@fitlab-ai/agent-infra 0.8.5 → 0.9.1

package/templates/.agents/skills/review-plan/reference/output-templates.en.md

@@ -4,9 +4,9 @@ Read this file before presenting the final review result to the user.
 
 ## Select exactly one output scenario
 
-Evaluate in this order (**manual-validation count does not participate in selection**):
-1. If `Blocker = 0`, `Major = 0`, and `Minor = 0`, use Scenario A, regardless of manual-validation count
-2. If `Blocker = 0` and (`Major > 0` or `Minor > 0`), use Scenario B
+Select from `stage-status` (**manual-validation and advisory counts do not participate**):
+1. If `stageStatus.canAdvance=true`, use Scenario A
+2. If `stageStatus.canAdvance=false` and there are no blockers, use Scenario B
 3. If `Blocker > 0` and the issues can be handled by one focused revision, use Scenario C
 4. If the technical plan needs major redesign, broad rewriting, or a restart, use Scenario D
 
@@ -16,7 +16,9 @@ Rules:
 - If `Blocker > 0`, never use an approved template
 - Never count manual-validation items as blocker / major / minor or use them to trigger Scenario B/C/D
 - The selected scenario must include all TUI command formats
-- The count line always shows 4 numbers: the first three (Blockers / Major / Minor) must be 0 to proceed; the fourth, `Human-decision` (`{h}`), is the number of rows in task.md `## 审查分歧账本` with `stage=plan` and `status=needs-human-decision` — a "pending human ruling" item that need not be zero and does not participate in scenario selection. When `{h} > 0`, before the selected scenario's "Next steps" commands you must expand each pending ruling per the "Pending human-decision pre-block" in `.agents/rules/next-step-output.md` and prompt to resolve them first
+- The count line shows 4 numbers. `Human-decision` (`{h}`) counts this stage's `needs-human-decision` rows; because those rows are unresolved, `{h} > 0` means `canAdvance=false`. Expand the "Pending human-decision pre-block" from `.agents/rules/next-step-output.md` and show revision and re-review paths only.
+
+For Scenarios B/C/D, follow the revision commands with re-review commands: `/review-plan {task-ref}`, `/{{project}}:review-plan {task-ref}`, and `$review-plan {task-ref}`.
 
 ### Scenario A: Approved with no findings
 
@@ -34,23 +36,18 @@ Next step - write code:
 Reminder: manual-validation items belong in the PR description manual verification checklist and should not trigger /plan-task.
 ```
 
-### Scenario B: Approved with findings
+### Scenario B: Changes requested (major / minor)
 
 ```text
-Task {task-id} technical plan review completed. Verdict: approved.
+Task {task-id} technical plan review completed. Verdict: changes requested.
 - Blockers: 0 | Major issues: {n} | Minor issues: {n} | Human-decision: {h}
 - Review report: .agents/workspace/active/{task-id}/{review-artifact}
 
-Next step - revise plan before coding (recommended):
+Next step - revise technical plan:
   - Claude Code / OpenCode: /plan-task {task-ref}
   - Gemini CLI: /{{project}}:plan-task {task-ref}
   - Codex CLI: $plan-task {task-ref}
 
-Or proceed directly to coding:
-  - Claude Code / OpenCode: /code-task {task-ref}
-  - Gemini CLI: /{{project}}:code-task {task-ref}
-  - Codex CLI: $code-task {task-ref}
-
 [When manual-validation > 0, append:]
 Reminder: manual-validation items belong in the PR description manual verification checklist and should not trigger /plan-task.
 ```

package/templates/.agents/skills/review-plan/reference/output-templates.zh-CN.md

@@ -4,9 +4,9 @@
 
 ## 选择唯一输出场景
 
-按以下顺序判断（**注意：manual-validation 数量不参与判断**）：
-1. 如果 `Blocker = 0` 且 `Major = 0` 且 `Minor = 0`，使用场景 A（不管 manual-validation 是否 > 0）
-2. 如果 `Blocker = 0` 且（`Major > 0` 或 `Minor > 0`），使用场景 B
+按 `stage-status` 结果判断（**注意：manual-validation 和 advisory 数量不参与判断**）：
+1. 如果 `stageStatus.canAdvance=true`，使用场景 A
+2. 如果 `stageStatus.canAdvance=false` 且无 blocker，使用场景 B
 3. 如果 `Blocker > 0`，且问题可以通过一次聚焦修复解决，使用场景 C
 4. 如果技术方案需要重大重构、大范围重写或整体重来，使用场景 D
 
@@ -16,7 +16,9 @@
 - 只要 `Blocker > 0`，就绝对不能输出通过模板
 - manual-validation 项绝对不能被计入 blocker / major / minor 计数，也不能用作触发场景 B/C/D 的依据
 - 所选场景中必须包含所有 TUI 命令格式
-- 计数行固定显示 4 个数字：前三项（阻塞 / 主要 / 次要）必须为 0 才进下一步；第四项 `人工裁决`（`{h}`）= task.md `## 审查分歧账本` 中 `stage=plan` 且 `status=needs-human-decision` 的行数，是「待人裁」项、不要求归零，也不参与场景判断。当 `{h} > 0` 时，必须在选定场景的「下一步」命令之前，按 `.agents/rules/next-step-output.md`「人工裁决待办前置块」逐项展开裁决项并提示先完成裁决
+- 计数行固定显示 4 个数字。`人工裁决`（`{h}`）是本阶段 `needs-human-decision` 行数；它属于未闭环账本状态，因此 `{h} > 0` 时 `canAdvance=false`，必须按 `.agents/rules/next-step-output.md` 的「人工裁决待办前置块」展开详情，并只输出修订与复审路径。
+
+场景 B/C/D 在修订命令后继续列出复审命令：`/review-plan {task-ref}`、`/{{project}}:review-plan {task-ref}`、`$review-plan {task-ref}`。
 
 ### 场景 A：通过且无问题
 
@@ -34,23 +36,18 @@
 提醒：manual-validation 项需在 PR description 的「待人工验证」清单中承接，不应触发 /plan-task。
 ```
 
