flower-trellis 0.6.13-beta.1 → 0.6.13-beta.3
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +42 -0
- package/enhancements/0.6/.agents/skills/trellis-check-all/references/findings-and-reporting.md +2 -2
- package/enhancements/0.6/.agents/skills/trellis-check-all/references/repair-and-recheck.md +1 -3
- package/enhancements/0.6/.agents/skills/trellis-check-all/references/reporting-and-disposition.md +4 -4
- package/enhancements/0.6/.agents/skills/trellis-push/SKILL.md +4 -4
- package/enhancements/0.6/.agents/skills/trellis-task-brief/SKILL.md +6 -1
- package/enhancements/0.6/.claude/skills/trellis-check-all/references/findings-and-reporting.md +2 -2
- package/enhancements/0.6/.claude/skills/trellis-check-all/references/repair-and-recheck.md +1 -3
- package/enhancements/0.6/.claude/skills/trellis-check-all/references/reporting-and-disposition.md +4 -4
- package/enhancements/0.6/.claude/skills/trellis-push/SKILL.md +4 -4
- package/enhancements/0.6/.claude/skills/trellis-task-brief/SKILL.md +6 -1
- package/enhancements/0.6/overrides/conflicts.json +8 -7
- package/enhancements/0.6/overrides/patches/skills/trellis-brainstorm/planning-handoff/content.md +1 -1
- package/enhancements/0.6/overrides/patches/workflow/intent-routing/phase-summary/guardrail-content.md +1 -1
- package/enhancements/0.6/overrides/patches/workflow/phase-ownership/phase-2-check-content.md +2 -2
- package/enhancements/0.6/overrides/patches/workflow/states-planning/common-content.md +1 -1
- package/enhancements/0.6/overrides/patches/workflow/task-brief-review/phase-1-activate-content.md +2 -2
- package/enhancements/0.6/scripts/confirmation.md +24 -0
- package/enhancements/0.6/scripts/confirmation.py +352 -0
- package/enhancements/MANIFEST.json +4 -2
- package/package.json +5 -4
- package/src/assets/flower_session_start.py +3 -3
- package/src/builtin-plugins/skill-garden/content-adapter.js +3 -0
- package/src/builtin-plugins/skill-garden/uninstall.js +43 -5
- package/src/cli.js +16 -5
- package/src/commands/confirm.js +217 -0
- package/src/commands/uninstall.js +2 -1
- package/src/lib/config-preserver.js +1 -1
- package/src/lib/confirmation-audit.js +27 -0
- package/src/lib/confirmation-claude.js +165 -0
- package/src/lib/confirmation-files.js +89 -0
- package/src/lib/confirmation-snapshot.js +79 -0
- package/src/lib/confirmation-state.js +253 -0
- package/src/lib/confirmation-witness.js +318 -0
- package/src/lib/copy-scripts.js +12 -5
- package/src/lib/global-trellis-sync.js +2 -10
- package/src/lib/npm-node-shim.js +36 -0
- package/src/lib/platform-patch-adapters.js +7 -0
- package/src/patches/bundles/flower-platform-integration.json +1 -0
- package/src/patches/platforms/claude/confirmation-hooks/patch.json +113 -0
package/README.md
CHANGED
|
@@ -124,6 +124,48 @@ flower-trellis -v
|
|
|
124
124
|
|
|
125
125
|
未指定平台时,交互模式会弹出多选菜单(默认勾选 Claude Code + Codex);也可直接传 `--claude` / `--codex` / `--cursor` / `--devin` / `--zcode` / `--trae` / `--omp` / `--grok` / `--kimi` / `--snow` 等指定,或用 `-y` 跳过菜单。`--windsurf` 仍作为 Devin 的旧别名透传给 Trellis。其余未识别的 flag(如 `-u`、`-f`、`--template`、`--with-statusline`)一律透传给 Trellis。
|
|
126
126
|
|
|
127
|
+
## 限时确认
|
|
128
|
+
|
|
129
|
+
Codex 和 Claude Code 共用 `ftl confirm`。走到确认点时,助手先展示完整方案、默认动作和等待时长,再调用命令。项目策略默认开启,日常无需手动调用命令。
|
|
130
|
+
|
|
131
|
+
- Linux Claude Code 2.1.293:普通安装或更新写入项目 Hook,照常运行 `claude`。现有 Hook、MCP 和权限规则保留。
|
|
132
|
+
- Linux Codex 0.167.0:使用原版 Codex 和原版 App Server。共享连接先接入本机连接见证,即使同一会话只开一个窗口,也用原连接判断关窗。普通 Flower 安装或更新不改现用连接。
|
|
133
|
+
|
|
134
|
+
Codex 连接见证由本机操作者显式启动。使用上游 App Server 的 Unix socket 绝对路径和私有目录里的空闲监听路径:
|
|
135
|
+
|
|
136
|
+
```bash
|
|
137
|
+
ftl confirm witness --upstream /绝对路径/app-server.sock --listen /私有目录/flower.sock
|
|
138
|
+
codex --remote unix:///私有目录/flower.sock
|
|
139
|
+
```
|
|
140
|
+
|
|
141
|
+
若本机服务已把见证安排在普通 Codex 的控制入口,日常仍用 `codex`。不要覆盖运行中服务的 socket。未接见证、非交互、未受信任的项目或尚未验证的平台保持人工确认。
|
|
142
|
+
|
|
143
|
+
