fix: elicitation scalar return, resource auto-serialization, Client.new() state, prompt errors

[strawgate]

Four client bugs found during a systematic audit of the FastMCP client API surface.

Elicitation scalar return: Handlers returning T directly for ctx.elicit("msg", str) crashed with "must be serializable as a JSON object." The type signature promises this works — now it does by auto-wrapping scalars into {"value": scalar} for ScalarElicitationType schemas.

async def handler(msg, rtype, params, ctx):
    return "Alice"  # Before: crash. After: works.

Resource auto-serialization: Resources returning dict, int, float, bool, or None crashed with TypeError despite the docstring promising auto-conversion. Now auto-serialized to JSON text.

@mcp.resource("data://info")
def info(): return {"key": "value"}  # Before: TypeError. After: JSON text.

Client.new() shared state: copy.copy() shared _task_registry and _submitted_task_ids between original and cloned clients, violating the "fresh session state" contract.

Prompt error message: function_prompt.py dropped the original exception message, unlike tool errors which preserve it.

Fixes #3856

🤖 Generated with Claude Code

[strawgate]

Client.new() shared state: copy.copy() shared _task_registry and _submitted_task_ids between original and cloned clients, violating the "fresh session state" contract.

@chrisguidry was this intentional?

[marvin-context-protocol]

Test Failure Analysis

(Updated to reflect the latest workflow run — 24323832944. Previous analysis about a Windows test timeout is no longer relevant.)

Summary: The Run static analysis workflow failed with two prek hook violations: an unused import caught by ruff-check, and a type-checker error caught by ty.

Root Cause:

1. Unused import (ruff-check): The Client.new() refactor in this PR removed the if not isinstance(self.transport, StdioTransport): guard, but left StdioTransport in the import list of src/fastmcp/client/client.py. Ruff auto-fixed it (removed the import), causing the hook to fail.

2. Type error (ty check): The new test test_client_new_resets_mutable_task_state seeds _task_registry with a lambda to verify isolation:

   client._task_registry["task-1"] = lambda: None  # type: ignore[assignment]

   But ty reports this as error[invalid-assignment], not error[assignment]. The # type: ignore[assignment] suppression doesn't match ty's error code. Looking at other patterns in the PR diff (e.g. # type: ignore[call-arg] # ty:ignore[unknown-argument]), the project uses a dual-comment style.

Suggested Solution:

1. src/fastmcp/client/client.py — Remove StdioTransport from the import block (line ~73):

   # Remove this line:
   StdioTransport,

2. tests/client/client/test_client.py — Add a ty:ignore suppression (line 777):

   # Before:
   client._task_registry["task-1"] = lambda: None  # type: ignore[assignment]
   # After:
   client._task_registry["task-1"] = lambda: None  # type: ignore[assignment]  # ty:ignore[invalid-assignment]

Detailed Analysis

Failure 1 — ruff-check

ruff check...................................................Failed
- hook id: ruff-check
  Found 1 error (1 fixed, 0 remaining).

Ruff's auto-fix diff:

--- a/src/fastmcp/client/client.py
+++ b/src/fastmcp/client/client.py
@@ -70,7 +70,6 @@ from .transports import (
     PythonStdioTransport,
     SessionKwargs,
     SSETransport,
-    StdioTransport,
     StreamableHttpTransport,
     infer_transport,
 )

The import was needed before this PR to satisfy the isinstance check in Client.new():

if not isinstance(self.transport, StdioTransport):
    new_client._session_state = ClientSessionState()

That condition was removed by the PR (always reset now), but the import was not cleaned up.

Failure 2 — ty check

ty check.....................................................Failed
- hook id: ty
  error[invalid-assignment]: Invalid subscript assignment with key of type `Literal["task-1"]` and value of type `() -> None` on object of type `dict[str, ReferenceType[ToolTask | PromptTask | ResourceTask]]`
  777 |     client._task_registry["task-1"] = lambda: None  # type: ignore[assignment]

_task_registry is typed as dict[str, weakref.ref[ToolTask | PromptTask | ResourceTask]]. Assigning a lambda is intentional in the test (just seeding state), but ty uses invalid-assignment as the error code while only assignment is suppressed. Other lines in this PR already use the dual-comment pattern (e.g. line 59: # type: ignore[call-arg] # ty:ignore[unknown-argument]).

Related Files

• src/fastmcp/client/client.py:73 — Unused StdioTransport import
• src/fastmcp/client/client.py:433-457 — Client.new() method that was refactored (the StdioTransport isinstance check was removed here)
• tests/client/client/test_client.py:777 — Test with # type: ignore[assignment] that needs a # ty:ignore[invalid-assignment] annotation

[chrisguidry]

Client.new() shared state: copy.copy() shared _task_registry and _submitted_task_ids between original and cloned clients, violating the "fresh session state" contract.

@chrisguidry was this intentional?

No, that does not sound intentional to me, I'd think we'd want each client to start fresh. If the server supports task listing then a client can ask for them, so we don't need to copy these between instances.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1759276a4b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[jlowin]

One more thing before merge — the Codex P2 about list[ResourceContent] is still open and I think it's real. The new JSON-native branch in convert_result catches pure list[ResourceContent] returns before the fallthrough to ResourceResult(raw_value), so json.dumps blows up on the ResourceContent items.

Minimal fix is to short-circuit list-of-ResourceContent ahead of the JSON path:

# list[ResourceContent] passes through — otherwise JSON-serialize below
if isinstance(raw_value, list) and all(
    isinstance(item, ResourceContent) for item in raw_value
):
    return ResourceResult(raw_value)

Happy to push it myself if easier — but flagging so it doesn't slip through.

[jlowin]

Thanks!