04-implicit-discovery.yaml 1.9 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364
  1. id: openagent-contextscout-implicit-discovery
  2. name: "OpenAgent: Implicit Discovery - Uses ContextScout Without Being Told"
  3. description: |
  4. Tests that OpenAgent PROACTIVELY uses ContextScout when encountering
  5. unfamiliar topics, even when NOT explicitly instructed to do so.
  6. This is the KEY test - agents should be smart enough to recognize
  7. when they need context discovery without the user saying "use contextscout".
  8. This validates:
  9. - Agent recognizes unfamiliar domain (registry system)
  10. - Agent PROACTIVELY delegates to ContextScout (not instructed)
  11. - Agent loads discovered context files
  12. - Agent provides accurate answer based on discovered context
  13. Expected Behavior:
  14. - SHOULD delegate to ContextScout automatically
  15. - MUST load discovered context files
  16. - SHOULD find registry-specific context
  17. - SHOULD NOT guess or fabricate information
  18. This test SHOULD FAIL if:
  19. - Agent doesn't use ContextScout (guesses instead)
  20. - Agent provides generic answer without loading context
  21. - Agent fabricates information without discovery
  22. category: developer
  23. agent: core/openagent
  24. prompts:
  25. - text: |
  26. How does the registry system work in this repository?
  27. What's the auto-detect process?
  28. approvalStrategy:
  29. type: auto-approve
  30. behavior:
  31. mustUseTools:
  32. - task # Must delegate to ContextScout
  33. - read # Must read discovered files
  34. minToolCalls: 3
  35. maxToolCalls: 25
  36. expectedViolations:
  37. - rule: approval-gate
  38. shouldViolate: false
  39. severity: error
  40. timeout: 90000
  41. tags:
  42. - openagent
  43. - contextscout-integration
  44. - implicit-discovery
  45. - proactive
  46. - critical
  47. # Expected outcome:
  48. # - Agent recognizes "registry system" as unfamiliar domain
  49. # - Agent delegates to ContextScout WITHOUT being told
  50. # - ContextScout finds .opencode/context/openagents-repo/core-concepts/registry.md
  51. # - Agent loads discovered files
  52. # - Agent provides accurate answer about registry and auto-detect