Output schemas are not documented. LLMs cannot predict what fields to expect from tool responses, forcing them to parse responses speculatively and lose context for chaining calls. For example, autovio_works_create returns a work object but the response shape is not declared in the tool definition.

Destructive and sensitive tools (autovio_projects_delete, autovio_works_delete, autovio_templates_delete, autovio_auth_register) lack confirmation or dry-run patterns. Descriptions do not warn the LLM of irreversibility. An agent could accidentally delete a project and all its works without explicit confirmation.

Error handling in tool implementations (src/tools/*.ts) uses generic ToolExecutionError with formatErrorDetail() but does not provide actionable recovery guidance in error messages. LLMs cannot determine if an error is retryable, user-fixable, or fatal, nor what to do next.

autovio_works_update and autovio_projects_update expose many optional fields with complex interdependencies (scenes, generatedScenes, editorState) but do not document which combinations are valid. LLMs may pass invalid field combinations resulting in server-side rejections.

Parameter descriptions for complex nested objects (styleGuide, content in templates) are minimal. styleGuide is documented as 'Brand style settings: tone, colors, tempo, camera style, must include/avoid' but does not specify which subfields are required, their types, or valid values. This forces LLMs to guess at the structure.

List tools (autovio_projects_list, autovio_works_list, autovio_templates_list) do not accept pagination parameters (limit, offset, cursor) and do not document result limits. Large projects with hundreds of works could produce responses that blow the context window.

Tool descriptions for compound workflows (autovio_works_generate_scenario, autovio_works_apply_template) state 'After creating, use...' or 'Call after creating a work' but do not explain prerequisite state. An LLM might call generate_scenario on a work that already has scenes, overwriting them without warning.