Skip to content

Changelog

All notable changes follow Keep a Changelog.

[0.1.1] - 2026-10-07

Fixed

  • Notifications never fired. Call.notify passed event twice to the template renderer; the resulting TypeError was swallowed by the "never break the pipeline" handler, so Slack, Teams and webhook notifications were silently dropped. Added a regression test.
  • workspace.telemetry.headers did not resolve ${VAR} placeholders through the secrets connector, so authenticated OTLP endpoints (Datadog, New Relic, Honeycomb, Langfuse, Dynatrace) could not be configured safely. Secrets are now resolved at use time and never written to disk.
  • Removed an unused import from the docker_export FastAPI template.

Added

  • Optional candidates list in a call's policy to restrict the tournament (for example to keep scheduled retrains off xlmr_finetune). Defaults to the selection rules; an explicit candidates= argument still wins.
  • Example notebooks (quickstart, five agent frameworks, eight connectors, three monitoring stacks), a quickstart GIF, vendor monitoring guides and a notebooks CI job.

[0.1.0] - 2026-10-06

Added

  • Profiler, tournament training (TF-IDF, multilingual-e5 + linear head, optional fine-tuned XLM-R, tabular models, custom candidates), calibration and precision-targeted thresholds.
  • @pecca.replace runtime with off / record / shadow / live modes, OpenTelemetry spans and exception isolation.
  • Governance: gates, approvals, evidence, retention, schedule, controls mapping; eight editable profiles (templates only, not legal advice).
  • Audit packs (Markdown, JSON, Confluence; PDF is untested).
  • Connectors: data sources (CSV, Parquet, Postgres, Databricks, Snowflake, BigQuery, OTel traces, LiteLLM, Databricks AI Gateway, MCP), registries (local, MLflow / Unity Catalog, W&B), notifiers, ticketers, schedulers, serving targets, secrets, docs publishers, identity.
  • CLI (pecca init|apply|connect|profile|train|status|eval|promote|approve|diff|audit|logs|doctor|plugins|scheduler|mcp|datasets), MCP server, agent-framework shims (LangGraph, Strands, Google ADK, OpenAI Agents SDK, Claude Agent SDK).
  • Documentation site, synthetic demo dataset and demo model, example project.

Known limitations

  • TF-IDF models are served from the pickled pipeline (ONNX output did not match scikit-learn); the e5 ONNX export covers the linear head only.
  • Extraction calls are profiled but not replaceable.
  • Remote MCP adapters do not support auth headers.