@fitlab-ai/agent-infra 0.9.11 → 0.9.13

package/templates/.agents/rules/issue-pr-commands.en.md

@@ -12,36 +12,23 @@ agent-infra-internal platform-pr inspect {task-id}
 
 ## Create or recover
 
-Write the title and body to files, then run:
-
 ```bash
 agent-infra-internal platform-pr create {task-id} \
   --agent {standard-agent-token} --base {target-branch} --head {current-branch} \
   --title-file {title-file} --body-file {body-file}
 ```
 
-The core performs exact upstream/head/base lookup first. One match is reused, zero creates, and multiple matches fail closed. Unknown create outcomes are reconciled by the same exact identity rather than retried blindly. `--dry-run` returns only the plan.
+The core performs exact upstream/head/base lookup. One match is reused, zero creates, and multiple matches fail closed. Unknown create outcomes are reconciled by exact identity. `--dry-run` returns only the plan.
 
-## Bind
+## Bind and synchronize
 
 ```bash
 agent-infra-internal platform-pr bind {task-id} --pr {pr-number} --agent {standard-agent-token}
-```
-
-Conflicting bindings fail without changing `task.md`.
 
-## Synchronize metadata
-
-```bash
 agent-infra-internal platform-pr sync {task-id} \
   --agent {standard-agent-token} --metadata --closing-issue
 ```
 
-The core copies type / `in:` labels, assignee, a specific milestone, and the closing association from the linked Issue. Permission-bound operations report `skipped` under top-level `degraded`; the Issue is never updated in reverse.
-
-## Status contract
+Conflicting bindings do not change `task.md`. The shared core computes one `in:` target from task-bound diff/PR files and the repository mapping. A unique closing Issue is converged before the PR; other metadata comes from the linked Issue. Zero or multiple closing Issues update only the PR and return `degraded`; `in:` labels are never copied back from the Issue or removed from unrelated resources. Permission-bound items degrade independently, partial side effects return `blocked` with `IN_LABEL_SYNC_PARTIAL`, and an already-created PR is never rolled back.
 
-- `planned|applied|no-op|degraded`: exit 0.
-- `failed`: exit 1.
-- `blocked`: exit 2 for authentication, network, or unknown outcome.
-- Metadata or summary failures never roll back an already-created PR.
+Statuses `planned|applied|no-op|degraded` exit 0, `failed` exits 1, and `blocked` exits 2.

