[loopx-agent]

The ordinary quota/heartbeat context adapter read the complete Goal file and sent its body to work_item.context.project, which deliberately leaves Goal consumption pending and discards that body. This removes the unused internal copy while retaining the full read, UTF-8 validation, path and SHA-256 revision.

The existing TypeScript context owner still requires the complete Goal read before dependent work. Selected Todo, acceptance and scoped User requirements, failure rejection and execution authority are unchanged. This is a source-adapter optimization; it adds no capability, setting or UI journey.

For a synthetic 4,200,108-byte Goal, the exact base/head adapter comparison through the native CLI and real TS runtime reduced internal request JSON from 4,205,594 to 5,466 bytes (File) and 4,205,634 to 5,506 bytes (SQLite). The entire Agent channel and typed projection were equal; CLI stdout stayed 23,849/23,889 bytes respectively. These figures measure internal transport only, not model tokens, latency, fees or long-horizon outcome benefit.

Validation on 2ceff9340af6b1f8f23ce9048d6cecc649c7dd8a:

• 18 focused Python tests passed across Goal transport, complete acceptance and selected-work/heartbeat/host delivery. The new transport-growth assertion failed on the base adapter before the fix.
• Real File/SQLite tests preserve full-file tail hashing, missing-file and invalid-UTF-8 rejection, fresh-guard recovery and legacy/full-input projection parity.
• Standard premerge: 6 direct and 11 selected checks passed, including public-boundary and semantic-vocabulary checks. Exact-scope change-quality receipt cqr_89d6ac1c15720f4ed241 recorded; the final aggregate premerge gate passed. The first run held for that missing receipt, which remains recorded as a failure.
• Diff-scoped semantic advisory found no supported new production vocabulary. It does not accept test paths outside loopx/; the new test was inspected separately.

The Python adapter remains the existing filesystem I/O boundary; TypeScript retains decision semantics. Bounded refactoring is the deletion itself, with no new wrapper or compatibility mode. PostgreSQL authority storage and UI behavior are untouched. The GitHub real-server PostgreSQL job passed; no local PostgreSQL suite or installed runtime qualification is claimed. Maintainer review and merge authorization remain separate.

中文：普通 quota/heartbeat 将完整 Goal 正文传给 TS 后立即丢弃，白白增加内部传输。本变更仅移除该副本，完整读取、UTF-8 校验、全字节摘要、原 Goal 必读命令、验收与失败拒绝均保留。合成大文件的内部请求约从 4.2 MB 降至 5.5 KB，完整 Agent 合同与最终 CLI 输出不变；不据此宣称模型 token、延迟、费用或长程效果改善。18 项定向测试与标准预检代码检查通过，提交正常 PR 供维护者审阅；未安装或合并。

CI follow-up for the same head (merge remains held):

• kernel-static-checks failed on 9 existing Ruff E701/E702 findings in test_quota_acceptance_references.py and test_linked_goal_sources.py. Independent lint runs on immutable base 8696bd6c3ba0f5169b92f83a6aa770c651065d45 and this head reproduce identical file/line/code/message diagnostics with unchanged configuration; the two changed Python files lint clean. The required CI result remains failed.
• The installed-page browser job timed out waiting for .goal-overview-source-note in navigation-sorting.mjs:262. That scenario is outside the changed paths, but a base browser comparison has not been run, so its cause is unattributed, not waived as inherited or flaky.
• The pytest job failed its upstream-result aggregation with checks failed and shards skipped; it did not report a failure from the new test. Dependent aggregate/merge gates remain failed.
• DCO, dependency review and the real-server PostgreSQL job passed. CI/installed-page qualification is incomplete; do not merge based solely on local premerge success.

中文补充：CI仍阻塞。9条Ruff错误已用不可变base/head独立同命令复现，新增Python代码检查干净；安装版浏览器在Goal overview source-note等待超时，尚无base浏览器比较，原因未归定。pytest为上游失败/分片跳过的汇总失败。真实PostgreSQL作业通过，但不能用本地预检替代完整CI或安装资格。

[loopx-agent]

Reviewer: model_agent · gpt-6.1-sol · OpenAI · runtime_reported · reasoning_effort=max

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

APPROVE。当前精确头 2ceff9340af6b1f8f23ce9048d6cecc649c7dd8a 没有阻断发现。 删除的是投影消费者未使用的内部副本，已按完整指令和失败恢复比较，未用请求变小代替语义证明。

动机

通过普通 quota/heartbeat 继续已有 Goal 的 Agent，需要完整目标、任务和停止条件，同时避免无用的内部重复传输。旧实现把整份 Goal 文件放进内部投影请求，但类型化消费者仍要求单独完整读取它；修改后仅传来源路径与全文件修订，Agent 收到的全部要求及必读命令保持相同。同一4,327,443字节文件的 File/SQLite 真实 CLI 对照中，内部请求从4,334,614/4,334,654字节降至7,085/7,125字节，完整 Agent channel 与外部输出大小保持相同，末尾停止条件仍属于完整必读来源。本轮不宣称模型 token、延迟、完整 Turn 成本或效果收益，不改变原始 Goal/任务权威、读取范围或执行许可，也不授予合并或安装权限。

Goal 文件的“可用”与 Agent 已经消费其全部意图是两件事。旧代码先读取文件、传整份正文，消费者却仍保留原完整读取命令，成本来自重复搬运；把必读命令也删掉才会变成丢要求的错误优化。此提案限定在前一种情况，不增加操作者准备、导航或确认轮次。

