fix(bootstrap): isolate objective prose from generated state grammar

[yilin-succeed]

Summary

Quote-isolate the Objective body at both state-writing boundaries so Markdown examples inside a goal objective can no longer open machine grammar and hide (or fake) the generated Todo sections. A fenced objective can currently bootstrap successfully while Todo readback returns no work.

• render_objective_markdown (bootstrap) quotes each objective line into a blockquote; render_state_markdown and project-registration _state_markdown embed it under ## Objective. The frontmatter objective field keeps the raw text unchanged.
• Generated-state grammar keeps full authority: no parser, visible_markdown_lines, region, authority, or transition changes. Quoted lines cannot open fences, comments, headings, or task lines, so both the Python and TypeScript readers stay protected by the writer boundary.
• Chat context readback (_active_state_section) strips the quote prefix so the Objective round-trips verbatim; non-quoted legacy lines keep their existing bullet flattening.

Scope: the two objective writers, one chat section readback, and focused tests. No Todo parser, authority, scheduler, provider, or public CLI interface changes. Repairing already-malformed live goal state stays out of scope.

Issue Or Task

Addresses #4401. I noticed #4402 covers the same issue with a quote + HTML-escape + frontmatter JSON-serialization approach; this PR is an alternative, smaller-diff isolation variant (quote-only, no escape/decode roundtrip, frontmatter untouched) offered for maintainer comparison. If #4402 is preferred, feel free to close this one.

Validation

• New regression tests/control_plane/test_objective_isolation.py (15 cases): inline and tilde fences, unclosed HTML comment, heading-like and task-like objective examples, plain-text and closed-multiline-fence controls, frontmatter fidelity, chat readback round-trip, and the project-registration writer.
  python3 -m pytest -q tests/control_plane/test_objective_isolation.py — 15 passed
• Focused suites around both writers and the reader — all green:
  python3 -m pytest -q tests/cli_commands/test_project_registry.py tests/control_plane/test_todo_next_action_settlement.py tests/control_plane/test_goal_terminal_no_followup.py tests/control_plane/test_todo_document_ownership.py tests/control_plane/test_todo_projection_recovery.py tests/test_chat_completed_todos.py tests/test_chat_server_cors.py tests/test_agent_onboarding_unconnected_project.py tests/test_chat_image_attachments.py tests/test_global_registry_write_serialization.py tests/test_state_file_containment.py — 229 passed
• ruff check on the changed files and python3 -m py_compile on the changed modules pass.

Signed-off-by: yilin-succeed 204474593+yilin-succeed@users.noreply.github.com

[huangruiteng]

详细中文评审

审查对象：e15ee864ba625de1ae73d25324c9168d941a7bf4（fix(bootstrap): isolate objective prose from generated state，base main，4 文件 +149/-8）。执行契约 policy_revision=3；结果 JSON 已通过 pr-review --check-result。

动机

生成状态文件把自己的语法和运营方写的 objective 混在同一层：## Objective 段落直接写入原始文本。于是 objective 里只要出现未闭合或 start-goal-collapsed 的代码围栏、未闭合 HTML 注释、一个 ## Agent Todo 示例或一行待办形状的文本，就会打开机器语法，把其后的生成段落吞掉。症状很糟：Todo 源静默消失，没有任何报错。本 PR 让 objective 逐行进入引用块，使这些字符只能是内容。

改动思路

在两个 renderer（loopx/bootstrap.py::render_state_markdown 与 loopx/control_plane/projects/registry.py::_state_markdown）里把 objective 逐行加 > ，并在 chat server 读回时把前缀剥掉，保证 frontmatter 与读回文本仍与原文一致。方向正确，边界也抓得准：围栏、注释、标题、待办形状四类都被覆盖。

具体改动

• render_objective_markdown（loopx/bootstrap.py:468）：逐行引用，空行写 >；frontmatter 保持原始 objective 不做二次改写。
• _state_markdown（loopx/control_plane/projects/registry.py:145）：注册路径复用同一渲染。
• _active_state_section（loopx/chat_server.py:132）：新增 > 前缀剥离分支，legacy 未引用状态仍走原有 bullet 剥离。
• 新增 tests/control_plane/test_objective_isolation.py（119 行，15 个用例）：五类边界 objective × 「Todo 段落仍可解析」「读回与原文一致」两组断言。

对主干的风险

本 head 过不了仓库门禁。 在 head e15ee864b 的独立 worktree 实跑 tests/canary/test_maintainability_ratchet.py：1 failed, 7 passed，报告 unreviewed finding: module_metric_budget:loopx/chat_server.py。原因很具体：loopx/chat_server.py 的已登记体积基线就是 1513 行（等于 main 当前值），本 PR 又给它加了 6 行（1519），于是变成一项未评审的控制面债务。远端对应 test-shard (4) fail，聚合作业 pytest 与 merge-gate 随之 fail。

