Skip to main content

WIP=1

WIP=1 means that only one atomic unit of work can be active at a time. It is the harness's core discipline and its main safeguard against model drift.

Cognitive constraint for language models​

A model's context window behaves like a developer's working memory:

  • When several objectives compete, attention shifts between scopes and restrictions become easier to miss.
  • With WIP=1, the model works against one bounded contract and knows what it may touch and how completion will be proved.

With versus without WIP=1​

With WIP=1Without WIP=1
One feature is completed with real evidenceMany features are started and none finishes
Blockers are recorded with an exact causeBlockers pile up without a record
Each session's scope is clear and predictableThe session jumps between tasks and creates chaotic diffs
Handoff describes exactly one next actionHandoff lists several vague items "in progress"

How it works​

The feature_list.json keeps the state of each feature:

{
"id": "feat-001",
"status": "in_progress",
"title": "Health endpoint"
}

Only one feature can be in_progress. When it finishes, it goes to passing (with evidence) or blocked (with an exact reason).

During execution​

If you discover adjacent work during a task:

  1. Add the new task as not_started in feature_list.json.
  2. Keep working only on the active task.
  3. The new task will be pulled in the next session.

Do not expand the scope silently. If the scope grows, stop, update the sprint contract, and ask for approval.

OpenCode and Oh My Pi with WIP=1​

  • Shared contract: WIP=1 limits each active unit and requires disjoint ownership when lanes run in parallel.
  • OpenCode: feature_list.json and progress.md preserve the active unit and next action between sessions.
  • Oh My Pi: todo keeps the active unit and the scheduler dispatches independent lanes through tasks[].

Overlapping lanes must be serialized or merged. Ownership is a prompt rule, not a lock; use worktrees when real isolation is required.

Next step​

See Evidence and Handoff to understand what counts as completion.