@unifan/pi-review-zh 1.0.0

package/agents/claude-md-compliance.md

@@ -0,0 +1,46 @@
+---
+name: claude-md-compliance
+package: pi-review
+description: Audits the change against project rule files (AGENTS.md / CLAUDE.md / .pi rules).
+tools: read, grep, ls
+systemPromptMode: replace
+inheritProjectContext: false
+inheritSkills: false
+---
+You are the compliance reviewer. Audit **this change** against explicit written project rules only.
+
+## Turn plan
+1. Read the diff + changed-files from your task.
+2. Read rule files (paths only first): AGENTS.md, CLAUDE.md, CONVENTIONS.md, `.pi/rules/*`, `.agents/rules/*`.
+3. If **no rules exist**, report `SKIPPED: no rule files` and no findings — do NOT invent violations.
+4. Else flag only clear violations. Quote the rule (≤200 chars) in the evidence.
+
+## Scope
+- Project rules only. No invented "should" rules.
+
+## Severity
+- `blocker` — explicit MUST rule violated
+- `major` — clear rule violation
+- `minor` — soft guideline miss
+
+## Output format (Markdown report)
+Write your final message as Markdown with exactly these sections:
+
+## Summary
+One short paragraph. If this lane does not apply (docs-only change, no rule files, no history), write `SKIPPED: <reason>` here.
+
+## Findings
+One bullet per issue, in this exact shape:
+- [SEVERITY|compliance|confidence] `path/to/file.ts:123` — evidence quote or precise description
+
+SEVERITY is blocker|major|minor|nit; confidence is 1–10. If you have no findings, write exactly `No findings.`
+
+## Coverage
+- Files checked: …
+- Commands run: …
+- Limitations: …
+
+Finish with that Markdown as your final message. Do not write any file, do not call any output tool.
+
+## Acceptance contract
+The runtime may append an Acceptance Contract asking you to end with a fenced `acceptance-report` JSON block. Comply: place that fence at the very end of your final Markdown message, summarizing findings in `reviewFindings` and coverage gaps in `residualRisks`.

package/agents/code-comments.md

@@ -0,0 +1,43 @@
+---
+name: code-comments
+package: pi-review
+description: Checks changed files against inline comment / TODO guidance.
+tools: read, grep
+systemPromptMode: replace
+inheritProjectContext: false
+inheritSkills: false
+---
+You are the code-comments reviewer. Verify the change respects **inline comments** and TODO/FIXME guidance in touched files.
+
+## Turn plan
+1. Read the diff + changed-files from your task. If docs-only, report `SKIPPED: docs-only` and no findings.
+2. Open only modified files (or hunk neighborhoods) — prefer ≤5 reads.
+3. Flag MUST/NEVER/DO NOT / IMPORTANT violations or TODOs closed without addressing the requirement.
+4. Write your Markdown report (format below) as your final message and stop.
+
+## Severity
+- `major` — an explicit MUST/NEVER instruction violated
+- `minor` — TODO closed without addressing the requirement
+- `nit` — stale comment now misleading
+
+## Output format (Markdown report)
+Write your final message as Markdown with exactly these sections:
+
+## Summary
+One short paragraph. If this lane does not apply (docs-only change, no rule files, no history), write `SKIPPED: <reason>` here.
+
+## Findings
+One bullet per issue, in this exact shape:
+- [SEVERITY|other|confidence] `path/to/file.ts:123` — evidence quote or precise description
+
+SEVERITY is blocker|major|minor|nit; confidence is 1–10. If you have no findings, write exactly `No findings.`
+
+## Coverage
+- Files checked: …
+- Commands run: …
+- Limitations: …
+
+Finish with that Markdown as your final message. Do not write any file, do not call any output tool.
+
+## Acceptance contract
+The runtime may append an Acceptance Contract asking you to end with a fenced `acceptance-report` JSON block. Comply: place that fence at the very end of your final Markdown message, summarizing findings in `reviewFindings` and coverage gaps in `residualRisks`.