第二项风险是归属：## Objective 段读取在仓库里已经有一个 owner（loopx/control_plane/goals/active_state_sections.py::active_state_sections，按 SECTION_HEADING_PATTERN 抽段，loopx/status.py、loopx/state_projection.py 都有同名模式）。本 PR 把新逻辑留成 chat server 的私有 helper，等于新增第二份段扫描实现。把读回搬到 goals 边界上下文可以一次解决体积与重复两项。

方向性风险：本 head 只做引用隔离，未覆盖 HTML 元字符与 U+0085/U+2028/U+2029 行分隔符；同一边界的另一份并行实现（#4402）覆盖了这两类。两 PR 目标是同一个洞，建议合并为一个节点交付，而不是让评审在两份实现间反复比对。

我的整体评价

REQUEST_CHANGES。修复方向正确、用例扎实（15 passed），change_proportionality=proportionate、authority_semantics=aligned、observable_semantics=intentional_change_validated；但 repository_reuse=unjustified_duplication 且存在未解决阻塞发现，两项都指向同一处修复。

最小修复路径（择一）：

1. 把 Objective 段读取从 loopx/chat_server.py 移出，复用/扩展 loopx/control_plane/goals/active_state_sections.py——同时清掉体积越界与重复实现；
2. 若要与并行 PR 收敛，直接关闭本 PR并让 #4402 承担该修复（#4402 已把规则搬进 active_state_metadata，且 ratchet 在本 head 为绿）。

验证：head e15ee864b 独立 worktree 内 pytest tests/control_plane/test_objective_isolation.py -q → 15 passed；pytest tests/canary/test_maintainability_ratchet.py -q → 1 failed（chat_server 体积越界）。远端 checks：test-shard (4)/pytest/merge-gate fail，其余 pass 或 skipping。

---

English verdict: REQUEST_CHANGES at e15ee864ba625de1ae73d25324c9168d941a7bf4. The direction is right and its own 15 boundary tests pass, but the head cannot pass the repository gate: the canary maintainability ratchet reports an unreviewed module_metric_budget finding on loopx/chat_server.py (1519 lines vs the registered 1513-line ceiling, i.e. +6 lines in an already-ceilinged module), and the remote test-shard (4), pytest and merge-gate checks fail for that reason. The second finding is ownership: section extraction already has an owner in loopx/control_plane/goals/active_state_sections.py, so the new private scanner in chat_server.py is a duplicate. Minimum repair: move the Objective section readback into the goals bounded context (clears both the size ceiling and the duplication), or consolidate with the parallel same-boundary PR #4402 and close this one.

[huangruiteng]

动机

REQUEST_CHANGES。 审查 head：cd921c260eaff2dc691745b9558a5d554bb7d93e。

这个 PR 要修的洞是真的：生成状态文件把机器语法和 operator 写的 objective 放在同一层，于是 objective 里一个未闭合代码围栏、未闭合 HTML 注释、一个 ## Agent Todo 示例或一行待办形状的文本就能打开机器语法、把后面的生成段落吞掉，而且静默无报错。方向、边界选择都对，上一轮 review 提出的两个问题（chat_server 体积越界、Objective 段读取的归属）在这个 head 里也都处理了。但现在的结论不是"还差一点"，而是这个修复已经用更好的实现合进 main 了。

改动思路

• 写入侧：render_objective_markdown（loopx/bootstrap.py:468）把 objective 逐行加 > ；registry._state_markdown 复用同一渲染。
• 读取侧：上一轮之后作者把段读取从 chat_server 私有 helper 提取为 loopx/control_plane/goals/active_state_sections.py::active_state_section_text，chat_server._active_state_section 改为委派（这是本 head 相对 e15ee864b 的唯一改动）。
• 第三个提交是空提交 chore(ci): re-trigger pull request workflows，用于重新触发 CI。

具体改动

关键代码讲解

1. render_objective_markdown："\n".join(f"> {line}" if line.strip() else ">" ...)——隔离语法，但不做转义。
2. active_state_section_text(state_text, heading, normalize_text=...)：> 前缀剥离走引用路径，其余行走 legacy - 扁平化；注释行与空行跳过。
3. tests/control_plane/test_objective_isolation.py：+119 行、6 个测试函数（含参数化共约 15 个用例）。

