Step 6: baseline browser-task generalist

[rockfordlhotka]

Summary

First v0.2 milestone (spec §9.1 step 6). Adds browser-task, the LLM-in-the-loop generalist capability built natively on Microsoft.Playwright (no MCP sidecar, no Stagehand port — Appendix A #16). Uses ref-annotated aria snapshots + aria-ref=eN locator resolution.

• New capability src/Foragent.Capabilities/BrowserTask/ with 7 [AIFunction] tools: snapshot, navigate, click, type, wait_for, done, fail. Budget enforced tool-side (default maxSteps=60, ceiling 150) + wall-clock CancellationTokenSource (default 120s, ceiling 600). Output is structured JSON {status, summary, result, steps, navigations}.
• Foragent.Browser extended with IBrowserAgentPage for ref-based interactions, plus an allowlist-scoped CreateSessionAsync(Func<Uri,bool>, ...) overload that aborts off-list document/subframe navigations inside the context's RouteAsync — before Playwright issues the request (spec §7.1).
• HostAllowlist parses bsky.app, *.example.com, and * patterns; empty lists reject.
• Program.cs now wires AddRockBotTieredChatClients (one model aliased across Low/Balanced/High per spec §3.7, Appendix #17); AddRockBotChatClient call removed (the tiered registration already wraps with RockBotFunctionInvokingChatClient).
• Microsoft.Playwright bumped 1.50 → 1.59 for AriaSnapshotMode.Ai (ref-annotated snapshots). Matching base-image bump deferred to release.
• browser-task added to ForagentCapabilities.Skills first; existing three specialists remain (spec removes post-to-site at step 7, not 6). deploy/rockbot-seed/ updated.

Verification

• dotnet build -c Release clean; dotnet test -c Release → 65 pass, 4 skipped (skipped are real-LLM integration benchmarks when FORAGENT_LLM_* is unset).

• Real-LLM benchmark (step 6 floor measurement, spec §9.1): 3 / 3 scenarios pass on first attempt against Azure AI Foundry Balanced.

  Scenario	Result	Wall-clock
  Click-through	✅ done	5 s
  Form submit	✅ done	8 s
  Multi-page nav	✅ done	7 s

• HTTP gateway smoke test via docker compose up:

  • /.well-known/agent-card.json advertises all four skills.
  • POST / with skill: browser-task / url: https://example.com / allowedHosts: ["example.com"] returns {"status":"done","result":"Example Domain","steps":1}.
  • Off-allowlist URL (https://evil.example/) is rejected: "Starting URL host 'evil.example' is not in the allowlist."

Framework observations surfaced (not worked around — see docs/framework-feedback.md)

1. AddRockBotTieredChatClients subsumes AddRockBotChatClient but this is undocumented; calling both can silently swap the wrapped registration.
2. No per-request MaximumIterationsPerRequest path through ChatOptions — Foragent enforces step budget tool-side via BrowserTaskState.BudgetExhausted.
3. Microsoft.Playwright C# bindings gate ref-annotated aria snapshots behind Mode = AriaSnapshotMode.Ai rather than a boolean Ref option.

Test plan

• dotnet build -c Release
• dotnet test -c Release
• Real-LLM benchmark scenarios (3 / 3 passing, floor established)
• docker compose up --build; curl smoke test via HTTP gateway
• Allowlist rejection verified end-to-end

🤖 Generated with Claude Code