PaperClipAI/tests/runner-e2e/chat-scenarios/steering.yaml

# yaml-language-server: $schema=./catalog.schema.json
# Planning scenarios; not registered executions or evidence of a pass.
schema_version: 1
feature: steering
status: planned
scenarios:
  - id: steering.queue-follow-up
    capabilities:
      - steering/queue
    context: >-
      Morgan is busy on the linked website task and cannot immediately consume a normal
      follow-up.
    query: >-
      Also add a contact section to the website.
    expected_outcome: >-
      The requirement is saved once in the correct task queue with its order and waiting state
      visible. Chat does not claim it has already been consumed. When the task becomes eligible,
      the normal continuation receives the saved requirement without dropping earlier work.
  - id: steering.edit-queued-message
    capabilities:
      - steering/edit-order-discard
    context: >-
      The person owns an undispatched queued message asking for a red page; queue mutations
      require their board session.
    query: >-
      Change my queued request from red to blue, and put it first.
    expected_outcome: >-
      The person reaches the supported queue editing controls with the intended message
      identified. The updated full text and order persist, while already-dispatched or
      concurrently changed messages reject stale edits visibly. The agent does not forge a human
      message edit through its own comment API.
  - id: steering.discard-queued-message
    capabilities:
      - steering/edit-order-discard
    context: >-
      A user-owned queued contact-section request has not been dispatched.
    query: >-
      Cancel the contact-section request I just queued.
    expected_outcome: >-
      The supported human queue control removes only that eligible pending message. History and
      queue state make clear that it will not be delivered. If it has already dispatched, the
      system says so and treats cancellation as new direction instead of pretending the consumed
      input vanished.
  - id: steering.steer-active
    capabilities:
      - steering/steer
    context: >-
      The active harness advertises same-turn steering; the person selects the supported Steer
      control for their saved message.
    query: >-
      Switch to a blue design while you are working.
    expected_outcome: >-
      The explicit steering action delivers the selected message to the compatible active run,
      with delivery state distinct from mere queue admission. Subsequent work reflects blue. If
      the provider rejects steering, the message and failure remain visible with a supported
      continuation option instead of silent loss.
  - id: steering.interrupt-with-direction
    capabilities:
      - steering/interrupt
      - steering/queue
    context: >-
      A task is running; the person uses the supported interruption control after saving new
      direction.
    query: >-
      Stop this attempt and start again with the simpler layout.
    expected_outcome: >-
      The current execution is stopped through its authorized control and accepted direction
      reaches one eligible continuation after cleanup. The old run's evidence remains
      inspectable. The system never runs both attempts as concurrent owners or delivers the new
      requirement twice.
  - id: steering.stop-versus-pause
    capabilities:
      - steering/stop-and-pause
      - task-coordination/tree-control
    context: >-
      A website task is running under a parent launch task. The person wants sustained
      suspension, not just cancellation of one turn.
    query: >-
      Pause all work on the launch until tomorrow.
    expected_outcome: >-
      The agent identifies the parent scope and guides the person to the supported task-tree
      pause control; a future resume needs a supported explicit schedule or later action. The UI
      distinguishes requested stop, confirmed stop and the persistent pause. Chat does not claim
      that one stopped run prevents future work by itself.
  - id: steering.preserve-draft
    capabilities:
      - steering/drafts
      - documents-artifacts/attachments
    context: >-
      The person types the query and uploads a brief but has not sent it; they switch tasks and
      reload before returning.
    query: >-
      Build the page from this brief, but keep the copy short.
    expected_outcome: >-
      The correct task's unsent draft and completed upload receipt survive the navigation
      without becoming an instruction on the other task. Sending saves one user message and
      correct attachment references. An uncertain submission is reconciled with persisted
      history before any resend duplicates the work.
  - id: steering.approval-delivery
    capabilities:
      - steering/approval-queue
      - questions-and-approvals/continuation
    context: >-
      A question is answered while ordinary comments wait in the queue and the worker is
      settling its prior turn.
    query: >-
      Use Acorn; that answers your naming question.
    expected_outcome: >-
      The bound answer is retained as an interaction response with its own delivery semantics,
      separate from editable queued comments. Normal admission delivers it once to eligible
      waiting work. Editing, reordering or deleting an ordinary comment cannot revoke or replay
      the accepted decision.