This release adds 3 notable features for engineering teams evaluating rollout.
✓ No known CVEs patched in this version
Affected surfaces
Summary
AI summaryAdds OpenRouter model routing, preserves reasoning metadata across turns, and fixes Codex browser‑auth replay failures.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Feature | Medium |
Add OpenRouter model routing for direct Agent usage and FastAPI overrides. Add OpenRouter model routing for direct Agent usage and FastAPI overrides. Source: llm_adapter@2026-06-10 Confidence: high |
— |
| Feature | Medium |
Preserve FastAPI file_urls provenance in replayed history for all endpoints. Preserve FastAPI file_urls provenance in replayed history for all endpoints. Source: llm_adapter@2026-06-10 Confidence: high |
— |
| Feature | Medium |
Normalize AG‑UI media content and keep file source snapshots valid across replay. Normalize AG‑UI media content and keep file source snapshots valid across replay. Source: llm_adapter@2026-06-10 Confidence: high |
— |
| Feature | Medium |
Allow handoff communication and default SendMessage to coexist for the same agent pair. Allow handoff communication and default SendMessage to coexist for the same agent pair. Source: llm_adapter@2026-06-10 Confidence: high |
— |
| Feature | Medium |
Preserve OpenRouter reasoning_details across all run types and replays. Preserve OpenRouter reasoning_details across all run types and replays. Source: llm_adapter@2026-06-10 Confidence: low |
— |
| Feature | Low |
Forward provider‑specific model settings from FastAPI clients, including LiteLLM reasoning configs. Forward provider‑specific model settings from FastAPI clients, including LiteLLM reasoning configs. Source: llm_adapter@2026-06-10 Confidence: high |
— |
| Feature | Low |
Surface LiteLLM thinking blocks as streamed reasoning events. Surface LiteLLM thinking blocks as streamed reasoning events. Source: llm_adapter@2026-06-10 Confidence: high |
— |
| Feature | Low |
Preserve OpenRouter reasoning_details across non‑streaming, streaming, encrypted, and replayed turns. Preserve OpenRouter reasoning_details across non‑streaming, streaming, encrypted, and replayed turns. Source: granite4.1:30b@2026-06-10-audit Confidence: low |
— |
| Dependency | Low |
Update runtime and development dependencies including OpenAI, Rich, websockets, and Ruff. Update runtime and development dependencies including OpenAI, Rich, websockets, and Ruff. Source: llm_adapter@2026-06-10 Confidence: high |
— |
| Bugfix | Medium |
Fix Codex browser‑auth replay failures by translating system-role input at final model call. Fix Codex browser‑auth replay failures by translating system-role input at final model call. Source: llm_adapter@2026-06-10 Confidence: low |
— |
| Bugfix | Low |
Strip stale OpenAI credential headers when copying base clients into OpenRouter swaps. Strip stale OpenAI credential headers when copying base clients into OpenRouter swaps. Source: llm_adapter@2026-06-10 Confidence: low |
— |
| Refactor | Low |
Route model‑only OpenRouter swaps through the OpenRouter endpoint instead of custom gateway clients. Route model‑only OpenRouter swaps through the OpenRouter endpoint instead of custom gateway clients. Source: granite4.1:30b@2026-06-10-audit Confidence: low |
— |
Full changelog
Minor release for stronger provider support, OpenRouter reasoning continuity, and safer FastAPI history replay.
This release is especially important for Agent Swarm CLI and OpenSwarm users who run Agency Swarm through FastAPI with provider-specific models. It adds first-class OpenRouter routing, preserves provider reasoning metadata across turns, and tightens several replay paths that matter for browser-auth, file attachments, AG-UI, and streamed responses.
What's Changed
Provider routing and reasoning
- Add OpenRouter model routing for direct
Agent(model="openrouter/...")usage and request-scoped FastAPI model overrides in https://github.com/VRSEN/agency-swarm/pull/655. - Preserve OpenRouter
reasoning_detailsacross non-streaming runs, streamed runs, encrypted-only reasoning, and second-turn replay in https://github.com/VRSEN/agency-swarm/pull/678. - Route model-only OpenRouter swaps through the OpenRouter endpoint instead of reusing unrelated custom gateway clients in https://github.com/VRSEN/agency-swarm/pull/678.
- Strip stale OpenAI credential headers when copying official-base clients into OpenRouter swaps in https://github.com/VRSEN/agency-swarm/pull/685.
- Forward provider-specific model settings from FastAPI clients, including LiteLLM reasoning settings for OpenAI, Anthropic, Gemini, and xAI paths, in https://github.com/VRSEN/agency-swarm/pull/654.
- Surface LiteLLM thinking blocks as streamed reasoning events in https://github.com/VRSEN/agency-swarm/pull/654.
FastAPI, Codex, and replay reliability
- Fix Codex browser-auth replay failures by translating replayed system-role input only at the final Codex-bound model-call boundary in https://github.com/VRSEN/agency-swarm/pull/651.
- Preserve FastAPI
file_urlssource provenance in replayed history for non-streaming, streaming, and AG-UI endpoints in https://github.com/VRSEN/agency-swarm/pull/676. - Normalize AG-UI media content parts and keep file source snapshots valid across replay in https://github.com/VRSEN/agency-swarm/pull/676.
- Allow handoff communication and default
SendMessagecommunication to coexist for the same agent pair in commit46ea8125without a pull request.
Package, CI, and dependency maintenance
- Bump Agency Swarm to
1.10.0for this minor release in https://github.com/VRSEN/agency-swarm/pull/677. - Improve live CI stability and route Dependabot PRs around secret-backed tests in https://github.com/VRSEN/agency-swarm/pull/670 and https://github.com/VRSEN/agency-swarm/pull/671.
- Stabilize file, tool-folder, instruction-loading, and code-interpreter assertions in https://github.com/VRSEN/agency-swarm/pull/659, https://github.com/VRSEN/agency-swarm/pull/661, https://github.com/VRSEN/agency-swarm/pull/662, https://github.com/VRSEN/agency-swarm/pull/663, and https://github.com/VRSEN/agency-swarm/pull/664.
- Update runtime and development dependencies, including OpenAI, Rich, websockets, datamodel-code-generator, fastmcp, griffe, ipykernel, and Ruff, in https://github.com/VRSEN/agency-swarm/pull/665, https://github.com/VRSEN/agency-swarm/pull/667, https://github.com/VRSEN/agency-swarm/pull/668, https://github.com/VRSEN/agency-swarm/pull/669, https://github.com/VRSEN/agency-swarm/pull/679, https://github.com/VRSEN/agency-swarm/pull/680, https://github.com/VRSEN/agency-swarm/pull/681, https://github.com/VRSEN/agency-swarm/pull/682, and https://github.com/VRSEN/agency-swarm/pull/683.
Compatibility Notes
- This is a minor release, not a major release. No intentional breaking public API change is included.
- OpenSwarm does not need its own release for these Agency Swarm fixes. Its Python dependency range can pick up
agency-swarm==1.10.0when the environment is upgraded or recreated. - A later patch release can continue polishing provider behavior without changing the
1.10.xline.
Full Changelog: https://github.com/VRSEN/agency-swarm/compare/v1.9.9...v1.10.0
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
Related context
Related tools
Earlier breaking changes
- v1.10.3 Prevents fresh installs with openai-agents==0.14.8 from resolving incompatible OpenAI >=2.45 versions.
Beta — feedback welcome: [email protected]