Pine/deepseek-harness
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases 1 Wiki Activity

feat(telemetry): add feedback-gated OTEL modes

Browse Source
This commit is contained in:
Turtle
2026-08-05 12:43:35 +08:00
parent b8d51704f3
commit c836fcd416
41 changed files with 635 additions and 182 deletions
Download Patch File Download Diff File Expand all files Collapse all files
packages/telemetry/session-telemetry-otel/README.md
+12 -4
View File
@@ -2,7 +2,7 @@
English | [中文](README.zh.md)
The OpenTelemetry backend for [the telemetry seam](../session-telemetry/) — the only entry a deployment loads. It composes the OTel JS SDK as-is (`LoggerProvider` → `BatchLogRecordProcessor` → OTLP/HTTP log exporter) and maps each record the seam hands over onto `logger.emit()`, under two instrumentation scopes: ledger records on `@deepseek-ai/dsh-session-telemetry-otel`, operational records on `@deepseek-ai/dsh-session-telemetry-otel/ops`. Resource identity (`service.name`/`service.version`) comes from `dsh-llm`'s `APP_IDENTITY`, the same source the attribution headers use.
The OpenTelemetry backend for [the telemetry seam](../session-telemetry/) — the only entry a deployment loads. Its `mode` decides whether the seam hands records over immediately, releases them only at recorded feedback, or keeps telemetry local. Uploading modes compose the OTel JS SDK as-is (`LoggerProvider` → `BatchLogRecordProcessor` → OTLP/HTTP log exporter) and map each handed-over record onto `logger.emit()`, under two instrumentation scopes: ledger records on `@deepseek-ai/dsh-session-telemetry-otel`, operational records on `@deepseek-ai/dsh-session-telemetry-otel/ops`. Resource identity (`service.name`/`service.version`) comes from `dsh-llm`'s `APP_IDENTITY`, the same source the attribution headers use.
## Config
@@ -10,6 +10,7 @@ The OpenTelemetry backend for [the telemetry seam](../session-telemetry/) — th
- id: telemetry-otel
name: '@deepseek-ai/dsh-session-telemetry-otel'
config:
mode: FULL # FULL (default), FEEDBACK_ONLY, or DISABLED
exporter: # passed verbatim to the SDK's OTLP/HTTP log exporter
url: https://collector.example.com/v1/logs
headers:
@@ -17,15 +18,21 @@ The OpenTelemetry backend for [the telemetry seam](../session-telemetry/) — th
processor: {} # optional; passed verbatim to BatchLogRecordProcessor
```
`exporter.url` is the one field this package validates itself — required, no default, must parse as `http(s)` — so a missing endpoint fails at plugin load (as does a non-positive-integer `processor.maxExportBatchSize`, which the SDK accepts but then hangs on at shutdown). Everything else is the SDK's option shape, owned and documented by the SDK, and both blocks pass through whole: every `OTLPExporterNodeConfigBase` field (`headers`, `timeoutMillis`, `compression`, `keepAlive`, …) reaches the exporter, and batching, export cadence (`scheduledDelayMillis`), retry, queue bounds, and loss policy under sustained failure are the SDK's documented behavior, tuned through the `processor` passthrough. The backend deliberately implements no `flush()`: the batch processor is the only flusher in the process, which is what makes `shutdown()`'s drain complete. Removing this block from `cordis.yml` is the opt-out: no residual state, no `enabled` flag.
| `mode` | Behavior |
|---|---|
| `FULL` | Default. Each projected record, including lifecycle ops records, is handed to the OTel SDK immediately. |
| `FEEDBACK_ONLY` | Each `feedback/record` releases the redacted, projected session prefix through that event. Later records wait for another feedback event and remain local if none arrives. |
| `DISABLED` | No coordinator, provider, processor, or exporter is constructed. No telemetry record leaves the process. A `feedback/record` logs `session telemetry is DISABLED; nothing will be shared and this feedback remains local`; the event remains in the local session log. |
`exporter.url` is required in `FULL` and `FEEDBACK_ONLY`, has no default, and must parse as `http(s)`; it is optional and unused in `DISABLED`. Uploading modes also reject a non-positive-integer `processor.maxExportBatchSize`, which the SDK accepts but then hangs on at shutdown. Everything else is the SDK's option shape, owned and documented by the SDK, and both blocks pass through whole: every `OTLPExporterNodeConfigBase` field (`headers`, `timeoutMillis`, `compression`, `keepAlive`, …) reaches the exporter, and batching, export cadence (`scheduledDelayMillis`), retry, queue bounds, and loss policy under sustained failure are the SDK's documented behavior, tuned through the `processor` passthrough. The backend deliberately implements no `flush()`: the batch processor is the only flusher in the process, which is what makes `shutdown()`'s drain complete.
