Skip to content

CHANGELOG.md

Latest commit

 

History

History
98 lines (68 loc) · 5.72 KB

CHANGELOG.md

File metadata and controls

98 lines (68 loc) · 5.72 KB

Changelog

All notable changes to @ultrathink-solutions/openclaw-logfire will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

0.3.0 - 2026-02-16

Added

  • llm_input and llm_output hook handlers for per-LLM-call observability
  • Individual gen_ai.chat child spans for each LLM invocation, correlated by runId
  • Token usage metrics now populated with real data from llm_output hook (recordTokenUsage() was previously stubbed but never called)
  • Model name tracking (gen_ai.request.model, gen_ai.response.model) on LLM and agent spans
  • Cache token tracking (openclaw.usage.cache_read_tokens, openclaw.usage.cache_write_tokens) on LLM spans
  • Cumulative token counts set on root invoke_agent span at session end
  • Inference detail events (gen_ai.client.inference.operation.details) now emitted when captureInferenceEvents is enabled

Changed

  • OpenTelemetry SDK upgraded from 1.x to 2.x — stable SDK 1.30→2.5, experimental 0.57→0.212. Only breaking change: Resource class constructor replaced with resourceFromAttributes(). Dependency ranges updated; consumers using OTEL 2.x will no longer see peer dep warnings.
  • Token usage metrics include actual model names instead of empty strings
  • Operation duration metrics include model name and provider from LLM hooks
  • Provider name on agent span is updated from llm_input when config providerName is unset
  • Zero-value token counts (e.g. 0 output tokens) are now correctly accumulated — fixed truthy checks that silently dropped zeros
  • LLM span closure wrapped in try/finally to prevent span leaks on exceptions
  • Prompt content capture now applies prepareForCapture() redaction and truncation (consistent with tool args)
  • isRecord runtime guard added to llm_input/llm_output hook registrations (consistent with all other hooks)
  • Orphaned span cleanup uses LIFO order (children before parent) for valid trace trees

Testing

  • Hook handler unit tests added — 113 tests across 11 files, coverage boosted from 28% to 67%

Known Limitations

  • after_tool_call hook not yet implemented — OpenClaw SDK passes sessionKey: undefined at all call sites, making span store lookup impossible. Per-tool error status is not available on tool spans. Tool errors are still captured at the agent level via the agent_end hook. Pending upstream fix.

0.2.0 - 2026-02-16

Fixed

  • Breaking: Align all hook handlers with the OpenClaw plugin SDK (event, ctx) two-argument contract. Previous version only captured the first argument and expected context data nested inside the event, which caused TypeError: Cannot read properties of undefined on every hook invocation.
  • Correct event property names: event.toolName (was event.tool?.name), event.params (was event.tool?.args), event.message (was event.result)

Removed

  • Token usage recording in agent_end — token data is not available in the agent_end event payload (restored in 0.3.0 via llm_output hook)
  • W3C trace context extraction from webhook headers in message_received — raw HTTP headers are not exposed in the message_received event metadata
  • Per-tool error recording in tool_result_persist — error details are not available in that hook's event payload

0.1.2 - 2026-02-16

Fixed

  • Read plugin config from api.pluginConfig instead of api.config — OpenClaw passes the full application config as api.config and the plugin-specific config as api.pluginConfig. Reading from api.config caused all settings (environment, providerName, etc.) to silently fall through to defaults.

0.1.1 - 2026-02-16

Fixed

  • Removed JSON Schema $schema declaration (draft 2020-12) from openclaw.plugin.json — OpenClaw's bundled Ajv only supports draft-07 and crashes on any CLI command when an incompatible $schema is present
  • Changed plugin manifest id from "logfire" to "openclaw-logfire" — OpenClaw resolves plugin config by manifest id, and the mismatch caused config entries to silently not pass through to the plugin

Changed

  • Breaking: Config entry key must now be "openclaw-logfire" (was "logfire") to match the corrected manifest id

Added

  • SKILL.md for ClawHub registry listing

0.1.0 - 2026-02-15

Added

  • Initial release
  • Full agent lifecycle tracing with OTEL GenAI semantic conventions
  • invoke_agent root spans with agent name, session key, workspace, and channel attribution
  • execute_tool child spans with tool arguments, duration, and result size
  • Error tracing with OTEL exception events (type, message, stack trace)
  • Token usage metrics (gen_ai.client.token.usage histogram)
  • Operation duration metrics (gen_ai.client.operation.duration histogram)
  • Pydantic Logfire OTLP HTTP export (US and EU regions)
  • Clickable Logfire trace URLs in agent logs
  • Secret redaction for API keys, platform tokens, JWTs, and credentials
  • W3C traceparent injection into HTTP commands for distributed tracing
  • W3C trace context extraction from inbound webhook headers
  • Configurable trace depth (tool input/output capture, message content, inference events)
  • Batch and simple span processor modes
  • Zero-config quickstart (just set LOGFIRE_TOKEN)