refactor(compact): turn-agnostic retention + dedicated agent/pre-request seam

Reform the compaction blueprint so a runaway turn survives and the design
stops drifting across review rounds:

- Drop in-flight-turn protection ("layer 2"). Retention is a uniform tail→head
  whole-unit walk; the only structural guard is step-alignment. A single turn
  that alone exceeds the window now compacts its own early closed steps instead
  of being retained verbatim (the failure mode that motivated this).
- Move auto-compaction off the agent/request waterfall onto a new awaited
  agent/pre-request loop seam, fired before history derivation. Compaction
  mutates the surface; the loop derives once from the result — no double-derive,
  and a listener structurally cannot act on not-yet-derived messages.
- Tighten compactIfNeeded to required (session, system, model, signal).
- Enforce a single-pass convergence invariant in resolveConfig: reject configs
  where summarizationMaxTokens + retainTokens exceeds the threshold, so a
  compaction can never immediately re-trigger.
- Document the crash vs recoverable failure taxonomy; core session repair stays
  compaction-agnostic (a log-only orphaned compact/start is inert).
- Wire dsh-compact-basic into examples/coding-agent and add a with-key
  compaction e2e (compaction's first real-world exercise + runaway net).
- Rewrite the RFC to encode the blueprint and move it to implemented/.

The runaway-turn snapshot is a named deferred follow-up: dsh-llm-replay cannot
yet serve the interleaved summarization model call.
This commit is contained in:
Hypatia May
2026-06-26 08:59:33 +08:00
parent aa9afcefc7
commit cec32faa4e
22 changed files with 724 additions and 424 deletions

View File

@@ -65,6 +65,16 @@
failures before moving on. Verify your work by running the code or
tests. Keep answers brief and factual.
# Automatic context compaction: when the derived history approaches the model's
# context window, summarize an older range into a checkpoint so a long-running
# or tool-heavy session keeps fitting. A leaf entry (needs ctx.llm + the
# agent-loop's `agent/pre-request` seam from the app above).
- id: compact-basic
name: '@deepseek-ai/dsh-compact-basic'
config:
contextWindow: 128000
retainTokens: 20480
# The subagent seam + BOTH in-process backends + two model-facing tools, as leaf
# entries after the app (which provides ctx.agents/ctx.tools). spawn (a fresh
# child) and fork (a child seeded with the parent's completed-turn prefix) are

View File

@@ -0,0 +1,95 @@
import { mkdtemp, rm, writeFile } from 'node:fs/promises'
import { tmpdir } from 'node:os'
import { join } from 'node:path'
import { afterEach, describe, expect, it } from 'vitest'
import type { Context } from 'cordis'
import { AgentId } from '@deepseek-ai/dsh-agent'
import { codingHarness, finalText, SYSTEM_PROMPT, waitForIdle } from './harness.ts'
/**
* The compaction smoke test: a real model runs a multi-step bash task with a
* deliberately tiny context window, so the auto-compaction listener fires
* MID-SESSION and summarizes the older history into a checkpoint. This is the
* first end-to-end exercise of the compaction seam (it is wired nowhere else),
* and the runaway-survival regression net — it proves a session that grows past
* the window keeps running rather than overflowing. Key-gated.
*
* Verifies the WORLD, not the agent's self-report: a compact/start…end pair
* landed in the real session log, the surface actually shrank (a replace node
* exists and shadowed older nodes), and the agent still produced a final answer
* after compaction (so the summarized history did not break the conversation).
*/
let workdir: string | undefined
let ctx: Context | undefined
afterEach(async () => {
await ctx?.fiber.dispose()
ctx = undefined
if (workdir !== undefined) await rm(workdir, { recursive: true, force: true })
workdir = undefined
})
describe.skipIf(!process.env.DEEPSEEK_API_KEY)('compaction: a long session compacts mid-flight and keeps running', () => {
it('summarizes older history into a checkpoint without breaking the task', async () => {
workdir = await mkdtemp(join(tmpdir(), 'dsh-compaction-'))
// A few files for the model to read, so multiple bash steps accumulate
// surface nodes (tool calls + results) and grow the history.
for (let i = 1; i <= 4; i++) {
await writeFile(join(workdir, `file${i}.txt`), `This is file number ${i}. `.repeat(40))
}
// Tiny window so a handful of steps crosses the threshold. The convergence
// invariant requires summarizationMaxTokens + retainTokens <= window *
// ratio = floor(8000 * 0.5) = 4000; 1500 + 2000 = 3500 <= 4000.
ctx = await codingHarness(workdir, {
compact: {
contextWindow: 8000,
thresholdRatio: 0.5,
retainTokens: 2000,
summarizationMaxTokens: 1500,
},
})
const agent = ctx.agentLoop.create(AgentId('e2e-compaction'), {
model: 'deepseek-v4-flash',
systemPrompt: SYSTEM_PROMPT,
})
agent.send([{
type: 'text',
text: 'Read file1.txt, file2.txt, file3.txt, and file4.txt one at a time using cat '
+ '(a separate bash command for each). After reading all four, tell me how many '
+ 'files you read and the number mentioned in file1.txt.',
}])
await waitForIdle(ctx, agent)
const events = [...agent.session.events]
// A compaction ran: the start…end bracket landed in the real log.
const starts = events.filter(e => e.type === 'compact/start')
const ends = events.filter(e => e.type === 'compact/end')
expect(starts.length).toBeGreaterThan(0)
expect(ends.length).toBe(starts.length) // every start was released
// It succeeded at least once: a compact/summary provenance event and a
// replace-op user/message (the surface mutation) both landed.
const summaries = events.filter(e => e.type === 'compact/summary')
expect(summaries.length).toBeGreaterThan(0)
const replaceNode = events.find((e) => {
const se = e as unknown as { type: string; surfaceOp?: unknown }
return se.type === 'user/message' && typeof se.surfaceOp === 'object' && se.surfaceOp !== null
})
expect(replaceNode).toBeDefined()
// The summary shadowed real older nodes (the surface shrank vs. the raw
// message-producing event count).
const summaryData = summaries[0]!.data as { shadowedSeqs: number[] }
expect(summaryData.shadowedSeqs.length).toBeGreaterThan(0)
// The conversation survived compaction: the agent produced a final answer
// that reflects the work (it read four files).
const answer = finalText(events).toLowerCase()
expect(answer.length).toBeGreaterThan(0)
expect(answer).toMatch(/\b(4|four)\b/)
}, 240_000)
})

