8.9 KiB
description, kind
| description | kind |
|---|---|
| Session-telemetry capture seam for deployments and backend authors choosing a reporting backend, mounting redaction rules, or implementing the backend contract. | package-library |
@deepseek-ai/dsh-session-telemetry
English | 中文
Summary
dsh-session-telemetry captures session activity for outbound reporting: it copies each session event into a telemetry record, lets a deployment redact it, and hands it to a reporting backend that implements the contract. Deployments do not load this package directly — they load exactly one backend (the shipped OpenTelemetry backend is dsh-session-telemetry-otel), which registers ctx.sessionTelemetry and composes the capture coordinator. The seam owns capture, redaction, and the sharing disclosure; batching, retry, queueing, and loss policy belong to the backend's SDK and stop at emit(). Every mounted backend discloses its deployment-selected sharing policy so acknowledgement surfaces can report whether and how a session is shared. The contract and capture behavior come first; the implementation internals live in a collapsible developer section below.
Table of Contents
- Use this package
- Understand the implementation
- Further Exploration
- Model Experience
- Known Limitations and Deferred Work
- Dev Note
Use this package
As a deployment, choose a backend, mount it, and add redaction rules when records must not leave the process as captured. As a backend author, implement the three-member contract and compose the coordinator with a capture mode.
Choosing and mounting a backend
Load exactly one backend plugin; it registers ctx.sessionTelemetry with the capture coordinator and its delivery pipeline. A duplicate load throws. The required sharing member reports the deployment mode, not per-session admission or delivery. A consumer may report "not configured" only when no telemetry service is mounted. The /feedback command confirms recording without reading this policy.
The backend contract
A backend implements three members: emit(record) must be a non-blocking enqueue because it runs synchronously on the session-event path; optional flush() is a fire-and-forget hint after a turn ends, which most backends omit in favor of their SDK's own batching schedule; shutdown() drains queued records and resolves when the SDK stops, and disposal awaits it. A backend that implements flush() must order concurrent flushes with the final shutdown() drain.
What gets captured
Capture runs in one of two modes. live capture follows session events as they are appended, replays already-live sessions at mount time, and records lifecycle markers; on-demand capture reads the canonical session log only when the backend requests a prefix through captureSession(session, throughSeq?). Coordinator options select whether stored history is included. Every canonical session event maps to one ledger record in order. An assistant/message or assistant/attempt record carries its complete embedded compact stream, including failed and retried output. Each ledger record also carries session.id, session.format_version, the numeric event identity, optional header facts, and a pre-mapped severity (error for tool/result.isError, turn/end error reasons, and agent-error; info otherwise).
The sharing disclosure
Every backend discloses its deployment mode through sharing: full, feedback-only, or disabled. A backend may additionally restrict eligible Sessions. This property is not a delivery receipt; handoff is a non-blocking enqueue, and batching, retry, and loss policy belong to the backend SDK.
Redacting records
Every outbound record passes the sessionTelemetry/record waterfall after the coordinator copies its canonical event. This package ships no rules: with no listener mounted, records reach the backend exactly as captured, so exported data is as clean as the rules a deployment mounts. Listeners stack by transforming next()'s return value; a throwing listener withholds that one record fail-closed. Redaction applies to the outbound copy only — the canonical session log is never rewritten.
Understand the implementation
Implementation internals — click to expand
This section explains the capture design; the observable behavior is fully covered in Use this package.
Design concept
The seam is built on one boundary: the harness's aspect ends at emit(). Complete event capture, redaction, and the handoff cursor live here; batching, retry, queueing, and loss policy are the reporting SDK's, deliberately not modelled or wrapped. The design and rejected alternatives are pinned in the revival Agent Note.
Source map
| File | Role |
|---|---|
src/index.ts |
Service Definition: SessionTelemetryBackend/SessionTelemetrySink contract, record vocabulary, session-telemetry/record waterfall declaration |
src/coordinator.ts |
Capture: live listeners, lifecycle-local on-demand replay, redaction, handoff cursor, containment |
Capture flow
Live capture registers Session events, flush hints, shutdown markers, and agent/error observers through the composing fiber’s effects. On-demand capture registers only the disposal effect and reads the requested canonical-log prefix under its history policy. Synchronous handlers contain failures so they cannot affect the agent loop or other listeners.
The handoff cursor
A module-scope WeakMap<Session, seq> records the highest sequence handed off, not delivered. Re-adopting the same object resumes after that cursor. Capture normally starts at firstLiveSeq; explicit includeHistory: true starts an unhanded object at seq 0, including restored or fork history. The backend owns capture authorization. Stored history does not itself authorize capture; the OTel backend waits for new explicit feedback. Receivers deduplicate repeated records by (session.id, session.format_version, event.seq).
Further Exploration
Read these pages when the seam contract is not enough. They move from the shipped backend to the subsystem reference and the decision evidence.
- OpenTelemetry telemetry backend — the shipped backend deployments load, with mode and exporter configuration.
- Session telemetry subsystem — the capability split and type declarations.
- Session telemetry revival decision — rationale, trade-offs, and rejected alternatives.
- Session package map — adjacent persistence, projection, title, and telemetry packages.
Model Experience
None, as the seam observes the session stream and hands redacted copies outward; it registers nothing model-facing.
KV Cache effect
None; the package neither assembles nor sends a provider request.
Known Limitations and Deferred Work
These limits define the delivery and data-protection guarantees a deployment gets. They are current package constraints.
- Best-effort delivery — the cursor marks handed-off, not delivered; a session torn down inside a reload window cannot be re-adopted, and whatever sits in a backend queue at crash time is lost. A durable outbox (spool, per-sink cursors, at-least-once) is deferred until a deployment states a crash-loss requirement.
- No built-in redaction rules — with no
sessionTelemetry/recordlistener mounted, records leave the process exactly as captured, including any credentials embedded in file contents or command output; a deployment exporting to a shared collector owns its rule set. - On-demand redaction uses current state — uncaptured events exist only in the canonical session log; a later
captureSession()deep-copies and redacts their current values with the policy mounted at that time, and there is no capture-time telemetry snapshot or durable pre-capture spool.
Dev Note
Working context for maintainers — click to expand
None.
Runtime invariant: No companion is published. The package's whole output is the backend handoff — a synchronous emit() call outside every authoritative event stream — and its capture side never appends session events, so no event/data relation exists for an independent companion to observe.