[loopx-agent]

Goal And Delivered Outcome

Native context delivery previously left an unmarked/mixed registered Goal as a separate progressive read. The exact native full/delta capture therefore could not carry or reuse that complete material. This extracts an independently usable step from #7430: the existing opt-in native path delivers the full original Goal before capture, verifies its digest, and reuses unchanged text through the existing codec.

Ordinary --context-delivery off keeps its progressive source behavior. Generated-region ownership, canonical Todo authority, stops, pending reads, capture identity and settlement remain with their existing owners.

Author Declaration

• Written by: model_agent — OpenAI Codex; the exact deployment model was not independently verified.
• Specification: the required-work-context contract at 2cdff65d886d13121ffd78dccd96dd7413ab9d01, plus the maintainer-requested independent C1 delivery boundary in WIP: converge governed work inputs and packaged heartbeat consumers #7430.

Criterion	Disposition	Owner / validation
Complete original Goal and source provenance before dependent work	implemented	Registered source IO → typed interaction owner; real native File/SQLite full-byte source test
Exact reuse and original-capture loss recovery	implemented	One shared Goal integrity predicate and existing context codec; typed and packaged lifecycle tests
Preserve progressive off behavior and rejection/replay boundaries	implemented	Existing activation and reader; source, quota, Host and settlement checks

Scope And Continuation

The new helper contains the existing Goal-integrity rule plus the complete-file witness, shared by fulfillment and delta eligibility. Python performs registered IO and passes the requested profile; it does not select work or implement a second validity rule. The internal complete_goal_v1 acknowledgement prevents an older runtime from silently ignoring that request.

This is complete within the packaged native source-delivery slice. #7430 retains the complete-contract writer/caller cutover and historical-reader retirement; #7560 covers the separate outcome-reading scope. Neither controller/executor sessions nor all C1 consumers are delivered here. No new configuration, serializer, authority store or frontend is added. Runtime installation and actual App adoption remain outside this PR.

Validation

• Tested revision: d980d621219b10353fdaeb058a1a0fd5714f4389, rebased on 2cdff65d886d13121ffd78dccd96dd7413ab9d01.
• Run state: finished.
• Input classes: synthetic.

Check kind	Result	Evidence / limitation
integration	passed	175 Python tests across Goal source transport/narrative, context CLI, quota projection, acceptance and Host context; 57 TypeScript context/interaction tests
static	passed	Control-plane typecheck, configured mypy (17 files), scoped Ruff, semantic advisory, bilingual diff and whitespace checks
regression_parity	passed	Same native Goal oracle fails on immutable main for both File/SQLite because the source is absent; passes on the fix. All 1,170 loaded baseline modules originate in that frozen tree
real_entrypoint	passed	Two source-matched wheels, isolated native CLI/hooks and File/SQLite: cold/warm, PostCompact and same-capture full recovery, all 24 clauses, tamper rejection, premature-spend refusal, refresh/replay, one debit and retained open work
real_backend	passed	Isolated File/SQLite use the owning store and CLI, preserving authority and once-only settlement; no active Goal was modified
integration	passed	Dashboard/Chat built and both wheels assembled; all 1,888 base / 1,889 head product members equal their frozen source. Initial missing frontend dependencies were repaired with the lockfile; no tracked dependency update
static	passed	Exact-scope CQR is valid with the cost warning below; standard premerge passed 4 direct and 19 selected automatic checks, with no failures or manual holds
manual	not_run	Actual model retention/decisions, tokens/cache/billing, installed App adoption, Windows native lifecycle and long-horizon outcomes

Full-repository pytest, actual-model costs, installed-App adoption and the platform matrix are not qualified by this scoped proposal; no merge readiness is claimed.

The added fixture initially omitted its Turn identity/capture root; those failures remain separate from the later semantic failing-before result. No runtime gate or budget was weakened.

Matched input cost

Same synthetic task, harness, path lengths and page budgets; one cohort with concurrent background work. Total UTF-8 stdout includes CLI, hook output and every required Goal read. These are engineering observations, not serialized model requests or token/latency/causal gains.

Backend / phase	Main	Proposed
File cold	97,188 B	98,347 B
File warm	59,782 B	52,099 B
File original-capture recovery	94,365 B	95,521 B
SQLite cold	97,302 B	98,465 B
SQLite warm	59,892 B	52,209 B
SQLite original-capture recovery	94,473 B	95,631 B

Warm delivery avoids the separate Goal cat; native CLI/hook call counts remain unchanged. Cold and loss recovery add about 1.2 KB, two native CLI/hook calls and one RPC, while removing the separate Goal read. This tradeoff is explicit; the broader C1 comparison and accepted-outcome/full-cost qualification remain open.

The matched private harness is not published; the durable synthetic regression is reproducible with:

uv run --extra test python -m pytest -q tests/control_plane/test_goal_state_context_transport.py tests/control_plane/test_context_delivery_cli.py
node --test tests/control_plane_ts/context_delivery.test.ts tests/control_plane_ts/interaction_contract.test.ts

Visual Evidence

UI impact: none. This changes the source delivered by an existing opt-in integration; no viewport or control changes. Existing activation, CLI readback and native lifecycle are reused. Packaged CLI observations do not certify installed App or actual model adoption.

Type / Area / Boundary

• Bug fix; bounded shared-owner refactor; docs and regression update.
• Control plane and host integration.
• Public-safe synthetic fixtures; no private state, local paths, raw trajectories, credentials or internal links.
• DCO sign-off; no paired RFC changed.
• No self-merge, installation, release, live model, paid experiment or manual CI run.

