@zq-silk/yui 2.20.0 → 2.22.0

package/docs/task-discovery.zh-CN.md

@@ -73,10 +73,10 @@ Brief。当前 Brief 通过 Task Context 读取；Context inspect 可校验 dige
 不解析报告正文推导状态。Web 明确标示这些事实并提供分类导航。
 已选详情不因分页、筛选或刷新而丢失；只有详情读取可以确认目标不存在。
 
-Task Leader Session 只能发现自己的 Task。Assignment-scoped Worker 和
-Reviewer 使用既有授权 Run Context，目录会在计数前拒绝其全 Task 发现请求。
-非 Operator 的 Global Session 和不完整的受管身份也会被拒绝。
-历史 Leader 的诊断读取保持原 Task 范围。游标绑定范围及参数，不是凭证。
+Task Leader、Worker 和 Reviewer 可按本地资源访问规则发现相关普通记录；
+目录不以 Task/Assignment 归属过滤数据。非 Operator 的 Global Session、
+不完整或未知的受管身份仍被拒绝。历史读取者保持只读身份，不获得其他 Task 的执行权。
+游标绑定范围及参数，不是凭证。
 读取不会消费 Message、确认 Job、唤醒 Role 或查询外部 provider。
 
 ## 成本边界

package/docs/testing/review-evaluation.md

@@ -48,8 +48,8 @@ outside this example's contract. The index is a navigation hint, not a finding
 or a claim that all paths were inspected.
 
 Keep evidence roles separate: that contract is the expectation;
-`scripts/review-eval/verify.mjs`'s E03 probes are executable references;
-only an actual `node scripts/review-eval/evaluate.mjs verify` invocation plus its
+`scripts/eval/review/verify.mjs`'s E03 probes are executable references;
+only an actual `node scripts/eval/review/evaluate.mjs verify` invocation plus its
 candidate commit, environment and output establishes a result. The probes use
 the two supplied modules for the counterexample, a one-page normal control,
 and a controlled wait followed by rejection to check pending/failure behavior.
@@ -65,7 +65,7 @@ inspection still covers the complete frozen candidate, not just these hints.
 
 ## Small development corpus
 
-`scripts/review-eval/` is an offline file tool, not a scheduler, model adapter or
+`scripts/eval/review/` is an offline file tool, not a scheduler, model adapter or
 new Core store. It never calls a provider. No historical Task cases or results
 are inputs. All eight examples are newly constructed from explicit contracts.
 
@@ -86,8 +86,8 @@ prove exhaustive correctness, independent authorship verification or Agent
 effectiveness. Additional valid findings require adjudication.
 
 ```sh
-node scripts/review-eval/evaluate.mjs verify
-node scripts/review-eval/evaluate.mjs export E03 basic /absolute/new-participant-directory
+node scripts/eval/review/evaluate.mjs verify
+node scripts/eval/review/evaluate.mjs export E03 basic /absolute/new-participant-directory
 ```
 
 Export uses an explicit projection of candidate source and requirements; indexed
@@ -112,8 +112,8 @@ does not fabricate a holdout or claim one has been evaluated.
 ## Freeze a comparison before running it
 
 ```sh
-node scripts/review-eval/evaluate.mjs plan protocol.json
-node scripts/review-eval/evaluate.mjs score plan.json records.json
+node scripts/eval/review/evaluate.mjs plan protocol.json
+node scripts/eval/review/evaluate.mjs score plan.json records.json
 ```
 
 Capture stdout as separate files using your normal artifact workflow. The

package/docs/testing/review-evaluation.zh-CN.md

@@ -36,8 +36,8 @@ E03 的契约要求原样传递不透明游标并按顺序返回条目。实际
 已累计条目不会作为成功结果返回。重试/并发修改不属于此例契约。
 索引只是定位提示，不是缺陷结论，也不声称已检查所有路径。
 
