Filesystem tools fail validation on weaker/custom models due to path/file_path parameter mismatch

Describe the bug

When using weaker or custom LLMs (e.g., smaller open-source models or custom endpoints) in deepagents and consumers like openwiki, the agent immediately crashes with a schema validation error on the first read_file call:

Root Cause

There are two contributing factors to this issue:

1. Schema Inconsistency: The ls tool defines its directory argument as path, while read_file, write_file, and edit_file define their target file argument as file_path. Models often generalize tool naming conventions and default to using path across all filesystem tools.
2. Confusing Prompt Examples: Both READ_FILE_TOOL_DESCRIPTION (in fs.ts) and the skills system prompt (in skills.ts) show documentation examples instructing the model to use path, e.g. read_file(path, limit=100). Less capable models strictly copy this behavior and fail the actual Zod schema validation because it expects file_path.

Proposed Solution

We should implement both a prompt correction and a defensive input normalization layer so that filesystem tools work reliably across all model tiers.

1. Prompt Corrections: Align the documentation examples in fs.ts and skills.ts to use file_path (as initiated in PR fix(deepagents): use file_path in read_file prompt examples #644).
2. Schema Input Normalization: Add a z.preprocess helper to read_file, write_file, and edit_file schemas to dynamically map path to file_path if the model passes path.

Implementation Example:

function normalizeFilePathInput(input: unknown): unknown {
  if (
    typeof input === "object" &&
    input !== null &&
    "path" in input &&
    !("file_path" in input)
  ) {
    const { path, ...rest } = input as Record<string, unknown>;
    return { ...rest, file_path: path };
  }
  return input;
}

// Wrap schemas
schema: z.preprocess(normalizeFilePathInput, z.object({
  file_path: z.string().describe("Absolute path to the file to read"),
  // ...
}))