Merge remote-tracking branch 'origin/master' into feat/read-image-context

.agents/notes/implemented/architecture/2026-08-10-session-log-version-mechanism.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-10-session-log-version-mechanism.md
+2026-08-10-session-log-version-mechanism.md: 25eb1230a254219c827b1d2750dba367b113f9f7
+2026-08-10-session-log-version-mechanism.zh.md: c47670f2de77773c17c9595eff442bf7f1e8ec3e

.agents/notes/implemented/architecture/2026-08-10-session-log-version-mechanism.md

@@ -0,0 +1,30 @@
+# Agent Note: Session log versioning — one integer, an upgrade chain, and a per-event ignorable marker
+
+Status: implemented
+
+English | [中文](2026-08-10-session-log-version-mechanism.zh.md)
+
+## Problem
+
+Session logs must be upgradable after release, and the runtime that ships first is the floor for every later decision: whatever refusal and degradation behavior is missing from the first released reader can never be added to the copies users already run. Release issue #1901 required at minimum that an old runtime reading a newer session format reports "unsupported" instead of misreading it. The pre-change reader did the opposite on both axes: `assertVersion` rejected any version mismatch with one direction-blind message, and the JSONL decoder passed unknown event types through untouched, so reconstruction silently skipped them — resuming a gutted session with no diagnostic at all.
+
+## Decision
+
+**One monotonic integer, no major/minor split.** Whether a version step is auto-upgradable is a property of that step — expressed by whether its upgrader exists — not something a two-level numbering scheme should promise in advance (you rarely know at design time whether the next change will turn out "major"). This matches the SQLite backend's `SCHEMA_VERSION` precedent.
+
+**The writer decides bumps, not the reader.** A bump is required exactly when an old runtime could no longer handle a new log with full semantic correctness. "Parses without error" is not the bar: silently skipping content that shapes reconstruction is a wrong read. Only structural changes qualify — header shape, event envelope, core event semantics, the surface mechanism (`SurfaceEventType` set, `SurfaceOp` variants). When unsure, bump: a near-identity upgrader is almost free, a missed bump silently corrupts old readers.
+
+**Read rules by direction.** Equal version: read normally. Newer than the reader: refuse, name the direction ("written by a newer harness — upgrade"), and point at the raw log artifact so the user can still see the text (`SessionFormatUnsupportedError`, distinct from `SessionPersistenceCorruptionError` because nothing is damaged). Older than the reader: convert in memory through the chain of n→n+1 upgraders for viewing; persist the converted log only when the session is actually continued (atomic temp-file replace, original kept as backup). A step whose upgrader cannot be written is left empty, which cuts off every version at or below it — those degrade to raw-text viewing.
+
+**A per-event `ignorable` marker covers vocabulary growth, so ordinary event additions never bump the version.** The event vocabulary is decided by which plugins are mounted, which a single version integer cannot describe. A reader meeting an unrecognized event type refuses to interpret the log unless the event carries `ignorable: true` in its envelope. The default is *required*: forgetting the marker over-refuses a resumable session (an inconvenience), while a default of ignorable would make the same mistake silently resume a gutted one (a safety failure). The architecture makes this sound: model-visible content flows only through the three `surfaceOp`-marked surface event types plus the `request/header`/`request/context` folds, so the dangerous unknowns are exactly the non-surface events that change how the rest of the log is read (`session/end-seed` is the existing example).
+
+## Consequences
+
+What shipped in v0 (release 0812): direction-aware refusal with the raw-log path; the unknown-event guard against a generated known-vocabulary list (`KNOWN_SESSION_EVENT_TYPES`, emitted by `gen-persistence-catalog` from every `SessionEventMap` merge and kept fresh by `verify-persistence-catalog`); the `ignorable` envelope field accepted by seed validation, both backends (a dedicated SQLite column, `SCHEMA_VERSION` 15), and the BFF wire schema. The upgrader chain itself is deferred until the first real v0→v1 step exists to test it against; writers do not yet set `ignorable` (no producer needs it), so `Session.append` gains that surface with its first user. Until a registration surface exists, an out-of-repo plugin's events refuse resume under first-party readers — the pre-release stance accepts that, and the refusal is loud rather than silent. The unknown-type guard is read-side only: `appendCore` keeps rejecting retired legacy shapes but does not vocabulary-check new types, because an append-time refusal would stall a live session's durability mid-flight, which costs more than a loud refusal at the log's next load. The JSONL backend additionally refuses a foreign version from the raw header line before validating today's header shape or decoding any event row, so a structurally different future format still reports the upgrade direction instead of "corrupt"; SQLite gates whole-file structure through its own `SCHEMA_VERSION` pragma first.
+
+## Alternatives considered
+
+- **Major/minor versioning** — the "is it convertible" bit lives on each step's upgrader, and pre-committing it into a number shape invites wrong promises.
+- **Default-ignorable unknown events** — inverts the failure mode of a forgotten marker from visible over-refusal into silent corruption.
+- **Auto-migrating on view** — rewriting the artifact on open turns a read into a destructive write: a converter bug corrupts logs at browse time, and a same-directory older runtime loses access because a newer one merely looked.
+- **Per-plugin runtime registration of known event types** — would make the known set composition-dependent, so a leaner same-version composition would refuse logs a fuller one wrote. The generated repo-wide list keeps same-version reads uniform; out-of-repo plugin events are outside it by construction, and a registration surface for them is deferred until such a consumer exists.

