flower-trellis 0.4.0-beta.0 → 0.4.0-beta.1

package/enhancements/0.6/.agents/skills/trellis-auto-loop/SKILL.md

@@ -14,35 +14,30 @@ description: "启动、恢复和推进 Trellis 自动任务循环。用于用户
 - 每完成一个 action，必须用 `record --action <next 返回的 action>` 精确写回结果；runner 会拒绝缺失或不匹配的 action。写回后立即再调用 `next`，直到 `done`、`blocked` 或需要用户决策。
 - 默认 profile 是 `commit-only`：自动推进到本地 commit，不 push、不发布、不归档。
 - 多任务只按用户显式给出的任务顺序执行；同一 worktree 不并发。
-- route 临时授权写在 auto-loop runtime 中，低于个人 `.trellis/.route-prefs.tmp`；不要写个人偏好。
-- planning start gate 会按有效 route 判断 JSONL 是否必需：inline / check-all-inline 可不因 seed-only JSONL 停住；个人默认 subagent 仍优先于 auto inline 授权并要求 curated context。
+- 启动 runner 前先完成 route 准备度判断：已有当前任务 runtime route 决策或个人 `.trellis/.route-prefs.tmp` 时可启动；没有时先进入 `trellis-route` 正常询问 / fallback，写入真实决策后再启动。
+- auto-loop 不默认写 `route_authorization`；如果未来显式传 route 参数，只能表示用户本次明确给出的临时策略，不能当成模型真实执行结果。
 - 代码提交必须走 `trellis-push` 的 commit-only 语义；不要裸 `git commit` / `git push`。
 
 ## 启动
 
 用户给了任务列表时，按原顺序传入；用户只说当前任务时，用 `task.py current --source` 的当前任务。
 
-子代理可用的 workflow 默认写入临时 route 授权：
+启动前对当前任务执行 route 准备度检查：
 
 ```bash
-python3 ./.trellis/scripts/auto_loop.py start \
-  --tasks <task> [<task> ...] \
-  --profile commit-only \
-  --route-implement subagent \
-  --route-check check-all-subagent
+python3 .agents/skills/trellis-route/scripts/route_state.py resolve --target implement
+python3 .agents/skills/trellis-route/scripts/route_state.py resolve --target check
 ```
 
-inline-only 场景改用：
+如果任一 target 返回 `status=miss`，先按 `trellis-route` 的正常 numbered fallback 询问用户并写入 runtime；不要替用户默认 inline 或 subagent。route 准备完成后启动 runner：
 
 ```bash
 python3 ./.trellis/scripts/auto_loop.py start \
   --tasks <task> [<task> ...] \
-  --profile commit-only \
-  --route-implement inline \
-  --route-check check-all-inline
+  --profile commit-only
 ```
 
-如果用户个人 `.trellis/.route-prefs.tmp` 存在，`trellis-route` 会优先使用个人偏好；auto 授权只负责 prefs miss 时减少询问。
+多任务队列中，当前任务切换到下一个任务且缺少该任务 route 决策时，回到 `trellis-route` 获取该任务真实选择，再继续 `next` / `record`。个人 `.trellis/.route-prefs.tmp` 会由 `trellis-route` 统一复用并写回 runtime。
 
 启动后立即运行：
 
@@ -67,12 +62,12 @@ python3 ./.trellis/scripts/auto_loop.py next
 | --- | --- | --- |
 | `refresh_brief` | 使用 `trellis-task-brief` 生成并展示 brief | `record --action refresh_brief --result ok` |
 | `start_task` | 执行返回的 `task.py start ...` 命令 | `record --action start_task --result ok` |
-| `run_implement` | 进入 Phase 2.1，先用 `trellis-route(target=implement)` 决定 inline/subagent，再实现 | `record --action run_implement --result ok` |
-| `run_check_all` | 进入 Phase 2.2，先用 `trellis-route(target=check)`，执行 check-all | `record --action run_check_all --result ok` |
-| `run_fix` | 根据 `last_failure` 修复，复用当前任务 implement route | `record --action run_fix --result ok` |
-| `run_recheck` | 复用当前任务 check route，重新 check-all | `record --action run_recheck --result ok` |
+| `run_implement` | 进入 Phase 2.1，先用 `trellis-route(target=implement)` 决定 inline/subagent，再实现 | `record --action run_implement --result ok --route-mode <mode> --route-source <source>` |
+| `run_check_all` | 进入 Phase 2.2，先用 `trellis-route(target=check)`，执行 check-all | `record --action run_check_all --result ok --route-mode <mode> --route-source <source>` |
+| `run_fix` | 根据 `last_failure` 修复，复用当前任务 implement route | `record --action run_fix --result ok --route-mode <mode> --route-source <source>` |
+| `run_recheck` | 复用当前任务 check route，重新 check-all | `record --action run_recheck --result ok --route-mode <mode> --route-source <source>` |
 | `run_spec_update` | 有代码/测试证据时用 `trellis-update-spec`；无必要更新也 record ok | `record --action run_spec_update --result ok` |
-| `commit_only` | 使用 `trellis-push` commit-only auto-loop 预授权路径，只提交当前任务可归属文件 | `record --action commit_only --result ok --commit <hash>` |
+| `commit_only` | 进入 `trellis-push` commit-only 语义：先由 AI 基于 task artifacts 与 git diff 生成提交计划，只提交当前任务可归属文件 | 执行成功后 `record --action commit_only --result ok --commit <hash>` |
 
