Crate open_loops 

open-loops: recupera contexto de trabalhos pausados. Architecture: docs/architecture/ (start at 00-overview.md)

Modules

cache

Distillation cache at /cache//@.md. Keying by the HEAD SHA makes the cache self-invalidate when the branch advances.

cli

Command definitions and module orchestration.

config

Config persisted at /config.toml. The base path comes from outside (main resolves OPEN_LOOPS_HOME or ~/.open-loops) so tests can inject a tempdir — nothing here reads environment variables.

distill

Distillation: builds the prompt with evidence (git + sessions) and calls the LLM via a configurable command (default “claude -p”). Injectable command means tests use cat and users can swap LLMs without changing code.

ignores

Loops discarded by the user (“not worth continuing”). Persisted at /ignores.toml, keys in “repo/branch” format.

index

SQLite-backed disposable index for cached scan and session data.

inventory

SHA-validated ahead/behind memo store.

output

Terminal rendering: loop inventory table and human-readable ages.

query

Query parsing and in-memory evaluation. Pure: no git, no I/O. Grammar lives in ADR 0003. This module turns a query string into a ScanPlan and decides whether a candidate loop matches it.

scanner

Repository and unmerged-branch discovery via git shell-out. Design decision: shell-out (not git2/gix) — simple and debuggable; the product performance bottleneck is the LLM, not git.

sessions

AI session sources. Each harness (Claude Code, and future ones like Codex, OpenCode) becomes an adapter of this trait — the rest of the code does not know the session format or location.

state

Runtime state at <base>/state.toml (separate from declarative config.toml). Holds the active @context chosen via the CLI.

worktrees

Worktree inventory: joins git worktree list with merged/idle/state signals.