GOAL
目标Goal
交付订单的目标保持不变。The goal—fulfill the order—stays fixed.
01 · 地图说直行，路口已经封了01 · The map says straight; the road is closed
仓 A 缺货的真实观察已经返回，Agent 应继续执行原计划，还是标记分支失效并生成可追溯的替代步骤？A real observation says warehouse A is out of stock. Should the agent keep following the old plan, or invalidate that branch and create a traceable alternative?
02拖入失败观察，让原计划分支失效并生成替代步骤与修订轨迹Drag in a Failed Observation to Invalidate a Branch and Generate an Alternative with Revision History
计划 v1 → 观察 → 计划 v2plan v1 → observation → plan v2
状态 AMetric A
—
—
状态 BMetric B
—
—
状态 CMetric C
—
—
所有概率、命中率、得分与输出均为固定种子的教学模拟，不代表真实模型。拖回主变量即可复核结论。All probabilities, hit rates, scores, and outputs are fixed-seed teaching simulations, not real model output. Move the main variable back to verify the conclusion.
03把刚才的变化拆成 4 个零件Break the Change into Four Parts
PlanningBRANCH
计划分支Plan branch
仓 A 预留建立在库存假设上。Reserve at A depends on a stock assumption.
OBSERVE
失败观察Failed observation
真实环境推翻原假设。The environment disproves the original assumption.
REVISION
修订轨迹Revision trace
记录失效原因、替代步骤与版本变化。Records invalidation reason, alternative action, and version change.
04最小机制：主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction
MECHANISM
plant+1 = revise(plant, observationt)
拖入观察是唯一变量；它把依赖库存假设的节点标红，并插入替代分支，同时保留
v1 → v2 轨迹。The dropped observation is the sole changing input. It turns the
assumption-dependent node red, inserts an alternative, and
preserves the v1 → v2 trace.
先分解Decompose目标 → 行动分支goal → action branches
读取观察Read observation检验分支假设test branch assumptions
修订计划Revise替代步骤 + 留痕alternative + trace
05注入“凭证过期”，让自动分支完全失效Inject Expired Credentials and Invalidate the Automated Branch
主动制造失败Create a failure边界实验Boundary experiment
替代步骤不是硬重试，而是转人工更新凭证；不能执行的分支必须明确失效。The alternative is not a blind retry; it hands credential renewal to a human. An unexecutable branch must be marked invalid.
尚未执行。先在主交互中观察正常机制。Not run yet. Observe the normal mechanism in the main interaction
first.
✗ 计划生成后就应原样执行到底。A generated plan should run unchanged to the
end.→ ✓ 新观察会让假设失效，计划必须可修订。New observations invalidate assumptions, so plans must be
revisable.
✗ 替代步骤覆盖旧计划即可。An alternative can overwrite the old plan.→ ✓ 要保留失效原因与 v1 → v2 修订轨迹。Preserve the invalidation reason and v1 → v2
trace.
✗ 任何失败都只需重试原步骤。Every failure only needs a retry.→ ✓ 缺货或权限失败可能需要换分支或转人工。Stock or authorization failures may require a new branch or
human handoff.
06 · 一句话带走06 · One line to keep