feat(replan): deliver core Goal and dense typed decision evidence

[huangruiteng]

Recent history slices can hide an earlier useful result behind repeated observations and newer peer activity. Project the full Goal objective separately from current task acceptance, and deliver Agent-scoped typed evidence that retains result/route diversity, folds duplicates, bounds display to 24 rows, and provides exact references plus omitted-history recovery.

Quota, PR review and handoff share the TypeScript context owner over the existing compact run index; Python adapts source IO. Retire the duplicate live evidence-log interface and mandatory diagnostic-read ritual, retaining historical readers. Agent-scoped history now filters the complete decoded source before the requested limit; bounded replan decision lookback and Goal-wide quota accounting remain separate. Preserve the current lazy runtime handler architecture.

Validation: 14 authority-derived identities across six Goal shapes, 221 synthetic domain observations each. Early results, 180 repeated observations, all distinct routes, exact-reference refusal, ordinary scoped history and the generated omitted read command are exercised through the actual CLI/index backend. Every scoped omitted read returns all 221 records without index writes or progress receipts. Focused context/history tests, extended regressions, TS/typecheck, real CLI base/head differential and all 19 latest-main premerge canaries pass.

Cost tradeoff: the same crowded quota fixture grows from 26844 chars/718 lines to 33523/806 (about 25% more characters); the documented crowded regression guard is 34000/830 and ordinary/non-replan guards are retained. Full source IO still grows with history, even though visible evidence rows are bounded. S11/T3 sustained model-effect and cost qualification remains open; this increment does not claim token savings, paid-model quality gains or new execution authority.

Final exact-head GitHub checks and published review determine merge readiness.

[mergify]

This pull request has merge conflicts with main and cannot be merged
until they are resolved. Please rebase or merge the base branch, @huangruiteng.

Choose the remote for the base repository, not an out-of-date fork.
For a fork clone, first inspect git remote -v; upstream must point
to https://github.com/loopx-project/loopx.git. If it is absent, add it
with git remote add upstream https://github.com/loopx-project/loopx.git.
Then run:

git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEAD

For a same-repository clone whose origin points to
https://github.com/loopx-project/loopx.git, use origin instead of
upstream for fetch/rebase. If you prefer merging the base, use
git merge <base-remote>/main and push normally.

Keep the DCO Signed-off-by trailer on every commit when you rebase.
https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[huangruiteng]

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

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

Exact head: a5e7a54a239b310198d09c7e750ebfe4dc269a4e. APPROVE 的结论只针对这一版本及以下明确的增量边界。

动机

跨多轮研究、工程或评审的 Agent，需要找回较早有效结果、重复尝试及尚未覆盖的方向。 一个 Agent 先得到有效结果，随后反复观察，同伴又产生较新记录；原先近期窗口藏住旧证据，现在上下文保留不同结果并折叠重复，还能沿生成的命令找回被省略的方向。

14 个身份、6 个 Goal 形状各有 221 条合成领域记录时，旧有效结果和重复次数保持，可见行数至多 24；生成的回查命令读齐本 Agent 的 221 条，拒绝跨 Agent 引用且不写进展回执。

本增量交付证据可用性和有界投影，不证明付费模型决策收益、多日运行收益或整个 TS 迁移验收，也不扩展任务、额度或外部账户权限。 现有 typed observation 没有领域级 hypothesis/反证完整性判断；完整模型收益与长期成本继续由既有 S11/T3 验收承担。

改动思路

由既有完整 compact run index 生成 Agent-scoped typed context，保留不同结果与尝试方向，合并重复观察，限量展示并提供精确引用和遗漏项回查。完整 Goal objective 与当前 task/acceptance 分开；局部任务不能冒充最终目标。quota、PR review 和 handoff 复用同一 TS owner，Python 保留 source/codec，旧重复 evidence-log writer 和强制读回仪式退休，历史 reader 保留。

复审修复了一处真实恢复缺口：生成的遗漏读取曾按 Goal 先截断，再过滤 Agent，还受 60 条 decision lookback 限制。现在完整 decoded source 先按 Goal/Agent scope，再应用调用者 limit；普通 history 与 replan decision window 分开，Goal-wide quota 计算仍用完整 Goal source。

具体改动

关键代码讲解

1. loopx/control_plane/work_items/replan_context_codec.py:15 的 replan_history_from_status 读取完整 compact index；source 缺失仍明确 snapshot 边界，不把有限切片当完整 authority。
2. loopx/control_plane/work_items/replan_context_codec.py:31 的 replan_evidence_rows 校验 Goal、Agent、时点与 typed observation，peer 不提供本 lane 的证据。
3. loopx/control_plane/work_items/replan_context.ts:74 的 projectReplanContext 按完整事实合并重复，保留结果/route 差异，至多 24 行；delivered 表示 host 提供上下文，不表示模型消费或验收完成。
4. loopx/history.py:324 的 collect_history 接收独立 scoped_agent_id，在 limit 之前过滤，runs/latest_runs/latest_status 一致；原 bounded decision lookback 与 Goal quota 输入保持独立。
5. loopx/control_plane/effect_runtime_handlers.ts 的两个 replan-context 方法继续 lazy load 同一现有 TS 模块，保留最新 main 的冷启动结构，没有恢复大批 static imports。

