wednesday docs
Docs/Core concepts

How a run works

A model on its own guesses. Wednesday wraps it in a process that cannot skip the checks: plan, build, verify, repair, review, receipt.

The lifecycle#

one run
request
  |
  v
plan -------> objective, workspace, files in scope, verification commands
  |
  v
build ------> engineering seats work in a controlled sandbox with explicit tools
  |
  v
verify -----> real tests + security scan; only declared project scripts
  |
  +-- fail --> repair (capped) --> verify again
  |
  v
review -----> acceptance criteria, diff, diagnostics
  |
  v
receipt ----> persisted locally; returned by wednesday_result

Plan#

The request becomes an objective, a workspace, the files in scope, and the commands that will prove the work. Constraints, acceptance criteria, test commands and required files can arrive with the MCP request.

The plan is not a chat answer. It is the input to the run, and the verification commands are part of it.

Build#

The engineering seats do the work in a controlled sandbox. Tools are explicit and fenced; the model does not get a free-form shell. Real files change, and the event log records what happened.

The seats are architect, planner, coder, tester, reviewer, vision, autocomplete, bug-hunter and security-audit. A crew may seat fewer; unstaffed seats fall back to their routing slot. See Crews & shapes.

Verify#

Verification runs real commands against real files. The model’s opinion about whether it worked does not count. If the workspace declares a gate — npm test, npm run typecheck, tsc, lint — Wednesday runs it, and a failing gate blocks exactly like a failing test.

Only scripts the project already declares are used. Nothing is guessed and nothing is downloaded. The full policy is on Verification.

Repair#

Failures feed back into the crew with a budget. The loop stops when green or when the repair cap or token/dollar budget is exhausted. A repair is not a retry of the same prompt: the failure output is part of the next attempt.

The cap and budgets are crew policy, set with wednesday team set repairs|tokens|usd.

Review and the receipt#

After green, the run returns a verdict: status, summary, files changed, tests, acceptance checks and diagnostics. The event log and task record persist under ~/.wednesday/state.

The receipt is a record, not a claim of correctness beyond what it ran. It tells you which commands passed, which acceptance criteria were checked, and what changed.

Nothing is committed for you. The run leaves the working tree for you to inspect and commit. Over MCP, the team runs unattended inside its workspace.

Why wait exists#

Clients kill a tool call on their own timeout, usually well under a minute. A long run delivered as one blocking call would be reaped and look like a crash. So the surface is start, poll or wait, then result.

the polling contract
wednesday_run  -> job id immediately
wednesday_wait -> up to 55s; returns done: false instead of hanging
wednesday_result -> the verdict, whenever the run is done

wednesday_wait also emits notifications/progress when the request carries a progressToken, so a client can show the team moving.

Limits of the lifecycle#

LimitConsequence
Jobs live in memoryA client or server restart loses job ids. Task records and the event log survive on disk.
One job at a timeA second run waits behind the first. This is deliberate: one state home, one workspace, one event stream.
No approval gate over MCPThe team runs unattended inside its sandbox. A client-side approval gate is future work.
No MCP elicitationThe team cannot ask you a blocking question through MCP yet.