fix(goal): derive titles and reuse existing agent frontier

[NIU-123370]

Summary

• Fix goal 标题由 /loopx 创建时固定用项目名，而非任务目标 #3577: derive a public-safe goal display title from /loopx <goal text>, while preserving an explicit --display-name override.
• Fix Guided existing-Agent takeover still requires unconditional Todo authoring #3586: when guided start reuses an agent with an open unblocked advancement Todo, project todo_delta: reuse_existing and continue that frontier instead of unconditionally planning and adding Todos.

Issue Or Task

• Closes goal 标题由 /loopx 创建时固定用项目名，而非任务目标 #3577
• Closes Guided existing-Agent takeover still requires unconditional Todo authoring #3586

Validation

• examples/bootstrap-command-pack-smoke.py
• loopx check on changed files (0 errors)
• tests/test_slash_command_install.py + tests/test_pi_goal_mode.py (49 passed)
• tests/test_host_loop_activation.py — 10 existing TraeX fixture/environment failures reproduced on clean base; 70 passed.

Type of Change

• Bug fix
• Test update

LoopX Area

• Control plane (goals, todos, quota, scheduler, registry, runtime)

Technical Direction

• Core control-plane hardening
• Target base branch: main

Boundary Checklist

• I did not commit private state, credentials, or local artifacts.
• I kept the change scoped to the linked issues.
• Every commit includes a DCO Signed-off-by trailer.

[huangruiteng]

详细中文评审

结论：REQUEST_CHANGES
评审 exact head： b9fd93075c06e4841a660e3e71a297ca9116367c

动机

这个 PR 同时修复两个 guided start 问题：一是让新 Goal 的公开展示名从用户目标文本派生，而不是退化为项目目录名；二是当解析出的 agent 已经拥有可执行的 advancement Todo 时，复用该 frontier，避免再次规划并写入重叠 Todo。两个目标都直接改善 /loopx <goal text> 的首次连接和接管体验，方向合理。

改动思路

展示名路径把 derive_goal_display_name 放在 bootstrap CLI 边界，使用现有 public_safe_compact_text；显式 --display-name 仍优先。frontier 路径新增 existing_runnable_todo_for_agent 读取 durable active state，再由 build_loopx_bootstrap_command_pack 投影 todo_delta=reuse_existing|add_new，最后 build_goal_todo_frontier_steps 决定生成 continue_existing_frontier，还是保留 plan_ranked_todos + write_ordered_todos。

期望正向路径：selected agent 有 status=open、task_class=advancement_task、且未被 gate/block 的 Todo → 复用其 todo id，不生成新 Todo。
期望负向路径：该 agent 只有 blocked/deferred Todo、continuous monitor、user action 或其他非 advancement 工作 → 不应把它当作可继续的 advancement frontier，仍应生成新的 advancement Todo 计划。

具体改动

exact base..head 为 6 个文件、+374/-35：生产代码约 +194/-35，durable bootstrap smoke +180。生产模块和调用点都已接入，不是 test-only scaffolding；但主要负向状态语义尚未被测试覆盖。

1. bootstrap.derive_goal_display_name：把 objective 压缩为 132 字符公开安全标题；bootstrap CLI 在没有显式 override 时调用它。
2. existing_agent_frontier.existing_runnable_todo_for_agent：解析 active-state agent Todo 并选择 resolved agent 的候选，这是 todo_delta 的权威分类点，也是当前 blocker 所在。
3. bootstrap_command_pack.build_loopx_bootstrap_command_pack：在 host identity 解析后读取候选，并把存在性直接映射为 reuse_existing。
4. existing_agent_frontier.build_goal_todo_frontier_steps：将 typed projection 翻译为 guided steps；reuse 分支删除 planner/write steps，add 分支保持旧行为。
5. bootstrap_command_pack.build_start_goal_guided_packet：把 frontier steps 放进 connect/bind 之后、refresh/activation 之前，顺序合理。

对主干的风险

[P1] 仅凭“claimed 且未 done”会把 blocked Todo 或 monitor 误判成 runnable advancement frontier

