v0.6.22 - HF Trace Export and Kimi 2.6

HF Trace Export

Export session traces in a Hugging Face Trace Viewer compatible format - and directly upload to a dataset. Available from /session export or fast-agent export --hf-dataset <repo>.

Kimi 2.6 and Kimi 2.6 Instant

Model aliases added for above models via Novita.

Other Changes

• Improvements to ACP session, error and system prompt handling.

Full Changelog: v0.6.21...v0.6.22

v0.6.17 - opus 4.7 and more

What's Changed

Opus 4.7 Support

Includes "summarized display" feature to show reasoning summaries.

• Adaptive reasoning is on by default. Set Reasoning Level and Task Budget with /model commands or
• via model string - e.g. opus?reasoning=xhigh&task_budget=64000.

Access earlier versions with presets opus46 and opus45 or the full model slug.

Note that task_budget is only available for Opus 4.7

Assistant Response Banner

Long responses now show a banner at the final assistant message, to make location in the scrollback buffer easier. Use fast-agent config display to disable (Stream Reprint Banner).

Other Changes

• Added Kimi25 instant mode alias
• Enhanced session snapshotting

• Bump pytest from 9.0.2 to 9.0.3 by @dependabot[bot] in #768
• Dev/0.6.17 by @evalstate in #769

colour themes, llama router, tweaks

• Experimental support for auto-configuration with llama.cpp as a router (shortcut added to model picker)
• Updated default colour scheme (all still ANSI)
• Fixes to hook display code, session new now resets counters.
• MCP Servers can now be supplied by the ACP Client

v0.6.13 Overdue Release (Note) - 🚨 IMPORTANT DEPENDENCY UPDATE

fast-agent 0.6.13 -- 🚨FastMCP Update to 0.3.2🚨

Please update to this version ASAP.

OpenAI Responses API fix and FastMCP v0.3.2

This release contains a fix for a server-side change on the OpenAI WebSockets API and important security fixes for FastMCP.

Major Features

• Tool Annotations (thanks @danieldagot). https://fast-agent.ai/agents/function_tools/#when-to-use-function-tools
• Remote MCP (Anthropic and OpenAI Responses). https://fast-agent.ai/mcp/#provider-managed-remote-mcp
• ACP Session Handling improvements.
• ACP compatibility fixes (0.8 > 0.9 upgrade)
• Anthropic on Vertex support
• Improvements to tool output rendering, structured outputs handling (thanks @iqdoctor )
• Colouration of apply_patch during streaming.
• GLM 5.1 support; gpt-oss alias back to Cerebras by default
• --prompt-file switch is now NOT interactive by default
• Document/Vision Attachments through the TUI file, MCP Resource and URL (use ^)
• Ability to denote tools as code style for improved rendering.

What's Changed

• raw usage storage, --prompt-file not interactive by @evalstate in #739
• fix markdown rendering at bottom-screen by @evalstate in #740
• Surface MCP server notifications to agents by @es617 in #741
• improve README by @evalstate in #743
• 0.6.10 / Vertex / ACP Fix by @evalstate in #745
• Feat/tool decorator by @danieldagot in #744
• ACP Session, Code (Monty Codegen) Highlighting/Tool Enhancements by @evalstate in #749
• fix: avoid gluing reasoning sentences across deltas by @iqdoctor in #704
• ui: prefer structuredContent preview for multi-block tool results by @iqdoctor in #703
• Dev/0.6.13 by @evalstate in #751
• Bump fastmcp from 3.1.1 to 3.2.0 by @dependabot[bot] in #752
• Bump aiohttp from 3.13.3 to 3.13.4 by @dependabot[bot] in #753
• Bump anthropic from 0.86.0 to 0.87.0 by @dependabot[bot] in #755
• Bump cryptography from 46.0.5 to 46.0.7 by @dependabot[bot] in #754
• Bump pygments from 2.19.2 to 2.20.0 by @dependabot[bot] in #756

