mirror of
https://github.com/paperclipai/paperclip.git
synced 2026-11-16 20:12:20 +01:00
- recovery.explain-stall and recovery.source-action: after three real Runner losses the task continues in one system-owned fresh run from saved work (#21135). The harness waits for that continuation to start, asks the question while it runs, then releases its brief and requires it to finish the original task with its saved result. - recovery.retry-exhausted: infrastructure loss no longer offers Retry, so the case uses a failure only the person can resolve (rejected model) and keeps the held-Retry-under-budget journey. - steering.edit-queued-message and agent-chat.multiple-handoffs: a task_history.read of the current revision by the running turn, with the change saved, an empty queue, a done task and no continuation, counts as consumption (#21175). One continuation still passes when the turn did not read the current revision. - companies.prompt-injection-guard (and fast-responses): wait for the save, check the sent and saved model, then read it back. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>