OpenWriterControle/evals/opencode/openagent/README.md

OpenAgent Evaluation Tests

Overview

Test suite for validating OpenAgent behavior against defined standards and critical rules.

OpenAgent Rules

OpenAgent must follow these critical rules:

1. Approval Gates - Request approval before bash/write/edit/task operations
2. Context Loading - Load required context files before execution
3. Delegation - Delegate when 4+ files or complex tasks
4. Stop on Failure - Never auto-fix errors, always report first
5. Tool Selection - Use appropriate tools for tasks

See: OpenAgent Specification

Test Categories

Conversational

Simple questions that require no execution:

• Code explanations
• Informational queries
• Analysis requests

Expected Behavior:

• No execution tools
• No approval required
• Response provided

Task

Simple file operations:

• File creation
• File editing
• Simple changes

Expected Behavior:

• Approval requested
• Appropriate tool selection
• Context loading (if needed)

Complex

Multi-file features and complex tasks:

• Architecture changes
• Feature implementation
• Complex refactoring

Expected Behavior:

• Approval requested
• Context loaded
• Delegation used (4+ files)
• Appropriate tool selection

Edge Cases

Error handling and special scenarios:

• Permission denials
• Missing files
• Error recovery
• Invalid inputs

Expected Behavior:

• Stop on failure
• Report errors (no auto-fix)
• Graceful handling

Test Cases

Test cases are defined in YAML files in test-cases/:

• approval-gates.yaml - Approval gate enforcement tests
• context-loading.yaml - Context loading compliance tests
• delegation.yaml - Delegation appropriateness tests
• tool-usage.yaml - Tool selection tests
• edge-cases.yaml - Error handling and special scenarios

Running Tests

# Run all OpenAgent tests
cd evals/framework
npm run eval -- --agent openagent --all

# Run specific test category
npm run eval -- --agent openagent --test approval-gates

# Run single test case
npm run eval -- --agent openagent --test approval-gates --case file-creation-with-approval

# Analyze specific session
npm run eval -- --agent openagent --session ses_xxxxx

Test Results

Results are stored in evals/results/YYYY-MM-DD/openagent/:

results/2025-11-21/openagent/
├── summary.json          # Overall summary
├── approval-gates.json   # Approval gate results
├── context-loading.json  # Context loading results
├── delegation.json       # Delegation results
└── report.md            # Human-readable report

Configuration

Configuration is in config.yaml:

agent: openagent
agent_path: ../../../.opencode/agent/openagent.md
test_cases_path: ./test-cases
sessions_path: ./sessions
evaluators:
  - approval-gate
  - context-loading
  - delegation
  - tool-usage
pass_threshold: 75
scoring:
  approval_gate: 40
  context_loading: 40
  delegation: 10
  tool_usage: 10

Recorded Sessions

The sessions/ directory contains recorded test sessions for regression testing:

sessions/
├── simple-question/
│   ├── session.json
│   └── expected.yaml
├── file-creation/
│   ├── session.json
│   └── expected.yaml
└── complex-feature/
    ├── session.json
    └── expected.yaml

Success Criteria

Overall

• Pass Rate: ≥ 90% of tests pass
• Average Score: ≥ 85/100
• Critical Violations: 0

Per Evaluator

• Approval Gates: 100% compliance (critical)
• Context Loading: ≥ 90% compliance
• Delegation: ≥ 80% compliance
• Tool Usage: ≥ 85% compliance

Adding New Tests

1. Create test case in appropriate YAML file:

- id: my-new-test
  name: "My New Test"
  description: "Test description"
  category: task
  input: "User prompt"
  expected_behavior:
    approval_requested: true
    tool_used: write
  evaluators:
    - approval-gate
    - tool-usage
  pass_threshold: 75

2. (Optional) Record a session for regression testing:

# Run OpenCode with the test prompt
opencode --agent openagent
> "User prompt"

# Copy session to sessions/
cp ~/.local/share/opencode/project/.../session/info/ses_xxxxx.json \
   sessions/my-new-test/session.json

3. Run the test:

npm run eval -- --agent openagent --test my-new-test

Continuous Integration

Tests run automatically on:

• Pull requests
• Commits to main
• Nightly builds

See: .github/workflows/eval-openagent.yml

Related Documentation

• Evaluation Framework
• OpenAgent Specification
• OpenCode Logging System

Metrics

Track these metrics over time:

• Pass rate trend
• Average score trend
• Violation frequency
• Model performance
• Cost per test