 失败时写回：
 
@@ -97,6 +92,10 @@ python3 ./.trellis/scripts/auto_loop.py record \
 
 runner 会按 3 轮 fix/recheck 预算决定继续、跳过当前任务或结束队列。
 
+route action 成功回写时必须带上 `trellis-route` 输出里的真实 `mode` / `source`，例如
+`--route-mode inline --route-source route-prefs` 或
+`--route-mode check-all-subagent --route-source trellis-route`；不要写 auto-loop 默认值。
+
 ## Commit-Only 预授权
 
 auto-loop 的 `commit-only` profile 是用户对“当前 run 内任务相关本地提交”的一次性预授权。只有同时满足这些条件时，`trellis-push` 可以跳过二次聊天确认：
@@ -104,11 +103,14 @@ auto-loop 的 `commit-only` profile 是用户对“当前 run 内任务相关本
 - `.trellis/scripts/auto_loop.py status` 显示当前 `run_status=running`，且 `outstanding_action.action=commit_only`、`outstanding_action.task` 等于当前活动任务。
 - profile 是 `commit-only`。
 - 模式是 commit-only，不 push、不 merge、不发布、不归档。
+- AI 已基于当前任务 artifacts、`git status`、`git diff` 和必要的文件内容生成提交计划，并能说明每个 planned file 为什么属于当前任务。
 - 提交计划只包含当前任务可归属文件；未识别 dirty 文件保留未提交并写入结果摘要。
 - 执行前复核 git 状态仍与计划一致。
 
 如果计划包含未识别 staged 文件、冲突、push/merge/release/archive、真实外部系统或生产数据效果，停止并把当前任务记为 blocked。
 
+不要用脚本基于时间差或 dirty baseline 猜测文件归属；语义归属由 AI 负责判断，`trellis-push` 负责展示/复核计划、精确暂存 planned files、执行本地 commit，并把 commit hash 回写 runner。预检不安全时只把当前队列项记为 blocked/skipped，并让 auto-loop 可继续后续任务。
+
 ## 状态与停止
 
 查看：

package/enhancements/0.6/.agents/skills/trellis-push/SKILL.md

@@ -8,7 +8,7 @@ description: "提交并推送配置仓库，可选合并到目标分支，并同
 
 核心原则：**先计划、一次确认、后执行**。在任何 `git add`、`git commit`、`git push`、`git merge` 之前，必须先展示完整执行计划，并获得用户确认。
 
-例外：当 `trellis-auto-loop` 已显式以 `profile=commit-only` 启动，且 runner 当前有 `outstanding_action.action=commit_only` 时，用户启动 auto-loop 本身就是本次 run 内任务相关本地提交的预授权。此时仍必须生成并自检完整计划，但如果计划只包含当前任务可归属文件、模式为 commit-only、不会 push/merge/release/archive，并且执行后会把 commit hash 回写 runner，可以跳过二次聊天确认直接执行。
+例外：当 `trellis-auto-loop` 已显式以 `profile=commit-only` 启动，且 runner 当前有 `outstanding_action.action=commit_only` 时，用户启动 auto-loop 本身就是本次 run 内任务相关本地提交的预授权。此时仍必须由 AI 生成并自检提交计划，但可跳过二次聊天确认；计划只允许 commit-only、不 push/merge/release/archive，并在成功后把 commit hash 回写 runner。预检不安全时只标记当前 auto-loop item blocked/skipped，后续队列仍可继续。
 
