Downgrading from Python 3.13 to 3.11.

Breaking Changes

• N/A

New Features

• Tool and prompt records are now physically keyed (on publish) to use a zlib-compressed version of its (logical) primary key.

Bug Fixes

• N/A

Improvements

• Users can now use our package in a Google Colab environment.

Adding the option to add post-commit hooks on init.

Breaking Changes

• N/A

New Features

• Adding support a add-hook-for option with agentc init, which will install a post-commit hook.

Bug Fixes

• Fixed bug with with_langgraph example where prompt was not being used.

Improvements

• Removing generation of Pydantic models (one less dependency).

Documentation updates + minor improvements.

Breaking Changes

• N/A

New Features

• Support for TTL added for DB logger.
• Rollover size for local logger has been exposed as a tunable parameter.

Bug Fixes

• Fixed issue with langgraph.graph.graph.CompiledGraph not being a valid type hint.
• Fixed issue with tuples not being a valid format for the LangChain callback.

Improvements

• Chat model cache collection creation for agentc_langchain now performs retries.
• Versions for all dependencies have been pushed up.

Other Changes

• Adding tutorial that walks through with_langgraph example.
• with_notebook no longer uses Command, but uses conditional edges.

Enhancing Agent Catalog's use as an agent tracing tool + updating various QOL.

Breaking Changes

• Provider and Auditor are no longer available. Users must use the newly provided Catalog and Span.
• Older catalog versions are not compatible (users must start fresh from a new catalog instance).
• "Jinja Prompt" and "Raw Prompt" no longer exist, both have been consolidated into "Prompt".
• agentc.langchain is no longer a valid import, users must import directly from agentc_langchain (note the underscore).
• agentc_langchain.chat.audit is now deprecated, use agentc_langchain.chat.Callback instead.
• CLI arguments and options have been revamped to share naming conventions across commands.

New Features

• New init command has been added for explicitly creating indexes, folders, and downloading models.
• Support has been added for structured content in prompts.
• All logs are now bound to some application nested identifier (a "Span").
  • Agent state can now be recorded (incoming and outgoing) using a context manager.
  • To record custom logs, the __setitem__ method can be used directly on a Span.
  • Span instances can be instantiated as "iterable" to directly access the log objects generated within that Span.
• LangChain Chat Models can now be cached using the same Couchbase instance Agent Catalog resides in.
• Support for LlamaIndex logging has been added (through their Callback machinery).
• Support for LangGraph logging has been added (mainly to record the results of tool calls).
• Support for OpenAI-Client-Compatible embedding models has been added (both for semantic record search and in the semantic search tool).
• Using openapi-pydantic instead of openapi3-parser (which has better support for OpenAPI specs with operation servers).

Bug Fixes

• Fixed issue with key length being too long with heavily nested paths (on publish).

Improvements

• Support for environment based secrets.
• Support for YAML-based JSON Schema in record files (affects SQL++ and semantic search tools).
• Prompts as YAML (for better tooling support).
• Environment-based configuration for CLI and SDK.
• CLI now supports colorized output (via click_extra).
• Documentation overhaul (new theme + more extensive API + CLI detail).

Other Changes

• Added a multi-agent example app with LangGraph that walks through some 'prompt engineering' + evals.

Version of Agent Catalog before 0.2.0 (to be placed on PyPi).

Breaking Changes

• Analytics UDFs have been updated to views (for logs).
• get_item should now be used in place of get_prompt_for and get_tools_for.

Bug Fixes

• Fix in generated code being malformed.
• Fix in agentc exec failing when snapshot is specified.
• Fix in SQL++ tool returning only one result.

New Features

• agentc clean is more granular (catalog vs. activity).
• Adding new agentc ls command to list all catalog artifacts.
• agentc publish can now publish logs.
• Support for a new .agentcignore file.
• Adding a new Makefile for building from source.

Improvements

• Updates to README + instructions + examples.
• More clarification on the use of agentc add.

Initial release for the pre-beta (private preview) Agent Catalog users.

• Support for Couchbase-backed management of the following:
  • SQL++ query-based tools
  • Semantic search tools
  • HTTP request tools
  • Python tools
  • Prompts
  • Jinja-templated prompts
• Introduction of the "agentc" CLI tool, with support for the commands below:
  • add, which creates a new tool or prompt.
  • clean, used to remove local and remote (on Couchbase) Agent Catalog artifacts.
  • env, used to display a user's working environment (w.r.t. Agent Catalog).
  • execute, used to execute a tool stored in the tool catalog.
  • find, used to find tools and prompts via semantic search, by name directly, and with support for annotations + version predicates.
  • index, used to generate local JSON files containing all prompts and tools.
  • publish, used to publish snapshots of the catalog to a Couchbase instance.
  • status, used to display the status of the local and remote catalogs.
  • version, used to display the current version of Agent Catalog.
• Introduction of the "agentc" package, with support for the following classes:
  • Provider, which serves versioned tools and prompts for use in an agent framework.
  • Auditor, which records all agent related activity to local log files and to a Couchbase instance.
• Three prebuilt tools for interfacing with data stored in Couchbase.
• Template agent project that combines Agent Catalog, ControlFlow (an agent framework), and Couchbase + a notebook for assessing quality with Ragas (an evaluation library).

Full Changelog: https://github.com/couchbaselabs/agent-catalog/commits/v0.1.0