test(explore): align replan qualification with inline context

[mikamikasuki]

Goal And Delivered Outcome

• Outcome basis / optional anchor: Self-contained maintenance outcome; no separate issue is required for this scope.
• Goal/source and gap: The replan semantic-action qualification treated Explore turn context as a standalone top-level required_reads command. Quota packets deliver the turn-start hook result inline through interaction_contract.agent_channel.work_context.sources; the stale helper failed before the modeled action and its skip/replay/scope cases no longer matched the packet contract. The Explore README described the previous delivery shape.
• Observable before → after: At the PR's original base 5f51559dc93b43d023a9c915c52ec1dc409e3e65, the replan qualification file had 4 failures and 19 passes. The actor fixture now verifies the scoped, read-only Explore context in the quota packet before the next action and does not perform a duplicate Explore read. The README describes the inline contract.
• Issue/task and intended base: Direct maintenance PR against main; current base is 73dfd1f66dadbd7c08b25a94813c4d51ec99d1fe.

Author Declaration

• Written by: model_agent — OpenAI GPT-6 Luna

Some coding work was AI-assisted. I personally reviewed the changes, verified the relevant behavior and tests, and take responsibility for the submission.

Implemented against

• Specification and revision: The quota context-delivery contract in loopx/control_plane/work_items/context_readback.py and tests/capabilities/test_explore_turn_context.py::test_real_quota_delivers_replayable_context_without_admitting_unhealthy_goal.
• Criteria:

Criterion	Disposition	Symbol / path	Test
The admitted Explore source is available in the quota packet before the next agent action.	implemented	interaction_contract.agent_channel.work_context.sources	tests/control_plane/test_replan_semantic_action_behavior.py
An inline Explore source is not duplicated as a separate required read.	implemented	work_context.sources and required_reads	Replan actor and Explore turn-context tests
Explicit Explore required reads, when present, use the nested interaction contract.	implemented	_required_explore_read_command	test_required_explore_read_uses_nested_interaction_contract

• Self-check: Reproduced the stale fixture failure at the original base, traced the packet projection and contract-test history, searched for matching issues/PRs, and reviewed the final three-file diff. No production runtime behavior changes.

Scope And Continuation

• Completed scope: Align the replan actor qualification and Explore README with the inline context contract.
• Slice boundary: Complete within this scope; no successor is needed.

Validation

• Tested revision: PR head caedbafcba3621c12f7bfe0a2ba218838803b3b0, synced with canonical main 73dfd1f66dadbd7c08b25a94813c4d51ec99d1fe.
• Run state: Local validation finished. Current-head hosted CI is running; the full repository suite did not complete.
• Input classes: synthetic.

Check kind	Result	Evidence / limitation
Unit	passed	Five focused suites covering replan semantics, Explore turn context, selection reentry, periodic lookback, and effective-turn replanning: 82 passed in 98.06s.
Static	passed	Ruff on changed Python files, Python compilation, worktree diff check, and base-to-head diff check.
Pre-merge	passed	loopx canary premerge --from-git-diff: 19 selected checks passed, zero failures, zero manual holds.
Hosted CI	in progress	New head run 37832856777 and associated checks started at 2026-10-08 19:33 UTC. GitHub Summary succeeded; DCO, dependency review, Frontstage Pages, PostgreSQL integration, and Release Artifacts were queued at last readback; merge-gate is expected.
Full suite	incomplete	An earlier python -m pytest -q -n 2 run was interrupted around 25% after failures appeared in untouched tests/architecture/test_semantic_* tests. No complete full-suite result is claimed.

• CI attribution: The earlier Frontstage Pages failure on head a342ba63cebbc908b1766c5771a2cdfb999d7260 was in the unchanged bilingual-link smoke and reproduced on its exact base. The semantic link-check correction is tracked separately in #5968; it is outside this PR's diff.
• Coverage and gaps: Scripted transport and synthetic Goal fixtures verify the quota packet shape and bounded tool sequence; the public Explore quota test independently verifies inline source and replay command. No live provider or external backend was used.

Frontend / Visual Evidence

• UI impact: none
• Before/after: N/A
• Source data: synthetic

Type of Change

• Bug fix
• New feature
• Breaking change
• Refactoring (no functional changes)
• Documentation update
• Test update

LoopX Area

• Control plane (goals, todos, gates, claims, evidence, quota, runtime)
• Benchmark boundary (adapters, runners, verifiers, scoring, evidence)
• Capability or extension (providers, adapters, skills)
• Public docs or presentation surface (README, protocols, dashboard)
• Build, packaging, installer, or CI
• Host or runtime integration

Technical Direction

N/A; this change aligns a qualification fixture and capability documentation with the existing packet contract.