-### 场景 B：通过但有问题
+### 场景 B：需要修改（major / minor）
 
 ```text
-任务 {task-id} 技术方案审查完成。结论：通过。
+任务 {task-id} 技术方案审查完成。结论：需要修改。
 - 阻塞项：0 | 主要问题：{n} | 次要问题：{n} | 人工裁决：{h}
 - 审查报告：.agents/workspace/active/{task-id}/{review-artifact}
 
-下一步 - 修订方案后编码（推荐）：
+下一步 - 修订技术方案：
   - Claude Code / OpenCode：/plan-task {task-ref}
   - Gemini CLI：/{{project}}:plan-task {task-ref}
   - Codex CLI：$plan-task {task-ref}
 
-或直接进入编码：
-  - Claude Code / OpenCode：/code-task {task-ref}
-  - Gemini CLI：/{{project}}:code-task {task-ref}
-  - Codex CLI：$code-task {task-ref}
-
 [当 manual-validation > 0 时，在最后附加一行：]
 提醒：manual-validation 项需在 PR description 的「待人工验证」清单中承接，不应触发 /plan-task。
 ```

package/templates/.agents/skills/review-plan/reference/report-template.en.md

@@ -40,12 +40,18 @@ Use this template when writing `review-plan.md` or `review-plan-r{N}.md`.
 **Description**: {details}
 **Suggested Fix**: {fix suggestion}
 
-### Minor Issues (optional improvements)
+### Minor Issues (low impact, closure required)
 
 #### 1. {Improvement point}
 **File**: `{file-path}:{line-number}`
 **Suggestion**: {improvement suggestion}
 
+## Non-blocking Advisories
+
+> Record only future optimizations that do not affect the current artifact's completeness, correctness, or acceptance. Advisories do not enter the disagreement ledger, finding counts, or verdict.
+
+- {future optimization}
+
 ## Manual Validation Items
 
 > Items the AI agent cannot close in the current execution environment; they do not participate in the next plan round. Maintainers carry them in the PR description as a "manual verification required" checklist.
@@ -61,7 +67,8 @@ Use this template when writing `review-plan.md` or `review-plan-r{N}.md`.
 
 ## Review Disagreement Ledger Writeback
 
-> Upsert each finding this round into the task.md disagreement ledger: append an `open` row for new findings (id prefix `PL-`, stage=plan); per the hand-back duty set the executor's prior-round responses to `confirmed` / back to `open` / `needs-human-decision`. State machine and evidence rules: `.agents/rules/review-handshake.md`.
+> Record the structured intents to submit: use `task-ledger finding-upsert` for new findings and `finding-review` for prior responses. The core allocates `PL-N` and validates transitions; do not edit task.md table rows.
+> Every finding escalated to `needs-human-decision` must include a self-contained detail block per `.agents/rules/human-decision-context.md`, with evidence pointing to that stable anchor.
 
 ## Evidence
 

package/templates/.agents/skills/review-plan/reference/report-template.zh-CN.md

@@ -40,12 +40,18 @@
 **说明**：{details}
 **修复建议**：{fix suggestion}
 
-### 次要问题（可选改进）
+### 次要问题（低影响但需闭环）
 
 #### 1. {改进点}
 **文件**：`{file-path}:{line-number}`
 **建议**：{improvement suggestion}
 
+## 非阻塞建议
+
+> 仅记录不影响当前产物完整性、正确性和验收的后续优化；不写入审查分歧账本，不计入 blocker / major / minor，也不影响结论。
+
+- {future optimization}
+
 ## 人工校验项
 
 > AI agent 在本执行环境无法闭环的项；不参与下一轮 refine。维护者在 PR description 中以「待人工验证」清单承接。
@@ -61,7 +67,8 @@
 
 ## 审查分歧账本回写
 
-> 把本轮每条 finding upsert 到 task.md `## 审查分歧账本`：新 finding 追加 `open` 行（id 前缀 `PL-`，stage=plan），对执行方上一轮响应按回交义务改 `confirmed` / 置回 `open` / `needs-human-decision`。状态机与证据规则见 `.agents/rules/review-handshake.md`。
+> 本段记录将提交的结构化意图：新 finding 用 `task-ledger finding-upsert`，上一轮响应复核用 `finding-review`；由核心分配 `PL-N` 并校验状态机，禁止手写 task.md 表格。
+> 凡升级为 `needs-human-decision` 的 finding，必须按 `.agents/rules/human-decision-context.md` 在本报告中提供自足详情块，并让 evidence 指向该稳定锚点。
 
 ## 证据原文
 

package/templates/.agents/skills/review-plan/reference/review-criteria.en.md

@@ -14,6 +14,7 @@ Follow the `design-review` step in `.agents/workflows/feature-development.yaml`.
 - [ ] Risks, migration, rollback, or compatibility handling are sufficient
 - [ ] The plan avoids over-design and unrelated scope expansion
 - [ ] The reviewer checked whether the executor missed any key design decision that should be upgraded to `[needs-human-decision]`
+- [ ] Every `needs-human-decision` detail produced this round follows the self-contained structure in `.agents/rules/human-decision-context.md`
 - [ ] Every blocker is backed by reproducible grep/sed/nl evidence; conclusions not directly verified are declared under Self-Doubt
 
