fix(adk): stop embedding the system prompt in the pod agent card

[TigerkidYang]

Problem

The agent pod's /.well-known/agent-card.json embeds the agent's entire system prompt: EnrichAgentCard unconditionally replaces the curated card's skills with adka2a.BuildAgentSkills(agent), and adk-go's buildDescriptionFromInstructions joins the agent description with the full instruction into the first skill's description. Anyone who can discover the agent can read how it is instructed to behave (fee policies, escalation rules, …), and the pod card diverges from the curated card the controller serves. The reporter re-confirmed this on 0.10.0 stable.

Fix

EnrichAgentCard no longer touches skills at all. The compiler-generated card is authoritative (it is also what the a2a-gateway serves from the pinned runtime revision), so the pod card now carries exactly the skills the compiler put there and nothing built from the ADK agent. The function still fills a missing description, the HITL extension and the default interface. adka2a is no longer imported here.

The pod still serves /.well-known/agent-card.json to direct-to-pod callers (the reporter's agentgateway setup, and the 0.10.x line), which is where this closes the leak.

Delegation is not affected: the remote-A2A tools in both runtimes (go/adk/pkg/tools/remote_a2a_tool.go, kagent-adk/_remote_a2a_tool.py) read only the card Description, and the reporter confirmed empirically that routing between sub-agents never depended on skills.

Tests

agentcard_test.go: an agent with an instruction and a sub-agent. A card without skills stays without skills and no instruction text reaches it; compiler-generated skills and description are kept verbatim; the description / HITL / default-interface enrichment still applies. go vet, go test -run TestEnrichAgentCard ./adk/pkg/a2a/ and go test ./adk/pkg/app/ pass.

Backport

The same function exists unchanged on release/v0.10.x, where gateway-fronted callers hit the pod card directly; happy to open the backport once this lands.

Fixes #2549

---

Prepared with Claude Code assistance; the diff was reviewed and the tests above were run locally before opening the PR.

🤖 Generated with Claude Code

[supreme-gg-gg]

The a2a-gateway now serves the public Agent Card from the pinned runtime revision and invokes the Actor directly over gRPC without discovering its card. I think we should we remove runtime skill enrichment instead of maintaining a custom deriveAgentSkills implementation. It appears unnecessary on this path and loses upstream support for workflow/custom agents and tool-derived skills. The compiler-generated card should remain authoritative.

[supreme-gg-gg]

what's "tool listing"? I don't recall such a mechanism in A2A

[TigerkidYang]

You are right, A2A has no such mechanism. The wording was inaccurate, and that code is gone now (see below).

[TigerkidYang]

Thanks @supreme-gg-gg, agreed. With the gateway serving the card from the pinned revision, the compiler-generated card should stay authoritative, and a custom deriveAgentSkills is not worth maintaining. Pushed 1 commit that drops skill enrichment entirely: EnrichAgentCard no longer touches skills and only fills a missing description, the HITL extension and the default interface.

The pod still serves /.well-known/agent-card.json to direct-to-pod callers (the reporter's agentgateway setup, and 0.10.x), so this is what closes the leak there. Tests now assert that a card without skills stays without skills and that compiler-generated skills are kept verbatim. PR description updated to match.