package/templates/.agents/rules/issue-pr-commands.zh-CN.md

@@ -39,7 +39,7 @@ agent-infra-internal platform-pr sync {task-id} \
   --agent {standard-agent-token} --metadata --closing-issue
 ```
 
-core 从关联 Issue 复制 type / `in:` labels、assignee、具体 milestone，并确保 closing association。权限不足返回逐项 `skipped` 和顶层 `degraded`；不得反向修改 Issue。
+core 根据 task-bound diff/PR files 与仓库映射计算唯一 `in:` target；唯一 closing Issue 按 Issue→PR 顺序收敛，其他 metadata 从关联 Issue 复制。零或多个 closing Issue 只更新 PR 并返回 `degraded`；不得从 Issue 反向复制 `in:`，也不清除非 `in:` labels。权限不足逐项返回 `skipped`，部分副作用返回 `blocked` 与 `IN_LABEL_SYNC_PARTIAL`。
 
 ## 状态语义
 

package/templates/.agents/rules/issue-sync.en.md

@@ -1,8 +1,8 @@
 # Issue Sync
 
-## Marker Registry
+> `--agent` values are defined in `.agents/rules/task-management.md` under “Collaborator Token Specification”.
 
-These hidden markers are the canonical registry for Issue synchronization:
+## Marker Registry
 
 | Key | Marker |
 |---|---|
@@ -12,8 +12,31 @@ These hidden markers are the canonical registry for Issue synchronization:
 | `summary` | `<!-- sync-issue:{task-id}:summary -->` |
 | `cancel` | `<!-- sync-issue:{task-id}:cancel -->` |
 
-Callers should refer to the marker key in skill prose and keep concrete marker strings in this rule or the platform adapter defaults.
+`pr-review` content is synced only as an Issue artifact comment (via the `artifact` / `artifactChunk` markers) and is never a `restore-task` recovery source; restore still accepts only an Issue number and reads only registered Issue markers, with no PR source.
+
+Comments use `platform-comment`; Issue resources use `platform-issue`:
+
+```bash
+agent-infra-internal platform-issue inspect {task-id}
+agent-infra-internal platform-issue create {task-id} --agent {standard-agent-token}
+agent-infra-internal platform-issue bind {task-id} --issue {number} --agent {standard-agent-token}
+agent-infra-internal platform-issue sync {task-id} --agent {standard-agent-token} {desired-state-flags}
+```
+
+The core owns status/in labels, assignees, milestones, Issue Type, pinned fields, requirements, state, capabilities, dry-run, retries, errors, and idempotency. Omitted flags preserve values; `none` explicitly clears them. Status labels converge to at most one, and ambiguous requirement identity fails closed.
+
+PR event `in:` synchronization uses `agent-infra-internal platform-pr sync-in-labels --pr <N> [--cwd <path>]`. PR files are the event evidence; a unique closing Issue is written and re-read before the PR. Zero or multiple closing Issues update only the PR and return `degraded`; unknown side effects or failed convergence return `blocked` with `IN_LABEL_SYNC_PARTIAL`.
+
+`planned|applied|no-op|degraded` exit 0; `failed` exits 1; `blocked` exits 2.
+
+Map material degraded/failed/blocked results to workflow warnings through the structured intent; callers must not edit warning rows directly:
+
+```bash
+agent-infra-internal task-warning {task-id} add \
+  --step issue-sync --severity {severity} --code {code} \
+  --target {target} --message {message} --action {action}
+```
 
-This code platform does not provide an Issue adapter. `platform-issue` returns structured no-op/degraded operations.
+## Backfill
 
-Issue metadata, labels, milestones, assignees, and comments are skipped for custom platforms unless you provide matching `.{platform}.en.md` rule templates and platform adapters. Continue writing local task artifacts normally.
+During completion, artifact comments are replayed with `--kind artifact --artifact <file> --backfill`. If the remote Issue has no matching marker, the core creates a comment with a historical-backfill notice. If a valid base or chunk marker set already exists, the core returns `no-op` and preserves the existing bodies and chunks. Marker conflicts still fail without writes.

package/templates/.agents/rules/issue-sync.zh-CN.md

@@ -1,8 +1,10 @@
-# Issue 同步
+# Issue 同步规则
+
+> `--agent` 取值见 `.agents/rules/task-management.md`「合作者 token 规范」。
 
 ## Marker 注册表
 
-以下隐藏标记是 Issue 同步的唯一权威注册表：
+以下 key 及格式是 Issue 评论身份的兼容契约；调用方只传 key/资源，不自行拼 marker。
 
 | Key | Marker |
 |---|---|
@@ -12,8 +14,65 @@
 | `summary` | `<!-- sync-issue:{task-id}:summary -->` |
 | `cancel` | `<!-- sync-issue:{task-id}:cancel -->` |
 
-Skill 正文应引用 marker key，具体 marker 字符串只保留在本规则或平台适配器默认值中。
+`prSummary` 属于 `.agents/rules/pr-sync.md`，本规则不实现 PR 聚合。
+
+`pr-review` 原文只通过 `artifact` / `artifactChunk` marker 同步为 Issue artifact 评论，不作为 `restore-task` 恢复来源；恢复仍只接受 Issue 编号并只读注册的 Issue marker，不新增 PR 来源。
+
+## 平台 intent
+
+平台 upstream、认证、capability、分页、marker 查找、幂等 create/update、分片、重试与错误分类由 typed platform core 统一处理：
+
+```bash
+agent-infra-internal platform-context resolve [--cwd <path>]
+agent-infra-internal platform-comment list --issue <N> [--cwd <path>]
+agent-infra-internal platform-comment owner <task-ref>
+agent-infra-internal platform-comment sync <task-ref> \
+  --kind task|artifact|summary|cancel --agent {standard-agent-token} \
+  [--artifact <canonical.md>] [--body-file <path|->] [--backfill]
+```
+
+- `applied|no-op|degraded` → exit 0；`failed` → exit 1；`blocked` → exit 2。
+- task 评论保持 `<details>` frontmatter 可逆格式；artifact 原文内联并在超过 profile 上限时使用 `artifactChunk`。
+- 相同 intent 重放必须收敛为 `no-op`；重复 marker 返回 `COMMENT_MARKER_CONFLICT` 且不写入。
+- 外部贡献者锁定统一使用 `platform-comment owner`；不同作者且无 triage 时返回 `COMMENT_OWNER_CONFLICT`。
+
+## 降级与告警
+
+平台结果不直接写 task.md。调用方在有关联任务时把关键失败映射为 workflow warning：
+
+- capability 不足：`IMPORTANT / PERMISSION_DEGRADED`
+- 评论同步永久失败：`ACTION_REQUIRED / COMMENT_SYNC_FAILED`
+- 网络重试耗尽：`ACTION_REQUIRED / NETWORK_RETRY_EXHAUSTED`
+
+使用 `agent-infra-internal task-warning {task-id} add ...` 登记；不得手写 warning 表。
+
+```bash
+agent-infra-internal task-warning {task-id} add \
+  --step issue-sync --severity {severity} --code {code} \
+  --target {target} --message {message} --action {action}
+```
+
+## Issue 元数据 intent
+
+所有 Issue 元数据写入统一使用：
+
+```bash
+agent-infra-internal platform-issue inspect {task-id}
+agent-infra-internal platform-issue create {task-id} --agent {standard-agent-token}
+agent-infra-internal platform-issue bind {task-id} --issue {number} --agent {standard-agent-token}
+agent-infra-internal platform-issue sync {task-id} --agent {standard-agent-token} {desired-state-flags}
