Merge branch 'codex/tool-json-schema-dsl' into codex/canonical-tool-output

packages/compact/compact-basic/README.md

@@ -1,6 +1,6 @@
 # @deepseek-ai/dsh-compact-basic
 
-The **basic compaction backend**: a `BasicCompactService` implementing the `@deepseek-ai/dsh-compact` seam with reusable `ctx.tokenMeter` pressure, token-budget retention, and summarization as a direct one-shot `ctx.llm.stream()` call (interceptable at `llm/stream`).
+The **basic compaction backend**: a `BasicCompactService` implementing the `@deepseek-ai/dsh-compact` seam with reusable `ctx.tokenMeter` pressure, token-budget retention, and summarization as a direct one-shot `ctx.llm.stream()` call that replays the conversation prefix to reuse the provider's KV cache (interceptable at `llm/stream`).
 
 This is the implementation tier of the compaction capability — see the [interface package](../compact/README.md) for the seam and the [capability-seam Agent Note](../../../.agents/notes/implemented/feature/2026-06-18-compaction-capability-seam.md) for the design.
 
@@ -13,7 +13,7 @@ This backend owns the compaction policy:
 - **Model-free pruning** — after pressure or canonical overflow qualifies, the optional [`ctx.toolResultPrune`](../compact-tool-result-prune/README.md) service rewrites oversized tool results before range selection. Compact-basic remeasures through `ctx.tokenMeter`, skips summarization when pressure becomes safe, and otherwise summarizes the pruned surface. Below-pressure post-step checks never prune.
 - **Retention** — compact the oldest whole surface units while preserving a recent tail and balanced tool-call/result cuts through the [`dsh-compact` boundary helpers](../compact/README.md#tool-pairing-boundaries). Turn boundaries do not protect old steps inside a runaway turn. An open indivisible tail declines until it closes. The optional pruner can repair an oversized closed tool unit when its text-bearing result is the removable bulk; indivisible non-tool units and non-prunable tool remainders remain out of scope.
 - **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 input transcript preserves non-text blocks as tagged placeholders; only returned text enters the checkpoint, excluding reasoning and tool calls that would leak private reasoning or create an orphaned call.
+- **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. 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. The serial `agent/post-step` listener checks pressure after successful output and tool work are durable but before `step/end`. Canonical provider overflow is handled through `agent/request-error` after the failed step closes.
 - **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.
@@ -96,30 +96,16 @@ Model-free pruning can avoid the auxiliary call entirely; otherwise it reduces t
 
 Replacing rather than append-only. Each checkpoint invalidates reuse from the first replaced history token; the unchanged request prefix before that range remains reusable.
 
-### Auxiliary summarizer user message
+### Auxiliary summarizer request
 
 #### What the model sees
 
-The summarization model receives exactly `Summarize this conversation history:` followed by a blank line, the data-dependent [`renderTranscript()`](../compact/README.md) output, another blank line, and `Summary:`. The conversation model never sees this private request or its reasoning; only returned text is stored.
+The summarization model receives the conversation replayed verbatim — the same system prompt, tool schemas, and messages the last routed request sent for the shadowed region — followed by one final user message: the compaction instruction below. The conversation model never sees this private request or its reasoning; only returned text is stored.
 
