Both tools are hook-based validators with minimal, domain-specific schemas. evidence-lint and turn-end-gate are READ_ONLY hooks that validate state transitions, not general-purpose agent tools. Naming is clear (verb_noun pattern), but descriptions are highly specialized and assume deep familiarity with build-it's hardened mode semantics. Input schemas are present but sparse: evidence-lint accepts tool_name, tool_input (with file_path, content/new_string); turn-end-gate accepts transcript_path and stop_hook_active. Neither tool has output schema documentation. Error handling is fail-open by design (intentional for hooks), but descriptions lack guidance on what LLMs should do when hooks reject actions. These are not traditional agent tools, they are Claude Code hooks that intercept and gate tool invocations, making them unsuitable for general LLM composition.

• Document output schema for both hooks: evidence-lint returns {hookSpecificOutput: {hookEventName, permissionDecision, permissionDecisionReason}} on deny, empty on allow; turn-end-gate returns {decision, reason} on block, empty on allow.
• Expand descriptions to explain the validation logic in plain language: 'evidence-lint: Validates that VERIFIED rows in reports have matching entries in the evidence ledger. Rejects writes if evidence is missing or unmatched.' Similar for turn-end-gate.
• Add parameter descriptions: tool_input should explain 'Object containing file_path (string) and either content (for Write) or new_string (for Edit)'. stop_hook_active should explain 'Boolean flag to prevent recursive hook invocation loops'.
• Document the fail-open behavior explicitly in descriptions: 'If the hook crashes, the action is allowed and the error is logged to .taskstate/hooks.log. This prevents hook bugs from blocking legitimate operations.'
• Consider whether these hooks should be exposed as MCP tools at all. If they are Claude Code hooks only, they should not be registered as MCP tools. If they must be MCP tools, redesign them as stateless validators that accept report/transcript content and return validation results.