.agents/notes/implemented/architecture/2026-08-10-session-log-version-mechanism.zh.md

@@ -0,0 +1,30 @@
+# Agent Note：Session log 版本机制：单调整数、升级器链、逐事件可忽略标记
+
+Status: implemented
+
+[English](2026-08-10-session-log-version-mechanism.md) | 中文
+
+## 问题
+
+Session log 在发布后必须能升级格式，而最先发布的运行时决定了此后一切的下限：第一个发布版的读取器缺少哪种拒绝和降级行为，用户手里已经装上的副本就永远补不上。发布 issue #1901 的最低要求是老运行时读到新 Session 格式时明确报不支持，而不是读错。改动前的读取器在两个方向上都做反了：`assertVersion` 对任何版本不匹配抛出同一条不区分方向的消息；JSONL 解码器把不认识的事件类型原样放行，重建时静默跳过，恢复出一个内容残缺的会话且没有任何诊断。
+
+## 决定
+
+**一个单调递增的整数，不分大小版本。**某一步能不能自动升级是那一步自己的属性，由它的升级器存在与否表达，不该由两级编号方案提前承诺（设计时很少能预知下一个变更算不算"大"）。这与 SQLite 后端 `SCHEMA_VERSION` 的先例一致。
+
+**升不升版本由写入方决定，与读取方能力无关。**当且仅当老运行时无法在语义上完全正确地处理新日志时才必须升版本。"解析不报错"不是标准：静默跳过影响重建的内容就是读错。只有结构性变更够得上这条线：header 形状、事件信封、核心事件语义、surface 机制（`SurfaceEventType` 集合、`SurfaceOp` 变体）。拿不准就升：近似恒等的升级器几乎没有成本，漏升一次会让老读取器静默读坏。
+
+**读取规则按方向区分。**版本相等：正常读。比读取器新：拒绝，说明方向（"由更新的 harness 写入，请升级"），并给出原始日志文件的路径，用户仍能看到文本（`SessionFormatUnsupportedError`，与 `SessionPersistenceCorruptionError` 区分，因为数据没有损坏）。比读取器旧：查看时经 n→n+1 升级器链在内存中逐级转换；只有会话真正被继续时才把转换落盘（临时文件原子替换，原文件留备份）。写不出升级器的那一步留空，这会切断该步及更早所有版本的升级路径，它们降级为只能看原文。
+
+**逐事件的 `ignorable` 标记吸收词汇表增长，普通的新增事件永远不用升版本。**事件词汇表由挂载了哪些插件决定，单个版本整数描述不了它。读取器遇到不认识的事件类型时拒绝解读日志，除非该事件的信封带 `ignorable: true`。默认为必需：忘写标记的后果是把一个本可恢复的会话拒绝过头（体验问题），而默认可忽略会让同样的疏忽静默恢复出残缺会话（安全事故）。架构保证了这条规则成立：模型可见内容只经三种带 `surfaceOp` 标记的 surface 事件加 `request/header`、`request/context` 折叠进入重建，危险的未知事件恰好是那些不进 surface 但改变日志其余部分解读方式的事件（`session/end-seed` 是现存例子）。
+
+## 影响
+
+v0（0812 发布）交付的内容：分方向的拒绝并带原始日志路径；基于生成的已知词汇清单（`KNOWN_SESSION_EVENT_TYPES`，由 `gen-persistence-catalog` 从所有 `SessionEventMap` 声明合并生成，`verify-persistence-catalog` 保证新鲜）的未知事件守卫；`ignorable` 信封字段被种子校验、两个后端（SQLite 专用列，`SCHEMA_VERSION` 升到 15）和 BFF 线上 schema 接受。升级器链本身推迟到第一个真实的 v0→v1 变更出现、有真实对象可测时再建；写入侧目前不写 `ignorable`（还没有生产者需要它），`Session.append` 的这一表面随第一个使用者一起落地。在注册表面出现之前，仓库外插件的事件在第一方读取器下无法恢复会话，预发布立场接受这一点，而且拒绝是显式的而非静默的。未知类型守卫只在读取侧生效：`appendCore` 继续拒绝已淘汰的 legacy 形状，但不对新类型做词汇检查，因为写入时拒绝会让活跃会话的持久化中途停摆，代价大于下次加载时的显式拒绝。JSONL 后端还会在校验当前 header 形状、解码任何事件行之前，直接从原始 header 行拒绝外来版本，因此结构完全不同的未来格式仍会报告升级方向而不是"损坏"；SQLite 则先由自己的 `SCHEMA_VERSION` pragma 把关整个文件的结构。
+
+## 曾考虑的替代方案
+
+- **大小两级版本号**：能否转换这一位信息属于每一步的升级器，把它预先固化进编号形状会做出错误承诺。
+- **未知事件默认可忽略**：把忘写标记的后果从可见的过度拒绝反转成静默损坏。
+- **查看时自动迁移落盘**：打开即改写把读操作变成破坏性写操作，转换器的 bug 会在浏览时损坏日志，同目录的旧版本运行时也会因为新版本只是看了一眼就失去访问能力。
+- **插件运行时注册已知事件类型**：会让已知集依赖插件组合，同版本的精简组合会拒绝完整组合写出的日志。生成的全仓库清单保证同版本读取行为一致；仓库外插件的事件按构造就在清单之外，为它们提供注册表面推迟到真有这样的消费者时再做。

