fix(periodic-report): name declared source adapters that failed activation

[songoow]

Goal / source

docs/reference/extensions.md says hook-adapter discovery "admits a factory only when the extension is enabled, doctor-ready, and authorized for every declared adapter permission" and that failures stay isolated. The periodic-report README says "a temporarily unavailable owner leaves the request pending; LoopX never falls through to another provider." Both hold. What was missing is that the failure was invisible.

Current gap

loopx/extensions/hook_adapters.py already records each failed adapter as a content-free ExtensionHookAdapterFailure, but discover_periodic_report_request_ports (request_action.py) built PeriodicReportRequestPorts(adapters=complete) and dropped discovery.failures. So:

• a declared Lark adapter whose factory import failed, or whose extension was disabled / doctor-stale / API-incompatible, produced periodic-report request source adapter is unavailable with no adapter id, no code and no next step — the same text as "no provider installed";
• an unreadable extensions/state.json returned an empty discovery, also indistinguishable from "nothing installed".

This is the silent-fallback shape AGENTS.md asks to avoid: control flow stayed fail-closed, but the actionable context was discarded.

Observable result

situation	before	after
--source-adapter-id probe-source, factory import fails	…unavailable: probe-source	…unavailable: probe-source is declared but failed activation (extension_hook_adapter_unavailable); inspect loopx extension doctor --execute``
no selector, only failed adapters	…unavailable	…unavailable: declared adapters failed activation: probe-source (extension_hook_adapter_unavailable); inspect …
adapter declared twice / missing a port	silently absent	ambiguous_adapter_ports / incomplete_adapter_ports named
extensions/state.json unreadable	empty catalog	discovery state_error_code = extension_state_unavailable; selection says extension state could not be read

Receipts, journal entries and schemas are unchanged; only refusal text gains identity and next action. ExtensionHookAdapterDiscovery.state_error_code and PeriodicReportRequestPorts.failed_adapters/state_error_code are defaulted, so existing constructors keep working.

Real-path readback (this head)

Installed an extension through loopx extension install --execute whose [[hook_adapters]].factory = "missing_extension.hooks:build_adapter", then called discover_periodic_report_request_ports on the real state file (no monkeypatch): adapters=(), failed={'probe-source': 'extension_hook_adapter_unavailable'}; both selection errors shown above reproduced verbatim.

Validation

• New test_request_ports_name_declared_adapters_that_failed_activation and test_request_ports_name_unreadable_extension_state; existing real-path discovery test extended with the unreadable-state case. All fail on base.
• test_lark_periodic_report_request + test_extension_runtime + test_capability_code_maps: 79 passed. Periodic-report suites (pending_intent, native_successor, turn_writeback, cadence_journal): 72 passed.
• ruff clean; semantic advisory: no new vocabulary carriers. README states the rule.

Rollback

Revert the single commit.

[loopx-agent]

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

Exact head: 51f6671. REQUEST_CHANGES — 两项 P2：显式 selector 隐藏 state-read loss；真实 disabled/doctor-stale 声明在 producer 前被跳过。

动机

从 Goal Channel 请求报告的用户和维护该来源适配器的操作者，需要区分未安装、声明无法激活和状态读取失败。以前导入失败的已声明 adapter 与空 catalog 都只报不可用；本改动让部分失败带上名称、无内容的错误码和修复提示，但显式选择下的坏状态文件、disabled/doctor-stale 声明仍退回旧通用提示。实际 CLI 已验证 factory 失败和未指定 adapter 时的状态错误更明确；完整承诺仍有两处已复现的诊断缺口，不能认定所有声明失败都已可观察。本 PR 不新增报告能力、开启扩展、运行 provider 修复、授予权限或外部写入；不认证真实 Lark、已安装前端或模型采用。

改动思路

保留 extension lifecycle/permission 与报告 journal 的原 owner；只把无内容的失败事实投影到现有拒绝信息，不新建启用策略、Python 授权源或恢复框架。当前边界是既有报告来源选择的诊断修复；在名称/原因传输完整后可独立交付，不需要把扩展升级、API 兼容或安装态报告旅程全部并入本 PR。 最强反例是测试手工供应 failure record，真实 discovery 却提前丢掉声明；formatter 正确并不证明操作者能得到其原因。当前 factory failure 的改善确有价值，不因 line count 或拒绝配额否定；完整宣称须与真实 producer/caller 一致。

