amiable-dev/stdtel

Make an external process report what it spent, so agent work can be costed end to end — reconcile its records against the provider's bill first, then emit std.artefact.activation spans with kind=external and prove them with stdtel-conform. Use when a tool an agent shells out to, or an MCP server, spends money on models and that spend is invisible in telemetry.

Decorate a skill, sub-agent or MCP server to the standards-telemetry contract so its cost can be attributed — adds the metadata block, picks a standard_id and policy_ids, sets telemetry.scope, and writes an overlay entry for artefacts you do not own. Use when onboarding a skill or agent to telemetry, when stdtel-validate fails, when something reports as unversioned, or when a long-running skill's cost needs rolling up.

Query standards-telemetry data to answer questions about skill cost, token usage, cache-hit rate, and delivery outcomes — tokens per merged PR, tokens per cycle-time hour, first-time policy pass rate. Use when asked what a skill cost, which skills are expensive, how a harness compares, or to pull numbers out of Tempo, Prometheus or the warehouse.

Set up standards telemetry with the user, for Claude Code and GitHub Copilot — installs the hooks, switches on each harness's own telemetry without content, and offers Claude Code's detailed view (OTEL_LOG_TOOL_DETAILS) only with the user's informed consent and after the collector has proven it drops content. Use when installing or re-installing stdtel, when the user asks to enable Copilot telemetry, or when third-party skills show up as "third-party" in cost data.

Emit JSON structured logs with the platform's required fields and no PII, shipped via the standard Alloy/Fluent Bit pipeline.