独立规格：docs/reference/protocols/goal-vision-replan-contract-v0.md，spec_revision e55489c7791f326abe63e7b9b2952a134d652cf7。Projection Contract 章节规定：Agent public-safe evidence、optional drill-down、不以读取 receipt 算进度，均有真实入口证据。Vision Continuation Audit 章节规定：当前权威证据优先，弱/旧/缺失证据不能当验收。S3/S6 是本次增量；S11/T3 模型 score gain、全调用者迁移及长期成本保留既有后续阶段。

对主干的风险

原反例：多 peer 下，一个 Agent 的 221 条记录被较新同伴记录挤掉；生成的成功回查只能给 0–180 条 top-level 或最多 60 条补充，多数旧方向缺失。修复后 14 个身份逐个沿 generated read_action 实际执行 CLI，全部读齐自身 221 条，普通 limit=10 正确，latest_status 不丢；所有 18 条 omitted 方向可回查。wrong-Agent exact ref 拒绝，读操作不改 index、不产进展或 read-gate receipt。

领域数据来自完整真实 authority 的 Goal/任务语境，结果观察是合成的：评审 head drift、lease/停止、产品重启、运营来源、决策未知背景、笔记索引、系统失败反例、工具冷暖启动、交接策略、发行人披露/反证、量纲与权益冲突、Bot 排队修正。14 个身份共 3094 条领域历史；旧 advanced 保留、180 次重复合并、42 个 distinct 仍在完整 novelty index，可见最多 24。上下文不会仅凭“盈利/已停止/已完成”文本授予领域结论或执行权。

修复入口聚焦 31 passed，扩展回归 179 passed，最终 premerge canary 19 项均通过；集合有重叠。TS 聚焦 44 passed、core typecheck、quota/review/handoff 实际入口与 CLI 差分通过。原完整差分为 102 baseline / 96 head、0 candidate-only；退休 6 项入口和 12 项 review-required intentional deltas 明确披露。两项 canonical periodic successor 原先在同一 immutable base 与原 head 具有相同 test id/guard 错误，已单独归因；不是 omitted-history 反例。当前 Goal 的原生评审/合并契约明确 wait_for_ci=false，已停止轮询；本次批准依据当前 head 的本地入口、反例、差分与 canary 验证。此前已观察到 DCO 和真实 PostgreSQL 检查成功，其余 GitHub CI 尚未全部完成，不宣称全绿；发布/部署和非要求版本路径的 skip 不当通过。用户已明确授权两 PR 自合并，仍须公开 exact-head review、无未解决 review thread 和原生 ready=true。

语义与 CI 对齐

默认上下文与 CLI 入口变化已披露，未宣称 opt-in/default-off。typed observation 决定结果分类，未通过 substring 或自然语言判断领域真假。evidence 提供、读取、进度与最终验收分层；旧 evidence-read 义务已退为可选诊断。snapshot 不等于全 source，未知 source 不当完成。manifest 漂移在修复后重生成并复跑，保留完整扫描根和真实分类；公开边界 clean。

成本明示：同样 crowded quota fixture 由 26844/718 增到 33523 chars/806 lines，约 25% 字符增长；相应 crowded regression guard 为 34000/830，普通/非 replan guard 保持原界。diagnose 43132/804 对应 44000/850。新增内容保留完整 Goal 与证据差异，重复观察被压缩且有可用 drill-down，增量预算不是无限豁免。每个领域 context 约 23.2–26.1k JSON 字符；本地 warm helper 约 12–15ms，仅是 helper 成本。完整索引 IO 随历史增长，未测付费模型 token/决策收益或端到端加速。

我的整体评价

有界且可回查的完整证据比单纯扩大最近窗口更适合长程工作。Future-facing pass 集中 TS 规则、退休重复 live entrypoint、保留真实持久化 reader，修复 scope-before-limit 并保持 lazy handler。预计效果正向，尤其证据稀疏或经常有 peer 并行活动的 lane；效率短期每次输入成本负向，长期减少重复探索的净收益尚待观察。不能把 context 到达等同模型采用。

English verdict: APPROVE — exact head a5e7a54 delivers diverse bounded evidence and a working Agent-scoped omitted-history recovery path, retains the current lazy runtime architecture, and preserves authority boundaries. Dense-context output cost is explicitly measured; sustained model-effect and multi-day cost claims remain outside this increment.