diff --git a/.github/workflows/pr-agent.yml b/.github/workflows/pr-agent.yml
index 63472d98..0fb35e77 100644
--- a/.github/workflows/pr-agent.yml
+++ b/.github/workflows/pr-agent.yml
@@ -2,8 +2,7 @@ name: PR-Agent
 
 on:
   pull_request_target:
-    # PR-Agent 通过 base repo 上下文读取 PR diff 并发布 Review，不 checkout 或执行 PR 分支代码。
-    # pull_request_target 允许 fork PR 使用仓库 secrets，因此 workflow 只运行固定 digest 的 PR-Agent 容器。
+    # Fork 审查需要目标仓库凭据；该 job 仅通过 GitHub API 读取 PR 内容，不 checkout 或执行 PR 分支代码。
     types:
       - opened
       - reopened
@@ -11,17 +10,13 @@ on:
       - review_requested
       - synchronize
   issue_comment:
-    # 手动命令只在 PR 评论中有意义；编辑后的命令也可重新触发。
     types:
       - created
       - edited
 
 permissions:
-  # 读取仓库内容和 PR diff。
   contents: read
-  # 更新 PR 描述、发布 PR Review 或修改 PR 相关元数据。
   pull-requests: write
-  # PR 评论在 GitHub API 中属于 issue comments，手动问答需要该权限。
   issues: write
 
 jobs:
@@ -29,21 +24,11 @@ jobs:
     if: >-
       github.event.sender.type != 'Bot' &&
       (
-        (
-          github.event_name == 'pull_request_target'
-        ) ||
+        github.event_name == 'pull_request_target' ||
         (
           github.event_name == 'issue_comment' &&
           github.event.issue.pull_request != null &&
-          contains(fromJSON('["OWNER", "MEMBER", "COLLABORATOR", "CONTRIBUTOR", "FIRST_TIME_CONTRIBUTOR"]'), github.event.comment.author_association) &&
-          (
-            github.event.comment.body == '/describe' ||
-            startsWith(github.event.comment.body, '/describe ') ||
-            github.event.comment.body == '/review' ||
-            startsWith(github.event.comment.body, '/review ') ||
-            github.event.comment.body == '/ask' ||
-            startsWith(github.event.comment.body, '/ask ')
-          )
+          contains(fromJSON('["OWNER", "MEMBER", "COLLABORATOR", "CONTRIBUTOR", "FIRST_TIME_CONTRIBUTOR"]'), github.event.comment.author_association)
         )
       )
     concurrency:
@@ -53,484 +38,15 @@ jobs:
     timeout-minutes: 20
 
     steps:
-      - name: Capture PR context
-        id: pr_context
-        env:
-          GH_TOKEN: ${{ github.token }}
-          REPO: ${{ github.repository }}
-          PR_NUMBER: ${{ github.event.pull_request.number || github.event.issue.number }}
-        run: |
-          set -euo pipefail
-          pull="$(mktemp)"
-          gh api "repos/${REPO}/pulls/${PR_NUMBER}" > "${pull}"
-          python3 - "${pull}" >> "${GITHUB_OUTPUT}" <<'PY'
-          import json
-          import re
-          import sys
-          from pathlib import Path
-
-          pull = json.loads(Path(sys.argv[1]).read_text(encoding="utf-8"))
-          title = pull.get("title") or ""
-          body = pull.get("body") or ""
-          labels = {item.get("name", "") for item in pull.get("labels") or []}
-          skip_pr_agent = "true" if "skip pr-agent" in labels or re.search(r"^(?:\[Auto\]|Auto)", title) else "false"
-          body = re.sub(
-              r"<!-- pr-agent-summary:start -->.*?<!-- pr-agent-summary:end -->",
-              " ",
-              body,
-              flags=re.DOTALL,
-          )
-          text = f"{title}\n{body}"
-          cjk_count = len(re.findall(r"[\u4e00-\u9fff]", text))
-          latin_words = len(re.findall(r"\b[A-Za-z][A-Za-z]{2,}\b", text))
-          if cjk_count >= 4 or latin_words < 8:
-              response_language = "zh-CN"
-              summary_heading = "PR-Agent 摘要"
-          else:
-              response_language = "en-US"
-              summary_heading = "PR-Agent Summary"
-          print(f"head_sha={pull['head']['sha']}")
-          print(f"changed_files={pull.get('changed_files') or 0}")
-          print(f"response_language={response_language}")
-          print(f"summary_heading={summary_heading}")
-          print(f"skip_pr_agent={skip_pr_agent}")
-          PY
-
-      - name: Prepare PR-Agent description markers
-        if: >-
-          steps.pr_context.outputs.skip_pr_agent != 'true' &&
-          (
-            github.event_name == 'pull_request_target' ||
-            github.event.comment.body == '/describe' ||
-            startsWith(github.event.comment.body, '/describe ')
-          )
-        env:
-          GH_TOKEN: ${{ github.token }}
-          REPO: ${{ github.repository }}
-          PR_NUMBER: ${{ github.event.pull_request.number || github.event.issue.number }}
-          SUMMARY_HEADING: ${{ steps.pr_context.outputs.summary_heading }}
-          CHANGED_FILES: ${{ steps.pr_context.outputs.changed_files }}
-        run: |
-          set -euo pipefail
-          payload="$(mktemp)"
-          python3 - "${payload}" <<'PY'
-          import json
-          import os