 **Common anti-examples**:
@@ -29,6 +30,12 @@ Follow the `design-review` step in `.agents/workflows/feature-development.yaml`.
 3. **Actionable**: suggest a concrete fix
 4. **Severity-based**: clearly distinguish blockers, major issues, and minor issues
 
+## Three-category Decision Tree
+
+1. A gap in constraints, risks, interface boundaries, validation strategy, or implementability is a formal finding. Assign blocker / major / minor by impact; severity does not control closure, so minor findings must also reach a terminal state.
+2. When there is no known defect but validation requires a real environment, permission, or human operation, classify it as manual-validation.
+3. Only a future optimization that does not affect the current plan's completeness, correctness, or acceptance is an advisory. Put it only in Non-blocking Advisories; do not add it to the ledger, finding counts, or verdict.
+
 ## Manual Validation Classification
 
 Some findings cannot be closed by an AI agent in the current execution environment, for example:

package/templates/.agents/skills/review-plan/reference/review-criteria.zh-CN.md

@@ -14,6 +14,7 @@
 - [ ] 风险、迁移、回滚或兼容性处理是否充分
 - [ ] 方案是否避免过度设计和无关扩张
 - [ ] 已复核执行方是否漏标应升级为 `[needs-human-decision]` 的关键设计决策
+- [ ] 本轮所有 `needs-human-decision` 详情均符合 `.agents/rules/human-decision-context.md` 的自足结构
 - [ ] 每条 blocker 都配可复现的 grep/sed/nl 证据，未直接验证的结论已在「自我质疑」声明
 
 **常见反例**：
@@ -29,6 +30,12 @@
 3. **可执行**：给出明确可落地的修复建议
 4. **按严重程度分类**：明确区分 blocker、major 和 minor
 
+## 三类审查项决策树
+
+1. 若问题揭示当前方案缺少约束、风险、接口边界、验证策略或可实施性，登记为正式 finding，并按影响赋予 blocker / major / minor；severity 只表示影响大小，minor 也必须闭环。
+2. 若没有已知缺陷，但只能依赖真实环境、权限或人工操作完成验证，归为 manual-validation。
+3. 仅当建议属于未来优化且不影响当前方案的完整性、正确性和验收时，归为 advisory。advisory 只写入「非阻塞建议」，不进入账本、问题计数或 verdict。
+
 ## 人工校验项分类
 
 某些发现项是 AI agent 在本执行环境**无法闭环**的，例如：

package/templates/.agents/skills/update-agent-infra/SKILL.en.md

@@ -41,7 +41,7 @@ The script outputs JSON to stdout. Parse and record the report.
 
 **Key fields**:
 - `error`: error message (if non-empty, stop and report)
-- `templateVersion`: current template source version
+- `templateVersion`: exact `v`-prefixed SemVer of the template source package (including prerelease or build metadata)
 - `templateRoot`: absolute path to the template file root directory
 - `templateSources.conflicts`: external template source conflicts; explicitly
   list these in the report so users know which files were ignored because a
@@ -55,7 +55,6 @@ The script outputs JSON to stdout. Parse and record the report.
 - `merged.pending`: list of merged files for AI to process
   - Each item has `target` (project-relative path) and `template` (template-root-relative path)
 - `registryAdded`: newly added file registry entries
-- `selfUpdate`: whether this is a self-update scenario
 - `configUpdated`: whether `.agents/.airc.json` was updated
 
 If `managed.conflicts` is non-empty, output every conflict and stop immediately before Phase B. Do not overwrite a conflicting file or advance its `files.managedBaselines` entry.

package/templates/.agents/skills/update-agent-infra/SKILL.zh-CN.md

@@ -35,7 +35,7 @@ node .agents/skills/update-agent-infra/scripts/sync-templates.js
 
 **关键字段**：
 - `error`：错误信息（如非空则停止并报告）
-- `templateVersion`：模板源当前版本
+- `templateVersion`：模板源包的精确 `v` 前缀 SemVer（可包含 prerelease 或 build metadata）
 - `templateRoot`：模板文件根目录绝对路径
 - `templateSources.conflicts`：外部模板源冲突列表；报告中必须显式展示，说明哪些文件因内置模板或后续外部源获胜而被忽略
 - `managed.written` / `managed.created`：已更新/新建的 managed 文件
@@ -47,7 +47,6 @@ node .agents/skills/update-agent-infra/scripts/sync-templates.js
 - `merged.pending`：需要 AI 处理的 merged 文件列表
   - 每项包含 `target`（项目中的目标路径）和 `template`（模板根目录下的相对路径）
 - `registryAdded`：新增的文件注册条目
-- `selfUpdate`：是否为自更新模式
 - `configUpdated`：`.agents/.airc.json` 是否已更新
 
 如果 `managed.conflicts` 非空，输出全部冲突并立即停止，不进入阶段 B；禁止覆盖冲突文件或推进其 `files.managedBaselines`。

package/templates/.agents/skills/update-agent-infra/scripts/sync-templates.js

@@ -13,7 +13,6 @@
  * Output: JSON report to stdout.
  */
 
-import childProcess from 'node:child_process';
 import crypto from 'node:crypto';
 import fs from 'node:fs';
 import os from 'node:os';
@@ -69,6 +68,7 @@ const DEFAULTS = {
       ".claude/commands/",
       ".codex/hooks.json",
       ".gemini/commands/",
+      ".git-hooks/check-large-files.cjs",
       ".git-hooks/check-version-format.sh",
       ".github/scripts/",
       ".github/workflows/metadata-sync.yml",
@@ -106,7 +106,7 @@ const AGENT_INFRA_SANDBOX_TOOL = 'agent-infra';
 const LEGACY_DEFAULT_SANDBOX_TOOLS = ['claude-code', 'codex', 'gemini-cli', 'opencode'];
 const DEFAULT_SANDBOX_TOOLS = [AGENT_INFRA_SANDBOX_TOOL, ...LEGACY_DEFAULT_SANDBOX_TOOLS];
 // Add a new identifier here only after shipping matching .{platform}. template variants.
-const KNOWN_PLATFORMS = new Set(['github']);
+const KNOWN_PLATFORMS = new Set(['github', 'none']);
 const KNOWN_LANGUAGES = new Set(['en', 'zh-CN']);
 
