fix: 所有问题都建 Issue、合并失败重试、push 与 PR 事件分流

1. 所有问题都建 Issue（原先只有阻断级才建）
   建 Issue 的判据从 blocking 改为 findings：任何等级的问题都创建/更新 Issue，
   等级体现在标题 [OCR][medium] 与标签 medium 上，阻断项在正文标注
   「（阻断合并）」。低等级问题不再丢失，也不会挡住合并。
   无任何发现时才关闭该分支的 Issue。

2. 合并失败自动重试
   Gitea 在算完 PR 可合并性之前会返回 405 Please try again later，
   原先直接放弃，晋级随机失败。现在对 405/409/5xx 按指数退避重试 4 次；
   权限不足、真实冲突等永久失败立即放弃并往 PR 留言说明。
   PR 已被合并（405 already merged）视为成功，幂等收尾。

3. push 与 PR 事件分流（本次新发现的 bug）
   合并路径原先按「是否找到关联 PR」判断，于是一个残留的 test→prd PR
   会让后续 push 被当成 PR 事件，走错分支并跳过晋级，日志还会给出
   「PR targets prd, which is not a checked branch」这种与实际不符的原因。
   现在按事件类型决定：trigger 以 pull_request 开头才走合并 PR 路径，
   push/manual 一律走晋级分支。关联 PR 仅用于评论归属。

验证：
- 直接 push test → 建出 issue #4（[OCR][medium]，标签 code-review,medium）
- 合并重试与幂等分支的判定表全部通过
- 事件分流判定表：push/manual → promote，pull_request.* → merge PR

README.md

@@ -243,6 +243,18 @@ push 时已经审过，重复审会在代码合并后凭空造出 Issue。
 该仓库所有由本服务创建的 open Issue 都会被评论并关闭——代码已经落地，
 上一轮的问题描述的是不再存在的状态。下一个问题会在下一次审查时重新开 Issue。
 
+**审查发现都会建 Issue**，与是否阻断合并无关。等级体现在 Issue 标题与标签上
+（`[OCR][medium] …`，标签 `medium`），阻断项在正文里额外标注「（阻断合并）」。
+这样低等级问题不会丢失，也不会挡住合并。
+
+**合并路径由事件类型决定**：push 事件永远走「晋级分支」，PR 事件才走「合并该 PR」。
+早先按「是否找到关联 PR」来判断，导致一个残留的 `test → prd` PR 会让后续
+push 误判成 PR 事件、跳过晋级。关联 PR 现在只用于评论归属，不参与路径选择。
+
+**合并失败会自动重试**：Gitea 在算出 PR 可合并性之前会返回 `405 Please try again later`，
+这类暂时性失败（405/409/5xx）按指数退避重试 4 次；权限不足、真实冲突等
+永久性失败立即放弃，并在 PR 上留言说明原因。若 PR 已被合并，视为成功。
+
 **审查范围**：PR 事件优先用 PR 的目标分支作为基准；push 事件优先用 webhook 里的 `before`，当 `before` 是新建分支的全零值或缺失时，退回到与管理分支的 merge-base。
 
 **审查背景**：仓库配置里的「审查背景 / 需求」是多行文本，既作为 OCR 的 `--background` 传给模型，也会完整记录到审查历史，方便回溯「当时是按什么需求审的」。

app/lib/gitea.js

@@ -345,6 +345,44 @@ export class GiteaClient {
   }
 }
 
+/**
+ * Gitea answers 405 "Please try again later" while it is still computing a pull
+ * request's mergeability, and 409 for transient lock contention. Both clear up
+ * on their own, so a merge is retried on those before giving up.
+ */
+export function isTransientMergeError(err) {
+  if (!err) return false;
+  const msg = String(err.message || "");
+  // Already merged is a final state, not something to retry.
+  if (/already merged/i.test(msg)) return false;
+  if (err.status === 405 || err.status === 409) return true;
+  if (err.status === 500 || err.status === 502 || err.status === 503 || err.status === 504) return true;
+  return /try again later|please try again|mergeability|not ready/i.test(msg);
+}
+
+/** True when a merge failed only because the pull request is already merged. */
+export function isAlreadyMerged(err) {
+  return /already merged/i.test(String(err?.message || ""));
+}
+
+/** Retry `fn` while it throws a transient error, with exponential backoff. */
+export async function retryTransient(fn, { attempts = 4, baseDelayMs = 1500, onRetry } = {}) {
+  let delay = baseDelayMs;
+  let lastErr;
+  for (let attempt = 1; attempt <= attempts; attempt += 1) {
+    try {
+      return await fn();
+    } catch (err) {
+      lastErr = err;
+      if (!isTransientMergeError(err) || attempt === attempts) throw err;
+      onRetry?.(attempt, err);
+      await new Promise((r) => setTimeout(r, delay));
+      delay = Math.min(delay * 2, 12000);
+    }
+  }
+  throw lastErr;
+}
+
 /** Derive the repository web URL from an API base URL. */
 export function webBaseUrl(apiBaseUrl) {
   return normalizeBaseUrl(apiBaseUrl);

app/lib/review.js

@@ -5,7 +5,9 @@ import { mkdir, rm, writeFile } from "node:fs/promises";
 import { existsSync } from "node:fs";
 import { join } from "node:path";
 
