SaulMoro · Oct 24, 2026 · Oct 23, 2026 · Oct 23, 2026 · Oct 23, 2026 · Oct 24, 2026
diff --git a/docs/dev/workflow-engine.md b/docs/dev/workflow-engine.md
diff --git a/docs/guide/workflow-advanced.md b/docs/guide/workflow-advanced.md
@@ -715,11 +715,11 @@ teamai workflow cancel <run> [step]
 teamai workflow resume <run> --tools default=claude
 ```
 
-`retry` accepts a leased, failed, or cancelled step, a refused approve with an answer, or an answered choose. Retry a HITL request to change its answer; its old answer stays in pass history. It frees an old lease and starts a new pass at a new revision. The local driver stops a command it was running and starts the new pass only after that command has exited. It resets the steps that depend on or read that step, including loop waits and item `after` dependencies. A HITL request among them that waits, or was answered and has not run, is asked again at a new pass. It reopens settled enclosing workflows, items, fan-out steps, and the run. If the reopened run is still running, the result includes a `resume` command. Run it to continue locally, or let your driver poll `next`. Unanswered skips in each affected container are re-derived; manual cancellations outside it stay cancelled. An independent standing end or break prevents retry.
+`retry` accepts a leased, failed, or cancelled step, a refused approve with an answer, or an answered choose. Retry a HITL request to change its answer; its old answer stays in pass history. It frees an old lease and starts a new pass at a new revision. The local driver stops a command it was running and starts the new pass only after that command has exited. It resets the steps that depend on or read that step, including loop waits and item `after` dependencies. A HITL request among them that waits, or was answered and has not run, is asked again at a new pass. It reopens settled enclosing workflows, items, fan-out steps, and the run. A step whose pull request delivery failed keeps its pass, and only its delivery runs again. If the reopened run is still running, the result includes a `resume` command. Run it to continue locally, or let your driver poll `next`. Unanswered skips in each affected container are re-derived; manual cancellations outside it stay cancelled. An independent standing end or break prevents retry.
 
 History stays in `passes` and the archive, so `length(steps.x.passes)` counts failed attempts too. Retrying a loop owner resets its iteration count. Retrying another step in its range keeps the owner's count. Retry keeps the top-level workflow's round count, so retry in its exhausted final round can fail `loop_exhausted` again.
 
-A single-step `cancel` stops pending, waiting, leased, or running work. Its open descendants stop too. Cancelled steps count as stopped, so dependents skip unless an `if` reads their outcome for recovery. The run normally fails. A whole-run cancel ends immediately with run outcome `cancelled`, without starting another loop round. Cancel does not bump the revision. A late report returns `lease_cancelled`; `next.cancelled` tells drivers which tools to stop.
+A single-step `cancel` stops pending, waiting, leased, running, or delivering work. Its open descendants stop too. Cancelled steps count as stopped, so dependents skip unless an `if` reads their outcome for recovery. The run normally fails. A whole-run cancel ends immediately with run outcome `cancelled`, without starting another loop round. Cancel does not bump the revision. A late report returns `lease_cancelled`; `next.cancelled` tells drivers which tools to stop.
 
 `resume` drives the run locally after a retry, response, Ctrl-C, or closed terminal. `workflow start --driver local` prints the run id on stderr when the run starts, with the `resume` command that continues it. It uses this machine's open local lease workspace first, then the workspace of the latest driver that called `next` when that driver was local on this machine, then the current directory. Resuming after a retry from a subdirectory therefore keeps the original local workspace. Only one local driver can drive a run at a time. A second resume returns `driver_running` with the owning process id; stop that driver or wait. Locks from dead processes are reclaimed automatically. A settled run returns `done`. Failed and cancelled local runs exit 1; successful control commands exit 0 even when their answer describes a failed or cancelled run.
 

diff --git a/docs/guide/workflow.md b/docs/guide/workflow.md
diff --git a/docs/guide/zh-CN/workflow-advanced.md b/docs/guide/zh-CN/workflow-advanced.md
@@ -715,11 +715,11 @@ teamai workflow cancel <run> [step]
 teamai workflow resume <run> --tools default=claude
 ```
 
