@deepseek-ai/dsh-session

September 1, 2026 · View on GitHub

English | 中文

Summary

dsh-session provides the append-only session log that records an agent's whole interaction history — the single source of truth every model-visible fact flows through. The LLM message history is derived from the log (deriveMessages()), never stored separately, so replay is re-derivation from the same events and compaction can shadow older surface entries without deleting history. The package also provides the in-memory store (ctx.sessions), the typed SessionEvent vocabulary that plugins extend by declaration merging, and the surface layer that orders message-producing events. Persistence is deliberately a separate concern: backends subscribe to session/event and flush on session/flush. Choose it as the foundation of any agent session; it runs no model calls itself.

Table of Contents


Use this package

Mount dsh-session wherever a session must exist. It creates and holds event-sourced Session instances in memory; durable storage is layered on by a persistence plugin that subscribes to the session/event feed.

Create and inspect sessions

ctx.sessions.create() builds a live session bound to the calling fiber; get(id) and list() find sessions, and fork() creates a child session from a stable prefix of a live one.

const session = ctx.sessions.create(sessionId, { meta: { cwd: '/workspace' } })
ctx.sessions.get(sessionId)      // the live session
ctx.sessions.list()              // every live session, in creation order

Append and derive

session.append(type, data, opts?) commits one typed event — it snapshots and freezes the payload, validates it as lossless JSON, and notifies observers. session.deriveMessages() projects the log into the Message[] the model sees, incrementally and cached:

session.append('user/message', { role: 'user', content: [{ type: 'text', text: 'hello' }], source: { kind: 'user' } },
  { surfaceOp: 'append' })
session.deriveMessages()         // the derived model history

Surface events (user/message, assistant/message, tool/result) must declare how they join the ordered surface. An Assistant message embeds the exact compact provider stream that produced it; assistant/attempt, boundaries, and other log-only events never produce a message.

Read the log

session.seq reads the current log length without materializing an array, and session.eventAt(seq) reads one accepted, deeply frozen event by sequence number. session.snapshotEvents(fromSeq?, toSeqExclusive?) materializes a frozen, stable snapshot of a half-open range; a complete current snapshot is cached until the next append. Callers that only need a length or one event use seq or eventAt().

Session log positions use two numeric types. SessionSeq identifies an existing event or inclusive event watermark; SessionLogOffset identifies a gap, prefix length, or read boundary and may equal the event count. SessionSeqCursor adds the -1 “no event yet” value, while OptionalSessionSeq uses null when absence is data. The constructors validate non-negative safe integers, and the brands disappear at runtime, so durable JSON and wire values remain ordinary numbers.

Fork a session

ctx.sessions.fork(source, boundary?, childSessionId?) selects source events through an inclusive boundary seq (default: the current last event), requires the prefix to end outside an open turn, and creates a live child session with lineage metadata. A tool-time delegation that must branch mid-turn clips to a completed prefix instead.

The logical SessionHeader.isSeeded field reports whether fork history exists without exposing a positional integer. Session.inheritedEventCount retains the exact checked SessionLogOffset; ownEvents() returns events at and after that cut, and isOwnSeq(seq) accepts only an existing child-owned position. A low-level seeded constructor must supply an explicit seed and inheritedEventCount because the constructor seed can contain child-owned setup events after the inherited prefix.

Flush durable state

ctx.sessions.flush(session) dispatches the awaited durability checkpoint: every persistence listener flushes and the call settles after all of them. A producer that needs an immediate durability barrier awaits it instead of assuming the write-behind drained.


Understand the implementation

Implementation internals — click to expand

This section explains how the package realizes the behavior above; the observable contract is covered in Use this package.

Design concept

The package is built on event sourcing: a Session is an append-only log of typed SessionEvents, and everything else — model history, transcripts, telemetry, titles, persistence — derives from that stream. The surface is a derived projection: an incremental manager validates append candidates, advances the ordered view from committed events, and tracks a replaceGeneration that bumps on every committed rewrite. Model-visible means logged: anything that reaches a model request must be reconstructable from the log. Each model attempt that reaches settlement commits one event: assistant/message carries the assembled model-visible message plus its compact timed stream, while assistant/attempt retains a failed, retried, cancelled, or stream-error attempt without adding model history. A hard process loss before settlement leaves no durable attempt stream.

Request headers

request/header stores a full canonical snapshot of the non-history request envelope with reason initial, resume, change, or series. An explicit message-series start or a surface replacement writes a series snapshot when the envelope is unchanged; a simultaneous change uses startsSeries: true. Same-series steps, retries, and ordinary later turns inherit the latest snapshot. adapterDefaults distinguishes values resolved by the adapter from explicit settings, and foldRequestHeader() selects the latest snapshot. This self-contained record supports partial-window rendering and exact reconstruction at the cost of growth per message series; the reconstructable-requests Agent Note owns the detail.

