This server has fundamental structural issues that limit its production readiness. Tool definitions are present with schemas and descriptions, but several critical quality gaps emerge: (1) All descriptions are in Japanese, limiting LLM language compatibility for most English-based agents. (2) Descriptions are brief but lack actionable guidance (e.g., 'get tasks' vs. 'Fetch all open tasks for this project; use project-id format: PROJ-123'). (3) Parameter descriptions are minimal and do not explain constraints, formats, or dependencies. (4) Output schemas are completely undocumented, handlers return plain JSON without schema guidance. (5) Error handling exists but does not classify errors as retryable/fatal or provide recovery guidance. (6) No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels (delete_task is DESTRUCTIVE). (7) No idempotency guarantees documented. (8) The server uses natural IDs (projectId, issueId as strings) but does not explain expected format or validation. Overall, the tool definitions are skeletal, they register but lack the depth and rigor expected of production tools.
All tool descriptions in Japanese; LLM integration assumes English or multilingual LLM support not guaranteed. Critical for international/English-primary agent adoption.
No output schemas documented. LLM cannot know what fields to expect from list_tasks, create_task, update_task, or add_comment responses. Forces LLM to infer structure or parse free-text JSON.
delete_task has DESTRUCTIVE risk but no destructiveHint annotation in tool definition. No confirmation/dry-run pattern offered. LLM has no way to warn user or implement safeguards.
Rewrite all tool descriptions in English (or add English translations). Include WHAT the tool does, WHEN to use it, and WHAT it returns. Target 50-150 characters per description. Example: 'list_tasks → Fetch all issues in a GitLab project. Returns up to 20 issues per page with ID, title, status, and URL. Use projectId (numeric or path-encoded name). Call first to discover available tasks.'
Add explicit output schemas to all tool definitions. Example for create_task: 'Returns {task_id: string, web_url: string, title: string, state: string, created_at: ISO8601}'. Document in handler or schema file.
Add format/pattern guidance to all string parameters. Example: 'projectId (string, format: PROJ-123 or numeric ID; required)', 'title (string, 1-255 chars)', 'body (string, supports Markdown, 1-5000 chars)'.
Add pagination to list_tasks: include limit (1-100, default 20), offset or page (0-indexed), and return total_count and has_more fields. Update schema and description.
Annotate tools with readOnlyHint/destructiveHint/idempotentHint in schema. Example: list_tasks gets readOnlyHint: true; delete_task gets destructiveHint: true, requiresConfirmation: true.
Implement dry-run or confirmation flow for delete_task. Return {requiresConfirmation: true, confirmationToken: 'xxx'} on first call; require confirmationToken on second call to execute deletion. Or add a separate 'confirm_delete_task' tool.
Refactor handler responses to return structured JSON instead of plain text. Example: {success: true, task: {id, web_url, title, state, created_at}} instead of {content: [{type: 'text', text: '✨ ...'}]}. This lets LLM extract data for chaining.
Score history
Overall score trend
↑ 4 points across a rubric change (v1 → v2)
46/100
Scored
Grade
Overall
Spec posture
Rubric
2026-09-22
F
46
2026-07-28+
v2
2026-03-09
F
42
-
v1
Parameter descriptions minimal or missing format guidance. Example: projectId is 'string' but no hint whether it's numeric (123), alphanumeric (PROJ-123), or URL-encoded. Forces LLM to guess or causes validation errors.
No pagination support in list_tasks. Schema lacks limit/offset/page_size parameters. For large projects, response could exceed context window. No mention of result limits or chunking.
Error messages in handlers are basic. Example: GitLab API error message is echoed but not classified as retryable/fatal/user-fixable. LLM cannot decide whether to retry, ask user, or stop.
No tool annotations (readOnlyHint, idempotentHint, destructiveHint) in schema definitions. LLM cannot infer tool safety or retry behavior. list_tasks should be marked readOnly; delete_task should be marked destructive.
Handler responses use plain text strings (emoji + message) rather than structured JSON. Example: 'create_task' returns 'type: text, text: "✨ タスクが作成されました..."' instead of structured {task_id, url, created_at}. LLM cannot extract data for downstream tool calls.
No idempotency guarantees documented. create_task may create duplicates if retried. delete_task will fail on second attempt (already deleted). LLM cannot safely retry on transient failures.
update_task description does not clarify that all of (title, description, state) are optional but at least one is required. LLM may call tool without any update params, expecting idempotent no-op, but API may reject.
update_task
Enhance error messages with classification and recovery guidance. Example: 'GitLab API 404 (retryable): Project 999 not found. Verify projectId is correct. Available projects: PROJ-1, PROJ-2. Call list_projects() to discover.'
Document idempotency guarantees. Example: 'list_tasks is idempotent (safe to retry). create_task creates a new issue on each call (not idempotent; use issue ID to check if already created). delete_task is irreversible.'
Add optional parameters to list_tasks for filtering (state: opened|closed|all; labels: array of strings) and sorting (sort_by: created_at|updated_at; direction: asc|desc). Include in schema description.
Clarify update_task behavior: 'All fields (title, description, state) are optional. Provide at least one to update. Omitted fields are not changed. Returns updated task object with same schema as create_task.'
Add a 'get_task' tool to fetch a single task by ID (faster than listing all and filtering). Input: {projectId, issueId}. Output: full task object. Enables chaining after create_task without redundant list call.
Consider adding search_tasks tool for filtering by title, state, or labels without fetching entire project. Reduces context size and improves performance for large projects.