-          import re
-          import subprocess
-          import sys
-
-          pull = json.loads(subprocess.check_output(
-              ["gh", "api", f"repos/{os.environ['REPO']}/pulls/{os.environ['PR_NUMBER']}"],
-              text=True,
-          ))
-          body = pull.get("body") or ""
-          start = "<!-- pr-agent-summary:start -->"
-          end = "<!-- pr-agent-summary:end -->"
-          heading = os.environ.get("SUMMARY_HEADING") or "PR-Agent 摘要"
-          changed_files = int(os.environ.get("CHANGED_FILES") or 0)
-          block = f"## {heading}\n\n{start}\npr_agent:summary\n{end}"
-          start_index = body.find(start)
-          end_index = body.find(end, start_index + len(start)) if start_index >= 0 else -1
-          owned_block = re.compile(
-              r"(?ims)^##\s+(?:PR-Agent\s+摘要|PR-Agent\s+Summary)\s*\n\s*"
-              r"<!-- pr-agent-summary:start -->.*?<!-- pr-agent-summary:end -->\s*"
-          )
-          if changed_files == 0:
-              updated = owned_block.sub("", body)
-              if updated != body:
-                  updated = updated.rstrip()
-          elif start_index >= 0 and end_index >= 0:
-              body = re.sub(
-                  r"(?im)^##\s+(PR-Agent\s+摘要|PR-Agent\s+Summary)\s*\n\s*(?=<!-- pr-agent-summary:start -->)",
-                  f"## {heading}\n\n",
-                  body,
-              )
-              start_index = body.find(start)
-              end_index = body.find(end, start_index + len(start))
-              updated = body[:start_index] + f"{start}\npr_agent:summary\n{end}" + body[end_index + len(end):]
-          elif start_index >= 0 or end in body:
-              updated = body
-          elif body.strip():
-              updated = f"{body.rstrip()}\n\n{block}\n"
-          else:
-              updated = f"{block}\n"
-          if updated != body:
-              with open(sys.argv[1], "w", encoding="utf-8") as handle:
-                  json.dump({"body": updated}, handle, ensure_ascii=False)
-          PY
-          if [ -s "${payload}" ]; then
-            gh api --method PATCH "repos/${REPO}/pulls/${PR_NUMBER}" --input "${payload}" >/dev/null
-          fi
-
-      - name: Update PR description
-        if: >-
-          steps.pr_context.outputs.skip_pr_agent != 'true' &&
-          steps.pr_context.outputs.changed_files != '0' &&
-          (
-            github.event_name == 'pull_request_target' ||
-            github.event.comment.body == '/describe' ||
-            startsWith(github.event.comment.body, '/describe ')
-          )
-        uses: docker://pragent/pr-agent:0.39.0-github_action@sha256:b253845caa8c7ff5ce8be78f32996647982bdd4890826a962b78eff2e385a825
+      - name: Run PR Review
+        uses: docker://ghcr.io/infinitypacer/pr-review-runner:latest
         env:
           GITHUB_TOKEN: ${{ github.token }}
           OPENAI_KEY: ${{ secrets.OPENAI_KEY }}
           OPENAI.API_BASE: ${{ secrets.OPENAI_API_BASE }}
-          config.model: 'gpt-5.6-terra'
-          config.fallback_models: '["gpt-5.5", "gpt-5.4"]'
-          config.custom_model_max_tokens: '1050000'
-          config.reasoning_effort: 'medium'
-          config.ai_timeout: '900'
-          config.response_language: ${{ steps.pr_context.outputs.response_language }}
-          config.large_patch_policy: 'clip'
-          config.ignore_pr_title: '["^\\[Auto\\]", "^Auto"]'
-          config.ignore_pr_labels: '["skip pr-agent"]'
-          github_action_config.auto_review: 'false'
-          github_action_config.auto_describe: 'true'
-          github_action_config.auto_improve: 'false'
-          github_action_config.pr_actions: '["opened", "reopened", "ready_for_review", "review_requested"]'
-          github_action_config.handle_push_trigger: 'true'
-          github_action_config.push_commands: '["/describe"]'
-          pr_description.generate_ai_title: 'false'
-          pr_description.publish_labels: 'false'
-          pr_description.publish_description_as_comment: 'false'
-          pr_description.publish_description_as_comment_persistent: 'false'
-          pr_description.enable_pr_diagram: 'false'
-          pr_description.enable_pr_type: 'false'
-          pr_description.enable_help_text: 'false'
-          pr_description.enable_help_comment: 'false'
-          pr_description.enable_semantic_files_types: 'false'
-          pr_description.collapsible_file_list: 'adaptive'
-          pr_description.add_original_user_description: 'true'
-          pr_description.use_description_markers: 'true'
-          pr_description.final_update_message: 'false'
-          pr_description.extra_instructions: |
-            Match the configured response language.
