Co-authored-by: ayattara <alassane.yattara@savoirfairelinux.com> Co-authored-by: Carlos Florencio <carlosmflorencio@gmail.com>
4.2 KiB
status, system, requirements
| status | system | requirements | |
|---|---|---|---|
| current | ui |
|
Backend-owned Cancel-turn Progress System Design
Purpose and boundaries
This design keeps the cancel control tied to the cancellation operation that the backend accepts, instead of to the browser component that requested it. The UI system owns the shared control's rendering and hydration behavior. The orchestrator and agent runtime retain ownership of execution cancellation, session lifecycle, and prompt termination.
Requirement mapping
| Requirement | Design sections |
|---|---|
REQ-UI-CANCEL-TURN-PROGRESS-001 |
Runtime projection and client contract, Control flow, Responsive behavior, Verification design |
Runtime projection and client contract
The orchestrator exposes a per-session cancellation_pending value and monotonically increasing
cancellation_revision. Task-session boot and REST payloads, subscription snapshots, and
session.cancellation_changed notifications carry both values. Clients merge a notification only
when its revision is current, so a delayed hydration response cannot replace a newer transition.
The shared chat cancel control combines this backend state with a short-lived local request flag. The local flag supplies immediate feedback before acceptance; once the backend accepts, the runtime projection is authoritative across remounts, route changes, and replacement pages. The projection is not a durable session lifecycle state and is not persisted across backend restart.
Control flow
- A user activates the existing
agent.cancelaction for a running session. - The backend guards the per-session cancellation operation, publishes pending, and starts or joins lifecycle cancellation.
- The current client and any later task-session hydration render the disabled loading control from the projection.
- After lifecycle and session reconciliation settle, the backend publishes a false value with the next revision; the shared control returns to its ordinary state.
The mock agent recognizes /e2e:cancel-hold in every mock-agent run. Its prompt waits at one
predictable cancellation boundary and produces no assistant output, allowing browser regressions to
observe the backend-owned pending operation without intercepting a browser request. The E2E profile
sets a three-second PromptCancelJoinTimeout baseline through private AgentctlStartupConfig, the
process adapter, and ACP. The two cancel-progress regressions scope a 12-second override through
their backend fixture; a zero value leaves ACP's normal three-second join bound in effect.
Responsive behavior
Desktop task navigation remounts the same shared cancel control after returning to a task. On mobile, the compact composer hydrates the same session projection after a full reload. Both surfaces preserve their existing composition, controls, and touch behavior; no viewport-specific cancellation state or new user-facing copy exists.
Failure and recovery
If a cancel request is rejected before backend acceptance, the local request flag clears and no pending projection is published. A missed live event is repaired by the next boot, REST, or subscription snapshot. On success, reconciliation, rejection, or a bounded cancellation timeout, the backend releases pending so a still-running session is retryable.
The E2E profile scopes the fixture timing and cancellation-join configuration to test runs. Production and non-E2E managed launch paths retain their existing cancellation timing and ACP default join bound.
Verification design
Mock-agent tests prove the hold does not emit assistant output. Configuration and process tests prove the managed E2E timeout reaches ACP and that zero keeps the default. Desktop Chromium repeats prove the disabled loading control remains visible through task navigation; mobile Chromium repeats prove the same behavior through reload hydration. Both flows wait for the authoritative pending value and eventual settlement rather than using a browser delay.