[cocolord]

An executing Turn currently imports PR reconciliation and its metadata/content providers while composing the hook, before observation permission is resolved. An unavailable optional provider can therefore fail a path that has no permission to use it.

Move the reconciler import into the existing producer provider try, after the invocation-time permission recheck. Both permission checks, revocation, provider ownership, cadence, result/error shape and unknown-read handling stay with their existing owners. The two internal test injection sites now patch the real reconciler owner; all their assertions and response-loss behavior remain intact.

This is a bounded default-closure increment in the small-kernel RFC. The paired checkpoint records the delta and its limits.

Validation:

• Six fresh-process import checks fail on the baseline and pass on the candidate; five permission/import-failure scenarios also fail/pass through separate installed Core environments. Four real CLI entry checks pass without importing the reconciler.
• Eight matched hook scenarios compare complete deterministic returns, permission checks and provider calls with a frozen clock. They cover denied/revoked/unavailable permissions, empty/observed/partial results and response loss.
• The original 74-case File/SQLite command returns 70 passes and the same four setup failures on both arms: the unchanged consumer fixture patches the removed turn.build_turn_envelope alias. These failures are retained.
• A separate current-producer/carrier fixture passes four cases per arm. Its 28 Python assertions and seven protocol-host assertions remain unchanged; Goal/Agent/Todo, complete requirements, original reader linkage, artifact/readback and once-only replay were separately checked. This qualifies the selected-work/effect/readback fixture.
• The current C1 source combinations each pass their committed 74-case PR lifecycle command. C1 already carries its journal/source consumer fixture update; this result is separate from the main baseline's retained four setup failures. A clean installed current-C1 candidate Core also passes authorized File/SQLite provider observation and Todo CLI readback. This qualifies the changed hook path, not the full C1 composition.
• Author quick validation passes six direct checks and four selected checks after local TypeScript dependencies are installed; the initial missing-dependency failure is retained.
• Two independent standalone Core wheels pass 15 execution/replay/recovery/refusal cases each and real File/SQLite authorized observation plus canonical Todo CLI readback. Source checks cover provider failures, lost responses, cadence, concurrency, revocation and full-source diagnostics. Focused hook coverage is 97%.

Observation window: the first granted invocation now includes deferred imports in the existing duration_ms window. The frozen-clock comparison proves deterministic semantics; real timing equivalence and performance gains are unqualified.

The current C1 head is #8266 at 1a3740cf9e6e. Its issue-fix provider modules do not overlap this change; the shared dispatcher remains its existing owner. This proposal does not certify the entire C1 composition, full D2/D5 closure, actual-model/context adoption, complete-cost improvement or the original ten acceptance requirements. Formal exact-head review, required validation and merge remain maintainer gates.

[cocolord]

Reviewer: model_agent; GPT-6; OpenAI; self_reported

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

动机 / Motivation

现有 native Turn 和 heartbeat 在组合 PR 等待观察时会经过这条路径，包括没有外部观察权限的任务。

旧路径在读取权限前导入 PR provider，provider 缺失会中断本来无须使用它的任务；新路径在两次权限检查成功后才加载 provider，并保留原失败结果。

已验证的局部结果是未授权、撤销和权限源不可用时不加载 PR provider；获权后的成功、未知效果和独立 Todo 回读保留现有语义。

本 PR 不完成完整 D2/D5 闭包、真实模型或上下文采用、完整成本改善、原十项验收、C1 默认切换或 split-package 发布。

完整 R1、实际模型恢复与上下文连续性、完整成本和原目标回归仍待原有 owner 验证。

这个问题影响普通执行入口的可用性：用户没有选择 PR 观察，仍可能因为该 provider 的导入失败而无法继续自己的任务。此次改动将可选代码的加载放回实际使用边界，无须用户增加配置、重复表达目标或确认。所确认的价值是隔离这条无关依赖；真实冷加载收益和长期任务收益仍需独立测量。

