You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
When a seller agent's runtime is active (acp serve start), running acp agent switch <other> stops the seller's WebSocket connection to the ACP server. If a job REQUEST is submitted to the seller during this window, the job lands server-side but the seller is offline. When the seller later reconnects (via acp agent switch <seller> → acp serve start), the pending job is not replayed. The job hangs in REQUEST phase indefinitely.
Reproduction
acp agent switch pqs-atlas-seller
acp sell create pqs_atlas_score
acp serve start
In a second terminal: acp agent switch pqs-buyer — seller runtime stops (unexpected side effect of switching the active agent)
Job is created server-side but the seller socket is offline. Job phase stays REQUEST.
acp agent switch pqs-atlas-seller → acp serve start — seller reconnects, but the pending job is not replayed. Verified with job 1003481568, which sat in REQUEST for >10 minutes with a live seller runtime, never progressing.
(b) Seller runtime replays pending REQUEST jobs on reconnect
(c) CLI warns the user when acp agent switch is about to stop an active runtime, with a --force override
Workaround we're using
Bypassed the CLI for job creation. POSTed directly to https://claw-api.virtuals.io/acp/jobs with the buyer's x-api-key while the seller runtime stays online and continuously bound to its active agent. Job 1003481624 completed successfully with this pattern:
Discovered while building a grade-gated prompt-scoring service on ACP for the Anthropic Built with Opus 4.7 hackathon (April 2026). Happy to contribute a PR if the fix direction is clarified — (a) seems like the lowest-risk change.
Environment
Observed behavior
When a seller agent's runtime is active (
acp serve start), runningacp agent switch <other>stops the seller's WebSocket connection to the ACP server. If a job REQUEST is submitted to the seller during this window, the job lands server-side but the seller is offline. When the seller later reconnects (viaacp agent switch <seller>→acp serve start), the pending job is not replayed. The job hangs in REQUEST phase indefinitely.Reproduction
acp agent switch pqs-atlas-selleracp sell create pqs_atlas_scoreacp serve startacp agent switch pqs-buyer— seller runtime stops (unexpected side effect of switching the active agent)acp job create <seller-wallet> pqs_atlas_score --requirements '{"prompt":"...","vertical":"general"}'REQUEST.acp agent switch pqs-atlas-seller→acp serve start— seller reconnects, but the pending job is not replayed. Verified with job1003481568, which sat inREQUESTfor >10 minutes with a live seller runtime, never progressing.Expected behavior
Any of:
acp agent switchpreserves running seller runtimes for non-active agentsREQUESTjobs on reconnectacp agent switchis about to stop an active runtime, with a--forceoverrideWorkaround we're using
Bypassed the CLI for job creation. POSTed directly to
https://claw-api.virtuals.io/acp/jobswith the buyer'sx-api-keywhile the seller runtime stays online and continuously bound to its active agent. Job1003481624completed successfully with this pattern:Context
Discovered while building a grade-gated prompt-scoring service on ACP for the Anthropic Built with Opus 4.7 hackathon (April 2026). Happy to contribute a PR if the fix direction is clarified — (a) seems like the lowest-risk change.