eval

name	eval
description	Use when evaluate a session for AI mistakes, user frustration, repeated errors, or scope drift. Use when user says eval session, what went wrong, diagnose session, session audit, post-mortem. Creates report in opencode-workflow/.scratch/evals/. Report only — never fix directly.

Eval

Analyze the current session for agent failures, user frustration, and repeated mistakes. Create an eval report in the opencode-workflow repo. Report only — never fix directly. After report, use grill-with-docs or grilling to plan fixes before implementing.

When to Activate

• User says "eval session", "what went wrong", "diagnose session"
• User expresses frustration ("wasting token", "kok gini", "kenapa", "lagi", "sudah gw bilang")
• Session has repeated tool failures or scope creep
• User wants to capture lessons learned

See REFERENCE.md for detailed content: examples, patterns, anti-patterns, and reference tables.

Self-Improvement Pipeline Integration

Part of self-improvement pipeline (see ADR-0003: Skills Recategorization & Memory Protocol).

Eval reports feed into aggregate → verify-evidence gate → skill fix cycle. Use YAML frontmatter format (see REFERENCE.md). Set findings[].target to the exact skill/agent name so verify-evidence can cross-reference.

Memory protocol: After creating report, store ov add-memory "[workflow:eval] <session-id>: <key findings, severity count>".