改动思路

现有 TypeScript owner 在这里不消费 Goal 正文，删除 Python 文件适配器传过去又被丢弃的 text 即可；完整文件仍读取、严格校验 UTF-8，并对全部字节算摘要，不需要摘要器、新预算或第二套选择规则。本 PR 只收敛现有可用性检查的内部传输；完整 Goal 的消费义务、选中 Todo/验收/User 来源、拒绝与恢复仍由现有 typed context owner 承担。

不修改现有 TypeScript 决策 owner，也不按标题猜哪些目标文字有价值。完整 Goal 文件仍是原来源，所选 Todo、验收与相关 User 要求仍通过各自完整 canonical adapter 交付。最有力的反例是：只有一个很小的请求，却忽略文件末尾的停止条件、非法字节或缺失来源。我用同一份可丢弃 File/SQLite 权威状态运行两版真实 CLI，完整比较 Agent channel；观察器只记录并转发实际 TS 请求，没有伪造投影结果。

具体改动

独立规格来自修改前的 docs/reference/required-work-context.md，固定版本 8696bd6c3ba0f5169b92f83a6aa770c651065d45。Registered Goal state 要求完整渐进式读 Goal，不能把可用性当消费；Already delivered sources 要求只移除已履行来源，失败仍需恢复和新 guard；Same contract in TurnEnvelope 要求普通包与签名宿主保持任务、验收及 User 范围。当前头三项均实现，依据分别是完整 channel 等价、真实失败恢复，以及既有签名宿主/选中任务回归。本次中英文补充说明没有反向取消旧规格。

关键代码讲解

• _source_content 的 Goal 分支仍 read_bytes()，再对完整 bytes 严格 UTF-8 解码，最后以全部原字节计算 SHA-256。返回只删 text，保留 path/revision；不是限长读取，也不是忽略解码错误或按正文摘要算版本。
• projectInteractionWorkContext 的481行分支没有变化：合法 Goal 可用性结果仍进入 pending，完整 cat 命令和其“读取原意图、验收与停止条件”的理由保留。它在这里没有读取被删字段；实际给它旧完整正文和新 bodyless 输入，完整投影结果相等。
• attach_work_context 仍捕获来源失败，送原 typed projector，work_context.complete=false 时关闭 dependent delivery。其余 selected Todo、acceptance、scoped User 分支均未改，来源元数据不是新的授权事实。

独立实测同一4,327,443字节文件：File 内部请求4,334,614降至7,085字节，SQLite为4,334,654降至7,125字节；各 provider 的小/大 Goal 请求大小在新头相同。两版完整 Agent channel 逐项相等，CLI stdout大小分别27,386/27,432字节保持相同，未归一化权限、指令、任务或来源修订。追加最终字节会改变摘要；非法 UTF-8 尾字节和来源缺失均拒绝，恢复文件后完整 native CLI 回到原 channel。

对主干的风险

当前头 18项 Python、62项 TypeScript 通过，包含完整任务到真实签名宿主、验收范围、相关/无关 User gate、选中正文变化后拒绝及 TurnEnvelope 原义。把新增真实 File/SQLite 回归复制到不可变基线执行，两例均在内部请求随 Goal 增长的断言失败；头版本通过。该灵敏度与独立完整 channel、尾字节及恢复证据共同证明删掉的是什么，不能仅靠尺寸测试证明不丢语义。

native standard premerge 的6项直接检查、2项 catalog、8项风险 smoke及1项公共边界检查通过，无人工 hold。按当前策略未查询、轮询或等待远端 CI，作者的其他 CI/性能声称未算作我的观测。此变更不修改 authority store，不需要借用活跃 Goal 做迁移实验；本轮真实 File/SQLite 路径通过，不声称 PostgreSQL 存储修改或安装态 App/飞书已获资格。

语义与 CI 对齐

这是已有内部投影字段的复用收敛，不新增类型词汇、自动启用、选择范围或权限。中英文仍明确“完整读取、全部字节摘要、来源错误拒绝、仅可用不等于消费、无模型收益推断”；实际 emitted reason/命令相同。消除序列化副本也没有放宽 Effect 请求硬界限，更没有修改任何冻结实验阈值。

另已读同一来源的开放提案 #7069（86094ab1f2ab7c6adc20c89763116887636dd93f）：它会新增真正消费 Goal 正文的路径，和当前“此消费者不用正文”的前提不同。两者触及同一 adapter/文档，不能不经整合重审就继承本结论或直接组合；这不是本精确头的隐藏消费者，也不要求它等待另一份提案。后续消费者改变时，应重新核完整来源和必读范围。

我的整体评价

APPROVE 的依据是接受范围内的具体内部冗余已删除，而完整目标消费、权威任务来源和恢复路径保持。长期要求、既有用户任务和零额外交互轮次均保留；仅从这项局部字节减少，不能推断实际模型采用、延迟、完整成本或任务效果改善。改动位置正确，现有 I/O 与 typed owner 足够，无需扩大重构。源码回滚不改历史或权威状态；当前没有未解决代码阻断项，合并权限、当前 readiness 和其他提案整合仍需各自成立。

English verdict: APPROVE at 2ceff9340af6b1f8f23ce9048d6cecc649c7dd8a (author-owned COMMENTED conclusion, not formal GitHub self-approval). Matched real File/SQLite CLI runs remove an unused4.3MB internal body while preserving the entire Agent channel, full required Goal read, byte-complete UTF-8/hash validation and actual failure/recovery.18 Python and62 TypeScript tests plus native local premerge pass. No model-token, latency, installed-adoption or total-cost benefit is claimed. #7069 would introduce a different body consumer and requires fresh integration review; no merge authority is granted here.