Loading ToolBench
Semantic Signal Engine for AI agents: decides what your agent reads, keeps a receipt for every byte it removes, and never invents a result.
OMNI has 19 tools with schemas and descriptions present, but quality is inconsistent. Naming is domain-specific (omni_*) but lacks action verbs, most names are nouns (omni_density, omni_query, omni_recall) rather than verb_noun patterns (analyze_density, search_index, retrieve_context). Descriptions are present but brief (avg ~60 chars, well below the 194-char baseline for A+ tools). Parameters lack depth: many have minimal descriptions, no enums for constrained inputs (e.g., omni_loop_memory 'action' accepts 'get|set|list|forget' but is not declared as an enum), and no documented ranges/formats. Output schemas are not documented, LLMs cannot predict what fields to expect. Two tools (omni_run, omni_cmd) execute shell commands with IRREVERSIBLE risk but lack dry-run, confirmation, or detailed error recovery guidance. Error handling is absent from visible code. Security concerns: omni_session and omni_loop_memory manage state without visible permission gates or audit trails.
Execute terminal tools (git, npm, cargo, docker, etc.) through OMNI's local semantic distillation engine to save 80-90% of token costs.
Analyze file context and dependencies
Analyze semantic density of text to determine distillation potential
Explain token savings from distillation
Identify recurring noise patterns in command output
Retrieve command and context history
Manage knowledge base entries
Tool names lack action verbs. 19 tools use noun-based names (omni_density, omni_query, omni_recall) instead of verb_noun patterns (analyze_density, search_index, retrieve_context). LLMs infer intent from names, noun-only names force description parsing and increase selection errors.
Descriptions are too brief (avg ~60 chars vs 194-char baseline). omni_verify, omni_density, omni_query, omni_recall, omni_context, omni_session lack context on WHEN to use them vs similar tools. Descriptions do not explain prerequisites or downstream dependencies.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2025-06-18+ | v2 |
Read/write persistent loop memory that survives session restarts. Use to remember patterns across loop iterations. Actions: get, set, list, forget.
Get current loop execution status
Query the semantic index for relevant context
Recall previously stored context and patterns
Store content for future recall and context building
Retrieve full content omitted by OMNI distillation (Hash from OMNI notice)
Execute a shell command and return distilled output
Search across stored context and history
Manage session state and lifecycle
Set context for loop execution and iteration tracking
Extract semantic signals from text
As a checker sub-agent, retrieve and evaluate the maker agent's recent work. Returns structured pass/fail with specific issues.
Constrained inputs not declared as enums. omni_loop_memory 'action' param accepts 'get|set|list|forget' but is typed as string with no enum constraint. omni_knowledge 'action' accepts 'get|set|list|delete' similarly. LLMs will hallucinate invalid values without enum enforcement.
Irreversible operations (omni_run, omni_cmd) lack dry-run, confirmation, or detailed error recovery. Shell command execution is IRREVERSIBLE, agents can delete files, corrupt repos, or trigger deployments. No visible confirmation pattern or rollback guidance.
Output schemas not documented. No visible schema definitions for tool responses, LLMs cannot predict return fields, forcing them to parse unstructured output or make assumptions. Breaks tool chaining (e.g., omni_retrieve returns what fields?).