+```
+
+`sync` 支持 status/in labels、assignee、milestone、Issue Type、pinned fields、需求复选框与 Issue state。省略 flag 表示 preserve，`none` 表示显式清空。适配层统一处理集合差集、动态 schema、权限降级、dry-run、重试、错误分类与幂等重放；SKILL 不得拼装平台 CLI、GraphQL 或权限分支。
+
+- `planned|applied|no-op|degraded` → exit 0
+- `failed` → exit 1
+- `blocked` → exit 2
+
+status labels 始终至多一个；关闭后不保留 status label。`in:` labels 只从项目允许映射与仓库实际 labels 交集产生。需求复选框以 task.md 原文为身份，歧义时 fail closed。
+
+PR event 的 `in:` 同步使用 `agent-infra-internal platform-pr sync-in-labels --pr <N> [--cwd <path>]`。PR files 是事件证据；唯一 closing Issue 按 Issue→PR 顺序写入并复读，零或多个 closing Issue 只写 PR 并返回 `degraded`。写入后副作用不明或复读未收敛必须返回 `blocked` 与 `IN_LABEL_SYNC_PARTIAL`。
 
-当前代码平台未提供 Issue 适配器；`platform-issue` 返回结构化 no-op/degraded operations。
+## 补发
 
-自定义平台会跳过 Issue 元数据、标签、里程碑、负责人和评论同步，除非你提供匹配的 `.{platform}.zh-CN.md` 规则模板和平台适配器。请继续正常写入本地任务产物。
+complete-task 按 artifact catalog 时间线遍历本地已有产物：task 使用 `--kind task`，其余使用 `--kind artifact --artifact <file> --backfill`；最后用 `--kind summary --body-file <path>` 原地同步交付摘要。artifact backfill 在远端缺少 marker 时创建带“历史产物补发”提示的评论；已存在合法 marker 集合时保持原评论和分片不变并返回 `no-op`；marker 冲突仍失败。该判定由 core 完成，不由 Skill 扫描评论或拼标题。

package/templates/.agents/rules/label-milestone-setup.en.md

@@ -1,5 +1,18 @@
 # Label and Milestone Setup
 
-This code platform does not provide built-in label or milestone setup.
+Read this rule before initializing labels or milestones, or before changing milestone metadata during release work.
 
-Initialization of remote labels and milestones is skipped for custom platforms unless you provide matching `.{platform}.en.md` rule templates and scripts. Local task workflow files remain usable without remote metadata setup.
+## Entry points
+
+Use the shared scripts rather than composing provider commands in a skill:
+
+```text
+bash .agents/skills/init-labels/scripts/init-labels.sh [--cleanup-stale-in]
+bash .agents/skills/init-milestones/scripts/init-milestones.sh "$ARGUMENTS"
+```
+
+`init-labels.sh` reads the repository's `labels.in` configuration, creates or updates the declared `in:` labels, and preserves unrelated labels. `--cleanup-stale-in` is destructive and may only be passed after explicit confirmation; it removes stale `in:` labels that are no longer declared.
+
+`init-milestones.sh` owns milestone inference and the provider-specific write path. Callers must pass the requested arguments through unchanged and must not rebuild milestone discovery, ordering, branch ancestry, or write logic.
+
+Both scripts return a stable status. A successful no-op is still success; a degraded result must be reported without claiming that remote metadata was changed. Failed or cancelled operations must not be reported as completed.

package/templates/.agents/rules/label-milestone-setup.zh-CN.md

@@ -1,5 +1,18 @@
 # Label 和 Milestone 初始化
 
-当前代码平台未内置标签或里程碑初始化支持。
+在初始化 label、milestone，或在发布工作中修改 milestone 元数据前，先阅读本规则。
 
-自定义平台会跳过远端标签和里程碑初始化，除非你提供匹配的 `.{platform}.zh-CN.md` 规则模板和脚本。即使没有远端元数据初始化，本地任务工作流文件仍可使用。
+## 入口
+
+SKILL 应调用共享脚本，不得自行拼装 provider 命令：
+
+```text
+bash .agents/skills/init-labels/scripts/init-labels.sh [--cleanup-stale-in]
+bash .agents/skills/init-milestones/scripts/init-milestones.sh "$ARGUMENTS"
+```
+
+`init-labels.sh` 读取仓库的 `labels.in` 配置，创建或更新声明的 `in:` label，并保留无关 label。`--cleanup-stale-in` 具有破坏性，只有在得到明确确认后才能传入；它会删除不再声明的过期 `in:` label。
+
+`init-milestones.sh` 负责 milestone 推断和 provider 写入路径。调用方应原样传递请求参数，不得重新实现 milestone 发现、排序、分支祖先判断或写入逻辑。
+
+两个脚本都返回稳定状态。成功的 no-op 仍算成功；degraded 结果必须如实报告，不能声称远端元数据已变更。失败或取消的操作不得报告为已完成。

package/templates/.agents/rules/local-artifact-repair.en.md

@@ -0,0 +1,53 @@
+# General Rule - Model-Driven Local Artifact Repair
+
+This rule applies when `analyze-task`, `plan-task`, or `code-task` handles **the same controlled local artifact** before its completed event, and when `review-analysis`, `review-plan`, or `review-code` handles a finalizer failure for **the same controlled review artifact**. It does not apply to `task.md`, the ledger, receipts, source code, Git, platform resources, or lifecycle state.
+
+## Pre-completion gate for analysis, plan, and code artifacts
+
+- `analyze-task`, `plan-task`, and `code-task` must call `task-artifact ... finalize-local` before publishing a completed event; pass only that call's `artifactSha256` and `semanticDigest` to the completed event.
+- `finalize-local` does not modify the artifact or task state, but it writes a one-shot local provenance intent in the repository workspace. After `failed`, the model may make one minimal edit in the same artifact only when `repairable=true` and the diagnostic explicitly describes a one-line replacement, then rerun the same call completely; count each byte-changing edit, up to 8.
+- The first repairable failure's semantic digest is retained in that intent; a later `passed` result must match the baseline, and the completed event must verify and atomically transition the same intent to `consumed` before writing `task.md` under the task lock. A consumption failure must not write the task; the consumed intent is durable and retryable, so no failure-prone post-write deletion is attempted. Do not replace a failed baseline with a newly computed digest or publish a completed event without the finalizer.
+- After `failed`, no progress, or a repeated diagnostic or fingerprint, do not publish a completed event. After `passed`, do not rescan or manually write summary data.
+
+## Authorization Boundary
+
+- The finalizer only reads facts, validates state, normalizes successful results, and performs atomic writes. It does not maintain a recoverable-error allowlist, infer repairability, or choose report content to delete.
+- The initial finalizer call is not counted as a repair attempt. Editing is allowed only after the mechanical safety gates pass and the model explicitly determines that the current problem can be solved by a minimal, explainable artifact change.
+- The model may edit only the one ordinary local artifact declared by the current skill, and the file must be inside the current task directory. It must not edit `task.md`, the review disagreement ledger, receipts, source code, other reports, or remote resources.
