Vir — Flow & Usage

Pipeline — every vir run

Scanner free

Walks ~/.claude/projects/**/*.jsonl. Computes SHA-256 of each file. Skips anything already in SQLite with the same hash — fully idempotent.

Filter free

Heuristic scorer — no LLM tokens spent. Score based on session length, tool call count, unique files touched, signal words (error, fixed, learned). Sessions below threshold 0.4 are skipped.

Scrubber free

Strips API keys, Bearer tokens, AWS credentials, emails, absolute paths with usernames before anything touches the LLM.

Classify Haiku

Haiku classifies the session: category (pattern / gotcha / decision / tool), topic (2–4 words), project, confidence (0–1). Sessions with confidence < 0.6 are skipped — no Sonnet call.

Distill Sonnet

Sonnet extracts durable knowledge: Summary, What Was Learned, Context, Related. Stored in SQLite content column. Rate-limited with 60/120/240s backoff on 429.

Write free

Writes typed markdown note to Obsidian vault: patterns/, gotchas/, decisions/, tools/. Updates index.md catalog and log.md. Wikilinks injected for Obsidian graph view clustering.

Commands

vir run

Run pipeline once. Processes only new/changed sessions. Used by the launchd daemon every 4 hours.

Haiku + Sonnet

vir run --full

Force reprocess all sessions. Required after schema migrations or first setup. Expensive one-time cost.

vir run --rewrite-only

Rewrite all vault notes from SQLite — no LLM calls. Use after changing writer format (e.g. adding wikilinks).

free

vir query "question"

TF-IDF retrieves top 8 notes, Sonnet synthesizes a direct answer. Citations included.

~$0.01

vir summarize <project>

Generate cross-session synthesis for a project: key patterns, gotchas, decisions, knowledge gaps.

~$0.02

vir lint

Three checks: orphan notes, stale notes (>90d with newer activity), contradiction detection between similar notes.

free + Haiku

vir dedupe

Interactive duplicate detection. Haiku finds candidates, you confirm each merge. Duplicates archived, not deleted.

Haiku

vir sync-claude --dry-run

Preview what would be injected into ~/.claude/CLAUDE.md and per-project CLAUDE.md files. Shows diff: added, upgraded, removed.

free

vir status

Knowledge heatmap: notes by project and category, confidence distribution, gaps, daemon status.

free

vir schedule install

free

Cost reference (Kie.ai)

Operation	Model	When	Cost
Scanner / Filter / Scrubber	—	every session	free
Classify	Haiku 4.5	sessions passing filter	~$0.001/session
Distill	Sonnet 4.6	confidence ≥ 0.6	~$0.02/session
First full run (226 sessions)	Haiku + Sonnet	one-time	~$1.50
Ongoing (4h daemon, ~8 sessions)	Haiku + Sonnet	per run	~$0.05/run
vir query	Sonnet 4.6	per query	~$0.01
vir summarize	Sonnet 4.6	per project	~$0.02
vir dedupe / lint --contradictions	Haiku 4.5	per pair checked	~$0.001/pair