## What leaves the machine
Records carry the complete `event.data` as the seam's `telemetry/record` waterfall returns it — user and assistant message content, tool arguments and results (command output, file contents), the full system prompt and tool schemas (`request/header`), todo text, compaction summaries, hook `stderrSummary`, and the session `cwd` (a local path). The seam ships no redaction rules: with no `telemetry/record` listener mounted, that is the raw captured copy, so a deployment exporting beyond a trusted boundary mounts its own rules (see [the seam README](../session-telemetry/README.md#the-redact-waterfall)). Provider credentials never appear regardless: adapter API keys are constructor parameters, not session events, so they are structurally absent from the log and therefore from telemetry.
In uploading modes, records carry the complete `event.data` as the seam's `telemetry/record` waterfall returns it — user and assistant message content, tool arguments and results (command output, file contents), the full system prompt and tool schemas (`request/header`), todo text, compaction summaries, hook `stderrSummary`, feedback text, and the session `cwd` (a local path). The seam ships no redaction rules: with no `telemetry/record` listener mounted, that is the raw captured copy, so a deployment exporting beyond a trusted boundary mounts its own rules (see [the seam README](../session-telemetry/README.md#the-redact-waterfall)). Provider credentials never appear regardless: adapter API keys are constructor parameters, not session events, so they are structurally absent from the log and therefore from telemetry. `DISABLED` does not construct the SDK pipeline or hand any capture to a backend.
## Field mapping
Seam record → SDK log record: `time` → `timestamp`/`observedTimestamp`; `severity` → `severityNumber`/`severityText` (INFO 9 / WARN 13 / ERROR 17); `body` → the structured log body; `attributes` verbatim. Receivers dedupe on `(session.id, event.seq)`, alert on severity, and detect crashes by `shutdown`-record absence (a session with activity, no `shutdown` ops record, gone stale ended uncleanly). The marker means telemetry stopped observing the session cleanly — emitted at the session's own disposal, or at application teardown for sessions still running then; a marker followed by more of that session's events is a telemetry reload, not a session restart. Streams are not self-contained across lineage: a resumed session continues its own id's stream from where the previous process left off, and a forked session's stream starts at its inherited boundary — its prefix lives in the parent's stream, stitched via `session.parent_id` + `session.seed_length`. One consequence of continuing rather than replaying: a turn left open mid-stream and never closed marks the previous process dying inside it. The local log is repaired with synthetic closers at resume, but those repairs are never exported — the wire stream stays faithful to what the crashed process actually shipped, and a later clean `shutdown` marker attests only to the resumed process's own exit.
Seam record → SDK log record: `time` → `timestamp`/`observedTimestamp`; `severity` → `severityNumber`/`severityText` (INFO 9 / WARN 13 / ERROR 17); `body` → the structured log body; `attributes` verbatim. Receivers dedupe on `(session.id, event.seq)` and alert on severity. In `FULL`, they may also detect crashes by `shutdown`-record absence: the marker is emitted at the session's own disposal or application teardown, and a marker followed by more events is a telemetry reload. In `FEEDBACK_ONLY`, a released prefix normally has no later `shutdown` marker, so its absence is not a crash signal. Streams are not self-contained across lineage: a resumed session continues its own id's stream from where the previous process left off, and a forked session's stream starts at its inherited boundary — its prefix lives in the parent's stream, stitched via `session.parent_id` + `session.seed_length`. A resumed local log may contain synthetic closers that were never exported; the wire stream stays faithful to records actually handed to the SDK.
## Model Experience
@@ -39,3 +46,4 @@ None; this package neither assembles nor sends a provider request.
- **Upstream experimental tree** — `@opentelemetry/sdk-logs` is still published from the upstream experimental tree; SDK API churn lands here and only here — the seam contract does not move.
- **No live-collector coverage** — every test exports to a local mock collector; the keyless Loader-composition e2e (`tests/loader-composition.e2e.ts`) covers the wire shape on every run, and behavior against a real OTLP deployment (auth, TLS, throttling) is the SDK exporter's documented territory.
- **Feedback-only memory** — each session retains deep-copied, redacted projected records in memory until feedback releases them or the session becomes unreachable. There is no durable pre-feedback spool; a crash before feedback uploads nothing.