refactor(agent-loop): simplify observable state machine

2026-07-24 21:18:48 +08:00
parent fb0ef82aa6
commit b73eb7663c
131 changed files with 2011 additions and 4292 deletions
@@ -10,7 +10,7 @@ Status: implemented
 - `streamBlocks()` — a "convenience view" that runs the chunks through a `BlockAssembler` and yields completed `ContentBlock`s in stream order ([index.ts:137-144](../../../../packages/llm/llm/src/index.ts)).
 - `generate()` — one fully-assembled `GenerateResult`, dispatched through a second `llm/generate` waterfall ([index.ts:151-157](../../../../packages/llm/llm/src/index.ts)).

-The only production consumer of the LLM service is the agent loop, and it uses `stream()` exclusively — feeding raw chunks through its own `BlockAssembler` so it can log chunks for replay fidelity while assembling in parallel ([packages/core/agent-loop/src/loop.ts](../../../../packages/core/agent-loop/src/loop.ts), the `ctx.llm.stream(req)` step). Grepping `streamBlocks` and `ctx.llm.generate` across `packages/*/src` and `examples/*/src` finds no production callers. The references are the service methods, docs, and tests; adapter tests use `generate()` as a convenient driver, but they can hand-drain `stream()` through the same assembler helper without preserving a public production API.
+The only production consumer of the LLM service is the agent loop, and it uses `stream()` exclusively — feeding raw chunks through its own `BlockAssembler` so it can log chunks for replay fidelity while assembling in parallel ([packages/core/agent-loop/src/agent.ts](../../../../packages/core/agent-loop/src/agent.ts), the `ctx.llm.stream(req)` step). Grepping `streamBlocks` and `ctx.llm.generate` across `packages/*/src` and `examples/*/src` finds no production callers. The references are the service methods, docs, and tests; adapter tests use `generate()` as a convenient driver, but they can hand-drain `stream()` through the same assembler helper without preserving a public production API.

 This is the [drop-mutable-session-summary](2026-06-19-drop-mutable-session-summary.md) pattern: assembled-view APIs with tested contracts, consumed by tests rather than production. They were built speculatively for consumers that do not care about token-level deltas, but the one real consumer cares about deltas precisely so it can persist high-fidelity replay data.

@@ -4,7 +4,7 @@ Status: implemented

 ## Problem

-The loop records every model token delta as a durable `assistant/chunk` session event AND emitted a parallel live `agent/stream-chunk` Cordis event carrying the identical data. In `packages/core/agent-loop/src/loop.ts` the two sat one line apart:
+The loop records every model token delta as a durable `assistant/chunk` session event AND emitted a parallel live `agent/stream-chunk` Cordis event carrying the identical data. In `packages/core/agent-loop/src/agent.ts` the two sat one line apart:

 ```ts ignore-check
 const chunkEvent = session.append('assistant/chunk', { turn, step, chunk })
@@ -4,7 +4,7 @@ Status: implemented

 ## Problem

-`agent/steering` was the last remaining transient mirror of a durable session event. The loop's steering drain appends the durable `steering/message { turn, content, source }` and, on the very next line, emitted `agent/steering(agent, turn, content, source)` — the identical fact as a fire-and-forget event (`packages/core/agent-loop/src/loop.ts`, `drainSteering`). It had zero production listeners: the only subscriber anywhere was a loop regression test asserting the emit carried `source` — the same fact the durable event already records one line above.
+`agent/steering` was the last remaining transient mirror of a durable session event. The loop's steering drain appends the durable `steering/message { turn, content, source }` and, on the very next line, emitted `agent/steering(agent, turn, content, source)` — the identical fact as a fire-and-forget event (`packages/core/agent-loop/src/agent.ts`, `drainOutbox`). It had zero production listeners: the only subscriber anywhere was a loop regression test asserting the emit carried `source` — the same fact the durable event already records one line above.

 `agent/steering` duplicated the immediately preceding durable `steering/message` with the same payload. `agent/queued` remains the live-only signal because it fires before persistence and covers work that may be cancelled before entering the log.

