A self-hosted focus system designed for humans and AI assistants.
FocusWith has 11 well-structured tools with explicit schemas and descriptions. All tools are registered in TOOL_SPECS with name, description, and input_schema. Descriptions are clear and action-oriented (avg ~120 chars, within 10-1024 baseline). Schemas use proper JSON Schema with types, enums, and constraints (e.g., priority 1-5, estimated_minutes 1-1440). However, parameter descriptions are MISSING, the schema defines constraints but the LLM cannot see WHY each parameter exists or what values mean. Output schemas are not documented, forcing LLMs to infer response structure. Error handling is absent from tool definitions. No tool annotations visible in schema (readOnlyHint, destructiveHint, idempotentHint) despite risk labels in metadata. Naming is strong (verb_noun pattern, clear intent). Composition is good, tools are single-responsibility and chainable (e.g., create_task → start_focus → control_focus).
Mark a task complete after the user says it is done.
Pause, resume, complete, or cancel the current focus session.
Create a project for a concrete outcome.
Create a reward for a custom uninterrupted-focus duration.
Create one actionable focus task, optionally inside a project.
Read the active timer, open projects, today's tasks, and daily/weekly statistics.
Parameter descriptions missing from all tools. Schemas define types and constraints but LLMs cannot see WHY each parameter exists or what values mean semantically. E.g., 'priority' has min/max but no description explaining priority scale or impact.
Output schemas not documented. Tools return complex nested objects (e.g., focus_context returns active_session, projects, open_tasks, stats, rewards) but LLMs have no schema to understand response structure, forcing inference and increasing hallucination risk.
Tool annotations missing from schema definitions. Risk labels exist in metadata (READ_ONLY, WRITE, DESTRUCTIVE) but are not exposed as readOnlyHint, destructiveHint, or idempotentHint in the input_schema, preventing LLMs from understanding safety implications.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | 2026-07-28+ | v2 |
Read uninterrupted-focus progress, reward targets, and earned rewards.
Redeem an earned reward after the user confirms they want it now.
Select a reward target; this restarts the current uninterrupted run.
Start a timer for a task, or a named free-focus session.
Move, reschedule, rename, or reprioritize an existing task.
No error handling guidance in tool definitions. Tools do not document what errors can occur, when they are retryable, or what the LLM should do next (e.g., 'task_id not found, call get_focus_context to list available tasks').
start_focus has ambiguous required parameters. Tool accepts task_id OR title but schema does not declare this mutual exclusivity or which is preferred. LLMs may pass both or neither, causing silent failures.