SwarmVault

Pick a task and get started

Install the CLI and initialize a separate directory. The first task supplies demonstration material, the second needs an already compiled codebase, and the third adds a proposed update to the first task’s vault.

1. Install the published CLI and check its version

   Copy install command

   Check node --version for Node.js ≥24 first. This installs the CLI globally; it does not install skills, MCP configuration or hooks into every project. The guide pins verified published release 3.20.0, while the main branch declares 3.21.0.

   npm install -g @swarmvaultai/[email protected]
   swarmvault --version

   Read the author’s setup guide

   Keep the vault separate from its inputs

   Copy install command

   Create an empty directory where the generated files should live, initialize it with the commands below, then read swarmvault.schema.md and swarmvault.config.json to confirm the default heuristic routing. If SWARMVAULT_OUT is set, check its destination: generated directories move there while configuration and the schema stay at the root. The first task can use the CLI directly. For automatic assistant discovery, separately run swarmvault install --agent claude --scope project, followed by swarmvault install status --agent claude --scope project. That install may also edit the project’s .gitignore and a strict-JSON tsconfig.json.

   mkdir my-vault
   cd my-vault
   swarmvault init
   swarmvault next

   Read the author’s initialization and agent setup guide

2. Give your AI this first task

   Copy prompt

   These are editorial examples. Replace bracketed fields with real paths or targets. Author-documentation URLs are included for assistants without the companion skill; install the CLI first. Run commands from the vault root and take all source and artifact IDs from actual output.

   Use SwarmVault 3.20.0 in the empty experimental vault initialized above. Read https://github.com/swarmclawai/swarmvault/blob/4ce0c7cb545c669f3cd35fc6d0300c6b088a52ea/skills/swarmvault/examples/quickstart.md, https://github.com/swarmclawai/swarmvault/blob/4ce0c7cb545c669f3cd35fc6d0300c6b088a52ea/skills/swarmvault/references/commands.md and https://github.com/swarmclawai/swarmvault/blob/4ce0c7cb545c669f3cd35fc6d0300c6b088a52ea/skills/swarmvault/references/artifacts.md, plus this vault’s schema and config. Confirm default heuristic routing with no cloud model or embedding service configured. Create three Markdown files in a new ../demo-inputs directory beside the vault. Label them “editorial demonstration data” and do not overwrite existing files: - overview.md: Paperbridge is a demonstration project that summarizes weekly project changes. Its owner is Lin Qing. - schedule.md: Paperbridge creates an internal draft every Tuesday at 15:00, Asia/Shanghai. The owner checks it before it is sent. - open-questions.md: Paperbridge’s document retention period has not been decided. From the vault root, run: swarmvault ingest ../demo-inputs --no-guide swarmvault compile swarmvault next swarmvault graph validate --strict Then run swarmvault query separately for “Who owns Paperbridge?”, “When does Paperbridge create its internal draft?” and “How long does Paperbridge retain documents?”. Keep the default save behavior. Check import counts and warnings, raw/sources, wiki/sources, wiki/graph/report.md, state/graph.json, state/retrieval, and the query output paths actually returned. Heuristic output lists relevant pages and excerpts; do not present it as completed deep reasoning. Deliver the original query output and a separate source-check table: Lin Qing is the owner, the schedule includes its timezone, and retention remains undecided. Trace each item to the corresponding original. Graph validation checks structure, not the truth of these facts. Mark unrun or failed steps honestly instead of writing a success record.

3. Check what you receive

   • The three ingested documents, wiki pages, graph and retrieval index in a separate vault, with actual paths.
   • Three saved query outputs and an item-by-item source check for the owner, schedule and unknown retention period.
   • Import warnings, graph structure validation and unresolved issues, separating command success from content verification.
   Read the author’s notes for this task