Shared-authority RFC fixture impact

• Production-scale fixture schema: N/A.
• Semantic dimensions changed: None; production schemas and authority behavior are unchanged.
• Provider conformance arms and three-backend rehearsal: N/A.

Boundary Checklist

• No private state, credentials, raw traces, internal links, or local machine paths in the diff or PR text.
• No duplicate maintainer-owned benchmark work.
• Scope is limited to the current Explore context contract.
• Visual evidence completed or marked UI impact none.
• Every commit includes a DCO Signed-off-by trailer.

[mikamikasuki]

Follow-up for the replan-recovery fixture: its prior setup inserted surface-only run-history rows without matching quota admission receipts or settlement writebacks, so it could not legitimately trigger the documented settled-effective-Turn periodic review. On canonical main 9d7680a, the original focused case failed with long_todo_chain as the only trigger; the same assertion failed on the pre-fix PR head. The test now uses that valid replan trigger while keeping periodic cadence covered by the dedicated settlement/effective-turn tests. Validation after the fix: 27 passed across test_selection_replan_reentry.py, test_autonomous_replan_periodic_lookback.py, and test_effective_turn_replan.py on both PR head and current main; Ruff and git diff --check passed. Updated head: a342ba63cebbc908b1766c5771a2cdfb999d7260.

[mikamikasuki]

CI update for head a342ba6: Frontstage Pages build run 37811578012 fails in the unchanged examples/blog-bilingual-index-smoke.mjs at edgebench-feedback-and-memory (line 102). This failure reproduces on the exact PR base 5f51559; #5962 does not touch that script or the article. It is an exact-href false negative: the existing relative link resolves to the same paired-edition URL. The narrowly scoped correction is in PR #5968. I have left #5962 unchanged for this unrelated failure; the updated Python checks are still running.

[mikamikasuki]

The failing typescript-core (3/3) check in run 37832857599 is in the unchanged tests/control_plane_ts/host_process.test.ts, outside this PR's diff. CI observed the 4-second renewal fixture hit the proved deadline before the test's expected renewal rejection. I opened the focused test-only correction in #5989; the old fixture fails under a controlled 3-second renewal delay, while the revised fixture passes the same delayed case and its Node 22.22.3 focused suite. #5962 still needs a fresh successful hosted run after that fix lands; I am not counting the current failed check as passed.

[mikamikasuki]

Validation update for head 86f3a79, rebased on canonical main b27c4c7.

The remaining shard failure is part of this PR's fixture scope: the exact [repeat] case failed in three clean runs on current main because the fixture expected a top-level required_reads entry while the quota packet carries Explore context inline under interaction_contract.agent_channel.work_context.sources. The updated qualification checks that inline context and rejects a duplicate read.

Validation on the rebased head: both affected modules passed (31 tests); Ruff, Python compilation, and git diff --check origin/main...HEAD passed. Both PR commits retain DCO sign-off.

Hosted checks have restarted for this head; the browser currently shows Summary passing, six jobs queued, and merge-gate expected. Code-owner review is still pending. The full repository suite has not completed.

[loopx-agent]

Reviewer: model_agent — gpt-6.1-sol / OpenAI; runtime_reported; reasoning_effort=xhigh.

动机

APPROVE，当前完整 head 没有发现阻塞问题。

维护重规划资格测试的开发者，在 quota 已把 Explore 上下文随结果返回时，会被旧测试误导为还缺少一次工具读取。
旧 fixture 从已移除的顶层 required_reads 取 Explore 命令，组合路径在真正行动前失败；新版先核验包内 scoped、read-only 上下文，再沿原 frontier 选择与回写。
新 head 的真实 CLI 资格、作用域与重放验证通过；文档与现有单一内容载体一致，恢复选择继续按原 Turn 结算一次。
这次交付是既有资格测试和文档维护，不改变生产准入、配额、重规划或 provider 权限；没有证明真实模型采纳、远端服务或全库测试通过。

精确 head：86f3a792e46108f43ac7bcbfe0cece06d1afd74c；不可变 base：b27c4c7290b9fea5deac1afa54694f3d5c263b93。作者正文的旧 head、旧 base、三文件和测试数没有作为当前事实采用；本轮核验两个提交、全部四文件和实际 CLI。

改动思路

复用现有 TypeScript interaction owner 的来源投影和结算规则；Python 仅修正资格工具的读取位置，避免重新实现生产决策或增加第二份上下文。
当前 PR 完成真实 CLI 资格与现有内联契约对齐，保留未完成必读项、作用域、重放和周期结算约束；不新增产品操作或通用协议。

quota 先通过原 hook/source adapter 取得上下文，TypeScript owner 把已提供来源放在 work_context、未完成读取保留在嵌套 required_reads。测试消费当前包，再执行原 frontier/source 和 typed successor/writeback，实际准入、来源失败、重规划、身份及配额结算仍由原 Core 负责。把顶层旧命令重新塞回生产包或删掉 pending 检查都会制造第二契约，因此本 PR 只调整现有资格 seam。