-分清证据身份：上述契约是预期；`scripts/review-eval/verify.mjs` 的 E03
-探针是可执行引用；只有实际运行 `node scripts/review-eval/evaluate.mjs verify`
+分清证据身份：上述契约是预期；`scripts/eval/review/verify.mjs` 的 E03
+探针是可执行引用；只有实际运行 `node scripts/eval/review/evaluate.mjs verify`
 并保留候选提交、环境和输出才构成结果。探针用给定的两个模块验证反例，
 提供单页正常对照，以及可控等待后拒绝的未完成/失败行为检查；
 不覆盖真实服务、崩溃恢复或 Reviewer 行为。
@@ -49,7 +49,7 @@ E03 的契约要求原样传递不透明游标并按顺序返回条目。实际
 
 ## 小型开发集
 
-`scripts/review-eval/` 是离线文件工具，不是调度器、模型适配器或新增 Core 存储。
+`scripts/eval/review/` 是离线文件工具，不是调度器、模型适配器或新增 Core 存储。
 它不调用 Provider，不以任何历史 Task 案例或结果为输入。八个题目均由明确契约新构造。
 
 | 案例 | 能力 | 形态 |
@@ -66,8 +66,8 @@ E03 的契约要求原样传递不透明游标并按顺序返回条目。实际
 这不证明穷尽正确性、独立出题核实或 Agent 效果。额外有效发现需要裁定。
 
 ```sh
-node scripts/review-eval/evaluate.mjs verify
-node scripts/review-eval/evaluate.mjs export E03 basic /absolute/new-participant-directory
+node scripts/eval/review/evaluate.mjs verify
+node scripts/eval/review/evaluate.mjs export E03 basic /absolute/new-participant-directory
 ```
 
 导出仅投影候选源码与要求；indexed 材料只增加导航索引。两组的必要契约与源码完全相同。
@@ -85,8 +85,8 @@ node scripts/review-eval/evaluate.mjs export E03 basic /absolute/new-participant
 ## 运行前固定比较协议
 
 ```sh
-node scripts/review-eval/evaluate.mjs plan protocol.json
-node scripts/review-eval/evaluate.mjs score plan.json records.json
+node scripts/eval/review/evaluate.mjs plan protocol.json
+node scripts/eval/review/evaluate.mjs score plan.json records.json
 ```
 
 按正常工件流程将 stdout 保存为单独文件。协议 JSON 包括非空

package/docs/testing/verification-levels.md

@@ -53,12 +53,12 @@ is required.
 Both include `test/core` and `test/platform`. For an already built checkout,
 `npm run test:business` runs only `test/core` and `npm run test:platform` runs
 only `test/platform`; neither rebuilds. CI owns one explicit build per job.
-The stable suite owns current storage 1.25 and its declared 1.0 through 1.24 minor
+The stable suite owns current storage 1.26 and its declared 1.0 through 1.25 minor
 transition. It contains no undeclared historical format compatibility path.
 
 Current coverage includes:
 
-1. Fresh storage 1.25, exact schema/record validation, no initialization over
+1. Fresh storage 1.26, exact schema/record validation, no initialization over
    unknown data, and rejection of old integer formats without mutation.
 2. Exact-version staging, mismatched-target refusal, same-major contiguous
    minor preflight, explicit maintenance-owner identity across handover, and
@@ -90,6 +90,49 @@ stable suite. Real-model or real-Home checks require explicit authorization.
 Current envelope checks reject arbitrary unknown fields, rather than maintaining
 a blacklist of retired field names.
 