-#### Token effect
-
-This is a separate model call with data-dependent input and `maxTokens`-capped output. Convergence retries can pay this cost more than once.
-
-#### KV Cache effect
-
-Independent of the conversation request cache. An auxiliary call can reuse an exact transcript prefix, while a different selected range or rendering invalidates reuse from its first changed token.
-
-### Auxiliary summarizer system prompt
-
-#### What the model sees
-
-The summarization model receives the checkpoint-writing instruction below.
-
-##### Auxiliary summarizer system prompt
+##### Compaction instruction (final user message)
 
 ```markdown
-You are a compaction engine for an AI coding assistant. Condense the conversation transcript into a structured checkpoint that lets another model resume the work with no loss of essential context.
+You are now acting as a compaction engine for this AI coding assistant. Condense the conversation ABOVE into a structured checkpoint that lets another model resume the work with no loss of essential context.
 
 Output EXACTLY the Markdown structure below: keep every section, in order. Use terse bullets, not prose paragraphs. Write "(none)" for an empty section — never drop a section.
 
@@ -150,17 +136,18 @@ Output EXACTLY the Markdown structure below: keep every section, in order. Use t
 Rules:
 - Preserve exact file paths, commands, error strings, identifiers, and function signatures.
 - Capture user feedback and explicit instructions faithfully, especially corrections.
-- Do NOT mention this summarization process or that the context was compacted.
-- If the transcript already contains a <compacted-summary> block, it is a PRIOR checkpoint. Do not copy it forward verbatim: preserve still-true facts, drop stale ones, and merge newer information into a single consolidated summary under the same structure.
+- Do NOT mention this summarization request or that the context was compacted.
+- Output only the checkpoint text: do not call any tool or take any other action.
+- If the conversation already contains a <compacted-summary> block, it is a PRIOR checkpoint. Do not copy it forward verbatim: preserve still-true facts, drop stale ones, and merge newer information into a single consolidated summary under the same structure.
 ```
 
 #### Token effect
 
-Fixed auxiliary input cost plus the data-dependent transcript on every summarization attempt.
+This is a separate model call: the replayed conversation prefix plus the fixed instruction as input, with `maxTokens`-capped output. Convergence retries can pay this cost more than once.
 
 #### KV Cache effect
 
-Prefix-stable for auxiliary calls while this instruction and the summarizer route are unchanged. Changing either starts a different prefix; transcript changes occur after the instruction.
+The replayed system prompt, tools, and shadowed-region messages match the conversation's last routed request byte-for-byte, so the provider's warm prefix cache is reused up to the trailing instruction; only that instruction, and the summary output, is uncached. Routing the summarizer to a different provider/model, or compacting a non-head range, forgoes this reuse.
 
 ## Known Limitations and Deferred Work
 

packages/compact/compact-basic/src/index.ts

@@ -22,6 +22,7 @@ import {
 } from './config.ts'
 import { compactSurfaceRegion, selectCompactableRange } from './region.ts'
 import { summarizeWithLlm } from './summarizer.ts'
+import type { SummarizationInput } from './summarizer.ts'
 import type {
   BasicCompactConfig,
   ModelCompactPolicyConfig,
@@ -205,15 +206,17 @@ export class BasicCompactService extends CompactService {
   }
 
   /**
-   * Summarize a rendered region through a direct one-shot `ctx.llm.stream()`
-   * call. Override this sole hook for a template or remote summarizer.
-   * @param text - plain-text conversation region to condense.
+   * Summarize the replayed conversation region through a direct one-shot
+   * `ctx.llm.stream()` call whose prefix reuses the conversation's own system
+   * prompt, tools, and messages so the provider's KV cache is not invalidated.
+   * Override this sole hook for a template or remote summarizer.
+   * @param input - replayed conversation prefix (system, tools, and leading messages) to condense.
    * @param agent - supplies routed-model history, fallback model, and session id.
    * @param signal - optional cancellation forwarded to the adapter.
    * @returns safe text summary blocks and exact auxiliary-call provenance.
    */
   protected async summarize(
-    text: string,
+    input: SummarizationInput,
     agent: Agent,
     signal?: AbortSignal,
   ): Promise<{ summary: ContentBlock[]; provider: string; model: string; maxTokens?: number }> {
@@ -221,7 +224,7 @@ export class BasicCompactService extends CompactService {
     const config = target === undefined
       ? this.config
       : resolveTargetPolicy(this.config, target)
-    return summarizeWithLlm(this.ctx, config, text, agent, signal)
+    return summarizeWithLlm(this.ctx, config, input, agent, signal)
   }
 
