OpenAgentsControl/evals/TESTING_CONFIDENCE.md

Testing System Confidence Assessment

Current State: Honest Evaluation

What Works Well ✅

Feature	Opencoder	OpenAgent	Notes
Agent Selection	✅ Verified	✅ Verified	Both agents correctly identified
Single Tool Calls	✅ Works	✅ Works	list, read, glob, bash all captured
Multi-Tool Chains	✅ Works	⚠️ Partial	glob→read works, but approval blocks chains
Event Capture	✅ 18-56 events	✅ 18-29 events	Real-time streaming works
Tool Verification	✅ Accurate	✅ Accurate	Tool names and inputs captured
File Cleanup	✅ Works	✅ Works	test_tmp/ cleaned before/after

What Needs Work ⚠️

1. OpenAgent Approval Workflow Issue

Problem: OpenAgent reads context but then stops and waits for text approval before executing write/edit tools.

Evidence:

Tool Call Details:
  1. read: {"filePath":".opencode/context/core/standards/code.md"}
  
Violations:
  - missing-required-tool: Required tool 'write' was not used

Root Cause: OpenAgent's system prompt requires text-based approval before execution. Single-prompt tests don't provide this approval.

Solution Options:

1. ✅ Use multi-turn prompts (already implemented for task-simple-001)
2. ⚠️ Need to update ALL openagent tests that expect write/edit to use multi-turn

2. Tool Flexibility

Problem: Agents sometimes use list instead of bash ls.

Solution: ✅ Fixed with mustUseAnyOf - allows alternative tools.

3. Approval Count Always 0

Observation: Approvals given: 0 even when tools execute.

Reason: The permission.request events are for tool-level permissions (dangerous commands), not text-based approval. OpenAgent's text approval is different.

Confidence Levels

Test Type	Confidence	Reason
Opencoder - Read Operations	🟢 HIGH	Works perfectly, verified
Opencoder - Multi-tool Chains	🟢 HIGH	glob→read verified
Opencoder - Bash/List	🟢 HIGH	Both tools work
OpenAgent - Read Operations	🟢 HIGH	Context loading verified
OpenAgent - Multi-turn Approval	🟡 MEDIUM	Works but needs more testing
OpenAgent - Write/Edit	🔴 LOW	Blocked by approval workflow
OpenAgent - Context→Write Chain	🔴 LOW	Stops after context read

Tests That Need Multi-Turn Updates

These openagent tests expect write/edit but use single prompts:

1. ctx-code-001.yaml - Expects read→write
2. ctx-code-001-claude.yaml - Expects read→write
3. ctx-docs-001.yaml - Expects read→edit
4. ctx-tests-001.yaml - Expects read→write
5. ctx-multi-turn-001.yaml - Already multi-turn ✅
6. create-component.yaml - Expects write

Recommended Actions

Immediate (High Priority)

1. Update openagent write/edit tests to multi-turn: ```yaml prompts:

   • text: "Create a file..."
   • text: "Yes, proceed" delayMs: 2000 ```

2. Add mustUseAnyOf where tools are interchangeable:

   behavior:
    mustUseAnyOf: [[bash], [list]]
   

Future Improvements

1. Add text content verification - Check agent's text output contains expected phrases
2. Add timing verification - Ensure context loaded BEFORE execution
3. Add file creation verification - Check test_tmp/ for expected files

Multi-Step Workflow Testing

What We CAN Test Now

1. Read chains: glob → read (verified ✅)
2. Context loading: read context file (verified ✅)
3. Multi-turn conversations: prompt → approval → execute (verified ✅)

What We CANNOT Test Yet

1. Full write workflows: Need multi-turn for openagent
2. Edit workflows: Need multi-turn for openagent
3. Delegation chains: task tool → subagent (not tested)

Summary

Agent	Simple Tasks	Multi-Step	Write/Edit	Confidence
Opencoder	✅	✅	✅	🟢 HIGH
OpenAgent	✅	⚠️	❌	🟡 MEDIUM

Bottom Line:

• Opencoder tests are reliable and working
• OpenAgent tests need multi-turn prompts for write/edit operations
• The framework itself is solid, but test cases need updating