full-send

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Full-send

Full-send

Read the rules of every route and apply them throughout. Use the status line after each phase. Take all build context from the plan folder. For a different request in the build thread, use the route selection in the linked skill.
阅读每个 route 的规则,并在整个过程中遵循这些规则。每个阶段结束后使用状态行。构建上下文全部取自 plan 文件夹。如果 build thread 中有其他请求,使用所链接 skill 中的 route 选择。

Preflight

预检

Stop if the t3-code MCP tools are absent. Check the sibling paths linked by this skill and its references. If one is missing, stop and print
bunx --bun skills add 0xr3ngar/bstack
. Load the setup settings in order: the request,
~/.config/bstack/preferences.md
, then the defaults and tool table. Use
orchestrator_capabilities
for each named model and effort. If a required model is unavailable, report the choices to the originating lead for a decision.
Read request.md, plan.md, plan-details.md and state.md. If the user requested plan-first review, verify their explicit go-ahead for the current plan before starting or resuming implementation. If it is missing, return the plan to the originating thread and wait. Use the state record and legacy migration. Read the origin and plan-reviewed base. If the plan or internal review is incomplete, have the originating thread complete that phase and resume automatically. Preserve any existing build progress. Recover missing files or thread IDs from the plan and launch records before reporting a blocker.
Get this thread's ID with
t3_thread_configure
. If done, return the recorded results. If an explicit pause is still in effect, preserve it. If another live thread has the handoff, identify it to the origin so it can collect the running build. Otherwise record this thread if needed and resume from the last completed step. Honor an explicit user stop. Accept a resume relayed by the originating thread.
Read the plan documents, repository agent, contributing and writing rules, and the proof procedure in research.md. Split by the plan's PR table, preserving order and bases. Write one build brief per PR in build/.
如果 t3-code MCP tools 不存在,则停止。检查此 skill 及其 references 链接的同级路径。如果缺少其中任何一个,则停止并输出
bunx --bun skills add 0xr3ngar/bstack
。按顺序加载 setup 设置:请求内容、
~/.config/bstack/preferences.md
,然后是 defaults and tool table。对每个指定的 model 和 effort 使用
orchestrator_capabilities
。如果所需 model 不可用,将可选方案报告给来源 lead,由其决定。
读取 request.md、plan.md、plan-details.md 和 state.md。如果用户要求先审查计划,则在开始或恢复实现前,核实用户是否已明确批准当前计划。如果没有明确批准,将计划退回来源线程并等待。按照 state record and legacy migration 操作。读取 origin 和 plan-reviewed base。如果计划或内部评审尚未完成,让来源线程完成该阶段,然后自动恢复。保留已有的构建进度。在报告 blocker 之前,先从计划和启动记录中找回缺失的文件或线程 ID。
使用
t3_thread_configure
获取当前线程的 ID。如果状态为 done,返回已记录的结果。如果显式暂停仍然有效,则继续保持暂停。如果另一个活动线程持有 handoff,则将其标识告知 origin,以便其接管正在运行的构建。否则,如有必要记录当前线程,并从最后完成的步骤恢复。尊重用户明确发出的停止指令。接受由来源线程转达的恢复请求。
读取计划文档、repository agent、贡献与写作规则,以及 research.md 中的验证流程。按照计划的 PR 表拆分任务,并保留顺序和 base。为每个 PR 在
build/
中编写一份 build brief。

