Skip to content

CHANGELOG.md

Latest commit

 

History

History
158 lines (86 loc) · 6.34 KB

CHANGELOG.md

File metadata and controls

158 lines (86 loc) · 6.34 KB

Release notes

0.0.16

New Features

0.0.15

New Features

  • Resize images returned from tool call (Anthropic) (#41)

Bugs Squashed

  • Anthropic messages: text content blocks must be non-empty (#44)
  • Invalid 'input[95].call_id': string too long. Expected a string with maximum length 64, but got a string with length 116 instead. (#45)
  • Fireworks generates tool_call.ids that result in Anthropic API calls failing (#42)
  • Claude server tool_result yield bug (#40)
  • fix truncation (#39), thanks to @RensDimmendaal

0.0.14

New Features

  • Anthropic messages: text content blocks must be non-empty (#44)
  • Resize images returned from tool call (Anthropic) (#41)

Bugs Squashed

  • Invalid 'input[95].call_id': string too long. Expected a string with maximum length 64, but got a string with length 116 instead. (#45)
  • Fireworks generates tool_call.ids that result in Anthropic API calls failing (#42)
  • Claude server tool_result yield bug (#40)
  • fix truncation (#39), thanks to @RensDimmendaal

0.0.12

New Features

  • Add CODEX_AUTH_PATH (#38)

0.0.11

New Features

  • rename tools to be as safepyrun/safecmd expects (#36), thanks to @RensDimmendaal
  • Add Mimo vendor, Opus 4-8, modern_llm preset, and fix Anthropic cache cost fallback (#35)

Bugs Squashed

  • Wrong roles in mk_msgs when fmt2hist ends with tool (#37)
  • Update 'gpt-5.3-codex-spark' meta to support tools (#34)
  • Model name changes from the server response invalidates model meta patches (#33)

0.0.10

New Features

  • MediaUrl for direct url handling without byte reading (#31)
  • Add timeout to mk_client (#26)
  • Add retry logic with exponential backoff to acomplete (#25), thanks to @ncoop57
  • Yield tool calls JIT (#19)

Bugs Squashed

  • force enable web_search for codex models (#22), thanks to @jackhogan

0.0.9

New Features

  • Add approx_pricing helper and fix Fireworks Kimi k2p6 model registration with pricing (#21)
  • Refactor model info into registry; add get_model_pricing; add new gemini models (#20)

Bugs Squashed

  • Fix accounts/fireworks/models/kimi-k2p5 registration (#23)
  • force enable web_search for codex models (#22), thanks to @jackhogan

0.0.8

New Features

  • Add finalize_usage to fix anthropic reasoning token tracking; Add debug brief mode (#18)

  • make old web search tool web_search_20250305 the default (#16)

  • fastllm chat debug mode (#15)

    • Image

  • Track reasoning tokens in Anthropic usage; handle token details in mk_msgs; fix _trunc_param escaping; add codex auth module (#13)

Bugs Squashed

  • 𝍁...𝍁 print() rstrip, summary truncation, non-ascii fix (#17)

  • claude-opus-4-7 thinking bug (#14)

0.0.7

New Features

  • AsyncChat callback system (#11)

Bugs Squashed

  • markdown='1' (#12)

  • stop status getting reset in streaming loop (#10)

    • Fix is to only check the stop condition if it's not met yet, e.g.
if not stop: stop = stop_and_trim(part_accum, d, stop_callables)

0.0.6

Bugs Squashed

  • model param in mk_client should be optional (#9)

0.0.5

Bugs Squashed

  • get_model_info strict param (#8)

    • If a model can't be resolved in get_model_info you can pass strict=False to get placeholder price values to avoid errors with AsyncChat. For example, we use strict=True in solveit and strict=False in shell sage where users can pass their custom models with custom base urls.

  • Unresolved model fixes (#7)

    • When a model and it's info can't be resolved:

  • Default to openai_chat api if not provided

  • Default max tokens to 32k

  • Set pricing to codex values

0.0.4

Bugs Squashed

  • Code fence tool fixes (#6)
    • _split_msg_on_fences fix which now correctly handles mixed msg content, e.g. thinking + text, tool use + text etc..
    • _fence_re new line start check
⚡