Problem (one or two sentences)

apply_diff stays in the advertised tool list and is sent to the model even when diff tools are disabled (diffEnabled=false). The model keeps calling it and hits a validation error (Tool "apply_diff" is not allowed …), blocking progress.

Context (who is affected and when)

• Any provider using native tools (OpenAI/OpenAI-compatible and others) when the user has turned off diffs in settings/profile.
• Reproducible in code-editing modes; switching modes does not help.

Reproduction steps

1. Set diffEnabled=false (profile/state).
2. Use a native-tools-capable provider; start a task in code (or similar) mode.
3. Observe: the tools array sent to the model still includes apply_diff.
4. The model calls apply_diff; extension validation rejects it with “Tool "apply_diff" is not allowed in code mode.”
5. Model retries the same tool call (reasoning shows the ban), causing a loop/no edits.

Expected result

apply_diff is omitted from advertised tools and not sent to the API when diffs are disabled; the model won’t call it.

Actual result

apply_diff is sent to the model even with diffs disabled; calls fail at validation, and the agent stalls/retries.

Variations tried (optional)

• Switching between editing modes — tool still advertised.
• Native tool protocol only; XML catalog correctly drops the description when diff is off, but native still sends the tool definition.

App Version

3.36.2

API Provider (optional)

OpenAI Compatible

Model Used (optional)

N/A

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)