deepseek-harness/docs/core-data-structures/persistence.md

Session Persistence

English | 中文

The durability seam for the event log. session.md describes the in-memory Session — the append-only SessionEvent log that is the source of truth. This page describes how that log is made durable: the abstract SessionPersistence service, its backends, the flush checkpoint, crash recovery, and the metadata header that travels alongside the log. The event vocabulary the log carries is enumerated, member by member, in the generated persistence log event catalog.

The seam is a textbook capability seam: one abstract service (dsh-session-persistence, ctx.sessionPersistence) defining locate/create/append, crash-repairing load, non-mutating inspect, and lightweight list/snapshot observation over the existing SessionEvent — no parallel persisted type — and two interchangeable backends that pass the same runPersistenceContract suite. See the session-persistence Agent Note.

The flush checkpoint

session/event is a synchronous notification; persistence plugins copy the event into a per-session controller and start an eager write without blocking the producer. Concurrent events share the current batch, and events admitted during that write trigger a follow-up batch. session/flush waits until no current or pending batch remains, so the loop still uses it as the ordering and error-observation checkpoint before claiming the next ordinary turn. A rejected eager write retains its events; an explicit flush retries them and reports failure through agent/error and the logger, never as a session event past the closed turn. Disposal performs the same final drain.

Crash recovery preserves an interrupted turn

A backend that reloads a log crashed mid-turn finds an open turn/start with no turn/end. It does not truncate — a single turn can be huge in a long-horizon task (many steps, large tool output), and those events were durably appended before the crash. Instead it closes the orphaned turn with a synthetic turn/end { reason: { kind: 'interrupted' } }, keeping the interrupted execution balanced without changing any standalone events before or after it. interrupted is the one TurnEndReason no loop emits (see session.md).

Repair applies only to cold sessions. For a live id, SessionPersistence.load(id) snapshots the in-memory log, waits until that snapshot is durable, and returns it with the stored header only when balanced; an open live turn rejects rather than receiving synthetic interruption boundaries. A coordinator-backed cold load reserves the id across backend reads and repair writes, so concurrent publication of a same-id live session rejects and rolls back. HMR also adopts a live prefix without closing its active turn.

SessionPersistence.inspect(id) is the observer counterpart to recovery: it returns a detached valid stored prefix without truncating a torn record, adding interruption closers, or publishing write state. Same-id serialization keeps it coherent with backend writes. Derived read models use inspect, never load, so observing a checkpointed open turn cannot mutate the log if live ownership begins concurrently.

SessionLocation — optional per-session artifact target

SessionPersistence.locate(meta) synchronously resolves a backend-owned independent artifact without reading, creating, or flushing it. JSONL returns the absolute transcript path inside its project/session directory; SQLite returns undefined because sessions share one database. A returned path can therefore name a file that does not yet exist or lacks the current unflushed turn; it is a location hint, not authorization or a freshness guarantee.

/**
 * A backend-resolved, per-session local artifact location. The path is an
 * absolute target path and can name an artifact that has not materialized yet.
 * Consumers must treat it as a location hint, never as an authorization token.
 */
interface SessionLocation {
  /** Backend-specific artifact kind, for example `jsonl`. */
  readonly kind: string
  /** Absolute path to this session's backend-owned artifact. */
  readonly path: string
}

SessionHeader — metadata beside the log

Per-session metadata travels separately from the event log: format version, cwd, lineage, and the seed boundary are storage concerns, not conversation events, so they stay out of SessionEventMap and never reach deriveMessages(). The header is attached to a Session via session.header.

Source: packages/core/session/src/types.ts

/**
 * Immutable validated storage metadata, kept outside the conversation event log.
 */
interface SessionHeader {
  /**
   * On-disk format version, stamped from {@link SESSION_FORMAT_VERSION} when the
   * session is created. A persistence backend rejects any other version on load
   * (no migration — see the constant).
   */
  readonly version: number
  /** The session's id (mirrors the {@link Session}'s id). */
  readonly id: SessionId
  /** Non-negative safe-integer Unix epoch milliseconds when the session was created. */
  readonly createdAt: number
  /** Absolute working directory the session was created in (if any). */
  readonly cwd?: string
  /** The session this one was forked from (seed lineage), if any. */
  readonly parentSession?: SessionId
  /**
   * How many leading events were inherited through a seed. Persisting this
   * boundary lets resume and replay distinguish parent history from child work.
   */
  readonly seedLength?: number
  /**
   * Coarse product classification for a session created as a subagent child.
   * This is presentation metadata, not proof that the child is continuable.
   */
  readonly origin?: 'subagent'
  /**
   * Delegation depth: absent (zero) for a top-level session, parent depth + 1
   * for a subagent child. Persisted so a recursion budget survives restart and
   * resume — a runtime-only depth would reset a resumed child to top-level.
   */
  readonly delegationDepth?: number
}

CreateSessionOptions — seeding and metadata

Creating a Session through the store takes a seed (initial replay or fork history) and meta (the storage-level fields the store folds into a SessionHeader). The store fills in version/id and defaults createdAt; the caller supplies the validated absolute cwd, the parentSession lineage, the seedLength seed boundary, the optional coarse origin, the delegationDepth, and — only when reconstructing a persisted session — the original createdAt to preserve it. origin: 'subagent' lets product navigation hide duplicate child rows; it does not prove that a descriptor is valid or that the child can resume.

/**
 * Options for creating a {@link Session} via the store. `seed` replays/forks
 * an existing event log; `meta` carries the caller-supplied storage fields the
 * store folds into a {@link SessionHeader}.
 */
interface CreateSessionOptions {
  /** Initial replay or fork history supplied at construction. */
  readonly seed?: readonly SessionEvent[]
  /**
   * Storage metadata read once before publication. `seedLength` is explicit
   * because a resumed seed contains the full stored log, not only its inherited prefix.
   */
  readonly meta?: {
    readonly cwd?: string
    readonly parentSession?: SessionId
    readonly createdAt?: number
    readonly seedLength?: number
    readonly origin?: 'subagent'
    readonly delegationDepth?: number
  }
}

Replay/fork is therefore ctx.sessions.create(id, { seed: seedEvents }); resuming a persisted session into a live agent is ctx.agents.resume({ resumeSessionId }).

Lightweight source revisions

Consumers of derived state compare a cheap opaque revision before loading a full event log. The persistence backend owns its representation and changes it transactionally with append or mutating load repair; callers compare it only for equality.

/**
 * Backend-owned token that identifies both one storage source and one revision
 * of a persisted session log.
 */
type SessionPersistenceRevision = Branded<'SessionPersistenceRevision'>

/** Lightweight immutable source identity returned without loading a full log. */
interface SessionPersistenceSnapshot {
  /** Detached metadata for one materialized session. */
  header: SessionHeader
  /** Opaque source-qualified token that changes whenever this stored log changes. */
  revision: SessionPersistenceRevision
}

The backends

Both implement the same abstract SessionPersistence (locate/create/append/load/inspect/list/listSnapshots over SessionEvent, with optional cancellation on observation methods) and pass runPersistenceContract, proving the seam is genuinely backend-agnostic:

• dsh-session-persistence-jsonl — an append-only logical JSONL log per session, stored as checksummed concatenated Zstandard frames by default or raw lines by configuration, with crash-safe atomic writes, interrupted-turn recovery, and a read/replay path.
• dsh-session-persistence-sqlite — node:sqlite, one row per SessionEvent. The row shape (session_id, seq, type, time, data, source_event_seqs, surface_op) maps 1:1 onto the event, including optional surface metadata, so there is no parallel persisted schema to keep in sync.