Worker whitelist (`SAF-DEBUG-ORIGIN-ONLY`)

Note: The "Current Status" column is module level (the module has been implemented), which does not mean that the behavior is complete. Behavioral gaps are not included in this list: /search, Telegram 429 backoff, and multi-instance re-delivery windows have all been implemented. SAF-DEBUG-ALLOWLIST (POST /debug/notify does not intercept any chat_id when the list is empty) has never been counted as a behavioral gap and was moved from docs/opengaps.md on 2026-09-29: its exposure in production deployments is provided by SAF-DEBUG-GATE two-factor mount with Worker routing safelist Hold on, the two doors are sufficient on their own, and there is no need for a third layer of lists, so there is no gap. This behavior and residual risks (direct connection to the backend origin can bypass both layers) are described in docs/deployment.md §2.1 and docs/design.md §7.6. The /search path of the worker module has been implemented with bfe0fd8.

7. Configuration and Security

7.1 Historical configuration reading (migrated to Redis)

Migration Target (C-REDIS-ONLY-STATE): Only REDIS_URL is retained for the production boot environment. Empty Redis First-time configuration must be done through a one-time bootstrap/admin session authenticated by CONFIG_ENCRYPTION_KEY; Unauthorized write access must not be provided. This key is only used for constant time comparison and does not echo, record, or write business configuration; The Redis ACL password is only responsible for the Redis connection itself, not any HTTP authentication credentials. After bootstrap is successful The administrator session hash and TTL are saved in Redis and can be restored after restarting; the business token is loaded from Redis at startup. Key GET will never be echoed; after bootstrap/administrator PUT is successful, the client will be built and replaced atomically, and subsequent requests will be immediate. Use new configuration, keep old instance on failure. This migration replaces the following stage 0 environment variable list. The stage 0 list is only for historical purposes. Compatibility instructions.

Actual implementation (ARCH-CONFIG-ENV): process startup read-only 2 process-level credentials plus listening port (src/main.rs:45 / :54 / :63, Add optional DEBUG_TOKEN in :107), directly std::env::var, do not use figment, do not use TOML/configuration file. Business configuration is no longer read from environment variables: The fields in the following table are all loaded by the config layer using serde deserialization through the Redis business configuration. And do fail-closed verification through validate_business_wire / validate_nonblank. Historically existing Config::from_env() full env parsing path removed (see docs/retired.md for registration), Therefore the current code reads zero variable names in the following table; the RUN_MODE variable itself has also been deleted entirely (the identifier is no longer in the code), webhook and reconcile already share the same set of routing tables, NG-SERVER-MODE is retained as a design non-target in docs/charter.md §8.

⚠️ The following table is the complete variable table of business configuration fields (the variable names are named after the original environment variables to facilitate comparison with the deployment document, but there is no longer any reading method in the code**). The current production deployment only requires REDIS_URL + CONFIG_ENCRYPTION_KEY two startup variables (src/main.rs); all other business fields are written by PUT /api/business-config through Redis business configuration (C-REDIS-ONLY-STATE). "Required" in the table refers to the fail-closed constraint (validate_nonblank, blank means rejection) of Redis business configuration, which is no longer required for the environment variable boot path.

Variable	Required	Default	Description
`PORT`	No	`8080`	Single listening port (`C-NO-TCP-EXPOSE`)
`BOT_TOKEN`	Yes	—	Telegram Bot Token(`SecretString`)
`TG_WEBHOOK_SECRET`	Yes	—	`/webhook/tg` Authentication: Request header `X-Telegram-Bot-Api-Secret-Token` (`SAF-AUTH-TG-WEBHOOK`; `SecretString`)
`CHAT_ALLOWLIST`	Yes	—	Comma separated integer chat ids; Whitelist Hard Constraint (`SAF-CHAT-ALLOWLIST`)
`JMAP_SESSION_URL`	Yes	—	Stalwart JMAP session URL (`REQ-JMAP-SESSION-URL`): You can fill in the service base address `https://host[:port]` or complete `…/.well-known/jmap`; the code is normalized to origin/base and then submitted to `jmap-client` (it is automatically appended `.well-known/jmap`), paths must not be repeated. HTTPS only, URL embedded credentials prohibited (`SAF-JMAP-URL`)
`JMAP_USERNAME`	Yes	—	Stalwart Account (Email)
`JMAP_PASSWORD`	Yes	—	App Password(`C-AUTH-APP-BASIC`; `SecretString`)
Push verification	No	—	Dynamically generated by Stalwart; the backend automatically writes back through `PushSubscription/set` and saves the short-term verification status in Redis
`REDIS_URL`	Yes	—	External Redis (User-Managed + AOF, `C-REDIS-MANAGED-AOF`; `SecretString`)
`RECONCILE_TOKEN`	Yes	—	`/reconcile` Authentication: `Authorization: Bearer` (`SAF-AUTH-RECONCILE`; `SecretString`). `/reconcile` route is always mounted, so required
`ACCOUNT_ID`	No	Empty	Single account (`REQ-SINGLE-ACCOUNT`); if left blank, the session main account will be used
`LLM_ENABLED`	No	`false`	Do not verify `LLM_*` on shutdown (`REQ-LLM-OPENAI-COMPAT`)
`LLM_ALLOW_NET`	No	`false`	AI outgoing network switch (`REQ-AI-EXTERNAL-CONSENT`)
`LLM_API_KEY` / `LLM_BASE_URL` / `LLM_MODEL`	Required when `LLM_ENABLED=true`	—	OpenAI-compatible
`LLM_MAX_RETRIES`	No	—	Fuse threshold (`REQ-AI-FUSE`). Non environment variables: managed by Redis running parameters (`OutboundConfig.max_retries`, fall back to default, see `docs/reference.md` §6.1)
`LLM_SUMMARY_TARGET_CHARS`	No (constant)	`300`	Summary target word count (`REQ-LONG-EMAIL`). Non configurable items: The `config` layer is hard-coded to 300 in the two constructors and does not enter the business configuration wire

Currently config.rs provides the above fields (including TelegramConfig / JmapConfig / LlmConfig) from Redis business configuration; process-level startup variables are read directly by src/main.rs without going through the config layer. The introduction of TOML/figment in the future will require separate decisions; the current documentation does not assume that the configuration file exists. The entrance authentication key (TG_WEBHOOK_SECRET/RECONCILE_TOKEN) must be valid after the business configuration is completed (fail-closed, SAF-AUTH-*). The verification code of JMAP Push is dynamically generated by Stalwart after the subscription is created and does not belong to the business configuration. When the startup boot variables are missing, the service enters configuration boot mode and remains accessible to SPA; business success will not be faked due to lack of startup variables, nor will entrance authentication be bypassed. JMAP session URL normalization (REQ-JMAP-SESSION-URL/SAF-JMAP-URL): JMAP_SESSION_URL accepts both the service base address (https://mail.example.com) and the full session URL (https://mail.example.com/.well-known/jmap). The code is unified and normalized to origin/base before calling jmap_client::Client::connect **——jmap-client will automatically append /.well-known/jmap, so there will be no duplicate paths (such as …/.well-known/jmap/.well-known/jmap). Constraints: HTTPS only (http denied); Disallow URL embedded username/password (credentials only via JMAP_USERNAME/JMAP_PASSWORD, C-AUTH-APP-BASIC); deny dangerous queries. Redirect whitelist: jmap-client denies all redirections by default, and Stalwart will jump /.well-known/jmap 307 to /jmap/session, so the origin host itself in the configuration is added to the trust list when connecting, and other hosts rewritten by the server are still rejected (follow_redirects([trusted_redirect_host(base)])). Known upstream risks: Stalwart's 307 Location will write user:pass into the URL, and the client will return the URL with credentials - so the transmission error string will not be logged additionally. (Status: D-G1-1 Code implemented and verified with cargo test -- --ignored jmap:: with a real Stalwart instance.)

7.2 Key Management

It is prohibited to write token/password into the warehouse or any configuration file (this project configuration does not enter the warehouse, ARCH-CONFIG-ENV).
Only 2 process-level variables are read during startup (REDIS_URL, CONFIG_ENCRYPTION_KEY, directly std::env::var from src/main.rs; if any one is missing, it will be downgraded to read-only setup mode); the business configuration is deserialized from Redis through the config layer (None ${VAR} interpolation/figment/config files, ARCH-CONFIG-ENV).
Use secrecy::SecretString to wrap it during runtime, and Debug to implement ***.
Log filtering: tracing field layer shielding Authorization/password/token.
For the runtime secret injection method (env / *_FILE / orchestrator secret), see deployment.md (C-NO-SECRET-IN-IMAGE).