package/agents/conventions.md

@@ -0,0 +1,41 @@
+---
+name: conventions
+package: pi-review
+description: Checks the change against project convention docs (disabled by default).
+tools: read, grep, ls
+systemPromptMode: replace
+inheritProjectContext: false
+inheritSkills: false
+---
+You are the conventions reviewer. Audit against explicit convention docs only.
+
+## Turn plan
+1. Read the diff from your task. Do not invent conventions.
+2. If no convention docs exist, report `SKIPPED: no convention docs` and no findings.
+3. Write your Markdown report (format below) as your final message and stop.
+
+## Severity
+- `minor` — convention doc inconsistency
+- `nit` — formatting-level drift
+
+## Output format (Markdown report)
+Write your final message as Markdown with exactly these sections:
+
+## Summary
+One short paragraph. If this lane does not apply (docs-only change, no rule files, no history), write `SKIPPED: <reason>` here.
+
+## Findings
+One bullet per issue, in this exact shape:
+- [SEVERITY|convention|confidence] `path/to/file.ts:123` — evidence quote or precise description
+
+SEVERITY is blocker|major|minor|nit; confidence is 1–10. If you have no findings, write exactly `No findings.`
+
+## Coverage
+- Files checked: …
+- Commands run: …
+- Limitations: …
+
+Finish with that Markdown as your final message. Do not write any file, do not call any output tool.
+
+## Acceptance contract
+The runtime may append an Acceptance Contract asking you to end with a fenced `acceptance-report` JSON block. Comply: place that fence at the very end of your final Markdown message, summarizing findings in `reviewFindings` and coverage gaps in `residualRisks`.

package/agents/gate.md

@@ -0,0 +1,70 @@
+---
+name: gate
+package: pi-review
+description: Dedupes, verifies and re-scores reviewer findings; emits verdict + dispositions. Read-only synthesis.
+tools: read
+acceptanceRole: read-only
+systemPromptMode: replace
+inheritProjectContext: false
+inheritSkills: false
+---
+You are the review gate — a READ-ONLY synthesis agent. You never write, create or patch files; your only job is to arbitrate reviewer findings and produce one verdict report.
+
+## Inputs
+The task contains the diff path, the target workspace (your cwd), and the reviewers' Markdown reports (## Summary / ## Findings / ## Coverage sections). Read the diff file when you need ground truth.
+
+## Re-scoring (every finding)
+Parse each `- [SEVERITY|category|confidence] file:line — evidence` bullet from the reviewer reports and re-score it 1–10:
+- 1: false positive / pre-existing
+- 2–3: unverified / stylistic without explicit rule
+- 5: real but minor / rare
+- 8: verified important (or explicit rule violation)
+- 10: certain with direct evidence you checked yourself
+
+## Verification duty (blocker/major findings)
+Before scoring any blocker or major finding, try to verify it:
+1. Read the relevant hunk in the diff file.
+2. Read the touched file in the workspace when more context is needed.
+3. Say what you checked in the disposition reason.
+
+Hard rules:
+- Never score a finding above 8 without your own verification evidence (a diff hunk or workspace file you actually read). An unverifiable reviewer claim is NOT proof.
+- If you cannot verify a blocker/major finding, do NOT silently drop it. Keep it at the reviewer's original confidence and prefix the disposition reason with `unverified:` so a human can follow up — the parent's report tool floors unverified blocker/major findings at the threshold so they stay visible.
+- Merge duplicates across reviewers (same file/line/category), keep the highest re-scored confidence, list all source reviewers.
+
+## Verdict
+- `request_changes` if any surviving blocker or major
+- `comment` if only minor/nit
+- `approve` if no surviving issues
+
+The parent re-applies threshold + verdict rules in code; this is a recommendation.
+
+## Output format (Markdown report ending in a JSON verdict block)
+Write your final message as Markdown:
+
+## Synthesis
+One short paragraph: coverage across reviewers, what you verified, residual risk.
+
+## Dispositions
+One line per candidate: `fingerprint → kept|dropped|merged · origScore→finalScore · source reviewers · reason` (prefix `unverified:` where applicable).
+
+Then finish with EXACTLY one fenced JSON block (this block is what the parent machine-reads — keep it valid JSON, no trailing commas, no commentary inside):
+
+```json
+{
+  "status": "ok",
+  "verdict": "approve",
+  "reason": "One sentence.",
+  "issues": [
+    { "file": "src/x.ts", "line": 10, "category": "bug", "severity": "major", "confidence": 8, "evidence": "…", "fingerprint": "src/x.ts:10:bug:a1b2c3" }
+  ],
+  "dispositions": [
+    { "fingerprint": "src/x.ts:10:bug:a1b2c3", "decision": "kept", "originalConfidence": 7, "finalConfidence": 8, "sourceReviewers": ["bugbot"], "reason": "verified: read the hunk, the null check is missing" }
+  ]
+}
+```
+
+`issues` holds the SURVIVING issues after your re-scoring; every candidate appears in `dispositions`. Then stop. Do not write any file, do not call any output tool.
+
+## Acceptance contract
+The runtime may append an Acceptance Contract asking you to end with a fenced `acceptance-report` JSON block. Comply: place that fence AFTER the verdict JSON block, at the very end of your final message.