• 触发条件： resolved agent 只有一个 status=blocked, task_class=advancement_task 的 Todo，或只有一个 status=open, task_class=continuous_monitor 的 Todo。
• 代码路径： existing_runnable_todo_for_agent 遍历 agent_todos.items，只检查 item.get("done")、不存在于 parser projection 的 item.get("blocking")，以及 claimed_by；它没有检查 typed status 或 task_class。该 item 被返回后，build_loopx_bootstrap_command_pack 在 839-843 行把 todo_delta 设为 reuse_existing，build_goal_todo_frontier_steps 随即只输出 continue_existing_frontier。
• 错误结果： guided start 删除 plan_ranked_todos 和 write_ordered_todos，却把不可执行的 blocked Todo 或仅观察的 monitor 当作当前 advancement frontier；agent 会被重新送回阻塞/观察通道，用户的新目标没有可运行 Todo。
• 位置： loopx/control_plane/goals/existing_agent_frontier.py:35-39。
• 最小修复： 复用现有 typed helpers，仅接受 todo_item_task_class(item) == TODO_TASK_CLASS_ADVANCEMENT 且 todo_item_is_actionable_open(item) 的 item，再匹配 claimed_by；不要新增 prose/substring 分类。
• 回归测试： 在现有 open advancement 正向用例之外，至少增加 blocked advancement 和 open continuous_monitor 两个负向用例，断言二者均得到 todo_delta=add_new，并保留 planner/write steps。

我独立运行了 examples/bootstrap-command-pack-smoke.py（通过）、changed-file ruff（通过）和 compileall（通过）。另外用 active-state parser 做了四组定向探针：open advancement 正确返回；deferred advancement 不返回；但 blocked advancement 和 open continuous_monitor 都被返回，直接复现上述 blocker。远端 DCO、dependency review、pytest、Windows、build 与 Sonar 全绿。当前 PR 仍为 mergeStateStatus=DIRTY；最新 main 的 bootstrap_command_pack.py 与 start_goal.py 均有内容冲突，即使 blocker 修复，也需要 rebase 后在新 exact head 重跑 required checks。

typed-state lens：todo_delta 本身是明确枚举值，没有 substring heuristic；问题是输入候选没有使用仓库已有的 typed task-class/status selector。domain-neutrality 和 guidance-vs-obligation 没有发现额外问题。展示名默认行为已由 smoke 明确覆盖；代码体量总体可接受，但建议把这组状态矩阵下沉成 focused test，使已很大的 bootstrap smoke 保持薄层集成验证。

我的整体评价

展示名路径和 guided-step 分层都清晰，frontier 复用也是值得合入的行为；相关 future-facing 边界已抽到 control_plane/goals，不需要更大重构。但当前唯一的分类函数没有实现自己声明的“open、unblocked、advancement”契约，会在真实 agent 同时拥有 monitor 或 blocked 工作时抑制新目标 Todo。这是核心行为 blocker。我请求修改；改用现有 typed selector、补齐两个负向用例并完成 current-main rebase 后即可重新评审。

English verdict: REQUEST_CHANGES for exact head b9fd93075c06e4841a660e3e71a297ca9116367c. The title derivation and projection architecture are sound, but existing_runnable_todo_for_agent accepts blocked advancement Todos and open continuous monitors because it never checks typed status/task class. Restrict reuse to actionable open advancement Todos, add the two negative cases, then rebase and rerun required checks.

[huangruiteng]

agent review出来的p1没必要修，直接解决冲突合并吧

[now-ing]

Re-reviewed against the updated head (b9fd930) — the
existing_agent_frontier.py extraction and the todo_delta or "add_new"
fallback are both solid improvements over the first revision; the fallback
now guarantees a full authoring path whenever no frontier resolves at
pack-assembly time. Two items from the earlier round are still open, plus
one new observation:

