fajarhide/omni

• **Archetype**: clawhub-plugin
• **Stars**: 49
• **Last pushed**: 2026-04-12
• **License**: MIT
• **Discovered**: 2026-04-12
• **Language**: Rust
• **Fork**: No

Summary

OMNI is a Rust-based terminal output filter that intercepts command output before it reaches the AI agent, removing noise and preserving signal. Claims up to 90% token reduction. Integrates via native hooks (pre-hook rewriter, post-hook distiller, session-start context, pre-compact summary) into Claude Code, OpenClaw, and MCP agents.

Core architecture: Registry (TOML filter definitions per tool) -> Scorer (context boost) -> Distiller (semantic filtering). Persistence via SQLite for session state and a "RewindStore" that archives raw output for on-demand retrieval. Ships with 35+ tool-specific filter definitions (vitest, jest, eslint, tsc, cargo, docker, kubectl, terraform, etc.) in TOML format.

Key features: DSL-based filter configuration (capture/action/output templates), session intelligence (tracks active files to avoid redundant context), RewindStore for lossless raw output access, omni stats / omni diff for token savings visualization.

Assessment

HIGH extractable value. OMNI's approach to token compression via tool-specific output filtering is directly relevant to babysitter's existing compression layer (packages/sdk/src/compression/). The TOML filter registry pattern, capture/action DSL, and RewindStore concept are all adaptable. The 35+ pre-built filter definitions for common dev tools are immediately useful.

**Extraction priority**: HIGH

Extractable Value: fajarhide/omni

Processes

None -- OMNI is a tool, not a workflow. Its value is entirely in plugin patterns.

Plugin Ideas

1. Smart Output Filter Plugin

• **Category**: Context & Memory
• **install.md**: Installs a hook-based output filter that intercepts tool execution results before they enter the agent context. Uses a TOML-based filter registry with per-tool definitions (capture patterns, action rules, output templates). Includes a RewindStore that archives raw output for on-demand retrieval when the agent needs full detail. Ships with pre-built filters for 35+ common dev tools (vitest, jest, eslint, tsc, cargo, docker, kubectl, terraform, npm, pip, etc.).

2. Token Savings Dashboard Plugin

• **Category**: Developer Experience & UX
• **install.md**: Adds babysitter compression:stats enhancement showing per-tool token savings over time. Tracks raw vs filtered output sizes, calculates cost savings, provides diff visualization comparing raw output to filtered version. Persists metrics in SQLite for historical analysis.

Library Mapping

Plugin Marketplace Mapping

Implicit Procedural Knowledge

• **TOML filter registry pattern**: Each dev tool gets its own TOML file defining capture patterns, confidence scores, and output templates. This is a clean, extensible pattern for babysitter's compression layer -- currently babysitter uses a density-filter engine with FNV-1a dedup, but OMNI's approach of tool-specific semantic filtering could complement it.
• **DSL capture/action/output model**: The omni_config.json shows a simple but powerful DSL: trigger (fast string match), capture (template extraction with named fields), action (keep/count), output (template with extracted fields). This could inform a babysitter plugin API for custom compression rules.
• **RewindStore pattern**: Archive raw output, serve filtered output by default, allow on-demand raw access. Prevents information loss while still saving tokens. Directly relevant to babysitter's compress-output command.
• **Session intelligence**: Tracking which files are actively being edited to avoid feeding the agent context it already knows. This is a form of deduplication that goes beyond babysitter's current FNV-1a approach.