NanoClaw
NanoClaw proposes opt-in container traces that include prompt and tool context
NanoClaw’s proposed telemetry skill would export turns, model calls, tools and deliveries through OTLP/HTTP, enabled per group. It also sends prompt and output text to the configured collector, making data handling as important as the promised visibility; live exporter validation remains outside the reported test suite.
NanoClaw received an OpenTelemetry skill proposal on September 13 that follows agent work across model calls, tools, subagents, compaction, background tasks and delivery. The addition is packaged as an opt-in skill, not a change that instruments every existing installation. Applying it copies runner files and adds pinned dependencies; a group must also contain an otel.json configuration before telemetry runs.
The proposal targets OTLP/HTTP collectors and carries both OpenInference and OpenTelemetry GenAI conventions. Phoenix is named as the reference target, with Jaeger and Langfuse also discussed. Those compatibility statements need to be read alongside a narrower verification boundary: the reported automated suite uses an in-memory exporter and explicitly excludes a real OTLP exporter test.
The facts
- The proposed skill traces turns, model calls, tools, subagents, compaction, background tasks and deliveries. - Span fields include cost, tokens with cache breakdown, failure class, and the skills and MCP servers used. - Without a group otel.json, telemetry packages are not imported; a missing package is intended to become a logged no-op. - Enabled traces include prompt and output text capped at 12,000 characters per attribute, tool arguments and output sizes; reasoning text requires a separate thinkingText opt-in. - The MCP server imports mailbox-backed telemetry state rather than starting an exporter in that process. - The author reports 148 focused tests and 309 host skill-policy tests, plus manual apply/remove and reapplication checks.
Why it matters
An agent can appear successful in chat while its delivery, background work or tool chain fails elsewhere. Tracing those legs can make the difference visible. But moving prompt text and tool arguments out of the container is a separate operational decision from collecting duration and token counts. A standards-based transport reduces collector lock-in; it does not decide whether the collector should receive the organization’s task data or how long it should retain it.
Current
Inspected on 2026-09-14. The NanoClaw stable-channel baseline is v2.3.0 published 2026-08-24T11:45:32Z. The primary source was open (not merged) when captured. Mainline merges, open proposals and packaged releases are distinct availability states; the release baseline does not establish that a proposal has shipped.
Evidence
Primary evidence: nanocoai/nanoclaw PR #3796 (https://github.com/nanocoai/nanoclaw/pull/3796). Supporting context: OpenTelemetry — OTLP transport specification (https://opentelemetry.io/docs/specs/otlp/); NanoClaw — official product README (https://raw.githubusercontent.com/nanocoai/nanoclaw/main/README.md). Public source bodies and linked documentation were inspected. Contributor test results remain attributed reports, not independently reproduced experiments.
Operator take
Treat collector selection, access control and retention as prerequisites, not cleanup after enabling traces. Inspect a sample span for secrets and private task material before routing production traffic. Verify one real collector round trip, then exercise shutdown, a missing dependency and a collector outage. The proposal caps shutdown draining at four seconds and intends instrumentation errors not to fail a turn; those are useful acceptance conditions to test in the intended deployment.
Caveat
This was an open skill proposal, not a shipped telemetry service or independent benchmark. Its body discusses backend compatibility while stating that a real exporter test is out of scope; the in-memory test results do not resolve that gap. The reported full runner run also notes a pre-existing host-file wiring error. No production collector or private prompt data was accessed for this article.
This was an open skill proposal, not a shipped telemetry service or independent benchmark. Its body discusses backend compatibility while stating that a real exporter test is out of scope; the in-memory test results do not resolve that gap. The reported full runner run also notes a pre-existing host-file wiring error. No production collector or private prompt data was accessed for this article.