 // Single source of truth for built-in TUI ids and owned path prefixes.
@@ -870,14 +870,6 @@ function isBinary(fp) {
   return false;
 }
 
-function gitUrl(dir) {
-  try {
-    return childProcess.execSync('git remote get-url origin', {
-      cwd: dir, encoding: 'utf8', stdio: ['pipe', 'pipe', 'pipe']
-    }).trim();
-  } catch { return null; }
-}
-
 function langSelect(rels, lang, allSet, project) {
   const sel = new Map();
 
@@ -1020,8 +1012,7 @@ function syncTemplates(projectRoot, templateRootOverride) {
     },
     ejected: { created: [], skipped: [] },
     merged:  { pending: [] },
-    configUpdated: false,
-    selfUpdate: false
+    configUpdated: false
   };
   const customTUIs = validateCustomTUIs(projectRoot, customTUIsConfig, report);
   const customTUICommandTargets = buildCustomTUICommandTargets(
@@ -1365,9 +1356,6 @@ function syncTemplates(projectRoot, templateRootOverride) {
     ([target, template]) => ({ target, template })
   );
 
-  const projUrl = gitUrl(projectRoot);
-  report.selfUpdate = !!(projUrl && /fitlab-ai\/agent-infra/.test(projUrl));
-
   const hasChanges = (
     report.managed.written.length +
     report.managed.created.length +

package/templates/.agents/skills/watch-pr/SKILL.en.md

@@ -12,16 +12,16 @@ After `create-pr`, continuously watch the PR's required CI checks: when everythi
 ## Behavior Boundaries / Key Rules
 
 - Only watch + self-heal the current PR's required checks; make no changes unrelated to the failing check.
-- Self-heal modifies code and `git push`es to the PR branch, but **local tests for the affected area must pass before pushing**; fix attempts have a hard cap (default 2); only self-heal locatable code-layer failures (lint / format / test / type / build), and always route non-code failures (network / permission / external service / flaky) to the help exit.
+- Self-heal publishes through Git workflow intents, but **affected tests must pass first**; the attempt cap and code-layer authorization remain unchanged.
 - The help exit is "produce-then-stop": end this round, output the blocker explanation, and wait for the user to trigger the next step — **never** ask mid-flow.
-- Bare numbers / `#NN` / `TASK-id` arguments are always resolved as task short ids (see `.agents/rules/task-short-id.md`); a PR number is passed only via `--pr <number>` / a PR URL / omission (current branch), never reusing the bare-number syntax.
+- Bare numbers / `NN` / `TASK-id` arguments are always resolved as task short ids (see `.agents/rules/task-short-id.md`); a PR number is passed only via `--pr <number>` / a PR URL / omission (current branch), never reusing the bare-number syntax.
 - After running this skill (task-anchored path), you must update task.md.
 
 Version stamp rule: before creating or updating `task.md` frontmatter, read `.agents/rules/version-stamp.md` and write or refresh `agent_infra_version`.
 
-## Task Argument Short-ID Alias
+## Task Context Resolution
 
-> If the `{task-id}` argument matches `^[#]?[0-9]+$` (a bare number or `#`-prefixed), first read the "SKILL argument parsing" section of `.agents/rules/task-short-id.md` to resolve it; subsequent commands treat `{task-id}` as the resolved full `TASK-YYYYMMDD-HHMMSS` form.
+> Keep `--pr <number>` and PR URLs on the existing PR-anchored path. Otherwise, the task path may omit the task ref and also accepts a legacy positional ref or `--task <ref>` / `-t <ref>`. Separate task scope from the full arguments, then call `agent-infra-internal task-context resolve {task-scope}` and bind `{task-id}` to the returned full `taskId`. A PR anchor and task scope are mutually exclusive.
 
 ## Step Start: Write the started Marker
 
@@ -39,22 +39,22 @@ After prerequisites pass and before this round's first artifact action, append a
 
 Resolve the target PR number `{pr#}` and an optional `{task-id}` via these deterministic branches:
 
-- Scenario A (argument omitted): use the current branch's PR number per `.agents/rules/pr-checks-commands.md`; then determine `{task-id}` via "Reverse-lookup task" below.
-- Scenario B (`#NN` / bare number / `TASK-id`, **task-anchored primary path**): when matching `^[#]?[0-9]+$`, resolve to the full `{task-id}` via "Task Argument Short-ID Alias" (on failure pass through the exit code; do not rewrite error handling); a `TASK-id` is used directly. Read `.agents/workspace/active/{task-id}/task.md` for `pr_number` as `{pr#}`; if `pr_number` is empty, follow "Error Handling" to prompt running `create-pr` first, then stop.
+- Scenario A (argument omitted): reverse-lookup the active task from the current branch, then read its `pr_number`.
+- Scenario B (omitted task ref, positional task ref, or `--task/-t`, **task-anchored primary path**): resolve the full `{task-id}` via "Task Context Resolution" and read `.agents/workspace/active/{task-id}/task.md` for `pr_number` as `{pr#}`; if `pr_number` is empty, follow "Error Handling" to prompt running `create-pr` first, then stop.
 - Scenario C (`--pr <number>` or a PR URL): use that PR number directly as `{pr#}`; then determine `{task-id}` via "Reverse-lookup task".
-- Reverse-lookup task (scenarios A / C): search `.agents/workspace/active/*/task.md` for a task whose `pr_number == {pr#}`; on a hit, take that `{task-id}` (task-anchored); on a miss, enter the "watch-only" degraded path (no `{task-id}`, skip steps 5/6).
+- Reverse-lookup task (scenarios A / C): use task context/query capabilities to find the unique active task bound to `{pr#}`. If none exists, stop and request binding first; the typed checks intent does not create a second taskless state machine.
 
