Loading ToolBench
A collection of AI-powered tools including an animated guiding sprite overlay that uses vision APIs to locate UI elements, a Dreams-to-Reality pipeline sentinel for monitoring video reconstruction, and a Claude Code context manager for session memory and attention-residuals-based context injection.
Critical gaps across all dimensions. Tool definitions lack proper schemas, most parameters are documented but input schemas are not visible in source code. Descriptions are present but generic (avg ~80 chars, well below 194-char baseline). No output schemas documented. Error handling absent. Security concerns: tools accept file paths and execute external pipelines without validation. Composition issues: tools mix concerns (e.g., watch_dtr_pipeline combines watching, validation, and execution). No tool annotations (readOnlyHint/destructiveHint). STDIO transport caps protocol readiness at 50.
Analyze a GLB mesh file and return vertex count, face count, watertightness, and file size statistics.
Capture current screen as PNG bytes for analysis by vision APIs.
Collect context chunks from session buffer, session log, archive, and memory files for attention-residuals scoring.
Read daily observation JSONL and generate a compact structured summary for session-log.md.
Guide sprite to UI element described by user. Captures screen, calls Vision API to find UI element coordinates, and animates sprite to that location.
Log a tool usage observation from PostToolUse hook data to daily JSONL observation file.
Animate sprite to target coordinates with configurable duration and easing curve.
No input schemas visible in source code. All 12 tools lack formal JSON Schema definitions with type constraints, enums, and validation rules. Parameter descriptions exist but are not machine-parseable constraints.
No output schemas documented. LLMs cannot plan downstream tool calls or extract required fields. E.g., analyze_glb returns stats but structure is undocumented; quality_gates returns status but fields are unclear.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 36 | <=2025-11-25 | v2 |
Pack highest-scoring context chunks into a token budget for session injection.
Evaluate mesh quality against configured thresholds and return pass/warn/degraded/failed status with actionable messages.
Score context chunks against a query signal using TF-IDF-like term overlap, recency decay, and source weighting (attention residuals algorithm).
Apply a visual style to the sprite from available style configurations (gemini_golden, blue_glow, hybrid_pulse, emerald_soft, purple_rings, red_alert, cyan_sharp).
Watch ~/dreams-to-reality/test_videos/ for new .mp4 files and auto-run the Dreams to Reality reconstruction pipeline with configurable parameters.
Path traversal and injection vulnerabilities. watch_dtr_pipeline, analyze_glb, generate_session_digest, and log_observation accept file paths as string parameters without validation. No sanitization against path traversal (../../../etc/passwd) or command injection.
Destructive operations lack confirmation or dry-run. watch_dtr_pipeline (WRITE risk) and generate_session_digest (WRITE risk) modify state without user confirmation. No idempotent hints or recovery guidance.
Descriptions are generic and under 100 chars (avg ~60 chars vs 194-char baseline). E.g., 'Capture current screen as PNG bytes for analysis by vision APIs' lacks WHEN to use, prerequisites, or return format. No dependency hints (e.g., 'Call this before guide_to').