Observability integrations (umbrella): W&B, MLflow, Langfuse, OpenTelemetry, Phoenix

[bordeauxred]

Why

ClawLoop already emits structured episodes, reward signals, and layer-state transitions. Teams running it in production or research invariably have an observability stack they want those signals landing in. Shipping first-class sinks for the common ones makes ClawLoop feel native in existing workflows instead of yet another dashboard to check.

Each item below is a small, self-contained adapter with a clear contract — ideal entry points for first-time contributors.

Integration stubs

• Weights & Biases sink — log per-iteration reward curves, playbook growth, layer state hashes. Pattern: clawloop.integrations.wandb.WandbSink(run_id=...) consuming the existing episode stream.
• MLflow tracking — iterations as runs, playbook entries as artifacts, reward signals as metrics. Same shape as W&B, different backend.
• Langfuse trace export — emit episode messages + tool calls as Langfuse traces so LLM-observability users can search/replay inside their existing UI.
• OpenTelemetry spans — one span per episode, nested spans per step/tool-call. Lets users pipe into any OTel-compatible backend (Honeycomb, Datadog, Tempo, etc.) without a bespoke adapter.
• Arize Phoenix export — trace + evaluation shape for teams already using Phoenix for LLM eval.

Contract

Each sink should:

1. Consume the existing Episode / EpisodeSummary / iteration-level events — no core changes.
2. Be an optional extra: uv sync --extra wandb etc., so core stays dependency-light.
3. Ship with a minimal example under examples/observability/ and a one-paragraph README section.
4. Fail soft — a broken sink never breaks a training run.

Why an umbrella?

Each integration is ~a day of work and independent of the others. Tracking them together shows intent; splitting them off keeps PRs reviewable.