Summary
Support per-file token budget configuration in .waza.yaml using glob patterns, extending the current task-level max_tokens in the behavior grader.
Motivation
Skills often have multiple files (SKILL.md, prompts, configs) with different token budgets. The current behavior_grader.max_tokens only checks total task output. A per-file budget config allows finer-grained token limits.
Proposed Implementation
.waza.yaml config
tokens:
limits:
defaults:
skill_md: 5000
prompt: 8000
overrides:
skills/complex-skill/SKILL.md: 7500
Integration with waza tokens check
- Read per-file limits from .waza.yaml
- Fall back to .token-limits.json for legacy support (with deprecation warning)
- Report per-file results
Acceptance Criteria
- .waza.yaml supports tokens.limits.defaults and tokens.limits.overrides with glob patterns
- waza tokens check reads per-file limits
- Legacy .token-limits.json fallback with deprecation warning
- Per-file reporting in both text and JSON formats
- Tests covering: defaults, overrides, glob matching, legacy fallback
Summary
Support per-file token budget configuration in .waza.yaml using glob patterns, extending the current task-level max_tokens in the behavior grader.
Motivation
Skills often have multiple files (SKILL.md, prompts, configs) with different token budgets. The current behavior_grader.max_tokens only checks total task output. A per-file budget config allows finer-grained token limits.
Proposed Implementation
.waza.yaml config
tokens:
limits:
defaults:
skill_md: 5000
prompt: 8000
overrides:
skills/complex-skill/SKILL.md: 7500
Integration with waza tokens check
Acceptance Criteria