7.3 Access control

Entry authentication (hard constraints SAF-AUTH-RECONCILE/SAF-AUTH-TG-WEBHOOK/SAF-AUTH-JMAP-PUSH): The three write paths must be authenticated first, fail-closed——
- /reconcile: Request header Authorization: Bearer <RECONCILE_TOKEN>;
- /webhook/tg: Request header X-Telegram-Bot-Api-Secret-Token == TG_WEBHOOK_SECRET;
- /push/jmap: Verify Redis short-term status according to pushSubscriptionId; StateChange is released only after the first verification is successfully written back by JMAP PushSubscription/set. The comparison uses the constant time algorithm (subtle, anti-timing side channel); verification failure will always be 401 and will not produce any side effects/status changes before the authentication is passed. After the business configuration is completed, the credentials for Webhook, Push, Reconcile and Push registration interfaces must be valid; when the startup boot variable is missing, the configuration boot mode will be entered and authentication will not be bypassed (§7.1).
Health probe (SAF-PROBE-PUBLIC): /healthz (ARCH-HEALTHZ) and /ready are public probes - no authentication, only returns health status, does not contain any sensitive information (does not echo configuration/key/internal error details).
- /healthz = liveness (process survival), long-term stable semantics.
- /ready does end-to-end probing: configuration integrity + Redis reachability + outbound read-only probing (GET {jmap_origin}/.well-known/jmap, Basic authentication with configuration; GET https://api.telegram.org/bot<token>/getMe; each PROBE_TIMEOUT = 3000ms, Parallel (tokio::join!), worst-case scenario 3s); all four pass 200 and readiness report ({"status":"ready","configured":...,"jmap":...,"telegram":...}, where jmap/telegram is the real probe result), any one fails 503 (standard error envelope {"error":"service_unavailable","request_id":<id>} + Retry-After: 30). The probe is read-only, read-only configuration status, and will not trigger business side effects such as email synchronization, nor will it echo token or third-party response content; refresh_business_config only reads Redis, and has no state to write. Therefore /ready requires the outbound egress to the JMAP host and api.telegram.org:443 to be reachable (if the egress requires a proxy, /ready is not available, see deployment.md).
chat whitelist (hard constraint SAF-CHAT-ALLOWLIST): CHAT_ALLOWLIST is a required configuration; any inbound event (action triggered by TG command/callback) must verify chat.id ∈ CHAT_ALLOWLIST before making any JMAP call, AI call or status change. If it is not in the whitelist, directly reject and terminate (prevent token can be called by anyone after being leaked). Phase 0 has completed the CHAT_ALLOWLIST parsing skeleton; the forced rejection logic has been implemented with Telegram channel access (telegram_webhook of src/notify.rs verifies the whitelist before any JMAP/AI/status operation, and terminates if rejected).
Command Minimization: Only necessary commands are exposed; writing operations such as sending emails must be confirmed twice (sending emails is not currently implemented, see §10.3).
Rate: No local token bucket is built on the outbound side; Telegram's outbound sending is retried according to the Redis operating parameter max_retries (see docs/reference.md §6.1 for default values and hard upper limits); currently only Redis current limiting is done for Push verification code writing (ratelimit:push-verify:*). 429 under the 30 msg/s limit of the Telegram server Automatic backoff in parameters.retry_after seconds (channel.rs, f4cae00): retry_after_ms is truncated to the 60s budget limit after parsing, and the exponential backoff is used when this field is missing or non-numeric backoff_delay_ms (starting from 250ms, capped 4s), the overall retry budget is 60s. There is still no local token bucket current limit - if the budget is exceeded, a failure will be returned directly and the upstream will try again.

7.3.1 Configuration Management API

GET / provides a SPA embedded in the Rust binary; /assets/config.js and /assets/styles.css provide page resources. The service does not read or write local files while running (C-NO-LOCAL-WRITE).
GET /api/status publicly returns { "ready": boolean, "mode": "configured" | "configuration-setup", "missing": string[], "version": string }, which only lists the missing environment variable names. When REDIS_URL or CONFIG_ENCRYPTION_KEY is missing, SPA only displays the configuration boot status and missing variables; the management session authorization area is displayed only after the service status is confirmed to be ready=true. version is the BUILD_VERSION (<git-sha or nogit>+<UTC build time>) baked by build.rs during compilation. The SPA no longer renders it, so the API is the only way to confirm a deployment landed.
POST /api/admin/session accepts Authorization: Bearer <CONFIG_ENCRYPTION_KEY> and returns { "session": "<opaque>", "expires_in": 1800 }; the admin session only saves the digest in Redis and expires after 1,800 seconds. POST /api/admin/session/revoke revokes the current session and returns 204 successfully.
The admin page only saves the session in JavaScript memory. Request settings credentials: omit, cache: no-store, without using cookies, localStorage or sessionStorage. The admin API accepts a valid admin session; the compatibility path also accepts WORKER_TOKEN. Worker transparently transmits the authentication header (SAF-LB-PASSTHRU) as it is.

GET /api/config and PUT /api/config only read and write non-sensitive running parameters:

{
  "jmap_timeout_ms": 15000,
  "telegram_timeout_ms": 10000,
  "llm_timeout_ms": 30000,
  "max_retries": 3
}

The three timeout units are all milliseconds; the default value, value range and hard upper limit of max_retries are only authoritative with docs/reference.md §6.1 (the table below and the example response body above are only for illustration, and the values are not repeated). max_retries represents the maximum number of retries beyond the first request. When there is no configuration in Redis, GET returns the default value object; PUT successfully returns the same object after saving and persists to Redis (C-REDIS-ONLY-STATE).
GET /api/business-config reads back saved configuration for management SPA pre-populated form: returns { "configured", "revision", "values", "secrets_present" }, where values is the value of 10 non-key fields and secrets_present is 6 Boolean existence of a key (bot_token, jmap_password, telegram_webhook_secret, reconcile_token, worker_token, llm_api_key) - Never return the key value itself (SAF-NO-SECRET-ECHO). Not yet saved, staleness returns 200 + configured: false, revision: 0, empty values and an existence boolean of all false. SPA therefore has only one code path and does not need to distinguish between first-time configuration and modified configuration.
PUT /api/business-config accepts BusinessConfigPatch (complete or partial), only replaces the fields that appear in the submission body, and the rest inherits the stored values (BusinessConfigPatch::apply); Push verification does not belong to Wire, is dynamically generated by Stalwart and written back by the backend. Verification is the only write gate, and the verification object is the merged configuration rather than the commit itself: patches are never verified individually, and existing complete wires are never discarded. validate_business_wire runs first, and returns the real 422 when rejected; if the merge result is legal, the configuration must be persisted (the merged Wire is persisted, not the request body), and then the client is built and the worker is switched to run. Therefore, when the dependency is not connected, 200 + { "persisted": true, "runtime_applied": false, "warnings": [{"component", "step", "detail"}] } is returned instead of 503 - the configuration has been saved and the fault is reported truthfully instead of being silently swallowed by an unreachable JMAP. runtime_applied is only true when the reload submission is successful; in both cases, the x-business-config-revision header is returned; the request body can carry revision (that is, the u64 returned by GET) as a control field - it does not participate in apply, does not enter Wire, and is not used as a saved setting. It is only used to determine whether this submission is based on the latest version. See the next article for the semantics of 409. The response never echoes configuration or keys. The key is "replaced only after submission": the omitted key inherits the existing value, so the form backfilled by GET only needs to submit the fields to be changed; but the backend does not distinguish between "unsubmitted" and "submitted empty string", and explicitly submitting "" will actually clear the key - "leave blank and keep" is a client contract implemented by SPA by discarding blank key fields before submission, not a server-side guarantee. chat_allowlist must be a non-empty list, empty arrays will be rejected by validate_business_wire (422). Incremental semantics only hold after existing configuration: if there is no existing configuration, there is no fallback value, and the patch must come with all required fields, otherwise 422 will be returned (the entry created from scratch is POST /api/bootstrap).
POST /api/business-config/preflight performs exactly the same validation and client build on the submission body, returning component-by-component conclusions, without writing, without touching the running worker: { "persisted": false, "validation": { "ok", "errors" }, "components": { "jmap": {...}, "llm": ... } }. Different from PUT, its submission body must be complete BusinessConfigWire (no patch merging), so SPA must first merge the value read back by GET into the local form before pre-verification. components is null when the verification has failed (rejected Wire does not need to probe the client again); when llm_enabled/llm_allow_net is false, llm is null instead of false ok. Authentication is consistent with the management session gate (config_authorized) and is deliberately not hung after DEBUG_TOKEN.
POST /api/bootstrap also accepts complete BusinessConfigWire (without patch merging), and shares the above semantics of "separation of persistence and error reporting": verification failure is 422, and build failure after persistence is still 200 + warnings. Because it does not merge, it is the only patch path that can create business configurations from scratch (PUT returns 422 when the existing configuration is missing and the patch is incomplete, see the previous article).
The running parameter errors of PUT /api/config use 401 (Unauthorized), 400 (JSON Invalid), 422 (Scope Error), 503 (Redis Unavailable); when GET read fails, 503 also indicates that Redis is unavailable. Business PUT uses 401 (unauthorized), 400 (JSON is invalid), 422 (the patch cannot be parsed, the saved configuration cannot be decrypted, the merge result verification fails, and the patch is incomplete when saving for the first time), 409 (the submitted revision has expired), 500 (merge result serialization failed) and 503 (Redis is unreachable, writing failed or reload submission failed); business GET uses 401, 400, 422 (the saved configuration cannot be decrypted) and 503 (Redis is unreachable) - note that Invalid is 422, not 503, and Unreachable is 503. Business PUT has 409: revision is a control field in the request body** (not a header) - SPA records revision when reading the configuration, and each subsequent incremental submission will bring back the same value; the saved revision and the requested revision are not equal and return 409 conflict, neither writing nor incrementing the revision, SPA receives 409 The saved value will be read back again and allowed to be changed by the operator, so the old page will not silently overwrite the new save. The request body without revision maintains last-write-wins, which is compatible with callers that never read revisions. This comparison is compare first and then write, and is non-atomic: if two requests compete at the same moment, both will pass, and the later one will overwrite the first one. The cost of optimistic locking without Lua scripts is this residual window, and the patch only takes the fields actually changed by the operator, so the window impact is limited. The remaining occurrences of 409 conflict are the competition of three locks (reconcile and acquire_lock of register_push / register_worker_token return false), which have nothing to do with business configuration writing. Dependent client build failure does not return an error status code, but 200 + warnings (see above). The remaining errors do not echo sensitive data. The management page displays a 503 status when Redis/session initialization is not completed; running parameters are disabled from being saved before re-reading successfully (C-REDIS-ONLY-STATE). When LLM is enabled, the business configuration requires an HTTPS Base URL, a non-empty API key, and a model name.

