deepseek-harness/docs/subsystems/compaction.md

Compaction

English | 中文

The compaction seam — a capability seam split like bash: interface (dsh-compact, ctx.compact), implementation (a backend such as dsh-compact-basic), and human consumer (dsh-command-compact). Compaction is one optional capability, not part of the agent-loop spine — so its vocabulary lives here, not in core.md. A tokenizer- or template-based backend is a sibling package implementing the same interface. Unlike bash, the interface necessarily depends on dsh-session and dsh-llm: its verbs act on an agent-owned Session, and its durable summary event uses the ContentBlock vocabulary (see the compaction capability-seam Agent Note).

Source: packages/compact/compact/src/types.ts

The compact/* session events

Compaction extends SessionEventMap with three event types via declaration merging. All three are log-only — they record the compaction lock and its provenance, and never join the surface. SurfaceEventType is deliberately NOT extended (only message-producing events reach the model), so the summary itself rides on a separate user/message with surfaceOp: { op: 'replace', start, end } — the only surface mutation performed by summary compaction. See the Agent Note for why reusing user/message is honest rather than a workaround.

Event	Payload	Role
compact/start	{ turn }	acquires the log-recorded lock; a number identifies the open automatic turn, while null identifies a standalone manual attempt
compact/summary	{ summary, rawOutput?, llmStreamCall?, shadowedRange, shadowedSeqs, shadowedTokenCount, provider, model, maxTokens?, usage? }	provenance: the safe summary projection, optional complete provider output and usage, an llmStreamCall: true marker when producing the result consumed exactly one call through this context's ctx.llm.stream() (which requires complete rawOutput), the shadowed surface-boundary pair (start/end seqs — a position span, not a numeric interval), the shadowed seqs in surface order, the estimated token count, and the summarize call's envelope (provider, model, plus its generation cap when one applied) — logged so the one-shot request is reconstructable from log + code (the reconstructability Agent Note); unmarked rawOutput does not identify the call path
compact/end	{ turn, error? }	releases the lock with the same numeric-or-null owner (error records an unsuccessful attempt)

The lock brackets the whole operation: compact/start is appended first, then summarization, the compact/summary provenance record, and the user/message replacement all land, and only then compact/end. Releasing the lock last turns a crash mid-operation into a detectable orphaned lock (a compact/start with no matching compact/end) rather than a compact/end that falsely claims compaction finished.

The markers are lock time points, not an exclusive container. An unrelated idle injection can appear between a standalone manual start and end while summarization is pending. The manual path revalidates only its selected positional span, so that injected context survives after the replacement checkpoint. A live unmatched start blocks every entry point; an unmatched start before a newer session/end-seed is stale evidence from a prior lifecycle and is ignored.

These variants are merged inside a declare module '@deepseek-ai/dsh-session' block, so — unlike the top-level types on the other subsystem pages — they are not pasted as a drift-checked ```ts type-equiv block (the verify-type-equiv extractor matches only top-level declarations by name). The payload table above is the catalog entry; follow the source link for the authoritative shapes.

CompactionResult

What a successful compaction returns to its caller: the bookkeeping-event seqs, safe summary projection, shadowed range and seqs, and estimated token count.

/** Result of a successful compaction operation. */
interface CompactionResult {
  /** The seq of the appended `compact/start` event. */
  startSeq: number
  /** The seq of the appended `compact/summary` event. */
  summarySeq: number
  /** The seq of the appended `compact/end` event. */
  endSeq: number
  /** The summary content blocks produced by the backend. */
  summary: ContentBlock[]
  /**
   * The surface-boundary pair that was shadowed: the seqs of the first
   * (`start`) and last (`end`) surface nodes of the replaced range. A
   * surface-POSITION span, not a numeric seq interval — after a prior replace
   * lands a fresh high-seq summary node at an older range's position, `start`
   * can be GREATER than `end`. {@link CompactionResult.shadowedSeqs} is the
   * authoritative set of shadowed nodes, in surface order.
   */
  shadowedRange: { start: number; end: number }
  /** The seqs of all shadowed surface nodes, in surface order. */
  shadowedSeqs: number[]
  /** Estimated token count of the shadowed content. */
  shadowedTokenCount: number
}

The service

Automatic callers state why policy is running; implementations may treat confirmed overflow more aggressively than ordinary pressure.

/** Why automatic policy is asking a backend to consider compaction. */
type CompactionTrigger = 'pressure' | 'context-overflow'

CompactService exposes compactIfNeeded(agent, trigger, signal) for automatic pressure or context-overflow policy, compactNow(agent, signal) for one useful idle-session reduction even below pressure, and compactRegion(...) for an explicit inclusive surface range. compactNow() runs as agent maintenance between turns, returns null without writing when no useful range exists, records a standalone turn: null bracket before summarization, and flushes a closed attempt before later queued prompts may derive from the new surface. Every backend marks its replacement user/message with COMPACT_CHECKPOINT_SOURCE; client and wire consumers import that value and isCompactCheckpointSource() from the cordis-free @deepseek-ai/dsh-compact/checkpoint subpath, while the package root re-exports both for host consumers. The predicate keeps checkpoint recognition independent of any one backend. Implementations must forward the supplied signal to summarization. The seam owns no pricing API: the singleton ctx.tokenMeter directly owns estimation and replay, while dsh-compact-basic owns retention, event sequencing, routed summarization calls, and their configuration.

