fix(compact-basic): replay conversation prefix so summarization reuses KV cache

Automatic compaction fires mid-conversation, right after the loop warmed
the provider's KV cache with the last routed request. The default
summarizer then issued a separate request whose prefix shared nothing
with that warm request — a bespoke summarizer system prompt followed by
the older history flattened to one rendered transcript string — so a
differing first token invalidated the entire cached prefix and every
compaction re-processed the whole replayed history twice.

Move the compaction directive from the FRONT (a fresh system prompt) to
the END (a trailing user message), and replay the last routed request's
own system prompt, tools, message prefix, and shadowed-region messages
verbatim via session.requestHeader() + deriveEventMessage. The auxiliary
call is now a genuine prefix-extension of the warm request, so the
provider reuses the cached tokens up to the trailing instruction.

SummarizationInput carries the replayed prefix instead of a flat string;
the now-unused renderTranscript/renderContentBlocks path is removed with
its spec. Cache reuse is best-effort (head compaction guarantees a hit;
a mid-range compaction or a differently-routed summarizer forgoes it),
correctness is not.

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

@@ -17,6 +17,7 @@ import type {} from '@deepseek-ai/dsh-compact-tool-result-prune'
 import { resolveConfig } from './config.ts'
 import { compactSurfaceRegion, selectCompactableRange } from './region.ts'
 import { summarizeWithLlm } from './summarizer.ts'
+import type { SummarizationInput } from './summarizer.ts'
 import type {
   BasicCompactConfig,
   ResolvedConfig,
@@ -135,19 +136,21 @@ 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 }> {
-    return summarizeWithLlm(this.ctx, this.config, text, agent, signal)
+    return summarizeWithLlm(this.ctx, this.config, input, agent, signal)
   }
 
   /**
@@ -236,7 +239,7 @@ 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'
 import type { ResolvedConfig } from './types.ts'
 
@@ -14,9 +14,15 @@ import type { ResolvedConfig } from './types.ts'
 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.',
   '',
@@ -47,14 +53,30 @@ 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[]
@@ -64,10 +86,12 @@ 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.
@@ -75,7 +99,7 @@ export interface SummaryResult {
 export async function summarizeWithLlm(
   ctx: Context,
   config: ResolvedConfig,
-  text: string,
+  input: SummarizationInput,
   agent: Agent,
   signal?: AbortSignal,
 ): Promise<SummaryResult> {
@@ -97,14 +121,16 @@ 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 },