1. item.get("blocking") still never matches (existing_agent_frontier.py:33).
   The items come from parse_active_state_todos, whose per-item dicts carry
   done and status but no blocking key — upstream models "blocked" as
   status: "blocked" (TODO_STATUS_BLOCKED, todo_done_for_status only
   treats done/deferred as terminal). So a status: "blocked" todo claimed by
   the agent still passes the filter and is returned as the runnable
   frontier, and the packet projects todo_delta: reuse_existing +
   continue_existing_frontier onto a blocked todo — which the new
   docstring ("Return one open, unblocked Todo") explicitly promises not to
   do. Filtering on
   normalize_todo_status(item.get("status")) == "blocked" closes it.
2. Default item_limit still truncates the lookup. The call
   parse_active_state_todos(state_text, goal=goal) omits item_limit, so
   the parser applies its default MAX_STATUS_TODOS_PER_ROLE cap before the
   frontier scan. With more todos per role, the agent's claimed todo can fall
   outside the capped items and the takeover degrades to add_new even
   though a perfectly good frontier exists. Passing item_limit=None matches
   the reuse intent.
3. (new, non-blocking) The reuse branch has no execution-time escape
   hatch. The reuse_existing decision is fixed at pack-assembly time from
   a state snapshot, and the projected steps carry only
   existing_todo_id + purpose — no todo add template. If the frontier
   todo gets completed or turns out not to cover the requested continuation
   by the time the model acts, a faithful model has no in-packet path to
   author the missing Todo. Attaching an add_new_command_template to the
   step (bounded escape hatch) would cover that gap without changing the
   default flow.

[NIU-123370]

Thanks @now-ing — all three items addressed in 3a6f975e (pushed to this branch):

1. item.get("blocking") never matches — fixed. The filter now uses normalize_todo_status(item.get("status")) == TODO_STATUS_BLOCKED, so a status: blocked todo claimed by the agent is skipped and the docstring promise ("one open, unblocked Todo") actually holds. Regression: test_start_goal_guided_skips_blocked_agent_frontier in examples/bootstrap-command-pack-smoke.py projects a blocked claimed todo and asserts todo_delta == "add_new" with no continue_existing_frontier step.
2. Default item_limit truncates the lookup — fixed. The call is now parse_active_state_todos(state_text, goal=goal, item_limit=None), so the frontier scan sees the full agent todo list instead of the MAX_STATUS_TODOS_PER_ROLE (12) prefix. Regression: test_start_goal_guided_reuse_scans_beyond_default_item_cap places the claimed frontier todo at position 14 and asserts reuse_existing + the correct existing_runnable_todo_id.
3. Execution-time escape hatch — done. The continue_existing_frontier step now carries add_new_command_template (the same todo add template as the write_ordered_todos path, hoisted to a shared constant) plus an add_new_escape_hatch note bounding its use to the case where the frontier todo is terminal or no longer covers the requested continuation at execution time. Default flow is unchanged.

Validation:

• examples/bootstrap-command-pack-smoke.py — passes; the new regressions fail against the pre-fix code (verified by reverting the fix).
• loopx check on changed files — 0 errors, 0 warnings.
• tests/test_slash_command_install.py + tests/test_pi_goal_mode.py — 52 passed.
• tests/control_plane -k "todo or goal" — 721 passed.

[NIU-123370]

Reopened and rebased onto current main (head ef9e19f4).

The branch had been deleted from the fork; restored it from refs/pull/3603/head and merged current main, which had moved 153 commits ahead. Six conflict hunks across loopx/bootstrap_command_pack.py and loopx/cli_commands/start_goal.py, resolved as follows:

• Four signature/call-site hunks were keep-both: this branch adds display_name, main added runtime_root_arg. Both are now threaded through.
• The todo-frontier hunk adopts build_goal_todo_frontier_steps(...) from this branch in place of the two inline steps main still carries.

Two regressions the mechanical resolution would have introduced, fixed in the merge commit:

1. build_goal_todo_frontier_steps hardcoded shell_arg(cli_bin) in add_new_command_template, which drops the --runtime-root flag main had since added to that same todo add template. Now renders via render_cli_command_prefix(cli_bin=..., runtime_root=command_pack.get("command_runtime_root")).
2. The build_start_goal_host_surface_selection_packet call lost display_name when the try/except wrapper from main was taken. Restored.

Validation on the merge head (local venv, Python 3.12):

• examples/bootstrap-command-pack-smoke.py — passes, including test_start_goal_guided_skips_blocked_agent_frontier and test_start_goal_guided_reuse_scans_beyond_default_item_cap.
• tests/test_slash_command_install.py + tests/test_pi_goal_mode.py — 52 passed.
• tests/control_plane -k "todo or goal" — 768 passed, 1165 deselected.
• ruff check on the five changed Python files — all checks passed.

mergeable is now true; mergeable_state is blocked pending required checks/review rather than conflicts. Issue #3577 is still open and main still resolves display_name from project_asset.display_name before any goal text, so the underlying bug is unfixed on main.