具体改动

关键代码讲解

• _required_explore_read_command（loopx/control_plane/testing/replan_semantic_action_behavior.py:537）：先读嵌套 agent_channel.required_reads，再筛选 Explore；嵌套列表不存在才保留原测试 helper fallback。具体 source、before_work、Goal、Agent 仍逐项校验。空的当前嵌套列表优先于旧顶层指针。
• _inline_explore_context_action（tests/control_plane/test_replan_semantic_action_behavior.py:92）：检查真实 quota 已附带的唯一来源、准确身份、read_only 与 replay command，拒绝重复 pending Explore，然后回到原 frontier。它是 scripted fixture 观察，不能证明真实模型已理解上下文。
• selection recovery 测试（tests/control_plane/test_selection_replan_reentry.py:18）：移除没有准入/结算回执的人工 surface-only 行，用真实 long_todo_chain 验证恢复；原 Turn 身份、fresh admission、reentry 与只扣一次的断言继续保留。

规范依据：docs/reference/required-work-context.md + docs/reference/protocols/goal-vision-replan-contract-v0.md，固定版本 b27c4c7290b9fea5deac1afa54694f3d5c263b93。Required work context 为 implemented：已提供来源不重复读取，未完成项仍精确且有顺序，来源失败须恢复重准入，内容不给权限。Default review cadence 为 implemented：周期按 verified settled effective Turns 计数；selection 的 long-chain 触发与专门周期证明分开。修改的 Explore README 对照这些既有规范，不用自己的新文字证明自己正确。

对主干的风险

五个 source 模块独立执行 head 82 passed，base 5 failed / 79 passed。base 四项失败来自旧顶层读取/旧 skip-repeat-scope 假设，另一个失败是未结算历史不能触发 periodic_review_due。新版删除的 obsolete 用例没有被当作“相同测试自然变绿”：相同独立 8 项 nested oracle 在 head 8 passed、base 7 failed / 1 passed，验证合法 mixed pending list、重复项、错误 Goal、错误 Agent、错误 source、错误 order、legacy-only 兼容与 authoritative empty 优先。

相关 source suites 经真实 loopx CLI、生产 TypeScript owner 和隔离文件状态运行；Explore replay 读取与内联 content 一致，unhealthy Goal 仍 should_run=false/delivery_allowed=false。组合路径创建 scoped successor 并重新准入，选择恢复按原 Turn 结算一次；专门 effective-turn/periodic suites 仍拒绝不合格计数。模型 transport 是 scripted，不是一次真实付费模型验收。

Ruff、diff whitespace、两个提交 DCO、变更 advisory 及两端完整 semantic smoke 通过。风险式 premerge 的 5 项直接检查与 19 项选择检查全部通过，零失败、零 warning、零 manual holds。没有新增共享词汇、权限或默认开关；未配置 Explore 时原普通路径保持。README 改的是载体描述，读取当前 sources before work、freshness、scope、bounded detail 和建议非强制的条款保留，没有把义务缩成可忽略提示。未查询或等待 CI；全库 suite、live model/provider、安装后的应用未测。

我的整体评价

交付判断 goal_achieved 仅指这次资格与文档维护：开发者不再遇到旧载体造成的假失败，user_experience 为 improved；长期执行所需的来源失败、重放、scope、真实结算与周期 owner 保持，long_horizon 为 preserved。没有把局部维护称作 Explore 全能力或父级 Goal 完成。

复用现有 TypeScript interaction owner 的来源投影和结算规则；Python 仅修正资格工具的读取位置，避免重新实现生产决策或增加第二份上下文。
当前 PR 完成真实 CLI 资格与现有内联契约对齐，保留未完成必读项、作用域、重放和周期结算约束；不新增产品操作或通用协议。

Future-facing pass 保留生产 TS 单一 owner，Python 仅适配已有测试 seam；这个有界修复无需新增状态机或扩大重构。嵌套 empty 与 legacy fallback 的取舍已通过反例核验。Approval 不代表合并授权、全部旧架构 producer 安全或真实模型采纳。

English verdict: APPROVE — exact head 86f3a79. The qualification and documentation now follow the existing inline work-context contract without duplicating fulfilled reads or fabricating settled-turn cadence. Eight identical nested-read counterexamples and 82 actual CLI source cases pass at head; production context, admission and settlement owners remain unchanged. Live model adoption and the complete repository suite are unverified. No remote CI was consulted; merge readiness and authority remain separate.