OpenAgentsControl/.opencode/prompts

Prompt Library System

Multi-model prompt variants with integrated evaluation framework for testing and validation.

---

🎯 Quick Start

Testing a Prompt Variant

# Test with eval framework (recommended)
cd evals/framework
npm run eval:sdk -- --agent=openagent --prompt-variant=llama --suite=smoke-test

# Test with specific model
npm run eval:sdk -- --agent=openagent --prompt-variant=llama --model=ollama/llama3.2 --suite=core-tests

# View results
open ../results/index.html

Using a Variant Permanently

# Switch to a variant
./scripts/prompts/use-prompt.sh --agent=openagent --variant=llama

# Restore default (canonical agent file)
./scripts/prompts/use-prompt.sh --agent=openagent --variant=default

---

📁 Structure

.opencode/
├── agent/                       # Canonical agent prompts (defaults)
│   ├── openagent.md            # OpenAgent default (Claude-optimized)
│   └── opencoder.md            # OpenCoder default
└── prompts/                     # Model-specific variants
    ├── navigation.md               # This file
    ├── openagent/              # OpenAgent variants
    │   ├── gpt.md             # GPT-4 optimized
    │   ├── gemini.md          # Gemini optimized
    │   ├── grok.md            # Grok optimized
    │   ├── llama.md           # Llama/OSS optimized
    │   ├── TEMPLATE.md        # Template for new variants
    │   ├── navigation.md          # Variant documentation
    │   └── results/           # Per-variant test results
    │       ├── default-results.json  # Default (agent file) results
    │       ├── gpt-results.json
    │       ├── gemini-results.json
    │       └── llama-results.json
    └── opencoder/              # OpenCoder variants
        └── ...

Architecture:

• Agent files (.opencode/agent/*.md) = Canonical defaults (source of truth)
• Prompt variants (.opencode/prompts/<agent>/<model>.md) = Model-specific optimizations
• Results always saved to .opencode/prompts/<agent>/results/ (including default)

---

🧪 Evaluation Framework Integration

Running Tests with Variants

The eval framework automatically:

• ✅ Switches to the specified variant
• ✅ Runs your test suite
• ✅ Tracks results per variant
• ✅ Restores the default prompt after tests

# Smoke test (1 test, ~30s)
npm run eval:sdk -- --agent=openagent --prompt-variant=llama --suite=smoke-test

# Core suite (7 tests, ~5-8min)
npm run eval:sdk -- --agent=openagent --prompt-variant=llama --suite=core-tests

# Custom test pattern
npm run eval:sdk -- --agent=openagent --prompt-variant=llama --pattern="01-critical-rules/**/*.yaml"

Auto-Model Detection

Variants specify recommended models in their YAML frontmatter:

---
model_family: llama
recommended_models:
  - ollama/llama3.2
  - ollama/qwen2.5
---

If you don't specify --model, the framework uses the first recommended model.

Results Tracking

Results are saved in two locations:

1. Main results: evals/results/latest.json (includes prompt_variant field)
2. Per-variant: .opencode/prompts/{agent}/results/{variant}-results.json

View in dashboard: evals/results/index.html (filter by variant)

---

📝 Creating a New Variant

Step 1: Copy Template

cp .opencode/prompts/openagent/TEMPLATE.md .opencode/prompts/openagent/my-variant.md

Step 2: Edit Metadata

---
model_family: oss
recommended_models:
  - ollama/my-model
status: experimental
maintainer: your-name
description: Optimized for my specific use case
---

Step 3: Customize Prompt

Edit the prompt content below the frontmatter for your target model.

Step 4: Validate

# Validate the variant exists and metadata is correct
cd evals/framework
npm run eval:sdk -- --agent=openagent --prompt-variant=my-variant --suite=smoke-test

Step 5: Test Thoroughly

# Run core test suite
npm run eval:sdk -- --agent=openagent --prompt-variant=my-variant --suite=core-tests

# Check results
open ../results/index.html

Step 6: Document Results

Update .opencode/prompts/openagent/navigation.md with:

• Test results (pass rate, timing)
• Known issues or limitations
• Recommended use cases

---

🎯 Available Variants

OpenAgent

Variant	Model Family	Status	Best For
default	Claude	✅ Stable	Production use, Claude models
gpt	GPT	✅ Stable	GPT-4, GPT-4o
gemini	Gemini	✅ Stable	Gemini 2.0, Gemini Pro
grok	Grok	✅ Stable	Grok models (free tier)
llama	Llama/OSS	✅ Stable	Llama, Qwen, DeepSeek, other OSS

See openagent/navigation.md for detailed test results.

