[loopx-agent]

Repeated history-read commands inflate replan context in ordinary quota responses: every visible evidence row repeats the same registry, runtime, Goal and Agent route. Multi-row scoped contexts now carry one evidence_read_action_template; each row retains its content-bound reference and complete evidence. The Agent substitutes that reference to execute the existing history command, with no additional lookup.

This extends the existing TypeScript replan-context projection consumed by quota and Turn callers. It preserves evidence ordering, coverage, source selection, required reads and authority. Single-row contexts retain their executable command; historical per-row contexts remain valid, and routes containing the placeholder literally retain per-row commands. The shared route participates in context identity. This changes the default multi-row presentation; it introduces no cross-turn baseline or incremental-context protocol.

Validation for 689b12c2dbb61950f66b715ed65322a568fb5a99 (runtime unchanged from 16d521de49bc8867d84599fcb41a731c38dc3a10), based on main 035a077767d1e9fb9b321e35181ff9ffd81f00ef:

• 89 focused TypeScript tests passed, including exact command expansion, historical contexts, route changes and malformed templates; control-plane TypeScript typecheck passed.
• 43 Python tests passed after rebasing. Real File and SQLite paths exercise ordinary quota should-run followed by the emitted history route, including unavailable evidence, wrong-Agent rejection and read-only behavior.
• Complete CLI output-budget file: 38 passed, one existing native-temp-path depth test skipped because this host cannot fit the qualified layout. The stale per-row-command assertion was updated; byte/shape limits were not relaxed.
• Standard premerge on the runtime commit: direct checks and 16 of 17 selected checks passed. The semantic-vocabulary check reports goal_boundary in 32 Python modules against a budget of 30; the same failure reproduces on the unmodified target main. This remains a validation hold, not a pass or merge authorization.
• Public-boundary scan is clean for all six changed files. The broader local registry check separately reports two existing unmigrated Goals; no runtime state was changed to bypass those errors.

An offline, lossless projection comparison over 31 saved solver decisions changes five packets, saving 24,215 serialized bytes across the set (about 1.0%; up to 6,885 bytes in one packet). This is a bounded reduction in repeated routing text, not measured model-token, task-outcome or wall-time improvement. Integrated live evaluation remains separate.

Placement and scope: an optional presentation field in the existing evidence-context contract; no new semantic vocabulary, capability, provider, state writer, frontend operation or Python decision owner. The adjacent reversible cleanup is route sharing itself. Required context and broader quota/Turn convergence remain unchanged.

[loopx-agent]

Reviewer: model_agent; gpt-6-astra; OpenAI; runtime_reported; reasoning_effort=xhigh

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

动机

普通 quota 和 handoff 的重规划调用者会收到多条历史证据，每条重复相同的补读命令。
此前每行重复 Goal、Agent 和路径参数；现在多行共用一次命令模板，选择行后仍能直接补读该条完整证据。
已验证展开后的命令、证据内容、顺序、覆盖和错误语义一致；确定性收益限于减少重复路由文本。
不改变证据选择、强制阅读、评测机会或写权限，不引入跨轮缓存，也不把字节减少视为模型效果提升。
模型对模板的采用和完整输入、补读、时间及任务收益仍需集成实验。

改动思路

共同路由可由已有 Goal、Agent 和命令前缀直接推导，放在现有 TypeScript 投影中即可，无需第二个状态或决策所有者。
本 PR 只合并同一响应内的重复补读路由；跨轮子树缓存与模型效果验证留在各自已有工作中。
入口仍是普通 quota 或完整 handoff，Python 仅传输已有历史，TypeScript 选择和投影证据，history CLI 解析具体引用。相比直接删掉命令，这个方案保留了按需访问；相比跨轮差分，它不依赖上轮缓存。没有新增操作、权限确认或强制阅读步骤。代价是调用者要替换一次明确的占位符，因此不能凭字节节省宣称模型一定更高效。