 支持多仓库（frontend / backend 等），merge 目标分支记录在 `.trellis/config.yaml` 的 `packages.<name>.merge_target` 中。
 
@@ -80,6 +80,8 @@ Step 6  输出结果
 
 任一条件不满足时，不得使用预授权。普通 `trellis-push` 继续走“展示计划并等待确认”的规则。
 
+auto-loop 预授权路径不使用额外提交 helper。AI 负责根据当前任务 artifacts、`git status`、`git diff` 和必要的文件内容判断文件归属，输出 planned files / retained files / commit message / 归属理由；`trellis-push` 负责边界复核：当前 action/profile/task 匹配、staged 区为空、无冲突、planned files 当前 dirty、planned files 不含 `.trellis/.runtime/`、`.trellis/.route-prefs.tmp`、其他任务目录或未解释文件。通过后只 `git add` planned files 并执行本地 commit，随后回写 `auto_loop.py record --action commit_only --result ok --commit <hash> ...`。预检失败时回写 `record --result blocked`，该 blocked 只影响当前 item；auto-loop runner 后续 `next` 应继续寻找后续 pending 任务。只有 merge/rebase 冲突、repo 状态不可读、脚本损坏或用户明确 stop 才停止整个 run。
+
 ### 0.2 发现候选 Git 仓库
 
 读取 `.trellis/config.yaml` 中的 `packages` 配置，但不要把 `git: true` 当作唯一仓库发现规则。`git: true` 主要用于 Trellis session context 展示独立包仓库状态；`trellis-push` 必须以实际 Git root 为准发现可提交仓库：

package/enhancements/0.6/.agents/skills/trellis-route/scripts/route_state.py

@@ -160,36 +160,63 @@ def _read_prefs(repo_root: Path) -> dict[str, str]:
     return prefs
 
 
+def _running_auto_run_paths(repo_root: Path) -> list[Path]:
+    """扫描当前项目内所有 running auto-loop run。"""
+    running: list[Path] = []
+    for path in sorted(_auto_loop_dir(repo_root).glob("auto-*.json")):
+        state = _read_json(path)
+        if state.get("status") == "running":
+            running.append(path)
+    return running
+
+
+def _auto_state_path(repo_root: Path, run_id: Any) -> Path | None:
+    """把 run id 转成状态文件路径，非法值返回 None。"""
+    if not isinstance(run_id, str) or not run_id.strip():
+        return None
+    return _auto_loop_dir(repo_root) / f"{run_id.strip()}.json"
+
+
 def _auto_route_mode(repo_root: Path, context_key: str, target: str) -> tuple[str | None, Path | None, str | None]:
-    """读取当前 auto-loop run 的临时 route 授权。
+    """读取当前 running auto-loop run 的临时 route 授权。
 
     auto 授权低于个人 `.route-prefs.tmp`，只在当前 session runtime 绑定了
-    `current_auto_run`，或全局 current 指针能指向唯一 running run 时生效。命中后
-    `resolve` 会写回 session runtime，后续压缩恢复仍走普通 route 决策路径。
+    running `current_auto_run`，或全局 current 指针能指向 running run 时生效。
+    如果这些指针 stale，则忽略 stale pointer，并 fallback 到唯一 running run。
     """
     session_path = _session_path(repo_root, context_key)
     session = _read_json(session_path)
-    run_id = session.get("current_auto_run")
-
-    if not isinstance(run_id, str) or not run_id.strip():
-        pointer = _read_json(_auto_loop_pointer(repo_root))
-        run_id = pointer.get("run_id")
-
     candidate_paths: list[Path] = []
-    if isinstance(run_id, str) and run_id.strip():
-        candidate_paths.append(_auto_loop_dir(repo_root) / f"{run_id.strip()}.json")
-    else:
-        for path in sorted(_auto_loop_dir(repo_root).glob("auto-*.json")):
-            state = _read_json(path)
-            if state.get("status") == "running":
-                candidate_paths.append(path)
-        if len(candidate_paths) != 1:
-            return None, None, "no-unique-auto-run"
-
-    path = candidate_paths[0]
+    stale_paths: list[Path] = []
+
+    for run_id in (session.get("current_auto_run"), _read_json(_auto_loop_pointer(repo_root)).get("run_id")):
+        path = _auto_state_path(repo_root, run_id)
+        if path is None:
+            continue
+        state = _read_json(path)
+        if state.get("status") == "running":
+            candidate_paths.append(path)
+        else:
+            stale_paths.append(path)
+
+    if not candidate_paths:
+        candidate_paths = _running_auto_run_paths(repo_root)
+    unique_paths = []
+    seen: set[str] = set()
+    for path in candidate_paths:
+        key = str(path)
+        if key in seen:
+            continue
+        seen.add(key)
+        unique_paths.append(path)
+    if len(unique_paths) != 1:
+        reason = "no-unique-auto-run"
+        if stale_paths and not unique_paths:
+            reason = "stale-auto-run-pointer"
+        return None, stale_paths[0] if stale_paths else None, reason
+
+    path = unique_paths[0]
     state = _read_json(path)
-    if state.get("status") != "running":
-        return None, path, "auto-run-not-running"
 
