[loopx-agent]

Goal And Delivered Outcome

quota should-run can repeat the same complete cumulative Goal acceptance in work context, planning/replan context and the Todo summary. These copies add no evidence but increase every consumer's input.

This proposal keeps one complete inline carrier and replaces only exactly equal copies with same-response references. Revisions, scope, sources, failures and JSON types must match; different or unavailable content stays inline. TurnEnvelope consumes the CLI representation, relocates the known reference to its signed work-context root and rejects missing/unavailable carriers. No requirements, authority, decisions or acceptance verdicts are removed.

• Outcome basis: demonstrated duplicate acceptance transport; the existing complete-context contract in docs/reference/protocols/turn-envelope-v0.md at b285c7c1b38bb1d426fca2ef6bd8ddfce177959a.
• Intended base: loopx-project/loopx:main.

Author Declaration

• Written by: model_agent — OpenAI Codex, GPT-6 family.
• Implemented against: the complete-context, reference and action-signature contract above; this PR defines the bounded CLI sharing extension.

Criterion	Disposition	Boundary and evidence
Keep complete acceptance and provenance in the same response	implemented	CLI projection; exact-copy and different-revision/scope/failure/source/type tests
Preserve executable action/signature and reject dangling envelope references	implemented	TurnEnvelope consumer; raw/CLI parity, missing/stale source and tamper cases
Qualify the actual CLI on isolated supported local providers	implemented	File and SQLite selection/replan calls with canonical cumulative acceptance
Demonstrate long-horizon model/task benefit	out_of_scope	Bytes and CLI measurements establish transport cost only; downstream outcome/cost comparison remains separate

Self-check: reviewed all five changed paths, consumers and the complete retained contract; ran focused regressions, real provider paths, type checking and standard premerge. The premerge failure below remains a hold.

Scope And Continuation

The Python change stays in the existing CLI presentation adapter after typed owner decisions; it adds no Python decision owner or extra RPC. Canonical acceptance, saved full-decision captures, decision construction and provider reads retain their original structures. Diagnostic detail options also share exact copies. The adjacent envelope consumer is updated in the same reversible slice; there is no generic reference resolver or new persistence mechanism.

Author implementation is complete within this scope. Review, the existing canary failure and required merge authorization remain open; this PR does not install or merge itself. Scientific scoring, tasks and experiments are unchanged. Model adoption and total task cost are not qualified here.

Validation

• Tested revision: 90d378a81 (the committed content matches the tested working tree).
• Run state: finished.
• Input classes: synthetic, public_fixture; private read-only diagnostics remain local and are not used for the public measurements below.

Check kind	Result	Evidence / limitation
unit / integration / real_entrypoint / real_backend	passed	52 Python tests across test_quota_acceptance_references.py, test_quota_cli_projection.py, test_quota_planning_horizon.py, test_replan_host_context_projection.py; includes disposable real File/SQLite CLI selection and replan paths
regression_parity	passed	42 turn_envelope.test.ts tests; complete source retention, equal raw/CLI action signatures, unavailable/missing-source rejection and signature sensitivity
static	passed	npm run typecheck:control-plane, diff hygiene and diff-scoped semantic advisory; no new state vocabulary
regression_parity	passed	Same synthetic Goal with 8 outcomes × 8 requirements, base b285c7c1b vs this head, same Python/dependencies, one warmup then 8 alternating fresh CLI processes per provider. File bytes 110,489 → 70,081; SQLite 110,513 → 70,105 (about 36.6% less). CLI medians: File 381 → 371 ms; SQLite 373 → 367 ms. Short local measurements do not establish a latency gain.
regression_parity	passed	Separate 40-repeat projection-plus-JSON comparison: two synthetic packets shrink 9.0% and 12.2%, while medians rise from 0.077 to 0.085 ms and 0.074 to 0.086 ms. Sharing has a small comparison cost.
integration	failed	Standard loopx canary premerge --from-git-diff --git-diff-base origin/main: quota-plan-smoke.py fails in its CLI fixture with canonical_authority_required. The same command failure reproduces on unmodified b285c7c1b; the changed code is on should-run projection, not quota plan. This is an unresolved baseline hold, not a passing skip.
static / integration	passed	Other premerge results: all direct checks, 4 other catalog canaries, all 8 risk-profile smokes and public/private boundary scan
real_backend	not_applicable	PostgreSQL authority-store behavior is unchanged; this proposal only changes presentation and its envelope consumer

Reproduce focused Python checks with uv run --extra test python -m pytest tests/control_plane/test_quota_acceptance_references.py tests/control_plane/test_quota_cli_projection.py tests/control_plane/test_quota_planning_horizon.py tests/control_plane/test_replan_host_context_projection.py -q. Run the TypeScript file with Node's test runner and strip-types support.

Frontend / Visual Evidence

UI impact: none. This repairs an existing agent CLI/envelope data journey; no frontend capability, configuration or visual surface is added.

Type / Area / Boundaries

• Bug fix, protocol documentation and regression tests; control plane / host integration.
• Shared-authority migration fixture impact: N/A; no provider, authority or persisted-state migration.
• Public scan passed; no private state, raw trajectories/logs, credentials, internal links or local operating paths are included.
• No maintainer-owned benchmark work duplicated; no paired RFC changed.
• All commits carry DCO sign-off. Review and protected-surface merge authorization remain separate gates.

