[loopx-agent]

Goal And Delivered Outcome

原生 context delivery 的首次读取保留了收件来源和读取入口，却漏掉已经生成的未完成委派摘要。一个请求即使已被采纳、关联 Todo 已完成，仍可能欠结果或等待回传；首次读取应直接呈现这个状态。

本次沿现有 TypeScript initialReadPointers 交付 source.content.unfinished_delegations 和 read_complete，保留计数、遗漏、不可用观测和下一步。完整请求正文仍从同一不可变 capture 按需读取，评估、执行或返回请求前仍须读全文。读取不会消费结果、完成请求或增加执行权限。

• Outcome basis: existing manager-context Turn agenda and native first-read contracts; target/base main at c334bbaf196505778149ae4edeefaa013d343259.
• Before → after: the original selector fails both new present/unavailable regressions; the final selector retains the existing summary and incomplete-read marker without expanding the full request body.

Author Declaration

• Written by: model_agent — Codex / GPT-6 family (OpenAI).
• Implemented against: loopx/capabilities/manager_context/README.md (compact unfinished-request agenda), and docs/reference/required-work-context.md (first-read selection), both at the base revision above.

Criterion	Disposition	Boundary	Evidence
Existing bounded unfinished agenda remains visible at Turn start	implemented	turn_envelope.ts::initialReadPointers	Present and unavailable first-read regressions; real File/SQLite CLI journeys
Full details stay expandable; observation does not settle work	implemented	Existing capture reader and collaboration receipts	Cold body recovery; byte-identical receipt files before/after reads; current-capture warm recovery

Scope And Continuation

Four files: one projection change, focused regressions and the existing English/Chinese reference paragraphs. Reuses the capability-owned bounded agenda and existing pointer helper; no new schema, configuration, state machine, permission or request lifecycle.

This is a complete, reversible author proposal for the first-read omission. The wider delegation follow-through, actual adoption, installed behavior and outcome/cost qualification remain open in their existing work. This PR does not prove model retention, better attention or lower total cost. The bounded agenda adds input bytes.

Validation

• Tested revision: 69dd8f5c068cf94b32ff1d33d64def89edff63f8.
• Run state: finished for the checks listed below; complete repository qualification remains held.
• Input classes: synthetic, public_fixture.

Check kind	Result	Evidence / limitation
regression_parity	passed	Two new tests fail against the unchanged base production selector and pass after the fix. Setup refusals and the corrected warm-delta test expectation are retained as failed intermediate observations.
real_entrypoint / real_backend	passed	37 Python tests across test_context_first_read.py, test_delegation_turn_context.py, test_context_read_reference.py; isolated real File/SQLite, two consecutive CLI guards, cold first read and warm recovery of the current capture. Existing host envelopes retained.
regression_parity	passed	5 default/off and unsupported-on context delivery cases; 9 TypeScript pointer-selection tests.
unit	passed	Complete npm run test:control-plane: 5,825 passed, 77 skipped, zero failed/cancelled out of 5,902 tests. Skips are not counted as passing coverage.
static	passed	Control-plane typecheck, configured Ruff roots, mypy (17 configured files), CLI output budgets, diff hygiene and committed RFC bilingual check. No RFC pair changed.
static	passed	Four-file public boundary scan: zero errors/warnings. Agent-only projection; no private operating context or generated artifacts committed.
integration	blocked	Goal-aware quick premerge executed five direct checks, three catalog canaries and one public-boundary check without code-check failures. Aggregate gate is quality_invalid_receipt, exit 1. The isolated managed runtime exited normally after the checks.
manual	blocked	Exact-scope change-quality review/record/verify completed. Receipt remains non-passing because full repository Python pytest has not run. Required validation is retained, not waived. Full CI and independent exact-head review remain pending.

The existing default-provider and capture contracts are unchanged. No authority-store schema, routing or promotion change; no new live PostgreSQL, installed host, cross-platform or sustained-effectiveness qualification is claimed. Dependency setup initially lacked tsc; installing declared dependencies and rerunning typecheck succeeded without lockfile changes.

Frontend / Visual Evidence

• UI impact: none; no frontend, Lark controls, documentation chrome or layout changed. Existing journeys still use the same typed owner. Screenshot/viewport evidence: N/A.

Type Of Change And Area

• Bug fix
• Documentation update
• Test update
• Control plane / host context integration

Shared-authority RFC Fixture Impact

N/A: no migration or shared-authority promotion claim; File/SQLite read-path evidence is scoped above.

Boundary Checklist

• Public-safe source, fixtures and PR body; no private state, credentials, raw traces, internal links or machine paths.
• Existing capability and typed owner reused; no benchmark work duplicated.
• Existing RFC bilingual gate passed; no paired RFC changed.
• UI impact marked none.
• Commit includes DCO Signed-off-by.

Reviewable proposal only. Keep unmerged until required qualification, exact-head review/readiness and maintainer authorization are satisfied.

