2026-06-18-compaction-capability-seam.md

Agent Note: Compaction as a capability seam (abstract contract + basic backend)

Status: implemented

English | 中文

Problem

A long-running agent conversation grows without bound. As the event log accumulates turns, the derived message history eventually approaches the model's context window — the model then truncates mid-response (max-tokens) or degrades. Compaction is the mitigation: replace a run of older history with a concise summary, keeping recent context intact.

The session surface was built as the foundation for exactly this — an ordered projection over the event log with a surfaceOp: { op: 'replace', start, end } operation purpose-built to shadow a range of entries and insert a replacement, with sourceEventSeqs listing every source event so replay can validate that the replacement cites every event it removes. What remained was the plugin that decides what to compact and produces the summary.

Two forces shape the design. First, compaction policy and reusable token measurement vary independently: measurement belongs to the LLM-family ctx.tokenMeter service, while summarization can be a model call, a template, or a remote service. Second, SurfaceEventType is closed to the message-producing event types (user/message, assistant/message, tool/result); only those may carry surfaceOp. A bespoke compaction/* event therefore cannot itself appear on the surface — the compiler and Session's always-on append/seed boundary reject surfaceOp on it.

Decision

Compaction is a capability seam with separate Service Definition and Service Provider roles

Per the capability-seams Agent Note, compaction ships as separate packages so the contract, the algorithm, and (later) the consumer API evolve independently:

1. Interface — @deepseek-ai/dsh-compaction: an abstract CompactionEngine owning the ctx.compaction key, the CompactionResult vocabulary, the compaction/* session events, the manual failure taxonomy, and the canonical checkpoint message source. It declares compactIfNeeded(), compactNow(), and compactRegion() as abstract — the contract states what compaction does, not how.
2. Implementation — @deepseek-ai/dsh-compaction-basic: a concrete BasicCompactionEngine that consumes ctx.tokenMeter and owns the tail→head retention walk, summarization via ctx.llm.stream(), the surface replacement, the lock, pre-step pressure, and canonical context-overflow recovery. summarize() is its sole subclass hook; pricing and replay stay with the meter.
3. Model-free companion — @deepseek-ai/dsh-compaction-tool-result-pruner: a concrete optional service that rewrites oversized current tool/result nodes before the backend selects a summary range. It is not a second compaction implementation and does not implement CompactionEngine.
4. Human consumer — @deepseek-ai/dsh-command-compact registers argument-free /compact through ctx.commands and calls the backend-independent compactNow() operation. It is direct human control, not a model-facing tool.

The contract depends on dsh-session and dsh-llm — a deliberate deviation

The capability-seams Agent Note states the Service Definition package "depends only on cordis" (true of dsh-shell, whose vocabulary is self-contained). Compaction cannot honor that: its verbs act on an agent-owned Session (compactRegion(start, end, agent)) and its output uses the content vocabulary (CompactionResult.summary: ContentBlock[]). There is no way to express the contract without naming Session/SessionEvent (from dsh-session) and ContentBlock (from dsh-llm).

This is not a coupling smell — it is the contract's domain. The "only cordis" guidance was always shorthand for "the interface depends only on what the contract genuinely names, and never on an implementation." dsh-session and dsh-llm are themselves interface/vocabulary packages, not implementations; dsh-compaction still imports no backend. The seam's real invariant — consumers and implementations evolve independently behind an abstract service — holds intact.

Three abstract operations, algorithm in the backend

Putting the full algorithm (the retention walk, token-summing, text extraction) as concrete methods on the interface recouples the contract to one strategy: a backend that wants a different retention policy or event sequence would have to fight inherited concrete code. Making all three operations abstract puts every how decision in the backend and keeps the interface a statement of what. Token measurement is not a compaction hook at all; the singleton service lets multiple consumers share one per-session replay fold.

