OpenClaw agent stopped responding after a tool call?

A documented OpenClaw failure can persist a tool call without its matching tool result. The provider then rejects every replay of that session, while the transport and gateway can still look healthy.

Preserve first: Save the affected session file and gateway error before changing history. Record the unmatched tool-call ID, if present; a healthy channel does not prove that the provider accepted the replay.

Match the failure before retrying.

Transport is healthy but the agent is silent

Messages can still reach the channel while the provider rejects the replayed history.

An orphaned tool call poisoned replay

Inspect logs for a tool_use or toolCall ID without the required immediate tool_result.

The same session fails after restart

Restarting alone can reload the same invalid history. Back up and quarantine the affected session before creating a clean one.

Safe recovery path.

Back up the affected session file and workspace before changing session state.

Inspect gateway logs for unmatched tool-call IDs, compaction pairing errors, or a stuck-session warning.

Move the invalid session out of the active session directory, remove only its stale lock if documented, and restart the gateway.

Rebuild the continuation from the workspace, Git state, decisions, failed attempts, and next verification instead of replaying the poisoned transcript.

What the next AI should receive.

Include the gateway and provider error, session identifier, unmatched call/result pair if observed, a backup location, and the last completed workspace operation. Keep session-history diagnosis separate from transport health.

How ShardStitch works with OpenClaw.

ShardStitch treats the transcript as one source, not the source of truth. It rebuilds a clean packet from the workspace and durable state, then labels recovered facts and inferred intent before OpenClaw or another agent resumes.

Source trail.

The OpenClaw report below concerns a specific session replay failure, not every messaging interruption.

FAQ.

Will restarting the gateway repair an orphaned tool call?

Not necessarily. Reloading the same invalid session can reproduce the rejection. Preserve the session and inspect the error before applying version-specific repair guidance.

Should I delete the session or its lock immediately?

No. Back up session and project state first. Change only the affected state when the installed version documents that remedy; do not remove unrelated session files.

Scope note: Back up session files before quarantine or lock cleanup. File names and service commands vary by OpenClaw version and operating system.