[chughtapan]

Summary

Moves wags onto stock fast-agent-mcp 0.10.46, which hard-pins mcp 2.3.0 and fastmcp 4.0.11, so wags can run current models (for example --model opus, which resolves to claude-opus-5-5, plus gpt-5.6 and gemini-3.5-flash). This removes the chughtapan forks of fast-agent, python-sdk, and fastmcp.

Breaking: RootsMiddleware removed

The forks existed mainly to allow non-file:// root URIs such as https://github.com/{owner}/{repo}. Upstream still requires file:// roots, and roots are deprecated in the 2026-07-28 spec (SEP-2577). RootsMiddleware, requires_root, their tests, docs, and the roots: config blocks are deleted.

Protocol changes (2026-07-28)

fast-agent now negotiates the 2026-07-28 protocol by default, in which servers cannot send requests to clients.

• ElicitationMiddleware returns an InputRequiredResult (SEP-2322) and reads the answer from the client's retry. Handshake-era clients still use ctx.elicit(). A decline or cancel now raises ToolError, because fastmcp 4 hides other middleware exceptions behind -32603 Internal server error.
• GroupsMiddleware sends tools/list_changed with ctx.send_notification, so the notification travels on the request stream. The previous session.send_tool_list_changed() was silently dropped for 2026-07-28 clients. This PR adds a test that fails on the old code.

Ports

• create_proxy now uses fastmcp.server.create_proxy. The roots-notification proxy subclass is removed.
• wags run is rewritten on the public FastAgent API, because run_async_agent no longer exists.
• mcp 2 renames: input_schema, is_error, max_tokens, and ClientRequestContext. This fixes a silent getattr(result, "isError", False) in the MCP-Universe benchmark that would have treated every tool error as a success.
• MessageSerializer.strip_server_prefix now calls fast-agent's get_resource_name. fast-agent names tools server__tool, but the old code split on the first -, so BFCL scored every call under its namespaced name (e.g. gorillafilesystem__cd) and failed every task.
• The AppWorld benchmark server moves to the mcp 2 low-level Server(on_list_tools=..., on_call_tool=...) API, and tool errors now set is_error=True. The BFCL server moves from FastMCP to MCPServer.

Docs

Examples now use current model names (gpt-6.1-sol, claude-opus-5-5, claude-sonnet-5-5), and the documented default model is corrected to gpt-4o-mini. A note explains that pytest's -p flag collides with fast-agent's --prompt-file, so plugin flags go in PYTEST_ADDOPTS.

Dependencies

• AppWorld: pinned to chughtapan/appworld@wags-dev, which is f31e671 plus a 2-line fix for pydantic 2.14. Without it, import appworld fails under fast-agent's pydantic==2.14.0 pin. Upstream fix: Forward pydantic>=2.14's origin argument in match_type patch StonyBrookNLP/appworld#235.
• Eval overrides: uv overrides now set the eval packages' conflicting exact pins to fast-agent's versions: mcp, openai, anthropic, google-genai, pydantic, pyyaml, tiktoken, filelock, anyio, click, and google-auth.

Test plan

• pre-commit (ruff, ruff-format) and mypy src/ tests/
• pytest tests/unit/ (91 passed) and pytest tests/integration/ (33 passed)
• Elicitation integration tests run on both the 2026-07-28 and handshake protocols
• Fast-agent 0.10.46 to wags proxy over stdio using the passthrough model:
  • group gating, then enable_tools, then tool refresh
  • proxied tool call
  • elicitation InputRequiredResult round-trip
• wags run interactive session connects and lists server tools
• BFCL server: lists tools and runs calls over stdio
• AppWorld server wiring checked over stdio with stubbed data loading
• AppWorld from the fork: LFS bundles install, appworld install succeeds, and the benchmark collects 90 tasks
• MCP-Universe imports load, and its benchmark plus BFCL's collect
• Live BFCL run with gpt-6.1-sol: multi_turn_base_0 and _1 pass. _2 fails because the model's calls don't match the ground truth; tool names serialize correctly.
• Live groups e2e run with gpt-6.1-sol. 1 of 6 passes; the others fail on the model's behavior within a single turn. With gpt-5, main fails all 6 the same way, so this is not a regression.
• Live Claude Opus 5.5 run (no Anthropic key available)
• Full AppWorld task run (the local data/ is missing base DBs)

🤖 Generated with Claude Code