Absorption engine for Claude with persistent skill graph, exposed as an MCP server.
SpongeBot defines 6 tools with complete input schemas and descriptions. Naming follows verb_noun convention (sponge_absorb, sponge_recall, sponge_skills, sponge_add_skill, sponge_boost, sponge_health). All tools have descriptions (avg 120 chars, within 10-1024 range). Schemas are well-formed with typed properties and enums where appropriate (e.g., mode enum in sponge_absorb). However, output schemas are not documented, LLMs cannot predict response structure. Error handling is absent: no recovery guidance, no actionable error messages, no categorization of retryable vs fatal failures. Parameter descriptions are adequate but lack format constraints (e.g., confidence range 0.0-1.0 stated in description but not as JSON Schema min/max). Tool composition is sound: each tool has one responsibility, and absorb→recall→skills→add_skill→boost forms a coherent chain. Security: no secrets exposed as parameters. Missing: pagination for sponge_recall (could return many results), dry-run for destructive operations (sponge_absorb, sponge_add_skill, sponge_boost modify state without confirmation).
Feed SpongeBot text, code, or documentation to learn from. Stored in the keyword-recall memory with a mode tag.
Add a new skill to SpongeBot's skill DAG. Skills are things it has learned to do.
Boost a skill's confidence score. Use this when a skill proves useful.
Show SpongeBot's subsystem health status — memory, skills, learning engine.
Search SpongeBot's memory by keywords. Returns matching entries with relevance scores.
List all skills SpongeBot has learned. Shows the skill DAG with confidence scores, types, and edges.
Output schemas not documented. LLMs cannot predict response structure for any tool. sponge_recall returns 'matching entries with relevance scores' but structure is undefined. sponge_skills returns 'skill DAG with confidence scores, types, and edges' but field names and types are opaque.
No error handling or recovery guidance. Tools provide no actionable error messages, categorization of retryable vs fatal failures, or suggestions for next steps. An LLM receiving a failure has no guidance on whether to retry, ask the user, or abandon the task.
sponge_recall lacks pagination. No limit enforcement or next_cursor documented. Large memory could return thousands of results, exhausting context window and degrading LLM reasoning.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | <=2025-11-25 | v2 |
Destructive operations (sponge_absorb, sponge_add_skill, sponge_boost) lack confirmation or dry-run support. Agents can modify state irreversibly without user approval or preview.
Parameter constraints not formalized in JSON Schema. confidence (0.0-1.0), amount (0.01-0.2), min_confidence (0.0-1.0) are described as ranges but lack JSON Schema min/max properties. LLMs cannot validate without parsing descriptions.