     auth = state.get("route_authorization")
     if not isinstance(auth, dict):

package/enhancements/0.6/.claude/skills/trellis-auto-loop/SKILL.md

@@ -14,35 +14,30 @@ description: "启动、恢复和推进 Trellis 自动任务循环。用于用户
 - 每完成一个 action，必须用 `record --action <next 返回的 action>` 精确写回结果；runner 会拒绝缺失或不匹配的 action。写回后立即再调用 `next`，直到 `done`、`blocked` 或需要用户决策。
 - 默认 profile 是 `commit-only`：自动推进到本地 commit，不 push、不发布、不归档。
 - 多任务只按用户显式给出的任务顺序执行；同一 worktree 不并发。
-- route 临时授权写在 auto-loop runtime 中，低于个人 `.trellis/.route-prefs.tmp`；不要写个人偏好。
-- planning start gate 会按有效 route 判断 JSONL 是否必需：inline / check-all-inline 可不因 seed-only JSONL 停住；个人默认 subagent 仍优先于 auto inline 授权并要求 curated context。
+- 启动 runner 前先完成 route 准备度判断：已有当前任务 runtime route 决策或个人 `.trellis/.route-prefs.tmp` 时可启动；没有时先进入 `trellis-route` 正常询问 / fallback，写入真实决策后再启动。
+- auto-loop 不默认写 `route_authorization`；如果未来显式传 route 参数，只能表示用户本次明确给出的临时策略，不能当成模型真实执行结果。
 - 代码提交必须走 `trellis-push` 的 commit-only 语义；不要裸 `git commit` / `git push`。
 
 ## 启动
 
 用户给了任务列表时，按原顺序传入；用户只说当前任务时，用 `task.py current --source` 的当前任务。
 
-子代理可用的 workflow 默认写入临时 route 授权：
+启动前对当前任务执行 route 准备度检查：
 
 ```bash
-python3 ./.trellis/scripts/auto_loop.py start \
-  --tasks <task> [<task> ...] \
-  --profile commit-only \
-  --route-implement subagent \
-  --route-check check-all-subagent
+python3 .agents/skills/trellis-route/scripts/route_state.py resolve --target implement
+python3 .agents/skills/trellis-route/scripts/route_state.py resolve --target check
 ```
 
-inline-only 场景改用：
+如果任一 target 返回 `status=miss`，先按 `trellis-route` 的正常 numbered fallback 询问用户并写入 runtime；不要替用户默认 inline 或 subagent。route 准备完成后启动 runner：
 
 ```bash
 python3 ./.trellis/scripts/auto_loop.py start \
   --tasks <task> [<task> ...] \
-  --profile commit-only \
-  --route-implement inline \
-  --route-check check-all-inline
+  --profile commit-only
 ```
 
-如果用户个人 `.trellis/.route-prefs.tmp` 存在，`trellis-route` 会优先使用个人偏好；auto 授权只负责 prefs miss 时减少询问。
+多任务队列中，当前任务切换到下一个任务且缺少该任务 route 决策时，回到 `trellis-route` 获取该任务真实选择，再继续 `next` / `record`。个人 `.trellis/.route-prefs.tmp` 会由 `trellis-route` 统一复用并写回 runtime。
 