.agents/notes/implemented/feature/2026-07-07-mcp-client-plugin.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-07-mcp-client-plugin.md
-2026-07-07-mcp-client-plugin.md: 756a5c4dc9f1152ecb1955d93b8dc47fcf07c661
-2026-07-07-mcp-client-plugin.zh.md: 9f45203c479379087c5a19eccce3ab7a339d4ed0
+2026-07-07-mcp-client-plugin.md: 077d978d8815d759574f89ed524ab3bc8c24a267
+2026-07-07-mcp-client-plugin.zh.md: 8f58c9359ca8447717cc353f97f1333370fa6fda

.agents/notes/implemented/feature/2026-07-07-mcp-client-plugin.md

@@ -153,13 +153,7 @@ Build the child environment from the subprocess seam's shared `scrubbedParentEnv
 
 ### Disconnection / crash
 
-No auto-reconnect. If the MCP server process exits or the transport closes:
-
-1. The effect disposes → all registered tools are unregistered (fiber-scoped disposers).
-2. Subsequent model calls to those tools → `ToolNotFoundError` → `isError: true`.
-3. Recovery: user edits `cordis.yml` (triggers HMR reload) or restarts the harness.
-
-This matches the ACP subagent pattern: "crash = terminal, report error, clean up, don't retry."
+A per-instance connection supervisor reconnects automatically after a lost connection with bounded exponential backoff and a per-outage attempt budget, re-running discovery on success; exhaustion unregisters the server's tools and stops until reload. The [auto-reconnect Agent Note](2026-08-06-mcp-client-auto-reconnect.md) owns that decision, including the `reconnect` config block and the `reconnect.enabled: false` opt-out that restores manual HMR/restart recovery.
 
