cookbook/05_agent_os/interfaces/agui/TEST_LOG.md
Tests not yet run. Run each file and update this log.
Status: PASS
Description: Agent With Media - AG-UI agent (Google Gemini) that accepts multimodal user input.
Result: AgentOS boots with the AG-UI interface; /config returns 200. Media sent through POST /agui reaches the Gemini agent, which describes it accurately - verified with an image via CLI and interactively in a browser. Multimodal input path verified end-to-end.
Status: PENDING
Description: Silent External Tools - Suppress verbose messages in frontends.
Status: PENDING
Description: Agent With Tools.
Status: PENDING
Description: Basic.
Status: PENDING
Description: Multiple Instances.
Status: PENDING
Description: Reasoning Agent.
Status: PENDING
Description: Research Team.
Status: PENDING
Description: Outbound state synchronization via STATE_SNAPSHOT + STATE_DELTA events. Emits initial and final STATE_SNAPSHOT events plus STATE_DELTA JSON Patch ops after each state-mutating tool call.
Status: PENDING
Description: Structured Output.
Status: PASS
Description: HITL over AG-UI - backend confirmation (Shape B). A send_email tool gated by requires_confirmation=True; the human approves or declines a drafted email in the dojo, and on approval agno runs the tool server-side. Closes the confirmation / user_input gap that external-execution-only HITL leaves open.
Result: Live A/B verified on the dojo (backend on 127.0.0.1:9001, himanshu/agui-hitl worktree). Prompt "send an email to [email protected]" -> the agent drafts a subject/body and pauses -> the confirmation card renders (TOOL_CALL_START send_email -> TOOL_CALL_ARGS -> TOOL_CALL_END -> RUN_FINISHED paused, no result). Confirm -> agno runs send_email (DB: confirmed=True, result "Email sent to [email protected] with subject ...") and the agent confirms the send. Reject -> tool not run (DB: confirmed=False, result=None) and the agent acknowledges the email was not sent - visibly distinct from accept. Backed by unit tests: 14/14 in tests/unit/os/interfaces/test_agui_hitl.py (emission for confirmation / user_input / external pauses, pause-type-aware resolution incl. the single-shape user_input contract that fails loud on a malformed payload, the duplicate-tool-name dedupe, sync + async stream wrappers, and the partial-answer resume guard) and 69/69 in the full os/interfaces suite with zero regression. A 4-mutation matrix confirms the load-bearing tests are non-vacuous. ruff check + mypy clean (0 new errors) on the changed files. (Secrets/keys redacted; recipient is a placeholder.)
Status: PASS
Description: HITL over AG-UI - user_feedback (multiple choice). An agent with UserFeedbackTools (ask_user) pauses to ask the human to pick from a fixed option set; the choice returns as a ToolMessage {"selections": {<question>: [<labels>]}} and the AG-UI interface resolves it via RunRequirement.provide_user_feedback(...), continuing server-side. Closes the user_feedback case the confirmation / user_input work deferred; emission is unchanged (an ask_user tool already carries requires_user_input=True, so it surfaces via the existing partition and is not double-emitted).
Result: Live round-trip verified against a local AgentOS (127.0.0.1:9001, agui-hitl-v0 worktree; keys/ids redacted). Turn 1 - prompt "help me pick a cuisine for dinner" -> the agent calls ask_user and pauses: TOOL_CALL_START (ask_user) -> TOOL_CALL_ARGS (one question, four options Italian / Mexican / Thai / Japanese, multi_select=false) -> RUN_FINISHED with no tool result. Turn 2 - resume with the trailing ToolMessage {"selections": {"<question>": ["Italian"]}} (tool_call_id echoed from Turn 1) -> RUN_STARTED -> TEXT_MESSAGE_CONTENT where the agent acts on the pick ("Great choice - Italian it is ...") -> RUN_FINISHED, no RUN_ERROR. Fail-loud guard also confirmed live: a selections map whose key does not match the paused question leaves the requirement unresolved and the resume raises Partial resume: 1 of 1 paused tool(s) unanswered instead of silently proceeding. Backed by unit tests: 18/18 in tests/unit/os/interfaces/test_agui_hitl.py (the four new cases: emit-exactly-once guard, selections resolution, the fail-loud {'selections': {...}} contract, empty-selections not-resolved) and 73/73 in the full os/interfaces suite with zero regression; handlers.py / resume.py untouched (one additive input.py branch). ruff check + ruff format + mypy clean on the changed files.