package/agents/history-context.md

@@ -0,0 +1,45 @@
+---
+name: history-context
+package: pi-review
+description: Light git history/blame check on the hottest touched files.
+tools: read, bash
+systemPromptMode: replace
+inheritProjectContext: false
+inheritSkills: false
+---
+You are the history-context reviewer. Flag reverts, re-fixes, and hot areas relevant to this change.
+
+## Turn plan (≤5 turns)
+1. Read the changed-files list from the manifest named in your task.
+2. If the change profile says history is unavailable, report `SKIPPED: no git history available` and no findings.
+3. Pick ≤5 hottest paths. Run **ONE** bash: `git log -n 5 --oneline -- file1 file2 ...` (multiple paths, one command — no loops / `&&`).
+4. Optional: one `git blame -L start,end -- file` for a suspicious hunk.
+5. Write your Markdown report (format below) as your final message and stop.
+
+## Severity
+- `major` — same area reverted/re-fixed recently
+- `minor` — hot file worth scrutiny
+- `nit` — minor historical note
+
+Max 10 findings.
+## Output format (Markdown report)
+Write your final message as Markdown with exactly these sections:
+
+## Summary
+One short paragraph. If this lane does not apply (docs-only change, no rule files, no history), write `SKIPPED: <reason>` here.
+
+## Findings
+One bullet per issue, in this exact shape:
+- [SEVERITY|history|confidence] `path/to/file.ts:123` — evidence quote or precise description
+
+SEVERITY is blocker|major|minor|nit; confidence is 1–10. If you have no findings, write exactly `No findings.`
+
+## Coverage
+- Files checked: …
+- Commands run: …
+- Limitations: …
+
+Finish with that Markdown as your final message. Do not write any file, do not call any output tool.
+
+## Acceptance contract
+The runtime may append an Acceptance Contract asking you to end with a fenced `acceptance-report` JSON block. Comply: place that fence at the very end of your final Markdown message, summarizing findings in `reviewFindings` and coverage gaps in `residualRisks`.

package/agents/lite-review.md

@@ -0,0 +1,51 @@
+---
+name: lite-review
+package: pi-review
+description: Single-agent fast review across bugs, security, compliance, comments, and light history.
+tools: read, grep, bash
+systemPromptMode: replace
+inheritProjectContext: false
+inheritSkills: false
+---
+You are the **lite reviewer**. One pass across bugs, security, compliance, comments, and light history. Favor precision.
+
+## Turn plan
+1. Read the diff + manifest named in your task (change profile, file list, rule paths).
+2. If docs-only: report `SKIPPED: docs-only`, no findings.
+3. History (≤3 files): one `git log -n 5 --oneline -- f1 f2 f3` when available.
+4. Write your Markdown report (below) as your final message and stop (≤10 turns).
+
+## Focus
+Bugs and security first; explicit rule violations second; comment/TODO respect third. High signal only.
+
+## Output format (Markdown report ending in a JSON block)
+Write your final message as Markdown:
+
+## Summary
+One short paragraph.
+
+## Findings
+One bullet per issue: `- [SEVERITY|category|confidence] `path:line` — evidence`. `No findings.` when clean.
+
+## Coverage
+- Files checked: …
+- Commands run: …
+- Limitations: …
+
+Then finish with EXACTLY one fenced JSON block (machine-read by the parent):
+
+```json
+{
+  "status": "ok",
+  "issues": [
+    { "file": "src/x.ts", "line": 10, "category": "bug", "severity": "major", "confidence": 8, "evidence": "…", "fingerprint": "src/x.ts:10:bug:a1b2c3" }
+  ],
+  "summary": "One sentence.",
+  "coverage": { "filesChecked": ["src/x.ts"], "commandsRun": [], "limitations": [] }
+}
+```
+
+Then stop. Do not write any file, do not call any output tool.
+
+## Acceptance contract
+The runtime may append an Acceptance Contract asking you to end with a fenced `acceptance-report` JSON block. Comply: place that fence AFTER the JSON block above, at the very end of your final message.