 ## Alternatives considered
 
@@ -173,7 +167,7 @@ Rejected. There is no foreseeable alternative MCP client implementation — MCP
 
 ### Auto-reconnect with exponential backoff
 
-Rejected for v1. Adds complexity (partial-availability state where tools are registered but temporarily non-functional), and stdio process crashes usually indicate a configuration problem that retrying won't fix. HMR already provides the manual recovery path. Can be added as a future `reconnect: boolean` config if needed.
+Rejected for v1: it added a partial-availability state (tools registered but temporarily non-functional), and stdio crashes often indicate configuration problems retrying cannot fix; HMR was the recovery path. Operational feedback reversed the deferral — the [auto-reconnect Agent Note](2026-08-06-mcp-client-auto-reconnect.md) implements it with a bounded per-outage budget and an opt-out.
 
 ### Bridge Resources and Prompts
 
@@ -211,4 +205,4 @@ Coverage is named per tier; each behavior lives at the cheapest tier that can ex
 - **MCP SDK stability**: the `@modelcontextprotocol/sdk` is still evolving; breaking changes require updating the bridge. The version is pinned, and the SDK is widely adopted (Claude Desktop, Cursor, VS Code) so breaking changes are unlikely to be silent.
 - **Tool schema quality**: MCP servers may expose poorly-described tools (vague descriptions, incomplete JSON schemas). The harness passes them through as-is — garbage-in-garbage-out; that is the server author's responsibility, not the bridge's.
 - **Stdio process management**: a misbehaving MCP server that ignores signals could wedge dispose. The Cordis fiber disposal has bounded quiescence; a stuck transport eventually times out at the framework level.
-- Crash recovery is manual (HMR edit or restart) — accepted for v1; a `reconnect` config remains open as future work.
+- Crash recovery is automatic within the [reconnect budget](2026-08-06-mcp-client-auto-reconnect.md); manual reload remains the path after exhaustion or with `reconnect.enabled: false`.

.agents/notes/implemented/feature/2026-07-07-mcp-client-plugin.zh.md

@@ -153,13 +153,7 @@ MCP 仅保证工具名在[单个服务器内](https://modelcontextprotocol.io/sp
 
 ### 断连 / 崩溃
 
-不自动重连。如果 MCP 服务器进程退出或传输层关闭：
-
-1. effect dispose → 所有已注册工具被注销（fiber 作用域的 disposer）。
-2. 后续模型对这些工具的调用 → `ToolNotFoundError` → `isError: true`。
-3. 恢复：用户编辑 `cordis.yml`（触发 HMR 重载）或重启 harness。
-
-这与 ACP subagent 模式一致：「崩溃即终态，报告错误，清理资源，不重试。」
+每个实例的连接监督器在连接丢失后以有界指数退避和单次故障尝试预算自动重连，成功后重新执行发现流程；尝试耗尽则注销该服务器的工具并停止，直到重新加载。[自动重连 Agent Note](2026-08-06-mcp-client-auto-reconnect.md) 拥有该决策，包括 `reconnect` 配置块和恢复手动 HMR/重启恢复的 `reconnect.enabled: false` opt-out。
 
 ## 曾考虑的替代方案
 
@@ -173,7 +167,7 @@ MCP 仅保证工具名在[单个服务器内](https://modelcontextprotocol.io/sp
 
 ### 指数退避自动重连
 
-v1 否决。引入复杂性（工具已注册但暂时不可用的部分可用状态），且 stdio 进程崩溃通常表明配置问题，重试无法修复。HMR 已提供手动恢复路径。如有需要，可在未来作为 `reconnect: boolean` 配置项添加。
+v1 否决：引入了部分可用状态（工具已注册但暂时不可用），且 stdio 崩溃往往表明配置问题，重试无法修复；HMR 曾是恢复路径。运营反馈扭转了该延期决定——[自动重连 Agent Note](2026-08-06-mcp-client-auto-reconnect.md) 以有界的单次故障预算和 opt-out 实现了自动重连。
 
