v1.20.5

This release improves stability for non-interactive sessions, updates the default Anthropic model to Claude Sonnet 4.5, and adds support for private GitHub repositories and standard agent directories.

What's New

• Adds support for using agent YAML files from private GitHub repositories
• Adds support for standard .agents/skills directory structure
• Adds deepwiki integration to the librarian
• Adds timestamp tracking to runtime events
• Allows users to define their own default model in global configuration

Improvements

• Updates default Anthropic model to Claude Sonnet 4.5
• Adds reason explanations when relevance checks fail during evaluations
• Persists ACP sessions to default SQLite database unless specified with --session-db flag
• Makes aliased agent paths absolute for better path resolution
• Produces session database for evaluations to enable investigation of results

Bug Fixes

• Prevents panic when elicitation is requested in non-interactive sessions
• Fixes title generation hanging with Gemini 3 models by properly disabling thinking
• Fixes current agent display in TUI interface
• Prevents TUI dimensions from going negative when sidebar is collapsed
• Fixes flaky test issues

Technical Changes

• Simplifies ElicitationRequestEvent check to reduce code duplication
• Allows passing additional environment variables to Docker when running evaluations
• Passes LLM as judge on full transcript for better evaluation accuracy

---

What's Changed

• docs: update CHANGELOG.md for v1.20.4 by @docker-read-write[bot] in #1577
• Fix cagent eval metric by @dgageot in #1579
• Update default anthropic model to sonnet 4.5 by @krissetto in #1580
• Support aliases to urls by @dgageot in #1581
• Bump dependencies by @dgageot in #1582
• Produce a session db for the evals by @dgageot in #1583
• Allow passing more env to docker run when running evals by @dgageot in #1585
• Fix flaky test by @dgageot in #1587
• Include reason in relevance check failures by @dgageot in #1584
• 🫡 by @krissetto in #1588
• Don't go negative in tui dimensions in collapsed sidebar by @krissetto in #1598
• Put back the sessions json. It was useful after all by @dgageot in #1597
• Add a timestamp to runtime events by @dgageot in #1590
• Fix current agent in TUI by @dgageot in #1596
• Support using agent yamls from private github repos by @dgageot in #1594
• Fix cagent acp session persitence by @krissetto in #1601
• cli: prevent panic when elicitation is requested in non-interactive s… by @Pnkcaht in #1567
• make aliased agent paths absolute by @krissetto in #1607
• Add deepwiki to the librarian by @dgageot in #1608
• Pass LLM as a judge on the full transcript by @dgageot in #1605
• fix: prevent title generation from hanging with Gemini 3 models by @dgageot in #1602
• Support for the .agents/skills standard by @krissetto in #1609

Full Changelog: v1.20.4...v1.20.5