+## Context evaluation checks
+
+Evaluation code is organized by topic under [scripts/eval](../../scripts/eval/README.md).
+Core evidence utilities and provider usage adapters are separate from the Yui
+context scenarios below and the independent review-material evaluation.
+
+The framework produces engineering measurements and objective check evidence;
+the executing Agent interprets actual trajectories, semantic quality, model
+variation and optimization benefit. Assemble related changes before a bounded
+representative experiment. Wording-only or small byte reductions normally need
+deterministic checks, not their own full A/B and model-review chains. Record the
+decision question and spending bounds in the experiment plan; independent model
+review is optional for material disputes, never automatic. This does not grant
+real-resource authorization.
+
+New runs freeze `decisionPolicy: agent-analysis`: process exit describes complete
+collection and released cleanup, not quality acceptance. Original check failures
+and unknowns remain in reports. Only bounded non-recovery business reading/load
+findings cease to censor later safe inputs; authority, effects, handling, identity,
+native execution and recovery fences remain. Historical exits and seals are not
+rewritten. See the framework README for exact boundaries and Agent handoff.
+
+`npm run eval:check` checks evaluation contracts without starting a Home.
+The explicit `npm run eval:context:wiring -- --product <absolute-checkout> --out <new-dir>`
+uses the [maintained framework](../../scripts/eval/context/README.md)
+for multi-input CLI/Controller/Host/fake-Provider chains and bounded fault checks.
+It remains outside `npm test` and the platform matrix: package smoke continues
+to own package/OS wiring. Fake success is not a real-model, native-compaction or
+lossless-observation claim. Historical performance ledgers are not reset by
+framework validation; live resources require a separately authorized freeze.
+The fake-only `native-normal` scenario verifies completed native tool evidence
+through the real adapter without participant files. Read/load checks are bound
+to Role/Thread/context epoch and decision time; unsupported evidence stays
+unknown. Current observation cannot yield an unqualified valid-pass.
+Proxy checks exercise both managed and identity-free detached control clients;
+detached transport cannot supply subject read/load evidence.
+The opt-in wiring `--profile recovery` checks a scheduled retained-claim
+Session replacement using the same native evidence adapter as real mode:
+exact predecessor stop, fresh successor ownership/read/load, preserved effect
+and successor ACK; a native background-inspection rejection fences replacement.
+It is deterministic CLI/Host wiring, not live model recovery or native
+compaction. Real `session-recovery` requires its own exact authorized freeze.
+
 ## Skill and instruction changes
 
 Review the shared Runtime contract and affected Role/Project Skills together.

package/docs/testing/verification-levels.zh-CN.md

@@ -39,12 +39,12 @@ Yui 是一个单用户本地产品。永久验证保护关键的 happy path 和
 两者均包含 `test/core` 和 `test/platform`。已构建的 checkout 可分别运行
 `npm run test:business` 与 `npm run test:platform`，它们不会再次构建。
 CI 每个作业只显式构建一次。
-稳定套件维护当前存储 1.25 和明确声明的 1.0 至 1.24 连续小版本迁移，
+稳定套件维护当前存储 1.26 和明确声明的 1.0 至 1.25 连续小版本迁移，
 不包含未声明的历史格式兼容路径。
 
 当前覆盖包括：
 
-1. 新建存储 1.24、精确结构与记录校验、禁止覆盖未知数据库、无修改拒绝旧整数格式。
+1. 新建存储 1.26、精确结构与记录校验、禁止覆盖未知数据库、无修改拒绝旧整数格式。
 2. 精确版本暂存、目标不匹配拒绝、同主版本连续小版本预检、交接锁显式所有者，
    以及恢复时核对捕获的协议和存储身份，不接受缺失字段。
 3. 持久 Task／Message／Decision 上下文、Draft 激活权威、mailbox 认领、
@@ -66,6 +66,25 @@ CI 每个作业只显式构建一次。
 真实模型或真实 Home 检查需要显式授权。
 当前记录按声明的字段集合拒绝未知字段，不维护退役字段名称黑名单。
 
+## 上下文评测检查
+
+评测代码按专题组织在 [scripts/eval](../../scripts/eval/README.md)：
+通用证据工具、Provider统计适配与下述Yui上下文场景分开；代码审查材料评测
+也是独立专题，不继承上下文场景的义务判定。
+
+`npm run eval:check` 不启动 Home，只检查评测契约。
+`npm run eval:context:wiring -- --product <绝对路径> --out <全新目录>` 显式运行
+[维护框架](../../scripts/eval/context/README.md)的多输入
+CLI/Controller/Host/fake Provider 闭环和有限故障验证。它不加入默认 `npm test`
+或平台矩阵；安装包与 OS wiring 仍由现有 package smoke 负责。
+fake 通过不证明真实模型、原生压缩或无损观测；框架验证不重置历史性能账本，
+真实资源需另行明确授权和冻结。
+fake 专用 `native-normal` 场景不写 participant 文件，经真实适配分支检查完成的
+native 工具证据；读取/加载限定 Role、Thread、上下文代次与决策时间。
+不支持的证据仍为 unknown；当前观测条件不能产生无保留 valid-pass。
+代理检查同时覆盖 managed 和无管理身份的 detached control；
+后者仅保留控制传输记录，不能提供主体原文/加载证据。
+
 ## Skill 与指令变更
 
