platform: v1.3.22
Immutable
release. Only release title and notes can be modified.
1.3.22 (2026-07-30)
Features
- a2a: full A2A v1.0 Task + Artifact support (#6967) (7469f17)
- a2a: group A2A routes/models into subdirs and make the AgentCard spec-conformant (#6969) (2b30700)
- a2a: name the provider on the agent card (#6976) (1369603)
- a2a: push notifications for disconnected clients (#6973) (0f92836)
- a2a: registry endpoint listing every card a credential can reach (#6977) (d1f748b)
- a2a: retention for finished A2A tasks (#6974) (a5295be)
- a2a: wake task streams across replicas with LISTEN/NOTIFY (#6975) (952655c)
- mcp-gateway: stop advertising unimplemented subscriptions, refuse removed methods (#6952) (c6782a7)
- mcp-gateway: subscriptions/listen and x-mcp-header support (#6953) (18695c1)
- mcp: bind credentials to their issuer, complete the result envelope, and finish the error-code migration (#6944) (47a64f3)
- mcp: Tasks extension for long-running tool calls — gateway and chat (#6956) (eebc51f)
- sandbox: raise the per-file cap to 50 MiB and sync large uploads via filesync (#6926) (a2ea594)
- sharing: share agents, MCP gateways, LLM proxies, skills, projects and models with named people (#6928) (216a1bb)
Bug Fixes
- a2a: reject contentless message turns instead of sending them to the provider (#6959) (7bc0ff3)
- apps: bind scaffolded apps to the agent's environment, accept Default tools as the app baseline (#6981) (031a132)
- apps: bound app tool calls at the LLM ceiling and cancel abandoned calls (#6963) (2feed0c)
- apps: let session replays show the images the live app CSP blocked (#6942) (6afdd4d)
- apps: let the review player load a submission as large as the gallery accepts (#6950) (f350153)
- deps: bump dompurify override to 3.4.12 to clear GHSA-cmwh-pvxp-8882 & GHSA-c2j3-45gr-mqc4 (#6940) (2c61608)
- frontend: cover /api/organization/members in the MSW handler set (#6948) (77af09b)
- github-copilot: verify model invocability at sync, surface an actionable model-not-supported error (#6964) (1dd549c)
- llm-cost: drop a Claude context-variant marker before recording a model (#6915) (b9623da)
- llm-cost: price a model the proxy discovers under a mismatched provider (#6946) (7e731bb)
- llm-cost: price the OpenAI Codex models the registry omits (#6957) (40fe031)
- llm-cost: stop pricing self-hosted models, and read the window vLLM reports (#6984) (e32495b)
- llm-proxy: classify client aborts as 499 and string-payload stream errors as upstream (#6971) (56d8e47)
- llm-proxy: mark status-less in-stream provider errors as upstream failures (#6968) (9be8744)
- oauth: forward the RFC 9207 iss parameter from the callback (#6949) (b015eed)
- teams: add the creator to the team they create, as its admin (#6931) (0765df4)
- white-label: resolve the deployment brand in user- and LLM-facing copy (#6965) (2f955b4)