open-loops

What did I start and not finish? Where did I leave off? What's the next step?

loops lists your paused work (unmerged branches across all your repos) and reconstructs resume context from your AI sessions and git — without you documenting anything.

All user-facing output is in English: CLI messages, errors, resume sections, and docs.

Install

# via cargo
cargo install open-loops

# via script (Linux/macOS)
curl -fsSL https://github.com/carvalhosauro/open-loops/releases/latest/download/open-loops-installer.sh | sh

# via Homebrew (requires carvalhosauro/homebrew-tap — see docs/distribution.md)
brew install carvalhosauro/tap/open-loops

Quickstart

# 1. register where your repos live
loops init ~/repo

# 2. inventory — open loops, most idle first (<5s, no LLM)
loops
# scanning git repositories…
# LOOP                    IDLE FOR  AHEAD  BEHIND
# my-app/feat/login            12d      3       1
# api/fix/timeout                 2d      1       0

# 3. audit evidence before distilling (no LLM call)
loops resume feat/login --dry-run
# scanning git…
# matching AI sessions…
#
# # my-app/feat/login
#
# **Confidence:** medium — AI sessions found but alignment uncertain — audit Sources before trusting
#
# ## Git
# ...
# ## AI sessions
# - abc123.jsonl (modified 2026-06-18) [in commit window, mentions branch]
#
# ---
# Dry run — LLM not invoked. Run without `--dry-run` to distill.

# 4. resume: why, done, remaining, next step + auditable sources
loops resume feat/login
# scanning git…
# matching AI sessions…
# distilling…
#
# # my-app/feat/login
#
# **Confidence:** medium — AI sessions found but alignment uncertain — audit Sources before trusting
#
# ## Why
# Adding OAuth login to unblock the onboarding flow.
#
# ## Done
# - Token validation middleware
# - Login form wired to `/auth/login`
#
# ## Remaining
# - Refresh token rotation
# - Error states on the form
#
# ## Next step
# Implement refresh token rotation in `auth/refresh.rs` and add a test for expiry.
#
# ## Sources
# - git: branch feat/login (HEAD a1b2c3d)
# - AI session: abc123.jsonl (modified 2026-06-18)

Progress lines go to stderr so you can pipe or redirect stdout without losing the distilled document. Long steps (distilling…) can take ~30–60s on a cold run; repeat calls are instant from cache.

State lives in ~/.open-loops/ — nothing is written inside your repos.

Audit before you trust

Every resume ships a confidence score and a ## Sources section — not metadata for debugging, but the audit trail you use to decide whether to trust the distillation.

Score	Meaning	What to do
`high`	AI sessions overlap branch commits and mention the branch name	Usually safe to continue
`medium`	Sessions matched heuristically	Read Sources; confirm sessions match this branch
`low`	No AI sessions matched — context from git only	Treat as a draft; verify Sources and diff yourself

Recommended flow when confidence is not high:

loops resume <branch> --dry-run — inspect matched commits and sessions (no LLM).
Check ## Sources in the full output — do those commits and sessions belong to this work?
Run loops resume <branch> only when the evidence looks right.

Demo

Record or replay locally:

cargo build --release
./scripts/demo.sh          # runs the quickstart flow in a temp dir
asciinema play docs/demo.cast   # replay the bundled recording

Docs

Full reference in docs/: setup · features · configuration.

License

MIT OR Apache-2.0.