refactor agent pre-step inbox lifecycle

This commit is contained in:
_Kerman
2026-07-31 19:21:16 +08:00
parent c2ff9ddec8
commit fcc2b5e282
267 changed files with 2052 additions and 1546 deletions

View File

@@ -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 packages/compact/compact-basic/README.md
README.md: 775355f1ac1a7c79c16f66a5b2489d73df7b960d
README.zh.md: bfa139596b5ef61c23d29575bdea5534fa82b158
README.md: eed9412e335a0b703f3d1d5cb7112de10a19e9a0
README.zh.md: 763736760e4e7a6acd34a0a085e0e3b685f0a8c9

View File

@@ -17,7 +17,7 @@ This backend owns the compaction policy:
- **Convergence** — retry head-checkpoint compaction up to `compactionRetries`; reject a summary that does not shrink its source, and throw if retries cannot return below threshold.
- **Summarization** — a direct `llm/stream` call uses the configured provider/model pair and cap, falling back to the latest logged request target and then the agent target, without running the loop-only `agent/request` seam. The call replays the conversation's own system prompt, tools, and shadowed-region messages verbatim and appends the compaction instruction as the final user message, so it reuses the provider's warm prefix cache instead of invalidating it. It sets `GenerateOptions.purpose` to `compaction`, which adapters may forward as request attribution (the DeepSeek adapter sends `x-deepseek-harness-compact: 1`) without touching the model-visible body. Only returned text enters the checkpoint, excluding reasoning and tool calls that would leak private reasoning or create an orphaned call.
- **Framing** — the replacement user message marks established checkpoint context with `<compacted-summary>` tags. The raw summary remains on the provenance event, and later automatic cycles merge the prior checkpoint.
- **Lifecycle** — `compactRegion()` mutates `agent.session` and records its start, summary, replacement, and end. After asynchronous summarization it rejects a changed surface-node snapshot, while unrelated log-only events may append without invalidating the selected span. The serial `agent/step` listener checks pressure before request derivation. A canonical provider overflow is offered through `agent/request-error` after the failed step; the plugin compacts there and returns a retry action only after durable surface progress.
- **Lifecycle** — `compactRegion()` mutates `agent.session` and records its start, summary, replacement, and end. After asynchronous summarization it rejects a changed surface-node snapshot, while unrelated log-only events may append without invalidating the selected span. The serial `agent/pre-step` listener checks pressure before request derivation. A canonical provider overflow is offered through `agent/request-error` after the failed step; the plugin compacts there and returns a retry action only after durable surface progress.
- **Overflow recovery** — provider-confirmed overflow needs no capacity metadata: it bypasses normal pressure and retention, prunes, then attempts one maximal balanced head reduction while leaving the newest indivisible unit. Retry is authorized whenever `surface.replaceGeneration` advances, including when pruning lands before later summary work throws. No replacement, an exhausted target-specific cap, cancellation, or an unknown/noncanonical error preserves the original provider failure.
- **Failure handling** — an unmatched `compact/start` is an inert crash marker because no summary replacement landed. A region failure records an error end; the surface remains unchanged unless pruning already landed. Operational pressure failures warn and continue, while overflow-recovery failure preserves the original provider error only when no earlier replacement advanced the surface. Cancellation remains authoritative after any progress.