 启动后立即运行：
 
@@ -67,12 +62,12 @@ python3 ./.trellis/scripts/auto_loop.py next
 | --- | --- | --- |
 | `refresh_brief` | 使用 `trellis-task-brief` 生成并展示 brief | `record --action refresh_brief --result ok` |
 | `start_task` | 执行返回的 `task.py start ...` 命令 | `record --action start_task --result ok` |
-| `run_implement` | 进入 Phase 2.1，先用 `trellis-route(target=implement)` 决定 inline/subagent，再实现 | `record --action run_implement --result ok` |
-| `run_check_all` | 进入 Phase 2.2，先用 `trellis-route(target=check)`，执行 check-all | `record --action run_check_all --result ok` |
-| `run_fix` | 根据 `last_failure` 修复，复用当前任务 implement route | `record --action run_fix --result ok` |
-| `run_recheck` | 复用当前任务 check route，重新 check-all | `record --action run_recheck --result ok` |
+| `run_implement` | 进入 Phase 2.1，先用 `trellis-route(target=implement)` 决定 inline/subagent，再实现 | `record --action run_implement --result ok --route-mode <mode> --route-source <source>` |
+| `run_check_all` | 进入 Phase 2.2，先用 `trellis-route(target=check)`，执行 check-all | `record --action run_check_all --result ok --route-mode <mode> --route-source <source>` |
+| `run_fix` | 根据 `last_failure` 修复，复用当前任务 implement route | `record --action run_fix --result ok --route-mode <mode> --route-source <source>` |
+| `run_recheck` | 复用当前任务 check route，重新 check-all | `record --action run_recheck --result ok --route-mode <mode> --route-source <source>` |
 | `run_spec_update` | 有代码/测试证据时用 `trellis-update-spec`；无必要更新也 record ok | `record --action run_spec_update --result ok` |
-| `commit_only` | 使用 `trellis-push` commit-only auto-loop 预授权路径，只提交当前任务可归属文件 | `record --action commit_only --result ok --commit <hash>` |
+| `commit_only` | 进入 `trellis-push` commit-only 语义：先由 AI 基于 task artifacts 与 git diff 生成提交计划，只提交当前任务可归属文件 | 执行成功后 `record --action commit_only --result ok --commit <hash>` |
 
 失败时写回：
 
@@ -97,6 +92,10 @@ python3 ./.trellis/scripts/auto_loop.py record \
 
 runner 会按 3 轮 fix/recheck 预算决定继续、跳过当前任务或结束队列。
 
+route action 成功回写时必须带上 `trellis-route` 输出里的真实 `mode` / `source`，例如
+`--route-mode inline --route-source route-prefs` 或
+`--route-mode check-all-subagent --route-source trellis-route`；不要写 auto-loop 默认值。
+
 ## Commit-Only 预授权
 
 auto-loop 的 `commit-only` profile 是用户对“当前 run 内任务相关本地提交”的一次性预授权。只有同时满足这些条件时，`trellis-push` 可以跳过二次聊天确认：
@@ -104,11 +103,14 @@ auto-loop 的 `commit-only` profile 是用户对“当前 run 内任务相关本
 - `.trellis/scripts/auto_loop.py status` 显示当前 `run_status=running`，且 `outstanding_action.action=commit_only`、`outstanding_action.task` 等于当前活动任务。
 - profile 是 `commit-only`。
 - 模式是 commit-only，不 push、不 merge、不发布、不归档。
+- AI 已基于当前任务 artifacts、`git status`、`git diff` 和必要的文件内容生成提交计划，并能说明每个 planned file 为什么属于当前任务。
 - 提交计划只包含当前任务可归属文件；未识别 dirty 文件保留未提交并写入结果摘要。
 - 执行前复核 git 状态仍与计划一致。
 
 如果计划包含未识别 staged 文件、冲突、push/merge/release/archive、真实外部系统或生产数据效果，停止并把当前任务记为 blocked。
 
+不要用脚本基于时间差或 dirty baseline 猜测文件归属；语义归属由 AI 负责判断，`trellis-push` 负责展示/复核计划、精确暂存 planned files、执行本地 commit，并把 commit hash 回写 runner。预检不安全时只把当前队列项记为 blocked/skipped，并让 auto-loop 可继续后续任务。
+
 ## 状态与停止
 
 查看：

package/enhancements/0.6/.claude/skills/trellis-push/SKILL.md

@@ -8,7 +8,7 @@ description: "提交并推送配置仓库，可选合并到目标分支，并同
 
 核心原则：**先计划、一次确认、后执行**。在任何 `git add`、`git commit`、`git push`、`git merge` 之前，必须先展示完整执行计划，并获得用户确认。
 