源码流程为 CLI request → installed catalog/activation → complete port选择或失败信息 → 原 journal/settlement。此路径说明源自代码，非额外运行证据。使用下面实测对照表而非未经渲染核验的图。

实际来源/选择	固定 base	当前 head
已启用、doctor通过、factory import失败；显式/省略id	通用 unavailable	名称、content-free code和doctor指针
状态坏JSON；省略id	通用 unavailable	明确 extension state could not be read
同一坏状态；显式id	通用 unavailable: id	仍相同，state loss被遮住
真disabled或真实runtime identity失效；两种id选择	通用 unavailable	仍相同，failure集合空
真空catalog；三个selector	原拒绝	全部完整输出相同，无效果

doctor提示不执行doctor或enable，不授扩展权限。原绑定/结算仍只认journal中的adapter，未增加fallback。声明可见与factory准入应复用既有owner分别处理；把“不能激活”表现成“未声明”会让修复路径失真。

具体改动

改动前规范为 docs/reference/extensions.md 的完整版本 70bc981，Manifest Contract L770–828，以及同revision loopx/capabilities/periodic_report/README.md 的 Request from a Goal Channel L59–96；实现前已读。完整PR为5文件+186/-8，两处Python owner、两处测试、一段README。Manifest Contract 与 Request from a Goal Channel implemented：原权限/完整port选择/固定journal owner保留。current-state-diagnostic 和 current-declared-failure-diagnostic 是本PR明确提出的当前交付要求（不是发明的roadmap编号），分别not_met：下面的真实反例仍失败。S8/S10只用于归属/诊断框架，不要求完成整个路线图，也不拿新README反向定义旧规则。

关键代码讲解

• ExtensionHookAdapterDiscovery L41 增加defaulted transient state_error_code；catalog读失败返回既有extension_state_unavailable词汇，不带异常内容。
• discover_extension_hook_adapters L78 仍由原activation验证后才导入factory。但L107的ready预过滤使disabled/stale连匹配声明与_failure都未进入，影响新诊断宣称。
• _unavailable_detail L85 集中命名failure，但L92 selected-id通用返回早于L94 state error；当前全局读损失因此只在省略selector时出现。
• discover_periodic_report_request_ports L185 保持complete bind/settle grouping，转交已有failure和ambiguous/incomplete标签；不改变合法selection和settlement_adapter。
• 新测试覆盖手填broken/partial/duplicate与省略id的state error，旧真实factory fixture扩展坏state；README承诺disabled/stale等都命名。手填记录不能证明真实生命周期failure producer已采用。

新state字段是派生诊断hint，不是新持久化权威；两类port标签局限于该formatter，不能驱动权限。既有extension_state_unavailable复用原resolver词汇，没有新capability、provider包或配置开关。Python在当前专门manifest/factory IO及展示边界，TS授权规则未复制。

对主干的风险

P2，显式selector遮住状态读取失败。 真实CLI安装合成extension后，把独立state改为坏JSON；省略id正确报state read loss，传--source-adapter-id probe-source或never-declared却只报unavailable: id。没有raw content、factory或journal效果；问题是可行动上下文丢失。将全局 state_error_code 在显式 selector 的通用 unknown-id 返回之前处理；保留真实 failed-id 的明确诊断、合法选择、现有错误 schema 与无效果拒绝。真实 CLI 覆盖破损 state＋已知/未知显式 id，不能只测 selector=None。 这不是要求改变selector授权或为错误状态选择另一provider。

P2，新声明失败投影没有真实disabled/stale producer。 README L78 明确列出disabled、doctor-stale且称never silently dropped；真实install后调用原disable，或修改doctor已验证的可执行文件使identity失效，catalog的ready=false，discovery在L107跳过。实际ports=()、failed_adapters={}、state_error_code=None；显式和省略id都是原通用unavailable。gate仍正确拒绝，不是权限绕过或新增执行回归，但当前完整改善未交付。在既有 discovery owner 中先识别匹配声明，再把生命周期未准入记录为现有 content-free failure；disabled/stale 分支禁止加载 factory 或执行 doctor。保留每个 hook 的 scope filter和完整端口规则。若当前只交付已有 factory failures，应明确缩小 README/PR/测试资格，而不能宣称 disabled/doctor-stale 已命名。