-import { GiteaClient, webBaseUrl, normalizeBaseUrl } from "./gitea.js";
+import {
+  GiteaClient, webBaseUrl, normalizeBaseUrl, retryTransient, isAlreadyMerged,
+} from "./gitea.js";
 import { OcrRunner, parseList, severityAtLeast } from "./ocr.js";
 import { parseUnifiedDiff, pickAnchorLine } from "./diff.js";
 import {
@@ -163,10 +165,33 @@ function renderSummaryBody({ repo, job, review, published, failed, blocking, not
   return lines.join("\n");
 }
 
-function renderIssueBody({ repo, job, blocking, summaryUrl }) {
+const SEVERITY_ORDER = ["critical", "high", "medium", "low"];
+
+/** Highest severity present in a finding list, or null when empty. */
+export function topSeverity(findings) {
+  for (const level of SEVERITY_ORDER) {
+    if (findings.some((f) => String(f.comment?.severity || "").toLowerCase() === level)) return level;
+  }
+  return null;
+}
+
+/** Count findings per severity, highest first, skipping empty levels. */
+function severityCounts(findings) {
+  const counts = new Map();
+  for (const f of findings) {
+    const level = String(f.comment?.severity || "unknown").toLowerCase();
+    counts.set(level, (counts.get(level) || 0) + 1);
+  }
+  return [...counts.entries()].sort(
+    (a, b) => SEVERITY_ORDER.indexOf(a[0]) - SEVERITY_ORDER.indexOf(b[0]),
+  );
+}
+
+function renderIssueBody({ repo, job, findings, blocking, summaryUrl }) {
+  const counts = severityCounts(findings);
   const lines = [
     SUMMARY_MARKER,
-    `自动代码审查在 \`${job.ref_name}\` @ \`${String(job.to_sha).slice(0, 10)}\` 上发现阻断级问题。`,
+    `自动代码审查在 \`${job.ref_name}\` @ \`${String(job.to_sha).slice(0, 10)}\` 上发现问题。`,
     "",
     `- 仓库：\`${repoSlug(repo)}\``,
     `- 分支：\`${job.ref_name}\``,
@@ -174,11 +199,16 @@ function renderIssueBody({ repo, job, blocking, summaryUrl }) {
   ];
   if (job.pr_number) lines.push(`- Pull Request：#${job.pr_number}`);
   if (summaryUrl) lines.push(`- 审查详情：${summaryUrl}`);
-  lines.push("", `### 阻断问题（${blocking.length} 条）`, "");
-  for (const item of blocking) {
+  lines.push(
+    `- 问题分布：${counts.map(([level, n]) => `${level} ${n}`).join(" · ")}`,
+    `- 其中阻断合并：${blocking.length} 条`,
+  );
+  lines.push("", `### 全部问题（${findings.length} 条）`, "");
+  for (const item of findings) {
     const c = item.comment;
     const where = c.start_line ? `${c.path}:${c.start_line}` : c.path;
-    lines.push(`- ${badge(c)}\`${where}\` — ${String(c.content || "").replace(/\s*\n\s*/g, " ").trim()}`);
+    const blocked = blocking.includes(item) ? " **（阻断合并）**" : "";
+    lines.push(`- ${badge(c)}\`${where}\`${blocked} — ${String(c.content || "").replace(/\s*\n\s*/g, " ").trim()}`);
   }
   lines.push("", `<sub>job #${job.id} · 由 gitea-codereview 生成</sub>`);
   return lines.join("\n");
@@ -436,22 +466,24 @@ export class ReviewEngine {
     }
 
