A pure-text roguelike game with MCP (stdio) tool interface. Players navigate branching narratives across multiple lives, with skills and memories persisting across deaths through a legacy system.
This server has critical definition gaps. Tool descriptions are present but generic (avg ~80 chars, below the 194-char baseline). Input schemas are partially visible but lack proper JSON Schema structure with type definitions for most parameters. The 'choose' tool accepts an integer but provides no enum of valid options or range constraints. 'recite' accepts both string and integer with ambiguous semantics. 'write_memory' and 'wish' have parameter descriptions but no type validation rules. No output schemas are documented. Error handling is absent, no recovery guidance, no categorization of retryable vs fatal errors. Tool names are verb-based (good), but parameter naming is inconsistent (e.g., 'option' vs 'wish_type'). No tool declares what it returns or how results chain to downstream calls.
Make a choice in response to a game event or scenario. Advances the narrative based on player selection.
Get a detailed report of the completed life, including deeds, fragments discovered, achievements, and statistics.
View the world legacy data including all deeds performed across lives, discovered truth fragments, achievements, and final crossing records.
Respond to the Lake crossing with a phrase. Used to answer the guardian at the crossing between lives with specific keywords (thirst, origin, lineage).
Begin a new life in the game. Initializes character with faction, augmentation level, and starting conditions.
Get current game state including life number, augmentation percentage, body health, skills, heat level, and current scene.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., what does 'status' return? Is it a string, object, or structured record?). Violates pattern:tool and pattern:response-shaper.
Input parameters lack proper JSON Schema type definitions and constraints. 'choose' accepts an integer but provides no enum of valid options (1-indexed, but how many?). 'recite' accepts both string and integer with no clarification of when to use each. 'wish' has 'value' as integer with no range or validation rules.
Tool descriptions are too brief (avg 45 chars, well below 194-char baseline). 'debrief', 'status', 'legacy' lack context for when to call them vs similar tools. No descriptions explain what the tool returns or how to use results in subsequent calls.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 42 | <=2025-11-25 | v2 |
Undo the last choice and return to the previous decision point.
Spend accumulated wish points to influence faction, augmentation level, or other game parameters. Costs vary by wish type.
Write or update a memory entry that persists across deaths. Limited to 10 entries of 10 characters each (or 10 words in English).
No error handling or recovery guidance. If 'choose' receives an invalid option, what happens? If 'write_memory' exceeds the 10-character limit, does it truncate, reject, or fail silently? No error messages guide the LLM to self-correct.
Parameter naming inconsistency. 'choose' uses 'option' (integer), 'recite' uses 'option' (string), 'wish' uses 'wish_type' and 'value'. Inconsistent naming forces LLMs to reason about field mappings and increases error likelihood.