改动思路 / Approach

现有 scoped capability resolver 持有权限，reconciler 持有 GitHub 观察，typed dispatcher 持有结果校验；把导入放在既有获权 provider try 内即可隔离可选代码，无需新增 owner 或 wrapper。

本次边界仅是 executing Turn 的可选 PR provider 加载及故障隔离；权限、cadence、原完整来源、capture、admission 和 settlement 都保留既有 owner。

既有 extend_turn_start_dispatch 会收集 capability 的观察注册，typed dispatcher 校验注册和结果。turn_start_hook 先读取当前 Goal/Agent 的权限，producer 调用时再读一次，避免注册后的撤销失效。只有第二次检查成功，才导入原 reconciler，并继续原 PR facts、cadence 和完整 canonical Todo 来源流程。注册校验、必要的工作义务及结果结算继续由现有控制面持有。

不改动会保留每次冷组合对可选代码的依赖。只在注册后导入仍跨过调用时撤销边界；增加缓存、skip flag 或第二个 dispatcher 会扩大状态和维护面。移动现有 import 到现有 provider try 内是足以修复这条路径的最小改动。原 Python 模块继续承担专门的 GitHub provider/transport，未新增 Python 决策 owner。

下表是固定 head 76502e5d18b46958a70a8cfed2218ddf3c0e2f3d 的源码分支说明，不能替代执行证据。

场景	provider 加载/调用	既有结果
注册时无权限	不加载、不调用	无 registration
注册后撤销	不加载、不调用	not_applicable
权限来源不可用	不加载、不调用	capability_source_unavailable，已知零读取/效果
调用获权、provider 成功	延迟加载并调用原 owner	empty / observed / partial
调用获权、导入或响应失败	进入原 provider 失败边界	failed，未知读取/状态效果保留 null

具体改动 / Changes

评审基准为修改前已接受的 small-kernel RFC，spec_revision 为 2b156284c4f44995797251dcc80b659ab1d673ac，重点对照其第 5、9、11 节。此次增加的 checkpoint 是交付记录，不是重新定义验收。最终 head 为 76502e5d18b46958a70a8cfed2218ddf3c0e2f3d：5 个路径，+136/-10；生产改动只有 import 的位置变化，另外两处已有测试将 fault injection 从 hook 内部 alias 迁到真正的 reconciler owner，原断言保持。两份 RFC 同时说明局部增量与计时边界。

关键代码讲解

1. _availability，第 42 行：先成功读取 scoped capability state，再显式交给原 resolver。Todo requirement、模块存在或 inherited hint 不能自行替代 provider IO 的权限来源。这条逻辑没有变化，负例验证权限来源不可用时没有读取或写入。
2. turn_start_hook，第 56 行：注册检查和 producer 内的 invocation 检查都保留。第 94 行才导入 provider，随后调用原 reconciler。导入失败进入原 except Exception，保留 dependency_observation_failed 以及未知效果；typed dispatcher 能识别该结果并维持原刷新语义。
3. _bounded_readback，第 21 行：继续在完整来源规划和效果完成后裁剪诊断展示。它不决定 PR target、工作选择或 cadence。已有大来源、响应丢失和 concurrent caller 用例检查这条边界，不能用 bounded view 的缺项推断 canonical source 不存在。

RFC criterion_id	本 head 的结论
I1	implemented：单一仓库、PR 和 CI 提案边界。
I2	implemented：权限、PR facts、typed result、canonical Todo/effect 继续复用原 owner。
I3	implemented：公开 hook contract 保留；非公开测试 alias 的调用方迁到原 owner。
I4	implemented：未授权或不可用的可选 provider 与核心执行隔离，无替代实现。
I5	out_of_scope：不改 Node 安装、缺失诊断或其资格。
I6	implemented：无新 schema、license 或公开/私有边界变化。
I7	implemented：未选 provider 不在权限检查前加载；无不受治理的执行 fallback。
I8	implemented：本局部路径保留原完整来源、要求、诊断与 replay；未据此完成整体 context 验收。
I9	implemented：记录准确 composition、冻结时钟和真实计时未资格化的范围。
V1	deferred：完整 presets 的 import/process/network 闭包仍待验证。
V2	deferred：完整 Minimal/Standard transition corpus 仍待验证。
V3	deferred：所有 supported commands、平台和 missing provider/runtime 仍待验证。
V4	deferred：真实模型任务的中断、重启和独立成果回读仍待验证。
V5	deferred：完整 initial/revised/compacted/lost-ACK host 场景仍待验证。
V10	deferred：重复配对的完整成本仍待验证。
R1	deferred：本 PR 的可选加载切片不完成 executable minimal 的所有退出项。
M3	deferred：split-package 发布与独立 clean-install journey 未交付。