 ### 2. Watch Required Checks
 
 Before running this step, read `reference/monitor-and-heal.md` and `.agents/rules/pr-checks-commands.md`.
 
-Using the watch command in `.agents/rules/pr-checks-commands.md`, poll `{pr#}`'s required checks (with an overall time cap, default 30 minutes), and classify the outcome per the "Outcome Classification" of `reference/monitor-and-heal.md` into the "all green" / "failure" / "pending" scenarios, routing to step 7 (green exit), step 3 (self-heal), or step 4 (help exit) respectively.
+Initialize the in-memory list `repairCommits=[]` for this run. Run `agent-infra-internal platform-checks watch {task-id} --interval-seconds 30 --deadline-seconds 1800`, then route its structured `checks.state` to the all-green, failure, or pending path.
 
 ### 3. Failure Self-Heal Loop
 
 Before running this step, read the "Self-Heal Decision Tree" of `reference/monitor-and-heal.md` and "Resolve a Failing Run id and Pull Logs" of `.agents/rules/pr-checks-commands.md`.
 
-For a failing check: first deterministically resolve its failing run and pull the failure logs per the rule, then classify the failure; before making any local fix, read `.agents/rules/debugging-guide.md` and locate the root cause via its four-phase flow, never blindly patching and retrying; only when it is a locatable code-layer failure, make a minimal local fix, run the relevant tests until they pass, then **stage, commit, and push the fix** (`git add` only the related files → `git commit` per `.agents/rules/commit-and-pr.md` → `git push` to the current PR branch, recording the commit SHA), and return to step 2 to re-watch. Count fix attempts; on reaching the hard cap (default 2) or when the run is unlocatable, go to step 4.
+For a locatable code-layer failure, make the minimum fix and run relevant tests. Then call `git-workflow commit` and `git-workflow push`; append only a remotely verified SHA to `repairCommits`.
 
 ### 4. Help Exit (Produce-Then-Stop)
 
@@ -77,7 +77,7 @@ Update `.agents/workspace/active/{task-id}/task.md`:
 - **Do not change** `pr_status` (keep `created`) or `current_step`
 - **Append** to `## Activity Log` (do not overwrite prior entries; `{N}` = number of existing Watch PR entries for this task + 1):
   ```
-  - {YYYY-MM-DD HH:mm:ss±HH:MM} — **Watch PR (Round {N})** by {agent} — {green: all required checks green / blocked: blocked: {summary}}
+  - {YYYY-MM-DD HH:mm:ss±HH:MM} — **Watch PR (Round {N})** by {agent} — {green: all required checks green, repair commits: {k} [{SHA summary}] / blocked: blocked: {summary}}
   ```
 
 ### 6. Verification Gate
@@ -87,7 +87,7 @@ Update `.agents/workspace/active/{task-id}/task.md`:
 Run the verification gate:
 
 ```bash
-node .agents/scripts/validate-artifact.js gate watch-pr .agents/workspace/active/{task-id} --format text
+agent-infra-internal task-verify {task-id} watch-pr.completed --format text
 ```
 
 Handle the result:
@@ -101,10 +101,12 @@ Keep the gate output in your reply as the verification evidence. Without current
 
 > On the task-anchored path, execute this step only after the gate passes.
 
-> **IMPORTANT**: All TUI command formats listed below must be output in full. Do not show only the format for the current AI agent. If `.agents/.airc.json` configures custom TUIs (via `customTUIs`), read each tool's `name` and `invoke`, then add the matching command line in the same format (`${skillName}` becomes the skill name and `${projectName}` becomes the project name). Before rendering the final output, read `.agents/rules/next-step-output.md` and apply both of its rules: (1) render `{task-ref}` in the "Next steps" commands as the short id `#NN` (falling back to the full TASK-id when unallocated or released); (2) append the `Completed at` line as the very last line of the user-facing output (this applies to every user-facing output — success, error, and early-return paths alike, not only the success path).
+> **IMPORTANT**: All TUI command formats listed below must be output in full. Do not show only the format for the current AI agent. If `.agents/.airc.json` configures custom TUIs (via `customTUIs`), read each tool's `name` and `invoke`, then add the matching command line in the same format (`${skillName}` becomes the skill name and `${projectName}` becomes the project name). Before rendering the final output, read `.agents/rules/next-step-output.md` and apply both of its rules: (1) render `{task-ref}` in the "Next steps" commands as the short id `NN` (falling back to the full TASK-id when unallocated or released); (2) append the `Completed at` line as the very last line of the user-facing output (this applies to every user-facing output — success, error, and early-return paths alike, not only the success path).
 