     // Issue lifecycle: one open issue per repository+ref, updated in place.
-    // When issue creation is disabled the service still closes any issue it
-    // previously opened once the ref is clean, so stale issues do not linger.
+    // Every finding gets an issue regardless of severity — severity is carried
+    // in the title and body so triage stays possible without losing low-level
+    // findings. Blocking only decides whether the merge is held back.
     let issueNumber = null;
     const labels = parseList(repo.issue_labels);
-    const issueTitle = `[OCR] ${repoSlug(repo)} · ${job.ref_name} 存在阻断级代码问题`;
+    const top = topSeverity(published);
+    const issueTitle = top
+      ? `[OCR][${top}] ${repoSlug(repo)} · ${job.ref_name} 有 ${published.length} 条待处理问题`
+      : `[OCR] ${repoSlug(repo)} · ${job.ref_name} 代码审查`;
+    const detailUrl = prNumber
+      ? `${webBaseUrl(this.config.giteaUrl)}/${repo.owner}/${repo.name}/pulls/${prNumber}`
+      : `${webBaseUrl(this.config.giteaUrl)}/${repo.owner}/${repo.name}/commit/${toSha}`;
     try {
       issueNumber = await this.upsertIssue({
-        client, repo, job, blocking, labels, title: issueTitle,
+        client, repo, job, findings: published, blocking, labels, title: issueTitle,
         createEnabled: Boolean(repo.create_issue),
-        summaryUrl: prNumber
-          ? `${webBaseUrl(this.config.giteaUrl)}/${repo.owner}/${repo.name}/pulls/${prNumber}`
-          : `${webBaseUrl(this.config.giteaUrl)}/${repo.owner}/${repo.name}/commit/${toSha}`,
-        body: renderIssueBody({
-          repo, job, blocking,
-          summaryUrl: prNumber ? `${webBaseUrl(this.config.giteaUrl)}/${repo.owner}/${repo.name}/pulls/${prNumber}` : "",
-        }),
+        summaryUrl: detailUrl,
+        body: renderIssueBody({ repo, job, findings: published, blocking, summaryUrl: detailUrl }),
       });
     } catch (err) {
       appendLog(`issue upsert failed: ${err.message}`);
@@ -597,7 +629,7 @@ export class ReviewEngine {
     return null;
   }
 
-  async upsertIssue({ client, repo, job, blocking, labels, title, body, createEnabled = true }) {
+  async upsertIssue({ client, repo, job, findings, blocking, labels, title, body, createEnabled = true }) {
     // Look up any open issue previously opened by this service for this
     // repository + ref, so reruns update it instead of stacking new issues.
     const openIssues = await client.get(`/api/v1/repos/${repo.owner}/${repo.name}/issues`, {
@@ -610,7 +642,8 @@ export class ReviewEngine {
           && String(i.title || "").includes(`${repoSlug(repo)} · ${job.ref_name}`),
       );
 
-    if (blocking.length === 0) {
+    // No findings at all: the ref is clean, so retire any open issue.
+    if (findings.length === 0) {
       if (existing) {
         await client.createIssueComment(repo.owner, repo.name, existing.number,
           `已在 \`${String(job.to_sha).slice(0, 10)}\` 上复查通过，关闭该 Issue。`);
@@ -626,9 +659,14 @@ export class ReviewEngine {
       return existing?.number ?? null;
     }
 
+    // Tag every issue with its severity so the list can be filtered by level,
+    // on top of whatever labels the repository configured.
+    const top = topSeverity(findings);
+    const labelNames = [...labels];
+    if (top && !labelNames.includes(top)) labelNames.push(top);
     // Gitea's issue API expects label IDs, so resolve names first.
-    const labelIds = labels.length
-      ? await client.ensureLabels(repo.owner, repo.name, labels)
+    const labelIds = labelNames.length
+      ? await client.ensureLabels(repo.owner, repo.name, labelNames)
       : [];
 
     if (existing) {
@@ -660,14 +698,24 @@ export class ReviewEngine {
     reviewIncomplete = false, statusState = "success",
   }) {
     if (!repo.auto_merge) return false;
-    if (!prNumber) {
+
+    // The merge path follows the event that produced this job, not whether a
+    // pull request happens to be linked. A push to `test` must promote the
+    // branch even when an open `test -> prd` pull request already exists;
+    // otherwise that leftover request hijacks the branch's own promotion.
+    const fromPullRequest = String(job.trigger || "").startsWith("pull_request");
+    if (!fromPullRequest) {
       return this.promoteBranch({ client, repo, job, toSha, appendLog, reviewIncomplete, statusState });
     }
+    if (!prNumber) {
+      appendLog("auto-merge skipped: pull request event without a pull request number");
+      return false;
+    }
     // A two-stage flow (feature -> test -> prd) legitimately merges into a
     // checked branch that is not the managed branch, so the gate is "is the
     // target one of the branches we protect", not "is it the managed branch".
     if (job.base_ref && !branchMatches(job.base_ref, repo.check_branches)) {
-      appendLog(`auto-merge skipped: PR targets ${job.base_ref}, which is not a checked branch`);
+      appendLog(`auto-merge skipped: PR targets ${job.base_ref}, which is not in check_branches (${repo.check_branches})`);
       return false;
     }
     if (reviewIncomplete) {
@@ -708,18 +756,31 @@ export class ReviewEngine {
     }
 
