docs(agent): propose context injection semantics

.agents/notes/proposed/architecture/2026-07-24-separate-context-injection-from-turn-execution.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
+2026-07-24-separate-context-injection-from-turn-execution.md: 652c3d410ab625d91a828f854bce302adcb0c9e0
+2026-07-24-separate-context-injection-from-turn-execution.zh.md: 1064e7a869ab9ea46c0145eb010119894a03aacf

.agents/notes/proposed/architecture/2026-07-24-separate-context-injection-from-turn-execution.md

@@ -0,0 +1,75 @@
+# Agent Note: Separate context injection from turn execution
+
+Status: proposed
+
+English | [中文](2026-07-24-separate-context-injection-from-turn-execution.zh.md)
+
+## Problem
+
+The agent API currently represents supplementary model-facing input in three overlapping ways: callers attach `HookContext[]` through `SendOptions.contexts`, interception and tool hooks return `additionalContexts`, and plugins call `agent.inject()`. These paths eventually write context into the same model history, but they carry different placement, metadata, admission, queue, and turn-lifecycle rules.
+
+Atomic attachment to an inbox message forces the loop to preserve context through prompt admission, steering conversion, cancellation, and terminal discard. `prompt-prefix` placement then combines context and the direct prompt into one event, requiring a model-hidden envelope so transcript consumers can recover what the user actually wrote. The result makes outbox entries, session projection, and UI replay responsible for a distinction that belongs to the producer.
+
+Idle `inject()` exposes a second mismatch. Injection does not request model execution, yet the current implementation opens and closes a zero-step `injection` turn solely to satisfy the turn-enclosure invariant and obtain a durability checkpoint. A turn therefore sometimes means “run the agent loop” and sometimes means “persist context without running it.”
+
+`HookContext` also names its producer rather than its role. The value may come from a native plugin, a hook bridge, prompt admission, or tool post-processing. Its stable meaning is simply additional model-facing context with provenance.
+
+## Proposal
+
+Make `inject()` the only caller-facing operation for adding supplementary model-facing input, and define a turn exclusively as one execution of the model loop.
+
+Remove `SendOptions.contexts`. A caller that owns context delivers it with `inject()` and independently submits the direct message with `send()` or `steer()`. Rename `HookContext` to `AdditionalContext`; retain only `content` and `source`, and remove placement and model-hidden metadata from this shared shape.
+
+Prompt and tool extension points may still return `additionalContexts`. These values are outputs of the extension point, not attachments captured from a caller's inbox item. Prompt admission runs before `run()` opens a turn. An allowed prompt enters the outbox together with its returned additional contexts; a blocked prompt writes neither and opens no turn. Tool-produced additional contexts enter the same outbox after the corresponding tool results.
+
+Every additional context becomes an independent `user/message` whose `source` records provenance. Remove `context/message`, prompt-prefix placement, the stable request delimiter, and the prompt envelope. Transcript and UI consumers distinguish direct user messages from injected context by `source`, not by recovering a hidden direct-prompt field from combined model content.
+
+## Injection lifecycle
+
+When a turn is open, `inject()` stages the context in the loop outbox. The loop drains the outbox at a safe step boundary, preserving tool protocol adjacency: a context accepted during an assistant tool-call batch appears only after that batch's complete ordered results. Taking the outbox as a whole makes steering and injected context accepted for one boundary visible to the same following request.
+
+When no turn is open, `inject()` appends its `user/message` immediately and starts a session flush. It does not increment turn numbering, emit `turn/start` or `turn/end`, change agent status, or run the model. The synchronous API still returns before the asynchronous flush settles; `whenIdle()` and agent disposal include outstanding idle-injection flushes in their quiescence boundary.
+
+A failed idle flush has no legitimate turn or step coordinates. It is reported through logging or a persistence-owned error surface, not by inventing an `agent/error` payload for a nonexistent turn. The in-memory event remains accepted and a later flush may retry persistence.
+
+The session invariant therefore permits `user/message` between turns while continuing to require turn enclosure for execution events, steering, assistant output, tools, and package-added events by default. Persistence, recovery, resume, fork, and compaction code must treat a valid out-of-turn `user/message` as committed session history rather than an interrupted or discardable turn tail.
+
+## Extension and caller semantics
+
+`PromptDecision.content` continues to replace only the direct prompt. `PromptDecision.additionalContexts` and tool-result `additionalContexts` retain FIFO order and individual provenance, but no longer select placement. A waterfall listener that delegates with `next()` must preserve downstream prompt content and additional contexts unless it intentionally returns replacements.
