fix(agent): surface pending capsule review details

[sunjiajunsunjiajun]

Summary

• require final responses to list every pending proposal and capsule candidate instead of reporting only a count
• present intent and explanatory text in the language currently used by the user while preserving proposal IDs, capsule names, code symbols, configuration keys, and file paths
• add a capture-skill fallback for responses that end before review is resolved
• document isolated Codex behavior validation plus the approved design and implementation plan

Scope

• no CLI changes
• no proposal or capsule storage changes
• translation is presentation-only and never rewrites stored records

Test plan

• python -m pytest -q (535 passed, 2 skipped)
• isolated Codex behavior check with a real pending proposal

Fixes #35

[Lum1104]

Should not have these hard code plain text only test~

[sunjiajunsunjiajun]

The substring assertions only verify the guidance text, so I ran an isolated host-agent behavior check for the original #35 failure.

Environment:

• temporary Git repository with a real .rgit/ store
• one open proposal, prop_0005bd86bc57
• two stored English candidates: clamp-negative-total and explicit-total-accumulator
• one choice-relevant knob: clamp_negative=false
• current PR guidance and the branch version of rgit-capture/SKILL.md loaded through an isolated skill home
• Codex gpt-5.6-sol, low reasoning
• Chinese user prompt, with approval/dismissal/resegmentation explicitly prohibited

Observed final response:

• listed the proposal id
• listed both stored candidate names
• rendered both one-line intents in Chinese
• included the clamp_negative knob
• preserved proposal id, candidate names, and config key verbatim
• did not replace the details with a candidate count

Post-run verification also confirmed that the proposal remained open and both stored English intents/knobs were byte-for-byte unchanged; no capsule or graph relation was created.

So the specific user-visible failure from #35 did not reproduce with the PR guidance. One narrower observation: the agent surfaced all review details but did not explicitly ask which names to keep, despite that wording in the skill fallback.

Actual final response from the isolated Codex run:

当前状态已核对，未修改任何文件，也未批准、跳过或改写 proposal。

- 分支：`main`
- 代码修改：`calculator.py`
- 未跟踪目录：`.rgit/`
- 已批准 / 应用 / 跳过的 capsule：无
- 图关系：尚未生成；需在 capsule 审批后处理

### 待审核 Capsule Proposal

Proposal：`prop_0005bd86bc57`

- `clamp-negative-total`
  调用方启用限制时，防止汇总结果低于零。
  关键参数：`clamp_negative`，默认值为 `false`。

- `explicit-total-accumulator`
  显式保存中间汇总结果，使后续的总计策略可以集中应用。

该 proposal 及其两个候选均保持原样、继续待用户审核。

[sunjiajunsunjiajun]

Updated the pending-review handoff in 51527d9 to keep it compact: no repeated decision menu, no separate status list unless requested, and one short paragraph that asks for the user's choice and states that the agent will execute the review.

Verification:

• full suite: 537 passed, 2 skipped
• isolated Codex behavior run: gpt-5.6-sol, low reasoning
• one open proposal with two stored English candidates
• Chinese user prompt
• automated checks confirmed the proposal remained open and stored candidates were unchanged
• final response contained one compact decision request, with no command-style reply menu

Sub-agent final response:

当前状态：`main` 分支，`calculator.py` 有未提交修改，`.rgit/` 未跟踪。本轮未修改代码，也未批准、丢弃、改写任何 proposal，尚未执行图关系处理。

### 待审核 Capsule

Proposal `prop_0005bd86bc57`：

- `clamp-negative-total`：调用方启用限制时，防止汇总值低于零。关键参数：`clamp_negative` 默认为 `false`。
- `explicit-total-accumulator`：显式保存中间汇总结果，便于集中应用后续汇总策略。

该 proposal 仍保持待审核；请决定保留哪些 capsule 名称、全部保留或全部丢弃，我会替你执行审核，保留的候选将被批准并存储为 capsules。

[Lum1104]

测试文件里面的硬编码文本要删掉～

[sunjiajunsunjiajun]

Removed the brittle text assertions and documented the isolated Codex behavior check instead. Full suite: 535 passed, 2 skipped.