Brief 默认开工,Check-All 默认修复当前全部待处置项并重检,Push 沿用已展示的提交和推送计划,明确 commit-only 时仅本地提交。Spec 未决事项和新业务取舍仍由用户决定。
|
|
144
|
+
|
|
145
|
+
在 `.trellis/config.yaml` 调整秒数;省略字段采用下列默认值,更新会保留该配置:
|
|
146
|
+
|
|
147
|
+
```yaml
|
|
148
|
+
confirmation_timeout:
|
|
149
|
+
enabled: true
|
|
150
|
+
brief:
|
|
151
|
+
enabled: true
|
|
152
|
+
base_seconds: 100
|
|
153
|
+
threshold_chars: 1000
|
|
154
|
+
increment_chars: 500
|
|
155
|
+
increment_seconds: 30
|
|
156
|
+
check:
|
|
157
|
+
enabled: true
|
|
158
|
+
base_seconds: 100
|
|
159
|
+
increment_seconds: 30
|
|
160
|
+
push:
|
|
161
|
+
enabled: true
|
|
162
|
+
seconds: 200
|
|
163
|
+
```
|
|
164
|
+
|
|
165
|
+
Brief 超过1000字后每500字加30秒,不足一档按一档;文件首尾空白不计字。检查第一条100秒、之后每条加30秒,零条不等待。项目或单类 `enabled: false` 可关闭;损坏配置和超出工具预算会保持人工确认,不截短等待时间。
|
|
166
|
+
|
|
167
|
+
在领取前提交回复、按 Esc、修改材料或关闭原窗口,都会使旧决定失效。原窗口与共享服务断开时也会取消,重开不会补跑。已开始的动作按实际结果报告。最小审计保存在项目 gitignored `.trellis/.runtime/confirmations/`。
|
|
168
|
+
|
|
127
169
|
## 并行工作区
|
|
128
170
|
|
|
129
171
|
先查看创建计划,再使用返回的指纹确认。任务参数可选,省略时只创建和准备工作区:
|
package/enhancements/0.6/.agents/skills/trellis-check-all/references/findings-and-reporting.md
CHANGED
|
@@ -63,7 +63,7 @@
|
|
|
63
63
|
|
|
64
64
|
按“接受后的增量展示”在对话中报告。不得为缩短回复、留存结论或同步任务状态新建 `check-report.md` 等附件或用链接代替应展示内容。
|
|
65
65
|
|
|
66
|
-
|
|
66
|
+
仅用户明确导出或交付约定时保存附件。Maven、auto-loop 沿用机器证据契约,不另建 Markdown;已有报告不删除或复制。
|
|
67
67
|
|
|
68
68
|
interactive 完成检查与允许的 DOC 修复后,先自然说明实际改动、行为影响和结论,再展示证据与问题:
|
|
69
69
|
|
|
@@ -157,7 +157,7 @@ interactive 完成检查与允许的 DOC 修复后,先自然说明实际改动
|
|
|
157
157
|
|
|
158
158
|
## 修复与重检
|
|
159
159
|
|
|
160
|
-
用户选择 `修复全部`、混合精确问题 ID
|
|
160
|
+
用户选择 `修复全部`、混合精确问题 ID 或其它明确修复范围,或取得有效 check claim 后,必须先读取 `references/repair-and-recheck.md` 执行修复续行;人工路径未选择时不加载。
|
|
161
161
|
|
|
162
162
|
修复完成后输出:
|
|
163
163
|
|
|
@@ -1,8 +1,6 @@
|
|
|
1
1
|
# Repair And Recheck
|
|
2
2
|
|
|
3
|
-
|
|
4
|
-
|
|
5
|
-
用户选择 `修复全部`、混合精确问题 ID 或其它明确修复范围后:
|
|
3
|
+
仅在用户选定当前报告修复范围或取得有效 check claim 后读取;否则保持只读。
|
|
6
4
|
|
|
7
5
|
本修复续行优先于新需求分类和 Active Task Scope Guard:归属依据是受检事项、报告证据和用户选定范围。问题跨模块、属于既有缺陷或未写入任务标题/Brief,不单独构成新需求;不得因此新建或纳管任务,也不得重走首次 Brief 确认。必要材料同步沿用 `trellis-task-brief` 的增量规则。
|
|
8
6
|
|
package/enhancements/0.6/.agents/skills/trellis-check-all/references/reporting-and-disposition.md
CHANGED
|
@@ -6,7 +6,7 @@
|
|
|
6
6
|
- validated auto-loop:读 `references/auto-loop-return.md`,匹配 action 的 record 成功后立即 next;不渲染交互报告。
|
|
7
7
|
- 普通 light strict pass:只执行下方简短格式,不读取上述细则。
|
|
8
8
|
|
|
9
|
-
|
|
9
|
+
默认只在对话中报告,不新建报告附件;确认材料只进 gitignored runtime。导出须用户明确要求或交付约定;Maven 与 runner 沿用机器证据。不得用链接代替应展示内容。
|
|
10
10
|
|
|
11
11
|
## Light 通过报告
|
|
12
12
|
|
|
@@ -40,10 +40,10 @@ untracked helper 不存检查证据或风险接受。通过且按下方 Gate 默
|
|
|
40
40
|
|
|
41
41
|
## Interactive Post-Check Stop Gate
|
|
42
42
|
|
|
43
|
-
非 validated auto-loop
|
|
43
|
+
非 validated auto-loop 先按本文件输出标准报告或接受状态的简短确认;第 2 项使用超时时先冻结再展示,不重贴:
|
|
44
44
|
|
|
45
45
|
1. 先识别当前事项最新有效的明确停止要求:`只检查`、`只读检查`、`检查后停止`、`暂不收尾`、`仅接受、不继续` 等。停止收尾,不运行 Update-Spec 或生成 Git 计划;同一事项后续明确继续可解除该限制。普通“检查一下”不等于只检查后停止,不从无关历史或摘要猜测限制。
|
|
46
|
-
2. 未处置 `CHK-*` / `FBK-*`、blocked、阻断型部分验证或其它未接受实质风险(`[上线后验证]`
|
|
46
|
+
2. 未处置 `CHK-*` / `FBK-*`、blocked、阻断型部分验证或其它未接受实质风险(`[上线后验证]` 除外)时,不运行 Update-Spec 或生成 Git 计划。没有停止/只读要求、全部可修复时,按 `.trellis/scripts/confirmation.md` 用 `ftl confirm` 冻结、展示报告和全部待处置 ID。匹配 `check / repair_all_and_recheck` 后修复、重检并记录结果;新业务取舍暂停。无有效 claim 时报告并停止等待人工选择;任何超时均不接受风险。其余普通代码问题始终由用户选择修复范围,Git 请求本身不授权修复。
|
|
47
47
|
3. 没有有效停止要求且 strict pass,或全部 findings 已有效接受时,默认同轮继续;通过条件沿用已加载 findings reference 的风险接受规则和本文件验证阶段,完整登记的 `[上线后验证]` 不阻断。报告或简短确认后同轮进入 Update-Spec,`no-op|written` 再到 Push,`needs-review` 停止。无需 direct Git intent 或额外回复“继续”,通过不授权 Git 写入。
|
|
48
48
|
4. 只从当前完成链证据识别 direct Git intent,保留 push / 用户主动 `commit-only` 选择;不从无关历史、摘要、dirty 状态或 auto-loop action 推断。风险接受沿用 findings reference 的证据规则。
|
|
49
49
|
|
|
@@ -51,7 +51,7 @@ untracked helper 不存检查证据或风险接受。通过且按下方 Gate 默
|
|
|
51
51
|
|
|
52
52
|
所有 interactive 标准报告都必须在末尾输出 `### 下一步`,并按以下首个命中分支给出一个明确主动作:
|
|
53
