Python Task Research System using AI for Task Breakdown and STORM-based Research. Provides tools for parsing input specifications into tasks, expanding tasks into subtasks, analyzing task complexity, and conducting STORM-based research.
The server defines 8 tools with explicit FastMCP registration. Tool names follow verb_noun conventions (parse_inputs, update_tasks, expand_task, etc.), which is good. However, descriptions are inconsistent in quality, parameter documentation is sparse, and output schemas are nowhere visible in the provided code. Several tools lack clear descriptions of what they return, making it difficult for LLMs to chain calls or plan downstream actions. Parameters mostly have types but lack rich descriptions explaining constraints, formats, or when to use them. Schema documentation is essentially absent, the code snippet doesn't show response structures, pagination info, or field definitions. Error handling is mentioned only in passing (try/except blocks with logging) but no structured recovery guidance is visible.
Analyzes task complexity and generates recommendations for subtask counts, saves report to file.
Expands all eligible pending tasks into subtasks.
Expands a task into subtasks (with AC). Use research=True for STORM workflow (requires confirmation via logs/UI if run manually).
Automatically fixes invalid dependencies (missing refs, self-deps, simple cycles).
Generates individual task files (phase_X_task_Y.txt).
Parses input specs to generate initial tasks (incl. phase) and overwrites the tasks file.
Output schemas are completely undocumented. No tool description specifies what is returned, its structure, or how to interpret it for downstream tool calls.
Parameter descriptions are sparse or missing. Most parameters lack guidance on valid ranges, formats, dependencies, or error cases. Example: 'num_subtasks' and 'num_tasks' are integers with no bounds specified; LLMs may pass absurd values.
Destructive operations (update_tasks, generate_task_files, expand_all_tasks with force=true, fix_dependencies) lack dry-run or confirmation steps. Agents can accidentally overwrite data without recovery options.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 41 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 0 | - | v1 |
Updates tasks (title, desc, details, phase, test strat) from ID based on prompt.
Validates task dependencies for issues like missing refs, self-dependencies, or circular dependencies.
Error handling is not visible in tool signatures or descriptions. Tools return success strings but have no documented guidance for failure cases (missing file, invalid ID, malformed input). LLMs won't know how to recover.
Tool composition: expand_task and expand_all_tasks overlap significantly; the distinction could be clearer. Consider renaming expand_all_tasks to expand_pending_tasks and documenting the 'eligible' and 'pending' criteria.
generate_task_files has no parameters but its description is vague (40 chars). Doesn't explain: Does it generate for all tasks? Only pending? What if tasks.json doesn't exist? Descriptions under 20 - 50 chars are typically too brief.