@@ -12,7 +12,7 @@ Steering carries real production traffic — the hook bridges' turn-continuation

 ## Decision

-`agent/steering` is removed from the agent event taxonomy: the declaration in `packages/core/agent/src/types.ts` (and its mention in the live-events JSDoc list there), the emit in `drainSteering` (whose then-unused `ctx` parameter went with it), the row in `packages/core/agent/README.md`, and the emit line in the loop-pseudocode blocks (the `packages/core/agent-loop/src/loop.ts` module doc and [architecture.md](../../../../docs/architecture.md)); the cordis catalog is regenerated without it. The one regression test pins source preservation on the durable `steering/message` event — the fact it pins lives on the log.
+`agent/steering` is removed from the agent event taxonomy: the declaration in `packages/core/agent/src/types.ts` (and its mention in the live-events JSDoc list there), the emit in `drainOutbox`, the row in `packages/core/agent/README.md`, and the emit line in the loop-pseudocode blocks (the `packages/core/agent-loop/src/agent.ts` module doc and [architecture.md](../../../../docs/architecture.md)); the cordis catalog is regenerated without it. The one regression test pins source preservation on the durable `steering/message` event — the fact it pins lives on the log.

 Three implemented Agent Notes stated the retention, and each is amended per [implemented/AGENTS.md](../AGENTS.md) to point here as the record of the removal: the [boundary Agent Note](2026-06-20-remove-agent-boundary-mirror-events.md)'s retained-list entry, the [stream-chunk Agent Note](2026-07-02-remove-stream-chunk-mirror.md)'s scope clause, and the [event-domain-semantics Agent Note](../architecture/2026-06-30-event-domain-semantics.md)'s transient-emit enumeration.

@@ -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
+2026-07-24-agent-loop-observable-state-machine.md: 7a3607c5bcf6edece07433acd766a6dece551d1c
+2026-07-24-agent-loop-observable-state-machine.zh.md: 7b53eb6087e99ce4c84caae7b94d275c0b6a10db
@@ -0,0 +1,58 @@
+# Agent Note: Collapse agent-loop events around the observable state machine
+
+Status: implemented
+
+English | [中文](2026-07-24-agent-loop-observable-state-machine.zh.md)
+
+## Problem
+
+The agent loop exposed its control flow as a large set of Cordis events. Separate `pre-step` and `post-step` checkpoints bracketed a step, `session-prefix` and `step-result` transformed request and response messages, `request-error` decided whether a failed request retried inside its turn, and `turn-continuation` plus `turn-stop` composed competing continuation decisions.
+
+Those events made internal phases public even when the durable session log already owned the corresponding turn and step facts. They also mixed two extension models: some listeners observed a boundary and issued an agent command, while others returned control decisions that the loop interpreted. Understanding the public machine therefore required reconstructing event order, waterfall precedence, and special terminal overrides together.
+
+Agent lifetime, whole-agent activity, inbox-item progress, and per-turn settlement are independent state dimensions. Treating them as one status or one linear callback sequence makes ordinary questions ambiguous: an agent can remain `running` across several turns, an accepted item can be discarded without opening a turn, and one turn can settle while later work keeps the agent active.
+
+## Decision
+
+The public contract exposes four orthogonal state dimensions:
+
+- Registration lifetime is the `agent/created` to `agent/disposed` interval. Disposal is the terminal registry edge, not an `AgentStatus`.
+- Whole-agent activity is `AgentStatus = 'idle' | 'running'`. Consecutive turns may share one `running` interval.
+- A FIFO-backed message progresses from `agent/inbox/enqueue` to exactly one `agent/inbox/dequeue` or `agent/inbox/discard`, correlated by `AgentMessageId`. The inbox events describe acceptance, claim, and removal rather than turn completion.
+- A claimed turn passes through prompt admission and zero or more request steps. An automatic retry closes the failed turn and immediately opens another; `agent/idle` reports only the terminal turn in that chain and remains distinct from the whole-agent transition to `status === 'idle'`.
