loopx-project · huangruiteng · Sep 12, 2026 · Sep 11, 2026 · Sep 11, 2026
@@ -2664,6 +2664,14 @@ The original direction remains; execution cards expand these stages rather than
 
 #### Durability execution cards
 
+Capability-gap consumers now share the TS requirement/resolution owner across
+legacy and canonical inputs, including quota's Monitor capability partition.
+The old Python missing-set and owner/repair decision builders are removed;
+source adaptation and read-only candidate ordering remain. This is T3 read-policy
+consolidation with disclosed resolution-priority/empty-source/identity corrections,
+not capability enablement, a durable permission receipt or D1–D3 qualification.
+The existing permanent Markdown projection and cutover holds remain unchanged.
+
 The T3 decision-dependency read policy now shares one TS owner for scope coverage,
 exact links and consistency diagnostics. Explicit gate recipients are independent
 of claim attribution; conflicting exact targets request repair rather than grant

@@ -2139,6 +2139,12 @@ T3 decision-dependency 读取策略现由同一 TS owner 解释 scope coverage
 
 **D1 — 资格化永久投影交付，可与 T1/T2 重叠推进。**
 
+能力缺口 consumer 在 legacy/canonical 输入上共用 TS requirement/resolution owner，
+包括 quota 的 Monitor 能力分流。删除 Python missing-set 与 owner/repair 决策 builder，
+保留来源适配和只读候选排序。这是披露修复优先级、空来源和 identity 修正的 T3 读取
+规则收拢，不是能力启用、持久权限回执或 D1–D3 资格化。永久 Markdown 投影与 cutover
+门禁不变。
+
 T2 的无 lease 原生 Monitor 观察与独立后继现由同一 canonical CAS／receipt 提交；
 route planner 本身仍不授予权限。CLI 将已提交回执交给既有 journal/outbox renderer，
 展示失败标为 pending，不回滚提交、不重新生成后继；quota 继续消费同一 v0 业务回执

@@ -487,6 +487,20 @@ delivery. This does not finish all T2 commands or authorize whole-Goal promotion
 
 **T3 — close remaining structured consumers, then remove their old reads.**
 
+Capability resolution now shares `agents/capability_gate.ts`: missing prerequisites,
+repair outputs, owner/agent resolution and blocked-Todo bindings have one typed
+owner. Quota planning v1 passes normalized requirements, not Python-computed
+missing lists; Monitor partitioning invokes the same rule in-process. The public
+gate uses one batch; exact-target recovery callers retain a bounded value-only
+cached bridge to that rule, not a second implementation. Python keeps legacy
+codecs, candidate source/eligibility and the shared profile/rank adapter.
+Disclosed corrections: a shared resolution binding names the highest-priority
+blocked Todo, display variants deduplicate by Todo identity, and an authoritative
+empty backlog never revives stale first-item diagnostics. Target capabilities
+remain repair outputs, not permission or installed capabilities. No new provider,
+source inventory, enablement or promotion is introduced; compact candidate-source
+limits and the remaining T3 consumers still require their own closure.
+
 Quota's scope/claim consumer now composes selection, bounded visibility and the
 existing resume planner in one `todo.quota_planning.project` call per source.
 `quota_selection.ts` replaces the Python claim-visibility module and the separate

@@ -419,6 +419,15 @@ summary 对候选关系批量调用，避免每对 Todo 一次 RPC；legacy comp
 状态快照 parity。T3 仍需处理旧 action-token fallback 路由及从压缩 summary 重建诊断的
 消费者，不把它们列为已迁移；不宣称 T1/T2、全部 T3 或持久化／promotion 完成。
 
+能力缺口与修复路由现由 `agents/capability_gate.ts` 统一解释执行前提、修复产出、
+owner/Agent 责任和受阻 Todo 绑定。Quota planning v1 传归一化的要求，而不是 Python
+算好的 missing；Monitor 分流在 TS 进程内复用同一规则。公共 gate 一次批处理，精确目标
+恢复调用保留有界、只缓存归一化值的桥接，不保留第二套判断。Python 继续负责 legacy
+codec、候选来源／资格和共享 profile/rank 适配。明确修正：共享缺口绑定最高优先级受阻
+Todo，同一 Todo 的不同展示不重复计算，权威空 backlog 不再复活陈旧 first-item。
+target capability 是修复产出，不是安装或授权。没有新 provider／inventory／enablement／
+promotion；压缩候选来源的上限和其余 T3 consumer 仍需分别闭合。
+
 列表过滤现改用 `compact_evaluated_todo_group`，不再用仅活动项重算 resume。
 初始解析／canonical 读取仍通过 TS owner 在完整来源上求值；过滤要求匹配的已求值
 条件，不能把归档中的已完成依赖变成丢失。共享合成 fixture 增补“有 scope 无 outcome”

@@ -157,6 +157,32 @@ The row may be rendered as a "task" card for operator familiarity, but API and
 state names should keep `todo` terminology to avoid implying a second runtime
 model.
 
+### Capability prerequisites and repair outputs
+
+The read-only capability gate distinguishes `required_capabilities` (execution
+prerequisites) from `target_capabilities` (what a repair task aims to provide).
+A target is not its own prerequisite, but declaring it neither enables it nor
+grants credentials, production access, write scope, a claim or a lease. A task
+repairing `network` can therefore be runnable while another task requiring
+`network` stays blocked until availability is observed.
+
+`agents/capability_gate.ts` owns missing-set evaluation, resolution responsibility
+and impacted-Todo bindings. Unknown missing capabilities retain Agent repair
+routing; `credentials` and `production_access` retain owner routing. Other runnable
+work remains available when one candidate is blocked. A shared capability binding
+uses the highest-priority affected Todo rather than whichever display row appears
+first, while preserving all affected Todo IDs. Display variants of one ID do not
+become separate tasks. An explicitly empty evaluated backlog is authoritative;
+the older first-item source is used only when the backlog field is absent.
+
+Quota's v1 planning request carries normalized requirement facts, so the same TS
+rule partitions due Monitors without per-Todo RPCs or trusting stale precomputed
+missing lists. Candidate source/eligibility and profile ranking remain adapters;
+this does not claim complete-inventory selection or change Monitor due/writeback
+and deferred-replan rules. Use `quota should-run --include-detail agent-todos` to
+inspect the existing full capability-gate detail; the default CLI keeps its
+compact representation. No configuration, UI control or storage migration is added.
+
 ## Handoff Notes
 
 Inter-agent handoff should appear as a typed note attached to existing todo,