53
|
|
|
54
|
-
1. 有未处置 findings:提示 `修复全部`、精确 ID、接受当前报告全部风险、`接受风险 <ID> 并继续` 或
|
|
54
|
+
1. 有未处置 findings:提示 `修复全部`、精确 ID、接受当前报告全部风险、`接受风险 <ID> 并继续` 或 `仅保留报告`;不逐项重复确认。受控分支说明默认动作和时长后按协议等待,不能写成已修复。
|
|
55
55
|
2. 有 blocked、阻断型部分验证或未标记 `[上线后验证]` 的实质风险:指出所需决策、授权或验证,完成后重跑 Check-All;不自行执行生产、外部或破坏性操作。
|
|
56
56
|
3. 有有效的明确停止要求且检查通过:说明已按要求停止,之后明确继续才恢复收尾。
|
|
57
57
|
4. 无有效停止要求且 strict pass / 已接受风险通过:说明本轮正在进入 `trellis-update-spec`,不要求用户再次回复“继续”;Git 计划确认仍由 Push 执行。
|
|
@@ -5,7 +5,7 @@ description: "按确认的精确文件范围提交普通变更或完成已就绪
|
|
|
5
5
|
|
|
6
6
|
# Trellis Push
|
|
7
7
|
|
|
8
|
-
`trellis-push` 是 Phase 3.4
|
|
8
|
+
`trellis-push` 是 Phase 3.4 唯一的代码提交入口。负责最小计划、精确提交推送、task 进度与 untracked 清理。
|
|
9
9
|
|
|
10
10
|
## 职责与条件入口
|
|
11
11
|
|
|
@@ -21,6 +21,7 @@ description: "按确认的精确文件范围提交普通变更或完成已就绪
|
|
|
21
21
|
| 普通 push 且有活动 task | [task-progress.md](references/task-progress.md);计划前读取,业务 Git 后执行 |
|
|
22
22
|
| taskStatus=completed 的恢复 | [completed-task-recovery.md](references/completed-task-recovery.md);先完成只读 preflight |
|
|
23
23
|
| validated runner 的内部 commit-only | [auto-loop-commit.md](references/auto-loop-commit.md);不 push、不执行 Step 5、不渲染交互模板 |
|
|
24
|
+
| 计划确认 | `.trellis/scripts/confirmation.md` |
|
|
24
25
|
|
|
25
26
|
未命中场景不加载其 reference;命中却缺失细则时阻塞,不凭记忆代替。内部 commit-only 只接受 runner 的合法预授权。
|
|
26
27
|
|
|
@@ -114,9 +115,8 @@ git rev-list --reverse @{u}..HEAD 2>/dev/null || true
|
|
|
114
115
|
|
|
115
116
|
确认前禁止 `git add`、`git commit` 或 `git push`。
|
|
116
117
|
|
|
117
|
-
|
|
118
|
-
|
|
119
|
-
reference 缺失、无法读取或缺少对应章节时停止并报告 `阻塞`,不得凭记忆重建、缩写或自制替代模板。
|
|
118
|
+
即将展示用户可见计划时,读取 `references/output-templates.md` 的“计划模板”和“共用展示规则”;缺失或不可读则 `阻塞`,不得凭记忆重建、缩写或自制替代模板。
|
|
119
|
+
按 `.trellis/scripts/confirmation.md` 使用 `ftl confirm` 冻结和展示完整计划;收到匹配 `push / execute_git_plan` 的领取结果才执行,否则等人工。
|
|
120
120
|
|
|
121
121
|
普通多仓只确认一次。计划已展示生成命令和预计 exact files 时,命令成功且没有出现预计列表外的新 dirty path 就沿用原确认;内容、hash 或统计变化不重问。其它计划边界变化仍按 Step 4 重新规划。
|
|
122
122
|
|
|
@@ -13,6 +13,10 @@ description: "从最新 prd.md、design.md、implement.md 生成、刷新、校
|
|
|
13
13
|
|
|
14
14
|
本例外沿用 runner 的 prepare、Open Questions、readiness 和 artifact 校验,不扩张任务或授权边界。schema 1 返回的 `refresh_brief` 同样在展示后交回 runner,由后续 `confirm_brief` action 等待人工确认。run 已停止、终态、损坏或 action/task 不匹配时不得免确认或代为推进;由 auto-loop owner 处理诊断。普通交互调用继续遵守以下确认规则。
|
|
15
15
|
|
|
16
|
+
## 原生会话超时确认
|
|
17
|
+
|
|
18
|
+
首次评审没有明确预授权、停止要求或未决事项时,读取 `.trellis/scripts/confirmation.md` 核验当前原生工具与项目策略。用 `ftl confirm` 冻结最新 Brief、在对话完整展示并前台等待;只有匹配 `brief / implement` 的领取结果,才在同轮启动任务。能力缺失或领取失败时照常展示并等人工。auto-loop 优先使用上方 runner 例外。
|
|
19
|
+
|
|
16
20
|
## 核心规则
|
|
17
21
|
|
|
18
22
|
- 每次运行都重新读取最新 `prd.md`、`design.md if present`、`implement.md if present`。
|
|
@@ -21,7 +25,7 @@ description: "从最新 prd.md、design.md、implement.md 生成、刷新、校
|
|
|
21
25
|
- 不要在 `brief.md` 里发明三件套没有表达的新需求。必填字段缺失时写“未明确”,并提示应补充三件套;没有明确技术机制时省略 `Technical Overview`,没有具体风险时省略 `Risks`。
|
|
22
26
|
- `Goal` 保持一句话,`Scope` 只写本轮纳入的行为或改动对象;延期事项归入 `Non-Goals`,不另设栏目,也不把范围、方案和验收写成同义复述。
|
|
23
27
|
- 首次开工评审或用户明确要求查看全貌时,写回 `brief.md` 后必须展示 brief 正文,不要只给文件路径;首次批准后的同步按下方增量规则展示变化。
|
|
24
|
-
- 首次实施批准前必须展示完整 brief
|
|
28
|
+
- 首次实施批准前必须展示完整 brief。除经 runner 校验的 Auto-Loop Action 或上方有效超时 claim 外,默认等待用户确认后再运行 `task.py start`;只有用户明确把当前任务或最终 Brief 与“展示后直接开始 / 不用再次确认 / 视为已确认”绑定时,才可在范围未变化的前提下免除第二次确认。
|
|
25
29
|
- “开始做吧”“按你建议来”“可以创建任务”等普通实现或建任务意图不是 Brief 预授权,不能据此跳过确认。
|
|
26
30
|
- 普通交互预授权只依赖当前对话中仍然明确可见的用户表达,不建立跨会话永久偏好,也不写 session runtime;auto-loop 恢复由上方 runner 例外所有。
|
|
27
31
|
- `in_progress` 阶段只读取 brief 和任务材料,不例行重新生成、展示或确认;新会话或压缩恢复同样如此。首次批准后的方案变化按下方增量规则处理,用户明确要求查看时再完整展示。
|
|
@@ -69,6 +73,7 @@ description: "从最新 prd.md、design.md、implement.md 生成、刷新、校
|
|
|
69
73
|
6. 写回 `<task>/brief.md`。如果文件已存在,仍用最新三件套派生内容覆盖旧正文。
|
|
70
74
|
7. 在对话中展示 brief 正文,并说明来源文件:
|
|
71
75
|
- 经校验的 auto-loop `refresh_brief`:展示后返回 runner,不进入下方交互分支;schema 1 的确认由后续 action 所有。
|
|
76
|
+
- 原生确认:按上方协议完整展示并等待;只有有效 claim 才同轮返回主 workflow 启动任务,并按实际结果记录 finish。取消或失效时转人工。已在当前对话完整展示的正文不重复输出。
|
|
