onecue-cli
v0.4.0
Published
The decision layer for coding agents. Captures decisions — what, why, and what they rule out, plus rejected approaches and discoveries — to the git-visible .onecue/decisions/ store, then surfaces the one relevant cue (or silence) per task. Claude Code and
Maintainers
Readme
onecue-cli
The decision layer for your coding agents — Claude Code and Devin today. It stops your agent re-deciding what you already decided: it captures decisions — what, why, and what they rule out, plus rejected approaches and discoveries — to the git-visible .onecue/decisions/*.md store, then surfaces the one relevant cue (or stays silent) at the start of each task. Everything stays on your machine.
One cue. Or silence.
Install
One command, run inside your project — installs the CLI, then writes the skill, hooks, and store for every detected agent:
curl -fsSL https://onecue.sh/install.sh | shOr by hand:
npm install -g onecue-cli
onecue install # skill + hooks + store for Claude Code / DevinRequires Node 20+. The package installs a single onecue binary.
Quick start
onecue remember "Keep the job queue in Postgres. Order and job commit atomically." --kind decision
onecue recall "add retries for failed order jobs"onecue install detects which agents are on your machine and writes the skill and hooks for each supported one found — .claude/ for Claude Code, .devin/ for Devin (both share the same hook format). Detected-but-unsupported agents are listed as planned; --harness claude,devin overrides detection. With hooks installed, relevant cues surface automatically on every prompt — no manual recall needed.
Devin also reads
.claude/hooks by default. To keep a single source in Devin, set"read_config_from": { "claude": false }in~/.config/devin/config.json.
Commands
| Command | What it does |
| --- | --- |
| onecue init | Create the git-visible .onecue/decisions/ store for this project |
| onecue install / uninstall | Add or remove OneCue's Stop + PostToolUse hooks and skill (decisions are kept) |
| onecue status | Initialization, hooks, skill, and decision counts |
| onecue doctor [--json] | Diagnose setup: runtime, store, per-event hooks, skill, hook targets, corrupt records, measured spend |
| onecue bench | Run the selection benchmark against a labeled fixture (violations / precision / correct silence) |
| onecue remember <text> | Save a decision — --kind, --reason, --files, --rejected |
| onecue recall <context> | Print the one cue this context would surface — --json |
| onecue list / show / forget | Inspect and manage decisions (forget tombstones — keeps the file, marks it status: forgotten, excludes it from recall) |
| onecue hook <event> | Internal hook entrypoint invoked by the agent's Stop / PostToolUse events |
Where it lives
Captured decisions live in the git-visible .onecue/decisions/*.md store — one Markdown file per decision, named <ISO-timestamp>-<kind>-<uuid8>.md, that you can read, diff, review, and commit with your code. forget tombstones a decision (keeps the file, marks status: forgotten) rather than deleting it. Session/event logs and spend accounting stay out of git under ~/.onecue/projects/<fingerprint>/. No account, no cloud, nothing leaves the machine.
Related
- onecue-skill — the zero-install Claude Code skill
- onecue-doctor — ask your agent to diagnose the setup