View File

@@ -10,6 +10,8 @@ import { LocalBashExecutor } from '@deepseek-ai/dsh-bash-local'
import * as ToolBash from '@deepseek-ai/dsh-tool-bash'
import * as LlmDeepSeek from '@deepseek-ai/dsh-llm-deepseek'
import SessionPersistenceJsonl from '@deepseek-ai/dsh-session-persistence-jsonl'
import { BasicCompactService } from '@deepseek-ai/dsh-compact-basic'
import type { BasicCompactConfig } from '@deepseek-ai/dsh-compact-basic'
/**
* Shared harness for the coding-agent e2e suites: the full plugin stack
@@ -21,7 +23,19 @@ export const SYSTEM_PROMPT = 'You are a coding agent. Your only tool is bash; '
+ 'do file operations with cat/grep/heredocs, check [exit code: N] markers, '
+ 'and report results briefly.'
export async function codingHarness(workdir: string, persistenceRoot?: string): Promise<Context> {
/** Options for {@link codingHarness}. */
export interface CodingHarnessOptions {
/** Durable JSONL persistence root (the resume suite needs it; others stay file-free). */
persistenceRoot?: string
/**
* Load {@link BasicCompactService} with this config so the compaction e2e can
* trigger compaction at a small, controlled history size. Omitted ⇒ no
* compaction plugin (the default suites run without it).
*/
compact?: BasicCompactConfig
}
export async function codingHarness(workdir: string, options: CodingHarnessOptions = {}): Promise<Context> {
const ctx = new Context()
await ctx.plugin(LlmService)
await ctx.plugin(SessionStore)
@@ -32,10 +46,13 @@ export async function codingHarness(workdir: string, persistenceRoot?: string):
await ctx.plugin(LlmDeepSeek, { models: ['deepseek-v4-flash'] })
await ctx.plugin(LocalBashExecutor, { cwd: workdir, timeoutMs: 30_000 })
await ctx.plugin(ToolBash)
// Compaction is opt-in: only the compaction e2e loads it, with a lowered
// contextWindow/retainTokens so a short real session crosses the threshold.
if (options.compact !== undefined) await ctx.plugin(BasicCompactService, options.compact)
// Durable JSONL persistence is opt-in: only the resume e2e needs it, and the
// other suites stay file-free. Loaded last so a resume's deferred
// `ctx.inject(['sessionPersistence'])` resolves once this is present.
if (persistenceRoot !== undefined) await ctx.plugin(SessionPersistenceJsonl, { root: persistenceRoot })
if (options.persistenceRoot !== undefined) await ctx.plugin(SessionPersistenceJsonl, { root: options.persistenceRoot })
return ctx
}

View File

@@ -38,7 +38,7 @@ describe.skipIf(!process.env.DEEPSEEK_API_KEY)('resume: continue a persisted ses
// Run 1: a fresh agent on a KNOWN session id learns a secret, then we
// dispose the whole context (simulating process exit) so only the JSONL
// log on disk survives.
ctx = await codingHarness(process.cwd(), root)
ctx = await codingHarness(process.cwd(), { persistenceRoot: root })
const first = ctx.agents.create({
agentId: AgentId('resume-1'),
sessionId: SESSION_ID,
@@ -52,7 +52,7 @@ describe.skipIf(!process.env.DEEPSEEK_API_KEY)('resume: continue a persisted ses
// Run 2: a brand-new context over the SAME root resumes the persisted
// session. The loaded event log seeds the live session, so the model sees
// run 1's exchange as conversation history.
ctx = await codingHarness(process.cwd(), root)
ctx = await codingHarness(process.cwd(), { persistenceRoot: root })
const resumed = (await ctx.agents.resume({
agentId: AgentId('resume-2'),
resumeSessionId: SESSION_ID,