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

docs(changelog): audit unreleased entries

Browse Source
This commit is contained in:
Mario Zechner
2026-04-20 23:53:53 +02:00
parent 14ad636fb5
commit 639da8a602
3 changed files with 42 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
packages/ai/CHANGELOG.md
View File

@@ -2,13 +2,22 @@
## [Unreleased]
### Added
- Added `PI_OAUTH_CALLBACK_HOST` support for built-in Anthropic, Gemini CLI, Google Antigravity, and OpenAI Codex OAuth flows, allowing local callback servers to bind to a custom interface instead of hardcoded `127.0.0.1` ([#3409](https://github.com/badlogic/pi-mono/pull/3409) by [@Michaelliv](https://github.com/Michaelliv))
### Changed
- Changed Bedrock Converse requests to omit `inferenceConfig.maxTokens` when model token limits are unknown and to omit `temperature` when unset, letting Bedrock use model defaults and avoid unnecessary TPM quota reservation ([#3400](https://github.com/badlogic/pi-mono/pull/3400) by [@wirjo](https://github.com/wirjo))
### Fixed
- Fixed `openai-completions` `compat.requiresThinkingAsText` assistant replay to preserve text-part serialization and avoid same-model crashes when prior assistant messages contain both thinking and text ([#3387](https://github.com/badlogic/pi-mono/issues/3387))
- Fixed Cloud Code Assist tool schemas to strip JSON Schema meta-declaration keys such as `$schema`, `$defs`, and `definitions` before sending OpenAPI `parameters`, avoiding provider validation failures for tool-enabled requests ([#3412](https://github.com/badlogic/pi-mono/pull/3412) by [@vladlearns](https://github.com/vladlearns))
- Fixed non-vision model requests to replace user and tool-result image blocks with explicit text placeholders instead of silently dropping them during provider payload conversion ([#3429](https://github.com/badlogic/pi-mono/issues/3429))
- Fixed OpenRouter Meta tests by switching `meta-llama/llama-4-maverick` to `meta-llama/llama-4-scout` to avoid type-check failures from model-catalog drift.
- Fixed direct OpenAI Chat Completions requests to map `sessionId` and `cacheRetention` to OpenAI prompt caching fields, sending `prompt_cache_key` when caching is enabled and `prompt_cache_retention: "24h"` for direct `api.openai.com` requests with long retention ([#3426](https://github.com/badlogic/pi-mono/issues/3426))
- Fixed OpenAI-compatible Chat Completions requests to optionally send aligned `session_id`, `x-client-request-id`, and `x-session-affinity` session-affinity headers from `sessionId` via `compat.sendSessionAffinityHeaders`, enabling cache-affinity routing for backends such as Fireworks ([#3430](https://github.com/badlogic/pi-mono/issues/3430))
- Fixed direct Bedrock runtime client construction to pass `model.baseUrl` through as the SDK `endpoint`, restoring support for custom Bedrock endpoints such as VPC or proxy routes ([#3402](https://github.com/badlogic/pi-mono/pull/3402) by [@wirjo](https://github.com/wirjo))
- Fixed OpenAI-compatible Chat Completions Anthropic-style prompt caching to apply `cache_control` markers to the system prompt, last tool definition, and last user/assistant text content via `compat.cacheControlFormat`, and enabled that compat for OpenCode/OpenCode Go Qwen 3.5/3.6 Plus models so prompt caching works there too ([#3392](https://github.com/badlogic/pi-mono/issues/3392))
## [0.67.68] - 2026-04-17