 Output per scenario:
-- "All green" + task-anchored: state that all required checks passed and the PR is ready to merge, then render the next step from the template below (`{task-ref}` becomes the short id):
+- "All green" + task-anchored: state that all required checks passed, then render exactly one exit based on whether this run created repair commits (`{task-ref}` becomes the short id):
+
+  `repairCommits.length == 0`:
 
   ```
   Next step - Complete and archive the task:
@@ -113,7 +115,15 @@ Output per scenario:
     - Codex CLI: $complete-task {task-ref}
   ```
 
-- "All green" + watch-only: state the PR is ready to merge; there is no linked task this run, so run `complete-task` against the relevant task (do not force a short-id command block when no `{task-ref}` is available).
+  `repairCommits.length > 0`:
+
+  ```
+  Next step - Re-run code review:
+    - Claude Code / OpenCode: /review-code {task-ref}
+    - Gemini CLI: /{{project}}:review-code {task-ref}
+    - Codex CLI: $review-code {task-ref}
+  ```
+
 - "Blocked": output only the step 4 blocker explanation; do not recommend a next-step command.
 
 ## Completion Checklist

package/templates/.agents/skills/watch-pr/SKILL.zh-CN.md

@@ -12,16 +12,16 @@ description: >
 ## 行为边界 / 关键规则
 
 - 仅监控 + 自愈当前 PR 的 required checks；不做与失败 check 无关的改动。
-- 自愈会修改业务代码并 `git push` 到 PR 分支，但**推送前必须本地跑通相关测试**；修复尝试有硬上限（默认 2）；仅对可定位的代码层失败（lint / format / test / 类型 / 构建）自愈，非代码层（网络 / 权限 / 外部服务 / flaky）一律转求助出口。
+- 自愈通过 Git workflow intent 发布修复，但**发布前必须本地跑通相关测试**；修复上限与代码层分类授权不变。
 - 求助出口是「产出后停止」语义：停止本轮、输出阻塞说明、等待用户主动触发，**不**中途提问。
-- 裸数字 / `#NN` / `TASK-id` 入参一律按任务短号解析（见 `.agents/rules/task-short-id.md`）；PR 号只走 `--pr <number>` / PR URL / 省略（当前分支），不复用裸数字语法。
+- 裸数字 / `NN` / `TASK-id` 入参一律按任务短号解析（见 `.agents/rules/task-short-id.md`）；PR 号只走 `--pr <number>` / PR URL / 省略（当前分支），不复用裸数字语法。
 - 执行本技能（任务锚定路径）后，必须更新 task.md。
 
 版本戳规则：创建或更新 `task.md` frontmatter 时，先读取 `.agents/rules/version-stamp.md`，并写入或刷新 `agent_infra_version`。
 
-## 任务入参短号别名
+## 任务上下文解析
 
-> 如果 `{task-id}` 入参匹配 `^[#]?[0-9]+$`（裸数字或带 `#` 前缀），先读取 `.agents/rules/task-short-id.md` 的「SKILL 入参解析」段执行解析；后续命令视 `{task-id}` 为解析后的全长 `TASK-YYYYMMDD-HHMMSS` 形式。
+> 入口允许省略 task ref，也接受旧位置 task ref 或 `--task <ref>` / `-t <ref>`。先从完整参数中分离 task scope 并原样保留其他业务操作数，再调用 `agent-infra-internal task-context resolve {task-scope}`；`{task-scope}` 为空、位置 ref 或 task flag 之一。只读取结构化结果的 `taskId`，后续把 `{task-id}` 绑定为该完整 `TASK-YYYYMMDD-HHMMSS`。解析失败时透传非零退出码，不自行扫描任务。
 
 ## 步骤开始：写入 started 标记
 
@@ -39,22 +39,22 @@ description: >
 
 按以下确定性分支解析出目标 PR 号 `{pr#}` 与可选 `{task-id}`：
 
-- 场景 A（省略入参）：按 `.agents/rules/pr-checks-commands.md` 取当前分支的 PR 号；随后按下方「反查任务」确定 `{task-id}`。
-- 场景 B（`#NN` / 裸数字 / `TASK-id`，**任务锚定主路径**）：匹配 `^[#]?[0-9]+$` 时按「任务入参短号别名」解析为完整 `{task-id}`（解析失败直接透传退出码，不重写错误处理）；`TASK-id` 直接采用。读 `.agents/workspace/active/{task-id}/task.md` 取 `pr_number` 作为 `{pr#}`；`pr_number` 为空时按「错误处理」提示先 `create-pr`，停止。
+- 场景 A（省略入参）：从当前分支反查 active task；定位后读取其 `pr_number`。
+- 场景 B（省略 task ref、旧位置 task ref 或 `--task/-t`，**任务锚定主路径**）：按「任务上下文解析」取得完整 `{task-id}`。读 `.agents/workspace/active/{task-id}/task.md` 取 `pr_number` 作为 `{pr#}`；`pr_number` 为空时按「错误处理」提示先 `create-pr`，停止。
 - 场景 C（`--pr <number>` 或 PR URL）：直接取该 PR 号为 `{pr#}`；随后按「反查任务」确定 `{task-id}`。
-- 反查任务（场景 A / C）：在 `.agents/workspace/active/*/task.md` 中查找 `pr_number == {pr#}` 的任务；命中则取该 `{task-id}`（任务锚定）；未命中则进入「仅监控」降级路径（无 `{task-id}`，跳过步骤 5/6）。
+- 反查任务（场景 A / C）：通过任务上下文/任务查询取得与 `{pr#}` 唯一绑定的 active task。未命中时停止并提示先绑定 PR；typed checks intent 不建立第二套无任务状态机。
 
 ### 2. 监控 required checks
 
 执行此步骤前，先读取 `reference/monitor-and-heal.md` 与 `.agents/rules/pr-checks-commands.md`。
 