72
77
|
- 无有效预授权:展示后结束当前回合,等待用户确认。
|
|
73
78
|
- 有有效预授权:先完整展示,再在同一回合返回主 workflow 执行 `task.py start`,不得省略展示步骤。
|
|
74
79
|
|
package/enhancements/0.6/.claude/skills/trellis-check-all/references/findings-and-reporting.md
CHANGED
|
@@ -63,7 +63,7 @@
|
|
|
63
63
|
|
|
64
64
|
按“接受后的增量展示”在对话中报告。不得为缩短回复、留存结论或同步任务状态新建 `check-report.md` 等附件或用链接代替应展示内容。
|
|
65
65
|
|
|
66
|
-
|
|
66
|
+
仅用户明确导出或交付约定时保存附件。Maven、auto-loop 沿用机器证据契约,不另建 Markdown;已有报告不删除或复制。
|
|
67
67
|
|
|
68
68
|
interactive 完成检查与允许的 DOC 修复后,先自然说明实际改动、行为影响和结论,再展示证据与问题:
|
|
69
69
|
|
|
@@ -157,7 +157,7 @@ interactive 完成检查与允许的 DOC 修复后,先自然说明实际改动
|
|
|
157
157
|
|
|
158
158
|
## 修复与重检
|
|
159
159
|
|
|
160
|
-
用户选择 `修复全部`、混合精确问题 ID
|
|
160
|
+
用户选择 `修复全部`、混合精确问题 ID 或其它明确修复范围,或取得有效 check claim 后,必须先读取 `references/repair-and-recheck.md` 执行修复续行;人工路径未选择时不加载。
|
|
161
161
|
|
|
162
162
|
修复完成后输出:
|
|
163
163
|
|
|
@@ -1,8 +1,6 @@
|
|
|
1
1
|
# Repair And Recheck
|
|
2
2
|
|
|
3
|
-
|
|
4
|
-
|
|
5
|
-
用户选择 `修复全部`、混合精确问题 ID 或其它明确修复范围后:
|
|
3
|
+
仅在用户选定当前报告修复范围或取得有效 check claim 后读取;否则保持只读。
|
|
6
4
|
|
|
7
5
|
本修复续行优先于新需求分类和 Active Task Scope Guard:归属依据是受检事项、报告证据和用户选定范围。问题跨模块、属于既有缺陷或未写入任务标题/Brief,不单独构成新需求;不得因此新建或纳管任务,也不得重走首次 Brief 确认。必要材料同步沿用 `trellis-task-brief` 的增量规则。
|
|
8
6
|
|
package/enhancements/0.6/.claude/skills/trellis-check-all/references/reporting-and-disposition.md
CHANGED
|
@@ -6,7 +6,7 @@
|
|
|
6
6
|
- validated auto-loop:读 `references/auto-loop-return.md`,匹配 action 的 record 成功后立即 next;不渲染交互报告。
|
|
7
7
|
- 普通 light strict pass:只执行下方简短格式,不读取上述细则。
|
|
8
8
|
|
|
9
|
-
|
|
9
|
+
默认只在对话中报告,不新建报告附件;确认材料只进 gitignored runtime。导出须用户明确要求或交付约定;Maven 与 runner 沿用机器证据。不得用链接代替应展示内容。
|
|
10
10
|
|
|
11
11
|
## Light 通过报告
|
|
12
12
|
|
|
@@ -40,10 +40,10 @@ untracked helper 不存检查证据或风险接受。通过且按下方 Gate 默
|
|
|
40
40
|
|
|
41
41
|
## Interactive Post-Check Stop Gate
|
|
42
42
|
|
|
43
|
-
非 validated auto-loop
|
|
43
|
+
非 validated auto-loop 先按本文件输出标准报告或接受状态的简短确认;第 2 项使用超时时先冻结再展示,不重贴:
|
|
44
44
|
|
|
45
45
|
1. 先识别当前事项最新有效的明确停止要求:`只检查`、`只读检查`、`检查后停止`、`暂不收尾`、`仅接受、不继续` 等。停止收尾,不运行 Update-Spec 或生成 Git 计划;同一事项后续明确继续可解除该限制。普通“检查一下”不等于只检查后停止,不从无关历史或摘要猜测限制。
|
|
46
|
-
2. 未处置 `CHK-*` / `FBK-*`、blocked、阻断型部分验证或其它未接受实质风险(`[上线后验证]`
|
|
46
|
+
2. 未处置 `CHK-*` / `FBK-*`、blocked、阻断型部分验证或其它未接受实质风险(`[上线后验证]` 除外)时,不运行 Update-Spec 或生成 Git 计划。没有停止/只读要求、全部可修复时,按 `.trellis/scripts/confirmation.md` 用 `ftl confirm` 冻结、展示报告和全部待处置 ID。匹配 `check / repair_all_and_recheck` 后修复、重检并记录结果;新业务取舍暂停。无有效 claim 时报告并停止等待人工选择;任何超时均不接受风险。其余普通代码问题始终由用户选择修复范围,Git 请求本身不授权修复。
|
|
47
47
|
3. 没有有效停止要求且 strict pass,或全部 findings 已有效接受时,默认同轮继续;通过条件沿用已加载 findings reference 的风险接受规则和本文件验证阶段,完整登记的 `[上线后验证]` 不阻断。报告或简短确认后同轮进入 Update-Spec,`no-op|written` 再到 Push,`needs-review` 停止。无需 direct Git intent 或额外回复“继续”,通过不授权 Git 写入。
|
|
48
48
|
4. 只从当前完成链证据识别 direct Git intent,保留 push / 用户主动 `commit-only` 选择;不从无关历史、摘要、dirty 状态或 auto-loop action 推断。风险接受沿用 findings reference 的证据规则。
|
|
49
49
|
|
|
@@ -51,7 +51,7 @@ untracked helper 不存检查证据或风险接受。通过且按下方 Gate 默
|
|
|
51
51
|
|
|
52
52
|
所有 interactive 标准报告都必须在末尾输出 `### 下一步`,并按以下首个命中分支给出一个明确主动作:
|
|
53
53
|
|
|
54
|
-
1. 有未处置 findings:提示 `修复全部`、精确 ID、接受当前报告全部风险、`接受风险 <ID> 并继续` 或
|
|
54
|
+
1. 有未处置 findings:提示 `修复全部`、精确 ID、接受当前报告全部风险、`接受风险 <ID> 并继续` 或 `仅保留报告`;不逐项重复确认。受控分支说明默认动作和时长后按协议等待,不能写成已修复。
|
|
55
55
|
2. 有 blocked、阻断型部分验证或未标记 `[上线后验证]` 的实质风险:指出所需决策、授权或验证,完成后重跑 Check-All;不自行执行生产、外部或破坏性操作。
|
|
56
56
|
3. 有有效的明确停止要求且检查通过:说明已按要求停止,之后明确继续才恢复收尾。
|
|
57
57
|
4. 无有效停止要求且 strict pass / 已接受风险通过:说明本轮正在进入 `trellis-update-spec`,不要求用户再次回复“继续”;Git 计划确认仍由 Push 执行。
|
|
@@ -5,7 +5,7 @@ description: "按确认的精确文件范围提交普通变更或完成已就绪
|
|
|
5
5
|
|
|
6
6
|
# Trellis Push
|
|
7
7
|
|
|
8
|
-
`trellis-push` 是 Phase 3.4
|
|
8
|
+
`trellis-push` 是 Phase 3.4 唯一的代码提交入口。负责最小计划、精确提交推送、task 进度与 untracked 清理。
|
|
9
9
|
|
|
10
10
|
## 职责与条件入口
|
|
11
11
|
|
|
@@ -21,6 +21,7 @@ description: "按确认的精确文件范围提交普通变更或完成已就绪
|
|
|
21
21
|
| 普通 push 且有活动 task | [task-progress.md](references/task-progress.md);计划前读取,业务 Git 后执行 |
