No output schemas documented for any of the 16 tools. Agent chaining and downstream tool parameter mapping is impossible. LLMs cannot plan multi-step workflows without knowing what fields to extract.

Enum parameters are documented in descriptions but not declared in JSON schemas. Tools like 'code_review' (focus: security|performance|quality|all), 'gencommit' (type: fixed|feat|docs|style|chore|refactor|test), 'start_feature' (template_profile: auto|guided|strict, requirements_mode: steady|loop), 'start_bugfix' (same), 'start_ralph' (mode: safe|normal), 'estimate' (experience_level: junior|mid|senior) all use string types with enum-like constraints documented only as text. LLMs hallucinate invalid values.

Numeric parameters lack min/max constraints. start_ralph has unbounded: max_iterations, max_minutes, confirm_every, confirm_timeout, max_same_output, max_diff_lines, cooldown_seconds. estimate has unbounded team_size. LLMs can pass absurd values (team_size=999999, max_iterations=1000000) that break intended limits.

String parameters have no length constraints or format declarations. 'code' parameter in refactor/gentest/fix_bug is unbounded, LLMs could pass multi-megabyte code blocks. No maxLength, no format hints (path, regex, date). Violates baseline: structured constraints prevent hallucination.

Parameter dependencies undocumented. start_feature and start_bugfix have mutually exclusive modes (requirements_mode: steady vs loop) with conditional parameters (loop_max_rounds only applies if loop=true), but this is not declared in schema or description. LLMs may pass both steady and loop_max_rounds together, causing confusion.

ask_user parameter structure is ambiguous. 'questions' is an array of objects with nested properties (question, context, options, required) but the schema does not define these nested object types. Does 'options' accept strings? Objects? Required is boolean but undescribed semantically.

No error handling patterns visible. No tool description mentions what to do on failure, how to recover, or what errors are retryable vs fatal. Pattern:recovery-guide and pattern:error-classification are completely absent.

package.json claims 21 tools but only 16 are defined in source code. Description says 'UI/UX Pro Max', 'Structured Output', 'Workflow Orchestration' but start_ui, ui_design_system, ui_search, sync_ui_data, start_product are missing from provided code. Cannot evaluate 5 tools. Score penalizes this incompleteness.

interview tool has ambiguous 'answers' parameter. Schema shows 'answers' as object but does not specify key format or value constraints. How does the agent know which question IDs to use? What format should answers have?

gencommit and git_work_report rely on git command execution but do not document error cases: What if git is not available? What if the repository is corrupted? What if the date range is invalid? No recovery guidance.