القائمة
Tune the agent harness system prompts to improve model tool-calling scores. Triggers on "tune harness", "improve scores", "train harness", "optimize prompts".
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
| name | harness-tune |
| description | Tune the agent harness system prompts to improve model tool-calling scores. Triggers on "tune harness", "improve scores", "train harness", "optimize prompts". |
| type | skill |
Iteratively improve src/main/engine/agent-harness.ts system prompts to push
model scores higher. Uses the evaluate → mutate → re-evaluate → keep/revert loop.
src/main/engine/agent-harness.ts — contains three prompt builders:
buildNativeSystemPrompt() — for models with native OpenAI tool calling (Devstral, Qwen, Llama, Gemma 4)buildSmallModelPrompt() — for sub-2B models (compact prompt, low temperature)buildXMLSystemPrompt() — for models that write XML tool calls (Gemma 3)ai/data/trainer/harness_evolution.jsonlnpx tsx e2e/train-harness.ts --model "<model-id>"build*Prompt() functionnpx electron-vite buildharness_evolution.jsonl| Symptom | Root Cause | Fix |
|---|---|---|
| Tools not called | Model ignores tool instructions | Stronger "you MUST use tools" language |
| Same tool called 5x | Model repeats instead of proceeding | Add "NEVER repeat a tool call" |
| Empty response after tools | Model generates tool calls but no text | Add "you MUST write a response" |
| Wrong data in response | Model hallucinated instead of reading result | Add "include SPECIFIC data from results" |
| XML in response text | Non-native model writes XML tool calls | Check if model should be on XML path |
| 90s timeout | Model generates 600+ tool calls | Verify MAX_TOOL_CALLS_PER_TURN stream break works |
In agent-runner.ts:
MAX_TOOL_ITERATIONS = 5 — max loop iterations per user messageMAX_TOOL_CALLS_PER_TURN = 3 — cap per model response (stream breaks at this)Before committing prompt changes, re-test Devstral and Qwen 3 4B to ensure they still hit 100%. Regression on top models is worse than marginal gains on weak ones.
Orchestrate LiteBench model testing — scan LM Studio for models, run the agent harness against each, collect scores, swap models, and produce a leaderboard. Triggers on "test models", "run benchmark", "run harness", "test all models", "benchmark models", "start testing".
Autonomous agent and skill trainer — transforms the current Claude session into a self-improving loop inspired by Karpathy's autoresearch. Activates on: "train", "evolve agent", "improve agent", "run training loop", "agent trainer", "evolve this agent", "train this agent", "self-improving agent", "auto-improve", "training loop", "agent evolution", "autotrainer", "improve skill", "evolve skill", "optimize skill", "skill training", "improve my skills". Targets agents (.claude/agents/) OR skills (.claude/skills/) — auto-detects which. For agents: spawns A/B variants, evaluates output quality, mutates config. For skills: uses Claude Code's built-in eval system (run_eval.py) to measure trigger accuracy, then mutates the skill description. Use when you want an agent or skill to improve itself over N cycles without manual intervention.
Download GGUF models from HuggingFace directly into LM Studio's model directory. Triggers on "download model", "get model", "grab model", "install model".