wednesday docs
Docs/Guides

Troubleshooting

Most failures are configuration or expectations, not the engine. Work through these in order before opening an issue.

Your client does not list Wednesday#

  1. Run wednesday setup --dry-run and confirm the client is detected.
  2. Run wednesday setup --client <id> to write it explicitly.
  3. Restart the client. MCP configuration is read at startup.
  4. If the config was reported as skipped, it is not strict JSON. Use the client’s own mcp add command.
  5. If the client spawns .cmd shims badly on Windows, use the absolute command that setup writes, or pass --command "node C:/path/to/Wednesday/bin/wednesday-mcp.js".

The server exits immediately#

Check the Node version: node --version. Wednesday requires 24 or newer. A server that starts and exits is usually a client launching a stale or missing binary, not a protocol failure.

Run the registered command by hand and send it a real initialize over stdin if you need to see the error. Do not add console.log to the server: stdout is the protocol.

Runs fail with no model#

  • Confirm the endpoint answers: curl <baseURL>/models.
  • Check ~/.wednesday/.env for the key and base URL.
  • For a local server, confirm it is listening on the expected port and serving an OpenAI-compatible route.
  • For a tunnel, confirm the URL is current. A stale ngrok URL fails fast; that is the expected behavior.

The crew changed but the run uses the old one#

Crews are adopted when the MCP server starts. A change made while the server is running does not affect that server. Reconnect your client, or restart it, to adopt the new crew.

wednesday_team reports in_sync: false when the running server and the disk disagree. That is the signal to reconnect, not a bug.

wednesday_wait returns done: false#

That is the contract, not an error. Clients reap tool calls under a minute, so wait closes its window at most 55 seconds and returns done: false. Call it again until the run finishes, then read wednesday_result.

The job id is gone after a restart#

Job ids live in memory. A client or server restart loses them. The Hermes task records and the event log survive on disk, so the work and its receipt are not lost. Use wednesday list, wednesday events and wednesday result — or wednesday_jobs in the current session for new runs.

A second run seems stuck#

Wednesday runs one job at a time. A second wednesday_run waits behind the first. This is deliberate: one state home, one workspace, one event stream. Check the first job with wednesday_status or wednesday_jobs.

The run is green but the change is wrong#

Green means no declared verification command failed. If no gate was declared, the tester passes with a note — that is not proof of correctness. Declare tests in the request or set a crew test command:

a real gate
wednesday team set test "npm test"

Where to look, and how to ask#

  • wednesday events --id <taskId> — the event log for one task.
  • ~/.wednesday/state/events.jsonl — the raw log.
  • wednesday result --id <taskId> --json — the receipt as data.
  • wednesday team show — the crew the engine will adopt.

If none of that explains it, open an issue with the version, the client, the command, and the smallest reproduction: github.com/lab1207/Wednesday/issues.