-按 `.agents/rules/pr-checks-commands.md` 的监控命令对 `{pr#}` 的 required checks 轮询（含总时长上限，默认 30 分钟），按 `reference/monitor-and-heal.md` 的「结果分类」分为「全绿」/「失败」/「挂起」三个场景，分别进入步骤 7 全绿出口、步骤 3 自愈、或步骤 4 求助。
+进入本轮时初始化内存列表 `repairCommits=[]`。调用 `agent-infra-internal platform-checks watch {task-id} --interval-seconds 30 --deadline-seconds 1800`，按结构化 `checks.state` 分为「全绿」/「失败」/「挂起」，分别进入步骤 7、步骤 3 或步骤 4。
 
 ### 3. 失败自愈循环
 
 执行此步骤前，先读取 `reference/monitor-and-heal.md` 的「自愈决策树」与 `.agents/rules/pr-checks-commands.md` 的「解析失败 run id 并拉日志」。
 
-对失败 check：先按规则确定性解析其失败 run 并拉取失败日志、判定失败类别；本地修复前先读取 `.agents/rules/debugging-guide.md`，按其四阶段流程定位根因，禁止盲目改代码重试；仅当属可定位的代码层失败时，本地最小化修复、运行对应测试通过后**暂存并提交本次修复再推送**（`git add` 仅相关文件 → 按 `.agents/rules/commit-and-pr.md` `git commit` → `git push` 到当前 PR 分支，并记录 commit SHA），再回到步骤 2 重新监控。修复尝试计数，达硬上限（默认 2）或 run 不可定位 → 转步骤 4。
+只对可定位代码层失败最小修复并测试；通过后依次调用 `git-workflow commit` 与 `git-workflow push`，仅记录远端复核成功的 SHA。达硬上限或 run 不可定位时转步骤 4。
 
 ### 4. 求助出口（产出后停止）
 
@@ -77,7 +77,7 @@ date "+%Y-%m-%d %H:%M:%S%z" | sed 's/\([+-][0-9][0-9]\)\([0-9][0-9]\)$/\1:\2/'
 - **不改** `pr_status`（保持 `created`）与 `current_step`
 - **追加**到 `## 活动日志`（不要覆盖之前的记录；`{N}` = 本任务已有 Watch PR 条目数 + 1）：
   ```
-  - {YYYY-MM-DD HH:mm:ss±HH:MM} — **Watch PR (Round {N})** by {agent} — {全绿：all required checks green / 阻塞：blocked: {简述}}
+  - {YYYY-MM-DD HH:mm:ss±HH:MM} — **Watch PR (Round {N})** by {agent} — {全绿：all required checks green, repair commits: {k} [{sha 摘要}] / 阻塞：blocked: {简述}}
   ```
 
 ### 6. 完成校验
@@ -87,7 +87,7 @@ date "+%Y-%m-%d %H:%M:%S%z" | sed 's/\([+-][0-9][0-9]\)\([0-9][0-9]\)$/\1:\2/'
 运行完成校验：
 
 ```bash
-node .agents/scripts/validate-artifact.js gate watch-pr .agents/workspace/active/{task-id} --format text
+agent-infra-internal task-verify {task-id} watch-pr.completed --format text
 ```
 
 处理结果：
@@ -101,10 +101,12 @@ node .agents/scripts/validate-artifact.js gate watch-pr .agents/workspace/active
 
 > 任务锚定路径仅在校验通过后执行本步骤。
 
-> **重要**：以下「下一步」中列出的所有 TUI 命令格式必须完整输出，不要只展示当前 AI 代理对应的格式。如果 `.agents/.airc.json` 中配置了自定义 TUI（`customTUIs`），读取每个工具的 `name` 和 `invoke`，按同样格式补充对应命令行（`${skillName}` 替换为技能名，`${projectName}` 替换为项目名）。 渲染最终输出前，先读取 `.agents/rules/next-step-output.md` 并落实其两类规则：(1) 「下一步」命令把 `{task-ref}` 渲染为短号 `#NN`（未分配/已释放时回退完整 TASK-id）；(2) 在面向用户输出的绝对最后一行追加 `Completed at` 收尾行（成功、错误、早退等任何面向用户输出都适用，不限于校验通过的成功态）。
+> **重要**：以下「下一步」中列出的所有 TUI 命令格式必须完整输出，不要只展示当前 AI 代理对应的格式。如果 `.agents/.airc.json` 中配置了自定义 TUI（`customTUIs`），读取每个工具的 `name` 和 `invoke`，按同样格式补充对应命令行（`${skillName}` 替换为技能名，`${projectName}` 替换为项目名）。 渲染最终输出前，先读取 `.agents/rules/next-step-output.md` 并落实其两类规则：(1) 「下一步」命令把 `{task-ref}` 渲染为短号 `NN`（未分配/已释放时回退完整 TASK-id）；(2) 在面向用户输出的绝对最后一行追加 `Completed at` 收尾行（成功、错误、早退等任何面向用户输出都适用，不限于校验通过的成功态）。
 
 按场景输出：
-- 「全绿」+ 任务锚定：说明所有 required checks 已通过、PR 可合入，并按下方模板渲染下一步（`{task-ref}` 替换为短号）：
+- 「全绿」+ 任务锚定：说明所有 required checks 已通过，并只按本轮是否产生修复 commit 渲染一个出口（`{task-ref}` 替换为短号）：
+
+  `repairCommits.length == 0`：
 
   ```
   下一步 - 完成并归档任务：
@@ -113,7 +115,15 @@ node .agents/scripts/validate-artifact.js gate watch-pr .agents/workspace/active
     - Codex CLI：$complete-task {task-ref}
   ```
 