语义与 CI 对齐

独立真实CLI共5状态×3selector的base/head对照，完整JSON/exit/stderr比较；9个空catalog/disabled/stale输出逐字相同，所有15拒绝前后runtime文件hash不变。factory显式/省略两个独立语义oracle在base失败、head通过；显式坏state及disabled/stale三项head oracle仍失败。原生install/doctor用合成无外部效果provider；source请求始终拒绝，无真实Goal、Lark消息或provider内容。全流程在各自隔离root，未借activeGoal。

当前七个完整module 151 passed/0 skipped，63.82s；固定base三个extension/codemap module 77 passed/0 skipped，25.42s。72个报告pending/native-successor/writeback/cadence用例验证原journal/retry/ACK边界。测试正向有效，未覆盖上面的实际failure组合。canary 5 direct/15 selected、0失败/0hold；Ruff、diff、5路径公开scan、advisory后全semantic通过。advisory只分析2生产路径/0 supported carriers，不证明动态/局部string语义；全semantic producer/cross-runtime未验限制保留。预算workload/limit/root未改。

检查了相关#6425 API revalidation、#6418 disabled upgrade、#6420参考manifest修复的当前frame；它们是各自owner的提案，不继承其验证，不要求先merge它们。这两处实际诊断缺口在本head独立成立。默认未安装/禁用仍无factory/admission/side effect；有意诊断文字改变已披露，generic owner无Lark私有payload/词汇或substring判定。新字段不叫权限或actor lifecycle；doctor是guidance，原权限/settlement是machine obligation。

未查询、轮询、等待或触发GitHub CI。CLI错误路径实际验证；真实Lark/安装前端/模型理解未执行，也不宣称新产品能力已完整发布。首次reviewer调用用了不存在的codemap路径、install位置参数，均保存并按仓库原CLI修正；不是PR失败，不以删断言变绿。

我的整体评价

REQUEST_CHANGES，只绑定 51f6671。归属与规模合理、部分实测改善成立；两处关键failure composition仍未达到当前明确宣称。保留 extension lifecycle/permission 与报告 journal 的原 owner；只把无内容的失败事实投影到现有拒绝信息，不新建启用策略、Python 授权源或恢复框架。当前边界是既有报告来源选择的诊断修复；在名称/原因传输完整后可独立交付，不需要把扩展升级、API 兼容或安装态报告旅程全部并入本 PR。 bounded future-facing pass已通过共同formatter收敛；下一步应在同owner让声明诊断与activation分离并修复优先级，不增加parallel readiness/state框架。最小重验是实际installed生命周期和坏state的已知/未知/省略selector对照，保留无效果/原owner/no fallback，再跑现有完整测试。合并/启用/对外写权限另行判断。

English verdict: REQUEST_CHANGES — 51f6671. Real install/discovery/CLI comparisons confirm the factory-failure and omitted-selector state diagnostics improve, while explicit selectors still hide unreadable extension state and disabled/stale declarations are dropped before failure production. Preserve admission/no-effect/no-fallback, fix these projection branches or accurately narrow the promised lifecycle scope, and exercise real installed cases. All151 current tests and risk checks pass; mocked failure records do not cover the two reproduced gaps. No new permission or production effect regression is alleged.

[songoow]

Addressed both P2s. The branch is rebased on f713e0f; the new head is 361ddf1, with two commits on top of the original one.

P2-1, explicit selector hid state loss (request_action.py, _unavailable_detail):

• The global state_error_code is now checked before the generic unavailable: <id> return.
• Bad state with an omitted, known (probe-source) or unknown (never-declared) selector now always reports extension state could not be read (extension_state_unavailable).
• A named failed id still gets its own diagnostic, and selection, the journal, settlement and no-fallback are unchanged.

P2-2, disabled/doctor-stale declarations dropped (hook_adapters.py):

• The ready pre-filter is gone. Each installed extension's declarations are first matched by phase, capability and hook.
• Non-ready ones become content-free failures with the existing readiness codes extension_disabled / extension_doctor_not_ready. Those codes are now constants in extensions/runtime.py, shared with optional-binding resolution.
• That branch never resolves activation, imports the factory or runs the doctor.
• Declared-only, not-installed manifests are still skipped, so they cannot be reported as "disabled".

