fix: coerce MCP tool args to match schema string types

[RomneyDa]

Summary

• Fixes MCP tools receiving parsed JSON objects instead of strings when the tool's input schema defines a parameter as type: "string" (e.g. write_file content parameter)
• Adds coerceArgsToSchema() that checks the tool's inputSchema and re-stringifies any values that were deeply parsed by JSON.parse() but should remain as strings
• Built-in tools already handled this via getStringArg(), but MCP tools with dynamic schemas had no equivalent fix

Test plan

• Added unit tests for coerceArgsToSchema() covering: object→string coercion, array→string coercion, no-op for already-string values, no-op for non-string schema types, no schema, mixed args, immutability
• All 49 tests in parseArgs.vitest.ts pass

Fixes #8510

---

Summary by cubic

Coerces MCP tool args to match string-typed schema params so tools don’t receive parsed objects/arrays. Applies before calling the MCP client to fix cases like write_file where JSON content must remain a string.

• Bug Fixes
  • Added coerceArgsToSchema() that re-stringifies only object/array values when the schema says type: "string", and used it before MCP callTool.
  • Added tests for object/array-to-string, skipping numbers/booleans, mixed args, and immutability.

^{Written for commit 42d2037. Summary will update on new commits.}

[continue]

Documentation Review

No documentation updates needed for this PR.

Reasoning: This PR is an internal bug fix that transparently handles edge cases in MCP tool argument parsing. The coerceArgsToSchema() function automatically re-stringifies values when the tool's schema expects type: "string" but receives a parsed JSON object (e.g., JSON file content passed to write_file).

This fix:

• Requires no user action or configuration changes
• Is applied automatically to all MCP tool calls
• Doesn't change how users configure or interact with MCP servers

The existing MCP documentation covers configuration, setup, and transport types—documenting internal argument coercion mechanics would add unnecessary implementation detail that doesn't help developers use Continue more effectively.

[cubic-dev-ai]

1 issue found across 3 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="core/tools/parseArgs.ts">

<violation number="1" location="core/tools/parseArgs.ts:51">
P2: This coercion is too broad for `type: "string"`: it also converts number/boolean inputs to strings, which can mask invalid tool arguments instead of only fixing deep-parsed object/array values.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}