7.4 TLS (code/dependent behavior)

The TLS provider is determined by the dependent feature: jmap-client 0.4.2 contains aws_lc_rs by default (and introduces rustls), and default-features = true is not "rustls default". The actual default-features = false, features = ["async","rustls"] of this project (aws_lc_rs/websockets is not enabled), and will ultimately be subject to Cargo.toml.
Certificate verification is mandatory by default and can be turned off only for testing.
For operation and maintenance details such as the system CA trust store in the image, internal CA injection, inbound webhook TLS termination, etc., see deployment.md.

7.5 Production red line: no database / no local writing / standard output log (deployment.md §0/§8.2)

This section summarizes the absolute red lines across code and deployment that you must not introduce any "seemingly convenient" local state into your code:

Redline ID	Meaning	Violation Example (Prohibited)
`C-NO-DB`	Production does not use any database; Redis is the only production state storage	Add SQLite deduplication table, add Postgres session
`C-NO-LOCAL-WRITE`	Disable writing of local files/directories; do not mount local volumes	Write logs to local files, write temporary directories, and place session cursors into local status files
`C-LOG-STDOUT-ONLY`	The log only writes stdout/stderr; the platform is responsible for collection	Introducing `tracing-appender`, `rolling-file`, `FileAppender`
`SAF-LOG-PURITY`	Log and Redis writing content is limited to structured events, counts, timestamps, desensitized summaries	Log printing JMAP text, AI prompt/completion, key original text, attachment content
`C-NO-STATEFUL-RECOVERY`	Production recovery does not rely on in-process state	Use `static Mutex<HashSet>` to store dedup and use memory LRU to store sinceState as the only recovery source

Verification: No local path constants / std::fs non-test calls / file log backends appear in the static inspection code; during runtime, the container /proc/mounts and docker inspect are used to confirm that there is no local volume mounting; see deployment.md §8.2 for the deployment checklist.

7.6 Why is the remote joint debugging interface absolutely closed by default?

The only reason for this surface to exist is to shorten the production troubleshooting path: when you cannot log in to the container and can only rely on external requests to observe the system, someone needs to be able to detect the JMAP/Telegram connectivity and current business configuration without releasing the version. The price is that its entry softness must be higher than the three write paths - in addition to read-only detection, it also retains a real outbound send, and reuses the business whitelist instead of an independent whitelist. Therefore, the design choice is "closed by default" instead of "enabled by default and blocked by the gateway": the process will only mount this set of routes (SAF-DEBUG-GATE) when "DEBUG_ENABLED is true or the command line has --debug"** and **a non-empty DEBUG_TOKEN is set; when the route is missing, the route does not exist at all in the router, and the request falls to axum general 404, not "Exists but 401" - the latter would leak the route's existence. There is no third state: there is no "open without configuration" fallback, and there is no configuration item to set it to be enabled by default. DEBUG_ENABLED uses environment variables instead of startup commands so that the switch can be changed once from the platform console without releasing the version - the startup command remains static. Add another layer of location constraints: it is not in the gateway's safe routing whitelist. Even if the backend is opened incorrectly, the platform entrance will be rejected by fail-closed. The only reachable path is to directly connect to the backend origin. These four things—two-factor mounting, 404 instead of 401, no default enabled fallback, and unreachable gateway—together form the architectural decision of “exposure is zero by default.” See deployment.md §2.1 for the enablement method, per-endpoint status code, and deployment confirmation checklist.

8. Error handling and observability

8.1 Unified error enumeration

Current implementation (full text of src/error.rs, 4 variants):

use thiserror::Error;

#[derive(Debug, Error)]
pub enum BotError {
    #[error("configuration error: {0}")]
    Config(String),
    #[error("I/O error: {0}")]
    Io(#[from] std::io::Error),
    #[error("JSON error: {0}")]
    Json(#[from] serde_json::Error),
    #[error("state error")]
    State(#[from] crate::state::StateError),
}

The design target form once included six additional variants: Jmap(#[from] jmap_client::Error), Telegram(#[from] teloxide::errors::RequestError), Storage(#[from] redis::RedisError), RateLimited, Unauthorized(i64), Llm(LlmErr); Telegram comes with teloxide They were abandoned without adoption, and the rest have not yet been implemented. For a complete list, see docs/retired.md.

8.2 Error → User Message Mapping

Low-level errors	Bot behavior
`Transport` / `ServerUnavailable` / `RateLimit`	Exponential backoff retry (`util::retry`), exceeding the threshold gives the user "temporarily unavailable, try again later"
`Forbidden` / `Unauthorized chat`	Silent rejection; log alert
`Method(InvalidArguments)`	Reply to "Wrong parameter + correct usage"
`Set(creation/update/destroy)`	Reply to the specific record-level failure reason
`StateMismatch`	Reset sinceState → Fill in full once (anti-lost/anti-repair)
`Llm(*)`	See §12.7 (Break/Fallback/Logo)

8.3 Observability

tracing (direct dependency) + tracing-subscriber (fmt + EnvFilter).
tracing (direct dependency) + tracing-subscriber (fmt + EnvFilter). Currently only recording startup events: src/main.rs 5 info!/warn! in total (missing startup variables, downgrade to setup mode, startup banner, debug endpoint enabled, JMAP service unavailable, downgrade × 2). Request-level events (Push callbacks, Reconcile pulls, channel pushes, LLM elapsed times) not implemented yet - there are no tracing calls in src/ except main.rs, and no named spans (no #[instrument] / span!).
Metrics (optional metrics crate): Number of Push callback arrivals, deduplication hit rate, Redis Streams backlog depth (pending), number of DLQ items, number of reconciliation items, JMAP request delay, push failure rate. Currently not connected to any metrics backend (metrics is not in Cargo.toml).
Graceful exit: Currently not implemented - There is no signal processing in src/ (no tokio::signal / ctrl_c), and there is no flush logic for Streams pending entries; the process is terminated when the container stops. Redis side C-NO-STATEFUL-RECOVERY ensures reconstruction from Redis after restart and does not rely on in-process state.
Reliability goals and strategies (Streams ACK/retry, idempotent deduplication, Push retry, reconciliation recovery, indicators/alarms, ≥99.9% notification availability and boundaries) see deployment.md §6.4/§6.5 (NFR-NOTIFY-SLA).