Expected manual failures use ManualCompactionErrorCode:

/** Expected failure classes for an explicit idle-session compaction request. */
type ManualCompactionErrorCode =
  | 'busy'
  | 'cancelled'
  | 'changed'
  | 'summary'
  | 'commit'
  | 'persistence'

changed and summary leave the conversation surface unchanged but still close and persist the failed attempt in the log. commit may follow partial mutation; persistence means the in-memory bracket closed but its flush failed. Cancellation remains separate and throws the exact abort reason after required cleanup.

Pressure compaction runs at serial agent/pre-step before request derivation. Once pressure or canonical overflow qualifies, compact-basic invokes optional ctx.toolResultPrune before range selection, remeasures through ctx.tokenMeter, and can advance the surface without a summary. Failed-request recovery runs through agent/request-error after the failed step closes and returns a retry action only when the surface replacement generation advances, even if later summary work throws after pruning; cancellation still wins. Region boundaries preserve tool-call/result pairing but not whole turns, allowing early closed steps of one oversized turn to compact. dsh-compact-basic owns thresholds, retained-tail policy, overflow caps, and failure handling.

The seam exports toolPairingBalancedBefore(session, seq) and toolPairingBalancedAfter(session, seq) for those edge checks. Both validate current surface membership and reject missing seqs and orphan results; the package contract owns their cache semantics.

Tool-result pruning outcomes

The optional tool-result pruning service reports each durable content replacement and the aggregate Unicode-code-point reduction. Its public result types live in compact-tool-result-prune/src/types.ts.

/** Provenance and size accounting for one landed surface replacement. */
interface PrunedEntry {
  /** Full-fidelity tool-result event shadowed by the replacement. */
  readonly originalSeq: number
  /** Newly appended pruned tool-result event. */
  readonly replacementSeq: number
  /** Tool call shared by the original and replacement. */
  readonly callId: CallId
  /** Original text size in Unicode code points. */
  readonly charsBefore: number
  /** Replacement text size in Unicode code points. */
  readonly charsAfter: number
}

/** Aggregate outcome of one stable-surface pruning pass. */
interface PruneResult {
  /** Replacements in the snapshotted surface order. */
  readonly pruned: readonly PrunedEntry[]
  /** Total Unicode code points removed across replacements. */
  readonly charsRemoved: number
}

Cordis surface

Generated from source by scripts/gen-cordis-catalog.ts (verified fresh by pnpm run verify-cordis-catalog in doc-sync; regenerate with pnpm run gen-cordis-catalog) — this section is byte-identical in both language sides of the page. Signature blocks use a ts cordis-catalog fence and keep the original source JSDoc; dispatch modes are defined in the primer, and the framework-inherited ctx surface lives in cordis-api/inherited.md.

ctx.compact — CompactService (abstract seam)

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. The replacement user message uses COMPACT_CHECKPOINT_SOURCE so consumers recognize it independently of the backend. Load one implementation per context as 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>

/**
 * Explicitly compact useful history even below automatic pressure thresholds.
 * Implementations synchronously start an idle task before any asynchronous
 * work, select a useful range without writing on a no-op, then
 * append a standalone `compact/start` before summarization. That durable
 * marker is the compaction lock until one `compact/end` attempt. Later waking
 * prompts remain accepted in FIFO order and start only after the optional
 * durability checkpoint and idle-task settlement. Context injected while the
 * summary runs may sit between the marker pair; only the selected span must
 * remain stable.
 *
 * @param agent - idle agent whose durable history should be compacted.
 * @param signal - cancellation scoped to this compaction request.
 * @returns the compaction result, or `null` when no safe useful range exists.
 * @throws {@link ManualCompactionError} for expected busy, agent-cancellation,
 * changed-span, summarization/shrink, commit-stage, or persistence failures;
 * an aborted request preserves its exact abort reason. Failed attempts remain
 * visible in the log.
 */
abstract compactNow( agent: ManualCompactAgentContext, 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`.
 * Its replacement user message must use {@link COMPACT_CHECKPOINT_SOURCE}.
 * 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>

Source: packages/compact/compact/src/index.ts:93

ctx.toolResultPrune — ToolResultPruneService

Deterministic head/middle/tail pruning for current tool-result surface nodes.

/**
 * Measure text content in Unicode code points; non-text blocks cost zero.
 * @param blocks - tool-result content to measure.
 * @returns total Unicode code points across text blocks.
 */
measureContent(blocks: readonly ContentBlock[]): number

/**
 * Replace an over-budget text middle while retaining rich-block order.
 * Text slicing is by Unicode code point, not UTF-16 code unit, so a retained
 * boundary cannot split a surrogate pair. Grapheme clusters may still split.
 * @param blocks - original tool-result content.
 * @returns pruned content, or `null` when the text is within budget.
 */
pruneContent(blocks: readonly ContentBlock[]): ContentBlock[] | null

/**
 * Prune every over-budget tool result from one stable current-surface snapshot.
 * Each replacement preserves the complete event data except for `content`,
 * points at the shadowed node for durable provenance and replay, and is
 * immediately preceded by a `compact/prune` shadow-price event pricing the
 * shadowed node through the injected token meter, so pure consumers can
 * subtract it without per-node state.
 * @param session - session whose current surface is rewritten.
 * @returns landed replacements and aggregate Unicode-code-point savings.
 * @throws when the session rejects a replacement; replacements committed
 * earlier in the pass remain durable.
 */
pruneSession(session: Session): PruneResult

Types: ContentBlock · Session

Source: packages/compact/compact-tool-result-prune/src/index.ts:44