What problem does it solve? When a Synapse agent process crashes or is force-killed mid-turn, the session's LangGraph checkpoint can be left in an inconsistent state: dangling tool calls without results, pending graph tasks that re-execute dangerous commands on resume, and a stale transcript projection. This Skill provides a safe procedure to detect that corruption and repair the session so it can be continued cleanly. ## Core Features & Use Cases - Read-only detection: Inspect the head checkpoint of a thread via SqliteSaver to find pending pregel tasks, dangling tool calls, and non-empty graph state without invoking the model. - Checkpoint repair: Seal dangling tool calls with cancellation ToolMessages using repair_thread_after_cancel, then clear residual middleware pending tasks with an END-node state update. - Transcript rebuild: Regenerate the derived transcript.sqlite projection from the repaired checkpoint so the TUI shows the full history. - Use Case: A user reports that reopening a session causes the agent to immediately execute an unrelated kill command. Use this Skill to detect the pending Send(node='tools') task, seal the dangling call with a crash-reason boundary, and rebuild the transcript before continuing. ## Quick Start Ask the agent to run the read-only detection script against the thread's checkpoints.sqlite file to confirm dangling tool calls, then execute the repair and transcript-rebuild scripts and verify the checklist.
