Decisions, not transcripts

Every 10 minutes Gari reads the new turns in your CLI logs and keeps only what mattered: decisions, open questions, your corrections, and what went well. Each card quotes your own words — the AI's claims never become memory.

reads: ~/.claude/projects (Claude Code)
writes: store/cards/YYYY-MM-DD.jsonl
model: distill_model: "haiku"

$ tail -4 store/cards/2026-10-08.jsonl
{"type": "decision", "text": "Pricing page offers one plan, monthly only",
 "quote": "monthly only for now", "tool": "claude", "project": "launch-site"}
{"type": "decision", "text": "Annual plan waits until 50 paying users",
 "quote": "annual after 50 users", "tool": "claude", "project": "launch-site"}
{"type": "pending", "text": "FAQ section is still empty — needs copy before launch",
 "quote": "do the faq later", "tool": "claude", "project": "launch-site"}
{"type": "correction", "text": "Don't add an annual tier to the pricing table",
 "quote": "no, drop the annual one", "tool": "claude", "project": "launch-site"}

# launch-site

Identity: marketing site for the v1 launch

Why it exists: first place new users land from the waitlist email

Success criteria: UNDEFINED — no metric in the cards yet

Valid decisions

One plan, monthly only
Annual plan waits until 50 paying users

Open pendings

FAQ section is still empty

store/wiki/launch-site.md · UNDEFINED slots become tomorrow's question

$ gari northstar
Gari north star — zero re-explaining (last 7 days vs previous 7)
  Re-explanations (repeat cards): 2  (prev 7)  ↓ good
  Context handoffs (briefings + answers): 38  (prev 21)  ↑ good

the one number every feature is judged against · example output

A report waiting at 9

Each morning Gari turns yesterday into what got decided, what's still open, and one step for today — plus a short review held to a senior product bar, not to yesterday's you.

when: report_hour: 9 (launchd)
where: reports/2026-10-09.md
model: deep_model: "sonnet"

reports/2026-10-09.md

Gari morning report — 2026-10-09

Overnight tidy-up is done. 14 cards from yesterday.

Decided

[launch-site] One plan, monthly only
[api] Rate limit stays at 60 req/min per key

Pending — pick these up today

(c41f09a2) FAQ section is still empty

Today's one step (north star)

Next step: write the three FAQ answers people asked about in the waitlist replies
Why: the launch is blocked on copy, not code

Mentor review

Critique: the pricing decision came before anyone wrote down who it's for. Today's drill: when you lock a price, write the buyer in one line first.

gari resolve c41f09a2 — close a pending item by its 8-character ID

Ask the fish

Click the pet and ask. Memory questions come back in seconds with where they came from. Planning questions go to a deeper model that lays out the forks before recommending one — and ends with a nudge when you're skipping something.

memory: ask_model: "haiku"
judgment: deep_model: "sonnet" + 11 methodology sources
also: gari ask "…" · Telegram gateway

Gari's chat panel answering a memory question with a nudge

Hand it off. It checks the work.

Say what you want, approve with "go". Gari packs the context into instructions, hands it to Claude Code in an isolated git worktree, then opens the real files before calling it done. Merging stays your call.

isolation: git worktree, never your working copy
done means: output verified in the repo
never: auto-merge, auto-push

zsh — ~/Projects/launch-site

$ gari do "add the three FAQ answers from the waitlist replies" --write
Gari: handing ~/Projects/launch-site to claude (write allowed)…
Delegated — I'll report with a notification when it's done. (background)
 
[worktree] The output is in an isolated copy (original untouched).
           After review, one `gari merge wt-20261009-1012` merges it.
 
$ gari merge wt-20261009-1012
Merged: gari/wt-20261009-1012 → launch-site (copy removed, branch cleaned up). Pushing is up to you.

Runs on the Claude API

One key in secrets.env. Text-only work — distilling, memory answers, judging nudges, wikis, the morning review — goes straight to the Messages API. Work that needs tools runs through Claude Code on the same key. Aliases like haiku resolve to the newest model.

metered: store/usage.jsonl, per call
budget: monthly_budget_usd warns once
config: prices_per_mtok for dollar totals

zsh — ~/gari

$ gari cost 7
Gari LLM usage — last 7 days: 412 calls, $3.8120
  ask-deep      31 calls  $1.4410
  distill      226 calls  $0.9032
  mentor         7 calls  $0.5121
  wiki          18 calls  $0.4380
  ask           96 calls  $0.3307
  nag-judge     34 calls  $0.1870
 
$ tail -1 store/usage.jsonl
{"kind": "distill", "model": "claude-haiku-…", "via": "api", "sec": 2.1,
 "tokens": {"in": 6182, "out": 311, "cache_read": 4096}, "ok": true}

illustrative numbers — your own appear after a day of use

Your AI coding sessions forget.
Gari keeps what you decided.