Tests (real installed lifecycle, no hand-built failure records):

• Discovery: install, then a positive control showing a ready extension reaches the factory loader. Then real disable, re-enable, and changing the doctor-verified executable. Assertions: named adapter and code, zero factory loads, state bytes unchanged, and an adapter declared for a different hook is not named.
• Real loopx periodic-report request --execute CLI over {disabled, doctor-stale, bad state} × {omitted, known, unknown selector}. Assertions: exit 1, the adapter or state-loss code, factory module not imported, file tree unchanged, no journal.
• Reverting either fix fails its tests.

Results:

• Extension and periodic-report modules: 382 passed.
• Broader 44-module set: 1428 passed / 1 failed. The failure is test_content_digest_single_owner.py::test_only_the_owner_module_states_a_whole_value_digest_shape (codex_sessions.py), which fails identically on clean main.
• canary premerge --from-git-diff: 5 direct / 18 selected / 0 failures / 0 holds, run with --timeout-seconds 600. The drift smoke takes about 285 s on both base and head, which exceeds the default 120 s limit.
• Semantic advisory: 0 candidates.

Scope wording:

• The README no longer claims "API-incompatible" is named, because nothing on this branch produces it; the installed-snapshot API recheck lives in fix(extensions): re-check installed snapshots against the runtime API and reject conflicting declarations #6425.
• The message text changed from "failed activation" to "was not activated", since a disabled extension never attempted activation.

中文：显式 selector 下坏 state 现在优先报告；disabled/doctor-stale 声明改为先匹配再以既有就绪码记为无内容失败，且不导入 factory、不运行 doctor。测试走真实安装/禁用/doctor 失效与真实 CLI；README 只承诺当前分支实际产生的代码。

[loopx-agent]

Reviewer: model_agent · gpt-6.1-sol · OpenAI · runtime_reported · xhigh

Exact head: 361ddf1. APPROVE.

动机

从已连接来源请求报告的用户，需要知道来源没安装、被禁用、doctor 资格失效，还是扩展状态无法读取。过去这些情况混成“来源不可用”，即使操作者显式指定来源，也可能看不到真正原因。本次把现有声明和读取失败传到原拒绝信息，真实 CLI 已显示对应名称、原因和修复指针；有效来源的选择、固定结算 owner 与无效果拒绝保持。这个切片完成既有来源选择的诊断修复，不扩大报告能力。真实 Lark、已安装前端、模型理解、全部安装升级故障和长期净收益未验证；不新增启用、发布或合并权限。

改动思路

保留原 extension lifecycle/permission 和报告 journal owner：discovery 区分“声明存在但未准入”与“状态根本读不到”，报告 ports 只把事实投影成可行动的错误。disabled/stale 分支先匹配 phase、capability、hook，再记录既有 lifecycle code，不导入 factory，不运行 doctor。ready 分支继续走原 activation、权限与完整端口校验；失败不能变成另一个 provider 的授权。

共享诊断事实并复用已有 readiness 状态，比再建启用策略、重试框架或 Python 授权源更小。当前 PR 边界是 discovery 到既有报告 request 的诊断传递及必要测试和文档，不新增 capability、设置或用户入口。runtime 中两处状态常量复用已完成有界重构；现有共享入口保留原调用与规则，没有进一步框架化的必要。

下面是实际 base/head CLI 对照；表格解释结果，不替代执行证据。

触发	common base	当前 head
disabled/stale；省略或已知 id	通用不可用	名称及 lifecycle code，未导入 factory
状态坏 JSON；省略/已知/未知 id	通用不可用	明确 state-read loss
ready 但 factory 缺失	通用不可用	命名失败和 doctor 指针
空 catalog、有效选择、未知结算 owner	原拒绝或原选择	原行为保持

具体改动

完整复核 common base f713e0f 上七文件 +565/-21，三个生产模块、两份文档和两个测试模块；三次提交均有 DCO。规范依据 docs/reference/extensions.md @ f713e0f 的 Manifest Contract：只有 enabled、doctor-ready、逐项授权才准入，失败隔离，implemented；同 revision 的 loopx/capabilities/periodic_report/README.md 的 Request from a Goal Channel：绑定 journal owner、不可临时换 provider、失败待恢复，implemented。当前修复也覆盖旧 review 指出的显式 selector/state-loss 与真实 lifecycle producer 缺口。