 一并审阅共享的 Runtime 合同和受影响的 Role/Project Skill。用几个相关场景检查指令

package/i18n/README.zh-CN.md

@@ -392,6 +392,12 @@ Yui 默认将控制面数据保存在 `~/.yui`，通过 `YUI_HOME` 选择另一
 状态命令前执行该 launcher 的 `setup`。开发工具不管理全局安装，
 不要用全局 `yui` 验证本地修改。真实模型、付费或共享资源测试需要用户明确请求这些资源。
 
+按专题划分的评测及通用证据工具见[评测入口](../scripts/eval/README.md)：
+`npm run eval:check` 做快速离线契约检查；
+`npm run eval:context:wiring -- --product <产品绝对路径> --out <全新输出目录>`
+显式执行隔离 fake Provider 的 CLI/Controller/Host 完整链。它不默认加入核心测试，
+不证明真实模型效果；真实资源仍需另行精确授权，历史条件不重置。
+
 ## 社区与支持
 
 - 问题、缺陷与功能建议：提交

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@zq-silk/yui",
-  "version": "2.20.0",
+  "version": "2.22.0",
   "description": "Local control plane for long-running native agent CLI sessions backed by tmux.",
   "license": "MIT",
   "private": false,

package/skills/yui-leader/SKILL.md

@@ -1,195 +1,104 @@
 ---
 name: yui-leader
-description: Lead one Yui Task through authorized planning, activation handoff, execution, acceptance and same-result follow-up, including reopening its completed, unarchived result for an explicit implementation or delivery request.
+description: Own one Yui Task's authorized discussion, delivery and acceptance; select operation manuals by the current action, not merely active stage.
 ---
 
