Loading ToolBench
Persistent, emotionally-indexed, provenance-aware memory infrastructure for AI applications; exposed as a Model Context Protocol server for Claude Desktop integration
CAMA exposes 4 tools with mixed quality. Tool naming follows verb_noun convention (cama_check_frame, cama_gather_counterweights, cama_backfill_embeddings, cama_recompute_rel_degrees), which is good. Descriptions are present and substantive (ranging 150 - 280 chars), exceeding the 10 - 1024 char baseline. However, input schemas are visible only for tools 1 - 2; tools 3 - 4 lack detailed parameter documentation in the provided source. Tool 1 (cama_check_frame) has excellent parameter descriptions with defaults and type info. Tool 2 (cama_gather_counterweights) similarly well-documented. Tools 3 - 4 show minimal schema detail, cama_backfill_embeddings has only 'batch_size' with a default, and cama_recompute_rel_degrees has empty input {}. Output schemas are not documented for any tool. Error handling guidance is absent, no recovery hints or categorization. Security: no secrets exposed as params (good), but no audit logging or permission gates visible. Composition: tools are single-purpose and well-separated (frame detection vs counterweight gathering vs maintenance), which is correct.
Backfill embeddings for memories that don't have them yet. Works with local or API embeddings. Returns JSON with backfilled count and remaining count.
Check a draft response for frame-capitulation patterns. Runs six detectors against the draft text: capitulation_imperative, reviewer_tone_adoption, self_deprecation_without_evidence, wellness_prompts, hedged_stop, evidence_absent_recommendation. Returns a JSON blob containing concerns (list of pattern_name, severity, location, excerpt, reasoning, suggested_pivot), max_severity, passed status, summary, and optionally counterweight_anchors with current repo-state facts.
Gather counterweight anchors from current repo state. Pulls recent-commit count, recent-PR count, test count, ruff lint status, and EVIDENCE.md row count from the actual repository. Returns each as a structured anchor naming the fact_type, value, source command, and detector relevance. Useful when the assistant wants to ground a response in measurable evidence without running the frame-capitulation detector.
Recompute all precomputed rel_degree values from edges. Updates the rel_degree field for all memories based on their edge counts in the graph.
Output schemas not documented for any tool. LLMs cannot plan downstream calls or extract fields without knowing response structure.
cama_recompute_rel_degrees has empty input schema ({}). No parameters documented; unclear what the tool does or how to invoke it.
cama_backfill_embeddings parameter 'batch_size' lacks description of valid range, constraints, or rationale. Unbounded integers risk absurd values.
No error handling guidance. Tools do not indicate what errors are retryable, user-fixable, or fatal. No recovery hints provided.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Maintenance tools (cama_backfill_embeddings, cama_recompute_rel_degrees) marked WRITE but lack confirmation/dry-run pattern. Agents could accidentally trigger destructive operations.