Build

构建

Use the PR owner procedure in both modes. Run the provider usage check before starting builders. If a design decision fails in the code, write evidence in build/ and return it to the originating lead. The lead resolves it from the feature request or reruns internal design review and resumes the build. Escalate only a genuine blocker or consequential choice the evidence cannot settle. Diagnose failures with root cause fix.
Record each transition as
pr <n> <branch> <thread or inline> <state>
. Use the newest line for each PR.
Threads. Start independent trunk PRs at once. For a dependent PR, wait for acceptance of the one below, then add its link to the brief. Use
t3_thread_launch
with a worktree from the assigned base and the implementer model. Send the first message from the route rules with
Route: change
and
Brief: <absolute path>
. Record started. Wait with
t3_thread_wait
, including delegated work. Read the report with
t3_thread_read
.
Record handed-back. Check the diff against the plan and, for a stacked PR, verify the branch sits on the current head below. Record accepted if both hold. Otherwise record started and send the correction with
t3_thread_send
, prefixed with
Route: change
. Wait again. Send fixes or restacks only after handback.
Inline. Build each PR yourself, in order, following the same procedure. Write the same state transitions. Check the diff and ancestry before accepting it.
Restack. After a lower head changes, restack each dependent PR lowest first. Record started before sending the instruction or running
gh stack
. In threads mode each PR owner rebases onto the new lower head, checks, pushes with
--force-with-lease
and hands back. In inline mode run the restack yourself and repeat the checks and handback.
Resume. Skip accepted PRs if the lower head is still their ancestor. Restack otherwise. For handed-back PRs inspect the diff. For started PRs in threads mode, send “continue” with the route line once the thread is idle and wait. In inline mode check out the recorded branch and resume at the first incomplete PR owner step. Record
built <prs>
once accepted.
在两种模式下都使用 PR owner procedure。启动 builder 前,运行 provider usage check。如果代码暴露出未决的设计决策问题,将证据写入
build/
并返回给来源 lead。lead 根据 feature request 解决该问题,或重新进行内部设计评审并恢复构建。仅在证据无法解决的真正 blocker 或重大选择时升级。使用 root cause fix 诊断失败。
每次状态转换都记录为
pr <n> <branch> <thread or inline> <state>
。每个 PR 以最新一行记录为准。
Threads. 同时启动相互独立的 trunk PR。对于 dependent PR,等待其下方 PR 被接受后,再把其链接添加到 brief。使用
t3_thread_launch
,传入基于所分配 base 的 worktree 和 implementer model。按照 route 规则发送首条消息,其中包含
Route: change
和
Brief: <absolute path>
。记录 started。使用
t3_thread_wait
等待,包括 delegated work。使用
t3_thread_read
读取报告。
记录 handed-back。根据计划检查 diff;对于 stacked PR,还要验证该分支以下方分支的当前 head 为基点。如果两项检查均通过,则记录 accepted。否则记录 started,并使用
t3_thread_send
发送更正消息,消息以
Route: change
为前缀。再次等待。仅在 handback 之后发送修复或 restack 指令。
Inline. 亲自按顺序构建每个 PR,并遵循相同流程。记录相同的状态转换。接受 PR 前检查 diff 和 ancestry。
Restack. 下方 head 发生变化后,从最低层开始 restack 每个 dependent PR。发送指令或运行
gh stack
前先记录 started。在 threads 模式下,每个 PR owner rebase 到新的下方 head,执行检查,使用
--force-with-lease
push,然后 handback。在 inline 模式下自行执行 restack,并重复检查和 handback。
Resume. 如果下方 head 仍是已接受 PR 的 ancestor,则跳过这些 PR;否则执行 restack。对于 handed-back PR,检查 diff。对于 threads 模式中的 started PR,线程变为空闲后,发送带 route 行的“continue”,然后等待。在 inline 模式下,检出已记录的分支,并从第一个未完成的 PR owner 步骤恢复。全部 accepted 后记录
built <prs>
。

Verify and review

验证与评审

For each stack, verify its top. In threads mode fetch and detach the lead checkout at the pushed head. In inline mode check out the top branch. Run every planned case on the real surface from research.md. Record the runs and gaps in build/verify.md. Record verified only for demonstrated behavior. If the repository has no real surface, say what remains unproved.
Run swarm in code mode on the whole change. Give the exact head and commit range per PR. Reviewers use read-only git. Fetch and update the brief with new heads before each later round. Return accepted fixes to the PR owner, mark started, restack dependents after handback and verify affected behavior again. Record each round's scores in state.md.
Return the draft PR URLs, heads, checks, every round's scores, real-surface results and remaining gaps to the originating thread. Include them in the last message for the waiting lead. Record done when complete. Leave the PRs in draft. Leave merges, ready state, reviews, posts and resolutions to the user. Later PR feedback follows the upkeep route.
对于每个 stack,验证其顶部。在 threads 模式下,fetch lead checkout,并将其 detach 到已 push 的 head。在 inline 模式下,检出顶部分支。在 research.md 指定的 real surface 上运行每个计划用例。将运行结果和缺口记录在
build/verify.md
。只有实际证明的行为才能记录为 verified。如果 repository 没有 real surface,说明仍未得到证明的部分。
对整个变更以 code 模式运行 swarm。为每个 PR 提供准确的 head 和 commit range。Reviewer 使用只读 git。在每一轮后续评审前,fetch 并用新的 head 更新 brief。将已接受的修复交回 PR owner,标记 started,在 handback 后 restack dependent PR,并再次验证受影响的行为。将每轮的分数记录在 state.md。
将 draft PR URL、heads、checks、每轮的分数、real surface 结果和剩余缺口返回给来源线程。在等待中的 lead 的最后一条消息中也包含这些信息。完成后记录 done。让 PR 保持 draft 状态。将 merge、ready 状态、reviews、posts 和 resolutions 留给用户处理。后续 PR 反馈遵循 upkeep route。