 ### 桥接 Resources 和 Prompts
 
@@ -211,4 +205,4 @@ v1 否决。它能防止跨服务器冲突，但无法将 MCP 注册与原生 ha
 - **MCP SDK 稳定性**：`@modelcontextprotocol/sdk` 仍在演进中；破坏性变更需要更新桥接。版本已固定，且该 SDK 被广泛采用（Claude Desktop、Cursor、VS Code），因此破坏性变更不太可能悄然发生。
 - **工具 schema 质量**：MCP 服务器可能暴露描述不佳的工具（模糊的描述、不完整的 JSON Schema）。harness 原样透传——垃圾进垃圾出；这是服务器作者的责任，不是桥接的。
 - **Stdio 进程管理**：行为异常的 MCP 服务器如果忽略信号，可能卡住 dispose。Cordis fiber 的 dispose 具有有界的完全停稳过程；卡住的传输层最终会在框架层面超时。
-- 崩溃恢复是手动的（HMR 编辑或重启）——v1 已接受；`reconnect` 配置作为未来工作保持开放。
+- 崩溃恢复在[重连预算](2026-08-06-mcp-client-auto-reconnect.md)内自动进行；耗尽后或配置 `reconnect.enabled: false` 时回退为手动重新加载。

.agents/notes/implemented/feature/2026-08-06-mcp-client-auto-reconnect.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-08-06-mcp-client-auto-reconnect.md
+2026-08-06-mcp-client-auto-reconnect.md: 99a8aec1abe3713822f8f17c17d8efaca5d61a4d
+2026-08-06-mcp-client-auto-reconnect.zh.md: 8d4dc935e6edee9a05f556774e743e48234ca709

