Stop mirroring the token stream as an agent event
停止将 token 流镜像为 agent 事件
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 }) chunkSeqs.push(chunkEvent.seq) ctx.emit('agent/stream-chunk', agent, turn, st
English
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/agent.ts the two sat one line apart:
const chunkEvent = session.append('assistant/chunk', { turn, step, chunk })
chunkSeqs.push(chunkEvent.seq)
ctx.emit('agent/stream-chunk', agent, turn, step, chunk) // ← the mirror
- Durable:
assistant/chunk: { turn, step, chunk }. - Live emit:
agent/stream-chunk(agent, turn, step, chunk)— sameStreamChunk, sameturn/step.
The only thing the emit added over the session event was the live Agent handle, and the sole consumer discarded it (its handler signature was (_agent, _turn, _step, chunk)).
This is the same duplication the boundary-mirror removal eliminated for turn/step boundaries: a consumer had two sources of truth for one durable fact, and every change had to touch both. That Agent Note deferred the chunk stream ("assistant/chunk persistence remains load-bearing, so the chunk stream could later be evaluated as a mirror, but that is a separate decision") rather than bundling it in. This Agent Note is that separate decision.
The premise the deferral hinged on is settled: chunk persistence is authoritative and staying. The proposal to stop persisting chunks and keep only a transient live stream event was rejected — high-fidelity replay, partial failed streams, and snapshot replay all depend on the persisted assistant/chunk feed. So assistant/chunk on session/event is the durable, load-bearing token stream, and agent/stream-chunk is a pure redundant mirror of it.
Decision
Remove agent/stream-chunk from the agent event taxonomy. The token stream is read off session/event as assistant/chunk, the same feed persistence and replay already use — session/event is the single live transcript stream (assistant chunks, turn/step boundaries, tool activity, todos).
Consumers. Persistence, replay, and interactive renderers consume the authoritative session stream directly. The automation-only ACP bridge emits committed assistant/message text rather than raw chunks, so it needs neither event. No production consumer requires an Agent-first token mirror.
Scope
Removed: agent/stream-chunk.
Not touched:
assistant/chunk(the durable session event) — the authoritative token stream, kept exactly as-is. This Agent Note removes the LIVE MIRROR, not the persistence (the persistence-removal proposal was separately rejected — see above).agent/steering— not touched by THIS decision (a control signal, not the token stream). Its durable twin issteering/message, and the mirror emit was removed by its own follow-up: Remove theagent/steeringmirror emit.agent/status,agent/error,agent/created/agent/disposed,agent/queued,agent/session-start— lifecycle/control events that are not transcript data and have no durable duplicate.
Alternatives considered
Remove the persistence and keep only a transient live stream — the inverse cut, rejected separately: high-fidelity replay, partial failed streams, and snapshot replay all depend on the persisted assistant/chunk feed. With that settled, the live emit is the redundant half of the pair.
Consequences
A plugin can no longer observe token deltas from an Agent-first event. It subscribes to session/event, filters assistant/chunk, and looks up the corresponding live handle directly with ctx.agents.get(session.id) when needed. No production consumer needed the live Agent at chunk time; this is the same acceptable trade the boundary-mirror removal made.
中文
问题
agent loop（智能体循环）将模型的每个 token delta 同时记录为持久的 assistant/chunk 会话事件，并发射一个携带相同数据的并行实时 agent/stream-chunk Cordis 事件。在 packages/core/agent-loop/src/agent.ts 中，二者仅相隔一行：
const chunkEvent = session.append('assistant/chunk', { turn, step, chunk })
chunkSeqs.push(chunkEvent.seq)
ctx.emit('agent/stream-chunk', agent, turn, step, chunk) // ← the mirror
- 持久事件：
assistant/chunk: { turn, step, chunk }。 - 实时发射：
agent/stream-chunk(agent, turn, step, chunk)——相同的StreamChunk，相同的turn/step。
实时发射相比会话事件唯一多出的东西是实时的 Agent 句柄，而唯一的消费方直接丢弃了它（其处理函数签名为 (_agent, _turn, _step, chunk)）。
这与移除边界镜像为轮次/步骤边界消除的重复相同：消费方面对同一持久事实的两个真源，每次变更都必须同时触及两者。该 Agent Note（agent 决策记录）没有把分片流一并纳入，而是推迟处理（“assistant/chunk 持久化仍承载关键约束，所以以后可以将分片流作为镜像评估，但那是一项独立决策”）。本 Agent Note 就是那项独立决策。
推迟所依赖的前提已经明确：分片持久化是权威的，且将保留。停止持久化分片、仅保留瞬态实时流事件的提案已被否决——高保真回放、部分失败的流以及快照回放都依赖持久化的 assistant/chunk 序列。因此 session/event 上的 assistant/chunk 是持久的、承重的 token 流，而 agent/stream-chunk 是它的纯冗余镜像。
决策
从 agent 事件分类体系中移除 agent/stream-chunk。token 流通过 session/event 以 assistant/chunk 的形式读取——持久化与回放已经使用的正是同一个序列。session/event 是唯一的实时 transcript（文本记录）流（assistant 分片、轮次/步骤边界、工具活动、todo）。
消费方。 持久化、回放和交互式渲染器直接消费权威的会话流。仅面向自动化的 ACP（Agent Client Protocol）桥接层发出已提交的 assistant/message 文本而非原始分片，因此两种事件它都不需要。没有生产消费方需要一个 Agent 优先的 token 镜像。
范围
移除：agent/stream-chunk。
未触及：
assistant/chunk（持久会话事件）——权威 token 流，原样保留。本 Agent Note 移除的是实时镜像，而非持久化（移除持久化的提案已单独遭到拒绝——见上文）。agent/steering——本决策未触及（它是控制信号，不是 token 流）。其持久孪生事件是steering/message，镜像发射由其自身的后续 Agent Note 移除：移除agent/steering镜像发射。agent/status、agent/error、agent/created/agent/disposed、agent/queued、agent/session-start——生命周期/控制事件，不是 transcript 数据，也没有持久副本。
曾考虑的替代方案
移除持久化、仅保留瞬态实时流——反向裁剪，已被单独否决：高保真回放、部分失败的流以及快照回放都依赖持久化的 assistant/chunk 序列。在此前提确定后，实时发射才是配对中冗余的那一半。
后果
插件不能再从 Agent 优先事件观察 token 增量。它需要订阅 session/event、过滤 assistant/chunk，并在需要时通过 ctx.agents.get(session.id) 直接查找对应的实时 handle。没有生产消费方需要在分片时刻取得实时 Agent；这与移除边界镜像所作的取舍相同，均可接受。