+- `changed=false`, an error code, or a format shape is diagnostic evidence, not automatic authorization. The model must judge each case using the complete diagnostic, artifact content, and context.
+
+## Non-bypassable Mechanical Safety Gates
+
+Before every model edit, confirm that:
+
+1. the finalizer returned a failure and no artifact operation was committed;
+2. the task, stage, artifact, review round, identity, and provenance still match;
+3. the target is an ordinary file inside the current task directory, with no concurrency conflict, permission error, I/O uncertainty, or target replacement;
+4. the change does not alter human-decision semantics, decision-detail or ledger identity, and does not involve task state, receipts, Git, platform, or any other external side effect. A known and verifiable pending human decision may remain when the model proves the edit is unrelated to that decision.
+
+If any condition fails, stop immediately without invoking a model edit. The model cannot bypass these gates.
+
+## Dynamic Convergence Loop
+
+1. Run the initial finalizer with the fixed task, stage, artifact, and orchestrated intent.
+2. On success, use that complete result for the existing completion gates; do not rescan the ledger or write summary fields manually.
+3. On failure with the safety gates passed, let the model read the structured diagnostic, current artifact, and required context, then decide whether to repair or stop. It must explain the scope and reason for the proposed change.
+4. If the model continues, perform one minimal edit to the artifact. Increment `repairAttempts` only when the file bytes actually change, then rerun the same finalizer intent completely.
+5. Re-run every safety check after each retry. A new, independent problem still limited to the same artifact may be offered to the model for another decision; the previous repairability decision must not be reused automatically.
+6. Stop immediately when the model cannot establish safety or identifies an environment, permission, concurrency, identity, provenance, unknown-state, uncertain human-decision semantics/details, or other non-local problem. A known human decision is not itself a stop condition.
+7. Stop when the diagnostic or artifact fingerprint repeats, no byte-level progress occurs, or the model cannot propose a verifiable minimal change.
+8. Allow at most 8 actual artifact edits per skill invocation as an emergency circuit breaker. This cap only prevents infinite loops and resource exhaustion; it is not a normal business stop condition and does not mean that at most eight problems may be repaired. Preserve the final structured diagnostic when the cap is reached.
+
+The repair count exists only in the in-memory context of the current skill invocation. Do not write it to `task.md`, the ledger, or a public receipt. Do not create different normal budgets by error code, skill, or problem type.
+
+## Completion Events and User Output
+
+- After the final complete finalizer result succeeds, publish the review completed event using that same result's provenance, ledger state, verdict, and counts. Generate cross-stage next-step commands only when `stageStatus.canAdvance=true` and the verdict is Approved; when `canAdvance=false`, still record the result and route to same-stage revision/review.
+- On failure, model stop, lack of progress, repeated diagnostics, or the emergency cap, do not publish a completed event, advance the lifecycle, generate cross-stage commands, or fabricate an approval.
+- Stopping advancement does not discard review results. User output must show the artifact path, the last safely readable summary/findings, actual `repairAttempts`, the last structured diagnostic, and the stop reason.
+- If the summary cannot be safely parsed, show only the artifact path and the raw structured diagnostic; do not infer counts or add a conclusion, and state that lifecycle advancement stopped while the artifact remains available.
+- A stop path may only suggest rerunning the current review skill or handling the issue manually. It must not invoke cross-stage helpers such as `code-task` or `complete-task`.
+
+## Lifecycle Isolation
+
+`complete-task`, alert closing, task restoration, and other task-lifecycle/task-finalization entry points involve task state, receipts, remote results, or archival side effects and do not consume this rule. They retain their own hard-stop, same-intent retry, and unknown-state semantics.

package/templates/.agents/rules/local-artifact-repair.zh-CN.md

@@ -0,0 +1,53 @@
+# 通用规则 - 模型驱动的本地产物修复
+
+本规则适用于 `analyze-task`、`plan-task`、`code-task` 在 completed 事件前处理**同一个受控本地产物**，以及 `review-analysis`、`review-plan`、`review-code` 在 finalizer 失败后处理**同一个受控 review artifact** 的场景。它不适用于 task.md、账本、receipt、源码、Git、平台资源或生命周期状态。
+
+## 分析、方案和代码产物完成前门禁
+
+- `analyze-task`、`plan-task` 和 `code-task` 必须在发布 completed 事件前调用 `task-artifact ... finalize-local`；只有同一次返回的 `artifactSha256` 和 `semanticDigest` 才能传给 completed 事件。
+- `finalize-local` 不修改产物或任务状态，但会在仓库工作区写入一次性的本地 provenance intent。返回 `failed` 时，只有 `repairable=true` 且诊断明确为单行替换，模型才可在同一产物中执行一次最小编辑，然后完整重跑同一调用；每次实际字节修改计一次，最多 8 次。
+- 首次可修复失败的 semantic digest 会保存在该 intent 中；后续 `passed` 必须匹配该基线，completed event 还必须在任务锁内验证并在写 task.md 前原子转换为 `consumed` 状态。消费失败时不得写任务；转换后的 intent 是可重试的 durable 记录，成功写入后不再执行可能失败的后置删除。不得用重新计算的新摘要替换失败基线，也不得绕过 finalizer 直接发布 completed event。
+- 返回 `failed`、无进展、诊断或指纹重复时不得发布 completed 事件；返回 `passed` 后也不得重新扫描或手工补写摘要。
+
+## 授权边界
+
+- finalizer 只负责读取事实、校验状态、规范化成功结果和原子写入；它不维护可修复错误码白名单，不推断 repairability，也不自动选择要删除的报告内容。
+- 首次 finalizer 调用不计入修复次数。只有在机械安全门通过后，模型明确判断当前问题能通过最小、可解释的 artifact 修改解决时，才允许编辑。
+- 模型只能修改当前 skill 声明的一个普通本地产物，且该文件必须位于当前 task 目录。不得修改 task.md、审查分歧账本、receipt、源码、其他报告或远端资源。
+- `changed=false`、错误码或某个格式形状只是诊断事实，不是自动批准。模型必须结合完整诊断、artifact 内容和上下文逐例判断。
+
+## 不可绕过的机械安全门
+
+在每次模型编辑前确认：
+
+1. finalizer 返回失败，且没有已提交的 artifact operation；
+2. 任务、stage、artifact、review round、身份和 provenance 仍然一致；
+3. 目标是当前 task 目录内的普通文件，没有并发冲突、权限错误、I/O 不确定性或目标被替换；
+4. 不改变人工裁决语义、裁决详情或账本身份，也不涉及任务状态、receipt、Git、平台或其他外部副作用；已知且可核验的 pending 人工裁决可以保留，但模型必须证明本次修改与该裁决无关。
+
+任一条件不满足，立即停止，不调用模型编辑。安全门不能被模型的语义判断绕过。
+
+## 动态收敛循环
+
+1. 使用固定的 task、stage、artifact 和 orchestrated intent 执行首次 finalizer。
+2. 成功时使用这一次完整结果进入既有完成前门禁；不得重新扫描账本或手工补写摘要。
+3. 失败且安全门通过时，模型读取结构化诊断、当前 artifact 和必要上下文，决定继续修复或停止，并说明修改范围和理由。
