SDK/Gateway: align cancel response, terminal event, and wait status

[BunsDev]

Live dogfood finding

During the live OpenMeow SDK cancel/stop dogfood run, the immediate cancel response, streamed terminal events, and run.wait() result did not agree.

Evidence doc:

• https://github.com/OpenCoven/open-meow-sdk/blob/main/docs/live-dogfood-2026-04-29.md
• Commit: 916d755 (docs: add live dogfood findings)

Observed cancel response:

{
  "ok": true,
  "abortedRunId": "4f4390d1-3434-4f78-af24-dbbd66c423d5",
  "status": "aborted"
}

But:

• no run.cancelled terminal stream event was observed during the cancel probe window
• run.wait() normalized to status: "completed" for the same aborted run

Impact

OpenMeow cannot rely solely on streamed terminal events or wait() to recover stop-button UI today.

The dogfood adapter now mitigates this by mapping a successful immediate cancel response to a deterministic synthetic cancelled UI state via reduceOpenMeowCancelResult().

Desired contract

Gateway/SDK should define one stable cancel semantic for app clients:

• successful cancel response should correlate with a terminal run.cancelled event, and/or
• agent.wait after abort should return cancelled / aborted, not completed, and/or
• SDK should explicitly document that the immediate cancel response is authoritative for UI state.

Acceptance

• OpenMeow stop button can transition from active/cancelling to idle from a stable SDK contract.
• run.wait() and terminal run event semantics are documented or aligned for cancelled runs.

[BunsDev]

Upstream OpenClaw PR opened for this contract gap:

• fix(gateway): align sessions abort wait semantics openclaw/openclaw#74751
• Commit: bae93f5407 fix(sdk): align run event and cancel semantics

The PR aligns abort semantics by emitting a lifecycle cancellation event for aborted chat/session runs and preserving the cancelled wait snapshot so agent.wait does not later report ok/completed after abort.

Fresh local verification before push:

• node scripts/run-vitest.mjs run --config test/vitest/vitest.gateway-server.config.ts src/gateway/server.chat.gateway-server-chat.test.ts → 22 passed
• node scripts/run-vitest.mjs run --config test/vitest/vitest.unit.config.ts packages/sdk/src/index.test.ts → 12 passed
• node scripts/run-vitest.mjs run --config test/vitest/vitest.e2e.config.ts packages/sdk/src/index.e2e.test.ts → 3 passed
• pnpm --filter @openclaw/sdk build → build complete

Leaving this issue open until the upstream PR lands and OpenMeow can re-run the stop-button dogfood against the merged SDK/Gateway.

[BunsDev]

Dogfood update from 2026-05-02:

Validated the cancel response, terminal wait status, and UI recovery behavior together.

Evidence:

• sessions.abort/cancel returned a successful abort response.
• OpenMeow reducer recovered the composer to an idle/can-send state immediately after cancel response.
• Follow-up run.wait() normalized the stopped run to cancelled rather than completed.
• Adapter tests and live probe agree on the expected UI-facing state.

Caveat:

• The raw wait payload still has a timeout/RPC-stop-shaped envelope underneath the normalized result. This does not block OpenMeow UI behavior, but it is worth clarifying in Gateway docs/internals.

Assessment:

• The P0 OpenMeow stop/cancel behavior is functionally satisfied at the SDK adapter level.
• Recommend app-shell signoff next.