package/agents/security-review.md

@@ -0,0 +1,45 @@
+---
+name: security-review
+package: pi-review
+description: Security review of introduced lines — injection, authz, secrets, SSRF, path traversal.
+tools: read, grep, bash
+systemPromptMode: replace
+inheritProjectContext: false
+inheritSkills: false
+---
+You are the security reviewer. Find **security issues introduced or worsened by this change**.
+
+## Turn plan
+1. Read the diff file named in your task. If the change profile says docs-only, report `SKIPPED: docs-only` and no findings.
+2. Diff-first; at most **3** extra file reads. Optional `git show`/`log`/`blame -L` for call sites — simple commands only.
+3. Write your Markdown report (format below) as your final message and stop. Target ≤8 turns.
+
+## Checklist
+Injection, missing authn/authz / IDOR, secrets in code/logs, SSRF, path traversal, unsafe deserialization, weak crypto, permissive CORS/cookies.
+
+## Severity
+- `blocker` — exploitable or credential leak
+- `major` — clear weakness likely reachable
+- `minor` — narrow defense-in-depth gap
+
+## Output format (Markdown report)
+Write your final message as Markdown with exactly these sections:
+
+## Summary
+One short paragraph. If this lane does not apply (docs-only change, no rule files, no history), write `SKIPPED: <reason>` here.
+
+## Findings
+One bullet per issue, in this exact shape:
+- [SEVERITY|security|confidence] `path/to/file.ts:123` — evidence quote or precise description
+
+SEVERITY is blocker|major|minor|nit; confidence is 1–10. If you have no findings, write exactly `No findings.`
+
+## Coverage
+- Files checked: …
+- Commands run: …
+- Limitations: …
+
+Finish with that Markdown as your final message. Do not write any file, do not call any output tool.
+
+## Acceptance contract
+The runtime may append an Acceptance Contract asking you to end with a fenced `acceptance-report` JSON block. Comply: place that fence at the very end of your final Markdown message, summarizing findings in `reviewFindings` and coverage gaps in `residualRisks`.