关键代码讲解

• _lifecycle_non_admission / discovery 从 catalog 的 installed/enabled/ready 派生诊断；匹配作用域后命名 disabled/stale，不执行 factory 或 doctor。catalog 读取失败保留单独 state_error_code。
• PeriodicReportRequestPorts._unavailable_detail 保留已知 failure 优先，再让全局 state loss 先于普通 unknown-id，省略、已知和未知 selector 都看得到读取失败。用“未激活”覆盖从未尝试激活的 disabled 状态，避免错误叙事。
• discover_periodic_report_request_ports 继续完整 bind/settle 分组和重复端口拒绝，再转交失败集合；失败记录不能替代 complete adapter，settlement 仍按已持久 owner 查找。
• runtime 状态常量 仅集中已有 extension_disabled / extension_doctor_not_ready 值。独立比较整个 runtime AST，代回两处常量后与 base 相同，没有复制新的 lifecycle 决策。

全仓调用扫描表明当前生产 discovery 消费者是该 request owner。失败集合与 state_error 是派生诊断，不新增持久状态或 actor lifecycle。Python 位于既有 manifest/factory IO 与显示边界；权限和内核权威未迁移。

对主干的风险

无当前阻塞发现。独立 Python 3.11.17 真实 CLI install/doctor、disable、可执行文件 identity 变异、坏状态、factory 缺失，对五种状态各跑省略/已知/未知 selector，共十五个拒绝。每次 runtime 全树 hash 不变，factory marker 不出现，doctor 计数不变，没有 Goal journal。实际 ready factory 的导入、单 provider 选择和结算 owner 查询有正向控制，防止“没导入”是空测试。base/head 有九处预期诊断变化、六处完整 JSON/exit/stderr 相等；未归一错误、授权或拒绝事实。

额外在真实旧提交 51f6671 运行同一 harness：disabled/stale 的省略/已知 id 及坏 state 的已知/未知 id 六个旧 P2 反例仍失败，当前全部修复。旧版已修的省略 state-loss 保留；factory 两处措辞由 failed activation 改为 not activated，不把额外格式变化混成新权限。

语义与 CI 对齐

156 项完整相关测试/0 skip，包括原 pending、successor、writeback、cadence journal 的重试与 ACK 边界；canary 5 direct / 18 selected 全通过，0 failure/hold；Ruff、diff、公开边界、advisory 后完整 semantic 通过。advisory 的零 supported carriers 不是无语义变化证明。现有 readiness code 复用；新增 transient port 诊断保持局部，没有私有组织文案或字符串型 Goal 分类。

feature-off 并未被省略：未安装保持原通用拒绝，disabled/stale 始终无 factory、doctor、provider 写入或 journal，新增的只是已披露诊断。doctor 指针是建议，原 admission/权限和 journal owner 是强制条件。没有 UI/设置变更，也没有真实 Lark、前端或模型采用资格；未查询、等待或触发 GitHub CI。最初 probe 误用无 main 的入口、再尝试同版本不同 manifest 重装，已按原 CLI 和独立 runtime 更正；这些测试准备错误没有计入产品失败或通过。

我的整体评价

APPROVE，仅绑定当前完整 head，不继承旧结论。长期维护改善在于失败能回到正确 owner，用户诊断改善有真实反例与读回；没有把测试数量当作整个报告旅程或长期收益。规模与已有 owner 合理，完整端口、scope filter、默认关闭和无 fallback 保持。旧正文仍有旧测试数、API-incompatible 及“revert single commit”描述，当前 README 和本评审明确最终范围；不宣称相关 API 重验或升级回执的其他 PR 已完成。合并 readiness 与维护者 authority 单独判断。

English verdict: APPROVE — 361ddf1. Full-scope re-review and real install/CLI probes verify both prior P2 fixes, preserved legal selection and fifteen effect-free refusals. Nine intended diagnostic deltas and six unchanged refusal outputs are recorded against the common base; six caller counterexamples fail on the prior head and pass here.156 tests and current risk checks pass. No live Lark, installed frontend, model adoption or merge authority is claimed.