[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 — 审核精确提交 90d378a81f644f55d0c056f266f2ab9d04d78b34。未发现本 PR 的阻断缺陷；标准 premerge 中一项已有 fixture 失败仍如实保留，合并权限与最终 readiness 单独判断。

动机

使用普通 quota CLI 获取下一步任务的 Agent。累计要求较多时，旧响应把相同完整验收复制到多个位置；新响应保留一份完整来源，其余位置在同一响应内引用，避免反复传送相同内容。完整要求、来源和失败信息仍可在同一响应内还原；独立 File/SQLite 对照及实际 TurnEnvelope 消费通过，未证明模型收益或整体任务效率。本 PR 不改变验收结论、调度、权限、预算、任务评分或持久状态，也不降低原始完整决策与捕获内容。

改动思路

选择保留现有 CLI 展示层与 TypeScript 签名消费者两个边界：权威验收与决策不迁移，不新增缓存、RPC、持久状态或通用引用解析器。本 PR 边界是同响应完整验收副本的展示复用及其既有 envelope 消费者；模型采用、任务收益和上游读取成本属于后续实验。先由已有类型化 owner 计算决定，再在 CLI 的展示阶段比较完整 JSON；只有版本、作用域、来源、失败信息及 JSON 类型全部一致时才复用。来源缺失或不同就保留完整内容，既不推断验收已通过，也不重新查询一个较新的版本。相邻的 envelope 消费者只迁移已知路径到自己的签名根。

具体改动

规范依据为 docs/reference/protocols/turn-envelope-v0.md，独立审查其变更前版本 fde154fc8147bc54d687b9cc22d6315069682de7。条款 complete-work-context：已实现完整来源与末尾要求保留；same-observation-and-authority：已实现同响应读取，原始捕获及权威状态不改；signed-context-preservation：已实现 CLI 引用进入 envelope 后的完整上下文和签名保持。上述条款均已通过实际 File/SQLite CLI、独立完整要求校验和类型化消费者验证。新增协议文字说明展示形式，不把缩减字节当作降低原要求的授权。

关键代码讲解

• loopx/control_plane/quota/cli_projection.py:708 的 _reference_inline_goal_acceptance 优先保留已加载的 work-context 来源，再选择 planning horizon 等可用完整副本。比较采用严格 JSON 编码，避免 Python 把 False 与 0 混为相等；只复制受影响的字典路径，不改传入决定。
• loopx/control_plane/quota/cli_projection.py:864 在普通 should-run 展示中接入该步骤，详情选项也保持同响应复用。quota.py 的 plan/status 分派仍走原有投影，原始 decision capture 不经过这个展示修改。
• loopx/control_plane/quota/turn_envelope.ts:839 的 turnActionProjection 验证唯一支持的引用形式、来源种类、加载状态和非空完整内容，然后移到 envelope 的 work_context。缺失、无关或不可用来源明确拒绝；原始完整输入仍兼容。

对主干的风险

主要风险是引用指错位置或隐藏旧要求。52 项 Python 回归和 42 项 TypeScript 测试通过；我另外在基线 fde154fc8147bc54d687b9cc22d6315069682de7 与本 head 上运行相同独立 synthetic File/SQLite CLI，完整要求哈希直接一致，实际 envelope 消费、同回合重入、之后的权威读回均通过。故意删掉末尾恢复／发布停止条款时，独立消费者校验失败，未把签名相等当作完整语义证明。

标准 premerge 的 6 项直接检查、4/5 catalog canaries、8 项风险 smoke 和公开边界检查通过；唯一失败是 examples/control_plane/quota-plan-smoke.py。我在未修改基线与本 head 重跑，并捕获实际 CLI 的完整规范化失败：6 个旧 fixture Goal 都缺 canonical authority，状态、迁移建议、退出码与全部细节相同。失败路径是 quota plan 的旧 fixture，未进入本次 should-run 改动；这属于 pre_existing_unrelated，保留其红色记录，不称整套 premerge 全绿。独立 canonical File/SQLite 覆盖了本 PR 的实际不变量。TypeScript 类型检查、差异检查及 RFC 双语检查均通过。按本 Goal 策略未查询或等待远端 CI。

语义与 CI 对齐

此次复用现有引用词汇和完整上下文契约，无新状态分类或执行权限。普通 CLI 的默认展示变化已经双语披露；TurnEnvelope 的启用条件不变，关闭它仍可读取完整普通上下文。没有 acceptance 或不相同的副本保持原路径。已有 fixture 失败的恢复责任仍在测试 fixture，不能通过削弱 canonical authority 或隐藏失败解决。

我的整体评价

该切片的展示复用目标已达到，整体规模与每回合重复传输的问题相称。长程要求保留判断为 preserved；CLI 使用体验为 improved，因为同一命令即可获得同一份完整要求，没有新增配置、用户确认或额外读取回合。现有 owner 复用和限定引用路径足够，无需新增通用框架。独立小 fixture 的 File 输出 68,809→65,651 字节、SQLite 68,837→65,681 字节只证明这一条件下的传输减少，不外推模型收益、任务得分或整体效率。保留原始捕获与持久状态使此变更可回滚；已知 baseline fixture 失败、最终合并 readiness 和实际后续实验采用继续分别记录。

English verdict: APPROVE at 90d378a. Complete acceptance, authority and signed consumers are preserved through real File/SQLite CLI/reentry checks, 52 Python and 42 TypeScript tests. One identical pre-existing quota-plan fixture failure remains disclosed; model utility and merge readiness are separate qualifications.