9. Test strategy

9.1 Levels

Layers	Location	Content
Unit	`#[cfg(test)] mod tests` in `src/*.rs`	`config` parsing, `worker` command routing, `ai` HTTPS verification, `state` Redis encapsulation, `channel` outbound, `notify` verification and deduplication, `main` startup branch
Contract	`src/domain/jmap.rs`	`JmapBackend` trait + `MockBackend`: validate domain verbs with preset data, not connected to the Internet
Integration	`src/channel.rs` test module	Use in-process `tokio::net::TcpListener` to start mock HTTP service, verify the path, status code and retry of outbound requests
Real (manual)	`src/domain/jmap/client.rs`	Marked with `#[ignore]`, you need to explicitly configure the JMAP test server before running `cargo test -- --ignored`

No integration test directory; the only test on the front end is web/config.test.mjs (Node native assertion).

9.2 mock strategy

JmapService holds the trait JmapBackend (list_emails / read_email / send_email / changes etc.), the production implementation wrapper jmap_client::Client (src/domain/jmap/client.rs), and the test MockBackend (src/domain/jmap.rs).
Telegram outbound does not introduce any framework mock: src/channel.rs starts an in-process TcpListener, uses AtomicUsize to record the number of requests, and verifies the URL, status code and timeout/retry behavior.
src/ai.rs asserts HTTPS enforcement (llm_requires_https) by constructing the configuration of LlmClient; summary logic validates on mock OpenAI compatible endpoint.
The decoupling of channels and domains relies on type constraints (channel SDK types do not appear in the public interface of domain.rs) and does not rely on mock channels.

9.3 Key invariant testing

sinceState is stored in external Redis: resumes the transfer from the correct cursor after simulating restart; restores the cursor from JMAP through reconciliation (reconcile, FLOW-RECONCILE) after simulating Redis clearing.
Notification deduplication: the same email_id will not be pushed repeatedly.
Text escaping: Message text containing <script> is escaped in HTML mode.
Security boundary assertion (hard, see docs/charter.md §3):
- VIEW/View original text path: mock AI endpoint zero request (LLM client not called);
- New email notification: There is no body content in the message (assertion leak);
- Long emails: if the text is > 4000 characters, the full text will not be sent;
- AI call prefix: LLM zero call before confirmation;
- The analysis results are not saved to disk: no new disk/Redis writing path is added after processing;
- AI fails 3 times → circuit breaker confirmation → fallback with "AI unavailable" logo.
- Entry Authentication (SAF-AUTH-*): /reconcile, /webhook/tg, /push/jmap returns 401 under missing or wrong credentials, and asserts no side effects (no Redis writes, no JMAP/AI calls) when authentication fails; correct credentials are allowed.
- Health Probe (SAF-PROBE-PUBLIC): /healthz returns 200 for process alive; /ready returns 200/503 for readiness - checks configuration integrity + Redis reachability + outbound read-only probe (GET {jmap_origin}/.well-known/jmap with Basic authentication, GET https://api.telegram.org/bot<token>/getMe, 3s each, parallel, about 3s at worst), any failure will return the standard error envelope (service_unavailable + Retry-After: 30), all will return the readiness report JSON; the response body does not contain sensitive information. Uptime Kuma is monitored by status code (/ready expects 200) and is not affected by changes in the response body.
- Channel decoupling: No channel SDK types appear in the public interfaces of domain modules (src/domain.rs / src/ai.rs / src/worker.rs); channel assembly is centralized in src/channel.rs.
- JMAP session URL (REQ-JMAP-SESSION-URL/SAF-JMAP-URL): both the base address and the complete …/.well-known/jmap are accepted and the normalized results are consistent**, the URL passed to Client::connect does not contain duplicate /.well-known/jmap**; http:// Rejected; Embedded credentials (https://user:pass@host) are rejected; Dangerous queries are rejected; Redirects only trust the origin host in the configuration (Stalwart will 307 jump to /jmap/session).
- JMAP multi-part original text (REQ-JMAP-RAW-MULTIPART): read_email splices the parts with part_id and bodyValue" for the multi-part text in the order of text_body; constructing the "no available part" use case assertion returns clear error (non-empty string).

10. Phased implementation plan

Prerequisite: rustup first installs the tool chain (stable). CI and release flows are in deployment.md.

10.0 Phase 0: Scaffolding and HTTPS entry skeleton (completed, `GATE-P0` passed)

Current Actual Boundaries: Single-port axum portal provides /webhook/tg, /push/jmap, /reconcile, /healthz, /ready. The entry authentication of the three write paths has been fail-closed (R1/SAF-AUTH-*); /reconcile has used JMAP Email/changes paging, Redis state:jmap:since and Redis solo flight lock, and the cursor will be advanced only after all events are successfully queued; /healthz is for liveness, /ready does end-to-end detection (configuration + Redis + Outbound read-only detection, about 3s at worst), but does not cover real message delivery acceptance.

Actual dependencies (Cargo.toml current, ARCH-DEPS-STAGE0/ARCH-DEPS-STAGE1 + ARCH-DEPS-STAGE4): axum 0.8 (single HTTPS entrance), async-trait, secrecy, subtle (constant time authentication comparison), serde, serde_json, thiserror, tokio, tracing, tracing-subscriber, url (JMAP_SESSION_URL normalized parsing), jmap-client =0.4.2 (default-features = false, features = ["async","rustls"]), redis 0.27 (Redis XPING/PING liveness detection, ARCH-STATE-REDIS), reqwest 0.13 (JMAP/TG HTTP client); dev-dependencies: tower 0.5 (routing test). teloxide is not introduced (ARCH-DEPS-STAGE4): Telegram channel is implemented in src/channel.rs using reqwest self-development.
Not yet introduced (Documents must not claim to have been used): teloxide and other Telegram Bot frameworks (Telegram outbound is sent directly by src/channel.rs and reqwest, and the evaluation record is in docs/retired.md). FIGMENT NOT USED: Configured for manual std::env parsing (ARCH-CONFIG-ENV, §7.1).
Stage 1 dependency status (ARCH-DEPS-STAGE1): jmap-client Introduced (currently =0.4.2, default-features = false, features = ["async","rustls"], version and features subject to Cargo.toml). ⚠️ Its default features ["async","websockets","aws_lc_rs"] includes WebSocket stack, so the default features must be turned off and websockets is not selected; JMAP only uses HTTPS short requests (Core/Mail) to comply with C-NO-LONG-CONN (no WS/SSE/long polling).
ACCOUNT_ID optional (REQ-SINGLE-ACCOUNT): leave blank → take the default/main account of the JMAP session; explicit values are used after verification; multiple accounts = multiple bot instances.
axum adopts 0.8 (ARCH-AXUM-08); if subsequent review decides to adjust the version, Cargo.toml will prevail and this section will be synchronized.
Currently SSE/WebSocket/long polling/SQLite/local volumes (C-NO-LONG-CONN/NG-SQLITE-PERSIST/NG-LOCAL-VOLUME) are not implemented; entrance authentication (R1/SAF-AUTH-*) is implemented as a fail-closed hard access control, and the Push, Streams worker and /reconcile business paths after authentication have been implemented.