独立验证（exact head cd921c260e，worktree /private/tmp/pr4418-r20）

• pytest -q tests/canary/test_maintainability_ratchet.py tests/control_plane/test_objective_isolation.py → 23 passed。上一轮 review 的体积阻塞已解除：loopx/chat_server.py 现在 1502 行（基线是 1513）。
• 但这个修复已经合入 main：并行 PR #4402（fix(goals): isolate Objective presentation and unify legacy readback，作者 ehz0ah）已于 2026-09-15T12:19:37Z 合并，而本 head 是在 15:52:46Z 才推送的。main 现在：loopx/bootstrap.py:518 用 markdown_blockquote(objective) 写入，loopx/control_plane/goals/active_state_metadata.py:37-38 的 markdown_blockquote 是 f"> {escape(line, quote=False)}"，读取用 unescape(line[2:])。
• 我在 main 上跑了一个恶意 objective 探针：\``\n## Agent Todo\n- [ ] evil\n<!-- open comment→ 写出的行是'> '`、`'> ## Agent Todo'`、`'> - [ ] evil'`、`'> <!-- open comment'`，读回得到 `' ## Agent Todo - [ ] evil <!-- open comment'，而且紧随其后的生成段 ## Agent Todo仍能解析为['', '- [x] real', '']。也就是说 main 的实现是**同一规则 + 额外转义 + 同一回归覆盖**（tests/control_plane/test_goal_objective_projection.py 已覆盖 open-comment、Unicode 分隔符、## Agent Todo` 示例、围栏示例、长 objective 读回）。
• 合并状态：mergeStateStatus=DIRTY，冲突文件正是 loopx/bootstrap.py、loopx/chat_server.py、loopx/control_plane/projects/registry.py 这三个 #4402 已经改过的文件。
• 远端检查：该 head 是 fork PR（yilin-succeed/loopx），statusCheckRollup 为空、没有任何 check run，因此必需检查无法观测（不是绿色）。

对主干的风险

[P1，重复实现] 本 PR 与已合并的 #4402 争同一个边界，且实现更弱。 定位：loopx/bootstrap.py:468-481 与 loopx/control_plane/goals/active_state_sections.py:30-57，对照 main 的 loopx/control_plane/goals/active_state_metadata.py:37-57。合入的后果是同一条规则出现两个同名读取器（active_state_metadata.active_state_section_text 与 active_state_sections.active_state_section_text）和两份写入路径，而仓库规则明确要求"一个规则一个 owner、compress rather than append"。最小修复：作为已合并变更关闭，或 rebase 后只保留 main 缺失的那部分（目前看不到这类内容）。

[P2，实现弱于 main] 引用隔离没有转义。 f"> {line}" 不处理 HTML 元字符与 U+0085/U+2028/U+2029 行分隔符；main 的 escape(line, quote=False) + 读回 unescape 覆盖了这两类（这正是上一轮 review 提到的并行实现差异）。若两者共存，同一输入会得到不同读回文本。最小修复：直接复用 active_state_metadata.markdown_blockquote，不要本地再写一份引用 helper。

主机侧状态（非代码结论）：分支与 main 冲突（DIRTY），且 fork 的 CI 未运行（零 check run），所以"必需检查"在本 head 是不可观测而非通过；这本身不是代码缺陷，但意味着即便代码方向可取，本 head 也不具备可合并状态。

我的整体评价

修复思路正确、边界清楚，这一轮也确实把上一轮指出的体积与归属问题都处理干净了（ratchet 8 passed、隔离用例 15 passed）。问题在于时机与收敛：#4402 已经用更完整（带转义）的实现把这个洞在 main 上补掉了，并且带了同边界回归测试，本 head 现在既和 main 冲突，又会把一条规则拆成两个 owner。因此建议不要继续在这个分支上叠改动，而是以"已被 #4402 取代"关闭；如果作者认为 main 的实现仍缺某个具体用例（例如某个分隔符或读回场景），请只把这个用例作为最小增量提交，而不是重做写入/读取两侧。未批准、未合并。

English verdict: REQUEST_CHANGES at cd921c2. The direction is correct and this head fixes both earlier blockers (the maintainability ratchet is clean again - loopx/chat_server.py is 1502 lines, and the Objective readback now lives in loopx/control_plane/goals/active_state_sections.py instead of a chat_server-private helper; 23 local tests pass). But the fix already shipped on main: the parallel PR #4402 merged at 2026-09-15T12:19:37Z, before this head was pushed at 15:52:46Z, and main now writes the objective with markdown_blockquote (escape(line, quote=False)) and reads it back through active_state_metadata.active_state_section_text with unescape. An independent hostile-objective probe on main shows the escaped round-trip and the generated ## Agent Todo section still parsing, and main's own test file already covers open-comment, Unicode separators, ## Agent Todo examples, fenced examples and long readback. Landing this head would add a second same-named section reader and a quoting writer without escaping, and the branch conflicts with main on exactly the three files #4402 touched. Minimum repair: close as superseded, or rebase and keep only a delta main lacks. Required checks are unobservable for this head (fork PR, zero check runs), so repository_required_checks is recorded unverified rather than green. No merge performed.