View File

@@ -17,7 +17,7 @@
- **收敛**:最多按 `compactionRetries` 重试头部检查点压缩;拒绝不能缩小源内容的摘要,如果重试仍无法回到阈值以下,则抛出异常。
- **摘要**:直接 `llm/stream` 调用使用已配置的提供方/模型对与上限,回退到最新已记录请求目标,然后再回退到 agent 目标,而不运行仅用于 agent loop 的 `agent/request` seam。该调用会逐字回放会话自身的系统提示词、工具与已遮蔽区域消息并将压缩指令作为最后一条 user 消息追加,从而复用提供方的热前缀 cache而非使它失效。它将 `GenerateOptions.purpose` 设为 `compaction`适配器可将其作为请求归因转发DeepSeek 适配器发送 `x-deepseek-harness-compact: 1`但不会触碰模型可见的请求体。只有返回的文本会进入检查点推理reasoning和工具调用都会被排除以免泄露私有推理或产生遗留调用。
- **框定**:替换 user 消息使用 `<compacted-summary>` 标签标记已建立的检查点上下文。原始摘要保留在溯源事件上,后续自动周期会合并之前的检查点。
- **生命周期**`compactRegion()` 会更改 `agent.session`,并记录开始、摘要、替换与结束。异步摘要后,如果表层节点快照已改变,它会拒绝操作,而不相关的仅日志事件可以追加,不会使已选 span 失效。串行 `agent/step` listener 会在派生请求之前检查压力。规范提供方溢出会在失败步骤之后经由 `agent/request-error` 交给本插件;插件在此执行压缩,并且只在表层取得持久进展后才返回重试动作。
- **生命周期**`compactRegion()` 会更改 `agent.session`,并记录开始、摘要、替换与结束。异步摘要后,如果表层节点快照已改变,它会拒绝操作,而不相关的仅日志事件可以追加,不会使已选 span 失效。串行 `agent/pre-step` listener 会在派生请求之前检查压力。规范提供方溢出会在失败步骤之后经由 `agent/request-error` 交给本插件;插件在此执行压缩,并且只在表层取得持久进展后才返回重试动作。
- **溢出恢复**:提供方已确认的溢出不需容量元数据。它会绕过常规压力与保留,执行剪枝,再尝试一次最大平衡头部缩减,并留下最新不可分单元。只要 `surface.replaceGeneration` 前进,就允许重试,包括剪枝在后续摘要工作抛出异常前已落地的情况。如果没有替换、目标特定上限已耗尽、已取消,或遇到未知/非规范错误,则保留原始提供方失败。
- **失败处理**:未配对的 `compact/start` 是不起作用的崩溃标记,因为没有摘要替换落地。区域失败会记录错误结束;除非剪枝已落地,否则表层保持不变。压力检查中的运行故障会发出警告并继续;只有此前没有替换推进表层时,溢出恢复失败才保留原始提供方错误。即使已经取得进展,取消仍具有最终决定权。

View File

@@ -11,7 +11,7 @@ import type { CompactionResult, CompactionTrigger } from '@deepseek-ai/dsh-compa
import type { Session } from '@deepseek-ai/dsh-session'
import { CONTEXT_WINDOW_EXCEEDED_CODE, assertNever } from '@deepseek-ai/dsh-llm'
import type { ContentBlock, LlmCallConfig } from '@deepseek-ai/dsh-llm'
import type { Agent } from '@deepseek-ai/dsh-agent'
import type { Agent, PreStepDecision } from '@deepseek-ai/dsh-agent'
// Type-only: makes the optional sibling service available to `ctx.get()`.
import type {} from '@deepseek-ai/dsh-compact-tool-result-prune'
import {
@@ -135,24 +135,26 @@ export class BasicCompactService extends CompactService {
)
}
ctx.on('agent/step', async (
ctx.on('agent/pre-step', async (
agent: Agent,
_turn: number,
_step: number,
signal: AbortSignal,
) => {
if (signal.aborted) return
try {
const result = await this.compactIfNeeded(agent, 'pressure', signal)
if (result !== null) logResult(result, 'step pressure')
} catch (error: unknown) {
if (error instanceof TargetPressureConfigError) {
if (this.warnedPressureConfigTargets.has(error.targetKey)) return
this.warnedPressureConfigTargets.add(error.targetKey)
_messages,
{ signal },
next,
): Promise<PreStepDecision> => {
if (!signal.aborted) {
try {
const result = await this.compactIfNeeded(agent, 'pressure', signal)
if (result !== null) logResult(result, 'step pressure')
} catch (error: unknown) {
if (error instanceof TargetPressureConfigError) {
if (this.warnedPressureConfigTargets.has(error.targetKey)) return next()
this.warnedPressureConfigTargets.add(error.targetKey)
}
const message = error instanceof Error ? error.message : String(error)
ctx.logger.warn(`step compaction failed: ${message}; continuing the turn`)
}
const message = error instanceof Error ? error.message : String(error)
ctx.logger.warn(`step compaction failed: ${message}; continuing the turn`)
}
return next()
})
ctx.on('agent/status', (agent, status) => {