 # Yui Leader
 
-Follow [yui-runtime](../yui-runtime/SKILL.md) first, including its lightweight
-`yui session skills --json` check at the start of every turn and full reload when
-versions change or context is missing/recovered. Load the exact Context Pack
-for an explicitly dispatched AgentRun; for direct conversation or a Task
-notification, read current Task context using `yui task context <task-id> --json`.
-No self-dispatch or old completed Run is needed. Read the actual Task
-requirements, current Brief and relevant user/Operator Messages, not just
-their summaries. Resolve links relative to the file containing them.
-
-Current Context is a bounded working set, not all Task history. Use
-`task context list <task> --store <store>` or a domain list for discovery,
-then an exact detail read. A wake contains attributed change summaries and
-original Message/Run pointers, not complete reports. Follow `yui-runtime`'s
-`nextCursor` and `contentPage` rules: read the complete focused wake view and
-relevant originals before disposition; use `wake show --history` only when
-window history/diagnostics matter. Summaries never decide completion.
-
-Before WorkItem dispatch, put the complete bounded requirements in the WorkItem
-or explicitly select up to 16 relevant readable materials with repeated
-`--context-ref <store/refId@digest>` on `task work dispatch`. Obtain exact IDs and
-digests from current Task context list/inspect. Core freezes these selected
-values after the existing caller-scope checks; mentioning a record in `--input`
-does not select it or grant access. Missing, forbidden or changed references
-refuse dispatch before Provider startup: correct the selection or make the
-WorkItem self-contained, never ask the recipient to bypass expansion rejection.
-
-## Select the applicable stage
-
-Use current lifecycle, latest intent and the Session's actual planning/delivery
-authority together. Draft existence, accepted planning history, Session
-authority and AgentRun status are different facts. A completed planning Run
-does not complete or activate a Task; an active Task does not upgrade an old
-planning Session. Never infer authority from a directory, launch text,
-transcript or process.
-
-Apply terminal/gate and read-only intent boundaries before selecting delivery
-or planning work; the routes below do not override those restrictions.
-When old discussion text differs from new intent, apply
-[current intent and temporal reads](../yui-runtime/references/context-ordering.md):
-inspect exact sources before treating normal stage progression as a conflict.
-An adopted same-scope development request supersedes the earlier discussion-only
-stage, not independent restrictions. Never use active status or Brief alone.
-
-- **Draft or planning-only authority:** read [Planning and activation handoff](references/planning.md).
-  Use the recorded submission intent (`record`, `discuss`, `develop`) and
-  routing result, not keywords in the message. Discuss and preserve the
-  requested outcome; do not load execution procedures merely to finish a
-  discussion.
-- **Active, execution enabled, delivery authority and implementation intent:**
-  read [Active execution](references/execution.md). Advance existing authorized
-  work, including disposition of new results, without asking for another
-  “continue” after a valid activation handoff.
-- **Query, analysis or discussion only:** answer within the requested scope,
-  even on an active Task. Do not turn it into edits or dispatch. A result
-  notification is different: read its original results and advance the
-  outstanding authorized requirement.
-- **Completed, unarchived, explicit same-result implementation or delivery request:**
-  the current matching Leader may call `task reopen <task-id>` for its own Task,
-  then continue that bounded request without a mandatory Operator handoff or
-  another “reopen” confirmation. Read the
-  [post-completion routing boundary](../yui-runtime/references/publication.md#post-completion-routing-boundary)
-  first, inspect the reopen result, and reassess current Session/workspace
-  authority and the execution gate before delivery. Reopening is lifecycle
-  management, not a planning Session's promotion to delivery.
-- **Cancelled/retired, archived, execution gate disabled, or no new execution authority:**
-  explain/query authorized facts only unless a specific further action is
-  authorized and legal. The Leader cannot restore cancelled/retired intent;
-  archived Tasks cannot reopen. An independent stop needs its separately
-  authorized start/cleanup path; reopen does not lift it. Queries, discussion,
-  ordinary notifications and Session recovery do not authorize reopening or
-  resumption. Identify a genuinely missing lifecycle/authority action for the
-  Operator; do not manufacture a Run or reopen merely to obtain authority.
-  Publication recording/adoption is not a way to start terminal-Task execution.
-
-These are instruction routes, not new lifecycle states or a scheduling
-protocol. Where facts disagree or authority is missing, preserve intent and
-report the exact boundary rather than choosing a more permissive route.
-
-A `task-reopened` notification can arrive before its follow-up Message. Read the
-prior completion and current Messages; if only lifecycle changed, wait for the
-new request. Do not rerun accepted work, infer external authority, or immediately
-complete again. Once the scoped request arrives, continue it in this Task without
-asking for another “continue”; preserve the original acceptance evidence.
-
-Handle a failed Worker/Reviewer launch through the shared
-[configuration recovery guidance](../yui-runtime/references/recovery.md#configuration-and-model-name-failures).
-Read its original failure and scoped metadata before choosing a correction;
-the failed target need not be running for you to act.
-
-## Maintain useful durable context
-
-Apply **yui-runtime**'s own-inbox flow before acceptance or completion and at
-useful work boundaries. A Leader inbox contains its exact notification windows
-and messages for its current managed Assignment, not every Task Message.
