Changelog

New Features:

Compression Manager Token Counting: Added support for token count
based compression across providers. Read more
here

Adding Knowledge - Content Hash Enhancement: Content hashing now
includes name and description fields for URLs, paths, and file_data to
ensure uniqueness. The same URL or path with different names or
descriptions will produce different content hashes, allowing multiple
distinct content items from the same source. Note: Existing content
previously added with names/descriptions may have different hashes than
what this version generates, which could
affect skip_if_exists and upsert behavior for that content.

Session storage in HITL flows: Add some parsing logic in
RunRequirement to ensure it is properly converted into JSON before
persisting the related run, to avoid DB errors or missing data.
Gemini Malformed Tool Calls: Fixes to the logic to retry Gemini
requests that end with a malformed tool call. Also introduce the
retry_with_guidance_limit parameter to adjust how many times these
retries should happen.

fix: Include content name and description fields for hash generation by @willemcdejongh in #5697
[fix] Serialize RunRequirement datetime for PostgreSQL JSONB storage by @edlsh in #5694
fix: AgentOS Knowledge reading async. by @willemcdejongh in #5700
fix: gemini retry by @ysolanky in #5695
chore: add workers to speed up release testing by @manuhortet in #5702
feat: Add token counting utility + Add support for it in Compression by @Mustafa-Esoofally in #5593
fix: csv reader by @ysolanky in #5706
fix: OpenAI responses test by @Mustafa-Esoofally in #5707
chore: v2.3.12 by @manuhortet in #5704

Full Changelog: v2.3.11...v2.3.12