Every gateway turn's prompt prefix injected three near-duplicate absolute paths:
currentTaskWorkspace + localWorkspace + remoteWorkspaceHint. localWorkspace is the
App's LOCAL thread dir (~/.xworkmate/threads/...) which the gateway agent cannot
access, and remoteWorkspaceHint duplicates currentTaskWorkspace. The conflicting
paths leave the agent unsure where to work and can block conversation continuation.
For gateway turns the prompt now carries only currentTaskWorkspace (the plugin owns
the artifact scope); localWorkspace is kept only for non-gateway (local agent runs
there); remoteWorkspaceHint is dropped when equal to currentTaskWorkspace. sessionKey
is kept (short, not a path). UI is unaffected (chat bubble shows the raw user message;
the prompt-debug parser only special-cases Execution context / Preferred skills /
Attached files). Tests updated; assistant_execution_target_test green (74).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Symptom: a gateway turn shows "任务运行中..." forever and 停止 has no effect,
even though the OpenClaw gateway has already finished (ACP_HTTP_CONNECTION_CLOSED).
- T3: add a hard deadline to the running-handle poll branch so the client no
longer polls forever when tasks.get keeps returning "running". Budget is
derived from taskLoadClass (10/30/60min, aligned with the bridge) + grace;
on timeout the turn lands in a recoverable `interrupted` state
(OPENCLAW_RUN_POLL_TIMEOUT) prompting the user to resend.
- T4: make 停止 locally authoritative — capture the association, mark the turn
aborted immediately (clears pending, exits the poll loop), then fire
tasks.cancel best-effort so a hung/failed cancel RPC can't block termination.
- T6: applyGatewayChatFailureInternal now authoritatively clears the pending
flag (both raw + normalized key). Previously runOpenClawGatewayQueuedTurnInternal's
finally never cleared it, leaving "error shown but still running".
Full cross-repo analysis + remaining TODO in docs/cases/06.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