.agents/notes/implemented/feature/2026-08-06-mcp-client-auto-reconnect.md

@@ -0,0 +1,48 @@
+# Agent Note: MCP client auto-reconnect with bounded backoff
+
+Status: implemented
+
+English | [中文](2026-08-06-mcp-client-auto-reconnect.zh.md)
+
+## Problem
+
+The [MCP client](2026-07-07-mcp-client-plugin.md) connected once at plugin load. When a stdio server crashed or was killed, its registered tools stayed visible but every call failed with `Not connected` until a human edited the config (HMR) or restarted the Host — v1 explicitly deferred reconnection. Long-running hosts (ACP automation, web) cannot be bounced because a child process died, and for stdio the harness composition is the only party that can respawn it. External feedback escalated this as a real operational gap (issue #1746).
+
+## Decision
+
+`packages/mcp/mcp-client/src/connection.ts` owns a per-instance connection supervisor; `apply()` shrinks to config resolution plus two effects (the `serverName` reservation and the supervisor's lifecycle). The supervisor owns the client/transport generations, the live tool registrations, and the reconnect loop.
+
+**Trigger.** The supervisor arms `client.onclose` per generation. The SDK fires it when the stdio child exits, so a crash is observed without polling. `StreamableHTTPClientTransport` fires `onclose` only for deliberate closes — it owns its internal SSE-stream recovery and surfaces request failures per call — so HTTP servers are effectively outside supervisor restarts; the package README records that limitation.
+
+**Generations without interleaving.** Each attempt builds a fresh transport and `Client` (the SDK binds a Protocol to one transport for life). One per-supervisor queue serializes every `syncTools` call — initial syncs and `list_changed` re-syncs across all generations — and an `isCurrent` fence makes stale generations inert, so no two syncs can interleave the dispose-previous/register-next swap (which would double-dispose one generation and leak another). The queue also closes a pre-existing race where two rapid `list_changed` notifications re-synced concurrently. The activation attempt, rather than the first queue entrant, explicitly owns strict startup registration: an early `list_changed` notification uses contained re-sync semantics and cannot consume `failOnStartupError`. Failure signals are idempotent per generation: a connect rejection racing its own transport close schedules exactly one retry. A failed attempt cannot enter backoff until both `Client.close()` settles and the transport reports `onclose`, which for stdio proves the child exited; a missing close signal stops reconnection after the SDK's bounded termination window instead of allowing two server processes to overlap. Disposal uses the same bounded close-signal barrier and reports an incomplete shutdown without ever restarting.
+
+**Bounded backoff with an outage budget.** Delays double from `initialDelayMs` up to `maxDelayMs`. One outage shares `maxAttempts` consecutive failed attempts; exhaustion unregisters the server's tools, logs at error level, and stops until disposal or reload. A connection that survives past the stability window — `maxDelayMs`, derived rather than a fifth tunable, as the longest configured backoff spacing — resets the budget, so an occasionally-crashing server recovers indefinitely while a crash loop whose connects briefly succeed cannot launder its budget into a restart storm.
+
+**Config and resolution.** Both transports accept `reconnect { enabled, initialDelayMs, maxDelayMs, maxAttempts }` with schemastery defaults (on, 500ms, 30s, 10). `resolveReconnectPolicy()` is the explicit resolve step: it re-judges every bound and cross-field constraint because programmatic construction may bypass Schemastery, and misconfiguration fails the plugin instance at load.
+
+**Observable states.** An initial or retry-attempt failure says `connection failed`; an established generation ending says `connection lost`. Retrying logs at warn with attempt count and delay, recovery at info, final failure and disabled recovery at error. During an outage the last good generation stays registered and calls against it fail — deterministic public names mean a recovered unchanged tool list reproduces identical definitions, keeping the model-visible schema prefix stable instead of flapping. With `reconnect.enabled: false` a lost connection keeps the v1 manual-recovery behavior.
+
+**Disposal.** Dispose flips the fence, cancels any pending timer, closes the current client, then awaits the in-flight attempt and the sync queue before unregistering — quiescence, not just a request to stop. The reconnect timer is unref'd so a waiting backoff never holds a finishing process open.
+
+## Alternatives considered
+
+**Consecutive-failure counter that resets on every successful connect.** Rejected: a crash-looping server whose connects briefly succeed would reset the budget each cycle and restart forever — exactly the restart storm the failure cap exists to prevent. The uptime-gated reset distinguishes a recovered server from a looping one without new configuration.
+
+**Reuse one SDK `Client` across reconnects.** The Protocol clears its transport on close and can technically connect again, but the SDK's own guidance is one connection per Protocol instance, and reuse carries notification handlers and negotiated capability state across server incarnations. A fresh `Client` per generation plus the `isCurrent` fence is unambiguous.
+
+**Unregister tools immediately on disconnect, re-register on recovery.** Rejected: a transient outage would flap the model-visible tool list (two schema-prefix invalidations per crash) for no information gain; failing calls already signal the outage, and the swap on recovery is atomic per generation. Tools are unregistered at final failure so a permanently dead server does not leak permanently broken tools.
+
+**Route Streamable HTTP request failures into the supervisor.** Rejected for now: the HTTP transport already reconnects its SSE stream with its own backoff, per-request errors do not imply a dead server, and there is no child process the harness could respawn. Transport close stays the single trigger.
+
+**Restart through Loader/HMR machinery instead of an in-plugin supervisor.** Rejected: the Loader owns config-driven recomposition, not runtime health. A plugin restarting itself through the Loader would conflate config generations with connection generations and lose the per-outage budget.
+
+## Testing
+
+Unit (`tests/reconnect.spec.ts`, mocked SDK): recovery swaps generations without duplication or leaks and serves post-recovery calls, diagnostics distinguish initial or retry failure from established connection loss, strict startup registration survives a pre-connect `list_changed` notification, failed initialization waits for the old generation's close signal and fails closed when that signal never arrives, disposal waits for the same signal with a bounded incomplete-shutdown path, the failure cap unregisters tools and stops, dispose cancels a pending backoff and quiesces an in-flight sync, a close after dispose schedules nothing, disabled mode keeps the v1 behavior, the stability window resets the budget while a crash loop exhausts it, double failure signals schedule one retry, stale generations and handlers are inert, and `resolveReconnectPolicy` rejects each invalid bound. E2E (`tests/mcp-client.e2e.ts`, keyless): the fixture server gained a `crash` tool that replies then exits; real-process tests prove a stdio crash recovers end to end and that unloading the plugin mid-outage stops reconnection promptly. Snapshot: deliberately none, per the original note's rationale — reconnection adds no new presentation shape, and a snapshot composition spawning a crashing server would make replays timing-dependent.
+
+## Consequences
+
+- A crashed stdio MCP server recovers without human intervention: bounded backoff, re-discovery, atomic generation swap. Default policy retries an outage for roughly 2.5 minutes before giving up.
+- Connection state is genuinely more intricate than connect-once — the partial-availability window v1 avoided now exists (registered tools failing during an outage), concentrated in one module with the invariants named.
+- `reconnect` is new config surface on both transports, and the stability window is deliberately derived from `maxDelayMs`; making it independently tunable is a compatible future change.
+- After final failure or with reconnect disabled, the plugin stays loaded with no (or failing) tools until reload — deliberate and logged, so a chronically broken server cannot restart forever.

