v0.6.0

Summary

Coding-posture owns general execution procedures. Review rubrics, test-design techniques, language guidance, and report templates remain optional companion guidance.

Changes

• Route by requested action: an authorized auth fix remains an implementation task with stronger evidence requirements.
• Require evidence before patching, diagnose new failures before another patch, and report missing reproduction or test seams honestly.
• Keep reviews scoped and findings-only unless edits were authorized; separate confirmed defects from uncertainty.
• Validate RED's failure reason, establish behavior baselines before refactoring, and compare performance under comparable workloads.
• Permit redundant-test removal when behavior coverage is preserved; continue prohibiting hidden failures, weakened gates, and fabricated verification.
• Respect read-only scope and existing authorization; finish the requested work without adding cleanup or repeated checks after completion.
• Expand answer evals and add disposable execution fixtures, informed by current model guidance. No model-specific variants or new quality score.
• Use an installed eval runner, fresh temporary inputs, and configurable model parameters; default to GPT-6.1 Sol at medium effort. Exclude local settings from packages.

Upgrade

Replace older copies of always-on-snippet.md after updating. Otherwise copied instructions can retain the old blanket test-deletion rule.

Mode names and installation paths are unchanged. No cc-thingz skills are changed or removed by this release; overlapping descriptions in older cc-thingz versions require a separate update there. Coding-posture needs no companion installation.

For eval automation, SKILL_EVAL_CLI must name one installed executable. SKILL_EVAL_ROOT now names a parent directory, never cleared by the script. Supply a compatible SKILL_EVAL_CONFIG when changing model/provider.

Changes since v0.5.0