+4. 模型选择继续时，只做一次最小 artifact 编辑。只有文件字节实际发生变化，才增加一次 `repairAttempts`，然后完整重跑同一个 finalizer intent。
+5. 每次重跑都重新执行全部安全校验。修复后出现新的、独立且仍局限于该 artifact 的问题，可以再次交给模型判断；不能沿用上一次“可修复”的结论。
+6. 模型无法证明下一步安全，或判断问题属于环境、权限、并发、身份、provenance、状态未知、人工裁决语义/详情不确定或其他非本地问题时，立即停止。已知人工裁决本身不是停止条件。
+7. 诊断重复、artifact 指纹重复、没有实际字节进展或模型没有提出可验证的最小修改时，立即停止。
+8. 每次 skill invocation 最多允许 8 次实际 artifact 修改作为紧急熔断。该上限只防止死循环和资源失控，不是正常业务停止条件，也不表示最多只能修复 8 个问题。达到上限时保留最后结构化诊断并停止。
+
+修复次数只存在于当前 skill invocation 的内存上下文，不写入 task.md、ledger 或公共 receipt。不得按错误码、技能或问题类型设置不同的正常预算。
+
+## 完成事件与用户输出
+
+- 最终一次完整 finalizer 成功后，必须使用同一次返回的 provenance、账本状态、verdict 和计数发布 review completed 事件；`stageStatus.canAdvance=true` 且结论为 Approved 时才可生成跨阶段 next-step 命令，`canAdvance=false` 时仍须登记结果并路由到同阶段修订/复审。
+- 失败、模型停止、无进展、诊断重复或紧急熔断时，不发布 completed 事件，不推进生命周期，不生成跨阶段命令，也不伪造通过结论。
+- 停止推进不等于丢弃审查结果。用户输出必须展示 artifact 路径、已安全读取的最后有效 summary/findings、实际 `repairAttempts`、最后结构化诊断和停止原因。
+- summary 无法安全解析时，只展示 artifact 路径和原始结构化诊断，不推算计数、不补写结论，并明确生命周期推进已停止但产物仍可查看。
+- 停止路径只提示重新运行当前 review skill 或进行人工处理；不能调用 `code-task`、`complete-task` 等跨阶段 helper。
+
+## 生命周期隔离
+
+`complete-task`、告警关闭、恢复任务和其他 task-lifecycle/task-finalization 入口涉及任务状态、receipt、远端结果或归档副作用，不消费本规则。它们继续使用各自的硬失败、同一 intent 重试和状态未知语义。

package/templates/.agents/rules/milestone-inference.en.md

@@ -1,5 +1,15 @@
 # Milestone Inference
 
-This code platform does not provide milestone inference; `platform-issue sync --milestone ...` returns a structured no-op/degraded result.
+> `--agent` values are defined in `.agents/rules/task-management.md` under “Collaborator Token Specification”.
 
-Milestone narrowing and reuse are skipped for custom platforms unless you provide matching `.{platform}.en.md` rule templates. Do not block task progress when no platform-specific milestone implementation is available.
+Milestones narrow through declarative Issue intents:
+
+```bash
+# create/import phase
+agent-infra-internal platform-issue sync {task-id} --agent {standard-agent-token} --milestone initial
+
+# code phase
+agent-infra-internal platform-issue sync {task-id} --agent {standard-agent-token} --milestone specific
+```
+
+`initial` prefers a valid task milestone, otherwise the lowest open `X.Y.x`, then `General Backlog`. `specific` narrows a release line to the highest open patch version using main/release-line ancestry. Missing facts or triage permission preserve the remote value and return skipped/degraded. `create-pr` reuses the Issue milestone in the 09/10 PR adapter.

package/templates/.agents/rules/milestone-inference.zh-CN.md

@@ -1,5 +1,27 @@
-# Milestone 推断
+# Milestone 推断规则
 
-当前代码平台未提供里程碑推断；`platform-issue sync --milestone ...` 返回结构化 no-op/degraded 结果。
+> `--agent` 取值见 `.agents/rules/task-management.md`「合作者 token 规范」。
 
-自定义平台会跳过里程碑收窄和复用逻辑，除非你提供匹配的 `.{platform}.zh-CN.md` 规则模板。没有平台专属里程碑实现时，不要阻塞任务推进。
+Milestone 继续按生命周期逐步收窄，但候选读取、分支判断、权限降级和写入统一由 `platform-issue` 处理。
+
+## 阶段 1：创建或导入
+
+```bash
+agent-infra-internal platform-issue sync {task-id} --agent {standard-agent-token} --milestone initial
+```
+
+优先使用 task.md 中存在且有效的显式 milestone；否则从 open milestones 选择最低 `X.Y.x` 版本线，最后回退 `General Backlog`。`triage=false` 或外部事实不足时保持远端值并返回 skipped/degraded。
+
+## 阶段 2：`code-task`
+
+```bash
+agent-infra-internal platform-issue sync {task-id} --agent {standard-agent-token} --milestone specific
+```
+
+版本线收窄规则保持不变：主干模式在当前版本线取最高 patch；多版本分支模式依据 release-line/main 祖先关系选择版本线。无法可靠判断时保持原值。完成门禁仍拒绝 `X.Y.x`。
+
+## 阶段 3：`create-pr`
+
+PR 只复用关联 Issue 的具体 milestone；该 PR 资源写入属于 09/10，不在 `platform-issue` 中实现。
+
+调用方不得重新实现 milestone 列表、排序、分支祖先或平台写入逻辑。

package/templates/.agents/rules/pr-sync.en.md

@@ -2,30 +2,45 @@
 
 > `--agent` values are defined in `.agents/rules/task-management.md` under “Collaborator Token Specification”.
 
-The model writes the semantic reviewer summary. Typed core owns canonical artifact selection, the marker, current HEAD, paginated comment lookup, and create/update/no-op reconciliation.
+The model writes the semantic reviewer summary. Typed core owns the task-intent digest, complete three-dot diff, canonical `pr-change-report.json` sidecar, report rendering, authoritative PR-head wrapping, paginated comment lookup, and reconciliation.
 
-## Inputs
+## Three-Layer Isolation
 
-```bash
-agent-infra-internal platform-pr summary-context {task-id}
-```
+- `sync-pr:{task-id}:summary` (`platform-pr summary-sync`) is an updatable reviewer summary; it is neither a process copy nor a formal Review.
+- The full `pr-review*` process copy lives in the Issue artifact comment (`platform-comment sync --kind artifact`) and is restored by `restore-task` under its Issue-only contract.
+- A formal PR Review (`platform-pr-review publish`) is the only formal conclusion carrier published to the PR, bound to the reviewed head SHA, and includes conclusion, findings, receipt, and the Issue artifact link.
 
-Use only the returned latest canonical `plan*`, `review-plan*`, `code*`, `review-code*`, and `manual-validation*` for lifecycle claims. The summary covers scope, tests, review history, and exactly one manual-validation state. It also includes a `### PR Code Changes` section based on authoritative base/head facts from `platform-pr inspect` and the complete `git diff --find-renames --numstat base...head`, with reconciled runtime, test, skill/rule, template, documentation, and other categories plus rename, mechanical-mirror, and unnecessary-change analysis.
+The three never cross: a regular PR comment does not carry the full process copy; an Issue artifact comment is not treated as a formal Review; a summary is not presented as a formal Review.
 
-## Publish
+```bash
+agent-infra-internal platform-pr summary-context {task-id}
 
-Write only the summary body to a file, without marker or SHA:
+agent-infra-internal platform-pr change-report {task-id} \
+  --agent {standard-agent-token} --mechanical-file {mechanical-report-file} \
+  --precheck-file {precheck-candidate-file}
 
-```bash
 agent-infra-internal platform-pr summary-sync {task-id} \
-  --agent {standard-agent-token} --body-file {summary-body-file}
+  --agent {standard-agent-token} --body-file {summary-body-file} \
+  --change-report-file .agents/workspace/active/{task-id}/pr-change-report.json \
+  --result {primary-result}
 ```
 