+
+The loop keeps five machine extension events. `agent/prompt-submit` admits, rewrites, or blocks a claimed prompt. `agent/step` is the single awaited between-steps checkpoint and runs before every request is derived. `agent/request` is the waterfall for the frozen call configuration; the configuration comes only from `await next()`, not from a duplicate positional argument. `agent/request-error` serializes ownership of awaited model-request recovery. `agent/stopping` runs when the turn otherwise has no work left; a listener that needs another step records real steering with `agent.steer()`, and the loop decides from that data after all listeners settle.
+
+Continuation and termination are data rather than returned control enums. Tool calls and accepted steering require another step. A tool result carrying `concludesTurn` ends the tool loop at its step. The loop does not expose general `ContinuationDecision` or terminal-stop return channels.
+
+A model-request failure closes its step, then enters `agent/request-error` with the exact error, normalized `LlmFailure`, and live turn signal. A listener that owns recovery repairs state, calls `agent.retry()`, and returns without delegating. The loop closes the failed turn and opens one retry turn over that state without an intervening idle notification; retry is not another step inside the failed turn. `agent/idle` reports the terminal outcome, and `agent/error` remains the live error notification for consumers that report failures independently of turn settlement.
+
+The event taxonomy removes `agent/pre-step`, `agent/post-step`, `agent/session-prefix`, `agent/step-result`, `agent/turn-continuation`, and `agent/turn-stop`. Durable turn and step boundaries remain session events. Model-facing additions use logged message channels, request configuration uses `agent/request`, response content is recorded as assembled, failed-request recovery uses `agent/request-error` plus `agent.retry()`, and end-of-turn continuation uses `agent/stopping` plus steering.
+
+## Alternatives considered
+
+**Keep the fine-grained event sequence.** This preserves a dedicated interception point for every internal phase, including request-only prefixes, assistant-message rewriting, post-step work, in-turn request recovery, and terminal stop overrides. It also makes the loop's private sequencing a permanent public contract and lets overlapping seams express conflicting decisions. The decision accepts the lost interception points in exchange for one boundary per supported extension responsibility.
+
+**Represent disposal as a third `AgentStatus`.** This gives retained handles a terminal status value but duplicates the registry lifecycle already expressed by `agent/disposed`. The decision keeps `AgentStatus` about live activity and makes registration lifetime a separate dimension.
+
+**Return a retry decision from `agent/request-error`.** A returned instruction duplicates the existing `agent.retry()` command and requires the loop to carry policy history across attempts. The waterfall remains useful for ordered ownership: an unhandled listener delegates, while a handling listener performs its awaited repair, calls `agent.retry()`, and stops delegation.
+
+**Mirror durable turn and step boundaries as agent events.** This gives live consumers a second event stream for the same facts. The decision keeps the session log as the source of truth and exposes only extension checkpoints or live-only facts that the durable stream cannot carry.
+
+## Consequences
+
+The observable machine is smaller and compositional: registration lifetime, activity, item progress, and terminal settlement can be followed independently. In particular, `agent/idle` does not imply `agent.status === 'idle'`; it reports the terminal turn of one drain chain, while `agent/status` reports whether the whole agent is active.
+
+Plugins no longer rewrite every phase of the loop. There is no request-only message prefix, assistant-message transform, post-step checkpoint, generic continuation enum, generic terminal-stop result, or in-turn request retry. Extensions use the remaining owned channels instead of recreating those phases.
+
+Continuation plugins publish durable steering rather than returning an unlogged reason. Recovery plugins act after the failed step and explicitly schedule another turn through `agent.retry()`. This makes every attempt a complete turn while keeping asynchronous repair and policy ownership at one narrow waterfall boundary.
+
+The inbox lifecycle complements, rather than replaces, the durable session log. `AgentMessageId` correlates acceptance with claim or discard; turn and step numbers, messages, tool activity, and terminal reasons remain session facts.