|
|
22
22
|
| taskStatus=completed 的恢复 | [completed-task-recovery.md](references/completed-task-recovery.md);先完成只读 preflight |
|
|
23
23
|
| validated runner 的内部 commit-only | [auto-loop-commit.md](references/auto-loop-commit.md);不 push、不执行 Step 5、不渲染交互模板 |
|
|
24
|
+
| 计划确认 | `.trellis/scripts/confirmation.md` |
|
|
24
25
|
|
|
25
26
|
未命中场景不加载其 reference;命中却缺失细则时阻塞,不凭记忆代替。内部 commit-only 只接受 runner 的合法预授权。
|
|
26
27
|
|
|
@@ -114,9 +115,8 @@ git rev-list --reverse @{u}..HEAD 2>/dev/null || true
|
|
|
114
115
|
|
|
115
116
|
确认前禁止 `git add`、`git commit` 或 `git push`。
|
|
116
117
|
|
|
117
|
-
|
|
118
|
-
|
|
119
|
-
reference 缺失、无法读取或缺少对应章节时停止并报告 `阻塞`,不得凭记忆重建、缩写或自制替代模板。
|
|
118
|
+
即将展示用户可见计划时,读取 `references/output-templates.md` 的“计划模板”和“共用展示规则”;缺失或不可读则 `阻塞`,不得凭记忆重建、缩写或自制替代模板。
|
|
119
|
+
按 `.trellis/scripts/confirmation.md` 使用 `ftl confirm` 冻结和展示完整计划;收到匹配 `push / execute_git_plan` 的领取结果才执行,否则等人工。
|
|
120
120
|
|
|
121
121
|
普通多仓只确认一次。计划已展示生成命令和预计 exact files 时,命令成功且没有出现预计列表外的新 dirty path 就沿用原确认;内容、hash 或统计变化不重问。其它计划边界变化仍按 Step 4 重新规划。
|
|
122
122
|
|
|
@@ -13,6 +13,10 @@ description: "从最新 prd.md、design.md、implement.md 生成、刷新、校
|
|
|
13
13
|
|
|
14
14
|
本例外沿用 runner 的 prepare、Open Questions、readiness 和 artifact 校验,不扩张任务或授权边界。schema 1 返回的 `refresh_brief` 同样在展示后交回 runner,由后续 `confirm_brief` action 等待人工确认。run 已停止、终态、损坏或 action/task 不匹配时不得免确认或代为推进;由 auto-loop owner 处理诊断。普通交互调用继续遵守以下确认规则。
|
|
15
15
|
|
|
16
|
+
## 原生会话超时确认
|
|
17
|
+
|
|
18
|
+
首次评审没有明确预授权、停止要求或未决事项时,读取 `.trellis/scripts/confirmation.md` 核验当前原生工具与项目策略。用 `ftl confirm` 冻结最新 Brief、在对话完整展示并前台等待;只有匹配 `brief / implement` 的领取结果,才在同轮启动任务。能力缺失或领取失败时照常展示并等人工。auto-loop 优先使用上方 runner 例外。
|
|
19
|
+
|
|
16
20
|
## 核心规则
|
|
17
21
|
|
|
18
22
|
- 每次运行都重新读取最新 `prd.md`、`design.md if present`、`implement.md if present`。
|
|
@@ -21,7 +25,7 @@ description: "从最新 prd.md、design.md、implement.md 生成、刷新、校
|
|
|
21
25
|
- 不要在 `brief.md` 里发明三件套没有表达的新需求。必填字段缺失时写“未明确”,并提示应补充三件套;没有明确技术机制时省略 `Technical Overview`,没有具体风险时省略 `Risks`。
|
|
22
26
|
- `Goal` 保持一句话,`Scope` 只写本轮纳入的行为或改动对象;延期事项归入 `Non-Goals`,不另设栏目,也不把范围、方案和验收写成同义复述。
|
|
23
27
|
- 首次开工评审或用户明确要求查看全貌时,写回 `brief.md` 后必须展示 brief 正文,不要只给文件路径;首次批准后的同步按下方增量规则展示变化。
|
|
24
|
-
- 首次实施批准前必须展示完整 brief
|
|
28
|
+
- 首次实施批准前必须展示完整 brief。除经 runner 校验的 Auto-Loop Action 或上方有效超时 claim 外,默认等待用户确认后再运行 `task.py start`;只有用户明确把当前任务或最终 Brief 与“展示后直接开始 / 不用再次确认 / 视为已确认”绑定时,才可在范围未变化的前提下免除第二次确认。
|
|
25
29
|
- “开始做吧”“按你建议来”“可以创建任务”等普通实现或建任务意图不是 Brief 预授权,不能据此跳过确认。
|
|
26
30
|
- 普通交互预授权只依赖当前对话中仍然明确可见的用户表达,不建立跨会话永久偏好,也不写 session runtime;auto-loop 恢复由上方 runner 例外所有。
|
|
27
31
|
- `in_progress` 阶段只读取 brief 和任务材料,不例行重新生成、展示或确认;新会话或压缩恢复同样如此。首次批准后的方案变化按下方增量规则处理,用户明确要求查看时再完整展示。
|
|
@@ -69,6 +73,7 @@ description: "从最新 prd.md、design.md、implement.md 生成、刷新、校
|
|
|
69
73
|
6. 写回 `<task>/brief.md`。如果文件已存在,仍用最新三件套派生内容覆盖旧正文。
|
|
70
74
|
7. 在对话中展示 brief 正文,并说明来源文件:
|
|
71
75
|
- 经校验的 auto-loop `refresh_brief`:展示后返回 runner,不进入下方交互分支;schema 1 的确认由后续 action 所有。
|
|
76
|
+
- 原生确认:按上方协议完整展示并等待;只有有效 claim 才同轮返回主 workflow 启动任务,并按实际结果记录 finish。取消或失效时转人工。已在当前对话完整展示的正文不重复输出。
|
|
72
77
|
- 无有效预授权:展示后结束当前回合,等待用户确认。
|
|
73
78
|
- 有有效预授权:先完整展示,再在同一回合返回主 workflow 执行 `task.py start`,不得省略展示步骤。
|
|
74
79
|
|
|
@@ -116,12 +116,13 @@
|
|
|
116
116
|
"values": [
|
|
117
117
|
"The existing `Interactive Post-Check Stop Gate` owns default continuation and explicit stopping.",
|
|
118
118
|
"without another next/continue reply or a direct Git prerequisite",
|
|
119
|
-
"
|
|
119
|
+
"repair needs a selected scope or a matching controlled-session `check` claim",
|
|
120
|
+
"timeout cannot accept risks",
|
|
120
121
|
"or replace Push confirmation"
|
|
121
122
|
]
|
|
122
123
|
},
|
|
123
124
|
"owner": "trellis-check-all",
|
|
124
|
-
"reason": "
|
|
125
|
+
"reason": "检查通过后默认推进,明确只检查仍停止;修复需人工范围或有效受控 claim,风险接受仍须人工,Git 归 Push。"
|
|
125
126
|
},
|
|
126
127
|
{
|
|
127
128
|
"id": "workflow-required-task-brief-review",
|
|
@@ -134,14 +135,14 @@
|
|
|
134
135
|