   /**
@@ -327,7 +330,7 @@ export class BasicCompactService extends CompactService {
     const session = agent.session
     return compactSurfaceRegion({
       meter: this.ctx.tokenMeter,
-      summarize: (text, owner, abort) => this.summarize(text, owner, abort),
+      summarize: (input, owner, abort) => this.summarize(input, owner, abort),
     }, session, start, end, agent, signal)
   }
 }

packages/compact/compact-basic/src/region.ts

@@ -5,20 +5,20 @@
  */
 
 import {
-  renderTranscript,
   toolPairingBalancedAfter,
   toolPairingBalancedBefore,
 } from '@deepseek-ai/dsh-compact'
 import type { CompactionResult } from '@deepseek-ai/dsh-compact'
+import type { Message } from '@deepseek-ai/dsh-llm'
 import type { TokenMeasurement, TokenMeterService } from '@deepseek-ai/dsh-token-meter'
 import type { Session, SessionEvent } from '@deepseek-ai/dsh-session'
 import type { Agent } from '@deepseek-ai/dsh-agent'
 import { frameSummary } from './summarizer.ts'
-import type { SummaryResult } from './summarizer.ts'
+import type { SummarizationInput, SummaryResult } from './summarizer.ts'
 
 interface RegionDependencies {
   readonly meter: TokenMeterService
-  summarize(text: string, agent: Agent, signal?: AbortSignal): Promise<SummaryResult>
+  summarize(input: SummarizationInput, agent: Agent, signal?: AbortSignal): Promise<SummaryResult>
 }
 
 /**
@@ -122,8 +122,8 @@ export async function compactSurfaceRegion(
       throw new Error('compaction: selected surface changed before summarization began')
     }
     const shadowedTokenCount = selected.reduce((total, node) => total + node.tokens, 0)
-    const text = renderTranscript(session.events, shadowedSeqs)
-    const { summary, provider, model, maxTokens } = await dependencies.summarize(text, agent, signal)
+    const summarizationInput = buildSummarizationInput(session, shadowedSeqs)
+    const { summary, provider, model, maxTokens } = await dependencies.summarize(summarizationInput, agent, signal)
 
     const currentMeasurement = dependencies.meter.measure(session)
     if (currentMeasurement.logRevision !== lockedMeasurement.logRevision) {
@@ -173,6 +173,34 @@ export async function compactSurfaceRegion(
   }
 }
 
+/**
+ * Reconstruct the last routed request's cacheable prefix for the shadowed
+ * region: its system prompt and tool schemas, then the request-only message
+ * prefix followed by the region's own derived messages in surface order. The
+ * summarizer appends only the compaction instruction after this, so the call
+ * is a genuine prefix of the conversation and reuses the provider's KV cache.
+ * @param session - session supplying the request header and per-node projection.
+ * @param shadowedSeqs - the surface-node seqs, in order, being compacted.
+ * @returns the replayed conversation prefix to condense.
+ */
+function buildSummarizationInput(
+  session: Session,
+  shadowedSeqs: readonly number[],
+): SummarizationInput {
+  const header = session.requestHeader()
+  const events = session.events
+  const regionMessages = shadowedSeqs
+    // shadowedSeqs are current surface seqs, so each is a valid log index.
+    // eslint-disable-next-line @typescript-eslint/no-non-null-assertion
+    .map(seq => session.deriveEventMessage(events[seq]!))
+    .filter((message): message is Message => message !== null)
+  return {
+    ...header?.system === undefined ? {} : { system: header.system },
+    ...header?.tools === undefined ? {} : { tools: header.tools },
+    messages: [...header?.messagePrefix ?? [], ...regionMessages],
+  }
+}
+
 /** Inspect the current turn boundary and latest compaction bracket once. */
 function inspectTurnTail(
   events: readonly SessionEvent[],

