# Claude Code: Hard Overrides to Stabilize Tool Use

When Claude Code's layered system prompts clash with our tool protocol, force a minimal, deterministic lane.

## Settings (recommended)
- **Stop sequences**: `
````, `

` (or whatever the IDE uses to inject formatting).
- **Max output tokens**: small (e.g., 512).
- **Temperature**: 0.1-0.3.
- **Disable auto-formatting / code fences** if possible.
- **Disable auto-tool use**; we trigger tools via explicit JSON only.

## System message (short)
Use the contents of `docs/prompts/qwen_tool_mode_system.txt` verbatim as the *final* system layer closest to the model.

## Runtime guardrails
- Reject any non-JSON output; send a short corrective user message: `OUTPUT MUST BE JSON. Please resend.`
- If repeated, send `{"final":{"content":{"error":"RESET_REQUIRED"}}}` back and restart the session.

## Registry injection
- Provide the tool list and JSON Schemas (kom.memory.*, kom.local.backup.*, acf.*).
- Keep it short; link to full schemas if the UI allows references.

## Troubleshooting
- If model keeps adding prose, tighten stop sequences and lower max tokens.
- If JSON keys drift, include a 2–3 line example of a **valid** `action` and a **valid** `final`.
- If it calls undefined tools, respond with a single tool error and re-present the allowlist.

## Fallback DSL
- Accept `@tool <name> {json-args}` and convert to a JSON `action` behind the scenes when necessary.