OpenCoder

Coming soon.

---

🔧 Advanced Usage

Custom Test Suites

Create custom test suites for your variant:

# Create suite
cp evals/agents/openagent/config/smoke-test.json \
   evals/agents/openagent/config/my-suite.json

# Edit suite (add your tests)
# Validate suite
cd evals/framework && npm run validate:suites openagent

# Run with your variant
npm run eval:sdk -- --agent=openagent --prompt-variant=my-variant --suite=my-suite

See evals/TEST_SUITE_VALIDATION.md for details.

Comparing Models

Test the same variant with different models:

# Test Llama 3.2
npm run eval:sdk -- --agent=openagent --prompt-variant=llama --model=ollama/llama3.2 --suite=core-tests

# Test Qwen 2.5
npm run eval:sdk -- --agent=openagent --prompt-variant=llama --model=ollama/qwen2.5 --suite=core-tests

# Compare results in dashboard
open evals/results/index.html

---

📊 Understanding Results

Dashboard Features

The results dashboard (evals/results/index.html) shows:

• ✅ Filter by prompt variant
• ✅ Filter by model
• ✅ Pass/fail rates per variant
• ✅ Test execution times
• ✅ Detailed test results

Result Files

Main results (evals/results/latest.json):

{
  "meta": {
    "agent": "openagent",
    "model": "ollama/llama3.2",
    "prompt_variant": "llama",
    "model_family": "llama"
  },
  "summary": {
    "total": 7,
    "passed": 7,
    "failed": 0,
    "pass_rate": 1
  }
}

Per-variant results (.opencode/prompts/openagent/results/llama-results.json):

• Tracks all test runs for this variant
• Shows trends over time
• Helps identify regressions

---

🚀 For Contributors

Creating a Variant for PR

1. Create your variant in .opencode/prompts/<agent>/<model>.md
2. Test thoroughly with eval framework
3. Document results in agent README
4. Submit PR with variant file only

PR Requirements

• ✅ Variant has YAML frontmatter with metadata
• ✅ Variant passes core test suite (≥85% pass rate)
• ✅ Results documented in agent README
• ✅ Agent file unchanged (unless updating default)
• ✅ No default.md files in prompts directory
• ✅ CI validation passes

Validation

# Validate your variant
cd evals/framework
npm run eval:sdk -- --agent=openagent --prompt-variant=your-variant --suite=core-tests

# Ensure PR uses default
./scripts/prompts/validate-pr.sh

---

🎓 Design Principles

1. Agent Files are Canonical Defaults

• Agent files (.opencode/agent/*.md) are the source of truth
• Tested and production-ready
• Optimized for Claude (primary model)
• Modified through normal PR process

2. Variants are Model-Specific Optimizations

• Stored in .opencode/prompts/<agent>/<model>.md
• Optimized for specific models/use cases
• May have different trade-offs
• Results documented transparently

3. Results are Tracked

• Every test run tracked per variant
• Dashboard shows variant performance
• Easy to compare variants

4. Easy to Test

• One command to test any variant
• Automatic model detection
• Results saved automatically

5. Safe to Experiment

• Variants don't affect default
• Easy to switch and restore
• Test before committing

---

📚 Related Documentation

• Eval Framework Guide - How to run tests
• Test Suite Validation - Creating test suites
• OpenAgent Variants - OpenAgent-specific docs
• Contributing Guide - Contribution guidelines

---

🆘 Troubleshooting

Variant Not Found

# List available variants
ls .opencode/prompts/openagent/*.md

# Check variant exists
npm run eval:sdk -- --agent=openagent --prompt-variant=your-variant --suite=smoke-test

Tests Failing

1. Check variant metadata (YAML frontmatter)
2. Verify recommended model is available
3. Run with debug flag: --debug
4. Check results in dashboard

Model Not Available

# Check available models
# For Ollama: ollama list
# For OpenRouter: check openrouter.ai/models

# Specify model explicitly
npm run eval:sdk -- --agent=openagent --prompt-variant=llama --model=ollama/llama3.2

---

💡 Tips

• Start with smoke-test - Fast validation (1 test, ~30s)
• Use core-tests for thorough testing - 7 tests, ~5-8min
• Check dashboard regularly - Visual feedback on variant performance
• Document your findings - Help others by sharing results
• Test with multiple models - Same variant may perform differently

---

🔮 Future Enhancements

• Automated variant comparison reports
• Performance benchmarking across variants
• Variant recommendation based on model
• Historical trend analysis
• A/B testing framework

---

Questions? See openagent/navigation.md or open an issue.