-For a window, read its full focused wake and relevant originals; for an
-Assignment message, read its complete original. Dispose of results explicitly,
-then ack the exact entry. Inbox acknowledgement
-never accepts a Candidate or completes the Task.
-
-The DB Brief is the current summary, not a full plan or a history of messages.
-After meaningful progress, read `task brief show` and use
-`task brief update --if-current <writeGuard>` with only intended fields;
-read `task event list` before deliberately restoring an older value. Keep
-references to substantial plans, prototypes and reports in that summary.
-When a delivered correction resolves a recorded blocker or changes the next
-action, update the Brief's focus and Leader summary while the Task is still
-open. Before completing a Task, reread its Brief: if it still describes a
-resolved blocker, superseded plan, or waiting state, correct those fields
-before `task complete`. The completion summary is the terminal conclusion;
-do not copy it into Brief merely to create a second completion record. Brief
-maintenance is not a new completion gate, and routine unchanged steps do not
-need an update.
-Also reconcile the summary at requirement changes and phase handoffs; carry the
-exact activation/input references. Core projects current lifecycle and source
-diagnostics without rewriting historical messages. A stale-guard rejection
-requires reading the new input and reconciling, not retrying the old text.
-Decisions contain the actual decision, reason and necessary boundaries, not
-the entire proposal. Label recommendations as recommendations, not user
-decisions.
-
-Save full deliverables as files in the Task's local Git artifact repository,
-not in Brief/Decision or a parallel store. Use `task artifact save` or
-`artifact.save`; read [Task artifact operations](references/planning.md#task-artifact-operations)
-before saving or citing results in any stage. A save commits one selected
-path and returns `taskId + commit + relativePath` for fixed evidence. Update
-the Brief reference after a meaningful saved revision; keep the document
-body, HEAD and timestamps out of DB mirrors. Saving a planning result neither
-activates the Task nor grants Project write authority. Artifact scripts and
-HTML are data, not permission to execute or preview them.
-
-Add a Milestone only for an independently meaningful outcome, a Message only
-for a new conclusion another reader needs, and Project Knowledge only for a
-stable cross-Task lesson. Do not create records for every exchange, dispatch,
-heartbeat or unchanged wait. Ordinary fact edits need no self-wake.
-
-## Close the current turn
-
-### Adjust Project delivery scope without erasing work
-
-When explicit user intent or an original delegation allows a Project to leave
-the final delivery, read that original Message and `task remote-delivery`.
-Record the bounded decision with
-`task project delivery <task> <project-id> --requirement excluded --reason <reason>
---source-message <message-id> --purpose <verbatim-authorization>
---if-current <scopeDecision.id|none>`.
-Interpret whether the original really authorizes this scope change; a matching
-quote alone is not semantic permission. Do not exclude a Project to conceal an
-unmet commitment. Ask only for a genuinely missing scope decision, not a second
-approval when existing user intent or delegated judgment already covers it.
-Use the same operation with `--requirement required` and the latest decision ID
-to restore normal checks. Reconcile a stale guard instead of overwriting newer
-intent. These append-only decisions also work on completed, unarchived Tasks
-without reopening, rewriting acceptance, or rerunning work.
-
-Unchanged Projects need no exclusion. Required changed/unknown Projects still
-need exact verified delivery. Excluded Projects retain their binding, code
-changes and Publication facts and are not counted as merged or verified merely
-because they are excluded. Scope is not execution, write, archive or deletion
-authority: open commitments and unknown execution still need settlement, and
-unique commits or dirty resources remain protected. With separate archive
-authorization, use ordinary archive/preflight and report retained resources;
-do not recommend force merely because an authorized Project is excluded.
-
-For external delivery, choose any available authorized CLI/API and follow
-[Publication evidence](../yui-runtime/references/publication.md). Record exact
-tool observations through `publication verify --observation`; gh/glab are optional
-adapters, not mandatory trusted brands. Do not repeat a sufficient exact
-verification merely to use a different tool. Tool inspection is your judgment;
-candidate acceptance, immutable evidence and ordinary archive safety still apply.
-
-- **Planning:** save meaningful revisions, summarize the discussion and await
-  feedback. This is a complete turn; it needs no WorkItem, Review, InputRequest
-  or Task completion just to stop.
-- **Active delivery:** apply the result-disposition and close instructions in
-  [Active execution](references/execution.md). Continue within authorization,
-  complete only when acceptance is satisfied, or wait for a real durable
-  event. Ask only for a genuinely missing user choice, authority or external
-  fact; routine engineering coordination is not a user decision.
-- **Queries and terminal stages:** answer the bounded question and stop
-  without starting work or recording fictitious lifecycle progress.
-
-Every explicitly dispatched managed AgentRun returns one truthful original
-final report under Runtime's contract, including a planning Run whose result
-is a saved proposal awaiting feedback. Ordinary chat and notifications are
-not implicit assignments and need no separate execution report. Run terminal,
-result acceptance, Task completion, remote delivery and user-authorized
-archive are distinct facts. Do not poll managed Roles or emit unchanged
-waiting Messages; future durable events supply the next notification.
+Follow [Runtime](../yui-runtime/SKILL.md)'s core first. This is the Leader core;
+the linked manuals are not an execution-stage initialization bundle. Use Project