**Phase 0 P0 Gate (GATE-P0) - Passed. It was 45 passed / 0 failed / 1 ignored when it was first frozen, and has continued to grow in subsequent rounds ④–⑥. The current baseline is 93 passed / 0 failed / 4 ignored (2026-09-28 retest; see the TEST_BASELINE line of docs/charter.md for the authoritative value). The following are the criteria at the time of stage 0, which are retained for historical records: **

cargo fmt --check passes (no format difference).
cargo clippy --all-targets -- -D warnings passes (zero warnings; crate-level allow is disabled).
cargo test passed (including routing/configuration minimum test).
The above three items are passed in the Debian rust:1-slim-bookworm container** (C-DEBIAN-SLIM).
The configuration is read as env-only (no figment/TOML); SecretString wraps the secret and Debug does not leak it; JMAP_PASSWORD security accessor is reserved.
RUN_MODE does not participate in route dispatch: webhook and reconcile share the same set of routing tables, reconcile is provided as an independent HTTP endpoint POST /reconcile. RUN_MODE only reads and verifies the value at startup, currently does not affect any runtime behavior, and leaves the differential side effects to be mounted here in subsequent stages (see the running mode description in §12.2).

Evolved criteria after phase 0 (current caliber, replacing the specific implementation descriptions in Articles 5 and 6 above):

Configuration reading is narrowed to 2 process-level variables (REDIS_URL, CONFIG_ENCRYPTION_KEY) read directly by src/main.rs; Business configuration is changed to be deserialized by Redis through the config layer, Config::from_env() and all legacy env parsers have been deleted (§7.1).

RUN_MODE The variable itself has been removed, the identifier is no longer in the code (registered in docs/retired.md); section 6 The conclusion of "not participating in route dispatch" is still true, but there is no need for variable carrying.

Security accessor jmap_password() for JMAP_PASSWORD has been removed; jmap_password is retained as a business configuration field name.

.gitignore exists (excluding target/, etc.; do not initialize git without authorization).
All new code has necessary comments referencing stable IDs; a single .rs ≤ 500 lines.

Run Monitor (GATE-UPTIME-KUMA):

Use Uptime Kuma HTTP(s) Monitor to check /healthz (process survival) and /ready (configuration/Redis/upstream reachable readiness), respectively expecting HTTP 200; /ready takes about 3s at worst, and the probe timeout needs to be set to ≥10s; /ready returns 503 and standard error envelope (service_unavailable + Retry-After: 30), Uptime Kuma is still determined according to the status code and is not affected by changes in the response body.
/healthz is pure liveness (unconditional 200); /ready checks configuration integrity + Redis reachability + outbound read-only detection (JMAP session GET, TG getMe, each 3s, parallel, about 3s at worst), the probe is read-only, does not echo tokens or third-party responses, does not trigger business side effects, and the report body does not contain sensitive information; because /ready is the heaviest link (possibly 3s), the platform side should give priority to using gateway-aggregated /healthz-worker as a survival detection to avoid high-frequency outbound requests.
No Prometheus, Exporter or additional indicator ports are introduced; the real Stalwart/Telegram end-to-end link still needs to be jointly debugged separately.

Production red line (throughout all stages):

No Database (C-NO-DB): Redis is the only production state store; no SQLite/Postgres/MySQL/embedded databases are introduced.
No local file writing (C-NO-LOCAL-WRITE): no log files, no data files, no temporary cache, no local volumes mounted.
Log only writes to stdout/stderr (C-LOG-STDOUT-ONLY): The platform is responsible for collection; disable the file log backend.
Log/Redis Content Constraints (SAF-LOG-PURITY): Only structured events, counts, timestamps, masked summaries; prohibited keys, JMAP email bodies, AI request/response, attachment content.
No in-process recovery (C-NO-STATEFUL-RECOVERY): Redis + JMAP reconciliation (FLOW-RECONCILE) will always be used after restart and continuation; the in-process cache is only optimized for performance, and it can be re-entered safely if it is lost.

Stage 1 Pushing the Boundary (BOUND-STAGE1):

Only enter phase 1 after passing all GATE-P0 (JMAP read only).
R1 (real portal authentication) has been implemented (SAF-AUTH-RECONCILE/SAF-AUTH-TG-WEBHOOK/SAF-AUTH-JMAP-PUSH): /reconcile verification Authorization: Bearer RECONCILE_TOKEN, /webhook/tg verification X-Telegram-Bot-Api-Secret-Token == TG_WEBHOOK_SECRET, /push/jmap Verify Redis short-term verification status by subscription ID; constant time comparison, failure 401 and no side effects. When the startup boot variable is missing, the service enters the configuration boot mode without forging business success; the corresponding entry is enabled after the business configuration is completed.
Still need to note: Local CI access control does not mean that the real Stalwart, Redis, and Telegram links have been accepted; real end-to-end testing and external monitoring configuration should be completed before public network deployment.
Phase 1 only does JMAP read-only (list_folders/list_emails/read_email), and does not introduce sending/push/AI.

10.1 Phase 1: JMAP Read Only (1.5d)

Prerequisite: GATE-P0 all passed (BOUND-STAGE1).

Depends on jmap-client 0.4.2 (Introduced, ARCH-DEPS-STAGE1; version/features is subject to Cargo.toml, Disable WebSocket feature); config.rs reuses the existing Redis business configuration skeleton (does not introduce an independent sample configuration file, see ARCH-CONFIG-ENV).
domain::jmap::client = real read-only adapter (MOD-JMAP-CLIENT, G1/D-G1-1 code implemented and verified with a real Stalwart instance by cargo test --ignored jmap:: (session connection → folder enumeration → mailing list → read by id)): implemented = wrapper Client::connect (Basic authentication C-AUTH-APP-BASIC), URL normalization (JMAP_SESSION_URL accepts the service base address or the complete …/.well-known/jmap, normalizes it to origin/base and then passes it in to avoid duplicate paths; only HTTPS, embedded credentials are prohibited and query, REQ-JMAP-SESSION-URL/SAF-JMAP-URL), Redirect trust (only trust the origin host in the configuration, Stalwart will 307 jump to /jmap/session, otherwise jmap-client will reject all redirects by default), account selection: ACCOUNT_ID Leave blank for session The main account is used after explicit value verification (REQ-SINGLE-ACCOUNT).
domain::jmap::JmapService::list_folders / list_emails / read_email (query/get; list_emails contains limit boundary; received_at parsing; read_email multi-part splicing, see REQ-JMAP-RAW-MULTIPART).
R1 entrance authentication is ready: /reconcile//webhook/tg//push/jmap and /api/push/register authentication (SAF-AUTH-*) has been implemented and passed the single test; subsequent changes must not relax or bypass the authentication.
#[ignore] Real test (GATE-G1-JMAP-READONLY): driven by environment variables, run the command cargo test -- --ignored jmap:: (--ignored is a libtest parameter and must be placed after --); only compile without execution using cargo test --no-run (compile the test target containing #[ignore]). Clearly skipping when the environment is missing and not leaking any keys; CI does not run real machine use cases by default, and mock tests continue to be retained. **Whether the real machine use case passes or not must be based on the actual --ignored run - it is not allowed to claim that "the real machine passes" without running it. **
Acceptance: cargo test (mock) is all green; cargo test -- --ignored jmap:: (when there is a real machine environment) can connect to the real server to list folders/emails and read the original text (it cannot be claimed to have passed before it is actually run).

10.2 Phase 2: Channel Adaptation (Completed)

src/channel.rs implements TelegramClient (send_text / send_notification, reqwest self-developed); domain type Notification is in src/domain.rs.
Telegram assembly uses reqwest to send directly to https://api.telegram.org/bot{token}/sendMessage (teloxide is not introduced; see docs/retired.md for evaluation records).
Inbound: POST /webhook/tg Verify secret token → Press update_id to deduplicate (dedup:tg:{update_id}) → Enter Redis Streams → 2xx; Outbound is the only Bot API endpoint used by Telegram.
Intent routing: parse_intent of src/worker.rs parses 6 types of intentions - Help / Consent { ttl, label } / Summary(email_id) / Search(query) / Query / Unknown, all using natural language trigger words (see AI authorization words in docs/reference.md Authorization state section). /search See §2.3 and §5.7 for details; Chinese search terms only perform prefix matching (search/find/retrieval + /search) to avoid overriding authorization and summary intent.
Rendered inside adapter: field Notification → TG Markdown/HTML + escaping.
Use external Redis for session storage (C-REDIS-ONLY-STATE; not using SQLite).
Acceptance: In the form of local webhook, the test client can get a reply when sending a message; the domain module does not import any channel SDK type.

