pi-codemcp 1.3.2 → 1.3.3

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-codemcp",
-  "version": "1.3.2",
+  "version": "1.3.3",
   "description": "Typed, sandboxed Code Mode access to configured MCP servers for Pi",
   "type": "module",
   "packageManager": "bun@1.3.10",

package/sidecar/executor.py

@@ -50,9 +50,10 @@ INSPECT_BYTE_LIMIT = 8 * 1024
 CHAIN_INPUT_EXTERNAL = "__codemcp_saved_chain_input"
 
 
-type FailureStage = Literal["preflight", "runtime", "timeout", "cancelled", "result"]
+type FailureStage = Literal["preflight", "arguments", "runtime", "timeout", "cancelled", "result"]
 type FailureKind = Literal[
     "preflight",
+    "argument_validation",
     "result",
     "result_reference",
     "sandbox_runtime",
@@ -487,7 +488,18 @@ class MontyExecutor:
                 raise RuntimeError(
                     f"Call limit exceeded: maximum {context.settings.max_calls} total calls"
                 )
-            validated = catalog.validate_arguments(name, arguments)
+            try:
+                validated = catalog.validate_arguments(name, arguments)
+            except (TypeError, ValidationError, ValueError) as error:
+                message = f"{spec.call}: invalid arguments: {error}"
+                context.failure = ExecutionFailureInfo(
+                    kind="argument_validation",
+                    server=spec.server,
+                    tool=spec.backend_name,
+                    retryable=False,
+                    message=message,
+                )
+                raise ValueError(message) from error
             if spec.kind == "saved_chain":
                 context.chain_calls += 1
                 return await context.call_tool(name, validated, context)
@@ -624,8 +636,10 @@ class MontyExecutor:
             context.metrics.runtime_ms += _elapsed_ms(runtime_started)
             message = error.display("type-msg").strip()
             lowered = message.lower()
-            stage: Literal["runtime", "timeout"] = (
-                "timeout"
+            stage: Literal["arguments", "runtime", "timeout"] = (
+                "arguments"
+                if context.failure is not None and context.failure.kind == "argument_validation"
+                else "timeout"
                 if (context.failure is not None and context.failure.kind == "upstream_timeout")
                 or "duration" in lowered
                 or "timed out" in lowered
@@ -728,7 +742,7 @@ class MontyExecutor:
     @staticmethod
     def _failure(
         context: ExecutionContext,
-        stage: Literal["preflight", "runtime", "timeout", "cancelled", "result"],
+        stage: FailureStage,
         error: str,
     ) -> ExecutionResponse:
         failure = context.failure or _execution_failure_info(stage, error)
@@ -749,6 +763,8 @@ def _execution_failure_info(
 ) -> ExecutionFailureInfo:
     if stage == "preflight":
         kind: FailureKind = "preflight"
+    elif stage == "arguments":
+        kind = "argument_validation"
     elif stage == "result":
         kind = "result"
     elif stage == "timeout":

package/sidecar/gateway.py

@@ -1511,6 +1511,7 @@ def _execution_failure_subtype(response: ExecutionResponse) -> str:
     failure = response.failure
     if failure is not None and failure.kind in {
         "result_reference",
+        "argument_validation",
         "sandbox_timeout",
         "upstream",
         "upstream_transport",

package/sidecar/sandbox_api.py

@@ -14,7 +14,7 @@ SANDBOX_FUNCTION_EXTERNALS = {
     EXPECT_INTEGER_NAME: "__codemcp_expect_integer",
 }
 
-STUB_IMPORTS = "from typing import Literal, Never, NotRequired, TypeAlias, TypedDict"
+STUB_IMPORTS = "from typing import Literal, Mapping, Never, NotRequired, TypeAlias, TypedDict"
 JSON_TYPE_STUBS = (
     "JsonScalar: TypeAlias = bool | int | float | str | None",
     'JsonValue: TypeAlias = JsonScalar | list["JsonValue"] | dict[str, "JsonValue"]',

package/sidecar/stats.py

@@ -98,6 +98,7 @@ DISTINCT_NAME_QUERIES = {
 FailureOutcome = Literal[
     "success",
     "preflight_rejection",
+    "argument_rejection",
     "result_refinement",
     "upstream_failure",
     "transport_failure",
@@ -996,6 +997,8 @@ def _operation_outcome(
         return "internal_error"
     if failure.stage == "preflight":
         return "preflight_rejection"
+    if failure.stage == "arguments":
+        return "argument_rejection"
     if failure.stage == "result":
         return "result_refinement"
     if failure.stage == "cancelled":

package/sidecar/tool_catalog.py

@@ -426,7 +426,8 @@ class ToolCatalog(BaseModel):
         for spec in specs:
             definitions.extend(spec.stub.split("\n\n") if spec.stub else [])
             facade_methods.setdefault(spec.namespace, []).append(
-                f"    async def {spec.method}(self, arguments: {spec.input_type_name}) "
+                f"    async def {spec.method}(self, arguments: "
+                f"{spec.input_type_name} | Mapping[str, JsonValue]) "
                 f"-> {spec.output_type_name}: ..."
             )
         facades: list[str] = []

package/src/execution-rendering.ts

@@ -114,6 +114,9 @@ function renderCompactFailure(
   if (stage === "preflight") {
     return theme.fg("warning", `✗ Preflight · code not run · ${summary}`);
   }
+  if (stage === "arguments") {
+    return theme.fg("warning", `✗ Argument validation · code run · ${summary}`);
+  }
   if (stage === "timeout") {
     return theme.fg("error", `✗ Timeout · stopped after ${summary}`);
   }
@@ -131,6 +134,9 @@ function failureHeading(stage: string, calls: number, chainCalls: number, theme:
   if (stage === "preflight") {
     return `${theme.fg("warning", theme.bold("Preflight failed"))}\n${theme.fg("muted", "Code was not executed; no upstream side effects")}`;
   }
+  if (stage === "arguments") {
+    return `${theme.fg("warning", theme.bold("Argument validation failed"))}\n${theme.fg("muted", "Code ran; no invalid MCP call was sent")}`;
+  }
   if (stage === "timeout") {
     return `${theme.fg("error", theme.bold("Execution timed out"))}\n${theme.fg("muted", `Stopped after ${summary}`)}`;
   }