shumengya/sproutclaw
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(ai): avoid double-counting reasoning tokens

Browse Source 
closes #3581
This commit is contained in:
Mario Zechner
2026-04-23 23:28:16 +02:00
parent ffa0f31239
commit c681d35d76
3 changed files with 48 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
packages/ai/CHANGELOG.md
View File

@@ -8,6 +8,8 @@
### Fixed
- Fixed OpenAI-compatible completion usage parsing to stop double-counting reasoning tokens already included in `completion_tokens` ([#3581](https://github.com/badlogic/pi-mono/issues/3581))
- Fixed long cache retention compatibility by adding `compat.supportsLongCacheRetention`, allowing Anthropic Messages and OpenAI-compatible proxies to explicitly disable long-retention fields while enabling long retention by default when requested ([#3543](https://github.com/badlogic/pi-mono/issues/3543))
- Fixed `openai-responses` compatibility by adding `compat.sendSessionIdHeader: false`, allowing strict OpenAI-compatible proxies to omit the underscore-containing `session_id` header while still sending other session-affinity headers ([#3579](https://github.com/badlogic/pi-mono/issues/3579))
- Fixed `anthropic-messages` tool streaming compatibility by adding `compat.supportsEagerToolInputStreaming`, allowing Anthropic-compatible providers to omit per-tool `eager_input_streaming` and use the legacy fine-grained tool streaming beta header instead ([#3575](https://github.com/badlogic/pi-mono/issues/3575))
- Fixed `supportsXhigh()` to recognize `openai-codex` `gpt-5.5`, preserving `xhigh` reasoning requests instead of clamping them to `high`.