对主干的风险 / Risks

最强反例是 warm 或预先导入的测试可以通过，冷进程仍可能在权限检查之前失败。此次 fresh-process negatives 在不可变 BASE 上保留 eager-import 失败，在最终 head 上通过；Root 也分别执行权限/CLI 9 个场景和 facade 1 个场景。八组 hook JSON 的完整内容相等，但脚本预先导入 provider，并冻结 monotonic_ns，所以它只证明确定性返回、权限检查与 provider 调用一致。首次获权调用的延迟导入现已纳入 duration_ms；真实计时等价与性能收益未验证。

主干原 74-case 命令在不可变 BASE 与候选均为 70 PASS / 4 FAIL。四个完整 case id 与 AttributeError: turn.build_turn_envelope 消息一致，失败发生在未改的 helper setup，尚未进入 Host adoption。该结果保留，不能写成 74 PASS。独立当前 carrier fixture 在每侧通过 4 个场景，28 个 Python assert 和 7 条 scripted-host assert 保持；另核对八份完整 saved source、journal、原 required marker、精确 PR provenance、artifact 和 replay。task_context 是投影，省略 CLI/control 字段，不能宣称它等于完整 raw work_contract 或证明真实模型采用。

两侧 standalone Core wheel 的 1,902 个产品文件与固定 Git、源码快照、wheel、installed bytes 和 RECORD 一致，ordinary/editable distribution 缺席。Root 再经原 provider 脚本仅追加 observation retention，独立执行 File/SQLite 各两侧：留下完整 hook、Todo CLI、事件和隔离状态，确认原 title/text/claim、精确 repository PR 和时间、一次 PR merge，以及零业务 quota spend/Todo complete。每个私有 fixture 内执行原 runtime owner shutdown；这不证明自然 process drain、所有 descendants 退出或完整进程成本。初次 Root replay 传入不含 tests 的打包源码目录而失败，改用不变的测试 fixture owner 后通过，旧失败保留。

最新 C1 为 #8266 的 1a3740cf9e6ebad337f17967452693a9050edc81。两份精确 source composition 的原 74-case 命令均通过，其 consumer fixture 的 22 行变化属于 C1 自身：改读 journal 保存的完整来源，并使用原 host-task-context owner。不能把这个通过结果归为本 PR 修复主干 setup。当前组合 candidate 的独立 Core 再通过 File/SQLite provider→事实→原 Todo CLI 回读；1,903 个产品文件与当前 composition/RECORD 一致。这里只资格化受影响的组合路径，未批准整个 C1 或默认切换。

语义与 CI 对齐

复用现有 typed status、null unknown-effect、scoped permission、source reader 和 PR cadence vocabulary；没有新 closed set、substring classifier、checked cache 或公共 skip flag。PR-specific 文案留在 provider 层，generic admission/settlement 义务不变，optional observer 不能把 must_attempt_work 变成建议。Root 还在当前 C1 Core 记录了实际 serving PID、redacted runtime locator 和 ps/lsof executable mapping，确认 Node v26.0.0 与 installed TS source 对应。该局部真实来源证据不认证自然 process drain。Root 最终标准 premerge 的 6 项 direct 与 15 项 selected 检查 PASS，包括双语、编译、maintainability、vocabulary 和公开边界。Goal change-quality policy 当前为 disabled，没有伪造 receipt。CI 的当前失败/未完成原因若未知，仍作为独立 diagnostic UNVERIFIED 保留；APPROVE 不代表所有 CI 绿，也不替代立即 merge-readiness。

