用STATE.md管理AI代理任务交接

YOUR AGENT CAN REMEMBER THE CHAT AND STILL LOSE THE JOB A long conversation can contain the goal, the abandoned plan, three failed fixes and a test result from before the last code change. All of it is available. Some of it is now wrong For a Dots-coordinated coding workflow, I'd keep a small task packet outside the conversation: > goal: what we're trying to finish > constraints: what must not change > decisions: what we chose and why > sources: the files to read now > evidence: what was checked, against which revision > next: the first unfinished step The worker reads that packet, then opens the current source files. It doesn't get a pasted copy of everything the previous session saw In the clip, the test receipt belongs to an older commit. The agent could happily carry on with a green result that says nothing about the current diff Hold the handoff. Rerun the relevant check. Replace the receipt. Keep the goal and decisions Claude Code already has context-management tools. The extra rule I'd add is that saved evidence needs a source revision, not just the word "passed." Try this before the next handoff: "Write STATE.md with the goal, constraints, decisions, current file paths, verification commands and results, and the next unfinished step. Tie test receipts to the code revision they checked. Before resuming, verify those references and flag anything stale. Keep the original requirements linked." A summary can also be wrong. Keep the source pointers so the next session can check it instead of trusting the summary because it is short morph: https://x.com/i/article/2106252350721560576