Seed Test Environment

Creates a sandbox data directory so the user can launch the Superagent dev build against a clean, populated state — without polluting their real install at ~/Library/Application Support/Superagent/.

Inputs

$ARGUMENTS — free-form. Parse it for:

• dir name (first word, optional) — defaults to superagent-test. Used as ~/Downloads/<dir-name>/.
• agent specs (the rest) — comma-separated slug:description pairs to seed. If absent, seed two default agents agent-a and agent-b instructed to use the x-agent (mcp__agents__*) tools on each other.
• --mode=web|electron flag — picks npm run dev (web/Vite) or npm run dev:electron. Defaults to electron.

Examples:

• /seed-test-env → defaults: dir superagent-test, agents A+B, electron
• /seed-test-env xagent-test → dir name only
• /seed-test-env xagent-test triager:Email triager,replier:Reply drafter → custom agents
• /seed-test-env xagent-test --mode=web → web dev server

If $ARGUMENTS is empty or ambiguous, briefly state the defaults you're using and proceed (don't ask). Confirm only if the target directory already exists with files (offer to delete-and-recreate vs abort).

Procedure

1. Locate the source settings

Source priority (settings.json):

1. ~/Library/Application Support/Superagent-Dev/settings.json — prefer this. The dev install has hardcoded direct LLM keys (Anthropic, OpenAI, OpenRouter, etc.) instead of relying on the platform LLM provider, which is flakey.
2. ~/Library/Application Support/Superagent/settings.json — fall back only if Superagent-Dev doesn't exist. The prod install often uses llmProvider: "platform" which depends on a flakey remote auth; tests want direct keys.
3. ~/.superagent/settings.json — last-resort fallback.

Pick the first that exists AND has a populated apiKeys.anthropicApiKey (or another working LLM key). If none have an LLM key, write the settings anyway and surface this loudly in the final summary.

2. Resolve target dir

TARGET=~/Downloads/<dir-name> (expand ~). If it already exists and contains files, prompt the user once: delete & recreate or abort. After confirmation, rm -rf and recreate.

3. Create structure

For each agent, create agents/<slug>/workspace/AGENTS.md. Write them before the first launch of the data dir: the app imports agent directories into its catalog table once, by a data migration when the database is first opened, and does not scan the directory again. An agent directory added to an already-launched data dir will not show up. Important: the AGENTS.md lives inside workspace/, NOT directly under the agent dir. The file-storage layer (src/shared/lib/utils/file-storage.ts:getAgentInstructionsPath) hard-codes that path — putting it elsewhere makes the agent invisible.

AGENTS.md uses YAML frontmatter delimited by ---. Required fields: name, createdAt (ISO string). Optional: description. Body is the agent's instructions. Mirror the format from src/shared/lib/services/agent-service.ts:createAgent. Example:

---
name: Agent A
createdAt: "2026-04-21T12:00:00.000Z"
description: Test agent A
---

You are Agent A. (instructions body)

4. Copy settings.json (curated)

Read prod settings.json and write a filtered copy to <TARGET>/settings.json.

Copy:

• apiKeys (anthropicApiKey, deepgramApiKey, openaiApiKey, composioUserId — whatever is present)
• llmProvider and any provider-specific config
• models (agentModel, browserModel, summarizerModel)
• app — but strip chromeProfileId (path won't apply) and any path-bound fields
• container (lima/docker runner config — without this the container won't start). Always override container.agentImage to superagent-container:latest so the test instance uses the locally-built dev image instead of pulling the published one from ghcr.io/skillfulagents/.... That tag is what npm run build:container produces. If the user has tagged it differently, they'll tell you — otherwise default to superagent-container:latest.
• skillsets (so the test instance has the same skill catalog)
• platformAuth token block if present (so platform LLM provider keeps working)
• auth, voice, shareAnalytics, customEnvVars if present

Skip:

• mounts (host paths won't apply to a fresh test agent)
• chromeProfileId
• Any session-specific or per-agent state

If a key is missing in prod, just omit it — don't invent values. If apiKeys.anthropicApiKey is missing, surface that loudly in the final summary so the user knows to add one.

5. Final output

Show:

1. Tree of what you created (one ls -R or equivalent, terse)
2. Settings keys copied (just the top-level keys, not values — never echo secrets)
3. Launch command — exact one-liner, e.g.:

   SUPERAGENT_DATA_DIR=/Users/iddogino/Downloads/<dir-name> npm run dev:electron
   

4. Notes — anything skipped or missing (e.g. "no Anthropic key in prod settings — add one before testing").

Migrations auto-run on first launch via migrate() in src/shared/lib/db/index.ts — no separate db:migrate step needed.

Implementation notes

• Use Bash + Write tools directly, don't print the commands and stop. Actually create the files.
• Never echo full secret values back to the user — show key names only.
• The dev:electron command is install-electron && electron-vite dev (the first part fetches the Electron binary if missing). The dev (web) command is vite. Both honor SUPERAGENT_DATA_DIR via getDataDir() in src/shared/lib/config/data-dir.ts.
• Don't modify the real prod data dir under any circumstances. All writes go to ~/Downloads/<dir-name>/ only.

When NOT to use

• The user wants to test against their real data — they should just run the app directly.
• The user wants to write Drizzle migrations or run integration tests — those use in-memory SQLite, not a real data dir.