compactIfNeeded(agent, trigger, signal) takes an explicit 'pressure' | 'context-overflow' trigger and cancellation. It reads only the latest durable routed request; no header means no work, while any routed provider/model target uses the singleton estimator. compactNow(agent, signal) requires an idle agent and performs one useful balanced reduction even below pressure, returning null without writes when none exists. compactRegion(start, end, agent, signal?) uses agent.session as its single session identity and keeps an optional signal for explicit callers. The default summarizer resolves its target from explicit config, the latest logged routed target, then agent options, and records the provider/model pair after any llm/stream routing. It replays the routed request's prefix and appends the compaction directive as a trailing user message so the provider's warm KV cache is reused — see the summary prefix-cache Agent Note. The result carries llmStreamCall: true because it consumed exactly one call through this context's LLM service; a subclass sets that marker only under the same condition, since retained rawOutput alone does not identify the call path. The call sets the provider-neutral GenerateOptions.purpose to compaction; adapters may map that purpose to model-hidden transport metadata, and the DeepSeek adapter sends x-deepseek-harness-compact: 1.

Automatic pressure runs after successful durable step work

Successful-call pressure runs at the next agent/pre-step, after the preceding response, tool results, buffered context, and steering are durable and before the next request is derived. dsh-compaction-basic measures the canonical logged request through ctx.tokenMeter, so the next request sees any replacement without a speculative envelope override. Once pressure qualifies, optional ctx.toolResultPruner rewriting runs before summary selection; compaction-basic remeasures the durable surface and skips summarization if pruning restores safe pressure.

Canonical provider context overflow takes a separate path. The failed step closes and agent/request-error receives the original request error. Compact-basic owns its per-agent overflow count, prunes before forcing one useful balanced reduction, and returns { kind: 'retry' } only if session.surface.replaceGeneration increases, including pruning-only progress when no summary range exists. The loop then closes the failed turn, opens a new numbered retry turn, and reconstructs its request from the durable log. No replacement, a recovery failure before any replacement, cancellation, an exhausted cap, or an unrelated error preserves the original provider failure. If pruning already advanced the generation before later summary work fails, recovery retries from that durable pruned surface unless cancellation or disposal wins. The complete lifecycle decision is in the after-call recovery Agent Note.

assistant/message → tool/result/context/steering → step/end
claim the next batch → await waterfall agent/pre-step  ⟵ pressure compaction before the next request
enter → next step/start

provider overflow → step/end
await waterfall agent/request-error  ⟵ forced compaction between attempts
retry → next numbered step/start      ⟵ derives from the replacement surface

Retention is turn-agnostic; tool-pairing balance is the only structural guard

Auto-compaction checks after every successful step, not once per turn. This is load-bearing for runaway-turn survival: a tool-heavy ReAct turn appends an assistant/message + a tool/result per step, so the surface grows within a turn. The next pre-step check can compact early closed tool pairs before continuation opens another step, and provider-confirmed overflow remains the backstop when a request crosses the limit first.

compactIfNeeded retains the smallest tail of whole surface units whose estimated size reaches the resolved retained-token budget and compacts older nodes. A unit is a complete closed step or one no-step message. If the token cutoff lands inside a step, retention expands until the cut is tool-pairing balanced. Balance is checked on surface order, not log sequence, because replacement summaries have new sequence numbers at old surface positions. dsh-compaction exports the before/after edge helpers; their per-session cache folds only appended surface-tail nodes while replaceGeneration is unchanged, does no event reads for log-only growth, and rebuilds current membership and balances after replacement. compactRegion rejects boundaries that split a tool call from its result. The in-flight turn receives no special retention.

A runaway turn thus compacts exactly like any other history: its early closed steps get summarized while its recent steps stay verbatim. When the only compactable content left is an un-splittable open tail step (its tool-calls have no results yet), compaction declines (null) and retries once that step closes.

Some single-unit overflow remains out of scope. Summary range selection cannot split an indivisible unit. The optional pruner can repair a closed tool pair when removable text-bearing tool-result content is the bulk and the pruned remainder fits. Envelope-only pressure, an oversized indivisible non-tool node such as a pasted user/message, and a tool unit whose non-prunable remainder is still oversized remain outside compaction; bounding those units is a separate concern.

Head-anchoring: one auto checkpoint, always at the head

Auto-compaction always starts at the surface head, merging the prior checkpoint with newly compacted history so only one automatic checkpoint remains. shadowedRange is therefore positional rather than a numeric sequence interval: a newer summary sequence may occupy an older surface position. shadowedSeqs records the authoritative surface order. Manual mid-range compaction may leave multiple checkpoints.