[huangruiteng]

详细中文评审

结论：REQUEST_CHANGES
评审 exact head： ef9e19f4742170d65fb3abd3461251abbe58fbf7

动机

这个 PR 修复 guided start 的两个真实缺口：从用户目标文本派生公开安全的 Goal 展示名，以及在接管既有 agent 时复用其当前 advancement frontier，避免重复规划和写入 Todo。方向正确，最新 head 也修复了上一轮指出的 blocked Todo 与默认列表截断问题。

改动思路

展示名由 bootstrap CLI 边界上的 derive_goal_display_name 统一派生，显式 --display-name 保持最高优先级。Todo 复用由 existing_runnable_todo_for_agent 从 durable active state 选择候选，再把结果投影成 todo_delta=reuse_existing|add_new；guided packet 据此生成 continue_existing_frontier，或保留 plan_ranked_todos 与 write_ordered_todos。这一分层符合现有 control-plane 边界。

具体改动

exact base..head 共 6 个文件、+498/-36。最新 head 已做到：使用 typed status=blocked 排除阻塞项；以 item_limit=None 扫描完整 Todo 列表；reuse step 携带 add-new escape hatch；main reconciliation 后仍保留 runtime_root。我运行了 bootstrap command-pack smoke（通过）、focused goal/slash 测试（50 passed）、git diff --check（通过），并复现 maintainability quality gate 的 3 个失败。

对主干的风险

[P1] open monitor 仍会被误判为可复用的 advancement frontier

existing_runnable_todo_for_agent 在 loopx/control_plane/goals/existing_agent_frontier.py:39-45 只排除了 done 与 typed blocked 项，然后按 claimed_by 返回第一个候选；它仍未校验 typed task_class=advancement_task。因此当 resolved agent 只有一个 open continuous_monitor Todo 时，函数会返回该 monitor，todo_delta 被设为 reuse_existing，guided packet 删除 planner/write steps，把用户的新推进目标送入观察通道。请复用现有 typed task-class helper，只接受 advancement Todo，并补一个 open continuous-monitor 负向回归用例，断言 todo_delta=add_new。

[P1] 本 PR 新增的模块复杂度债务使 required pytest gate 失败

本地执行 tests/canary/test_maintainability_ratchet.py 与 tests/control_plane/test_m6_quality_gates.py 得到 3 failed / 9 passed；三处失败都指向本 diff 修改的 loopx/bootstrap_command_pack.py：module_metric_budget 有 1 个 unreviewed finding。远端 pytest 也以相同的 3 个 maintainability-ratchet 失败结束。该失败由当前 diff 跨过模块预算触发，不是继承自 main，不能作为无关红灯跳过。请在现有 owner 边界内收敛/抽取本次新增逻辑，使 ratchet 重新通过；不要仅新增例外来掩盖增长。

typed-state lens：blocked 状态现在使用 typed normalizer，改善正确；但 frontier 的 task-class 仍缺少 typed gate。domain-neutrality、behavior-change disclosure、guidance-vs-obligation 没有发现额外 blocker。future-facing pass 已有合适的 control_plane/goals seam，不需要扩大重构；当前只需把分类契约补完整，并让 command-pack 模块回到维护预算内。

我的整体评价

展示名链路、reuse projection 与 escape hatch 的结构都值得保留，上一轮的两项状态/列表问题也已实质修复。但 exact head 仍会把 monitor 当作 advancement frontier，且 required pytest gate 因本 diff 引入的维护性债务失败，因此现在不能批准或合并。修复 typed task-class 选择、补 monitor 负向用例、收敛模块复杂度并让 required checks 全绿后，再按新 exact head 复审。

English verdict: REQUEST_CHANGES for exact head ef9e19f4742170d65fb3abd3461251abbe58fbf7. The latest head fixes blocked-Todo filtering and full-list scanning, but it still reuses an open continuous monitor because the frontier selector does not require typed task_class=advancement_task. The PR also introduces an unreviewed module_metric_budget finding in bootstrap_command_pack.py, causing three required pytest failures locally and remotely. Fix both blockers and rerun checks on the new exact head; do not merge this head.