packages/compact/compact-basic/src/summarizer.ts

@@ -6,7 +6,7 @@
 
 import type { Context } from 'cordis'
 import { BlockAssembler } from '@deepseek-ai/dsh-llm'
-import type { ContentBlock, FinishReason, GenerateOptions } from '@deepseek-ai/dsh-llm'
+import type { ContentBlock, FinishReason, GenerateOptions, Message, ToolSchema } from '@deepseek-ai/dsh-llm'
 import type { Agent } from '@deepseek-ai/dsh-agent'
 
 interface SummaryConfig {
@@ -19,9 +19,15 @@ interface SummaryConfig {
 const SUMMARY_OPEN_TAG = '<compacted-summary>'
 const SUMMARY_CLOSE_TAG = '</compacted-summary>'
 
-/** Fixed structure required from the auxiliary summarization call. */
-const SUMMARIZE_SYSTEM_PROMPT = [
-  'You are a compaction engine for an AI coding assistant. Condense the conversation transcript into a structured checkpoint that lets another model resume the work with no loss of essential context.',
+/**
+ * The summarization directive, delivered as the FINAL user message after the
+ * replayed conversation rather than as a distinct summarizer system prompt.
+ * Keeping the conversation's own system prompt, tools, and message prefix in
+ * front of it makes the auxiliary call a genuine prefix of the last routed
+ * request, so the provider's KV cache is reused instead of invalidated.
+ */
+const COMPACTION_INSTRUCTION = [
+  'You are now acting as a compaction engine for this AI coding assistant. Condense the conversation ABOVE into a structured checkpoint that lets another model resume the work with no loss of essential context.',
   '',
   'Output EXACTLY the Markdown structure below: keep every section, in order. Use terse bullets, not prose paragraphs. Write "(none)" for an empty section — never drop a section.',
   '',
@@ -52,14 +58,30 @@ const SUMMARIZE_SYSTEM_PROMPT = [
   'Rules:',
   '- Preserve exact file paths, commands, error strings, identifiers, and function signatures.',
   '- Capture user feedback and explicit instructions faithfully, especially corrections.',
-  '- Do NOT mention this summarization process or that the context was compacted.',
-  `- If the transcript already contains a ${SUMMARY_OPEN_TAG} block, it is a PRIOR checkpoint. Do not copy it forward verbatim: preserve still-true facts, drop stale ones, and merge newer information into a single consolidated summary under the same structure.`,
+  '- Do NOT mention this summarization request or that the context was compacted.',
+  '- Output only the checkpoint text: do not call any tool or take any other action.',
+  `- If the conversation already contains a ${SUMMARY_OPEN_TAG} block, it is a PRIOR checkpoint. Do not copy it forward verbatim: preserve still-true facts, drop stale ones, and merge newer information into a single consolidated summary under the same structure.`,
 ].join('\n')
 
 /** Framing that makes the replacement user message established context. */
 const CHECKPOINT_PREAMBLE =
   'This is an automatically generated checkpoint condensing an earlier span of the conversation to free up context. Treat the captured context as established background and build on it without restating it. Continue the task directly from the messages that follow, without acknowledging this checkpoint.'
 
+/**
+ * The replayed conversation surface the summarizer condenses. Reproducing the
+ * last routed request's system prompt, tools, and leading messages verbatim
+ * lets the auxiliary call reuse the provider's warm prefix cache; the trailing
+ * compaction instruction is then the only novel input.
+ */
+export interface SummarizationInput {
+  /** The conversation's own system prompt, reused for prefix-cache alignment; absent for a system-less request. */
+  readonly system?: string
+  /** The conversation's tool schemas, reused for prefix-cache alignment; absent when the request carried none. */
+  readonly tools?: readonly ToolSchema[]
+  /** The request prefix followed by the shadowed region, in surface order, that precedes the compaction instruction. */
+  readonly messages: readonly Message[]
+}
+
 /** Safe summary content plus the exact auxiliary call envelope recorded in provenance. */
 export interface SummaryResult {
   summary: ContentBlock[]
@@ -69,10 +91,12 @@ export interface SummaryResult {
 }
 