我的整体评价 / Verdict

没有发现阻断本切片的生产问题。problem_context 为 justified_increment：普通任务免于加载无关 provider 的可用性依赖，现有权限、work/cadence、原来源与结果继续由同一 owner 持有。long_horizon 结论为 preserved，依据是 scoped source/requirement/replay 和 once-only effect/settlement 路径；user_experience 在这条 unavailable-provider 路径为 improved，未新增交互轮次。完整长期效果、产品旅程和性能不由这批局部证据推定。

现有 scoped capability resolver 持有权限，reconciler 持有 GitHub 观察，typed dispatcher 持有结果校验；把导入放在既有获权 provider try 内即可隔离可选代码，无需新增 owner 或 wrapper。

本次边界仅是 executing Turn 的可选 PR provider 加载及故障隔离；权限、cadence、原完整来源、capture、admission 和 settlement 都保留既有 owner。

已检查 hook、turn_start composition、quota/native callers 和 RFC append；此次内部加载修改没有改变 rendered presentation。

所触及范围无视觉呈现待验；真实模型、完整 Settings/产品旅程和平台资格不在本局部结论中。

同域的有界整理已体现为删除 eager internal alias，并让测试使用原 reconciler owner；没有保留需要维护的第二套 adapter。当前 exact-head、完整来源/负例、installed Core 与 C1 affected-path 证据支持此实现边界。合入仍须独立读回 review、thread closeout、未变 head 和 readiness，并在已给出的 maintainer 授权下处理相应 CI/policy hold。完整原目标和原十项保持开放，完成前仍需完整回归。

English verdict: APPROVE

The exact head 76502e5d18b46958a70a8cfed2218ddf3c0e2f3d isolates optional PR-provider loading behind the existing invocation-time permission check and failure owner. Source, negative, real File/SQLite, installed Core and current-C1 affected-path evidence support this bounded increment. Historical setup failures and diagnostic CI limits remain visible; full R1, actual-model adoption, total-cost improvement and C1 default cutover are not qualified.

[cocolord]

Merge decision for exact head 76502e5d18b46958a70a8cfed2218ddf3c0e2f3d:

The complete author-owned COMMENTED approval conclusion was published and read back at this head: #8821 (review) . Current policy26 result/body verification passes all19 required evidence items. Approval closeout is clear, with no blocking reviews or unresolved review threads; COMMENTED is not formal GitHub self-approval.

Required local validation passed: standard premerge6direct/15selected, final-head bilingual gate, cold provider/permission negatives, source File/SQLite observation/cadence/response-loss/full-source cases, independently installed Core execution/recovery/refusal and full Todo CLI/facts retention, and the current-C1 affected source/Core path. The immutable main original74 retains70PASS/4unchanged setupFAIL; the current carrier fixture preserves its original assertions. Current-C174/74 uses its own committed fixture correction and does not repair those main failures. CQQ is disabled for this Goal; no receipt is claimed.

Immediate readiness is ready=false only for failed/pending/incomplete status checks. Its normalized snapshot is49checks:33success/10failure/6pending, from51raw rows with2superseded. Unknown CI causes remain unresolved; this is not an all-green statement. The head is OPEN, unchanged and MERGEABLE, and review threads are complete with0unresolved.

Under the existing maintainer authorization for independently reviewed positive changes, use admin bypass for this CI/GitHub policy hold. This bypass does not replace review, bilingual/required local checks, thread resolution or exact-head validation. No unresolved finding or conflict is bypassed. Read back the merge commit, source tree, sign-off and main before continuing #8822. Full D2/D5, original-ten acceptance, full regression, actual-model adoption, complete cost and C1 default cutover remain open.