parallel-impl-agent-dies-mid-stream-verify-working-tree

What problem does it solve? When you fan out multiple parallel implementation agents and one stalls mid-stream, the harness may still report "completed (exit code 0)" even though the agent wrote nothing. Integrating on that false status ships a hole in your codebase. This Skill provides a reconciliation protocol to detect dead agents and recover correctly. ## Core Features & Use Cases - Claimed-vs-actual reconciliation: Check git status and grep each agent's target file for the expected change before integrating, instead of trusting the completion status. - Three recovery variants: Covers the shared-file case (finish the dead agent's task yourself), the checkpointed-pipeline case (reconcile against origin/<branch> and open PRs, re-dispatch only missing stages, never use resumeFromRunId), and the reviewer/analyst case (SendMessage a finalize-only resume to the same agent id). - 3-strike escalation rule: If the same stage dies three times, stop re-dispatching and do it in the main loop in small, individually verified steps. - Use Case: You dispatched 3 parallel agents to edit disjoint files; one returns API Error: Response stalled mid-stream with a "completed" notification. Before committing, you grep each target file, find one agent wrote nothing, and finish its task yourself in the main loop. ## Quick Start Ask the assistant to verify each parallel agent's claimed work against the actual working tree with git status and grep before integrating their combined output.