local-tool-reliability

name	local-tool-reliability
description	Use when local Copilot/Qwen tool calls fail, malformed tool JSON appears, tools are skipped, or LLM Gateway agent mode is flaky.
argument-hint	[tool-call symptom or error message]

Local Tool Reliability

Debug and stabilize local-model tool calling before changing application code.

Procedure

1. Separate server health from agent-tool health.
   • Confirm llama-server is running and /v1/models responds.
   • Send a tiny direct chat-completions request to confirm basic generation.
   • If the server is healthy but Copilot fails, suspect gateway, prompt size, or tool-set bloat next.
2. Check local-model gateway settings.
   • github.copilot.llm-gateway.agentTemperature: 0.0 for tool mode.
   • Start with parallelToolCalling: false for Qwen.
   • Keep enableToolCalling: true unless isolating a basic chat failure.
   • Use a long enough requestTimeout for large local prompts.
3. Check Qwen llama-server launch settings.
   • Use Jinja chat templates (--jinja).
   • Preserve Qwen thinking when needed for multi-turn tool use (preserve_thinking: true).
   • Use the qwen3_coder tool parser where available.
   • Keep KV cache at q8_0 when reliability matters.
4. Trim the tool surface.
   • Disable MCP servers and extension tools the task does not need.
   • Prefer one well-named tool over overlapping tools.
   • Make argument names and types unambiguous.
   • Use absolute paths.
   • Prefer plain-text or search-replace edit formats over giant JSON payloads.
5. Reduce the prompt before blaming the parser.
   • Start a fresh chat and retry with a minimal reproduction.

Output Contract

Return:

• Whether the failure is server, gateway, parser, tool-set, or context-size.
• The smallest reliability setting change to try.
• A validation step.
• Whether the change trades speed for correctness.

References

• Evidence and rationale