the-ai-alliance/cube-harness/profiling

investigator-profiling

When to pick this recipe

Pick this when the failure smell is quantitative — token spend, retry counts, context-window thrash, budget exhaustion — not semantic. Common triggers:

• Agent looped on the same diagnostic for many steps.
• LLM call count is suspiciously high relative to budget.
• Many timeouts or empty completions in summary_stats.

Output: findings.json with the narrowed ProfilingOutput schema: the standard base fields PLUS profile_signal and suggested_budget_change. primary_blame is constrained to agent_scaffolding, model_capability, or none — other categories are out of scope for this recipe.

When NOT to pick this

• Semantic failures (wrong tool choice, misread task): use general_blame.
• Agent-loop pathologies that aren't budget-shaped (e.g. response/action

mismatches): use agent_scaffolding.