platform: v1.3.19

1.3.19 (2026-07-28)

Features

• chat: surface Claude Opus 5 default-on thinking (#6849) (f3cebef)
• chat: surface MiniMax thinking via reasoning_details translation (#6834) (d054175)
• chat: surface upstream Archestra thinking via openai-compatible client (#6837) (d3b3562)
• dock the hackathon submission review player in the chat panel (#6847) (4bfd5df)
• guardrails: default a coding CLI's native tools to Allow always at discovery (#6871) (605c310)
• guardrails: explain sensitive-context blocks, filter observed tools by user/client, session logs export (#6869) (fc68c5e)
• hackathon: one-minute final cut, cutting that shrinks the upload, and a limit that stops contradicting itself (#6902) (c6b1924)
• llm-logs: attribute knowledge base interactions to their connector (#6865) (9e39517)
• mcp-gateway: serve the 2026-07-28 MCP revision alongside 2025-11-25 (#6907) (8ac7a31)
• model provider subscriptions (#6864) (f71cdb9)
• models: team-scoped model access (#6806) (c895408)
• rbac: impersonation permission + sensitive-data log hygiene (#6877) (e876e87)
• resource oauth management (#6872) (d14e0a9)
• sandbox: provision Dagger engines in code, one per organization (#6430) (ad1dbe0)
• sandbox: raise run_command input limit to 256 KB (#6863) (ff7a28f)
• skills: add Org/Team/Personal scope filter to the Skills page (#6875) (8bb344d)

Bug Fixes

• address top recurring production error hotspots (#6874) (f2a083e)
• apps: make LLM and enhancement failures name their cause (#6898) (d6f5c50)
• chat: bundle LLM provider logos instead of loading them from models.dev (#6857) (eba873c)
• chat: stop token-bucket rejections from failing every chat request (#6885) (c39572d)
• chat: surface Azure AI Foundry open-model thinking in chat (#6840) (361cccd)
• chat: surface Kimi and vLLM thinking via openai-compatible provider (#6831) (3f370f0)
• chat: Surface OpenAI reasoning in Chat by requesting summaries (#6836) (53bc7ac)
• chat: surface Perplexity reasoning by extracting <think> blocks (#6838) (f58ca0c)
• chat: surface Zhipu AI GLM thinking as reasoning in chat (#6832) (a077c63)
• classify recurring external-failure errors across MCP runtime, chat, and model listing (#6887) (9137ff1)
• deps: clear four HIGH CVEs still shipping in the platform image (#6852) (a1bd849)
• deps: patch brace-expansion CVE-2026-14257 in the MCP base image (#6853) (0c1cf36)
• deps: raise shell-quote override to >=1.9.0 to clear GHSA-395f-4hp3-45gv (#6909) (483835b)
• frontend: expand advanced settings before asserting the primary-key switch (#6868) (3702306)
• frontend: never surface raw HTML error pages in error toasts (#6856) (ea54533)
• frontend: stop the limits page crashing when a limit is created or deleted (#6867) (906e90e)
• gemini: tolerate blocked and nonconforming generateContent responses (#6892) (28ce49a)
• hackathon: harden ai generated prompt and description in recorder (#6846) (043cb1e)
• hackathon: keep the recording replay in its recorded order (#6900) (a7c582f)
• hold the replay clock until the app frame really renders (#6850) (9fe86f5)
• interactions: harden tool-arg JSON parse and last tool-call id (#6884) (4c2a7be)
• llm-cost: exclude subscription usage from default user limits, detect ChatGPT Codex credentials (#6886) (9db0c01)
• llm-cost: skip subscription-billed interactions in cost-limit accumulation (#6876) (f094bbb)
• llm-providers: surface OpenRouter reasoning deltas in chat (#6835) (81ede65)
• llm-proxy: forward Bedrock reasoning deltas to streaming clients (#6833) (414bb4d)
• llm-proxy: harden ollama-native content and Bedrock stopReason schemas (#6894) (da8579e)
• llm-proxy: record interactions for streams that end without usage (#6896) (eaa45e6)
• llm-proxy: tolerate omitted message.content and unknown finish_reasons (#6881) (f46abf0)
• make replay start fast and visibly loading, top-align review layout (#6851) (4cfb267)
• mcp-catalog: roll the shared deployment on a multitenant execution edit (#6870) (6d0c8ef)
• models: price dated snapshot ids and carry reseller capabilities (#6895) (a94cab2)
• observability: classify outbound connectivity failures and group chat MCP-gateway outages (#6899) (dd14e41)
• projects: don't list instructions.md twice in a project chat's Files panel (#6854) (1032dcf)
• sandbox: bound the memory an engine's sandboxes can hold (#6862) (84a66c0)
• sandbox: tear down an environment's Dagger engine when it is deleted (#6861) (3c4d479)
• sso: honour the org default role on first-time SSO login (#6866) (7678df2)
• ui: readable page headers and tables on phone and tablet widths (#6873) (ec69a53)

Performance Improvements

• db: kill registry-list author-privilege N+1 and cache gateway knowledge-source presence (#6901) (4da1b91)

Miscellaneous Chores

• mcp-catalog: remove evaluation-pipeline fields from catalog data + client (#6811) (649d7ed)
• platform: tighten authorization and untrusted-input handling (#6878) (7bb6e78)