10.3 Phase 3: Sending Email + Status Management (Not implemented)

send_email (draft + submission_set) Currently not implemented: JmapBackend only has read-only verbs.
The sending process was originally planned to use a multi-step FSM to collect to/subject/body, which was never implemented (see the Dialogue FSM section of docs/retired.md).
/flag /unseen Keyword flag unimplemented.
Phase 3 is an unimplemented scope, not a gap: docs/opengaps.md only registers matters that are "not yet completed/yet to be verified". Future stages that are not scheduled for implementation are not included in this list.

10.4 Phase 3.5: LLM facade + rollback (1.5d, not executed)

Not implemented, see docs/retired.md.

10.5 Implemented capabilities: real-time push (Push callback + Streams worker + external Cron reconciliation)

src/notify.rs is a single tiled file of nearly 2000 lines, with no submodules (notify::push_handler / notify::worker / notify::reconcile / mod_streams / mod_sincestate do not exist in the following table):

register_push (POST /api/push/register): Accepts an explicit HTTPS callback URL and creates a subscription; writes back verificationCode with JMAP PushSubscription/set when receiving a Stalwart push.
jmap_push (POST /push/jmap): Verify subscription ID + verificationCode → remove duplicates → enter Redis Streams → 2xx.
worker (POST /worker): Redis Streams consume → Email/changes increment → notify → advance sinceState → XACK.
reconcile (POST /reconcile): External HTTPS Cron trigger (FLOW-RECONCILE); also serves as cursor reconstruction after Redis loss. No CLI subcommand form.
sinceState writes external Redis (state.rs, C-REDIS-ONLY-STATE; not SQLite/file NG-SQLITE-PERSIST/NG-LOCAL-VOLUME).
Target acceptance: Send a test email to Stalwart. After Push callback + worker, Telegram receives the push in ~ seconds; it will not be resent after restarting; it will be reissued by reconciliation after clearing Redis. The current code gate has passed, but end-to-end verification still needs to be completed in real Stalwart, Redis and Telegram environments, and the goals here cannot be regarded as verified facts.

10.5.1 Redis Streams event guarantee boundary (FLOW run semantics)

The Push event is consumed by Streams and delivered to Telegram. Its key path delivery semantics are as follows:

**The only authoritative source for keys and TTL is docs/reference.md (section Redis Keys and TTL). **The tables will not be repeated in this section - repeated tables Will be modified: push:subscription:{id} was previously marked as 7d, and push:registration:{...} was marked as 360s registration solo lock. Both are wrong, and the real 360s solo lock lock:push-register:{sha256(callback_url)} was missing the entire line. has now converged to Single authority table, this section only retains four "whys" that affect design judgment:

Reconciliation lock TTL 300s, deliberately larger than the upper limit of 120s for a single page, to avoid lock expiration during the lock period, causing the same account to enter reconciliation repeatedly (REQ-RECONCILE-IDEMPOTENCY, SAF-RECONCILE-LOCK).
push registration solo lock TTL 360s, deliberately larger than the upper limit of 300s for a single outbound request, otherwise a slow registration will cause repeated requests (SAF-AUTH-JMAP-PUSH).
Reconciliation deduplication 24h is intentionally designed: the same email will only be notified once within 24h (MOD-DEDUP, see 11.5 of this article for the scope).
If the configuration key is missing, it will be regarded as failed (fail-closed) and the business path will not be entered.

Delivery process (notify::worker, one XREADGROUP batch ≤10 items)

Read batch → process one by one → XACK.
Write delivery:inflight (60s) for lease during processing; write delivery:committed (7d) after success.
Entries that crash within the inflight lease window are recycled and retried by another instance through XAUTOCLAIM (the idle threshold is derived from the running timeout configuration, see docs/reference.md §6.3 for the formula and upper and lower limits) - at most repeated and not lost.
retry_or_dlq (state.rs): If the retry count (max_attempts.max(1)) does not reach the upper limit, stay in the source stream and try again; when the upper limit is reached, use a single Lua script to atomically INCR+XADD (into DLQ) + XACK (source stream confirmation), ensuring that there will be no "source has ACKed but is neither in the source nor in the DLQ" gap.

Alarm Boundary (current implementation, see §8.3 Monitoring Conventions)

Use Uptime Kuma HTTP(s) Monitor without introducing Prometheus/exporter.
Indicators such as the number of DLQs, inflight backlog depth, and number of reconciliations have not yet been automatically reported (§8.3 is marked as optional). Current operation and maintenance needs to be viewed directly through Redis: XLEN messageweave:dlq:*, XPENDING, etc.
If you need automated alerts, it is recommended to add health checks on /ready or the reconciliation portal in Uptime Kuma, and manually review the DLQ depth.

Known boundaries (closed, not gapped; docs/opengaps.md currently has no outstanding items)

Multi-instance repeated delivery window (** has been closed, 6c99ce5**): XAUTOCLAIM idle threshold is no longer scaled by a fixed value, but is derived from the running timeout configuration - (max_retries + 1) × (jmap + telegram + llm timeout) × 2 is the upper limit of a single item, multiplied by the batch size, the lower limit is 300s, and the upper limit is 6h (see docs/reference.md §6.3). The early claim window is closed under both single-instance and multi-instance deployments; single instances are not affected, and multi-instances can be repeated at most and will not be lost.

Audit Opinion → Closing (2026-09-26)

[Repair-2] When the enqueue failed, dedup released best-effort, which may have caused a 24h silent loss event → Fixed to claim_dedup_and_enqueue (Lua atoms: SET NX EX must succeed before XADD), there is no intermediate failure window between claim and enqueue.
[Should be revised-1] Email/changes relies on newState for continued transmission, jmap-client 0.4.2 does not have upToId → has been changed to "Same as sinceState, double maxChanges one by one to expand the window (upper limit 4096), only advance new_state when the window cannot be expanded" to avoid missing batches when advancing by page; newState The semantics have been confirmed to be load-bearing (docs/opengaps.md §2), and the callback verification round-trip and backlog emptying in the real Stalwart environment follow the path of baseline → increment /changes → newState (docs/deployment.md §4.1, §6.3.1).
The rest of the low-risk items have been closed: the empty value of the unknown stream is changed to Err (fail-closed, enter retry/DLQ); push:disable is cleared by forget_push_subscription; the lower limit of SET NX EX TTL is tightened to .max(1); XAUTOCLAIM idle threshold is scaled according to the batch size; malformed stream entries without payload are determined by ack_malformed is moved out of PEL via XACK; CSPRNG changes the owner-token to "time + PID + counter" and no longer uses constants.
Unscheduled (Phase 5 "Search + Search Snippet") ** Closed (bfe0fd8)**: /search uses email_query(Filter::text) + SearchSnippet/get. For the upper limit of highlight degradation and truncation, see §2.3 and §5.7; text-level highlighting cannot be done in the locked version (see §2.3 remarks). Stage 5 Backlog cleared.

10.6 Total estimated time

~10.5 man-days (excluding waiting for user confirmation and real joint debugging and troubleshooting).

11. Archive of historical issues and decisions (product/architecture category, all have conclusions)

⚠️ Deployment/platform decisions have been all confirmed (single account / App Password+Basic / Redis hosting + AOF / platform HTTPS URL / external Cron reconciliation / Q-DEP-A platform URL and certificate configurator / Q-DEP-B scheduler selection), which have been filed in the Confirmed Decisions section of docs/deployment.md and will not be repeated in this article.

11.1 Authentication method (confirmed)

Confirmed: App Password + Basic (C-AUTH-APP-BASIC). No master password, no OAuth2 Bearer (no OIDC required).
Affects: Credentials::Basic construct; no need for OAuth client/token auto-refresh module.

11.2 Real-time channel (decided, see deployment.md)

Channel = JMAP Push HTTPS callback + external Cron reconciliation and draining (C-NO-LONG-CONN/C-HTTPS-INBOUND); EventSource/SSE/WebSocket are all non-target (NG-POLLING-SSE). Push registration is triggered explicitly through protected POST /api/push/register; external Cron must call /reconcile (incremental enqueue) and /worker (drain concurrent notification) at the same time - Only adjusting reconciliation will make the notification never sent (deployment.md §6.3.1).
Stalwart side access has been verified in the real environment: the built-in role has the PushSubscription permission, and registration, verification and round-trip and real callback delivery are all passed. The match between the number of callback retries and the idempotent key TTL/reconciliation interval is an operation and maintenance parameter item, not an acceptance threshold: docs/deployment.md §5 provides retry suggestions, §6.3.1 provides a value for the reconciliation interval, and §7 explains that the idempotent key TTL must cover the TG rate-limit fallback upper limit. docs/opengaps.md currently has no outstanding items.

