Gate browser tool requests on server support

[neubig]

Summary

• cache /server_info tool metadata during the compatibility probe and gate browser-tool requests on usable_tools
• only request browser_tool_set when the env flag allows it and the connected server advertises browser support
• update mocks and tests for the new browser gating behavior, then merge the latest main

Verification

• npm test
• npm run typecheck
• npm run build

Depends on OpenHands/software-agent-sdk#3028 for the /server_info.usable_tools metadata. When the server does not advertise tool metadata, the frontend keeps the previous allow-all behavior.

This PR was created by an AI agent (OpenHands) on behalf of the user.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
agent-server-gui	Ready	Preview, Comment	May 1, 2026 7:30pm

[all-hands-bot]

🟢 Good taste - Clean feature detection pattern with proper backward compatibility.

[IMPROVEMENT OPPORTUNITIES] (Optional)

Consider adding a test case for backward compatibility when available_tools is missing:

it("allows all tools when server does not provide available_tools metadata", async () => {
  server.use(
    http.get("/server_info", () =>
      HttpResponse.json({
        uptime: 0,
        idle_time: 0,
        version: "1.17.0",
        // No available_tools field
      }),
    ),
  );

  await OptionService.getConfig();

  expect(isAgentServerToolAvailable("browser_tool_set")).toBe(true);
  expect(isAgentServerToolAvailable("terminal")).toBe(true);
});

This would explicitly verify the fallback behavior for older servers that don't advertise capabilities.

[RISK ASSESSMENT]

⚠️ Risk Assessment: 🟢 LOW

This is a clean infrastructure improvement that adds proper server capability detection. The change:

• Only gates tools based on actual server support
• Maintains full backward compatibility with older servers via sensible fallback (allow all tools when metadata is missing)
• Has good test coverage for the new gating behavior
• Makes no changes to agent behavior, prompts, or evaluation-sensitive code

The module-level cache pattern is appropriate for a frontend that connects to a single backend per session. Cache invalidation happens correctly on errors.

VERDICT:

✅ Worth merging: Solid implementation with proper fallback behavior. The optional test suggestion above would strengthen the backward compatibility coverage, but existing tests are sufficient.

KEY INSIGHT:

Elegant additive feature detection that degrades gracefully for older servers - new servers get fine-grained tool gating, old servers preserve existing env-flag behavior.

[neubig]

@OpenHands QA to confirm that this PR works with OpenHands/software-agent-sdk#3028 in the case that the browser is available or unavailable.

[openhands-ai]

I'm on it! neubig can track my progress at all-hands.dev

[neubig]

QA complete.

I found one integration mismatch while checking against OpenHands/software-agent-sdk#3028: that PR currently exposes /server_info.usable_tools, while this frontend branch was only reading available_tools. I fixed that in b548f3b by preferring usable_tools, falling back to available_tools, and preserving the old "allow all tools" fallback when neither field is present.

What I verified:

• Browser advertised: against a live checkout of software-agent-sdk PR #3028 (273e911), /server_info.usable_tools included browser_tool_set, isAgentServerToolAvailable("browser_tool_set") returned true, and the frontend kept browser_tool_set in the conversation start payload.
• Browser withheld: I ran the same SDK checkout with the /server_info.usable_tools response patched to omit browser_tool_set (this sandbox has /usr/bin/chromium, so I had to emulate the no-browser environment). In that case, isAgentServerToolAvailable("browser_tool_set") returned false, and the frontend omitted browser_tool_set from the conversation start payload.
• Regression coverage: added tests for usable_tools, legacy available_tools, and the no-metadata fallback.
• Repo verification: npm test, npm run typecheck, and npm run build all pass after the fix.

AI-generated PR update/comment by OpenHands on behalf of the user.

[openhands-ai]

Final summary:

Completion status

• The PR comment request was fully addressed.
• QA was completed for both cases requested:
  • browser support advertised by the SDK server
  • browser support not advertised by the SDK server
• The integration mismatch discovered during QA was fixed on this branch.
• The fix was committed and pushed (b548f3b).
• The PR description was updated to match the actual SDK contract.
• A QA results comment was posted on PR Gate browser tool requests on server support #64.

New changes since my last summary

• No additional code changes were made after the previous update.
• I only performed final cleanup/confirmation:
  • stopped the temporary QA servers used for live verification
  • confirmed the branch is clean and synced with origin

Conciseness check

• The final code changes remain focused and minimal:
  • compatibility handling for usable_tools with available_tools fallback
  • targeted tests for the metadata cases
  • mock metadata alignment
  • one AGENTS.md memory note
• No extraneous working-tree changes remain.

[neubig]

Follow-up update: per review, I removed the frontend fallback to available_tools and merged the latest main before pushing.

Current behavior on this branch:

• use /server_info.usable_tools when present
• omit browser_tool_set if it is not advertised there
• keep the existing allow-all behavior only when the server does not advertise any tool metadata

I re-ran:

• npm test
• npm run typecheck
• npm run build

AI-generated PR update/comment by OpenHands on behalf of the user.

[neubig]

This still needs live testing.