+
+Caller-driven injection and hook-produced additional context deliberately have different admission ownership. A hook's additional contexts materialize only after that hook allows the prompt or tool result. A caller that invokes `inject(context)` and then `send(prompt)` has already committed context independently; if prompt admission later blocks the prompt, the injected context remains in history. Callers requiring all-or-nothing domain behavior must perform their own preparation before either operation or expose a domain-specific admission seam.
+
+Cross-session references follow the ordinary composition: the host prepares the snapshot, injects it with session-reference provenance, then sends or steers the readable direct prompt. The target log contains two simple messages, so later source mutation cannot change replay and transcript consumers do not need a prompt envelope. This supersedes the attachment mechanism in the [cross-session reference decision](../../implemented/feature/2026-07-21-cross-session-references.md) while retaining its snapshot and trust-boundary rules.
+
+This proposal preserves the caller-owned framing decision from [unwrapped injected content](../../implemented/simplification/2026-07-20-unwrap-injected-content-envelopes.md), the one-item turn rule from [one send, one turn](../../implemented/simplification/2026-07-17-one-send-one-turn.md), and narrows the [turn-enclosure decision](../../implemented/architecture/2026-06-15-turn-enclosure-invariant.md) so turns enclose execution rather than every session event.
+
+## Alternatives considered
+
+**Keep `SendOptions.contexts` as an atomic attachment.** This preserves all-or-nothing delivery when prompt admission blocks, but it keeps context inside inbox lifecycle state and requires every queue transition and observation event to carry it. The generic agent API should not encode a domain transaction that most callers can express as context injection followed by message delivery.
+
+**Keep a distinct `context/message` session event.** A separate event makes the out-of-turn exception narrower, but user-role model input would again have two event types with identical projection. `user/message.source` already carries the distinction needed by policy, transcript, and replay consumers.
+
+**Keep one-shot turns for idle injection.** This retains universal turn enclosure and a convenient flush boundary, but it makes turn counts and turn observers report work that never ran the model. Durability is an independent session concern and can be awaited without fabricating execution.
+
+**Keep `prompt-prefix` as an optional placement.** Prefix baking can make the context and request appear in one provider message, but it introduces a second representation of the direct prompt and spreads placement handling across admission, steering, logging, replay, and UI code. Producers that require textual framing may include it in their own context content.
+
+**Let hooks call `inject()` directly instead of returning additional contexts.** Direct injection would erase the extension point's admission ownership: a listener could append context before a downstream listener blocks the operation. Returning `additionalContexts` keeps the waterfall result authoritative while sharing the same post-admission outbox path.
+
+## Acceptance criteria
+
+- `SendOptions` and steering inbox records contain no attached contexts; `agent/queued` reports only the retained message and steering facts.
+- `AdditionalContext` replaces `HookContext` across prompt interception, tool execution, hook bridges, guards, and context producers, with only `content` and `source`.
+- Prompt-prefix placement, prompt envelopes, and `context/message` are absent from public types, durable events, projection, and UI replay.
+- Idle `inject()` appends and flushes one sourced `user/message` without a turn or model call; `whenIdle()` and disposal await the flush.
+- Active-turn injection and hook-produced contexts drain at safe boundaries after complete tool-result batches and before the request that consumes them.
+- Blocked prompt admission opens no turn and appends neither the prompt nor hook-produced additional contexts; independently injected caller context remains.
+- Unit, persistence/resume, invariant, ACP/TUI replay, and keyless assembled-application snapshots cover the new event order and durability semantics.
+
+## Risks
+
+- Allowing one surface event outside turns weakens a simple invariant and may expose hidden assumptions in persistence scanning, crash repair, forking, compaction, and session queries.
+- Consecutive user-role messages replace one baked prompt message; provider adapters and cache behavior must accept and preserve that ordering.
+- `inject()` followed by a blocked `send()` leaves context without its intended direct prompt unless the caller accepts the independent-commit contract.
+- A synchronous injection API cannot return flush failure. Logging alone is less structured than `agent/error`, while adding a new persistence event solely for this case may create another unnecessary seam.
+- Removing attachment, placement, metadata, envelopes, and a durable event type is a broad pre-release migration that must update every producer and consumer atomically.