Confirmed: Single account implementation (REQ-SINGLE-ACCOUNT); multiple accounts temporarily use multiple bot instances (respective tokens/configurations), no multiple accounts single instance (so no need for chat→account routing and JmapService pooling). 2026-09-28 Users have made this decision, and the boundaries and future changes to be made are registered in docs/opengaps.md §3.
Deployment shape = webhook-only + universal HTTPS-only Docker + external Cron reconciliation (deployment.md C-NO-LONG-CONN/NG-LONG-POLLING/NFR-RECONCILE-INTERVAL).
Multi-instance LB/HA (confirmed, ARCH-LB-WORKER): You can optionally deploy the same image on multiple serverless platforms, share the same Redis, and use free Cloudflare Worker as the only entrance and failover; the Worker code is located in the subdirectory cloudflare-worker/ (not a native Rust binary). For deployment, see docs/deployment.md for Worker Deployment section. The trust model is transparent transmission (the backend still fails-closed verification, SAF-LB-PASSTHRU), the backends share the same set of secrets** (C-LB-SHARED-SECRETS), /reconcile Redis lock single instance (SAF-RECONCILE-LOCK), and the Worker provides aggregated health view (MOD-HEALTH-AGG). Redis single point of failure is not within the scope of this solution (NFR-HA-MULTI-INSTANCE, solved externally by the user). Both active-active and active-standby are available. See the Multi-Instance Deployment section of docs/deployment.md for details.

11.4 Early questions answered by code (no longer pending confirmation)

The following questions were listed in the form of Q6-Q30 during the design phase. The answers were given when the code was implemented, so they are no longer "items to be confirmed". Only the conclusions are recorded here:

Message format (Q6): Outbound only has send_text (src/channel.rs:114), and its payload structure SendMessage (src/channel.rs:60) has only two fields chat_id + text, no parse_mode - Telegram renders it as plain text after receiving it, without using HTML or MarkdownV2.
Long emails/original text sent directly (Q7, Q29): There is no policy of "do not send the full text if the text exceeds 4000 characters". read_email pulls the full text and sends it to LLM; when AI fails, it falls back to the first 300 characters and no "truncated" mark (see §12.3/§12.5).
Attachments (Q8, Q28): Completely unimplemented. The domain model only has has_attachment: bool, outbound only sendMessage, no send_document and no Blob/get (see §12.3).
Push Scope (Q9): Pull all changes for reconciliation, no folder/sender/keyword filtering is done on the Bot side, and there is no Sieve dependency (Q11 therefore has no impact).
Summary Aggregation/Timed Aggregation (Q10): Not implemented, only /reconcile.
Multiple Sending Identities (Q12): Identity is not used, single account (REQ-SINGLE-ACCOUNT).
unsafe (Q13): There is no unsafe in src/, but #![forbid(unsafe_code)] is not added either.
Monitoring (Q15): Implemented as /healthz + /ready two HTTP probes, without introducing Prometheus/Exporter; see deployment.md for the running platform configuration.
LLM Provider/Network License (Q25, Q26, Q30): LLM_BASE_URL is specified by the deployer (only verifies https); LLM_ENABLED and LLM_ALLOW_NET default are false, both must be true at the same time to construct the client, otherwise the llm field is None (Option<Arc<LlmClient>>, no noop() implementation exists in the warehouse), no detection (see §12.2).
Break Cooling/Threshold (Q27): Not applicable - the fuse itself is not implemented (see §12.4).

There are no pending decisions regarding deployment/platform decisions: Q-DEP-A (who configures the platform URL/domain name and certificate, is determined by the deployment environment at the time of release) and Q-DEP-B (which scheduler is used for external Cron, does not limit the implementation) have both been decided and are archived in the Confirmed Decisions section of docs/deployment.md and will not be repeated in this article.

12. AI auxiliary capabilities: architecture, confirmation threshold, failure fallback

Seven needs confirmed by users (this section is designed accordingly, and all subsequent expressions shall be subject to this):

AI will only access the text after being explicitly requested and confirmed (REQ-AI-CONFIRM / REQ-AI-EXTERNAL-CONSENT): Only when the user explicitly initiates the "analyze/summarize/translate" intention** and confirms it again**, the text is allowed to be fed to LLM; Only when the user explicitly allows it, the email text is sent to the external AI (default is not sent out); other scenarios (including viewing the original text)** are ignored AI**, taken directly from JMAP.

View original text always JMAP direct access (REQ-VIEW-DIRECT): Intent::Query and any "read text" action go to JmapService::read_email, not LLM.

Long emails are prohibited from directly sending the complete original text (REQ-LONG-EMAIL): When the text exceeds the threshold (default 4000 characters, see §12.3, adjustable), Bot does not send the full text directly, but prompts "The text is longer, please view it on a computer" or "Select AI summary"; the AI summary target is about 300 words, no hard mathematical truncation on the code side (by LLM Naturally generated ~300 words).

Analysis/summary results are not persistent (REQ-ANALYSIS-EPHEMERAL): no cache, no Redis, no disk retention; every request is taken and discarded (stateless and naturally friendly).

Attachments are pulled on demand (retained) (REQ-ATTACH-ONDEMAND): No preloading; Blob/get is only triggered when the user clicks to download, single file ≤50MB is sent, and the limit is exceeded by the download_url link.

Non-AI fallback (reserved) that requires user confirmation after AI fails about 3 times (REQ-AI-FUSE): 3 consecutive failures → circuit breaker + popup confirmation to the user; select "Close for 60s" to directly follow the rule template + original text direct path during the cooling period; select "Restore AI" to do a detection.

OpenAI-compatible environment variables (reserved) (REQ-LLM-OPENAI-COMPAT): LLM_API_KEY / LLM_BASE_URL / LLM_MODEL etc.

This section only does design and does not write business code.

12.1 Module division (current implementation)

There is only one file for LLM related code:

src/ai.rs   # LlmClient（唯一实现）：summarize() 打 OpenAI 兼容 /chat/completions

The configuration is loaded in LlmConfig in src/config.rs, and the runtime parameters are in OutboundConfig in src/state.rs (delivered by RuntimeConfigProvider). The LLM capability only exists in the src/ai.rs file; no configuration/fallback/policy/audit layering - circuit breaker, rule fallback, and audit span are all not implemented (see §12.4–12.7). summarize() has only one call site: the /summary authorized summary branch of worker.rs. New email notification does not call LLM - send_notification only renders the From: / Subject: / Received: three lines of metadata, and the body is never sent out (see docs/reference.md §5.4).

12.2 LLM variable list (current implementation)

The LLM_* names in the following table are no longer environment variables**: Config::from_env() After deletion, the code reads zero from them, All loaded via Redis business configuration (§7.1). The "default" in the table is the fallback value in the config layer constructor.

Variable Type Default Purpose

LLM_ENABLED bool false Master switch; off by default, no verification when not enabled LLM_API_KEY/LLM_BASE_URL/LLM_MODEL

LLM_ALLOW_NET bool false Runtime outbound permission; LlmClient is constructed only if LLM_ENABLED && LLM_ALLOW_NET are both true, otherwise the llm field is None (src/main.rs:135-141)

LLM_API_KEY string Can be omitted Bearer token; required only when LLM_ENABLED=true

LLM_BASE_URL URL Can be omitted Required only when enabled, and must be https, otherwise AiError::InvalidEndpoint

LLM_MODEL string Can be omitted Required only when enabled; transparently passed to model in /chat/completions

LLM_SUMMARY_TARGET_CHARS int 300 LlmClient::max_chars, perform character truncation on the returned summary text (there is truncation on the code side)

max_retries (running parameters) int See docs/reference.md §6.1 Issued by OutboundConfig (RuntimeConfigProvider), see docs/reference.md §6.1 for the hard upper limit; LLM is shared with Telegram outbound

llm_timeout_ms (running parameter) int See docs/reference.md §6.1 Issued by OutboundConfig; unit milliseconds, for the lower limit and value range, see docs/reference.md §6.1

