MCP tools
Seven coarse tools, all prefixed wednesday_. There is no tool for a seat, and the test suite asserts that none can appear.
Transport and handshake#
Wednesday speaks MCP over stdio (newline-delimited JSON-RPC 2.0). The client is the interface: it starts the process on demand and speaks the protocol on stdin/stdout.
The server exposes tools, one prompt (run), and progress notifications. It exposes no resources yet; artifacts as resources are an obvious next addition, not part of the current surface.
The run lifecycle#
wednesday_run -> { job_id, state, request, workspace, elapsed_ms, next }
wednesday_status -> { job_id, state, elapsed_ms, progress[], done, result? }
wednesday_wait -> same shape; done: false when the wait window closes
wednesday_result -> same shape; result populated at a terminal state
The wire shape does not change between status, wait and result: the outside model reads it blindly and rewords it for the operator.
| Field | Meaning |
|---|---|
job_id | The in-memory job id. |
state | The job state (queued, running, completed, failed, cancelled). |
request | The operator request, verbatim. |
workspace | The absolute workspace the run is using. |
elapsed_ms | Wall time since the job started. |
done | Whether the run has reached a terminal state. |
progress | The last few team steps, newest last. |
result | The verdict object, when the run is finished. |
error | Present when the job itself failed. |
hermes_task_id | The persisted task id, when one exists. |
wednesday_run#
Hand a request to the team. Returns immediately with a job id and validates the workspace in the same call.
| Argument | Type | Meaning |
|---|---|---|
request | string, required | The operator’s request in their own words. One sentence is enough. |
workspace | string | Absolute path to the project. Defaults to the server’s own directory, which most clients set to the project root. |
role | string | A seat-role hint (backend, frontend, docs...). |
constraints | string[] | Hard rules the team must respect. |
acceptance_criteria | string[] | What counts as done. |
tests | string[] | Commands that must pass before the run counts as verified. |
required_files | string[] | Files the run must touch or produce. |
A bad workspace fails in the same call, before a job is queued. The tool description tells the host model not to attempt the request itself.
wednesday_status#
Cheap and safe to poll: state, elapsed time, and the last few team steps. No side effects.
| Argument | Type | Meaning |
|---|---|---|
job_id | string, required | The id returned by wednesday_run. |
wednesday_wait#
Block until the run finishes or the wait window closes. This is the normal way to follow a run: start, wait, wait, then read the result.
| Argument | Type | Meaning |
|---|---|---|
job_id | string, required | The id returned by wednesday_run. |
timeout_seconds | number | How long to block, at most 55 seconds. Defaults to 45. |
When the window closes, the result is done: false — not an error. Call it again. If the request carries a progressToken, the server emits notifications/progress while it waits.
wednesday_result#
The verdict: status, summary, files changed, tests, acceptance checks and diagnostics. Report it faithfully, failures included.
The tool description also tells the host model not to substitute its own implementation for a failed run. That is a prompt-level guard, not a protocol one: MCP cannot force the host model to behave.
wednesday_cancel#
Stop a queued or running job. The team stops between stages, and the result is marked cancelled. A job cancelled while queued is never submitted.
wednesday_jobs#
This session’s runs, newest first. Useful after a client restart to find a run that is still going. No arguments.
The registry is in memory: a server restart loses the ids, though the Hermes task records and the event log survive on disk.
wednesday_team#
Read-only: the crew the operator selected (seats, models, policy) and whether this server adopted it. No arguments.
Both adopted and selected are reported. If they differ, in_sync: false means the crew changed after the server started — reconnect to adopt it. There is no tool to change the crew.
The run prompt#
Most clients surface the prompt as a slash command:
- OpenCode:
/wednesday:run <request> - Claude Code:
/mcp__wednesday__run <request>
The prompt is the piece that makes the hand-off stick. It tells the outside model, in as many words, not to do the work itself, not to “help” by editing files, and not to replace a failed run with its own implementation.
Errors and edges#
- Missing request — the tool returns a readable error, not a crash.
- Bad workspace — fails in
wednesday_runbefore a job is queued. - Unknown tool or method — a JSON-RPC error, as the protocol requires.
- stdout discipline — the server writes protocol messages only. A stray log line breaks the client; the server test asserts stdout carries nothing else.