Three commands in.
Decisions on every prompt.

01 / Set up once

One command. Everything local.

1. 1Run the installer inside your repository

   Inside a git repo it installs the onecue CLI and runs onecue install — which writes the skill, the recall hooks, and the local store for each agent it detects. Nothing else to wire up.

   $curl -fsSL https://onecue.sh/install.sh | sh

   Decisions: .onecue/decisions/ (git-visible) · Logs: ~/.onecue/projects/<fingerprint>/ · Hooks: .claude/settings.local.json + .devin/config.local.json · Skill: .claude/, .devin/, .agents/ (Codex), .cursor/, .gemini/, .opencode/ — each under skills/onecue

   Prefer npm? npm i -g onecue-cli, then onecue install. Skill only, no hooks? npx skills add onecueai/onecue-skill.

   Devin also reads .claude/ hooks by default — set read_config_from.claude: false in ~/.config/devin/config.json if you want a single source.

02 / Use it

Capture once. Recall when it matters.

1. 2Save something worth keeping

   From the terminal — or just tell Claude Code “remember that we chose Postgres for the job queue” and the skill captures it.

   $onecue remember "Keep the job queue in Postgres — order and job commit atomically." --kind decision
   • decisiona choice with the reason it was made, and what it rules out
   • insighta non-obvious lesson or root cause
   • open_loopunfinished work or a pending constraint
   • discoveryan undocumented fact about the repo or system
   • referencean external pointer with a known future use

2. 3Let recall happen

   With hooks installed, each new prompt gets the decisions that bear on it — usually one — silence when nothing applies. Check what would surface, or browse the store, any time:

   $onecue recall "add retries for failed jobs"

   onecue list · onecue show <id> · onecue forget <id>

03 / Verify

Check the setup before blaming the memory.

onecue status shows which agents were detected and which have hooks; the doctor verifies the whole chain.

1. 4Run the doctor

   Runtime, store, per-agent hooks and skill, and decision-file integrity. Exits non-zero on warnings, so it fits CI.

   $onecue doctor

   No CLI? The companion skill ships the same checks as a script: onecue-doctor

Reference

Every command.

onecue init

Create the git-visible .onecue/decisions/ store

onecue install

Install skill + hooks for detected agents [--harness claude,devin,cursor,codex]

onecue uninstall

Remove only OneCue's hooks (decisions are kept)

onecue status

Initialization, hooks, and decision counts

onecue statusline

One-line status for the agent status bar (decisions, cues this session, local)

onecue doctor

Diagnose the setup, decision graph, and spend report [--json]

onecue bench

Run the decision-selection benchmark [--json]

onecue remember <text>

Record a decision [--kind …] [--reason …] [--files a,b] [--rejected] [--rules-out a,b] [--supersedes <id>] [--conflicts-with <id>]

onecue recall <context>

Print the cue this context would surface [--json]

onecue list

List decisions [--kind <k>] [--json]

onecue show <id>

Show one decision (id or prefix)

onecue forget <id>

Tombstone one decision (file kept, excluded from recall)

onecue hook <event>

Invoked by the agent harness, not by hand

One canonical store: with the CLI, decisions live in .onecue/decisions/ and commit with the repo. Without it, the skill falls back to .onecue/memories/. Session logs stay in ~/.onecue/projects/. onecue doctor checks both.