     try {
-      await client.mergePullRequest(repo.owner, repo.name, prNumber, {
+      await retryTransient(() => client.mergePullRequest(repo.owner, repo.name, prNumber, {
         style: repo.merge_method || "squash",
         title: pr.title,
         deleteBranch: Boolean(repo.delete_branch),
         headCommitId: toSha,
         mergeWhenChecksSucceed: repo.auto_merge_mode === "when_checks_succeed",
+      }), {
+        attempts: 4,
+        baseDelayMs: 2000,
+        onRetry: (attempt, err) => appendLog(
+          `PR #${prNumber} merge deferred (attempt ${attempt}): ${err.message}`,
+        ),
       });
       appendLog(`auto-merge requested for PR #${prNumber}`);
       await this.closeRepoIssues({ client, repo, appendLog, reason: `PR #${prNumber} 已合并` });
       return true;
     } catch (err) {
+      if (isAlreadyMerged(err)) {
+        appendLog(`PR #${prNumber} was already merged; treating as success`);
+        await this.closeRepoIssues({ client, repo, appendLog, reason: `PR #${prNumber} 已合并` });
+        return true;
+      }
       appendLog(`auto-merge failed: ${err.message}`);
+      await client.createIssueComment(repo.owner, repo.name, prNumber,
+        `自动合并失败：${err.message}\n\n审查提交：\`${toSha}\`。可稍后重新推送，或在界面手动合并。`).catch(() => {});
       return false;
     }
   }
@@ -858,12 +919,18 @@ export class ReviewEngine {
       // branches diverge a little more on every promotion and eventually
       // conflict. A real merge keeps the shared ancestry, so the next
       // promotion only carries the new commits.
-      await client.mergePullRequest(repo.owner, repo.name, index, {
+      await retryTransient(() => client.mergePullRequest(repo.owner, repo.name, index, {
         style: "merge",
         title: pr.title,
         deleteBranch: false,
         headCommitId: toSha,
         mergeWhenChecksSucceed: repo.auto_merge_mode === "when_checks_succeed",
+      }), {
+        attempts: 4,
+        baseDelayMs: 2000,
+        onRetry: (attempt, err) => appendLog(
+          `promotion PR #${index} merge deferred (attempt ${attempt}): ${err.message}`,
+        ),
       });
       appendLog(`auto-merge requested for promotion PR #${index}`);
       await this.closeRepoIssues({
@@ -872,7 +939,17 @@ export class ReviewEngine {
       });
       return true;
     } catch (err) {
+      if (isAlreadyMerged(err)) {
+        appendLog(`promotion PR #${index} was already merged; treating as success`);
+        await this.closeRepoIssues({
+          client, repo, appendLog,
+          reason: `${job.ref_name} 已合并到 ${target}（PR #${index}）`,
+        });
+        return true;
+      }
       appendLog(`auto-merge failed: promotion PR #${index}: ${err.message}`);
+      await client.createIssueComment(repo.owner, repo.name, index,
+        `自动合并失败：${err.message}\n\n审查提交：\`${toSha}\`。可稍后重新推送，或在界面手动合并。`).catch(() => {});
       return false;
     }
   }

app/server.js

@@ -219,14 +219,18 @@ async function handlePush(payload, cfg) {
     }
   }
 
+  // A push is always handled as a push, never as a pull request. An open pull
+  // request whose head happens to be this branch is only linked for comment
+  // routing; letting it decide the merge path would make a leftover promotion
+  // pull request hijack the branch's own promotion logic.
   const scoped = repo.review_scope === "pr";
-  const pr = scoped ? null : await findOpenPullRequestForRef(cfg, repo, refName, toSha);
-  if (scoped && !pr) {
+  const linkedPr = scoped ? null : await findOpenPullRequestForRef(cfg, repo, refName, toSha);
+  if (scoped && !linkedPr) {
     return { queued: 0, reason: "review_scope=pr and no open pull request" };
   }
-  // Base the review on what the push would merge into: the open PR's target
-  // when there is one, otherwise the managed branch.
-  const baseRef = pr?.base?.ref || repo.managed_branch;
+  // Compare against what this branch merges into: its linked pull request's
+  // target when it has one, otherwise the branch it promotes into.
+  const baseRef = linkedPr?.base?.ref || repo.managed_branch;
 
   if (findJobBySha(db, repo.id, toSha)) {
     return { queued: 0, reason: `commit ${toSha.slice(0, 10)} already queued or running` };
@@ -240,7 +244,7 @@ async function handlePush(payload, cfg) {
     baseRef,
     fromSha: before,
     toSha,
-    prNumber: pr?.number ?? null,
+    prNumber: linkedPr?.number ?? null,
   });
   logger.info(`queued job #${jobId} for ${owner}/${name} ${refName}@${toSha.slice(0, 10)}`);
   return { queued: 1, jobId };