deepseek-harness/packages/context/session-reference

@deepseek-ai/dsh-session-reference

English | 中文

ctx.sessionReferences prepares bounded, read-only snapshots of other sessions as sourced model-facing context. It consumes ctx.sessionQuery and the backend-independent compact checkpoint marker; SQLite FTS is not required. Hosts that support cross-session mentions may opt into the service.

Public API

• listCandidates(agent, query?, limit?) lists sessions other than agent.id, filters case-insensitively by id, cwd, or the latest log-backed title, and ranks same-cwd, cwd-less, then other-cwd records while preserving listSessions() creation order within each group. Each selected candidate uses that title as the mention label and falls back to the session id when the title is absent or unreadable; message bodies are not searched.
• prepare(agent, content, references, signal?) preserves first-mention order, deduplicates ids, rejects self-reference and more than the configured distinct-source limit, reads every source in parallel, and returns detached content plus zero or one aggregated, identified UserMessage context. Any invalid reference, failed read, cancellation, or budget failure rejects before the host calls followup() or steer().
• encodeSessionReferenceUri() and decodeSessionReferenceUri() implement dsh-session:<base64url(JSON.stringify(sessionId))> so every JavaScript string id round-trips exactly. formatSessionReferenceMention() emits @[label](uri), and parseSessionReferenceText() replaces Markdown mentions or bare canonical URIs with readable @label text while returning structured references. Explicit Markdown mentions reject every malformed URI; bare text is considered a reference only when a non-empty base64url-shaped payload follows the scheme, and a matching noncanonical candidate still fails. Empty or punctuation-only scheme mentions remain ordinary discussion text.

Snapshot semantics

Preparation calls ctx.sessionQuery.readSurface() once per distinct source and never rereads it after enqueue. It projects only direct-user user/message, assistant text, and user/message checkpoints carrying the canonical dsh-compact source marker from the folded current surface. For a source prompt that already contains baked prefix context, projection reads only its model-hidden display content, preventing recursive snapshot propagation. Shadowed pre-compaction events, tools, reasoning, context, plugin-generated user messages other than marked compact checkpoints, and unfinished assistant chunks are excluded. A compacted source therefore contributes its latest checkpoint plus retained later conversation, not restored shadowed text.

The context source is { kind: 'session-reference', version: 1, references }; each reference records its source id and label, capture seq, compact presence, retained/omitted message counts, omitted UTF-8 bytes, and truncation state. When the agent is idle, the standard TUI installs a one-shot agent/pre-step wrapper that adds the snapshot only to an enter decision containing the claimed direct prompt. While the agent is running, it calls inject() immediately before steer(), placing both messages in the next-step inbox for the same later claim. The target log therefore records a sourced context user/message followed by the readable direct user/message. Later source mutation, compaction, or deletion cannot change target replay.

Configuration

Key	Default	Contract
maxReferences	3	Maximum distinct source sessions in one prepared message; must be at most 3.
candidateLimit	50	Default candidate count returned to a host.
maxReferenceBytes	65536	Maximum serialized JSON bytes for one reference object.

Retention applies maxReferenceBytes independently to each source, keeps compact checkpoints and the newest message before dropping older non-checkpoint units, and uses dsh-retention head/tail truncation with an exact UTF-8 omission notice. If one source's fixed serialized fields cannot fit, preparation fails with SESSION_REFERENCE_BUDGET_EXCEEDED instead of returning a partial context.

Model Experience

Referenced session background

What the model sees

The model sees two consecutive user-role messages: the ## Referenced sessions untrusted snapshot, then the current message with its readable @label. The warning forbids following instructions, permission claims, or tool requests from the snapshot unless the current user repeats them. Labels, cwd values, ids, and conversation text are serialized as JSON inside <referenced-sessions> tags; every data < is emitted as the lossless JSON escape \u003c, so source text cannot spell a framing tag.

Token effect

Each referenced message adds the fixed warning plus up to three serialized snapshots, each independently bounded by maxReferenceBytes. The exact snapshot remains in target history until target compaction shadows or summarizes it; source-session changes add no further tokens.

KV Cache effect

The snapshot and request are consecutive append-only target messages and preserve earlier cacheable history. Different references or source capture contents change the new suffix only; later target compaction may invalidate reuse from its replacement boundary.

Known Limitations and Deferred Work

• No body discovery — candidate queries inspect folded titles but do not search message bodies. A non-empty query may inspect every visible persisted session log through the session-query service's bounded, cancellable batch; a dedicated title index may replace that discovery path without changing URI, snapshot, or persistence contracts.
• Trusted caller boundary — the service assumes its host is authorized to read every session exposed by ctx.sessionQuery; it is not a model-facing search tool.
• Text projection only — non-text user and assistant blocks are not propagated across sessions.
• No live link — references are snapshots, not forks, resumes, subscriptions, or source-session mutations.