.agents/notes/proposed/architecture/2026-07-24-separate-context-injection-from-turn-execution.zh.md

@@ -0,0 +1,75 @@
+# Agent Note: 将上下文注入与轮次执行分离
+
+Status: proposed
+
+[English](2026-07-24-separate-context-injection-from-turn-execution.md) | 中文
+
+## 问题
+
+agent API 目前用三种相互重叠的方式表示面向模型的补充输入：调用方通过 `SendOptions.contexts` 附加 `HookContext[]`，拦截钩子和工具钩子返回 `additionalContexts`，插件则调用 `agent.inject()`。这些路径最终都会把上下文写入同一份模型历史，但各自携带不同的放置、元数据、准入、队列和轮次生命周期规则。
+
+将上下文原子附加到收件箱消息后，agent loop（智能体循环）必须让上下文跟随提示词准入、steering（中途引导）转换、取消和终止丢弃的完整生命周期。`prompt-prefix` 放置方式又会把上下文与直接提示词合并为一个事件，因此 transcript（文本记录）消费方需要依赖模型不可见的封套，才能还原用户实际输入。这样一来，outbox 条目、会话投影和 UI 回放都必须处理本应由生产方负责的区分。
+
+空闲状态下的 `inject()` 还暴露了另一处语义错位。注入并不请求模型执行，但当前实现仅为了满足轮次封闭不变量并获得持久性检查点，就会打开并关闭一个零步骤的 `injection` 轮次。于是，轮次有时表示「运行 agent loop」，有时却表示「不运行 agent，仅持久化上下文」。
+
+`HookContext` 的名字也描述了生产方，而非该值的职责。它可能来自原生插件、hook bridge、提示词准入或工具后处理；其稳定含义只是带来源信息的额外模型上下文。
+
+## 提案
+
+将 `inject()` 设为调用方添加补充模型输入的唯一操作，并把轮次严格定义为一次模型循环执行。
+
+移除 `SendOptions.contexts`。拥有上下文的调用方通过 `inject()` 交付上下文，再独立使用 `send()` 或 `steer()` 提交直接消息。将 `HookContext` 重命名为 `AdditionalContext`；这个共享结构只保留 `content` 和 `source`，移除放置方式与模型不可见元数据。
+
+提示词和工具扩展点仍可返回 `additionalContexts`。这些值是扩展点的输出，而不是从调用方收件箱条目捕获的附件。提示词准入在 `run()` 打开轮次之前执行。提示词获准后，它与返回的额外上下文一同进入 outbox；提示词被阻止时，两者都不写入，也不打开轮次。工具产生的额外上下文则在对应工具结果之后进入同一个 outbox。
+
+每项额外上下文都成为独立的 `user/message`，并由 `source` 记录来源。移除 `context/message`、prompt-prefix 放置方式、稳定请求分隔符和提示词封套。transcript 与 UI 消费方通过 `source` 区分直接用户消息和注入上下文，无需从合并后的模型内容中恢复隐藏的直接提示词字段。
+
+## 注入生命周期
+
+轮次打开时，`inject()` 将上下文暂存在 loop outbox 中。agent loop 会在安全的步骤边界排空 outbox，同时保持工具协议要求的相邻关系：在助手工具调用批次期间接纳的上下文，只能出现在该批次所有有序结果之后。系统整体取走 outbox，确保同一边界接纳的 steering 和注入上下文对后续同一次请求可见。
+
+没有打开的轮次时，`inject()` 会立即追加对应的 `user/message` 并启动会话刷新。它不会增加轮次编号、发出 `turn/start` 或 `turn/end`、改变 agent 状态，也不会运行模型。同步 API 仍会在异步刷新完成前返回；`whenIdle()` 和 agent dispose（资源释放）会把尚未结束的空闲注入刷新纳入静止边界。
+
+空闲刷新失败时不存在合法的轮次或步骤坐标。系统通过日志或持久化所属的错误接口报告该失败，而不是为不存在的轮次伪造 `agent/error` 载荷。内存中的事件仍已接纳，后续刷新可以重试持久化。
+
+因此，会话不变量允许 `user/message` 位于两个轮次之间，同时继续要求执行事件、steering、助手输出、工具事件以及默认的包扩展事件均受轮次边界约束。持久化、恢复、resume、fork、压缩和查询逻辑必须把合法的轮次外 `user/message` 当作已提交会话历史，而不是中断轮次或可丢弃的日志尾部。
+
+## 扩展点与调用方语义
+
+`PromptDecision.content` 仍只替换直接提示词。`PromptDecision.additionalContexts` 和工具结果的 `additionalContexts` 保留 FIFO 顺序及各自来源，但不再选择放置方式。waterfall（瀑布式事件）监听器调用 `next()` 委托时，必须保留下游返回的提示词内容和额外上下文，除非它有意返回替代值。