具体改动

关键代码讲解

1. projectReplanContext（loopx/control_plane/work_items/replan_context.ts:514）：先按原规则选择和生成完整证据，再在有 Agent 且实际展示多行时共用命令路由；模板参与 context identity。空结果、单行、无 Agent 以及前缀含字面占位符的情况保留既有形式，遗漏证据的补读入口没有变化。
2. validateReplanContext（同文件425行）：历史逐行格式继续可读；新字段出现时验证非空、恰好一个占位符和完整末尾引用参数。该校验没有授予执行权，也不替代真正 history 的 Goal/Agent 范围校验。
3. test_quota_shared_evidence_command_reads_exact_history_and_preserves_scope 经真实 File/SQLite quota 和 history 路径验证回读，原有长历史测试改为展开模板。独立复核另外用同一31条合成历史比较主干与本提交，逐条检查展示的24条补读、完整尾部和只读性，而非只检查字段存在。

规范依据：docs/reference/protocols/agent-scoped-evidence-ledger-v0.md，修订 035a077767d1e9fb9b321e35181ff9ffd81f00ef，按变更前的 Acceptance 判断。本次布局调整依据已接受的无损精简方向，未以修改后的文档自证语义。
验收映射：可读证据、确定性选择、精确补读、显式错误、阅读不记进展、合成验证 均通过当前投影、真实回读、反例及未改语义写入边界核查；模型效果边界明确保留，确定性测试不等于长程效果提升。六个文件为16行生产增量、现有测试扩展和协议说明，没有新服务或状态存储。

对主干的风险

主要风险是模板展开后错误地丢失范围、选择另一个记录，或旧上下文不能再读取。独立基线/首提交16d521de各完成 File、SQLite 两个真实路径，每个路径24次精确补读；错误 Agent、删去 Agent 参数以及删除合成索引后的旧引用均显式失败。完整证据和完整错误负载在限定根路径及派生context identity归一化后保持一致。27项独立 TypeScript 测试通过，涵盖旧格式、非法模板、字面占位符、单行和无 Agent；最终head只追加既有CLI预算测试的模板展开断言，生产文件及上述测试的源保持不变，原始回读证据明确沿用16d521de而不冒充最终head重跑。独立运行最终head整份CLI预算文件得到38通过、1项native temp无法容纳depth-12布局的既有跳过；该布局条件在基线同现且未被本补丁修改。另保留作者当前提交89项TS、43项Python和typecheck证据，不把它们标成我的独立运行。

语义与 CI 对齐

不查询或等待远端CI。作者标准premerge直接检查通过，所选检查16/17通过；唯一失败是 goal_boundary Python读取模块32超过预算30。我在固定主干和精确head用相同命令再次得到完全相同的失败输出；相关Python读取模块与扫描器不在此补丁中。它属于既有无关失败，保留记录及独立合并资格限制，不提高预算或称全绿。没有新枚举；沿用已有证据契约。未改PG存储/权限路径，当前真实验证覆盖File/SQLite；模型采用与总成本仍未测量。

我的整体评价

精确提交 689b12c2dbb61950f66b715ed65322a568fb5a99 的这项增量可以批准：它在已有所有者中可逆地减少重复路由，完整证据、可选阅读和后续恢复路径保留。长程继续工作的契约与用户操作体验判断为 preserved，不能据此判断新实验的分数会更好。共同路由可由已有 Goal、Agent 和命令前缀直接推导，放在现有 TypeScript 投影中即可，无需第二个状态或决策所有者。本 PR 只合并同一响应内的重复补读路由；跨轮子树缓存与模型效果验证留在各自已有工作中。
本 PR 修改 quota JSON 与协议说明；handoff 文本仍读取同一 summary，没有渲染界面变化。模型实际选择并展开模板的采用情况尚未验证。
这里是评审批准，不是main合并授权、全套检查全绿或集成实验已启动。独立实验仍需最终源版本、接收端和资源准入。

English verdict: APPROVE — 689b12c2dbb61950f66b715ed65322a568fb5a99. Independent base/head File and SQLite quota-to-history comparisons preserve complete evidence, exact scoped reads and failure behavior after documented route expansion. 27 focused TS tests passed. The identical pre-existing 32-versus-30 vocabulary ratchet remains a separate validation/merge hold. Live model adoption, total cost and benchmark gains are unverified.