Rollback: use the existing --context-delivery off for a new invocation. Preserve and finish an in-flight decision through its original capture and settlement identity; disabling presentation is not permission to replace that capture.

[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)

动机 / Motivation

开启 native context delivery 的 Agent 在连续回合中需要读取完整 Goal。以前未标记或混合 Goal 留在另行读取命令中；现在原文进入同次 capture，暖态可精确复用。完整原文、未知章节、注释和最终停止条件能够在原捕获中读取与恢复。本 PR 不完成其余 C1 切换，也不证明模型保留、安装采用或 token 与长程收益。其余 C1 writer/caller 收敛、历史 reader 退出及实际模型完整成本仍由 #7430 承接。

未发现阻塞问题；批准本次准确 head d980d621219b10353fdaeb058a1a0fd5714f4389 的源码交付切片。原渐进读取本身并未丢掉要求，但正文不在 native capture 内，就无法由同一基线精确复用。这个具体缺口值得修复，不能仅凭字节减少判断价值。

改动思路 / Approach

完整性规则属于已有 TypeScript interaction/context owner；Python 保留 registered-source IO，共享校验消除履行与复用的规则分叉。本次边界是已开启 native 路径的完整 Goal 来源交付与 exact reuse，普通关闭态继续渐进读取。只修改 codec 无法复用根本没有送入 capture 的正文；另起缓存或第二套来源规则又会扩大维护面。现在沿现有 native activation、registered-source、typed fulfillment 和 capture/restore 链接入，无新配置或额外确认轮次。

路径	本次行为	资格边界
关闭 native	保留完整渐进读取	不自动启用
开启且来源有效	原文入同次 capture，正文可 exact reuse	当前 witness 和未知字段仍明文
来源无效或上下文失效	拒绝错误履行或恢复原 capture	不重跑准入、不新授执行权

具体改动 / Changes

规范依据为 docs/reference/required-work-context.md，spec_revision 2cdff65d886d13121ffd78dccd96dd7413ab9d01；先按该不可变版本的条文判断，再核本次双语扩展。complete_original_goal 已实现完整原文交付；original_capture_recovery 保持同捕获恢复；source_failure 保持失败拒绝；default_off 保持原关闭态。没有把未来 C1 或模型收益要求冒充本 PR 已完成。

关键代码讲解

• _source_content 复用原文件读取与严格区域 parser，只在 native 请求下补全部未标记原文；没有用标题或摘要判断哪些要求可删。
• completeGoalNarrative 是从既有 owner 移出的共享校验：完整原文必须匹配 UTF-8 digest，不能夹带排除区域或假装 partial read 完整。
• projectInteractionWorkContext 在 native 完整性请求下拒绝仍 pending 的 Goal；Python 还核内部 complete_goal_v1 ACK，避免旧 runtime 静默忽略请求。
• splitContext 复用同一 predicate。只有已核验正文可以继承，revision、full-source witness 和未知当前字段不借旧值。

全 diff 为 10 路径 +264/-69，包含现有 predicate 搬迁、6 个生产文件、3 个测试文件及双语协议说明。共享小模块有两个实际消费者；无需再抽额外 wrapper。

对主干的风险 / Risks

最强反例是“尾部停止条件被截断却仍被当完整来源复用”。独立合成 fixture 使用 CRLF、中文、emoji、24 条约束及末尾 stop；同一 oracle 在不可变 base 的 File/SQLite 均因 Goal 仍 pending 而失败，在 head 两项通过。每项实际读取 9 页，核对全部 1,856 原始字节；暖态完整重建、epoch 丢失恢复全文、错误 capture digest 拒绝，原文件不变。所有已加载产品模块均核到对应冻结源码。

关闭态另有 5 组相同完整 source-project 输入，base/head 输出逐字节相等，包括 required reads 和全部指令。PASS：292 Python、57 TypeScript、typecheck、standard premerge 的 6 个直接与 19 个选中检查，0 failure/hold。SKIP：68 个需显式原生 Codex binary 的既有 boundary/capture 场景；未作为通过，也未启动真实模型。相关 source、adapter、reader/lifecycle 的实际 File/SQLite 覆盖单列，未声称完整原生平台矩阵。未查询、轮询或等待 CI。

语义与 CI 对齐

这是既有来源投影的字段扩展，TS 继续持有完整性判断，Python 保留 IO。complete 表示来源交付条件，不是模型阅读证明；完整要求、先读后执行、fresh guard、停止与结算义务没有删减。冷态和丢失恢复可能增加传输/校验成本，作者披露的取舍不能外推为 token 或端到端效率收益。

我的整体评价 / Verdict

交付判断是 justified_increment：long_horizon 与 CLI 旅程在这个来源/恢复边界得到改善，剩余 C1 和真实采用继续留在原工作。没有修改渲染页面；检查的是 quota.py、context_readback.py 与 context_delivery.ts 的现有 CLI/native 来源路径。实际模型消费、已安装 App 和 Windows 原生生命周期未验证；未据此宣称体验或费用收益。保留 marked lineage 与关闭态兼容有实际消费者和持久 capture 依据；共享 predicate 的收敛合理，没有新增状态 authority。代码批准、原生 closeout/readiness、维护者合并和安装是分别判断的阶段。

English verdict: APPROVE d980d621219b10353fdaeb058a1a0fd5714f4389; complete native Goal delivery and exact recovery pass independent File/SQLite counterexamples; actual model and installed adoption remain unverified.