Source map

FileRole
src/index.tsPlugin entry: SessionStore service, store lifecycle, fork, flush
src/types.tsSessionEventMap, SessionEvent, UserMessage, SessionHeader, TurnEndReasonMap
src/surface.tsOrdered surface projection, replacement validation, deriveEventMessage
src/request-header.tsrequest/header folding and reconstruction
dsh-util-valuesShared lossless JSON validation and detached snapshots
src/repair.tsCold repair of crash-orphaned logs
src/invariant.tsInvariant companion: seq, turn/step enclosure, tool call/result pairing

Append validation

Every append uses the shared iterative snapshotJsonValue() pass, which reads, validates, and copies each nested value once, so a stateful getter cannot supply one value to validation and another to storage. Non-lossless-JSON payloads (BigInt, cycles, sparse arrays, -0, exotic prototypes) are rejected at the append site, before any backend flush. The append path constructs each SessionSeq; surface events additionally validate marker shape, cited source-event sequences, and complete shadowed-node coverage for replacements.

Derived history

deriveMessages() caches each surface node's projection once and returns a fresh array per call over shared, deep-frozen messages; each of the three surface event types (user/message, assistant/message, tool/result) projects its own message kind — user content verbatim, the assembled assistant message with its provider and model, or a user-role tool result. Embedded Assistant streams and assistant/attempt events remain replay and diagnostic data only. A surface rewrite rebuilds the projection — there is no raw-log fallback, so the surface is the single source of derived history.

The request header

The loop logs a full canonical request/header snapshot (call config, adapter defaults, rendered system prompt, assembled tool schemas) at each loop-instance boundary and on change; foldRequestHeader(events) reconstructs it by selecting the latest snapshot, making every conversation request a pure function of the log. Route metadata (request/context) is separate logged state appended only when the provider, model, or capacity differs.


Further Exploration

The package-level contract is enough for most consumers; read these when you need the surrounding domain.


Model Experience

Derived message history

What the model sees

The model receives the complete messages from user/message, assistant/message, and tool/result surface entries verbatim — identities, roles, sources, and content blocks are the same values established at creation, and projections never mint identities. Direct prompts and injected context remain separate user/message events whose sources preserve their provenance. Embedded streams, assistant/attempt, boundaries, and other log-only facts add no message.

Token effect

Appended surface entries are resent on later steps. A replace surface operation removes the shadowed entries from future inputs without deleting their raw log records.

KV Cache effect

Appended surface entries preserve reusable prefixes. A replace operation invalidates reuse from the first shadowed message even though the underlying event log stays append-only.

Crash-repair result

What the model sees

If recovery finds an assistant tool request with no durable tool/call, its synthetic TOOL_NOT_STARTED result says The tool call was interrupted before the Harness recorded it as started. Retry it if it is still needed. If a durable tool/call has no result, its TOOL_OUTCOME_UNKNOWN result says The tool call was interrupted after it was recorded, but no result was durably recorded. Its outcome is unknown. Decide whether to retry from the tool semantics: retry only if the operation is read-only or idempotent; if it may have side effects, first verify external state or ask the user. Do not retry blindly.

Token effect

Zero tokens in an intact session. Each repaired call adds its retained risk-specific error text on resume.

KV Cache effect

Append-only; newly visible content follows the reusable request prefix and does not invalidate existing KV-cache entries.

Logged request header

What the model sees

The session reconstructs the system prompt, tool schemas, call config, and session prefix that the loop actually sent. Header events do not add a second copy to message history; the prefix is prepended outside deriveMessages().

Token effect

Zero duplicate tokens from logging. The reconstructed prefix, system text, and schemas still incur their normal per-request cost.

KV Cache effect

Logging causes no invalidation, and exact reconstruction preserves request-prefix identity. A later header with changed prefix, prompt, or schemas may invalidate reuse from its first difference.

Known Limitations and Deferred Work

These limits define when the session store needs special care. They are current package constraints, not a task backlog.

  • fork() cuts only at stable boundaries of live sessions — the selected prefix must end outside an open turn and the source must be in the store; forking a persisted-but-unloaded session is excluded from the fork API.
  • SESSION_FORMAT_VERSION names only the current logical representation — historical headers and events remain in adjacent format packages, while persistence publishes a complete supported chain before constructing Session; equal-version unknown events still require the envelope's explicit ignorable marker (mechanism).
  • TurnEndReasonMap omits the ACP-named refusal / max_turn_requests variants — producer-gated: they land when an adapter or the loop first emits them.
  • No session tree beyond fork — a pi-style entry tree over branched sessions is deferred unless a consumer needs more than boundary-based forking.

Dev Note

Working context for maintainers — click to expand

None.