View File

@@ -1369,8 +1369,11 @@ describe('default one-shot summarizer', () => {
})
describe('automatic listener and loader composition', () => {
function postStep(ctx: Context, owner: Agent, signal = SIGNAL): Promise<unknown> {
return agentEvents(ctx, owner).serial('agent/step', 1, 1, signal)
function preStep(ctx: Context, owner: Agent, signal = SIGNAL) {
return agentEvents(ctx, owner).waterfall(
'agent/pre-step', [], { turn: 1, step: 1, signal },
() => Promise.resolve({ kind: 'enter' as const, messages: [] }),
)
}
function recover(
@@ -1394,23 +1397,23 @@ describe('automatic listener and loader composition', () => {
return Object.assign(new Error(message), { code: CONTEXT_WINDOW_EXCEEDED_CODE })
}
it('compacts post-step above threshold using the durable routed model and remains idle below it', async () => {
it('compacts before a step above threshold using the durable routed model and remains idle below it', async () => {
const ctx = createContext()
const compact = new TestCompactService(ctx, {
thresholdRatio: 0.5,
retainTokens: 180,
})
const pressured = conversation(4)
await postStep(ctx, agent(pressured, 'unconfigured-agent-fallback'))
await preStep(ctx, agent(pressured, 'unconfigured-agent-fallback'))
expect(pressured.events.some(event => event.type === 'compact/summary')).toBe(true)
const small = conversation(1)
await postStep(ctx, agent(small, MODEL))
await preStep(ctx, agent(small, MODEL))
expect(small.events.some(event => event.type === 'compact/start')).toBe(false)
expect(compact.calls).toHaveLength(1)
})
it('skips post-step pressure when the step signal is already aborted', async () => {
it('skips pre-step pressure when the step signal is already aborted', async () => {
const ctx = createContext()
const compact = new TestCompactService(ctx, {
thresholdRatio: 0.5,
@@ -1419,8 +1422,8 @@ describe('automatic listener and loader composition', () => {
const pressured = conversation(4)
const compactIfNeeded = vi.spyOn(compact, 'compactIfNeeded')
await expect(postStep(ctx, agent(pressured, MODEL), AbortSignal.abort('step aborted')))
.resolves.toBeUndefined()
await expect(preStep(ctx, agent(pressured, MODEL), AbortSignal.abort('step aborted')))
.resolves.toEqual({ kind: 'enter', messages: [] })
expect(compactIfNeeded).not.toHaveBeenCalled()
expect(pressured.events.some(event => event.type === 'compact/start')).toBe(false)
@@ -1437,7 +1440,7 @@ describe('automatic listener and loader composition', () => {
compact.error = 'temporary failure'
const session = conversation(4)
await expect(postStep(ctx, agent(session, MODEL))).resolves.toBeUndefined()
await expect(preStep(ctx, agent(session, MODEL))).resolves.toEqual({ kind: 'enter', messages: [] })
expect(warnings).toContainEqual(expect.stringContaining('temporary failure'))
expect(session.events.some(event => event.type === 'compact/summary')).toBe(false)
})
@@ -1457,8 +1460,8 @@ describe('automatic listener and loader composition', () => {
})
const session = conversation(4)
await postStep(ctx, agent(session, MODEL))
await postStep(ctx, agent(session, MODEL))
await preStep(ctx, agent(session, MODEL))
await preStep(ctx, agent(session, MODEL))
expect(warnings).toEqual([
expect.stringContaining(`no context capacity for ${MODEL}/${MODEL}`),
@@ -1475,8 +1478,8 @@ describe('automatic listener and loader composition', () => {
})
const session = conversation(4)
await postStep(ctx, agent(session, MODEL))
await postStep(ctx, agent(session, MODEL))
await preStep(ctx, agent(session, MODEL))
await preStep(ctx, agent(session, MODEL))
expect(warnings).toEqual([
expect.stringContaining('retainTokens (500) must be less than threshold tokens 500'),
@@ -1759,7 +1762,7 @@ describe('automatic listener and loader composition', () => {
retainTokens: 180,
})
const session = conversation(4)
await postStep(ctx, agent(session, MODEL))
await preStep(ctx, agent(session, MODEL))
const summaries = session.events.filter(event => event.type === 'compact/summary').length
expect(summaries).toBe(1)
expect(await recover(ctx, agent(session, MODEL), overflow())).toBe(false)
@@ -1774,7 +1777,7 @@ describe('automatic listener and loader composition', () => {
retainTokens: 180,
})
const session = conversation(4)
await postStep(ctx, agent(session, MODEL))
await preStep(ctx, agent(session, MODEL))
expect(session.events.some(event => event.type === 'compact/start')).toBe(false)
expect(await recover(ctx, agent(session, MODEL), overflow())).toBe(false)
})
@@ -1803,7 +1806,7 @@ describe('automatic listener and loader composition', () => {
await fiber.dispose()
const session = conversation(4)
await postStep(ctx, agent(session, MODEL))
await preStep(ctx, agent(session, MODEL))
expect(session.events.some(event => event.type === 'compact/start')).toBe(false)
expect(await recover(ctx, agent(session, MODEL), overflow())).toBe(false)
})