-The core wraps `<!-- sync-pr:{task-id}:summary -->` and current `<!-- last-commit: ... -->`, then creates, updates in place, or returns no-op. Duplicate markers fail deterministically. Callers never assemble shell/heredoc comment commands.
+Use only canonical inputs returned by the context command. The precheck must contain file evidence for six fixed checks; `formalReview` is always false and `needs-review` routes to `review-code`. Write the plain body with exactly one `<!-- canonical-pr-change-report -->` placeholder, without a report heading/JSON, marker, or SHA. The core validates the task-intent digest, bound PR identity, complete patch SHA, and mechanical totals before rendering the report, then owns `<!-- sync-pr:{task-id}:summary -->`, authoritative `<!-- last-commit: ... -->`, create/update/no-op behavior, and duplicate-marker conflicts. Missing, stale, invalid, or bypassed report input never publishes; summary failure never rolls back a PR or commit.
 
-`create-pr` does not roll back the PR on summary failure; `commit` records a warning without rolling back; manual validation refreshes after its canonical artifact is written.
+The canonical comment envelope produced by the core is:
+
+```markdown
+<!-- sync-pr:{task-id}:summary -->
+<!-- last-commit: {git-head-sha} -->
+## Review Summary
+{manual-validation-section}
+### Key Technical Decisions
+### Review History
+### Test Results
+### PR Code Changes
+```
 