.agents/notes/implemented/feature/2026-08-06-mcp-client-auto-reconnect.zh.md

@@ -0,0 +1,48 @@
+# Agent Note: MCP client auto-reconnect with bounded backoff
+
+Status: implemented
+
+[English](2026-08-06-mcp-client-auto-reconnect.md) | 中文
+
+## 问题
+
+[MCP 客户端](2026-07-07-mcp-client-plugin.md)在插件加载时仅连接一次。stdio 服务器崩溃或被终止后，其已注册的工具仍然可见，但每次调用均以 `Not connected` 失败，直到人工编辑配置触发 HMR（热模块替换）重载，或重启 Host——v1 明确推迟了重连机制。长时间运行的 Host（ACP 自动化、Web）不能因为子进程死亡就被重启；而对于 stdio 传输，harness 组合层是唯一能重新拉起子进程的一方。外部反馈将此升级为真实的运维缺口（issue #1746）。
+
+## 决策
+
+`packages/mcp/mcp-client/src/connection.ts` 拥有一个逐实例的连接监督器；`apply()` 收缩为配置解析加两个副作用（`serverName` 预留和监督器的生命周期）。监督器负责管理 client/transport 代、活跃的工具注册以及重连循环。
+
+**触发条件。** 监督器在每一代上挂载 `client.onclose`。SDK 在 stdio 子进程退出时触发该回调，因此崩溃无需轮询即可感知。`StreamableHTTPClientTransport` 仅在主动关闭时触发 `onclose`——它内部拥有自己的 SSE（Server-Sent Events）流恢复机制，并将请求失败以逐调用方式暴露——因此 HTTP 服务器实际上不在监督器的重启范围内；包 README 记录了该限制。
+
+**代隔离，无交错。** 每次尝试构建一个全新的 transport 和 `Client`（SDK 将一个 Protocol 绑定到一个 transport 上终身使用）。每个监督器内部有一个队列将所有 `syncTools` 调用串行化——跨所有代的初始同步和 `list_changed` 再同步——`isCurrent` 栅栏使过时的代变为惰性，从而确保不会有两次同步交错执行 dispose 上一代/注册下一代的切换（否则会对同一代执行两次 dispose 并泄漏另一代）。该队列还消除了一个先前存在的竞态：两次快速的 `list_changed` 通知同时触发重新同步。严格启动注册由激活尝试本身显式拥有，而非由首个入队者拥有；提前到达的 `list_changed` 采用故障隔离的再同步语义，不能消费 `failOnStartupError`。失败信号按代幂等：一次连接拒绝与其自身 transport 关闭竞态时，仅调度恰好一次重试。失败尝试只有在 `Client.close()` 结算且 transport 报告 `onclose` 后才能进入退避；对 stdio 而言，`onclose` 证明子进程已退出；若关闭信号始终未到，则在 SDK 的有界终止窗口结束后停止重连，而不是允许两个服务器进程重叠运行。dispose 使用同一个有界关闭信号屏障；若关停未完成则予以报告，且绝不重启。
+
+**有界退避与故障预算。** 延迟从 `initialDelayMs` 起逐次翻倍，上限为 `maxDelayMs`。一次故障期间共享 `maxAttempts` 次连续失败尝试的预算；耗尽后注销该服务器的工具、以 error 级别记录日志并停止，直到 dispose 或重新加载。连接在存活超过稳定窗口——即 `maxDelayMs`，作为最长退避间隔从配置推导得出而非作为第五个独立调参项——之后重置预算；因此偶尔崩溃的服务器可无限恢复，而连接短暂成功后立即再次崩溃的循环无法将其预算洗白为重启风暴。
+
+**配置与解析。** 两种传输均接受 `reconnect { enabled, initialDelayMs, maxDelayMs, maxAttempts }` 配置，Schemastery 默认值为（启用、500ms、30s、10）。`resolveReconnectPolicy()` 是显式的解析步骤：它重新校验每个边界值和跨字段约束，因为程序化构造可能绕过 Schemastery，配置错误在加载时即令插件实例失败。
