v1.2.1

Highlights

• Mistral voice for ag2.live: offline and realtime speech-to-text (MistralTranscriber, MistralRealtimeTranscriber) and text-to-speech (MistralTTSConfig, MistralStreamingTTSConfig), on top of mistralai 3.2 with its new chat parameters.
• OpenAIResponsesConfig accepts a reasoning option for effort and summaries.
• New search toolkits: FirecrawlToolkit and KeenableSearchToolkit.
• New compaction_verifier extension: measures whether a CompactStrategy makes the agent's next actions worse, not just how many tokens it saves.
• TypeSafe can locate objects in images: ResponseSchema(BoundingBox) sends a Box question.
• ACP and the network layer keep multimodal input parts instead of flattening them to text.
• Fixes for Ollama, Gemini and xAI history, paginated file listings, and atomic SQLite and Redis knowledge appends.

Upgrade notes

• Mistral uses httpx2. mistralai 3.x moved from httpx to httpx2, so MistralConfig.async_client is now an httpx2.AsyncClient and the extra requires mistralai>=3.2.0,<4. An httpx.AsyncClient still works at runtime but no longer type-checks, and httpx.HTTPError does not catch httpx2 errors.
• ElevenLabs now requires elevenlabs>=2.73.0,<3.
• Input parts are no longer joined into one string in ACP and the network layer. ACP sends each part as its own content block, based on what the agent advertises; before, every non-text part was dropped. ChannelAdapter.render_envelope and extract_turn_input can now return an Input or a sequence of them, so check custom channel adapters.
• MCP tool calls are pinned to the configuration used for discovery, so a Variable(default_factory=...) can no longer route a call to a different server than the one that advertised the tool.
• OpenAI Chat Completions rejects files uploaded to another provider with UnsupportedInputError, like the Responses, Anthropic and Gemini mappers.

What's Changed

• chore(deps): bump j178/prek-action from 3.0.0 to 3.0.1 in the github-actions group by @dependabot[bot] in #3405
• chore(deps): bump the uv group with 16 updates by @dependabot[bot] in #3406
• fix(ollama): serialize multipart tool results as text by @ch-z-hc in #3374
• feat(openai): add reasoning option to OpenAIResponsesConfig by @vvlrff in #3407
• feat(mistral): bump mistralai to 3.2 and expose new chat parameters by @vvlrff in #3408
• fix: keep input parts separate instead of joining them into a string by @vvlrff in #3413
• fix(annotations): resolve falsy Variable and Inject values from context by @kabishou11 in #3414
• fix(events): fall back to media type for unknown binary kinds in sizing by @kabishou11 in #3416
• fix(eval): apply threshold keys to passthrough feedback by @BichengWang in #3423
• fix(gemini): preserve function call IDs in history by @ch-z-hc in #3417
• fix(shell): check paths after '=' against the ignore list by @kabishou11 in #3415
• fix(xai): preserve conversation history order by @ch-z-hc in #3418
• feat(extensions/search): add Firecrawl search and scrape toolkit by @JuampiHernandez in #3362
• feat(live): add MistralTranscriber for offline speech-to-text by @vvlrff in #3419
• feat(live): add Mistral TTS configs by @vvlrff in #3420
• fix(typesafe): bump typesafe-sdk to 0.7.5 and handle unexpected answe… by @vvlrff in #3424
• fix(elevenlabs): support elevenlabs 2.73 and expose Scribe/TTS options by @vvlrff in #3437
• fix(files): enumerate all file listing pages by @ch-z-hc in #3429
• fix(a2a): accept empty text, raw and url parts in part_to_input by @kabishou11 in #3440
• fix(openai): reject files from another provider in Chat Completions messages by @kabishou11 in #3441
• test(ag-ui): make the pre-question client call test deterministic by @vvlrff in #3444
• fix(knowledge): remove empty parent directories after DiskKnowledgeStore.delete by @kabishou11 in #3439
• fix(mcp): pin tool calls to the discovered configuration by @ch-z-hc in #3435
• feat(typesafe): support Box questions with a BoundingBox response schema by @vvlrff in #3436
• fix(events): honor explicit media types for local paths by @ch-z-hc in #3430
• fix(xai): preserve file creation timestamps by @ch-z-hc in #3431
• fix(gemini): honor the configured HTTP client for files by @ch-z-hc in #3427
• fix: make SQLite knowledge append atomic across stores by @liangzhechen655 in #3388
• fix(eval): accept os.PathLike in human_labels by @BichengWang in #3376
• fix: notify knowledge watchers of both renamed paths by @liangzhechen655 in #3389
• fix: normalize knowledge store lock paths by @liangzhechen655 in #3390
• fix: make Redis knowledge appends atomic across connections by @liangzhechen655 in #3391
• fix(response): accept Literal schemas by @ch-z-hc in #3432
• test(config): cover ToolSearchTool rejection in unsupported-provider tests by @BichengWang in #3375
• chore(deps): bump the github-actions group with 4 updates by @dependabot[bot] in #3453
• test(xai): share one gRPC Files server helper across file tests by @vvlrff in #3448
• feat(live): add Mistral realtime transcriber by @vvlrff in #3421
• chore(deps): bump aiohttp from 3.14.3 to 3.14.4 by @dependabot[bot] in #3458
• feat(extensions): add mi4afa probing failure attribution by @wuliang211 in #3277
• feat(extensions/search): add Keenable search toolkit by @ilya-bogin-keenable in #3373
• docs(extensions/search): link Keenable API docs for request and response fields by @vvlrff in #3463
• feat(extensions): add compaction_verifier — measure whether a CompactStrategy makes the agent's next actions worse by @wuliang211 in #3306
• fix(ollama): keep tool call IDs unique across turns by @ch-z-hc in #3426
• chore: bump version to 1.2.1 by @vvlrff in #3467

New Contributors

• @ch-z-hc made their first contribution in #3374
• @JuampiHernandez made their first contribution in #3362
• @wuliang211 made their first contribution in #3277
• @ilya-bogin-keenable made their first contribution in #3373

Full Changelog: v1.2.0...v1.2.1