test(a2a): cover the A2AAgent component in Internal mode (C1)

[rafaelgiln]

Spec C1 of the A2A inventory in docs/core-functionality/a2a/a2a-coverage-scope.md — planned in #1195, never written. Closes the first [ ] bullet of QA-CHECKLIST.md §16.2 (line 650).

Problem

A2A has two halves: serving an agent and consuming one. #1349 promoted 7 specs covering the server; the A2AAgent component — the entire client half — has zero coverage.

mode=Internal is the half that is reachable today: it lists locally published agents in a dropdown and calls one over the same /jsonrpc endpoint the server specs already prove. It does not depend on the loopback-SSRF question that blocks C2/External (LE-1904 class) — that constraint applies only to pointing the component at an external URL, including this instance's own.

What to prove

Two flows in one test:

1. Flow A — the published passthrough (createRunnableChatFlowViaApi → PATCH flow_type=agent, a2a_enabled=true). Same shape the server specs publish.
2. Flow B — a second flow carrying an A2AAgent node with mode=Internal.

Observable	Expected
agent_name_selected dropdown in Flow B	lists Flow A's published agent (refreshable input, populated from local published agents)
Running Flow B	the Response output contains the per-run sentinel Flow A echoes verbatim

No LLM on either side — the passthrough echoes Chat Input to Chat Output, so the sentinel is causal evidence the remote call actually executed the other flow's graph, not a "something came back" check.

Component facts from the surface map (measured)

lfx/components/models_and_agents/a2a_agent.py — display_name = "A2A Agent", name = "A2AAgent", icon = "bot".

Inputs: mode (TabInput Internal | External), agent_name_selected (Dropdown, refreshable, populated from local published agents), agent_url (External only), agent_card (DataDisplayInput), message (Multiline, tool_mode=True), api_key (SecretStr), timeout (Int, seconds). One output: Response.

Two things to watch:

• The dropdown is refreshable — it may need an explicit refresh after Flow A is published, since the node was added before. Assert the option is present rather than assuming the initial render has it.
• agent_name_selected is a name, not an id. Two flows with the same name would make the assertion ambiguous — give Flow A a per-run unique name.

Preconditions already available

• createRunnableChatFlowViaApi — the LLM-free passthrough plus id-scoped teardown.
• requireA2aEnabled() — first call, as in every spec of the area.
• The server specs already prove the endpoint Flow B will call, so a failure here isolates to the component, not the transport.

Deliverables

• tests/tests-automations/regression/core-functionality/a2a/a2a-client-agent-internal.spec.ts
• docs/core-functionality/a2a/a2a-client-agent-internal.md (spec doc first — SDD)
• QA-CHECKLIST.md §16.2 line 650 → [x] with the spec reference
• Tags: @stable @components @a2a (plus @workspace if the spec drives the canvas to build Flow B). LLM-free, so no provider dependency.
• Both flows deleted id-scoped in finally.

Decision to make during PLAN

Whether Flow B is built through the canvas (drag the A2AAgent in, set the dropdown) or via API from a fixture JSON. The canvas route also covers the dropdown population — which is half the observable — so it is the stronger option, but it is slower and pulls in the sidebar-add primitive. Scout it live before committing.

Out of scope

• C2 / External mode (LE-1845) — blocked until the loopback-SSRF question is answered; tracked separately.
• C3 / as a Tool (LE-1963) — the only LLM-dependent row of the area; tracked separately.