[loopx-agent]

Reviewer: model_agent; model=gpt-6.1-sol; provider=OpenAI; runtime_reported; reasoning_effort=max

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

当前精确 head 可以批准；CI 未作为批准条件。change-quality qualification 仍是独立的未通过合并门禁，本结论不授予合并权限。

动机 / Motivation

受管理 Turn 唤醒、通过原生文件首读查看当前工作的 Agent。例如一个委托已被采纳，关联 Todo 也已完成，但仍欠原路答复。旧首读只给来源和完整读取路线，隐藏已经生成的待办委托议程，Agent 可能因此漏掉返回步骤；当前首读直接保留该议程及不完整读取标记，完整请求仍须在行动前展开。File、SQLite 的真实 CLI 冷启动和当前 Turn 的 warm recovery 都能读到该议程；两轮观察不改写委托回执。这个切片修复首读的信息遗漏，不证明模型已理解或采用议程、委托已经完成、实际长期效果提升或完整成本下降。

改动思路 / Approach

沿用已有 TypeScript 首读选择器和已有委托投影，只把当前 capture 中两个既有字段加入地址集合；Python 继续负责摘要校验、只读分页和当前捕获恢复，不新增生命周期、权限或配置 owner。当前 PR 只有四个文件 +97/-9：生产选择器新增两条字段选择及两行说明，其余是已有用例扩展和中英参考说明。已有投影按至多 20 个可见请求、10000 行字节收敛，保留遗漏计数、不可用观测和完整读取路线。复用这一 owner 比新增生命周期或展开全部请求更贴合问题；增加输入字节的代价仍需保留，不能把可见性等同成本收益。

正向路径是：现有回执分类出仍欠答复的请求，Turn hook 捕获议程，首读选择同一 capture 中的值，然后调用者在行动前展开完整请求。负向路径中，来源不可用或有遗漏时仍显式保留；损坏 capture、选择 profile 或游标会按原校验拒绝，恢复只读当前 Turn，不重新入场或重放效果。

具体改动 / Changes

关键代码讲解

1. initialReadPointers 通过现有 own-property 地址遍历加入 read_complete 与 unfinished_delegations，字段不存在时不新增值；它不推导完成状态，也不修改回执。
2. capturedDecisionSelection 继续把 initial-read 和发现模式接到同一选择器；普通显式地址选择、未知 schema 和失败语义不变。
3. read_context_capture 是沿用的实际消费者：逐页校验原 SHA 和选择 profile，再返回当前值。selected EOF 仍不表示完整 Inbox 或模型阅读；完整 brief 可以从同一 capture 按需读取。

两个测试文件扩展现有用例：available/unavailable 保留整份摘要与 false 标记；File、SQLite 各自走真实 CLI 的 host envelope 和 native 冷启动／当前捕获恢复，两次读取前后委托 JSON 字节相同。文档的英文、中文补充也保留了“完整请求在评估、执行、回传前读取”的次序。

依据是已存在的manager-context 议程与 partial-read 契约和原生首读契约，两份文档在所核基线与当前目标主干之间未变。对应已有条款标签：unfinished delegation agenda → implemented，保留采纳后工作、待返回及遗漏／不可用事实；read_complete=false → implemented，完整请求仍须展开且观察不消费结果；Native first_read_command → implemented，冷首读和 warm current recovery 走同一当前来源，不把选择 EOF 当完整输入。

对主干的风险 / Risks

最强风险是把部分议程误读成完整 Inbox，或把 Todo 完成、采纳、queued return 当作已经返回成果。当前只搬运已有 typed 投影，不新增字符串分类或状态 owner；整个摘要、未完成标记和原读取路线一起保留。真实 overflow、回执生命周期、损坏来源与选择 profile 的负例通过，但 Agent 仍可能忽略正确输入，不能据此认证模型遵循。

明确 off、unsupported projection 和默认惰性 cache 查找用例通过，原 host envelope 在基线与 head 的两例均通过；没有改变 native 激活默认、权限、调度、预算或效果。实际调用者是已有 managed quota context reader，属于内部投影修复，没有新设置或前端操作。这是现有 agent context 的文本投影，没有页面、布局或设置变化；检查了实际 CLI 返回，不适用截图审查。

语义与 CI 对齐

新增选择复用既有词汇，没有新 Enum、Literal、closed set 或 schema；原 read/adopt、work done、queued、delivered、consumed 与 accepted 区别保留。既有文本中的完整读取义务没有降为可选，机器执行的 SHA／profile／权限校验也没有放松。CI 未查询或等待。标准预合并的 5 direct、5 catalog、8 risk 与 1 public-boundary 命令均 PASS，但整体 FAIL（exit 1），原因是当前 quality_stale_receipt，merge_gate_passed=false；没有改写或豁免 receipt。作者另有完整 Python 资格未完成的说明，也没有当成通过。

我的整体评价 / Verdict

交付判断为这个有实际调用者的首读遗漏修复已完成。long_horizon=preserved：原生命周期和验收边界保留，长期采用与收益未知；user_experience=improved：调用者第一次有效读取就能看见已有委托事实和后续路线。两条现有字段选择无需额外框架或相邻重构，维护成本与已复现问题相称，兼容层也没有增加。

独立验证：当前 39 项 pytest PASS（9.01 秒，零失败、零跳过），9 项 TypeScript 选择器测试、类型检查、Ruff、diff／双语检查、CLI 输出预算以及上述 19 个预合并命令均 PASS。同一组最终新用例在精确基线 c334 得到 4 个预期 FAIL、2 个原 host envelope PASS；head 的同例通过。旧 native 例在冷首读就失败，旧 warm 流程未执行到第二轮；head 的当前 Turn warm recovery 实际执行通过。真实模型理解和采用、安装版读回、跨平台 native 支持、完整 Python 仓库测试、长期跟进效果与完整费用净益均未在本轮验证。主干 change-quality receipt、后续独立合并资格和授权仍由相应 owner 满足；新 head 或具体新证据需要重新核验。

English verdict: APPROVE this exact head for the bounded first-read projection repair; CI is not an approval prerequisite. Change-quality qualification and merge authorization remain unmet, separate gates.