v0.3.1

What's Changed

• feat(policies): ToolGuard UI viewer modal and per-policy enable/disable (#419, #421) by @offerakrabi in #426
• knowledge(ui): redesign env-presets panel as Carbon ContainedList by @dolev31 in #383
• feat(rits): direct RITS connections with .env-driven model swapping by @Sergey-Zeltyn in #430
• chore(frontend): retire duplicate WorkspacePanel (#364) by @sami-marreed in #431
• fix: deliver HITL response in tool-approval e2e resume helper by @sami-marreed in #428
• fix(test): normalize unicode whitespace in tool-approval e2e assertions by @sami-marreed in #436
• fix: reject invalid generated Python before sandbox exec by @sami-marreed in #434
• fix(watsonx): correct gpt-oss-120b context profile to prevent negative max_tokens by @sami-marreed in #437
• feat: add MiniMax M3 provider by @octo-patch in #435
• refactor: split manage_routes into focused submodules by @sami-marreed in #429
• fix: confine ingest_knowledge file paths to workspace (CWE-22/73) by @sami-marreed in #447
• fix: harden watsonx context budget and skills syntax validation by @sami-marreed in #441
• fix(frontend): finalize stream and clear loading state on Stop by @sami-marreed in #450
• feat(policy): reuse pre-generated ToolGuard code across environments via policy export/import by @offerakrabi in #443
• fix(llm): only send top_p when explicitly configured by @Sergey-Zeltyn in #464
• fix(knowledge): collection-management & reindex state-machine hardening + tile/chip UX (closes #402 #398 #397 #396 #387) by @dolev31 in #410
• fix(cuga-lite): don't stop mid-task after stating a plan; warn on suspicious tool arguments; correct made-up tool names by @haroldship in #416
• refactor(test): replace run_tests.sh with global pytest entry point by @sami-marreed in #440
• refactor: break up ManagePage save logic so each setting saves in its own place by @sami-marreed in #465
• docs: point agents at CONTRIBUTING AI issue/PR commands by @sami-marreed in #485
• feat: add CODEOWNERS for required path reviews [skip ci] by @sami-marreed in #488
• Feature/toolguard generate batch by @offerakrabi in #473
• fix(security): stop false-positives on words like "requests" in tool args by @sami-marreed in #489
• feat: slash command for skill invocation with autocomplete + reasoning UX by @illeatmyhat in #235
• fix: route workspace tree to host when sandbox_mode is native/local by @sami-marreed in #490
• ci: wire watsonx vars and secrets into github actions jobs by @sami-marreed in #515
• feat(knowledge): inline citations & sources + simplified embeddings config UX by @dolev31 in #468
• fix(browser_planner): supply blank img when vision on but no screenshot by @segevshlomovIBM in #509
• fix: preload Docling ONNX heron layout for airgapped CPU ingest by @sami-marreed in #522
• feat(cuga_lite): first-class local Ollama / gpt-oss support by @segevshlomovIBM in #504
• fix(cuga_graph): default AgentState.input/url so empty-query resume doesn't crash by @segevshlomovIBM in #508
• fix(knowledge): make the embedding-model cache durable and self-healing by @dolev31 in #497
• fix(observability): use local OpenLit pricing in airgapped envs by @sami-marreed in #513
• Support MiniMax China endpoint selection by @octo-patch in #495
• fix(llm): rebind watsonx async client across pytest loops by @sami-marreed in #534
• feat: preserve agent configs from DB on startup when storage is prod by @sami-marreed in #514
• feat(knowledge): open a cited PDF at the cited page by @dolev31 in #498
• fix(272): CugaLite executes code that guesses at undeclared tool output shapes by @haroldship in #417
• feat(manage): configure max tokens and advanced LLM params by @sami-marreed in #512
• fix(appworld): auth/token lifetime, sandbox clock freeze, schema types, and prompt scoping by @sami-marreed in #545
• Sandbox timeout threw away everything a code block had done, so the agent just re-ran the same doomed loop by @haroldship in #493
• fix(deps): pin mcp[cli] below 2.0 by @segevshlomovIBM in #567
• fix(cuga-lite): persist set/frozenset variables across code steps by @sami-marreed in #572
• perf(startup): lazy imports — 66% faster SDK cold-start, sub-5ms import cuga (#486) by @offerakrabi in #535
• fix(cuga-lite): preserve stdout on error and hint assignment-RHS tool names by @sami-marreed in #548
• fix(context): summarization fixes for dead setting, silent truncation, Gemma window by @Sergey-Zeltyn in #565
• chore: require parent epic when filing agent issues by @sami-marreed in #576
• fix(cuga-lite): JSON-safe set/frozenset variable storage for stream events by @sami-marreed in #575

New Contributors

• @octo-patch made their first contribution in #435
• @illeatmyhat made their first contribution in #235

Full Changelog: v0.3.0...v0.3.1