[huangruiteng]

详细中文评审

结论：APPROVE
评审 exact head： e615152137f41a826da3b804ad79fbefc8f5b139

动机

这个 PR 修复 guided start 的两个真实缺口：从用户目标文本派生公开安全的 Goal 展示名，以及在接管既有 agent 时复用其当前可执行 advancement frontier，避免重复规划和写入 Todo。两项都直接改善 /loopx <goal text> 的首次连接与连续推进体验。

改动思路

展示名继续由 bootstrap 边界统一派生，显式 --display-name 保持最高优先级；默认派生标题不再重复写入生成命令。Todo 复用则集中在 start_goal_todo_delta.py：从 durable active state 扫描完整 Todo 集合，只接纳 actionable-open 的 advancement_task，再按 resolved agent 过滤 claim；无法证明存在 runnable frontier 时保持 fail-closed，回到正常规划和 authoring。

具体改动

exact base..head 共 7 个文件、单一主题。最新 head 已完成上一轮要求的两项修复：

1. existing_runnable_agent_frontier 使用 typed todo_item_is_actionable_open 和 TODO_TASK_CLASS_ADVANCEMENT；blocked、deferred、resume-blocked、blocker、continuous monitor 与 peer-claimed Todo 都不会进入可复用 frontier。
2. parse_active_state_todos(..., item_limit=None) 扫描完整列表，不会因默认展示上限漏掉后部的 actionable Todo。
3. project_prompt.render_goal_start_bootstrap_command 承担紧凑命令渲染，只有显式 override 才加入 --display-name，避免从 objective 派生出的同一文本在命令中重复出现。
4. reuse 分支保留有界 add_new_command_template escape hatch；默认仍是比较并复用 frontier，只有请求未覆盖时才新增 Todo。
5. command-pack 模块增长已收敛到现有 owner 边界，维护性 ratchet 不再报新增债务。

对主干的风险

未发现仍需阻塞合入的问题。typed-state lens 已覆盖 open advancement 正向，以及 blocked/deferred/resume-blocked/blocker/monitor/peer-claimed 负向；分类不依赖 substring denylist。domain-neutrality、behavior-change disclosure 与 guidance-vs-obligation 均未发现问题。相关 future-facing pass 已完成：命令渲染下沉到 project_prompt.py，continuation 分类集中到 start_goal_todo_delta.py，没有引入第二套状态权威。

独立验证：

• tests/test_slash_command_install.py、tests/test_pi_goal_mode.py、tests/control_plane/test_start_goal_compact_projection.py：110 passed；唯一未执行成功的 linked-worktree 用例在创建临时 git commit 时被本机工作时段全局 commit hook 拦截，与 PR 代码无关。
• tests/canary/test_maintainability_ratchet.py + tests/control_plane/test_m6_quality_gates.py：12 passed。
• changed-file Ruff：通过。
• bootstrap command-pack smoke 的非 git-commit 路径及新增状态矩阵均由 focused tests 覆盖；完整 smoke 的临时 commit 同样仅被上述本机 hook 拦截。
• exact head 的远端 Sign-off、dependency review、pytest、build、Windows PowerShell 与 SonarCloud 均为 success。

我的整体评价

这个版本已把展示名默认行为、可复用 frontier 的 typed 语义、完整列表扫描和维护性边界一起收敛好。旧 blocker 均已实质修复，负向状态矩阵足够防止 monitor/blocked work 再次吞掉用户的新推进目标，当前 exact head 可以批准。批准不等于合并；本次只发布 review 结论。

English verdict: APPROVE for exact head e615152137f41a826da3b804ad79fbefc8f5b139. The prior blockers are resolved: reuse is restricted to actionable open advancement Todos, monitor/blocked/deferred/resume-blocked/blocker/peer-owned cases fail closed, full-list scanning avoids projection-cap misses, and derived display names are no longer duplicated in generated bootstrap commands. Focused behavior and maintainability gates pass, changed-file lint passes, and all required remote checks are green. Approval is review-only; no merge was performed.