fix(tool-call): convert INVALID_PARAMS to soft error for Devin/Windsu…

[altasol]

fix(tool-call): convert INVALID_PARAMS to soft error for Devin/Windsurf compatibility

Summary

Some MCP clients (Devin/Windsurf) treat hard JSON-RPC -32602 errors as transport failures, disconnect and respawn the server, hiding the actual parameter validation message from the agent. Convert INVALID_PARAMS to CallToolResult with isError=true so agents see validation messages and can fix parameter names.

Test plan

• cd rust && cargo test -- --test-threads=1 (the suite shares process-global state; CI serializes it too)
• cd rust && cargo clippy --all-targets --all-features -- -D warnings
• cd rust && cargo fmt --check
• If cookbook/packages changed: relevant npm test / build steps

Notes for reviewers

Fixes Devin MCP connection reset issue on Devin local mode.

Failed to call ctx_patch from lean-ctx
Failed to connect to MCP server 'lean-ctx'. Please try again.

thus model stops using ctx_* because of continues failures and not reliable tools.
Resulting soft error, allows model to get response and understand issue, and MCP connection stays in tact.

Contributor License Agreement

First-time contributors: a bot will ask you to sign our one-time
CLA (it keeps lean-ctx
Apache-2.0 and free for individual developers — see §8). You sign once by
replying to this PR with: I have read the CLA Document and I hereby sign the CLA

I have read the CLA Document and I hereby sign the CLA

[github-actions]

All contributors have signed the CLA. ✅
_{Posted by the CLA Assistant Lite bot.}

[altasol]

Relates to and resolves #1409 also, reduces mcp connection restarts and contention.

[altasol]

I have read the CLA Document and I hereby sign the CLA

[altasol]

lean-ctx MCP: Validation Errors Misreported as Connection Failures

Summary

Devin/Windsurf's MCP client misreports JSON-RPC -32602 (INVALID_PARAMS) error
responses from the lean-ctx MCP server as "Failed to connect to MCP server",
making parameter validation errors look like infrastructure/transport failures.

This causes agents to retry, respawn MCP processes, and waste turns debugging
connectivity instead of fixing the parameter names.

Encountered Errors Example

ctx_patch with wrong params (anchor/content instead of line/hash/new_text)

Called:

{
  "name": "ctx_patch",
  "arguments": {
    "op": "insert_after",
    "path": "/path/to/file.md",
    "anchor": "some text",
    "content": "new text"
  }
}

Server response (JSON-RPC error):

{
  "jsonrpc": "2.0",
  "id": 2,
  "error": {
    "code": -32602,
    "message": "missing 'line'"
  }
}

Devin MCP client log:

[warning] MCP operation failed on cached service for 'lean-ctx',
  retrying with fresh connection: Mcp error: -32602: missing 'line'
[info] Connecting to MCP server 'lean-ctx'
[info] Starting stdio MCP server 'lean-ctx': "/opt/homebrew/bin/lean-ctx" []
[info] MCP server 'lean-ctx' connected successfully
[warning] Failed to connect to 'lean-ctx' for tool call: Mcp error: -32602: missing 'line'

Agent sees: Failed to connect to MCP server 'lean-ctx'. Please try again.

What should have happened: Agent sees: missing 'line' — insert_after requires line (integer) + hash (string) + new_text (string). Use ctx_read(mode="anchored") to get line:hash anchors.

and similar, when model is not aware how to exactly call the tools.

[altasol]

real life test:

now after error:

MCP tool 'ctx_patch' returned an error: [ { "type": "text", "text": "replace_unique requires non-empty old_text (old_string also accepted)" } ] Available tools on server `lean-ctx`: [ {…

Model corrects it self:
Let me use the correct parameter names for ctx_patch.

[yvgude]

Merged to main — thanks @altasol! 🎉

The fix is clean and the real-world Devin testing confirms the improvement. I trimmed the comment block slightly for consistency with our code style, but the logic is unchanged.

This will ship in the next release.