+
+**可观测状态。** 初始尝试或重试尝试失败时记录 `connection failed`，已建立的代结束时记录 `connection lost`；重试的 warn 日志包含尝试次数和延迟，恢复以 info 级别记录，最终失败和禁用重连时的断连以 error 级别记录。故障期间，上一个正常代保持注册，对其工具的调用返回失败——确定性公开名称意味着恢复后未变化的工具列表会复现相同的定义，保持模型可见 schema 前缀稳定而非反复抖动。设置 `reconnect.enabled: false` 后，断连保持 v1 的手动恢复行为。
+
+**资源释放。** dispose 翻转栅栏、取消待执行的定时器、关闭当前 client，然后等待正在进行的尝试和同步队列完成后再注销工具——完全停稳，而非仅发出停止请求。重连定时器使用 unref，因此等待中的退避不会阻止进程正常退出。
+
+## 曾考虑的替代方案
+
+**连续失败计数器，每次成功连接即重置。** 否决：连接短暂成功后立即崩溃的循环服务器会在每个周期重置预算并永远重启——恰恰是失败上限旨在防止的重启风暴。基于运行时间的重置能区分已恢复的服务器与循环崩溃的服务器，无需新增配置。
+
+**跨重连复用同一个 SDK `Client`。** Protocol 在关闭时清除其 transport，技术上可以再次连接，但 SDK 自身的指导方针是每个 Protocol 实例对应一次连接，且复用会将通知处理器和已协商的能力状态带入新的服务器实例。每代创建全新 `Client` 加 `isCurrent` 栅栏的方式无歧义。
+
+**断连时立即注销工具，恢复时重新注册。** 否决：短暂故障会使模型可见工具列表抖动（每次崩溃触发两次 schema 前缀失效），而无任何信息增益；失败的调用已足以标示故障，恢复时的切换按代原子执行。工具仅在最终失败时注销，确保永久死亡的服务器不会泄漏永久失效的工具。
+
+**将 Streamable HTTP 请求失败路由到监督器。** 暂不采纳：HTTP 传输已使用自己的退避机制重连其 SSE 流，逐请求错误并不意味着服务器已死，且 harness 没有可重新拉起的子进程。transport 关闭仍是唯一触发条件。
+
+**通过 Loader/HMR 机制重启，而非使用插件内监督器。** 否决：Loader 负责配置驱动的重组合，而非运行时健康管理。插件通过 Loader 重启自身会混淆配置代与连接代，并丢失逐故障预算。
+
+## 测试
+
+单元测试（`tests/reconnect.spec.ts`，mock SDK）：恢复在不产生重复或泄漏的前提下切换代并服务恢复后的调用、诊断区分初始或重试尝试失败与已建立连接丢失、严格启动注册在连接前收到 `list_changed` 通知后仍然生效、初始化失败会等待旧代的关闭信号，若该信号始终未到则停止重连、dispose 同样等待同一关闭信号，并在有界等待到期时报告关停未完成、失败上限注销工具并停止、dispose 取消待执行的退避并使进行中的同步完全停稳、dispose 后的关闭不调度任何操作、禁用模式保持 v1 行为、稳定窗口重置预算而崩溃循环耗尽预算、双重失败信号仅调度一次重试、过时的代和处理器为惰性、`resolveReconnectPolicy` 拒绝每个无效边界值。E2E（`tests/mcp-client.e2e.ts`，无需密钥）：fixture 服务器新增了一个 `crash` 工具（先回复再退出）；真实进程测试证明 stdio 崩溃端到端恢复，以及在故障期间卸载插件能立即停止重连。快照：刻意不做，原因与原 Agent Note 相同——重连不引入新的展示形态，而在快照组合中 spawn 崩溃服务器会使回放依赖时序。
+
+## 后果
+
+- 崩溃的 stdio MCP 服务器无需人工干预即可恢复：有界退避、重新发现、原子代切换。默认策略对一次故障大约重试 2.5 分钟后放弃。
+- 连接状态确实比一次性连接更复杂——v1 刻意回避的部分可用窗口现已存在（故障期间已注册工具返回失败），集中在一个模块中并命名了所有不变式。
+- `reconnect` 是两种传输上的新配置表面，稳定窗口刻意从 `maxDelayMs` 推导；将其设为独立可调参数是兼容的未来变更。
+- 最终失败后或禁用重连时，插件保持加载状态但无（或失败的）工具，直到重新加载——行为是刻意的且有日志记录，确保长期故障的服务器不能永远重启。