"type": "required-literal",
|
|
135
136
|
"values": [
|
|
136
137
|
"<!-- BEGIN skill-garden patch workflow-phase-1-activate v0.6 -->\n#### 1.4 Activate task `[required · once]`\n\nBefore changing task status, load `trellis-task-brief`",
|
|
137
|
-
"Unless `trellis-task-brief` validates an explicit preauthorization for the current final Brief",
|
|
138
|
+
"Unless `trellis-task-brief` validates an explicit preauthorization for the current final Brief or a matching controlled-session `brief` policy-timeout claim",
|
|
138
139
|
"Ordinary implementation or task-creation intent is not confirmation.",
|
|
139
140
|
"After initial approval, follow that owner's incremental change rules",
|
|
140
|
-
"After a later confirmation, or in the same turn when that explicit preauthorization remains valid"
|
|
141
|
+
"After a later confirmation, or in the same turn when that explicit preauthorization remains valid or a matching single-use policy-timeout claim succeeds"
|
|
141
142
|
]
|
|
142
143
|
},
|
|
143
144
|
"owner": "trellis-task-brief",
|
|
144
|
-
"reason": "Phase 1.4
|
|
145
|
+
"reason": "Phase 1.4 由 trellis-task-brief 完整展示并校验人工、预授权或匹配材料的一次性超时 claim。"
|
|
145
146
|
},
|
|
146
147
|
{
|
|
147
148
|
"id": "brainstorm-required-planning-handoff",
|
|
@@ -157,7 +158,7 @@
|
|
|
157
158
|
"<!-- BEGIN skill-garden patch brainstorm-planning-handoff v0.6 -->\n## Quality Bar",
|
|
158
159
|
"## Planning Handoff",
|
|
159
160
|
"<!-- BEGIN skill-garden patch brainstorm-planning-readiness v0.6 -->\n8. Load `trellis-task-brief`",
|
|
160
|
-
"Only valid explicit preauthorization
|
|
161
|
+
"Only valid explicit preauthorization or a matching controlled-session policy-timeout claim under `trellis-task-brief` permits same-turn start within unchanged scope.",
|
|
161
162
|
"The final Brief must cover a one-sentence Goal, a Scope limited to included changes, Non-Goals, Context & Decisions, Acceptance, and one-hop Next Step",
|
|
162
163
|
"Ordinary implementation intent expressed before the final artifacts and full Brief are shown authorizes planning only",
|
|
163
164
|
"Once initial Brief approval exists, plan updates use `trellis-task-brief`'s incremental change rules.",
|
|
@@ -182,7 +183,7 @@
|
|
|
182
183
|
"<!-- BEGIN skill-garden patch brainstorm-planning-handoff v0.6 -->\n## Quality Bar",
|
|
183
184
|
"## Planning Handoff",
|
|
184
185
|
"<!-- BEGIN skill-garden patch brainstorm-planning-readiness v0.6 -->\n8. Load `trellis-task-brief`",
|
|
185
|
-
"Only valid explicit preauthorization
|
|
186
|
+
"Only valid explicit preauthorization or a matching controlled-session policy-timeout claim under `trellis-task-brief` permits same-turn start within unchanged scope.",
|
|
186
187
|
"The final Brief must cover a one-sentence Goal, a Scope limited to included changes, Non-Goals, Context & Decisions, Acceptance, and one-hop Next Step",
|
|
187
188
|
"Ordinary implementation intent expressed before the final artifacts and full Brief are shown authorizes planning only",
|
|
188
189
|
"Once initial Brief approval exists, plan updates use `trellis-task-brief`'s incremental change rules.",
|
package/enhancements/0.6/overrides/patches/skills/trellis-brainstorm/planning-handoff/content.md
CHANGED
|
@@ -14,7 +14,7 @@ Do not start implementation merely because the user originally asked for impleme
|
|
|
14
14
|
|
|
15
15
|
## Planning Handoff
|
|
16
16
|
|
|
17
|
-
Once the Quality Bar is satisfied, load `trellis-task-brief`, refresh `brief.md` from
|
|
17
|
+
Once the Quality Bar is satisfied, load `trellis-task-brief`, refresh `brief.md` from final planning artifacts, and display the full Brief in chat or through that owner's verified foreground controller. By default, end the current turn and wait for the user's planning review confirmation. Only valid explicit preauthorization or a matching controlled-session policy-timeout claim under `trellis-task-brief` permits same-turn start within unchanged scope.