-For a linked task, the caller records a structured warning:
+Canonical inputs include `manual-validation.md`. The manual-validation section is one of `### ✅ Manual Validation Passed`, `### ⚠️ Manual Verification Required`, or `### ✅ No Manual Verification Needed`; `complete-manual-validation` refreshes the same summary in place. The PR code-change section uses authoritative base/head facts from `platform-pr inspect` and the complete `git diff --find-renames --numstat base...head`, reconciles runtime, test, skill/rule, template, documentation, and other categories, and explains renames, mechanical mirrors, and potentially unnecessary changes. The caller supplies the placeholder; the core renderer supplies the canonical `### PR Code Changes` section.
 
 ```bash
 agent-infra-internal task-warning {task-id} add --step {step} --severity WARNING \

package/templates/.agents/rules/pr-sync.zh-CN.md

@@ -2,7 +2,15 @@
 
 > `--agent` 取值见 `.agents/rules/task-management.md`「合作者 token 规范」。
 
-PR 摘要的语义聚合由模型完成；canonical 产物选择、marker、当前 HEAD、分页评论查找和 create/update/no-op 由 typed core 负责。
+PR 摘要的语义聚合由模型完成；任务意图 digest、完整三点 diff、canonical `pr-change-report.json` sidecar、报告渲染、marker、权威 PR head、分页评论查找和 create/update/no-op 由 typed core 负责。
+
+## 三层隔离
+
+- `sync-pr:{task-id}:summary`（`platform-pr summary-sync`）是可更新的 reviewer 摘要，不是过程原文，也不是正式 Review。
+- `pr-review*` 完整过程原文在 Issue artifact 评论（`platform-comment sync --kind artifact`），由 `restore-task` 按 Issue-only 契约恢复。
+- 正式 PR Review（`platform-pr-review publish`）是唯一发布到 PR 的正式结论载体，绑定被审 head SHA，含结论、finding、receipt 与 Issue artifact 链接。
+
+三者不串用：PR 普通评论不承载完整过程原文；Issue artifact 评论不当作正式 Review；摘要不冒充正式 Review。
 
 ## 聚合输入
 
@@ -10,26 +18,52 @@ PR 摘要的语义聚合由模型完成；canonical 产物选择、marker、当
 agent-infra-internal platform-pr summary-context {task-id}
 ```
 
-只使用返回的 canonical latest `plan*`、`review-plan*`、`code*`、`review-code*`、`manual-validation*`。模型据此生成 reviewer 摘要正文，至少包含：
+只使用返回的 canonical latest `plan*`、`review-plan*`、`code*`、`review-code*`、`manual-validation*`（包括 `manual-validation.md`）。模型据此生成 reviewer 摘要正文，至少包含：
 
 - 变更摘要与实现范围
 - 测试结果
 - 审查历程与当前结论
 - 人工校验三态：`### ⚠️ 需人工校验`、`### ✅ 人工验证已通过` 或 `### ✅ 无需人工校验`
-- `### PR 代码增减`：基于 `platform-pr inspect` 的权威 base/head 和完整 `git diff --find-renames --numstat base...head`，按运行时代码、测试、Skill/规则、模板、文档及其他分类核算，并说明 rename、机械镜像与疑似不必要变化
+- 变更摘要、实现范围、测试结果、审查历程和人工校验三态
+- 一个 `<!-- canonical-pr-change-report -->` 占位符；`### PR 代码增减` 由 core renderer 基于权威 diff 生成，调用方不自行写标题或统计
 
