platform: v1.3.43

1.3.43 (2026-08-25)

Features

• agents: routed wizard and detail pages for agents, LLM proxies and MCP gateways (#7352) (fe4bfbd)
• chat: move conversations between projects (#7431) (5d3c339)
• costs: add personal usage and billing-aware cost analytics (#7396) (154cb19)
• Cursor as a first-class LLM proxy client (#7375) (fc24673)
• frontend: agent card parity with gateway cards, and chip Skills as New (#7440) (ec4c6a6)
• frontend: give Skills and Plugins one sidebar row with tabs (#7439) (222dc17)
• frontend: refine entity and MCP alert workflows (#7401) (78bd745)
• frontend: rework Personal Settings and give settings a shared side nav (#7347) (83f0c12)
• frontend: unify loading states (#7427) (d103937)
• knowledge: add per-passage contextual retrieval (#7414) (4866bce), closes #7155
• knowledge: add pluggable retrieval backend (#7428) (c3d2e16)
• knowledge: BM25 keyword ranking in plain SQL (#7330) (865ea4c), closes #7158 #7162
• knowledge: cancel running connector syncs (#7415) (e8453f8)
• knowledge: Cohere direct multimodal embedding, and make retrieved images reach the user (#7355) (ead7007)
• knowledge: parent/child chunk indexing (#7436) (920c734)
• knowledge: scope a knowledge query to a subset of a connector's documents (#7437) (597bdb5)
• knowledge: Vertex AI multimodal embedding (multimodalembedding@001) (#7380) (f3fadce)
• knowledge: Voyage AI embeddings, the first embeddings-only provider (#7363) (67faa38)
• logs: disambiguate agents in the LLM and MCP log filters (#7361) (bfcec32)
• models: let admins set a model's context window and max output tokens (#7351) (f049554)
• one bulk actions bar, and a bulk API to back it (#7377) (4922f96)
• plugins: add marketplace management and client delivery (#7402) (81ba678)
• settings: compact layouts and improve table bulk selection (#7418) (0226fee)
• settings: give each built-in catalog its own settings page (#7350) (4cb811b)
• skills: add live MCP Skills (#7398) (cdd66d5)
• skills: move skill usage to a tab and say who ran the skill (#7444) (796e712)
• ui: unify responsive search and filter bars (#7408) (448d6ea)

Bug Fixes

• agents: make the member default agent an explicit pin, so the org default reaches everyone (#7359) (dbc3a5d)
• agents: render system prompts around malformed expressions (#7421) (85e3b02)
• auth: restore built-in SSO registration and account linking (#7413) (ab52fa6)
• bedrock: carry redacted extended thinking end to end (#7385) (d1c4738)
• chat: let a message be queued while context compaction is running (#7371) (59e2da1)
• chat: let the composer's microphone listen in the speaker's language (#7354) (fbd3c2c)
• chat: make stop cancel queued turns (#7411) (e6a4ad2)
• chatops: identify subagent inference errors (#7417) (39a0cb2)
• chatops: make the 🔇 reaction reliably mute a thread — and say that it did (#7442) (03ba46f)
• costs: stop repeating the same day on cost chart time axes (#7353) (563c2d6)
• deps: swap npm's bundled tar for CVE-2026-73566 in the MCP base image (#7394) (2044b5a)
• docker: bump moby/go-archive to 0.3.0 for CVE-2026-17106 (#7376) (f84cb7e)
• files: serve sandbox PDFs inline so the file viewer can preview them (#7364) (b31675f)
• frontend: even spacing down the agent detail page, and one voice for its reach-this-agent channels (#7447) (f9e86a5)
• frontend: hide dismissed MCP alerts from server details (#7406) (5046f53)
• frontend: let list-page filter bars wrap so the search box survives a phone width (#7356) (d7ac6a7)
• frontend: one loading indicator, fixed in place, and a refresh that restores the page (#7443) (5bad33b)
• frontend: polish embedding settings layout (#7432) (bf67b7c)
• frontend: proxy /v2 to the backend so A2A calls return JSON, not the HTML shell (#7365) (023f615)
• frontend: search pages in command palette (#7412) (e57b9f1)
• frontend: slim the Overview row, unstrand the MCP server tabs, and stop Connect naming itself twice (#7445) (4df1da8)
• frontend: three UI paper-cuts across files, costs, and plugins (#7435) (668b067)
• gemini: reach Vertex AI models the configured region does not serve (#7378) (cd9aa45)
• knowledge: keep retrieved-image parts to model-facing callers, inline base64 for the rest (#7358) (09b143b)
• knowledge: let "select all that match" reach a whole corpus (#7381) (1a0b455)
• knowledge: polish the connector detail tables and redesign the knowledge bases list (#7392) (1647222)
• knowledge: surface silent ingestion loss (#7426) (a1c59e1), closes #7319
• llm-models: disambiguate colliding model names at read time (#7374) (4bd1dd2)
• llm-models: stop two models sharing one name in the model pickers (#7372) (cb4b74c)
• llm-providers: stop badging non-token-billed models as free (#7373) (075057e)
• llm-proxy: hand back a direct call in dispatch mode too, rather than ending the turn (#7386) (ca087c6)
• llm-proxy: keep the model's own output on the remaining six adapters (#7389) (3e8aa24)
• llm-proxy: keep the model's own output when a refusal replaces the turn (#7388) (a7f591a)
• llm-proxy: learn a client's gateway label instead of assuming it (#7387) (14b4b39)
• llm-proxy: make a refused turn identifiable, and record the reasoning it produced (#7393) (0ac0595)
• llm-proxy: record the reasoning openai, zhipuai and bedrock turns produced (#7403) (92e44b3)
• llm-proxy: see through a client alias when unwrapping a run_tool dispatch (#7384) (7151140)
• llm-proxy: stop a blocked tool call from ending an agent's turn (#7382) (8b45829)
• llm-proxy: stop dropping Claude prompt-cache reads from reported usage (#7360) (fb5359b)
• logging: apply redaction to captured task logs (#7430) (6a75cb5)
• policy-subagent: resolve LLM through shared chain (#7416) (8ccb159)
• settings: keep the save bar floating for the whole settings page (#7362) (685e675)
• settings: stop the selected theme's checkmark overlapping its label (#7357) (a062223)

Performance Improvements

• logs: speed up session summaries and harden history reads (#7419) (bfee313)

Documentation

• clarify agent access vs MCP server access precedence (#7345) (e76d6c9)

Dependencies

• bump the rust-dependencies group in /platform/archestra-rs with 3 updates (#7366) (44cabb5)

Code Refactoring

• llm: remove the optimization rules feature (#7395) (b6a0ae1)

Miscellaneous Chores

• clean up log tables (#7425) (8b09a05)
• dev: bootstrap husky from Tilt so pre-commit hooks actually run (#7407) (b9872b2)
• polish guardrails, registry, and skills tables (#7429) (a9d4f2a)
• polish registry and plugin tables (#7433) (e5ed2f6)
• redesign OAuth callback experience (#7420) (42fa7fb)
• rust: declare the rust-analyzer component in rust-toolchain.toml (#7405) (093af4b)
• unify table and card layouts across entity pages (#7409) (c1475c7)