Feature Request: Implement OpenTelemetry-based Observability

[minimAluminiumalism]

Description

Summary

Add OpenTelemetry based observability to Crush, enabling standardized tracing, metrics, and logging capabilities. This would allow users to monitor usage patterns, track costs, debug issues.

Currently, Crush only sends anonymous metrics to PostHog (data.charm.land). This proposal adds support for user-controlled, standards-based telemetry export.

Similar issue mistralai/mistral-vibe#63

I am willing to implement this feature.

Proposal

Let's impelement the tracing and metrics function firstly.

Metrics

GenAI Standard Metrics:

• gen_ai.client.token.usage - Token consumption by model/type (input/output/cache)
• gen_ai.client.operation.duration - API call latency

Crush-specific Metrics:

• crush.session.count - Session starts
• crush.tool.call.count - Tool invocations (by tool name, success status)
• crush.tool.call.latency - Tool execution duration
• crush.api.request.count - LLM API requests (by provider, model, status)
• crush.api.request.latency - LLM API latency
• crush.agent.duration - Agent run duration
• crush.agent.turns - Conversation turns per agent run

Tracing

Distributed tracing for agent workflows:

• Parent span for each user prompt → agent run
• Child spans for LLM API calls, tool executions
• Automatic error recording and status propagation

References

Gemini CLI Telemetry - Reference implementation
Claude Code Monitoring

[meowgorithm]

Hey! Thanks for submitting this. In order to keep the issue tracker clean, I'm going to transfer it to a discussion (for further discussion). If we decide it's something that should be implemented we'll convert it back to an issue.

[minimAluminiumalism]

@meowgorithm I've already found a PR #1981. Perhaps we could discuss this issue further?

[kwiksher]

#1981 was closed, why unmerged?

[frzifus]

An option like this to track would be awesome. especially when running crush inside a cluster without manual user input.