attentive

Attention router for Claude Code. Learns which files matter per prompt, injects them as context via hooks so Claude starts each turn already focused.

How it works

1. Hooks intercept Claude Code lifecycle events:

   • user-prompt-submit — routes attention, injects HOT/WARM files
   • stop — records which files were actually used, trains learner
   • session-start — dashboard, project switch detection

2. Learner builds word→file associations from session history (TF-IDF weighted). After enough data, it predicts which files you'll need before you ask.

3. Router scores every known file per prompt (7-phase pipeline: decay → co-activation → pinned floors → demoted penalty → learner boost → cache stability → truncation).

4. Tiers determine injection strategy:

   • HOT (≥0.8) — full file content
   • WARM (≥0.25) — table of contents (function signatures)
   • COLD (<0.25) — evicted

Install

cargo install --path crates/attentive

Setup

# Initialize config + hooks
attentive init

# Bootstrap learner from existing Claude Code sessions
attentive ingest

Commands

Command	Description
init	Initialize config and install Claude Code hooks
ingest	Bootstrap learner from session JSONL files
benchmark	Measure token reduction on current repo
status	Show config and learner state
diagnostic	Check dependencies and health
history	View turn history with filters
report	Generate token usage report
compress	Compress observations via Claude API
graph	Analyze file dependency graph
plugins	Manage plugins

Workspace crates

Crate	Purpose
attentive	CLI binary and hook implementations
attentive-core	Router, attention state, config, tiers
attentive-learn	TF-IDF learner (word→file associations)
attentive-telemetry	Path resolution, JSONL I/O, turn records
attentive-plugins	Plugin system (burn rate, loop breaker, verify-first)
attentive-index	SQLite index with fastembed semantic search
attentive-compress	Claude API observation compression
attentive-repo	Git repo analysis

State files

All state is project-scoped under ~/.claude/projects/<project-hash>/:

• learned_state.json — learner associations
• attn_state.json — current attention scores
• session_state.json — session metadata

Global config: ~/.claude/attentive.json

License

MIT