Repository navigation
Changes
fast-agent 0.10.49
• Z.ai defaults to GLM‑5.3. A bare zai model and the zaiglm alias now select zai.glm-5.3 (previously glm-5.2).
Use zaiglm52 or zai.glm-5.2 to stay on GLM‑5.2. A zai.default_model setting still takes precedence.
• The glm alias moves to GLM‑5.3 on Hugging Face (hf.zai-org/GLM-5.3:zai-org). glm52 keeps its GLM‑5.2 route.
• GLM‑5.3 reasoning is always on. It supports low, high and max only, and defaults to max. On GLM‑5.2,
reasoning=off or none turned reasoning off. On GLM‑5.3 those settings log a warning and max is used.
• medium reasoning now maps to high on models that don't support medium (GLM‑5.3, Kimi K3, DeepSeek). It
previously fell back to the model default, usually max. This matches the existing minimal→low and xhigh→max
handling.
• Provider safety refusals are now saved as responses instead of raising. When a provider declines a request
under its safety policy, the turn ends with a SAFETY stop reason and an explanatory message instead of an
exception. The provider codes covered are content_filter, content_policy_violation and cyber_policy. This
applies to Responses and OpenAI-compatible chat providers. Responses requests in auto mode no longer retry a
refusal over SSE.
New
• GLM‑5.3 and GLM‑5.3‑Flash on Hugging Face through Together, Novita and Z.ai. New aliases are glm53 and
glm53flash, and the HF picker has an entry for each route. Reasoning is streamed and replayed exactly across
turns and tool calls. Flash accepts images.
• Trackio is back as an optional extra: uv pip install 'fast-agent-mcp[trackio]'. The gepa extra includes it
again. Automatic CPU and GPU logging is off by default.
• PTC call records in ATIF exports. execute_python observations now record each tool call made from inside the
Python program. Each record has the tool name, a snapshot of its arguments, and its outcome (completed,
raised, in_flight or cancelled), under ptc_calls (fast-agent.ptc-calls/v1). This does not cover direct
Python file access. Arguments can contain secrets, so review exports before sharing. See the export
reference.
Fixes
• Provider failures are recorded correctly. When Responses retries run out, websocket errors and timeouts are
now saved as error responses with the provider error code attached. Previously they escaped as exceptions.
• Request metadata is no longer modified. Building a provider request no longer changes the caller's
RequestParams.metadata. Native Z.ai was adding tool_stream and thinking to extra_body, which carried over
into later requests.
Dependencies
• openai 3.40.0 and google-genai 2.33.0
• Minimum versions: azure-identity 1.27.0, boto3 1.43.119, trackio 0.44.0
- Dev/0.10.49 @evalstate (#1047)
- fix: register file-backed tool and hook modules before execution @LJK-ikun (#1035)
- fix: preserve full-object template placeholders on name collisions @BlueLi66 (#1038)
- build(deps): bump dnspython from 2.8.0 to 2.9.0 @dependabot[bot] (#1046)
- fix(acp): preserve newline command separators @skulitom (#1042)
- build(deps): bump authlib from 1.8.0 to 1.9.1 @dependabot[bot] (#1045)
- build(deps): bump virtualenv from 21.14.9 to 21.14.13 @dependabot[bot] (#1044)
- feat(ptc): programmatic tool calling with /tools tool-surface view @evalstate (#1034)
- fix: report poll wait budget exactly instead of re-deriving it from monotonic floats @evalstate (#1033)
- fix: keep commas in MCP URL query strings when splitting --url lists @ningyixi (#1024)
- fix: preserve text attachments with MIME parameters @BlueLi66 (#1025)
- fix: resolve local document file URIs @BlueLi66 (#1027)
- Align MCP skills wire models with the 2026-10-15 spec revision @olaservo (#1029)