Variable	Type	Default	Purpose
`LLM_ENABLED`	bool	`false`	Master switch; off by default, no verification when not enabled `LLM_API_KEY`/`LLM_BASE_URL`/`LLM_MODEL`
`LLM_ALLOW_NET`	bool	`false`	Runtime outbound permission; `LlmClient` is constructed only if `LLM_ENABLED && LLM_ALLOW_NET` are both true, otherwise the `llm` field is `None` (`src/main.rs:135-141`)
`LLM_API_KEY`	string	Can be omitted	Bearer token; required only when `LLM_ENABLED=true`
`LLM_BASE_URL`	URL	Can be omitted	Required only when enabled, and must be `https`, otherwise `AiError::InvalidEndpoint`
`LLM_MODEL`	string	Can be omitted	Required only when enabled; transparently passed to `model` in `/chat/completions`
`LLM_SUMMARY_TARGET_CHARS`	int	`300`	`LlmClient::max_chars`, perform character truncation on the returned summary text (there is truncation on the code side)
`max_retries` (running parameters)	int	See `docs/reference.md` §6.1	Issued by `OutboundConfig` (`RuntimeConfigProvider`), see `docs/reference.md` §6.1 for the hard upper limit; LLM is shared with Telegram outbound
`llm_timeout_ms` (running parameter)	int	See `docs/reference.md` §6.1	Issued by `OutboundConfig`; unit milliseconds, for the lower limit and value range, see `docs/reference.md` §6.1

None LLM_TEMPERATURE / LLM_MAX_TOKENS / LLM_TIMEOUT_SECS / LLM_MAX_RETRIES: The request body only has model / messages / max_tokens (= max_chars * 2), no temperature, and no timeout switch independent of llm_timeout_ms. The API key is only read from the configuration, without entering the source code or logging (SAF-LOG-PURITY). For the runtime secret injection method, see deployment.md (C-NO-SECRET-IN-IMAGE).

12.3 Text acquisition strategy and long email processing (requirement 1/2/3/5)

View original text = JMAP direct access: Intent::Query goes through JmapService::read_email with any "read text" action, without going through LLM; LLM is not on the viewing path.
Preliminary threshold for AI to contact the text: Only when the user initiates a clear "analysis/summarization/translation" intention** and confirms it again** (such as clicking the [AI Summary] button / executing /summarize), the text will be input as LLM. Unconfirmed → LLM cannot see the text.
No "long message protection" logic: no 4000 character threshold, no Preview type, no [continue to view original] button. The digest path after authorization is to hand over the full text to LLM. When LLM fails, it will fall back to the first 300 characters, and there will be no truncation mark in the whole process.
The fallback text does not have any annotation: fallback() is body.chars().take(300) and returns directly without appending the words "Truncated" or "AI Unavailable".
The attachment function is not implemented at all: src/domain/jmap.rs clearly annotates "no mutation, attachment, AI, or streaming APIs", read_email only returns the body text; the only attachment information in the domain model is has_attachment: bool. Telegram only calls sendMessage outbound, does not exist send_document / Blob/get streaming download / download button.
The input for AI summary is only the email body text, and there is no attachment channel.

12.3.1 AI authorization period and Redis TTL (`REQ-AI-CONSENT`)

AI authorization must have a duration explicitly selected by the user; optionally temporary once (one authorization window, 3600 seconds), today (86400 seconds), 7 days (604800 seconds), or until revoked (up to 365 days, or 31536000 seconds). Trigger words include matches according to the original text (temporary/once/today/7 days/until I undo/long-term, and /ai on, /ai yes, /ai off), and English aliases are not supported.
Redis's consent key only saves the expiration Unix time for the chat id, and uses exactly the same number of SET EX seconds as the selected period; does not use the implicit 30-day default, and does not save the body or summary.
Expiration is guaranteed by Redis TTL and expiration verification when reading. The digest request returns to metadata mode and prompts "Authorization has expired"; the user can re-select the expiration date before authorizing again. /ai off deletes the key immediately.

12.4 Failure detection (current implementation)

LlmClient::summarize (src/ai.rs) Only retries, no circuit breaker:

Timeout llm_timeout_ms, number of retries max_retries (see docs/reference.md §6.1 for unit, default value, lower limit and hard upper limit; both items share the same configuration with Telegram outbound).
Only retry the last failure: 429 and 5xx will be retried; other 4xx (including 401/403) will return an error immediately and will not be retried.
Retries exhausted or all timed out → Err(AiError::Response); URL parsing failed or non-https → Err(AiError::InvalidEndpoint); JSON deserialization failed or missing choices[0].message.content → Err(AiError::Response).
AiError has only three variants: InvalidEndpoint / Request / Response. None Subdivision type for 429/401/Timeout.

12.5 Non-AI fallback (current implementation)

fallback() of worker.rs just takes the first 300 characters of the text, nothing more:

The AI authorization is valid and the LLM call fails → Silently replace it with these 300 characters, without prompting the user, and without adding the "AI Unavailable" mark.
AI is not authorized → Do not send summary, reply with Sender — Subject + "Authorization has expired or has not been authorized yet".
Therefore, there is no fuse half-open state, no inline-button to confirm rollback, no 60s cooling period, and no fuse count written to Redis (there are no LLM-related status keys on Redis).

12.6 Stateless/Containerized Adaptation

Summary No persistence, no caching: worker.rs calls LLM immediately every time, and the result is that it does not write to Redis and does not write to disk. Consistent with C-REDIS-ONLY-STATE.
The log does not record prompt plain text and API key (SAF-LOG-PURITY) - but the current log does not record any LLM call events (see §8.3).
The only valid running parameters are llm_timeout_ms, max_retries, and LLM_SUMMARY_TARGET_CHARS (corresponding to LlmClient::max_chars, which truncates the summary output). Does not exist The LLM_MAX_RETRIES environment variable.
The request body only contains model / messages / max_tokens (max_tokens = max_chars * 2); no temperature, no system prompt - the body is sent directly as the only user message.

12.7 Error handling and observability (continued from §8)

Not implemented BotError::Llm variant (BotError only has Config/Io/Json/State four), LLM errors are swallowed in place within worker.rs and downgraded to fallback text, not propagated upward. Not implemented tracing span for llm.call (no prompt hash, response length, latency record). These two items are documented in docs/retired.md, along with 12.4's circuit breaker/half-open design, 12.5's rules rollback, and "AI not available" logo.

"Phase 3.5" in the 12.8 plan has not been implemented

The original plan was to insert "Phase 3.5: LLM Facade + Fallback (1.5d)" between Phases 3 and 4, delivering the three modules of llm::client / llm::policy / llm::fallback, plus a 4000-character long email threshold, "AI Unavailable" logo and 3 circuit breakers + Redis share count. This plan did not perform as designed: LLM capabilities ended up in a single file src/ai.rs (see §12.1), with no facade, no policy, no fallback module, and no logo and threshold branches. Retired parts are documented in docs/retired.md. The original set of Q25-Q30 design questions in this section have also been invalidated - their respective conclusions have been given when the code is implemented, and the answers can be found in §11.4.

Attachment: Research basis (can be reviewed)

stalwartlabs/jmap-client main Branches (the paths are relative to the root of the warehouse): stalwartlabs/jmap-client/src/lib.rs (URI/Method/DataType/Error), stalwartlabs/jmap-client/src/client.rs (Authentication/Connection/event _source), stalwartlabs/jmap-client/src/email/, stalwartlabs/jmap-client/src/email_submission/helpers.rs, stalwartlabs/jmap-client /src/event_source/, stalwartlabs/jmap-client/src/push_subscription/ (create/verify/update_types/destroy), stalwartlabs/jmap-client /src/core/error.rs, stalwartlabs/jmap-client/Cargo.toml, stalwartlabs/jmap-client/README.md, stalwartlabs/jmap-client/examples/.
crates.io: jmap-client metadata.
stalwartlabs/mail-server main branch: stalwartlabs/mail-server/crates/common/src/auth/credential.rs (Password/AppPassword/ApiKey), stalwartlabs/mail-server/crates/http/src/auth/authenticate.rs (AccessScope Permission tailoring), stalwartlabs/mail-server/crates/jmap/src/push/, stalwartlabs/mail-server/api/v1/openapi.yml (securitySchemes: basicAuth/bearerAuth/liveToken 60s).
Project directory /home/okabe/Repo/messageweave/ (see §10.0 and ../AGENTS.md §3.3 for toolchain requirements).
For deployment/platform related research basis (lambda_runtime/worker/aws-sdk, etc.), see deployment.md.

109 KiB Raw Blame History Unescape Escape