+
+调用方主动注入与钩子产生的额外上下文具有不同的准入归属。钩子的额外上下文只会在该钩子允许提示词或工具结果后落入日志。调用方执行 `inject(context)` 后再执行 `send(prompt)` 时，上下文已独立提交；后续提示词准入即使阻止该提示词，注入上下文仍保留在历史中。需要领域级全有或全无语义的调用方，必须在执行任一操作前自行完成准备，或提供领域专用的准入 seam。
+
+跨会话引用使用普通组合方式：宿主先准备快照，以会话引用来源调用 `inject()`，再发送或 steer 可读的直接提示词。目标日志包含两条简单消息，因此来源会话后续变化不会改变回放，transcript 消费方也不需要提示词封套。本提案取代[跨会话引用决策](../../implemented/feature/2026-07-21-cross-session-references.md)中的附件机制，但保留其快照与信任边界规则。
+
+本提案保留[移除注入内容封套](../../implemented/simplification/2026-07-20-unwrap-injected-content-envelopes.md)确立的调用方自主管理框架原则，以及[一次 send、一个轮次](../../implemented/simplification/2026-07-17-one-send-one-turn.md)确立的单条目轮次规则；同时收窄[轮次封闭决策](../../implemented/architecture/2026-06-15-turn-enclosure-invariant.md)，使轮次约束执行过程，而不是约束所有会话事件。
+
+## 曾考虑的替代方案
+
+**保留 `SendOptions.contexts` 作为原子附件。** 提示词准入阻止消息时，这种方式能保留全有或全无交付，但也会让上下文继续成为收件箱生命周期状态的一部分，并迫使每次队列转换和观察事件携带它。大多数调用方都可以通过先注入上下文、再交付消息来表达需求，通用 agent API 不应内置领域事务。
+
+**保留独立的 `context/message` 会话事件。** 独立事件可以缩小轮次外事件的例外范围，但面向模型的 user-role 输入会再次拥有两个投影完全相同的事件类型。`user/message.source` 已能为策略、transcript 和回放消费方提供所需区分。
+
+**为空闲注入保留一次性轮次。** 这种方式能保留通用轮次封闭和方便的刷新边界，却会让轮次计数与轮次观察方报告从未运行模型的工作。持久性是独立的会话关注点，无需伪造执行即可等待。
+
+**保留 `prompt-prefix` 可选放置方式。** 前缀烘焙可以让上下文和请求位于同一条提供方消息中，但它会引入直接提示词的第二种表示，并把放置处理扩散到准入、steering、日志、回放和 UI 代码。需要文本框架的生产方可以直接把它写入自身上下文内容。
+
+**让钩子直接调用 `inject()`，而不是返回额外上下文。** 直接注入会破坏扩展点的准入归属：下游监听器阻止操作之前，上游监听器就可能已经追加上下文。返回 `additionalContexts` 能维持 waterfall 结果的最终权威性，同时复用准入后的 outbox 路径。
+
+## 验收标准
+
+- `SendOptions` 与 steering 收件箱记录不再包含附加上下文；`agent/queued` 只报告保留的消息和 steering 事实。
+- `AdditionalContext` 在提示词拦截、工具执行、hook bridge、guard 和上下文生产方中取代 `HookContext`，且只包含 `content` 与 `source`。
+- 公共类型、持久事件、投影和 UI 回放中均不存在 prompt-prefix 放置方式、提示词封套与 `context/message`。
+- 空闲 `inject()` 在不产生轮次或模型调用的情况下，追加并刷新一条带来源的 `user/message`；`whenIdle()` 和 dispose 会等待该刷新。
+- 活跃轮次注入和钩子产生的上下文会在完整工具结果批次之后的安全边界排空，并在消费它们的请求之前进入日志。
+- 被提示词准入阻止的消息不会打开轮次，也不会追加提示词或钩子产生的额外上下文；调用方此前独立注入的上下文仍保留。
+- 单元测试、持久化与 resume 测试、不变量测试、ACP/TUI 回放测试，以及无需密钥的组装应用快照覆盖新的事件顺序和持久性语义。
+
+## 风险
+
+- 允许一个表层事件位于轮次之外，会削弱一条简单不变量，并可能暴露持久化扫描、崩溃恢复、fork、压缩和会话查询中的隐含假设。
+- 两条连续的 user-role 消息会取代一条烘焙后的提示词消息；提供方适配器和缓存行为必须接受并保留这一顺序。
+- 如果调用方不能接受独立提交契约，`inject()` 后跟一个被阻止的 `send()` 会留下缺少预期直接提示词的上下文。
+- 同步注入 API 无法返回刷新失败。只记录日志的结构化程度低于 `agent/error`，但仅为此场景增加新的持久化事件也可能产生另一个不必要的 seam。
+- 移除附件、放置方式、元数据、封套和一种持久事件类型，是一次影响面较广的预发布迁移，必须原子更新所有生产方和消费方。