fix: handle JSON contents of create_new_file

[mcowger]

Handle case where JSON content was parsed into an object by the tool call parser. If the arguments to the tool call are valid JSON (e.g. the model attempts to create a .json file) the earlier call to JSON.parse() will have deeply parsed the returned contents. If that has happened, convert back to string.

Fixes: #8972

Description

[ What changed? Feel free to be brief. ]

AI Code Review

• Team members only: AI review runs automatically when PR is opened or marked ready for review
• Team members can also trigger a review by commenting @continue-review

Checklist

• I've read the contributing guide
• The relevant docs, if any, have been updated or created
• The relevant tests, if any, have been updated or created

Screen recording or screenshot

NA

Tests

Added 6 tests to core/tools/parseArgs.vitest.ts for new check.

---

Summary by cubic

Fixes failures when create_new_file receives JSON content parsed into an object by ensuring contents is always a string. Prevents crashes when creating .json files. Fixes #8972.

• Bug Fixes
  • In getStringArg, JSON.stringify object/array values to safely handle parsed JSON.
  • Throw a clear error for non-string values after conversion.
  • Added tests covering contents object/array handling and edge cases.

^{Written for commit e04c21f. Summary will update automatically on new commits.}

[continue]

Keep this PR in a mergeable state →

Learn more

---

All Green is an AI agent that automatically:

✅ Addresses code review comments

✅ Fixes failing CI checks

✅ Resolves merge conflicts

[continue-development-app]

Keep this PR in a mergeable state →

Learn more

---

All Green is an AI agent that automatically:

✅ Addresses code review comments

✅ Fixes failing CI checks

✅ Resolves merge conflicts

[continue-development-app]

Keep this PR in a mergeable state →

Learn more

---

All Green is an AI agent that automatically:

✅ Addresses code review comments

✅ Fixes failing CI checks

✅ Resolves merge conflicts

[github-actions]

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

[RomneyDa]

@mcowger could we add this special handling in createFileImpl instead of here?

OR, I think you could just remove special handling for contents and always make sure it's a string within getStringArg!

[continue]

Reviewed the PR changes. This is an internal bug fix in the tool argument parser that handles JSON content correctly when creating JSON files. The changes don't require documentation updates since:

1. This fixes internal parsing logic, not user-facing behavior
2. The existing documentation correctly describes the create_new_file tool functionality
3. The fix is transparent to users - it restores expected behavior rather than changing the tool's API or capabilities

The fix ensures that when models create .json files with JSON content, the parser correctly handles the case where the content gets parsed into an object and converts it back to a string.

[mcowger]

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

[mcowger]

@mcowger could we add this special handling in createFileImpl instead of here?

OR, I think you could just remove special handling for contents and always make sure it's a string within getStringArg!

Good call. Removed the special case handling.

[cubic-dev-ai]

No issues found across 2 files

[cubic-dev-ai]

No issues found across 2 files

[RomneyDa]

@mcowger this looks good to me. I haven't been able to replicate the failing case locally - it seems like claude and openai models mostly escape JSON strings correctly. Could you share which model you find this error often happens with?

EDIT I see M2 in the issue but wasn't sure which model that was. Model + provider combo would be helpful

[mcowger]

I can replicate on Synthetic.new. It doesn't happen for me on OpenRouter either.

Happy to provide a key privately for you to test so you dont have to give them money.

[RomneyDa]

@mcowger which model are you working with?
EDIT this was a silly question, I see M2 on openrouter now
https://openrouter.ai/minimax/minimax-m2

[mcowger]

@mcowger which model are you working with?

Minimax M2 (note, I cannot replicate this on OpenRouter). I cannot replicate this via GLM4.6 on the same provider, so it appears to be an interaction between the inference and the model and continue.

[mcowger]

I captured traces of the differences in the output:

https://gist.github.com/mcowger/3aea3a2314d9f5540e7bbf68d9cafcd7

GLM outputs double escaped strings, while M2 does not.

That should be enough to replicate with?

[RomneyDa]

This seems like could be a valid issue for other models as well, and if the tool implementation is looking for a string arg specifically, like the create new file tool, this fix seems fine.

I guess you would still see issues with MCP tools when they expect string JSON and receive JSON object. The solution might be to detect "type": "string" in the tool definition and use getStringArg to replace at the time of calling the MCP tool. Do you use MCP tools at all? I could open a separate issue for that

[RomneyDa]

Got it, thanks for the traces. It's no clear why the model would escape differently between providers (as far as I understand this would only happen because of something on the provider side) but I could see this happening in other situations as well and this fix won't hurt.

[mcowger]

Agree its a provider side issue, and I've filed a bug with them.

I dont use MCPs much, but I agree it could impact them as well.

[RomneyDa]

This looks good for now, can address MCP on a separate issue if it comes up

[sestinj]

🎉 This PR is included in version 1.35.0 🎉

The release is available on:

• npm package (@latest dist-tag)
• GitHub release

Your semantic-release bot 📦🚀

[sestinj]

🎉 This PR is included in version 1.7.0 🎉

The release is available on:

• npm package (@latest dist-tag)
• GitHub release

Your semantic-release bot 📦🚀