-`retry` 接受 leased、failed 或 cancelled 步骤，也接受已回答并拒绝的 approve 和已回答的 choose。重试 HITL 请求可以更改答案，旧答案保留在 pass 历史中。它释放旧租约，以新修订号启动新 pass（本地驱动会停止正在运行的命令，并等它退出后才启动新 pass），重置依赖或读取该步骤的后续步骤，包括循环等待和 item 的 after 依赖（其中正在等待、或已回答但尚未运行的 HITL 请求会在新 pass 中重新询问），并重新打开已结算的外层工作流、item、fan-out 步骤和运行。如果重新打开后运行仍在进行，返回结果会包含 `resume` 命令。执行该命令继续本地运行，或让驱动轮询 `next`。每个受影响容器中未回答的跳过步骤会重新计算；范围外手动取消的步骤保持取消。无关步骤已到达的有效结局或 break 会阻止重试。
+`retry` 接受 leased、failed 或 cancelled 步骤，也接受已回答并拒绝的 approve 和已回答的 choose。重试 HITL 请求可以更改答案，旧答案保留在 pass 历史中。它释放旧租约，以新修订号启动新 pass（本地驱动会停止正在运行的命令，并等它退出后才启动新 pass），重置依赖或读取该步骤的后续步骤，包括循环等待和 item 的 after 依赖（其中正在等待、或已回答但尚未运行的 HITL 请求会在新 pass 中重新询问），并重新打开已结算的外层工作流、item、fan-out 步骤和运行。交付 pull request 失败的步骤保留原来的 pass，只重新运行交付。如果重新打开后运行仍在进行，返回结果会包含 `resume` 命令。执行该命令继续本地运行，或让驱动轮询 `next`。每个受影响容器中未回答的跳过步骤会重新计算；范围外手动取消的步骤保持取消。无关步骤已到达的有效结局或 break 会阻止重试。
 
 历史保留在 passes 和 archive 中，因此 `length(steps.x.passes)` 也统计失败尝试。重试循环所有者会重置其迭代数；重试范围内的其他步骤则保留所有者的计数。重试保留顶层工作流的 round，耗尽最后一轮后重试仍可能再次 `loop_exhausted`。
 
-指定步骤的 cancel 停止 pending、waiting、leased 或 running 工作，同时停止尚未结束的后代。cancelled 属于 stopped，因此依赖步骤会跳过，除非其 if 读取取消结果来恢复；运行通常失败。取消整个运行立即以 `cancelled` 结束，不启动下一轮。cancel 不增加修订号。迟到的报告返回 `lease_cancelled`，`next.cancelled` 告诉驱动停止哪些工具。
+指定步骤的 cancel 停止 pending、waiting、leased、running 或 delivering 工作，同时停止尚未结束的后代。cancelled 属于 stopped，因此依赖步骤会跳过，除非其 if 读取取消结果来恢复；运行通常失败。取消整个运行立即以 `cancelled` 结束，不启动下一轮。cancel 不增加修订号。迟到的报告返回 `lease_cancelled`，`next.cancelled` 告诉驱动停止哪些工具。
 
 在重试、回答、Ctrl-C 或关闭终端后，用 resume 继续本地运行。`workflow start --driver local` 在运行开始时就把运行 id 和继续它的 `resume` 命令输出到 stderr。它优先使用本机未结本地租约的工作目录，其次使用最近调用 `next` 的驱动工作目录（该驱动必须是本机本地驱动），最后使用当前目录。因此重试后即使从子目录继续，也会保留原来的本地工作目录。每个运行同时只能有一个本地驱动。重复 resume 返回 driver_running 和持有者进程号，请停止该驱动或等待；死进程留下的锁会自动回收。已结算运行返回 done。本地运行以 failed 或 cancelled 结束时退出码为 1；控制命令成功时退出码为 0，即使返回的运行状态为 failed 或 cancelled。