-例外：当 `trellis-auto-loop` 已显式以 `profile=commit-only` 启动，且 runner 当前有 `outstanding_action.action=commit_only` 时，用户启动 auto-loop 本身就是本次 run 内任务相关本地提交的预授权。此时仍必须生成并自检完整计划，但如果计划只包含当前任务可归属文件、模式为 commit-only、不会 push/merge/release/archive，并且执行后会把 commit hash 回写 runner，可以跳过二次聊天确认直接执行。
+例外：当 `trellis-auto-loop` 已显式以 `profile=commit-only` 启动，且 runner 当前有 `outstanding_action.action=commit_only` 时，用户启动 auto-loop 本身就是本次 run 内任务相关本地提交的预授权。此时仍必须由 AI 生成并自检提交计划，但可跳过二次聊天确认；计划只允许 commit-only、不 push/merge/release/archive，并在成功后把 commit hash 回写 runner。预检不安全时只标记当前 auto-loop item blocked/skipped，后续队列仍可继续。
 
 支持多仓库（frontend / backend 等），merge 目标分支记录在 `.trellis/config.yaml` 的 `packages.<name>.merge_target` 中。
 
@@ -80,6 +80,8 @@ Step 6  输出结果
 
 任一条件不满足时，不得使用预授权。普通 `trellis-push` 继续走“展示计划并等待确认”的规则。
 
+auto-loop 预授权路径不使用额外提交 helper。AI 负责根据当前任务 artifacts、`git status`、`git diff` 和必要的文件内容判断文件归属，输出 planned files / retained files / commit message / 归属理由；`trellis-push` 负责边界复核：当前 action/profile/task 匹配、staged 区为空、无冲突、planned files 当前 dirty、planned files 不含 `.trellis/.runtime/`、`.trellis/.route-prefs.tmp`、其他任务目录或未解释文件。通过后只 `git add` planned files 并执行本地 commit，随后回写 `auto_loop.py record --action commit_only --result ok --commit <hash> ...`。预检失败时回写 `record --result blocked`，该 blocked 只影响当前 item；auto-loop runner 后续 `next` 应继续寻找后续 pending 任务。只有 merge/rebase 冲突、repo 状态不可读、脚本损坏或用户明确 stop 才停止整个 run。
+
 ### 0.2 发现候选 Git 仓库
 
 读取 `.trellis/config.yaml` 中的 `packages` 配置，但不要把 `git: true` 当作唯一仓库发现规则。`git: true` 主要用于 Trellis session context 展示独立包仓库状态；`trellis-push` 必须以实际 Git root 为准发现可提交仓库：

package/enhancements/0.6/.claude/skills/trellis-route/scripts/route_state.py

@@ -160,36 +160,63 @@ def _read_prefs(repo_root: Path) -> dict[str, str]:
     return prefs
 
 
+def _running_auto_run_paths(repo_root: Path) -> list[Path]:
+    """扫描当前项目内所有 running auto-loop run。"""
+    running: list[Path] = []
+    for path in sorted(_auto_loop_dir(repo_root).glob("auto-*.json")):
+        state = _read_json(path)
+        if state.get("status") == "running":
+            running.append(path)
+    return running
+
+
+def _auto_state_path(repo_root: Path, run_id: Any) -> Path | None:
+    """把 run id 转成状态文件路径，非法值返回 None。"""
+    if not isinstance(run_id, str) or not run_id.strip():
+        return None
+    return _auto_loop_dir(repo_root) / f"{run_id.strip()}.json"
+
+
 def _auto_route_mode(repo_root: Path, context_key: str, target: str) -> tuple[str | None, Path | None, str | None]:
-    """读取当前 auto-loop run 的临时 route 授权。
+    """读取当前 running auto-loop run 的临时 route 授权。
 