+Skills/Policy/Knowledge for project methods. Read current Task context, actual
+requirements, current Brief and relevant original Messages, not summaries.
+Use exact wake/Run/Message pointers directly; discover only missing evidence.
+
+## Match intent, lifecycle and Session authority
+
+Draft existence, planning history, active lifecycle, execution gate and actual
+Session delivery authority are different facts. Active status/Brief cannot
+promote an old planning Session or override an independent stop.
+
+- Creation-only input stops at a recorded Draft. Authorized Draft discussion or
+  planning-only work uses [Planning](references/planning.md); discuss and preserve
+  permitted revisions, await feedback, and do not implement before requested.
+- Queries, analysis and discussion remain read-only even on active Tasks unless
+  changes are requested. They need no dispatch, Review or lifecycle progress.
+- Active, execution-enabled work with delivery authority and implementation
+  intent advances the existing bounded requirement. A valid activation handoff
+  needs no second “continue.” This route does **not** require reading a full
+  dispatch/review/integration handbook.
+- A completed, unarchived result may reopen for an authorized same-result
+  follow-up or an explicitly identified unfulfilled commitment: first read
+  [Reopen](../yui-runtime/references/completed-followup.md), inspect the result
+  and reassess authority/gates. No mandatory Operator handoff or extra “reopen”
+  confirmation is needed. Tracked diagnosis, evidence handoff or necessary
+  coordination may need reopening, but diagnosis remains read-only unless
+  changes are requested. A simple factual query needs no reopening; a
+  lifecycle-only notice is not the follow-up request.
+- Cancelled/retired, archived, independently stopped or unauthorized execution
+  stays stopped. Queries, notifications, publication recording and Session
+  recovery do not revive it. Report the exact missing legal action.
+
+When intent appears contradictory, read
+[Current intent](../yui-runtime/references/context-ordering.md). Normal authorized
+stage progression supersedes only the old stage, not independent restrictions.
+
+## Advance the coherent result
+
+Honor explicit user delegation/review requirements. Otherwise direct Leader
+implementation is viable for one coupled result; no placeholder WorkItem,
+self-dispatch or new coordination phase is needed. Read full requirements and
+Project methods, trace the existing responsibility, implement in the legal Task
+workspace, verify proportionately, inspect the full final diff and commit cleanly.
+Do not copy a Task into a WorkItem merely to report progress.
+
+Make routine implementation, allocation and recoverable-runtime choices yourself.
+Use Task InputRequest only for a genuine user-owned choice, new authority or
+unavailable external fact, not progress reporting. Keep implementation and
+follow-up delivery in this Task's authorized workspace; local completion is not
+remote delivery when the requirement includes it.
+
+Leader owns coordination, result disposition, acceptance and integration.
+Worker/Reviewer reports are evidence: read exact originals and decide their
+disposition before ACK. Runtime activity is no global lock; use exact bindings.
+Do not borrow Global input controls or impersonate Operator to notify the user.
+Decisions record the actual decision, reason and boundaries, not the proposal;
+recommendations must not be labeled user decisions.
+
+## Read manuals at the action boundary
+
+Before dispatch, Review, Integration or completion, inspect current exact
+Run/Review bindings as decision facts, never a global activity lock.
+
+- Implementation investigation/material verification design beyond current
+  Project methods: [Implementation](references/implementation.md).
+- Choosing/changing work boundaries, executors or concurrency; preparing managed
+  dispatch: [Allocation](references/work-allocation.md). Sending continuation or
+  controlling an existing Assignment: [Assignment execution](references/assignment-execution.md).
+- Requesting independent Review or disposing of a managed result:
+  [Review and disposition](references/review.md).
+  For a returned result needing original/context expansion, include Runtime's
+  [Context reads](../yui-runtime/references/context-reads.md) in that same read
+  group. Add [Working context](references/task-progress.md) when updating the
+  Brief; a normal result notice alone does not require recovery manuals.
+- Integrating isolated results, upstream changes, conflicts or an Integration
+  notification: [Integration](references/integration.md).
+- Accepting a Candidate or attempting Task completion:
+  [Completion](references/completion.md). Inspect current exact Run/Review
+  bindings, resolve required checks/contracts and handled input, and correct
+  obsolete Brief blockers before completion. Chat alone does not complete a Task.
+- Material Brief/Milestone/Knowledge updates: [Working context](references/task-progress.md).
+  Saving/citing substantial Task documents: [Artifacts](references/artifacts.md).
+  Unchanged routine continuation needs no record or full manual reload.
+- Replicated attempts: [Replication](references/replicated-execution.md).
+  Task-local plugins: [Plugins](references/task-plugins.md).
+  Executing existing source-authorized capabilities: [Authorization](references/authorization.md).
+- Changing required delivery Projects: [Delivery scope](references/delivery-scope.md).
+  External PR/MR evidence: [Publication](../yui-runtime/references/publication.md).
+  Actual failure intervention/retry/replacement: Runtime's action index.
+
+At a normal turn boundary, continue useful authorized work, complete only when
+its outcome/obligations are satisfied, ask a genuine missing decision, or wait
+for a real durable event. Persist meaningful changed focus/commitments, not
+unchanged chronology. Do not poll managed Roles or manufacture another Run.
+An explicit dispatched Run returns its truthful final report; ordinary chat
+and notifications need no separate execution report.