|
|
18
18
|
|
|
19
19
|
Ordinary implementation intent expressed before the final artifacts and full Brief are shown authorizes planning only; it cannot be reused as the final review confirmation.
|
|
20
20
|
|
|
@@ -1 +1 @@
|
|
|
1
|
-
- Task creation approval is not implementation approval; task implementation waits for `task.py start` after artifact review.
|
|
1
|
+
- Task creation approval is not implementation approval; task implementation waits for `task.py start` after artifact review. `trellis-task-brief` owns any verified controlled-session timeout claim.
|
package/enhancements/0.6/overrides/patches/workflow/phase-ownership/phase-2-check-content.md
CHANGED
|
@@ -6,8 +6,8 @@ For untracked work, route reads only the personal pref helper and the dispatch p
|
|
|
6
6
|
|
|
7
7
|
Before interactive Check-All begins, run `python3 ./.trellis/scripts/pre_check_state.py clear`. A missing, subject-mismatched, or already-cleared preference is a no-op; a damaged runtime is reported diagnostically but safely defaults to checking.
|
|
8
8
|
|
|
9
|
-
Check-All selects light/full depth from intent, actual behavioral-contract impact, and runtime context. It is audit-only and collect-all by default: classify main-path issues as `CHK-*`, fallback-path issues as `FBK-*`, and low-risk factual drift as `DOC-*`. Assign P0/P1/P2 after
|
|
9
|
+
Check-All selects light/full depth from intent, actual behavioral-contract impact, and runtime context. It is audit-only and collect-all by default: classify main-path issues as `CHK-*`, fallback-path issues as `FBK-*`, and low-risk factual drift as `DOC-*`. Assign P0/P1/P2 after classification. Report all items before code changes; repair needs a selected scope or a matching controlled-session `check` claim under Check-All's reporting owner. Low-risk `DOC-*` auto-remediation remains the sole audit-stage write exception, unless a validated auto-loop owns continuation.
|
|
10
10
|
|
|
11
|
-
The existing `Interactive Post-Check Stop Gate` owns default continuation and explicit stopping. Unless the user explicitly requests check-only, read-only checking, stopping after the check, or no further completion work for this subject, a strict pass or a pass with every remaining finding explicitly and validly accepted by the user continues to `trellis-update-spec` in the same turn, without another next/continue reply or a direct Git prerequisite. Both paths require no blocker, no blocking partial verification, and no unaccepted material residual risk outside explicitly recorded `[上线后验证]`. A user reply accepting all findings from the current report covers the full report, including P0, without a fixed phrase or repeated ID confirmation; partial acceptance must identify a unique subset. Do not infer acceptance or Git write authorization from unrelated history, summaries, or dirty state. Preserve accepted findings in existing evidence and use the reporting reference's incremental display rules. `[上线后验证]` items remain visible but do not block.
|
|
11
|
+
The existing `Interactive Post-Check Stop Gate` owns default continuation and explicit stopping. Unless the user explicitly requests check-only, read-only checking, stopping after the check, or no further completion work for this subject, a strict pass or a pass with every remaining finding explicitly and validly accepted by the user continues to `trellis-update-spec` in the same turn, without another next/continue reply or a direct Git prerequisite. Both paths require no blocker, no blocking partial verification, and no unaccepted material residual risk outside explicitly recorded `[上线后验证]`. A user reply accepting all findings from the current report covers the full report, including P0, without a fixed phrase or repeated ID confirmation; partial acceptance must identify a unique subset. Do not infer acceptance or Git write authorization from unrelated history, summaries, or dirty state. Preserve accepted findings in existing evidence and use the reporting reference's incremental display rules. `[上线后验证]` items remain visible but do not block. Unresolved findings return to Check-All's reporting owner for manual disposition or a verified timeout repair selection; timeout cannot accept risks. Checks do not create the Git plan or replace Push confirmation.
|
|
12
12
|
|
|
13
13
|
After authorized repairs, return through the same route and re-run Check-All. Finishing Phase 2.2 or entering the commit chain does not itself force another Full: reuse still-valid Full evidence plus complete focused recheck evidence. Run Full again only when the user explicitly requests it, a hard-full signal appears, scope cannot be closed, a baseline or related contract changes, an unknown dirty path appears, or prior evidence becomes stale or invalid.
|
|
@@ -2,5 +2,5 @@ Planning is not implementation permission. Load `trellis-brainstorm` and stay in
|
|
|
2
2
|
A created task or default `prd.md` is not enough to start implementation. Lightweight tasks may remain PRD-only; complex tasks require `prd.md`, `design.md`, and `implement.md`.
|
|
3
3
|
Before changing artifacts, routing, or files, follow `Request Triage`: resolve selected Check-All repairs first; apply the Active Task Scope Guard only to new work.
|
|
4
4
|
If the latest current-request switch says no task/direct edit, call `task_intent.py discard --task <current-task>` only for a task auto-created by intent routing. Continue only on `status=discarded`; keep manual or historical tasks unchanged.
|
|
5
|
-
Before the first implementation approval, refresh and display `brief.md` through `trellis-task-brief
|
|
5
|
+
Before the first implementation approval, refresh and fully display `brief.md` through `trellis-task-brief`; its owner validates manual confirmation, preauthorization, or a matching controlled-session policy-timeout claim. After initial approval, use that owner's incremental change rules instead of repeating whole-Brief approval on every artifact update.
