Enhance command-based content type inference and distillation logic

[fajarhide]

PR Auto Describe

Summary

This PR overhauls Omni's core pipeline and CLI stack, replacing the legacy content-type classification model with command-aware routing, cleaning up deprecated code, and streamlining user-facing reporting and performance.

---

Key Changes

1. Replaced legacy content-based pipeline routing with command-first profile resolution
2. Rewrote omni stats CLI: removed --by-type flag, added --project per-project analytics
3. Deleted unused distillers and deprecated classifier/composer modules
4. Updated hook logic with auto-sibling command detection for pipe mode

---

Detailed Breakdown

Core Pipeline Overhaul

• Removed entire src/pipeline/classifier.rs legacy content classifier
• Rewrote distiller system: removed content_type() from Distiller trait, replaced get_distiller() with command-based distill_with_command() in src/distillers/mod.rs
• Deleted unused distillers: infra.rs, log.rs, tabular.rs
• Updated all distiller tests to use command strings instead of ContentType, updated snapshot outputs for docker/kubectl/nginx workflows
• Rewrote collapse.rs to use CollapseMode enum instead of ContentType for content-aware collapsing
• Removed src/pipeline/composer.rs, moving rewind/learning logic directly into hook code

CLI & Stats Updates

• src/cli/stats.rs: Removed --by-type mode, added --project flag for per-project stats, reworked default view to show top commands, updated JSON output to return command stats instead of content-type data
• Replaced hardcoded $3/1M token cost calculation with config-driven pricing via new src/guard/config.rs
• Renamed all Indonesian-named test functions in init.rs, session.rs, stats.rs to English
• Deleted old truncate_commands() function and associated tests
• Updated bench targets in benches/pipeline.rs from bench_classify to bench_registry_resolve testing scorer::score_with_command

Hook & Runtime Changes

• Rewrote src/hooks/pipe.rs and src/hooks/post_tool.rs: removed classifier calls, added auto-sibling command detection for pipe mode, updated rewind/route logic and auto-learn triggers
• Added src/guard/config.rs for loading user pricing/config settings
• Updated src/main.rs to detect pipe commands via env vars, added background DB cleanup thread
• Added new MCP omni_search tool in src/mcp/server.rs
• Removed Intelligence Consistency check from src/cli/doctor.rs

---

Notes

• Distiller snapshot outputs have been updated for docker, kubectl, and nginx workflows to match new command-based distillation logic

---

Breaking Changes

1. Public API Break: Removed pipeline::classifier::classify, distillers::get_distiller(), and deprecated ContentType enum usage across core code
2. CLI Break: Removed --by-type flag from omni stats; JSON output now returns commands instead of content_types
3. Distiller Trait Break: Distiller trait no longer includes the content_type() method
4. Hook Logic Break: Pipe mode now requires command context, and pipeline results no longer include ContentType fields

Last updated: 2026-04-12 04:30:58