Claude Code

Yoke supports Claude Code CLI, desktop Code, and the editor surface. Shared project instructions come from AGENTS.md; Yoke skills live in .agents/skills/yoke/ and Claude discovers them through .claude/skills/yoke. See Native discovery for the supported engine floor and configuration checks.

Start work

Open a session in the project folder and run /yoke onboard after setup. Use /yoke steer next to review direction and staff work. The command reference lists the other Yoke skills.

Run yoke doctor run --session to check this session's registration, Claude sign-in, and permission to run Yoke commands. Follow its named recovery. Claude can load hooks installed during a conversation on the next prompt; the readiness check tells you whether a new session is needed.

The CLI's own authentication is separate from Yoke sign-in. Check it with claude auth status --json; use claude auth login when signed out. Setup can find Claude's bundled CLI even when it is absent from PATH.

Keep the built-in AGENTS plugin enabled. A conflicting CLAUDE.md can suppress shared instructions under Claude's default configuration; the discovery guide explains how to load both without duplicating Yoke rules.

Messages and launches

Hooks deliver orientation and messages through the shared Yoke engine. Supported launch, resume, and wake behavior varies by surface: the installed .yoke/harness/claude/manifest.json owns those capabilities. CLI worker launches use the CLI surface; desktop chats require operator-driven re-entry. See Sessions.

Long commands and waits

Run yoke watch <wrapper> --print-streaming-pair ... first, then run the command it prints. background-wake arms the command and a Monitor tail; in-turn runs foreground Bash with the printed timeout. yoke watch --help lists the wrappers. Main sessions and subagents follow different steps; if you are unsure which you are, follow the subagent steps.

A Monitor defaults to five minutes. It is capped at thirty minutes in interactive sessions and ten in headless -p runs (agent_wake in the manifest). When it expires, re-arm the same tail: the command is still running, so never start it again beside itself.
Relay launches set BASH_DEFAULT_TIMEOUT_MS and BASH_MAX_TIMEOUT_MS before Claude starts (cli.launch_environment in the manifest), and subagents inherit them. Exporting them inside Bash is too late: relaunch instead.
An announced QA-wait cadence means silence past it is a stuck run; the deadline still ends the wait. With no wrapper, Monitor a filter over the captured progress and failures, read it at most once a minute, and file a note about the missing wrapper.
Stop only a local check that has run past about one minute: stop it cleanly, keep its partial capture, commit and let CI run it. Never stop a CI watcher.
To keep a claim through a stop, arm the Monitor first, then run yoke sessions touch --mode parked --reason claim as the turn's last action.
A subagent runs foreground Bash through to the end. It never uses background Bash, Monitor or ScheduleWakeup; if the launch settings are missing, it asks its parent to relaunch.

Hooks and session recovery

Every settings.json hook entry uses the nested {hooks: [{type, command}]} form; one flat entry turns off all hooks. Check with yoke watch doctor -- --only project-hook-config-validity and repair the whole file, not one entry.
A passing SessionEnd (sleep, reload, disconnect) releases no claims; your next hook event brings the session back. If your claims did not come back, run yoke claims work acquire --item PREFIX-N --reason session-reactivation-recovery.
Stopping a session needs authority: yoke sessions terminate SESSION-ID --reason R. Keep a session with no claim alive with yoke sessions keepalive hold.
Read evidence for this session's current episode only with yoke events query --session SESSION-ID --current-episode. A nonzero elided_prior_episode_rows means read again without the flag; it does not mean the events are absent.