 /**
- * Run the default direct `ctx.llm.stream()` summarization call.
+ * Run the default cache-reusing `ctx.llm.stream()` summarization call: replay
+ * the conversation prefix, then append the compaction instruction as the final
+ * user message so the provider's warm prefix cache is reused.
  * @param ctx - context providing the LLM service.
  * @param config - resolved backend configuration.
- * @param text - rendered transcript region to summarize.
+ * @param input - replayed conversation prefix (system, tools, and leading messages) to condense.
  * @param agent - supplies routed-model history, fallback model, and session id.
  * @param signal - optional cancellation forwarded to the adapter.
  * @returns safe text-only summary blocks and exact call provenance.
@@ -80,7 +104,7 @@ export interface SummaryResult {
 export async function summarizeWithLlm(
   ctx: Context,
   config: SummaryConfig,
-  text: string,
+  input: SummarizationInput,
   agent: Agent,
   signal?: AbortSignal,
 ): Promise<SummaryResult> {
@@ -102,14 +126,16 @@ export async function summarizeWithLlm(
   }
 
   const assembler = new BlockAssembler()
+  const messages: Message[] = [
+    ...input.messages,
+    { role: 'user', content: [{ type: 'text', text: COMPACTION_INSTRUCTION }] },
+  ]
   const options: GenerateOptions = {
     provider: target.provider,
     model: target.model,
-    messages: [{
-      role: 'user',
-      content: [{ type: 'text', text: `Summarize this conversation history:\n\n${text}\n\nSummary:` }],
-    }],
-    system: SUMMARIZE_SYSTEM_PROMPT,
+    messages,
+    ...input.system === undefined ? {} : { system: input.system },
+    ...input.tools === undefined ? {} : { tools: [...input.tools] },
     maxTokens: config.maxTokens,
     sessionId: agent.session.id,
     ...signal === undefined ? {} : { signal },

packages/compact/compact-basic/tests/compact-basic.spec.ts

@@ -3,6 +3,7 @@ import { Context } from 'cordis'
 import BasicCompactService from '@deepseek-ai/dsh-compact-basic'
 import type { BasicCompactConfig } from '@deepseek-ai/dsh-compact-basic'
 import { selectCompactableRange } from '@deepseek-ai/dsh-compact-basic/src/region.ts'
+import type { SummarizationInput } from '@deepseek-ai/dsh-compact-basic/src/summarizer.ts'
 import { toolPairingBalancedAfter, toolPairingBalancedBefore } from '@deepseek-ai/dsh-compact'
 import {
   resolveCompactSpec,
@@ -16,6 +17,7 @@ import type {
   GenerateOptions,
   LlmFailure,
   LlmModelContext,
+  Message,
   StreamChunk,
 } from '@deepseek-ai/dsh-llm'
 import { Session, SessionId } from '@deepseek-ai/dsh-session'
@@ -67,6 +69,21 @@ function agent(session: Session, model?: string): Agent {
   return { session, options: model === undefined ? {} : { provider: model, model } } as Agent
 }
 
+/** Flatten every text fragment the summarizer received, recursing tool-result blocks. */
+function summarizedText(input: SummarizationInput): string {
+  const collect = (blocks: readonly ContentBlock[]): string =>
+    blocks.map(block =>
+      block.type === 'text' ? block.text
+        : block.type === 'tool-result' ? collect(block.content)
+          : '').join('\n')
+  return input.messages.map(message => collect(message.content)).join('\n')
+}
+
+/** A minimal replayed prefix carrying one user message of the given text. */
+function promptInput(text: string): SummarizationInput {
+  return { messages: [{ role: 'user', content: [{ type: 'text', text }] }] }
+}
+
 /** Closed two-message turns followed by one open turn for durable compaction events. */
 function conversation(turns = 4, text = 'fixture '.repeat(40).trim()): Session {
   const session = new Session(SessionId(`conversation-${turns}`))
@@ -182,14 +199,14 @@ class TestCompactService extends BasicCompactService {
   summaryModel = 'summary-model'
   error: unknown
   mutateDuringSummary: (() => void) | undefined
-  calls: Array<{ text: string; signal: AbortSignal | undefined }> = []
+  calls: Array<{ input: SummarizationInput; signal: AbortSignal | undefined }> = []
 