Approximate convergence invariant

resolveConfig supplies usable defaults: threshold ratio 0.8, retained-tail ratio 0.16, empty summarization provider/model overrides, maxTokens: 8192, compactionRetries: 1, maxOverflowRetries: 1, and auto: true. Optional exact provider/model policies partially override the top-level defaults; pressure scales ratios against capacity from the route-owning LLM adapter, while retainTokens can replace ratio retention. Retention must remain below the resulting threshold. Convergence remains dynamic because provider output caps can be spent on hidden or surfaced reasoning tokens and summary size is unpredictable. If pressure remains over threshold, compactIfNeeded() re-compacts the head checkpoint up to the configured retry count, but each committed summary must be smaller than what it shadows. Overflow needs no capacity metadata and bypasses threshold and retained-tail policy for one maximal balanced head reduction, leaving the newest indivisible unit. The ownership split is specified by the routed model context and compaction policy Agent Note.

Surface replacement: compaction/* events are log-only; one user/message carries the summary

Because SurfaceEventType is closed, the summary cannot ride on a compaction/* event. The backend instead appends a single user/message with source: COMPACT_CHECKPOINT_SOURCE and surfaceOp: { op: 'replace', start, end } whose content is the (framed) summary and whose sourceEventSeqs covers the shadowed entries and the bookkeeping events. The interface exports that source and isCompactCheckpointSource() so consumers recognize a persisted or cloned checkpoint without depending on backend package identity. The compaction/* events record the lock, summary, selected range, shadowed seqs, token count, and model call without joining the surface. The surface mutation sits inside the lock — compaction/end is the last event appended:

compaction/start    → log-only. Acquires the lock.
[summarize older range via the backend]
compaction/summary  → log-only. Records the raw summary, local-call marker, range, shadowed seqs, and token count.
user/message     → canonical checkpoint source + surfaceOp { op:'replace', start, end }.
                   THE surface mutation (framed summary).
                   deriveMessages() renders it as a user-role message.
compaction/end      → log-only. Releases the lock (carries `error` on a recoverable failure).

deriveMessages() then yields [summary_as_user_message, ...retained_entries]. Reusing user/message is honest rather than a workaround: a summary genuinely is user-role context.

Checkpoint framing + incremental merge (backend-private)

The basic backend wraps the summary as established checkpoint context and tags it for incremental merging on the next cycle. The raw summary remains on compaction/summary. Framing is backend policy; the seam promises that one replacement user message carries the possibly framed summary and uses the canonical checkpoint source.

Blocking via a log-recorded lock, plus a crash/recoverable failure taxonomy

The compaction/start … compaction/end bracket is justified by two roles:

1. Crash-detectable orphan plus recorded summary inputs (primary). Summarization is a slow model call persisted after compaction/start. A crash mid-summarization leaves a compaction/start with no matching compaction/end — a detectable orphan. Releasing the lock last (rather than first) converts the crash window from silent corruption into that detectable orphan.
2. Prevents concurrent compaction. Every automatic, manual, and explicit-range entry point refuses a live unmatched compaction/start. The bracket is the single lock; no process-local mutex duplicates it.

The lock excludes another compaction, not unrelated facts. Its markers are time points rather than an exclusive container, so durable inbox splices may appear between a standalone manual start and end. Automatic work requires whole-surface stability inside its turn. Manual work revalidates only the selected positional span, letting append-only context outside it remain visible after replacement.

The lifecycle boundary makes crash state unambiguous:

• Current lifecycle: a dangling compaction/start after the newest session/end-seed is the live durable lock and reports busy.
• Later lifecycle: a newer constructor-written session/end-seed proves that the older unmatched start is stale, so resume, fork, and adoption do not remain wedged by a dead writer.
• Recoverable failure: once start lands, the backend makes exactly one compaction/end { error } attempt. Summary or stability failure leaves the conversation surface unchanged while preserving the failed attempt in the log. If the close append fails, the unmatched start remains intentionally blocking.

compaction/end keeps its error? field (mirroring tool/result's self-contained error — one event tells success from failure without correlating a sibling). There is no separate compaction/error event.