New Contributors

• @es617 made their first contribution in #741
• @danieldagot made their first contribution in #744

Full Changelog: v0.6.7...v0.6.13

v0.6.7

What's new

• Reduce Markdown streaming flickering
• Fix Opus/Sonnet 4.6 structured content API warning
• Display README files from card pack installs/updates (--pack) option
• Fix by @phucly95 to fix per-agent skill filtering
• Fix by @peachgabba22 to fix MCP connectivity for non-persistent Servers.
• Suppress uvloop / prompt_toolkit warnings on Python 3.14

New Contributors

• @phucly95 made their first contribution in #729
• @peachgabba22 made their first contribution in #737

Full Changelog: v0.6.1...v0.6.7

v0.6.1

What's Changed

• GPT-5.4-mini/nano support
• Remove SSE server support / migrate to FastMCP3 by @evalstate in #724

Full Changelog: v0.6.0...v0.6.1

v0.6.0 ❤️llama.cpp

fast-agent 0.6.0

llama.cpp support

• Added llama.cpp model discovery, import, and launch-oriented picker flows to make local model setup much easier. Use with fast-agent model llamacpp.
• Reads model settings (available context window size) and modalities, and creates a configuration file (model overlay). If you want to include sampling parameters from the server use --include-sampling-defaults.

Model Overlays

• Added local model overlays so teams can define reusable local model aliases and runtime settings outside the built-in catalog.
• Expanded model selection and switching UX across CLI, prompt UI, and ACP, with better cataloging, diagnostics, and reference handling.
• https://fast-agent.ai/models/model_overlays/

Anthropic 1M Context, Google Native Refresh

• Support for Anthropic 1M context window for 4.6 class models. Older models still require the extension flag e.g. claude-sonnet-4-5?context=1m
• Improved Google Gemini support and general Gemini provider behavior.
• Added support for Anthropic 1M-context workflows and broader Anthropic capability improvements.
• Introduced a dedicated skills CLI and reworked the skills subsystem into clearer service and operations layers.
• Improved ACP/MCP command handling, auth passthrough, session behavior, and command-surface parity.
• Upgraded prompt and interactive shell experiences, including better completions, tool display, history views, and input/runtime structure.
• Added more detailed history summaries, response telemetry, tool tracking, and streaming behavior coverage.
• Refactored core runtime, agent loading, validation, and model resolution code to support the newer workflows more cleanly.

⚠️WARNING

• Model Aliases are now known as Presets (e.g. kimi25).
• Configuration Aliases are now known as References (e.g. $system.default, $system.fast). Configuration files will need to be updated if they defined aliases.
• Large internal refactorings -- no major API changes but be careful upgrading regardless.

What's Changed

• Support local Codex auth.json override by @iqdoctor in #709
• Feat/acp registry by @evalstate in #711
• Dev/0.5.9 by @evalstate in #712
• Harden interrupted session tool-loop recovery by @evalstate in #713
• Upgrade and clean to ty 0.0.21. by @evalstate in #714
• Add skills CLI and extract standalone-ready skills core by @evalstate in #715
• Add history timing summaries and streaming telemetry by @evalstate in #716
• mcp: report lower-layer startup errors accurately by @iqdoctor in #717
• Fix/ty0.23 by @evalstate in #718
• Bump pyjwt from 2.10.1 to 2.12.0 by @dependabot[bot] in #720
• Dev/0.5.12 by @evalstate in #719
• Feat/model overlay by @evalstate in #722
• Feat/model overlay followup by @evalstate in #723

Full Changelog: v0.5.7...v0.6.0

v0.5.7

What's Changed

OpenAI Responses Improvements and gpt-5.4 support

• Support for gpt-5.4 (codexplan alias updated) and gpt-5.3-chat-instant (chatgpt alias)
• Support for assistant "phase"
• SDK version update
• Service Tier selection fast and flex for supported models (use Shift+TAB in UI for convenience)
• WebSocket is now the default transport - (use ?transport=sse for SSE connections) NB: OpenResponses remains SSE

Shell Improvements

• read_text_file and write_text_file are supplied with shell access (Qwen 3.5/ACP compatibility). Configurable with write_text_file_mode (auto|on|off|apply_patch)
• write_text_file is replaced with apply_patch for Responses models
• apply_patch output is coloured post streaming
• read_text_file previews enhanced and syntax highlight aware.
• Skill prompt generation now uses explicit read-tool naming (read_text_file vs read_skill) based on
  runtime capability.