   override async summarize(
-    text: string,
+    input: SummarizationInput,
     _agent: Agent,
     signal?: AbortSignal,
   ): Promise<{ summary: ContentBlock[]; provider: string; model: string; maxTokens?: number }> {
-    this.calls.push({ text, signal })
+    this.calls.push({ input, signal })
     this.mutateDuringSummary?.()
     if (this.error !== undefined) throw this.error
     return {
@@ -720,8 +737,8 @@ describe('optional model-free tool-result pruning', () => {
 
     expect(await compactIfNeeded(compact, session)).not.toBeNull()
     expect(compact.calls).toHaveLength(1)
-    expect(compact.calls[0]!.text).toContain('tool result middle pruned')
-    expect(compact.calls[0]!.text).not.toContain('result 1 '.repeat(300))
+    expect(summarizedText(compact.calls[0]!.input)).toContain('tool result middle pruned')
+    expect(summarizedText(compact.calls[0]!.input)).not.toContain('result 1 '.repeat(300))
   })
 
   it('retains the original compact-basic behavior without the optional plugin', async () => {
@@ -758,7 +775,7 @@ describe('compaction region transaction', () => {
     expect(result.shadowedSeqs).toEqual(before.slice(0, 4))
     expect(result.shadowedTokenCount).toBeGreaterThan(0)
     expect(compact.calls[0]).toMatchObject({ signal: SIGNAL })
-    expect(compact.calls[0]?.text).toContain('fixture user 1')
+    expect(summarizedText(compact.calls[0]!.input)).toContain('fixture user 1')
     const summary = session.events.findLast(event => event.type === 'compact/summary')
     expect(summary?.data).toMatchObject({
       shadowedSeqs: result.shadowedSeqs,
@@ -776,6 +793,25 @@ describe('compaction region transaction', () => {
     expect(replay.deriveMessages()).toEqual(session.deriveMessages())
   })
 
+  it('replays the latest routed header prefix so the summarizer reuses the cache', async () => {
+    const compact = service()
+    const session = conversation(3)
+    const tools = [{ name: 'do_thing', description: 'd', parameters: { type: 'object' } }]
+    const messagePrefix: Message[] = [{ role: 'user', content: [{ type: 'text', text: 'SESSION PREFIX' }] }]
+    session.append('request/header', {
+      header: { config: { provider: MODEL, model: MODEL }, system: 'CONVERSATION SYSTEM', tools, messagePrefix },
+      reason: 'resume',
+    })
+    const nodes = session.surface.nodes
+    await compact.compactRegion(nodes[0]!, nodes[1]!, agent(session, MODEL), SIGNAL)
+
+    const { input } = compact.calls[0]!
+    expect(input.system).toBe('CONVERSATION SYSTEM')
+    expect(input.tools).toEqual(tools)
+    expect(input.messages[0]).toEqual(messagePrefix[0])
+    expect(summarizedText(input)).toContain('fixture user 1')
+  })
+
   it.each([
     ['start missing', 9_001, undefined, /start seq 9001 not found/],
     ['end missing', undefined, 9_002, /end seq 9002 not found/],
@@ -989,11 +1025,11 @@ class ScriptedAdapter extends LlmAdapter {
 
 class ExposedCompactService extends BasicCompactService {
   runSummarize(
-    text: string,
+    input: SummarizationInput,
     owner: Agent,
     signal?: AbortSignal,
   ): Promise<{ summary: ContentBlock[]; provider: string; model: string; maxTokens?: number }> {
-    return this.summarize(text, owner, signal)
+    return this.summarize(input, owner, signal)
   }
 }
 
