mirror of
https://github.com/deepseek-ai/deepseek-harness
synced 2026-08-15 21:04:50 +00:00
a0268d2509
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.
88 lines
3.6 KiB
TypeScript
88 lines
3.6 KiB
TypeScript
/**
|
|
* Compaction service seam (`ctx.compact`): implementations decide when to
|
|
* compact and replace a history range with one summary node by subclassing
|
|
* {@link CompactService}. This interface necessarily depends on session and LLM
|
|
* vocabulary; the rationale is in the
|
|
* [compaction Agent Note](../../../../.agents/notes/implemented/feature/2026-06-18-compaction-capability-seam.md).
|
|
* @module @deepseek-ai/dsh-compact
|
|
*/
|
|
|
|
import { Context, Service } from 'cordis'
|
|
import type { Session } from '@deepseek-ai/dsh-session'
|
|
import type { CompactionResult } from './types.ts'
|
|
|
|
export type { CompactionResult } from './types.ts'
|
|
export { toolPairingBalancedAfter, toolPairingBalancedBefore } from './tool-pairing.ts'
|
|
|
|
/** Why automatic policy is asking a backend to consider compaction. */
|
|
export type CompactionTrigger = 'pressure' | 'context-overflow'
|
|
|
|
/** Minimal agent context compaction needs without depending on the agent package. */
|
|
export interface CompactAgentContext {
|
|
session: Session
|
|
options: { provider?: string; model?: string }
|
|
}
|
|
|
|
declare module 'cordis' {
|
|
interface Context {
|
|
compact: CompactService
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Abstract compaction service. Implementations own trigger policy, retention,
|
|
* and summarization, and may consume a separate measurement service. A
|
|
* successful run replaces the selected surface span with one summary node and
|
|
* prevents concurrent compaction of the same session. Load one implementation
|
|
* per context as `ctx.compact`.
|
|
*/
|
|
export abstract class CompactService extends Service {
|
|
constructor(ctx: Context) {
|
|
super(ctx, 'compact')
|
|
}
|
|
|
|
/**
|
|
* Consider automatic compaction for one explicit trigger. Pressure policy
|
|
* uses the latest durable routed request, while context-overflow policy may
|
|
* force a useful balanced reduction even below the normal threshold. Return
|
|
* `null` when no safe range can be compacted. A single oversized retained
|
|
* unit or request envelope cannot be repaired through surface compaction.
|
|
*
|
|
* @param agent - agent context owning the session surface and routing options.
|
|
* @param trigger - normal pressure or provider-confirmed context overflow.
|
|
* @param signal - cancellation signal; model-backed implementations must forward it.
|
|
* @returns the compaction result, or `null` if no compaction was needed.
|
|
*/
|
|
abstract compactIfNeeded(
|
|
agent: CompactAgentContext,
|
|
trigger: CompactionTrigger,
|
|
signal: AbortSignal,
|
|
): Promise<CompactionResult | null>
|
|
|
|
/**
|
|
* Forcibly compact a range of surface nodes into a single summary node.
|
|
* `start` and `end` name an inclusive span by surface position, not numeric seq
|
|
* order; replacements can make visible seqs non-monotonic. Both edges must be
|
|
* balanced so assistant tool calls remain paired with their results. A model-
|
|
* backed implementation forwards cancellation and rejects active, missing,
|
|
* reversed, or unbalanced ranges. The target session is `agent.session`.
|
|
* Use {@link toolPairingBalancedBefore} and {@link toolPairingBalancedAfter}
|
|
* for the edge checks.
|
|
*
|
|
* @param start - first surface seq, inclusive.
|
|
* @param end - last surface seq, inclusive.
|
|
* @param agent - context whose session is mutated and whose routing options guide summarization.
|
|
* @param signal - optional cancellation; model-backed implementations must forward it.
|
|
* @throws when compaction is active or the range is missing, reversed, or unbalanced.
|
|
* @returns the appended event seqs, summary, replaced range, and token accounting.
|
|
*/
|
|
abstract compactRegion(
|
|
start: number,
|
|
end: number,
|
|
agent: CompactAgentContext,
|
|
signal?: AbortSignal,
|
|
): Promise<CompactionResult>
|
|
}
|
|
|
|
export default CompactService
|