|
|
6
6
|
After status becomes `in_progress`, enter Phase 2 through `trellis-route(target=implement)` instead of editing or dispatching directly.
|
package/enhancements/0.6/overrides/patches/workflow/task-brief-review/phase-1-activate-content.md
CHANGED
|
@@ -1,12 +1,12 @@
|
|
|
1
1
|
#### 1.4 Activate task `[required · once]`
|
|
2
2
|
|
|
3
|
-
Before changing task status, load `trellis-task-brief`. For the first implementation approval, refresh `<task>/brief.md` and display
|
|
3
|
+
Before changing task status, load `trellis-task-brief`. For the first implementation approval, refresh `<task>/brief.md` and fully display it through that owner. Unless `trellis-task-brief` validates an explicit preauthorization for the current final Brief or a matching controlled-session `brief` policy-timeout claim, stop and wait for planning review confirmation. Ordinary implementation or task-creation intent is not confirmation.
|
|
4
4
|
|
|
5
5
|
After initial approval, follow that owner's incremental change rules: synchronize affected artifacts and resolve only new decisions, without repeating whole-Brief approval for a document refresh or an already authorized change. Existing approval with no unresolved change decisions permits activation without another confirmation.
|
|
6
6
|
|
|
7
7
|
Lightweight tasks need `prd.md`; complex tasks also need `design.md` and `implement.md`. Sub-agent routes require real entries in both JSONL manifests.
|
|
8
8
|
|
|
9
|
-
After a later confirmation, or in the same turn when that explicit preauthorization remains valid, run:
|
|
9
|
+
After a later confirmation, or in the same turn when that explicit preauthorization remains valid or a matching single-use policy-timeout claim succeeds, run:
|
|
10
10
|
|
|
11
11
|
```bash
|
|
12
12
|
python3 ./.trellis/scripts/task.py start <task-dir>
|
|
@@ -0,0 +1,24 @@
|
|
|
1
|
+
# 前台限时确认
|
|
2
|
+
|
|
3
|
+
本协议用于 Brief 首次开工、Check-All 问题处置、Push 精确计划。项目默认启用,由助手在当前原生窗口直接运行 `ftl confirm`。原版 Codex 需能核验原连接,Claude Code 需能核验当前终端和项目 Hook;能力缺失、配置关闭或失联时沿用人工。已校验的 auto-loop 继续用 runner。Spec 歧义、未决业务取舍、风险接受和独立权限审批仍由人工决定。
|
|
4
|
+
|
|
5
|
+
1. **检查能力。** 用当前主对话的内置 Shell 调用 `ftl confirm status`。只在 `protocol=1`、`bound=true`、`ready=true`、`enabled=true` 且本类 `kinds.<brief|check|push>=true` 时继续。命令缺失、诊断异常或 `manual=true` 时,简述原因,照常完整展示并等待人工。
|
|
6
|
+
2. **准备材料。** 按当前流程的原模板形成完整正文,保留全部范围、风险与生成命令。Brief 的 `file` 必须是最新的 `<task>/brief.md`;检查和 Git 计划用 `body` 提供完整正文,或用已冻结的项目内相对 `file`。可用 `python3 -X utf8 .trellis/scripts/confirmation.py materialize` 从 stdin 冻结较长正文,结果只写入 gitignored runtime。可选 `basis` 填项目内证据路径;检查 `problem_ids` 列全当前待处置 CHK/FBK/DOC;提交明确 `mode: "push"|"commit-only"`,单仓用 `git_paths`,多仓用 `repositories: [{"root":"仓库路径","git_paths":["精确文件"]}]` 按计划顺序列全。
|
|
7
|
+
3. **前台等待。** `prepare` 从 stdin 读取单行 JSON,输出 `request_id`、`review`、`seconds`、`kind` 和 `action`,此时未计时。把 `review` 原样作为一条普通、可见的助手正文发送,再调用 `ftl confirm <request-id>`。不要加导语、引用、代码围栏、截取末段或用链接替代正文。Claude 把正文和 Bash 调用放在同一条助手消息,Bash 用 `timeout: 600000` 和 `run_in_background: false`;Codex 使用内置 `exec_command`,若返回会话 ID,只用 `write_stdin` 空输入等同一进程,Code Mode 若先让出则等待同一单元。不要用 `&`、后台任务、定时唤醒或重复运行等待命令。Shell 工具输出不是展示回执。
|
|
8
|
+
4. **核对并执行。** 仅当该前台命令成功返回同一 `request_id`、`source=policy_timeout`、`decision=continue` 且 `kind`、`action` 均匹配时,本次范围选择才被领取。立即按原 owner 的前置条件与精确范围执行;用户回复、取消、材料变化、预算不足或错误时停止旧决定并转人工。不得重试已取消或已领取的决定,也不能把它带到另一个回合。真实执行结果用 `ftl confirm finish <request-id> <completed|interrupted|needs_verification>` 记录;失联后核实已发生的副作用。
|
|
9
|
+
|
|
10
|
+
`prepare` 例子(用实际材料替换 JSON;分隔符保持单引号,避免 Shell 展开):
|
|
11
|
+
|
|
12
|
+
```bash
|
|
13
|
+
ftl confirm prepare <<'FTL_CONFIRM_SPEC'
|
|
14
|
+
{"kind":"brief","file":".trellis/tasks/实际任务/brief.md"}
|
|
15
|
+
FTL_CONFIRM_SPEC
|
|
16
|
+
```
|
|
17
|
+
|
|
18
|
+
| kind | 默认时长 | action | 匹配后的动作 |
|
|
19
|
+
| --- | --- | --- | --- |
|
|
20
|
+
| brief | 正文 1000 字内 100 秒;每多 500 字加 30 秒,不足一档按一档 | implement | 对已展示的 Brief 运行 `task.py start` |
|
|
21
|
+
| check | 第一条 100 秒,每多一条加 30 秒;零条不等待 | repair_all_and_recheck | 修复当前全部待处置项并重检;新业务取舍暂停 |
|
|
22
|
+
| push | 200 秒 | execute_git_plan | 按当前精确计划提交并推送;明确 commit-only 时仅本地提交 |
|
|
23
|
+
|
|
24
|
+
自动决定记为 `policy_timeout`,人工明确回复记为 `human`。项目或单类可在 `.trellis/config.yaml` 的 `confirmation_timeout` 中设置 `enabled: false`;缺失使用默认,损坏或超过预算时人工确认,不截短等待。运行记录不入 Git,也不能在窗口重开时补跑。
|