From b19ddb215e5fb1961ac14c3aa2aeedb9a1378b20 Mon Sep 17 00:00:00 2001 From: John Preston Date: Fri, 28 Aug 2026 02:21:01 +0400 Subject: [PATCH] [ai] Require a dead leaf before relaunching its phase Two runs of one phase append to the same artifact silently, so the surviving verdict is decided by write order instead of judgment. --- .claude/ai-workflow-adapter.md | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/.claude/ai-workflow-adapter.md b/.claude/ai-workflow-adapter.md index 720128072d..5f025ae706 100644 --- a/.claude/ai-workflow-adapter.md +++ b/.claude/ai-workflow-adapter.md @@ -58,6 +58,15 @@ This file adapts harness mechanics and removes unnecessary text normalization. same stateful worker, resume that id; never create a duplicate performer or duplicate an agent whose writes may still be in flight. Never launch a nested `claude` process from Bash. +- Before relaunching a phase whose leaf stalled, establish that the original + leaf is finished. A leaf left running still lands its writes, and two runs of + one phase append to the same artifact with no conflict and no warning, so the + verdict that survives is decided by write order rather than by judgment. When + the original cannot be confirmed dead, point the relaunch at a distinct + artifact and reconcile the two deliberately. A phase artifact carrying two + passes that disagree is this failure, not a reviewer changing its mind: + re-read it before acting, and never treat the last block as authoritative + merely because it is last. - If the first real leaf Agent is rejected before work begins because nested delegation is unavailable, use the shared same-session fallback. Do not treat mere presence of the Agent tool as a successful delegation probe.