     auto 授权低于个人 `.route-prefs.tmp`，只在当前 session runtime 绑定了
-    `current_auto_run`，或全局 current 指针能指向唯一 running run 时生效。命中后
-    `resolve` 会写回 session runtime，后续压缩恢复仍走普通 route 决策路径。
+    running `current_auto_run`，或全局 current 指针能指向 running run 时生效。
+    如果这些指针 stale，则忽略 stale pointer，并 fallback 到唯一 running run。
     """
     session_path = _session_path(repo_root, context_key)
     session = _read_json(session_path)
-    run_id = session.get("current_auto_run")
-
-    if not isinstance(run_id, str) or not run_id.strip():
-        pointer = _read_json(_auto_loop_pointer(repo_root))
-        run_id = pointer.get("run_id")
-
     candidate_paths: list[Path] = []
-    if isinstance(run_id, str) and run_id.strip():
-        candidate_paths.append(_auto_loop_dir(repo_root) / f"{run_id.strip()}.json")
-    else:
-        for path in sorted(_auto_loop_dir(repo_root).glob("auto-*.json")):
-            state = _read_json(path)
-            if state.get("status") == "running":
-                candidate_paths.append(path)
-        if len(candidate_paths) != 1:
-            return None, None, "no-unique-auto-run"
-
-    path = candidate_paths[0]
+    stale_paths: list[Path] = []
+
+    for run_id in (session.get("current_auto_run"), _read_json(_auto_loop_pointer(repo_root)).get("run_id")):
+        path = _auto_state_path(repo_root, run_id)
+        if path is None:
+            continue
+        state = _read_json(path)
+        if state.get("status") == "running":
+            candidate_paths.append(path)
+        else:
+            stale_paths.append(path)
+
+    if not candidate_paths:
+        candidate_paths = _running_auto_run_paths(repo_root)
+    unique_paths = []
+    seen: set[str] = set()
+    for path in candidate_paths:
+        key = str(path)
+        if key in seen:
+            continue
+        seen.add(key)
+        unique_paths.append(path)
+    if len(unique_paths) != 1:
+        reason = "no-unique-auto-run"
+        if stale_paths and not unique_paths:
+            reason = "stale-auto-run-pointer"
+        return None, stale_paths[0] if stale_paths else None, reason
+
+    path = unique_paths[0]
     state = _read_json(path)
-    if state.get("status") != "running":
-        return None, path, "auto-run-not-running"
 
     auth = state.get("route_authorization")
     if not isinstance(auth, dict):

package/enhancements/0.6/overrides/workflow-states/in_progress-inline.md

@@ -2,6 +2,7 @@
 HIGHEST PRIORITY SKILL-GARDEN STATE GUARD (in_progress-inline):
 This state block is a breadcrumb; the top-level skill-garden hub is the source of truth for route details.
 Before the first implement route, read `<task>/brief.md` if present and restate the task brief in chat. If it is missing, read the task artifacts and suggest backfilling brief; do not silently rely on memory.
+Before procedural or high-impact actions, run `.trellis/scripts/spec_router.py` with a short query describing the intended action and read any matched SOP/spec before acting.
 Inline workflow-state is not an inline route decision. At Phase 2.1/2.2, reuse only an explicit target-matched `route_decision`; otherwise MUST load/read/use `trellis-route(implement|check)` (or its local `SKILL.md`) to resolve session runtime state/prefs, write the resolved decision, or show numbered fallback and wait.
 Plain preferences, ordinary summaries, `codex-mode`, raw `.runtime` files, and empty/old prefs are not route evidence by themselves; only `trellis-route` may validate runtime route state.
 User reselect/override/use-X-this-time/clear-default wins over remembered route evidence, runtime state, and prefs.

package/enhancements/0.6/overrides/workflow-states/in_progress.md

@@ -2,6 +2,7 @@
 HIGHEST PRIORITY SKILL-GARDEN STATE GUARD (in_progress):
 This state block is a breadcrumb; the top-level skill-garden hub is the source of truth for route details.
 Before the first implement route, read `<task>/brief.md` if present and restate the task brief in chat. If it is missing, read the task artifacts and suggest backfilling brief; do not silently rely on memory.
+Before procedural or high-impact actions, run `.trellis/scripts/spec_router.py` with a short query describing the intended action and read any matched SOP/spec before acting.
 At Phase 2.1/2.2, reuse only an explicit target-matched `route_decision`; otherwise MUST load/read/use `trellis-route(implement|check)` (or its local `SKILL.md`) to resolve session runtime state/prefs, write the resolved decision, or show numbered fallback and wait.
 Plain preferences, ordinary summaries, `codex-mode`, raw `.runtime` files, and empty/old prefs are not route evidence by themselves; only `trellis-route` may validate runtime route state.
 User reselect/override/use-X-this-time/clear-default wins over remembered route evidence, runtime state, and prefs.