Core session repair stays compaction-agnostic — deliberately. interruptedTurnClosers is never taught about compaction/*. The general session/end-seed lifecycle boundary supplies the evidence the compaction owner needs; the compaction invariant and backend interpret it without adding plugin-specific repair to core.

Alternatives considered

• The full algorithm as concrete interface methods — rejected because it recouples the contract to one retention strategy. All three operations are abstract; reusable measurement is a separate LLM-family service and summarize() is basic's sole hook.
• Compaction on agent/request or a compaction-specific loop callback — rejected because the former observes a provisional request and the latter couples generic lifecycle to compaction policy. Pre-step replay of the prior durable request plus canonical overflow recovery covers successful and rejected calls.
• A compact boolean or untyped request metadata map — rejected because multiple auxiliary call kinds would become mutually exclusive flags, while an open bag would discard compiler-checked vocabulary. One typed purpose discriminant extends with additional call kinds without adding another GenerateOptions field.
• A separate compaction/error event — rejected: compaction/end keeps an error? field, mirroring tool/result's self-contained error — one event tells success from failure without correlating a sibling.
• Teaching core turn-repair about compaction/* — rejected: the general end-seed boundary already distinguishes prior-lifecycle history, and patching core for every future xxx/start … xxx/end pair is exactly the coupling the capability-seam architecture exists to avoid.

Consequences

• Packages: packages/compaction/compaction supplies the interface, compaction-basic supplies the backend, compaction-tool-result-pruner supplies optional deterministic rewriting, and command-compact supplies human /compact. packages/llm/token-meter owns replay-aware measurement independently.
• Automatic extension points: agent/pre-step (@mode waterfall) handles pressure before request derivation and agent/request-error (@mode waterfall) handles final request failures after the failed step closes. The pre-step payload carries the claimed batch, turn, step, and signal (see the payload-object events decision), with no compaction-only prompt/prefix payload.
• SessionEventMap gains compaction/start / compaction/summary / compaction/end by declaration merging (merge-extensible); SurfaceEventType is not touched. These are session events, not cordis Events, so the event-taxonomy gate needs no entry.
• dsh-compaction owns COMPACT_CHECKPOINT_SOURCE, isCompactCheckpointSource(source), toolPairingBalancedBefore(session, seq), and toolPairingBalancedAfter(session, seq). The marker identifies replacement summaries across backend implementations. The cached surface-edge checks prevent compactRegion and compactIfNeeded from splitting a tool-call/result pair, validate current membership by seq, answer both edges from one per-cut balance sequence, and reject stale or missing seqs and orphan results.
• dsh-session validates positional replacement, complete cited source-event coverage, and content-only single-node tool/result rewrites through its one surface manager. Its invariant companion treats fresh appended tool results as executions that require an open step and pending call, while the compaction companion owns numeric-turn versus standalone-null bracket relations.
• Wiring: examples/tui-agent/cordis.yml loads zero-config dsh-token-meter, dsh-compaction-tool-result-pruner, dsh-compaction-basic, then dsh-command-compact; service-wide defaults make the composition usable without repeated numeric policy.

Testing

• Unit: Real Loader and invariant plugins cover whole-unit retention, pruning configuration and replay, rich-block ordering, metadata preservation, convergence, both compaction/end outcomes, open-tail refusal, pruning-only and summarized overflow recovery, generation proof, caps, and original-error preservation.
• Loop: Tests pin pre-step after the preceding step/end and before the next step/start, actual agent/request routing, closed failed steps, fresh retry numbering, and complete thrown/in-band overflow → compaction → reconstructed retry composition.
• Manual: Maintenance serialization, marker ordering, injection retention, live/stale orphan classification, cancellation, close/flush failures, command mapping, and the queued TUI journey are pinned without a model key.
• With-key e2e: A real model and bash session with lowered limits triggers compaction, records a complete compaction/start…end pair, shrinks the surface, and finishes the task.
• Snapshot: The assembled context-overflow scenario derives the auxiliary call from compaction/summary only when llmStreamCall: true proves that the local LLM service consumed it; canonical reconstructed blocks pin the complete recovery without provider delta partitioning.