@@ -1025,7 +1061,7 @@ describe('default one-shot summarizer', () => {
       maxTokens: 321,
     })
     const session = conversation(1)
-    const output = await compact.runSummarize('transcript', agent(session, 'fallback'), SIGNAL)
+    const output = await compact.runSummarize(promptInput('transcript'), agent(session, 'fallback'), SIGNAL)
 
     expect(output).toEqual({
       summary: [{ type: 'text', text: 'public summary' }],
@@ -1040,7 +1076,68 @@ describe('default one-shot summarizer', () => {
       signal: SIGNAL,
       sessionId: session.id,
     })
-    expect(adapter.lastOptions?.system).toContain('## Primary Request and Intent')
+    const instruction = adapter.lastOptions?.messages.at(-1)?.content[0]
+    expect(instruction?.type === 'text' ? instruction.text : '').toContain('## Primary Request and Intent')
+  })
+
+  it('replays the conversation prefix and appends the instruction as the final message', async () => {
+    const { adapter, compact } = await summarizerHarness([{ type: 'text', text: 'summary' }])
+    const tools = [{ name: 'do_thing', description: 'd', parameters: { type: 'object' } }]
+    const prefix: Message = { role: 'user', content: [{ type: 'text', text: 'earlier turn' }] }
+    await compact.runSummarize({
+      system: 'REPLAYED SYSTEM',
+      tools,
+      messages: [prefix],
+    }, agent(conversation(1), MODEL))
+
+    expect(adapter.lastOptions?.system).toBe('REPLAYED SYSTEM')
+    expect(adapter.lastOptions?.tools).toEqual(tools)
+    const messages = adapter.lastOptions?.messages ?? []
+    expect(messages[0]).toEqual(prefix)
+    const last = messages.at(-1)?.content[0]
+    const lastText = last?.type === 'text' ? last.text : ''
+    expect(lastText).toContain('Condense the conversation ABOVE')
+    expect(lastText).toContain('## Primary Request and Intent')
+  })
+
+  it('applies the routed model policy without changing the replayed prefix', async () => {
+    const { ctx, compact } = await summarizerHarness(
+      [{ type: 'text', text: 'unused default summary' }],
+      undefined,
+      MODEL,
+      {
+        auto: false,
+        maxTokens: 111,
+        modelPolicies: [{
+          provider: MODEL,
+          model: MODEL,
+          summarizationProvider: 'policy-summary',
+          summarizationModel: 'policy-summary',
+          maxTokens: 222,
+        }],
+      },
+    )
+    const policyAdapter = new ScriptedAdapter([{ type: 'text', text: 'policy summary' }])
+    ctx.llm.registerAdapter(['policy-summary'], policyAdapter)
+    const prefix: Message = { role: 'user', content: [{ type: 'text', text: 'warm prefix' }] }
+
+    const output = await compact.runSummarize({
+      system: 'WARM SYSTEM',
+      messages: [prefix],
+    }, agent(conversation(1), 'fallback'))
+
+    expect(output).toMatchObject({
+      provider: 'policy-summary',
+      model: 'policy-summary',
+      maxTokens: 222,
+    })
+    expect(policyAdapter.lastOptions).toMatchObject({
+      provider: 'policy-summary',
+      model: 'policy-summary',
+      maxTokens: 222,
+      system: 'WARM SYSTEM',
+    })
+    expect(policyAdapter.lastOptions?.messages[0]).toEqual(prefix)
   })
 