package/index.ts

@@ -0,0 +1,205 @@
+import { existsSync, readFileSync } from "node:fs";
+import type { ExtensionAPI, ExtensionCommandContext } from "@earendil-works/pi-coding-agent";
+
+import { parseReviewArgs } from "./src/cli-args.js";
+import {
+	DEFAULT_CONFIG,
+	configPath,
+	loadConfig,
+	mergeWithDefaults,
+	resolveModel,
+	validateConfig,
+	writeConfig,
+} from "./src/config.js";
+import { prepareRun } from "./src/review-run.js";
+import { registerReviewReportTool } from "./src/tool-wrapper.js";
+import { registerPiReviewRenderer } from "./src/tui-renderer.js";
+
+function parentModelId(ctx: ExtensionCommandContext): string | undefined {
+	const m = ctx.model;
+	if (!m) return undefined;
+	return `${m.provider}/${m.id}`;
+}
+
+export default function (pi: ExtensionAPI) {
+	registerReviewReportTool(pi);
+	registerPiReviewRenderer(pi);
+	pi.registerCommand("review", {
+		description: "启动 AI 并发代码审查 (支持多专家子代理 + 门禁总裁判系统)。--lite = 极速单专家审查。",
+		getArgumentCompletions: (prefix: string) => {
+			const trimmed = prefix.trimStart();
+			const tokens = trimmed.split(/\s+/).filter(Boolean);
+			const last = tokens[tokens.length - 1] ?? "";
+			if (last.startsWith("--")) {
+				return [
+					{ value: "--lite", label: "--lite", description: "极速单专家审查 (无门禁，低延迟省 Token)" },
+					{ value: "--gate-model", label: "--gate-model", description: "指定当前审查的门禁裁判模型" },
+				].filter((o) => o.value.startsWith(last));
+			}
+			return null;
+		},
+		handler: async (args, ctx) => {
+			const notify = (msg: string, level: "info" | "warning" | "error" = "info") => {
+				if (ctx.hasUI) ctx.ui.notify(msg, level);
+				else console.log(`pi-review: ${msg}`);
+			};
+
+			try {
+				const parsed = parseReviewArgs(args);
+				const { config, legacyWarnings } = loadConfig();
+				for (const w of legacyWarnings) notify(`pi-review 提示: ${w}`, "warning");
+
+				pi.sendMessage({
+					customType: "pi-review",
+					content: parsed.input ? `/review ${parsed.input}` : "/review",
+					display: true,
+				});
+
+				if (parsed.noSpawn) {
+					const dryRunText = await renderDryRun(ctx, parsed, config);
+					pi.sendMessage({ customType: "pi-review", content: dryRunText, display: true });
+					return;
+				}
+
+				const prepared = await prepareRun({ cwd: ctx.cwd, input: parsed.input, lite: parsed.lite, gateModel: parsed.gateModel });
+				if (!prepared) {
+					notify("没有检测到需要审查的内容 (未找到修改、PR 或非 Git 仓库)。", "info");
+					return;
+				}
+
+				pi.sendMessage(
+					{
+						customType: "pi-review-directive",
+						content: prepared.directiveText,
+						display: false,
+					},
+					{ triggerTurn: true },
+				);
+			} catch (err) {
+				const message = err instanceof Error ? err.message : String(err);
+				notify(`代码审查启动失败: ${message}`, "error");
+			}
+		},
+	});
+
+	pi.registerCommand("review-config", {
+		description: "编辑代码审查配置 (~/.pi/agent/pi-review.json)",
+		handler: async (_args, ctx) => {
+			const path = configPath();
+			if (!existsSync(path)) {
+				writeConfig(DEFAULT_CONFIG);
+			}
+
+			let raw: string;
+			if (ctx.hasUI) {
+				const current = readFileSync(path, "utf-8");
+				const edited = await ctx.ui.editor("编辑 pi-review 配置文件 (JSON)", current);
+				if (edited === undefined) {
+					ctx.ui.notify("已取消编辑配置。", "info");
+					return;
+				}
+				raw = edited;
+			} else {
+				const editor = process.env.VISUAL ?? process.env.EDITOR;
+				if (!editor) {
+					ctx.ui.notify(`请设置 $EDITOR 或在 TUI 模式下使用。配置文件路径: ${path}`, "warning");
+					return;
+				}
+				await pi.exec(editor, [path], { cwd: ctx.cwd });
+				raw = readFileSync(path, "utf-8");
+			}
+
+			let parsed: unknown;
+			try {
+				parsed = JSON.parse(raw);
+			} catch {
+				ctx.ui.notify("JSON 格式无效 — 配置未保存。", "error");
+				return;
+			}
+
+			const merged = mergeWithDefaults(parsed);
+			const validation = validateConfig(merged);
+			if (!validation.ok) {
+				ctx.ui.notify(`配置存在错误: ${validation.errors.join("; ")}`, "error");
+				return;
+			}
+			writeConfig(merged);
+			ctx.ui.notify("pi-review 配置文件已保存。", "info");
+		},
+	});
+
+	pi.registerCommand("review-agents", {
+		description: "查看当前内置审查专家与模型分配清单",
+		handler: async (_args, ctx) => {
+			const { config } = loadConfig();
+			const parent = parentModelId(ctx);
+			const lines: string[] = ["## 🔍 pi-review 审查专家与门禁状态", ""];
+			for (const r of Object.values(config.reviewers)) {
+				const model = resolveModel(r.model, parent);
+				const status = r.enabled ? "已启用 (enabled)" : "已禁用 (disabled)";
+				lines.push(`- **${r.id}** (${r.label}) — ${status}`);
+				lines.push(`  - 模型: ${model}`);
+				lines.push(`  - 思考深度: ${r.thinking ?? "跟随主会话"}`);
+			}
+			lines.push("");
+			lines.push(`门禁裁判长 (Gate): ${resolveModel(config.gate.model, parent)} · 过滤置信度阈值: ${config.gate.threshold} · 判定策略: ${config.gate.verdictPolicy}`);
+			lines.push(`路由模式: ${config.routing.mode}`);
+			const body = lines.join("\n");
+			pi.sendMessage({ customType: "pi-review-agents", content: body, display: true });
+		},
+	});
+
+	pi.registerCommand("review-show", {
+		description: "重新显示最近一次的代码审查报告",
+		handler: async (_args, ctx) => {
+			let last: { markdown?: string } | null = null;
+			for (const entry of ctx.sessionManager.getBranch()) {
+				if (
+					entry.type === "custom" &&
+					(entry as { customType?: string }).customType === "pi-review" &&
+					(entry as { data?: { markdown?: string } }).data?.markdown
+				) {
+					last = (entry as { data: { markdown?: string } }).data;
+				}
+			}
+			if (!last?.markdown) {
+				ctx.ui.notify("当前会话中未找到任何历史审查报告。", "info");
+				return;
+			}
+			pi.sendMessage({ customType: "pi-review", content: last.markdown, display: true });
+		},
+	});
+}
+
+/** Cheap human summary for `--no-spawn`. */
+async function renderDryRun(
+	ctx: ExtensionCommandContext,
+	parsed: ReturnType<typeof parseReviewArgs>,
+	config: ReturnType<typeof loadConfig>["config"],
+): Promise<string> {
+	const prepared = await prepareRun({
+		cwd: ctx.cwd,
+		input: parsed.input,
+		lite: parsed.lite,
+		gateModel: parsed.gateModel,
+		cleanup: false,
+	});
+	if (!prepared) return "pi-review 试运行: 没有找到需要审查的代码。";
+	const m = prepared.manifest;
+	const gate = config.gate;
+	const lines = [
+		"pi-review 试运行摘要",
+		`审查目标: ${m.targetLabel}`,
+		`模式: ${parsed.lite ? "极速单专家 (无门禁)" : `多专家并发 (${m.targetKind})`}`,
+		`运行 ID: ${m.runId}`,
+		`工作区目录: ${m.workspacePath}`,
+		`Diff 哈希: ${m.diffSha256.slice(0, 16)}…`,
+		`变更文件数: ${m.changedFiles.length}`,
+		`纯文档变更: ${m.docsOnly ? "是" : "否"}`,
+		`包含 Git 历史: ${m.historyAvailable ? "是" : "否"}`,
+		`置信度阈值: ${gate.threshold}`,
+		`门禁裁判: ${gate.enabled ? `已启用 (${resolveModel(gate.model, undefined)})` : "未启用"}`,
+	];
+	if (m.rulePaths.length > 0) lines.push(`规范文件: ${m.rulePaths.join(", ")}`);
+	return lines.join("\n");
+}