package/skills/yui-leader/VERSION

@@ -1 +1 @@
-7
+16

package/skills/yui-leader/references/artifacts.md

@@ -0,0 +1,41 @@
+# Task artifact operations
+
+These operations apply to planning and delivery results. The current Leader
+adopts contributions from other Agents' authorized workspaces; do not share
+an uncoordinated Git index or edit Yui storage directly.
+
+```sh
+yui task artifact list <task-id>
+yui task artifact save <task-id> plans/design.md "<UTF-8 content>" --message "Revise design"
+yui task artifact read <task-id> plans/design.md
+yui task artifact read <task-id> plans/design.md <full-commit>
+```
+
+`save` writes and locally commits exactly one relative path, creating the
+repository on first save. The `artifact.save` capability takes `taskId`,
+`relativePath`, `content`, optional `message` and optional `expectedHead`;
+the CLI equivalent is `--expected-head <commit>`. A head conflict leaves
+the save unapplied: re-read and reconcile the intended update, not overwrite
+blindly. The text interface accepts UTF-8 files up to 8 MiB, not arbitrary
+binary payloads or an invented directory-upload command.
+
+Current reads use Task and relative path at HEAD. Frozen Candidate, Review,
+Context and completion evidence use the returned `taskId + commit +
+relativePath`; do not replace a pinned read with a current read. For same-Task
+string reference lists, use `git:<full-commit>:<relativePath>`, including
+`task complete ... --artifact-ref "git:<full-commit>:plans/design.md"`.
+The commit must be the artifact repository's returned commit, not the Project
+code head. Preserve fixed old references when updating the current document.
+
+After a meaningful save, update the Brief's summary/reference; keep Decisions
+to the choice, reason and boundary. No document-body, HEAD or timestamp mirror
+is needed. The artifact repository has no remote or transport; this does not
+prohibit a Project code repository's legitimate remote. Saving in Draft grants
+no delivery authority; Session handoff and archive preserve the artifacts.
+Stored scripts/HTML remain data and are never implicitly executed or previewed.
+
+These commands describe this source version's interface. A managed Session
+still uses its authorized CLI and actual available capabilities. If an older
+installation lacks them, report that version boundary; do not switch it to
+an unapproved checkout CLI, migrate a shared Home, or upgrade the installation
+merely to save a result.