+
+## Related
+
+- [Unify agent delivery on send(target × wakeup) and coalesce injected context into user/message](../architecture/2026-07-22-unified-send-and-coalesced-user-messages.md)
+- [Remove implicit batching from ordinary sends](2026-07-17-one-send-one-turn.md)
+- [Microkernel event taxonomy](../architecture/2026-06-11-microkernel-event-taxonomy.md)
+- [Bounded LLM request recovery](../architecture/2026-06-21-bounded-llm-request-recovery.md)
+- [Reconstructable requests](../architecture/2026-07-05-reconstructable-requests.md)
@@ -0,0 +1,58 @@
+# Agent Note: 围绕可观察状态机收拢 agent loop（智能体循环）事件
+
+Status: implemented
+
+[English](2026-07-24-agent-loop-observable-state-machine.md) | 中文
+
+## 问题
+
+agent loop 曾将其控制流暴露为大量 Cordis 事件。`pre-step` 和 `post-step` 两个独立检查点分列步骤前后，`session-prefix` 和 `step-result` 分别变换请求消息与响应消息，`request-error` 决定失败的请求是否在当前轮次内重试，`turn-continuation` 与 `turn-stop` 则组合相互竞争的继续执行决策。
+
+即使持久会话日志已经记录了对应的轮次与步骤事实，这些事件仍会将内部阶段公开。它们还混用了两种扩展模型：部分监听器观察边界并发出 agent 命令，另一些监听器则返回由循环解释的控制决策。因此，要理解公开状态机，必须同时还原事件顺序、waterfall（瀑布式事件）优先级和特殊的终止覆盖规则。
+
+agent 生命周期、agent 整体活动状态、收件箱条目的进度以及每轮次的结算，是彼此独立的状态维度。若将它们视为一个状态或一条线性回调序列，常见问题就会产生歧义：agent 可以在多个轮次之间持续保持 `running`；已接受的条目可以不启动轮次就被丢弃；一个轮次可以完成结算，而后续工作仍让 agent 保持活动。
+
+## 决策
+
+公开契约暴露四个正交的状态维度：
+
+- 注册生命周期是从 `agent/created` 到 `agent/disposed` 的区间。dispose（资源释放）是注册表的终止边界，而不是一种 `AgentStatus`。
+- agent 整体活动状态为 `AgentStatus = 'idle' | 'running'`。连续多个轮次可以共用同一个 `running` 区间。
+- 由 FIFO 支撑的消息从 `agent/inbox/enqueue` 开始，最终必然进入 `agent/inbox/dequeue` 或 `agent/inbox/discard` 二者之一，并通过 `AgentMessageId` 关联。收件箱事件描述接受、领取和移除，而不是轮次完成。
+- 已领取的轮次经过提示词准入和零个或多个请求步骤。自动重试会关闭失败轮次并立即开启另一个轮次；`agent/idle` 只报告该重试链的终态轮次，且仍不同于 agent 整体转换到 `status === 'idle'`。
+
+循环保留五个状态机扩展事件。`agent/prompt-submit` 对已领取的提示词执行准入、改写或阻断。`agent/step` 是步骤之间唯一需要等待的检查点，在每次派生请求前运行。`agent/request` 是冻结调用配置所用的 waterfall；配置只能来自 `await next()`，不再通过重复的位置参数提供。`agent/request-error` 串行确定需要等待的模型请求恢复由谁负责。当轮次原本已经没有剩余工作时，`agent/stopping` 运行；需要再执行一个步骤的监听器使用 `agent.steer()` 记录真实的 steering（中途引导），循环在所有监听器完成后根据这份数据作出决定。
+
+是否继续和终止执行由数据表达，不再由返回的控制枚举表达。工具调用和已接受的 steering 要求再执行一个步骤。携带 `concludesTurn` 的工具结果会在其所属步骤终止工具循环。循环不再暴露通用的 `ContinuationDecision` 或终止停止返回通道。
