Hooks (Claude Code only)

⚠️ Platform: these hooks target Claude Code. They use Claude Code's hook schema (hooks.json with Stop / PostToolUse / PostToolUseFailure events and the ${CLAUDE_PLUGIN_ROOT} variable). They will not run under OpenCode or other harnesses — those use a different hook mechanism.

Running CAKE2 without these hooks is fully supported: the agents still generate and evaluate operators correctly. The hooks only add convenience automation (stop-gating + debug-skill suggestions); their absence degrades gracefully (nothing blocks, no auto-suggestions).

What each hook does

Event	Script	Purpose
Stop	check_eval_done.sh	Blocks the agent from stopping until evaluation has passed. No-op unless CAKE_OUTPUT_DIR is set by an external batch runner (manual runs always approve).
PostToolUse (Bash)	debug_skill_advisor.py --event PostToolUse	On compile/runtime/precision failures detected in Bash output, suggests the relevant debug skills.
PostToolUseFailure (Bash)	debug_skill_advisor.py --event PostToolUseFailure	Same advisor, on tool-call failures.

Other scripts here (not wired into hooks.json)

These are repo-quality linters, invoked manually or via the optional .pre-commit-config.yaml (a local pre-commit hook, also Claude-Code-independent):

• skills_linter.py — validate SKILL.md frontmatter / token budget.
• agents_linter.py — validate agent definition files.
• _yaml_utils.py — shared YAML helpers for the linters.

Porting to OpenCode

Not provided. An OpenCode port would need the equivalent events expressed in OpenCode's hook format (cf. the cross-platform run-hook.cmd wrapper used by other cannbot community plugins) and is out of scope for this plugin today.