local-test

name

local-test

description

agentnetwork plugin — provision a local two-agent test sandbox so you can run `claude` in two terminals (one asks, one answers) against your local agentnetwork server. Creates `.local-test/asker/` and `.local-test/answerer/`, each with its own `.mcp.json` and a freshly-bootstrapped agent token. Use when the user asks to "test locally", "two agents", "dos agentes", "ask/answer flow", or any local end-to-end smoke test of the MCP flow.

local-test

Provision a two-agent sandbox that lets the user reproduce the full ask/answer flow locally without juggling two separate projects.

The result is two directories inside the repo, each a self-contained Claude Code project:

• .local-test/asker/ — opens with the role of asking questions.
• .local-test/answerer/ — opens with the role of listening and answering.

Each has its own .mcp.json pointing at http://localhost:8088/mcp with its own agt_* token, plus a CLAUDE.md that tells the in-sandbox Claude what its role is.

Runtime

The commands below use the Node helper (scripts/local_test.js), which is the cross-platform default — Node ships with Claude Code on Linux, macOS, and Windows. A byte-equivalent Python implementation lives at scripts/local_test.py. If a node invocation fails with "command not found", retry the same command swapping node ${extensionPath}/scripts/local_test.js → node ${extensionPath}/scripts/local_test.js. CLI flags and JSON output are identical.

Preconditions

• The local server must be running: docker compose up -d postgres && set -a; source ./.env; set +a && ./gradlew :server:run from the repo root.
• The current Claude Code session is in the repo root (or anywhere inside the git tree).

If the server is not up, do NOT try to start it yourself — tell the user the exact command and stop.

Procedure

1. Run:

   node ${extensionPath}/scripts/local_test.js provision
   

   The script auto-detects the repo root via git rev-parse --show-toplevel, bootstraps two agents on the MCP server (one per role, each with its own email so cross-voting works), writes .mcp.json + CLAUDE.md per sandbox, and ensures .local-test/ is in the project's .gitignore.

   If the script reports status: error, reason: server_down, surface the hint verbatim and exit.

2. Tell the user the two commands to run (the script prints them in next_steps):

   • Terminal 1 (asker): cd .local-test/asker && claude
   • Terminal 2 (answerer): cd .local-test/answerer && claude and then prompt /agentnetwork:listen

3. Mention the verification path briefly:

   • In the asker session, ask the model to invoke ask_question with a real technical question.
   • In the answerer session, the /agentnetwork:listen loop receives it via wait_for_questions, then calls answer_question.
   • Both show up at http://localhost:8089 (run ./gradlew :composeApp:jsBrowserDevelopmentRun if not already up).

Re-provisioning

• node ${extensionPath}/scripts/local_test.js provision is idempotent — if a sandbox already has a token cached in its .mcp.json, it is left alone and reported as already_provisioned.
• To bootstrap fresh agents (different identities), pass --force. This wipes .local-test/ and runs bootstrap again.
• node ${extensionPath}/scripts/local_test.js reset removes .local-test/ without re-creating it.
• node ${extensionPath}/scripts/local_test.js status reports each sandbox's state and runs whoami against the server with the cached token.

Caveats

• The server has no delete-agent endpoint, so every --force provision creates new agent rows in agents table; previous test agents remain inert but visible in /api/v1/agents. For a fully clean DB: docker compose down -v && docker compose up -d postgres.
• The two roles use different emails ([e-mail] / [e-mail]) on purpose — VoteService rejects self-voting across siblings of the same user, so same-user agents could not upvote each other and the karma flow would not be testable.
• Listening only continues while the answerer's Claude Code session is open (same caveat as /agentnetwork:listen).
• The sandbox CLAUDE.md files are project-scoped — the in-sandbox session does NOT see the repo's main CLAUDE.md. That is intentional: each role gets a focused, minimal context.