+
+模型请求失败会先关闭当前步骤，再携带准确错误、标准化 `LlmFailure` 和仍有效的轮次信号进入 `agent/request-error`。负责恢复的监听器修复状态、调用 `agent.retry()`，并停止继续委托。循环会关闭失败轮次，并基于该状态开启一个重试轮次，中间不发布空闲通知；重试不是失败轮次内的另一个步骤。`agent/idle` 报告终态结果；对于需要脱离轮次结算单独报告失败的消费方，`agent/error` 仍作为实时错误通知保留。
+
+事件分类体系移除了 `agent/pre-step`、`agent/post-step`、`agent/session-prefix`、`agent/step-result`、`agent/turn-continuation` 和 `agent/turn-stop`。持久的轮次与步骤边界仍由会话事件记录。面向模型的新增内容使用有日志记录的消息通道，请求配置使用 `agent/request`，响应内容按组装后的原样记录，失败请求恢复使用 `agent/request-error` 加 `agent.retry()`，轮次结束时是否继续则使用 `agent/stopping` 加 steering 表达。
+
+## 考虑过的替代方案
+
+**保留细粒度事件序列。** 这样可以为每个内部阶段保留专用拦截点，包括仅用于请求的前缀、助手消息改写、步骤后处理、轮次内请求恢复以及终止停止覆盖。但这也会使循环的私有执行顺序成为永久的公开契约，并允许相互重叠的 seam 表达彼此冲突的决策。当前决策接受这些拦截点的缺失，以换取每项受支持的扩展职责仅对应一个边界。
+
+**将 dispose 表示为第三种 `AgentStatus`。** 这样会让仍被持有的句柄得到一个终止状态值，但也会重复表达 `agent/disposed` 已经体现的注册表生命周期。当前决策让 `AgentStatus` 只表示活动中 agent 的状态，并将注册生命周期作为独立维度。
+
+**让 `agent/request-error` 返回重试决策。** 返回指令会与现有的 `agent.retry()` 命令重复，还要求循环跨尝试携带策略历史。Waterfall 只需保留有序归属能力：未处理的监听器继续委托，负责处理的监听器完成可等待的修复、调用 `agent.retry()`，然后停止委托。
+
+**将持久的轮次与步骤边界映射为 agent 事件。** 这样会为同一事实向实时消费方提供第二条事件流。当前决策将会话日志保留为真源，仅暴露扩展检查点或持久事件流无法承载的纯实时事实。
+
+## 影响
+
+可观察状态机更小，也更容易组合：注册生命周期、活动状态、条目进度和终态结算可以分别追踪。尤其是，`agent/idle` 并不意味着 `agent.status === 'idle'`；前者报告一次排空链的终态轮次，`agent/status` 则报告整个 agent 是否处于活动状态。
+
+插件不再能够改写循环的每个阶段。不再提供仅用于请求的消息前缀、助手消息变换、步骤后检查点、通用的继续执行枚举、通用的终止停止结果或轮次内请求重试。扩展改用剩余的归属明确的通道，而不是重新构造这些阶段。
+
+负责继续执行的插件发布可持久化的 steering，而不是返回未记录到日志中的原因。恢复插件在失败步骤结束后处理错误，并通过 `agent.retry()` 显式安排另一个轮次。这样，每次尝试都会成为完整轮次，同时异步修复和策略归属集中在一个狭窄的 waterfall 边界。
+
+收件箱生命周期用于补充持久会话日志，而非取代它。`AgentMessageId` 将接受操作与领取或丢弃操作关联起来；轮次编号与步骤编号、消息、工具活动和终止原因仍属于会话事实。
+
+## 相关内容
+
+- [统一通过 send(target × wakeup) 交付 agent 消息，并将注入上下文合并到 user/message](../architecture/2026-07-22-unified-send-and-coalesced-user-messages.md)
+- [移除普通发送中的隐式批处理](2026-07-17-one-send-one-turn.md)
+- [微内核事件分类体系](../architecture/2026-06-11-microkernel-event-taxonomy.md)
+- [有界 LLM 请求恢复](../architecture/2026-06-21-bounded-llm-request-recovery.md)
+- [可重建的请求](../architecture/2026-07-05-reconstructable-requests.md)