-- 「全绿」+ 仅监控降级：说明 PR 可合入；本次无关联任务，请对相应任务运行 `complete-task`（无 `{task-ref}` 可渲染时不强行输出短号命令块）。
+  `repairCommits.length > 0`：
+
+  ```
+  下一步 - 重新代码审查：
+    - Claude Code / OpenCode：/review-code {task-ref}
+    - Gemini CLI：/{{project}}:review-code {task-ref}
+    - Codex CLI：$review-code {task-ref}
+  ```
+
 - 「阻塞」：仅输出步骤 4 的阻塞说明，不推荐下一步命令。
 
 ## 完成检查清单
@@ -127,7 +137,7 @@ node .agents/scripts/validate-artifact.js gate watch-pr .agents/workspace/active
 
 ## 停止
 
-完成检查清单后立即停止。全绿出口等待用户运行 `complete-task`；阻塞出口等待用户裁定。
+完成检查清单后立即停止。全绿出口等待用户运行所选的 `complete-task` 或 `review-code`；阻塞出口等待用户裁定。
 
 ## 注意事项
 

package/templates/.agents/skills/watch-pr/config/verify.json

@@ -15,8 +15,7 @@
       ]
     },
     "activity-log": {
-      "expected_action_pattern": "(Watch PR|Watching PR) \\(Round \\d+\\)",
-      "freshness_minutes": 30
+      "expected_action_pattern": "(Watch PR|Watching PR) \\(Round \\d+\\)"
     }
   }
 }

package/templates/.agents/skills/watch-pr/reference/monitor-and-heal.en.md

@@ -30,8 +30,8 @@ For each failing check, decide "self-heal" vs "ask for help" in this order:
    - Before fixing, run `git status -s` to record the working tree and ensure only changes related to this failure are included.
    - Locate and make a minimal fix per the logs (touch only code / tests / config related to that failure).
    - Run the relevant tests: prefer the local command for the failing job; otherwise read the project `test` skill and select its declared core or full validation command. If neither command is known, ask for help. **Do not commit or push before tests pass.**
-   - After tests pass, publish the fix: per `.agents/rules/commit-and-pr.md` stage only the related files (`git add <paths>`, avoid `git add -A` pulling in unrelated changes) → create the fix commit (`git commit`) → `git push` to the current PR branch.
-   - Record the fix commit SHA, increment the fix count, and return to SKILL step 2 to re-watch.
+   - After tests pass, call `git-workflow commit` and `git-workflow push`, then verify the remote SHA.
+   - Append the fix commit SHA to this run's `repairCommits`, increment the fix count, and return to SKILL step 2. When checks turn green, an empty list routes to `complete-task`; a non-empty list routes to `review-code`.
    - Never make unrelated "drive-by" optimizations; never loosen / skip the failing assertion to "make it green".
 
 ## Help Report Template

package/templates/.agents/skills/watch-pr/reference/monitor-and-heal.zh-CN.md

@@ -30,8 +30,8 @@ unknown: help
    - 自愈前先 `git status -s` 记录当前工作树，确保后续只纳入与本次失败相关的改动。
    - 在本地按日志定位并最小化修复（只动与该失败相关的代码 / 测试 / 配置）。
    - 运行对应测试：优先失败 job 对应的本地命令；否则读取项目 `test` skill，选择其中声明的 core 或 full 验证命令。两者均未知时进入求助出口。**测试通过前不得提交或推送。**
-   - 测试通过后发布修复：按 `.agents/rules/commit-and-pr.md` 仅暂存本次相关文件（`git add <相关路径>`，避免 `git add -A` 卷入无关改动）→ 创建修复 commit（`git commit`）→ `git push` 到当前 PR 分支。
-   - 记录本次修复的 commit SHA，修复计数 +1，回到 SKILL 步骤 2 重新监控。
+   - 测试通过后发布修复：把相关路径、message 与 expected HEAD 写入 intent JSON，调用 `git-workflow commit`，再调用 `git-workflow push` 并复核远端 SHA。
+   - 将本次修复的 commit SHA 追加到当前运行的 `repairCommits`，修复计数 +1，回到 SKILL 步骤 2 重新监控；全绿时列表为空走 `complete-task`，非空走 `review-code`。
    - 绝不执行与失败无关的「顺手优化」；不放宽 / 跳过失败的断言来「修绿」。
 
 ## 求助报告模板

package/templates/.agents/templates/task.en.md

@@ -22,6 +22,24 @@ pr_status: pending             # PR status: pending (default) | created (PR crea
 
 [Describe the task clearly and concisely.]
 
+## Task Input
+
+<!-- Populated by create-task from information already present in the current request and necessary prior discussion. Leave missing categories empty; do not infer them. -->
+
+### Sources
+
+### Confirmed Facts and Evidence
+
+### Constraints
+
+### Confirmed Decisions
+
+### Candidate and Rejected Options
+
+### Acceptance Criteria
+
+### Open Questions
+
 ## Context
 
 - **Related Issue**: #XXX
@@ -62,7 +80,12 @@ pr_status: pending             # PR status: pending (default) | created (PR crea
 
 ## Human Rulings
 
-<!-- Humans record rulings for needs-human-decision decisions here and flip matching HD- rows in the Review Disagreement Ledger to human-decided. -->
+<!-- Use ai decide <task-ref> <ordinal|ledger-id> [--needs-implementation true|false] <decision> for needs-human-decision rulings; code-stage rulings must explicitly declare whether implementation is required. -->
+
+## Implementation Inputs
+
+| id | ledger_id | decision_evidence | stage | needs_implementation | decided_at | status | consumed_by |
+|----|-----------|-------------------|-------|----------------------|------------|--------|-------------|
 
 ## Workflow Warnings