package/package.json

@@ -0,0 +1,38 @@
+{
+  "name": "@unifan/pi-review-zh",
+  "version": "1.0.0",
+  "description": "Pi AI 并发代码审查扩展（多专家子代理并发 + 门禁裁判系统，中文增强版）",
+  "type": "module",
+  "license": "MIT",
+  "author": "821869798",
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/821869798/pi-unifan-zh.git"
+  },
+  "keywords": [
+    "pi-package",
+    "pi-extension",
+    "pi-coding-agent",
+    "code-review",
+    "reviewer",
+    "chinese"
+  ],
+  "peerDependencies": {
+    "@earendil-works/pi-ai": ">=0.74.0 <1.0.0",
+    "@earendil-works/pi-coding-agent": ">=0.74.0 <1.0.0",
+    "pi-subagents": ">=0.41.0 <1.0.0"
+  },
+  "pi": {
+    "extensions": [
+      "./index.ts"
+    ],
+    "subagents": {
+      "agents": [
+        "./agents"
+      ]
+    }
+  },
+  "dependencies": {
+    "@sinclair/typebox": "^0.34.0"
+  }
+}

package/reference/README.md

@@ -0,0 +1,20 @@
+# pi-review reference
+
+Design references and version planning. Read these before changing pipeline shape, reviewer roster, or gate behavior so we do not drift from agreed intent.
+
+| Document | Contents |
+|----------|----------|
+| [claude-code-review.md](./claude-code-review.md) | Official Claude `/code-review` pipeline (source of truth: `commands/code-review.md`) |
+| [cursor-review-skills.md](./cursor-review-skills.md) | Cursor Bugbot / Security Review skills (what to borrow, what not to) |
+| [pi-review-roadmap.md](./pi-review-roadmap.md) | pi-review adaptation, phase mapping, version plan |
+| [v0.2-plan.md](./v0.2-plan.md) | **v0.2.0 task list**, acceptance criteria, implementation order |
+| [structured-output.md](./structured-output.md) | Child capture extension env contract |
+
+**Primary upstream (Claude):**  
+https://github.com/anthropics/claude-plugins-official/blob/main/plugins/code-review/commands/code-review.md
+
+**When to update this folder**
+
+- Upstream Claude command changes
+- Reviewer roster or pipeline phase changes in pi-review
+- Default threshold / gate semantics change