   it('resolves the latest routed provider/model before the AgentOptions pair', async () => {
@@ -1050,7 +1147,7 @@ describe('default one-shot summarizer', () => {
       header: { config: { provider: 'routed', model: 'routed' } },
       reason: 'initial',
     })
-    const output = await compact.runSummarize('history', agent(session, 'fallback'))
+    const output = await compact.runSummarize(promptInput('history'), agent(session, 'fallback'))
     expect(output.provider).toBe('routed')
     expect(output.model).toBe('routed')
     expect(adapter.lastOptions?.provider).toBe('routed')
@@ -1084,7 +1181,7 @@ describe('default one-shot summarizer', () => {
     await ctx.plugin(LlmService)
     void new TokenMeterService(ctx)
     const compact = new ExposedCompactService(ctx, { auto: false })
-    await expect(compact.runSummarize('history', agent(new Session(SessionId('model-less')))))
+    await expect(compact.runSummarize(promptInput('history'), agent(new Session(SessionId('model-less')))))
       .rejects.toThrow(/no provider\/model available for summarization/)
   })
 
@@ -1092,7 +1189,7 @@ describe('default one-shot summarizer', () => {
     const { adapter, compact } = await summarizerHarness([{ type: 'text', text: 'summary' }])
     const session = new Session(SessionId('headerless-summary'))
 
-    await expect(compact.runSummarize('history', agent(session, MODEL))).resolves.toMatchObject({
+    await expect(compact.runSummarize(promptInput('history'), agent(session, MODEL))).resolves.toMatchObject({
       provider: MODEL,
       model: MODEL,
     })
@@ -1109,7 +1206,7 @@ describe('default one-shot summarizer', () => {
       session: new Session(SessionId(`incomplete-${String(options.model)}`)),
       options,
     } as Agent
-    await expect(compact.runSummarize('history', owner))
+    await expect(compact.runSummarize(promptInput('history'), owner))
       .rejects.toThrow(/no provider\/model available for summarization/)
   })
 
@@ -1124,7 +1221,7 @@ describe('default one-shot summarizer', () => {
       const { compact } = await summarizerHarness([], finish)
       let thrown: unknown
       try {
-        await compact.runSummarize('history', agent(conversation(1), MODEL))
+        await compact.runSummarize(promptInput('history'), agent(conversation(1), MODEL))
       } catch (error: unknown) {
         thrown = error
       }
@@ -1136,7 +1233,7 @@ describe('default one-shot summarizer', () => {
 
   it('rejects empty or reasoning-only successful output', async () => {
     const { compact } = await summarizerHarness([{ type: 'reasoning', text: 'private' }])
-    await expect(compact.runSummarize('history', agent(conversation(1), MODEL)))
+    await expect(compact.runSummarize(promptInput('history'), agent(conversation(1), MODEL)))
       .rejects.toThrow(/no text summary content/)
   })
 })
@@ -1304,7 +1401,7 @@ describe('automatic listener and loader composition', () => {
     expect(await recover(ctx, agent(session, MODEL), overflow())).toEqual({ action: 'retry' })
     expect(session.events.some(event => event.type === 'compact/summary')).toBe(true)
     expect(compact.calls).toHaveLength(1)
-    expect(compact.calls[0]!.text).toContain('tool result middle pruned')
+    expect(summarizedText(compact.calls[0]!.input)).toContain('tool result middle pruned')
   })
 
   it('retries from a durable prune when later overflow summarization throws', async () => {

packages/compact/compact-basic/tests/compact-loop-repro.spec.ts

@@ -81,7 +81,12 @@ class OverflowRecoveryAdapter extends LlmAdapter {
   }
 
   override async * stream(options: GenerateOptions): AsyncIterable<StreamChunk> {
-    if (options.system?.includes('You are a compaction engine')) {
+    // The cache-reusing summarizer replays the conversation prefix and marks
+    // its call only by the compaction instruction in the trailing user message.
+    const trailing = options.messages.at(-1)?.content
+      .map(block => (block.type === 'text' ? block.text : ''))
+      .join('') ?? ''
+    if (trailing.includes('acting as a compaction engine')) {
       this.summaryRequests.push(options)
       yield { type: 'block-start', index: 0, blockType: 'text' }
       yield { type: 'block-end', index: 0, block: { type: 'text', text: 'RECOVERY CHECKPOINT' } }