Agent Setup

Jumpspace works best when the coding agent does not need you to remember every command. Install repo-local guidance once, then ask the agent to use Jumpspace before changing code.

Install All Supported Guidance

npx @jumpspace/cli add-skill --all

Today this installs supported Codex and Claude Code guidance. The operation is idempotent and non-destructive: Jumpspace creates missing files, appends managed blocks, or updates clearly marked Jumpspace-managed blocks.

Install One Agent

npx @jumpspace/cli add-skill --codex
npx @jumpspace/cli add-skill --claude

Use --json when another tool needs to inspect exactly what changed.

Install One Pipeline Skill

npx @jumpspace/cli add-skill jumpspace-bootstrap --agent claude
npx @jumpspace/cli add-skill jumpspace-work --agent codex
npx @jumpspace/cli add-skill review --agent claude
npx @jumpspace/cli add-skill intent-review --agent codex

Named installs include the reference jumpspace-workflow skill and the selected pipeline skill. Use this when you want a smaller agent packet for a specific workflow.

Verify Setup

npx @jumpspace/cli release install-doctor --json
npx @jumpspace/cli task doctor --json

release install-doctor checks whether the jumpspace on your PATH matches the current repo build. doctor checks repo-local task memory health.

First Prompt To Try

Read the Jumpspace guidance for this repo. Then run the appropriate Jumpspace
commands to understand existing intent memory before editing code.

Once likely edit paths are known, the agent should usually run:

npx @jumpspace/cli intent check --for <paths...> --json

If the repo intentionally uses the advanced task graph, it can also run:

npx @jumpspace/cli task scan
npx @jumpspace/cli task find <keywords> --json --compact
npx @jumpspace/cli task work <TASK_ID> --json

When CI says changed files match active intents, ask the agent to use the jumpspace-intent-review skill. The agent should run jumpspace intent validate --since <ref> --json and jumpspace intent verify --since <ref> --json, inspect local diff hunks, and report possible_violation only with quoted intent and diff evidence.

What Humans Still Approve

Agents can gather evidence, draft plans, suggest links, and run checks. Humans should still approve new task blocks, durable plans, risky repairs, and any claim that a task is complete.