ContextQMD
Back to Cherry Studio

Observability

docs/references/ai/observability.md

2.0.05.8 KB
Original Source

Observability

The src/main/ai/observability/ subsystem: OTel tracing, the local span projection, and the sink registry. "Trace / telemetry" is the user-facing surface; this doc covers the whole subsystem.

What's instrumented

Every AI SDK call run through Cherry produces an OpenTelemetry span tree:

chat.turn                                      (root, created by context provider)
├── ai.streamText                              (AI SDK auto)
│   ├── ai.streamText.doStream                 (AI SDK auto)
│   ├── ai.toolCall (per tool invocation)      (AI SDK auto)
│   └── ai.streamText.<step>                   (AI SDK auto)
└── attributes: topicId, modelName, …          (set by AiTurnTrace / AdapterTracer)

AI SDK's experimental_telemetry produces the inner spans; Cherry owns the root span through AiTurnTrace so it lands in the same observability path without going through the AI SDK adapter.

The main-process observability boundary is src/main/ai/observability:

  • core/ creates Cherry-owned turn roots and common cs.* attributes.
  • adapters/aiSdk/ interprets AI SDK child spans.
  • adapters/claudeCode/ interprets Claude Code OTLP spans and logs.
  • storage/ keeps the in-memory span projection and JSONL-compatible history.
  • sinks/ defines the extension point for local and future external export.

Local history flush

The container traceId is persisted on the topic / agent-session row (one trace tree per container), but the span tree is first collected in the main-process TraceStorageService memory store. The durable history file is written by the stream terminal path:

  • PersistentChatContextProvider attaches a TraceFlushListener to normal chat turns.
  • AgentSessionRuntimeService attaches the same listener to agent-session:${sessionId} turns, including queued follow-up turns.
  • On the topic-level terminal event (done, paused, or error), TraceFlushListener calls TraceStorageService.saveSpans(topicId).
  • Flush errors are logged as warnings and do not affect message completion.

Collection and persistence are main-process only. Spans live in TraceStorageService's in-memory store and are flushed to the JSONL history file on the terminal event. The renderer trace viewer (TracePage) reads the persisted spans on demand through the trace.getData IPC — it never collects spans itself.

AdapterTracer

src/main/ai/observability/adapters/aiSdk/adapterTracer.ts wraps the OTel Tracer returned by the global provider. On every startSpan / startActiveSpan it:

  1. Patches span.end() to also call AiSdkSpanAdapter.convertToSpanEntity(...) and hand the result to the observability sink registry.
  2. Stamps trace.topicId and trace.modelName so the main-side TraceStorageService can key spans per topic.

AdapterTracer is intentionally only for AI SDK child spans:

  • buildTelemetry (runtime/aiSdk/params/buildTelemetry.ts) — passed to AI SDK as experimental_telemetry.tracer. Captures every AI SDK auto-span. Returns undefined (no telemetry, no tracer) when there is no topicId or developer mode is off — see below.

AiSdkSpanAdapter

src/main/ai/observability/adapters/aiSdk/aiSdkSpanAdapter.ts converts an OTel span into the SpanEntity shape TraceStorageService stores and persists:

  • Reads span name, attributes, events, status, links.
  • Recovers AI SDK's hierarchical attribute conventions: ai.xxx is a level, ai.xxx.yyy is a sub-level under it.
  • Normalises usage attributes across the base and LLM spans, preferring AI SDK v6 keys with legacy and semantic-convention fallbacks: input from ai.usage.inputTokens / ai.usage.promptTokens / gen_ai.usage.input_tokens (and the single ai.usage.tokens embeddings shape), output from ai.usage.outputTokens / ai.usage.completionTokens / gen_ai.usage.output_tokens, plus ai.usage.totalTokens, ai.usage.cachedInputTokens, and ai.usage.reasoningTokens (emitted as completion_tokens_details.reasoning_tokens).

Claude Code Agent SDK spans do not go through AiSdkSpanAdapter; they are converted by src/main/ai/observability/adapters/claudeCode/ClaudeCodeOtlpAdapter.ts.

Sensitive data capture & redaction

Cross-referenced from ClaudeCodeTraceBridgeService.prepareTrace.

The Claude Code OTLP bridge runs only when developer mode is enabled. When it does, it intentionally turns on verbose Claude Code telemetry:

  • OTEL_LOG_USER_PROMPTS — user prompt text
  • OTEL_LOG_TOOL_DETAILS / OTEL_LOG_TOOL_CONTENT — tool calls and their content
  • OTEL_LOG_RAW_API_BODIES — raw API request/response bodies

These payloads land in span attributes that TraceStorageService persists as plaintext JSONL trace files on disk, so a trace can contain secrets (authorization headers, API keys embedded in raw bodies) alongside the prompt and tool content.

Redaction is deliberately not done. Stripping secrets would mean parsing arbitrary OTLP attribute structures across the ingest path and would risk dropping legitimate trace data. The accepted tradeoff is that capture is local-only and developer-gated; turning that into a redaction/threat-model guarantee is a deferred decision. Treat exported trace files as sensitive.

Developer-mode gating

Dev mode only. The span projection (TraceStorageService) is built and persisted in the main process; the renderer trace viewer (TracePage) reads it on demand via the trace.getData IPC. Outside developer mode buildTelemetry returns undefined, so no tracer is attached at all and the AI SDK emits no spans — there is nothing to project, and the viewer shows an empty trace.

Where to read more