v0.3.0

CUGA Agent v0.3.0

A major step forward for multi-agent orchestration, policy enforcement, knowledge, and enterprise integrations — with stronger reliability across WatsonX, MCP, and the SDK.

---

Highlights

• ToolGuard — Runtime policy enforcement for tool calls, plus UI-driven guard generation
• Supervisor & A2A — Full policy runtime (approval + output formatting), A2A v0.3 exposure, two-CUGA demo, and HITL round-trip
• Skills & memory — Configurable skills root, workspace uploads, SDK skill discovery, and semantic user-fact memory in CugaLite
• Knowledge — Faster ingest, OCR, session isolation, and live UX improvements
• Enterprise-ready — WatsonX context fixes, improved AI config, configurable LLM timeouts, and multi-user context for Evolve

---

New features

Agents & orchestration

• Supervisor delegation, load tests, and stream concurrency improvements — #326 (@sami-marreed)
• Full supervisor policy runtime — Tool approval and output formatter — #349 (@sami-marreed)
• A2A v0.3 — Expose CUGA over A2A with a working two-CUGA demo — #323 (@laredo)
• A2A human-in-the-loop — Approval round-trip for the simple runner — #361 (@laredo)
• Sandbox-aware execution and Carbon side panel refresh — #225 (@sami-marreed)
• Semantic memory — User facts for CugaLite — #180 (@gaodan-fang)

ToolGuard & policies

• ToolGuard runtime enforcement for tool calls — #329 (@offerakrabi)
• Generate tool guards from the UI — #370 (@offerakrabi)

Skills & SDK

• Configurable skills root, workspace uploads, and SDK skill discovery — #366 (@sami-marreed)
• Token usage tracking in the CUGA SDK for trajectory support — #236 (@haroldship)

Knowledge

• Ingest performance, OCR, session isolation, and live UX — #352 (@dolev31)

Configuration & integrations

• Configurable LLM HTTP timeouts — #346 (@sami-marreed)
• Multi-user context (user_id, namespace_id, session_id) propagated to Evolve MCP — #194 (@gjt-prog)
• Improved WatsonX AI configuration — #200 (@Tissandier)

Examples & documentation

• Travel Agent — Multi-agent corporate travel planning demo — #201 (@offerakrabi)
• Knowledge demo example — #156 (@dolev31)
• README — Date AppWorld and WebArena #1 benchmark claims — #348 (@sami-marreed)

---

Reliability & bug fixes

WatsonX & LLM

• Resolve WatsonX context window overflow (#189) — #207 (@offerakrabi)
• Honor TOML temperature / top_p; add sampling-config INFO log — #405 (@Sergey-Zeltyn)
• Policy conflict resolution uses json_schema for structured output (gpt-oss / WatsonX) — #414 (@segevshlomovIBM)
• Default WatsonX model to openai/gpt-oss-120b — #423 (@segevshlomovIBM)

MCP & tools

• Fix missing MCP server authentication for SSE transport — #208 (@Tissandier)
• Regression tests for SSE transport auth (#191) — #234 (@offerakrabi)
• Sanitize dashed MCP tool names to valid Python identifiers — #233 (@offerakrabi)
• Propagate auth query params to MCP transport URLs — #345 (@sami-marreed)
• Include nested tool schemas in code-agent prompt — #237 (@offerakrabi)
• Cap bind_tools count and shortlist over the limit — #203 (@haroldship)

CugaLite & execution

• Recover code from unclosed ```python fences — #252 (@haroldship)
• Reset task_todos state on fresh conversation — #315 (@Sergey-Zeltyn)
• Serialize pandas DataFrame/Series and numpy scalars in variable collector — #224 (@Sergey-Zeltyn)
• Normalize numpy types at producer boundaries for checkpointing — #341 (@sami-marreed)
• Configurable sandbox code-execution timeout (default 30s) — #365 (@Sergey-Zeltyn)
• Serialize ToolGuard objects before returning generation response — #409 (@offerakrabi)

Supervisor, SDK & UI

• Expose A2A subagent endpoint in /api/config/subagents — #339 (@syhcode)
• Handle Pydantic model iteration in _agent_card_description — #308 (@syhcode)
• SDK HITL resume and tool-approval denial crash — #380 (@sami-marreed)
• Fall back to base_tools in ChatAgent.execute_tool when tools is None — #340 (@sami-marreed)
• Show Carbon loading indicator during pre-stream window — #265 (@nir-mashkif)
• Gate browser planner multimodal content on per-model vision support — #251 (@haroldship)

Policies, tracing & evaluation

• Load .cuga policies in demo mode — #223 (@offerakrabi)
• Propagate Langfuse callbacks to nested LLM calls — #277 (@haroldship)
• Make TraceScopedLangfuseCallbackHandler override reachable — #358 (@haroldship)
• Skip Langfuse retry loop when credentials are missing — #328 (@Sergey-Zeltyn)

---

Performance

• Default to planning on any CugaLite task with dependent steps — #331 (@Sergey-Zeltyn)
• Knowledge ingest performance improvements — #352 (@dolev31)
• Supervisor delegation and stream concurrency optimizations — #326 (@sami-marreed)

---

Security

• Bump Playwright to ≥1.59.0 (CVE-2025-9611, CVE-2025-59288) — #212 (@haroldship)
• Remediate Dependabot alerts — #344 (@sami-marreed)
• Resolve high-severity Dependabot alerts; drop milvus-lite — #393 (@sami-marreed)

---

Architecture & maintenance

• Extract shared agent logic into cuga_agent_core module — #250 (@sami-marreed)
• Remove dead/unused backend files — #298 (@sami-marreed)
• Remove dead code from frontend workspaces — #300 (@sami-marreed)
• Replace loguru %s/%d format strings with f-strings — #389 (@Maqbool61)
• E2E test suite for skills components — #232 (@Iftach-Shoham)
• Reduce Windows stability tests to a single smoke case — #424 (@sami-marreed)

---

New contributors

First contributions to CUGA — thank you!

Contributor	PR
@Tissandier	#208
@nir-mashkif	#265
@syhcode	#308
@laredo	#323
@Iftach-Shoham	#232
@Maqbool61	#389
@segevshlomovIBM	#414

---

Full changelog: v0.2.28...v0.3.0