 保留人工校验项时，每条写明校验内容、定位和只能人工完成的原因。
 
+core 最终包装出的 canonical 评论结构为：
+
+```markdown
+<!-- sync-pr:{task-id}:summary -->
+<!-- last-commit: {git-head-sha} -->
+## 审查摘要
+{manual-validation-section}
+### 关键技术决策
+### 审查历程
+### 测试结果
+### PR 代码增减
+```
+
+先用机械脚本和模型预检生成输入，再写入任务目录固定 sidecar：
+
+```bash
+agent-infra-internal platform-pr change-report {task-id} \
+  --agent {standard-agent-token} --mechanical-file {mechanical-report-file} \
+  --precheck-file {precheck-candidate-file}
+```
+
+precheck 必须覆盖六项固定检查并提供文件证据；`formalReview` 固定为 `false`，`needs-review` 路由到 `review-code`。sidecar 是当前 head 的可丢弃派生缓存，不是 lifecycle artifact 或 Issue artifact。
+
 ## 发布
 
 把聚合后的纯正文写入 `{summary-body-file}`，不要自行添加 marker 或 commit SHA：
 
 ```bash
 agent-infra-internal platform-pr summary-sync {task-id} \
-  --agent {standard-agent-token} --body-file {summary-body-file}
+  --agent {standard-agent-token} --body-file {summary-body-file} \
+  --change-report-file .agents/workspace/active/{task-id}/pr-change-report.json \
+  --result {primary-result}
 ```
 
-core 会包装唯一 `<!-- sync-pr:{task-id}:summary -->` 和当前 `<!-- last-commit: ... -->`，分页查找 PR 普通评论：不存在则创建、正文变化则原地更新、无差异则 no-op、重复 marker 则稳定失败。正文通过文件输入，调用方不得拼 shell/heredoc。
+正文必须恰好包含一次 `<!-- canonical-pr-change-report -->`，不得自行拼接报告标题/JSON、marker 或 last-commit。core 会重新校验任务意图 digest、绑定 PR identity、完整 patch SHA 和机械统计，由 renderer 生成报告段，再包装唯一 `<!-- sync-pr:{task-id}:summary -->` 和权威 `<!-- last-commit: ... -->`，分页查找 PR 普通评论：不存在则创建、正文变化则原地更新、无差异则 no-op、重复 marker 或报告失配则稳定失败。正文通过文件输入，调用方不得拼 shell/heredoc。
 
 ## 结果回传
 

package/templates/.agents/rules/release-commands.en.md

@@ -1,5 +1,44 @@
-# Release Commands
+# Release Platform Commands
 
-This code platform does not provide a release-note adapter.
+Release-note platform operations use a typed internal intent. Callers consume structured JSON and do not interpret platform commands, raw fields, identity email rules, or authentication errors.
 
-Local `agent-infra-internal platform-release-notes stage --notes-file "{notes-file}"` still normalizes an external file and returns its digest. Use `context` to receive a structured `PLATFORM_RELEASE_NOTES_UNSUPPORTED` no-op and do not probe another platform client. Publish with `--expected-sha256` remains a no-op until an adapter is provided.
+## Collect Release-note Context
+
+```bash
+agent-infra-internal platform-release-notes context \
+  --from-tag "v{prev-version}" \
+  --to-tag "v{version}" \
+  --branch "{branch}" \
+  --history-limit 3
+```
+
+The result contains `history`, `pullRequests`, `closingIssues`, and commit `authors` with normalized `login` / `resolution` values. `status: no-op` with error code `PLATFORM_RELEASE_NOTES_UNSUPPORTED` means the current platform does not support remote release-note operations.
+
+### Identity Safety Boundary
+
+- Render an identity as `@login` only when `resolution` is `platform-user` or `platform-noreply` and `login` is non-empty.
+- Exclude `resolution: unresolved` identities from the release-note contributor list. Never infer a login from a name, email address, domain, brand, or same-named platform account, and never substitute a guessed personal or organization account.
+- Do not publish unresolved identity emails or identity-verification TODOs in Release notes. Record any required investigation outside the release workflow without creating a platform mention.
+
+## Stage Release Notes
+
+```bash
+agent-infra-internal platform-release-notes stage \
+  --notes-file "{notes-file}"
+```
+
+The command accepts only an external regular file, normalizes it atomically, and returns SHA-256 for its exact bytes.
+
+## Publish Release Notes
+
+Call this only after user confirmation:
+
+```bash
+agent-infra-internal platform-release-notes publish \
+  --tag "v{version}" \
+  --title "v{version}" \
+  --notes-file "{notes-file}" \
+  --expected-sha256 "{preview-sha256}"
+```
+
+The command recomputes the digest before platform access and performs no platform write on mismatch. On a match it updates an existing Release or creates it when missing; `--dry-run` returns only the planned operation. Exit code `0` means success, `1` means a stable failure, and `2` means a network or platform block.