Tool Result passthrough

• Option in request_params and model card to bypass LLM processing of tool outputs (intended for use with code-mode style subagents and MCP Servers).

Other UI/UX Improvements

• service_tier, reasoning, verbosity, web_search and web_fetch are toggled with Shift+TAB, F6-F9 keys)
• Added fast-agent model setup and fast-agent model doctor to manage model configuration and aliases
• Added Codex login link in the model picker
• fast-agent model setup now guides users through missing, invalid, and configured aliases
• fresh model setup now defaults to $system.default
• Use ##<agent> to send a message to an agent with silent return to the input buffer.
• Interactive UI toolbar/model chip was upgraded with service-tier, reasoning, verbosity, web-search, and
  web-fetch indicators.

New prompt shortcuts and controls:

• Shift+Tab service-tier cycle
• F6 reasoning cycle
• F7 verbosity cycle
• F8 web search toggle
• F9 web fetch toggle

Internal / Refactor / Tests

• Shared formatting and marketplace-argument parsing helpers for command handlers.
• Unified A3 header rendering utility used across CLI/command surfaces.
• Docs layout switched to a docs git submodule pointer.
• Broad test expansion for passthrough flow, filesystem runtime, apply_patch display/preview, quiet hash-
  send, display suppression, model controls, and ACP response-mode handling.
• Improved cancellation/interrupt handling across interactive, CLI runner, ACP prompt handling, and tool-
  loop history reconciliation.
• Better post-cancel recovery: stale progress row cleanup, clearer cancelled-turn messaging, and safer
  task cancellation state handling.
• Improved read_text_file and shell result display behavior (cleaner summaries, truncation behavior,
  display suppression integration).
• Refined MCP connection/reconnect handling and stream transport tracking hooks.
• Result export reliability improvements for one-shot message runs (including transient result history
  capture when needed).

v0.5.6

What's Changed

• Huge improvements to Markdown Streaming Performance
• Major WebSockets improvements for Responses API.
• Minimax M2.5 and Qwen 3.5 Support
• Model Picker
• MCP experimental sessions demonstrations
• Other display enhancements

• Dev/0.5.1 by @evalstate in #693
• Bump requests from 2.32.3 to 2.32.4 by @dependabot[bot] in #690
• Bump urllib3 from 2.3.0 to 2.6.3 in /docs by @dependabot[bot] in #688
• Dev/0.5.2 by @evalstate in #694
• fix websocket continuation, more geneourous shell token limits by @evalstate in #695
• Bump requests from 2.32.3 to 2.32.4 by @dependabot[bot] in #699
• migrate to new schema by @evalstate in #698
• markdown optimistic case improvements by @evalstate in #700
• Dev/0.5.4 by @evalstate in #697
• fix double print by @evalstate in #701
• Feat/model selector by @evalstate in #702

Full Changelog: v0.5.1...v0.5.6

fast-agent 0.5.1 codex and config

Codex and Config

• GPT-5.3-codex support via API Key
• All OpenAI Responses models can now be configured to use WebSocket (add ?transport=ws to the model string e.g. fast-agent -x --model responses.GPT-5.2?transport=ws)
• Improvements to WebSocket handling/planning
• fastagent.config.yaml and AgentCards can both use target style configuration for MCP
• MCP Experimental Sessions support and demos

Full Changelog: v0.5.0...v0.5.1