package/enhancements/0.6/overrides/workflow-states/no_task.md

@@ -3,6 +3,7 @@ HIGHEST PRIORITY SKILL-GARDEN STATE GUARD (no_task):
 Creating or resuming a task is not implementation permission.
 After PRD is ready and the task is started, the next implementation action is Phase 2.1 `trellis-route(implement)` unless a valid current-task implement route decision already exists.
 If no active task exists, scan `.trellis/tasks/*/task.json` once per session for in-progress tasks with `last_push_snapshot`; surface `completed_steps` + `next_step` and suggest rebinding the active task before resuming.
+Before procedural or high-impact actions, run `.trellis/scripts/spec_router.py` with a short query describing the intended action and read any matched SOP/spec before acting.
 Do NOT call the harness built-in plan mode (`EnterPlanMode` / `ExitPlanMode`) for Trellis planning. It is not a substitute for Trellis task-creation consent, Trellis planning, or the route gate. For complex work, classify the turn, ask for task-creation consent, then use `trellis-brainstorm`.
 If the turn is a meta edit to Trellis itself and Trellis tracking would be overkill, say so and ask to skip Trellis; never silently swap built-in plan mode in for the consent gate.
 <!-- END skill-garden workflow-state no_task v0.6 -->

package/enhancements/0.6/overrides/workflow-states/planning-inline.md

@@ -4,5 +4,6 @@ Planning is not implementation permission.
 Complete prd.md + required context first.
 If the active workflow later routes to sub-agent execution, required context includes real curated entries in both `implement.jsonl` and `check.jsonl`; the seed `_example` row alone is not ready.
 Before `task.py start`, use `trellis-task-brief` to refresh `brief.md` from the latest task artifacts and display it in chat for review.
+Before procedural or high-impact actions, run `.trellis/scripts/spec_router.py` with a short query describing the intended action and read any matched SOP/spec before acting.
 After status becomes in_progress, next action = `trellis-route(implement)`, not direct edits.
 <!-- END skill-garden workflow-state planning_inline v0.6 -->

package/enhancements/0.6/overrides/workflow-states/planning.md

@@ -4,5 +4,6 @@ Planning is not implementation permission.
 Complete prd.md + required context first.
 For sub-agent-dispatch platforms, required context includes real curated entries in both `implement.jsonl` and `check.jsonl`; the seed `_example` row alone is not ready.
 Before `task.py start`, use `trellis-task-brief` to refresh `brief.md` from the latest task artifacts and display it in chat for review.
+Before procedural or high-impact actions, run `.trellis/scripts/spec_router.py` with a short query describing the intended action and read any matched SOP/spec before acting.
 After status becomes in_progress, next action = `trellis-route(implement)`, not direct edits.
 <!-- END skill-garden workflow-state planning v0.6 -->

package/enhancements/0.6/overrides/workflow.md

@@ -6,10 +6,25 @@
 
 **Priority**: This hub overrides any conflicting Trellis workflow, skill, or command text for the scoped behaviors below.
 
-**Scope**: Phase 1.4 task brief handoff, Phase 2.1 implement routing, Phase 2.2 check/check-all routing, current-task route reuse, post-check stop, auto-loop commit-only preauthorization, Phase 3.4 trellis-push, explicit Phase 3.5 finish-work bookkeeping, and push-progress recovery. State blocks should keep one short skill-garden sentinel; long-form rules live here.
+**Scope**: project knowledge discovery for procedural/high-impact actions, Phase 1.4 task brief handoff, Phase 2.1 implement routing, Phase 2.2 check/check-all routing, current-task route reuse, post-check stop, auto-loop commit-only preauthorization, Phase 3.4 trellis-push, explicit Phase 3.5 finish-work bookkeeping, and push-progress recovery. State blocks should keep one short skill-garden sentinel; long-form rules live here.
 
 **Mechanical rule**: use this hub as the source of truth. Do not add separate top-level skill-garden override sections or multiple skill-garden sentinels inside the same `workflow-state:*` block.
 
+#### Project Knowledge Discovery
+
+Before procedural or high-impact actions, run project knowledge discovery:
+
+```bash
+python3 ./.trellis/scripts/spec_router.py "<short query describing the intended action>"
+```
+
+Build the query from the current user request plus relevant immediate context:
+the intended action, commands about to run, files or systems involved, package/layer,
+and domain words such as release, publish, deploy, migration, config, CI, workflow,
+hooks, rollback, data fix, or destructive command.
+
+Read any matched SOP/spec files before acting; if nothing matches, continue normally.
+
 #### Task Brief Handoff
 
 Before Phase 1.4 `task.py start`, use `trellis-task-brief` to refresh `<task>/brief.md` from latest task artifacts, display it in chat, and wait for user confirmation.