-            Summarize the change goal, key implementation details, compatibility impact, tests, and notable risks.
-            Use 2-4 bullets for small pull requests and 4-8 bullets for larger changes.
-            Avoid file lists and local command transcripts.
-
-      - name: Remove unfilled PR-Agent description marker
-        if: >-
-          always() &&
-          steps.pr_context.outputs.skip_pr_agent != 'true' &&
-          (
-            github.event_name == 'pull_request_target' ||
-            github.event.comment.body == '/describe' ||
-            startsWith(github.event.comment.body, '/describe ')
-          )
-        env:
-          GH_TOKEN: ${{ github.token }}
-          REPO: ${{ github.repository }}
-          PR_NUMBER: ${{ github.event.pull_request.number || github.event.issue.number }}
-        run: |
-          set -euo pipefail
-          payload="$(mktemp)"
-          python3 - "${payload}" <<'PY'
-          import json
-          import os
-          import re
-          import subprocess
-          import sys
-
-          pull = json.loads(subprocess.check_output(
-              ["gh", "api", f"repos/{os.environ['REPO']}/pulls/{os.environ['PR_NUMBER']}"],
-              text=True,
-          ))
-          body = pull.get("body") or ""
-          placeholder = "<!-- pr-agent-summary:start -->\npr_agent:summary\n<!-- pr-agent-summary:end -->"
-          owned_block = re.compile(
-              r"(?ims)^##\s+(?:PR-Agent\s+摘要|PR-Agent\s+Summary)\s*\n\s*"
-              r"<!-- pr-agent-summary:start -->\s*pr_agent:summary\s*<!-- pr-agent-summary:end -->\s*"
-          )
-          if placeholder in body:
-              updated = owned_block.sub("", body).rstrip()
-              if updated != body:
-                  with open(sys.argv[1], "w", encoding="utf-8") as handle:
-                      json.dump({"body": updated}, handle, ensure_ascii=False)
-          PY
-          if [ -s "${payload}" ]; then
-            gh api --method PATCH "repos/${REPO}/pulls/${PR_NUMBER}" --input "${payload}" >/dev/null
-          fi
-
-      - name: Analyze PR review
-        id: review_analysis
-        if: >-
-          steps.pr_context.outputs.skip_pr_agent != 'true' &&
-          (
-            github.event_name == 'pull_request_target' ||
-            github.event.comment.body == '/review' ||
-            startsWith(github.event.comment.body, '/review ')
-          )
-        uses: docker://pragent/pr-agent:0.39.0-github_action@sha256:b253845caa8c7ff5ce8be78f32996647982bdd4890826a962b78eff2e385a825
-        env:
-          GITHUB_TOKEN: ${{ github.token }}
-          OPENAI_KEY: ${{ secrets.OPENAI_KEY }}
-          OPENAI.API_BASE: ${{ secrets.OPENAI_API_BASE }}
-          config.model: ${{ github.event_name == 'issue_comment' && 'gpt-5.6-sol' || 'gpt-5.6-terra' }}
-          config.fallback_models: '["gpt-5.5", "gpt-5.4"]'
-          config.custom_model_max_tokens: '1050000'
-          config.reasoning_effort: 'xhigh'
-          config.ai_timeout: '900'
-          config.response_language: ${{ steps.pr_context.outputs.response_language }}
-          config.large_patch_policy: 'clip'
-          config.ignore_pr_title: '["^\\[Auto\\]", "^Auto"]'
-          config.ignore_pr_labels: '["skip pr-agent"]'
-          config.publish_output: 'false'
-          github_action_config.auto_review: 'true'
-          github_action_config.auto_describe: 'false'
-          github_action_config.auto_improve: 'false'
-          github_action_config.pr_actions: '["opened", "reopened", "ready_for_review", "review_requested"]'
-          github_action_config.handle_push_trigger: 'true'
-          github_action_config.push_commands: '["/review"]'
-          github_action_config.enable_output: 'true'
-          pr_reviewer.num_max_findings: '4'
-          pr_reviewer.require_score_review: 'false'
-          pr_reviewer.require_tests_review: 'false'
-          pr_reviewer.require_security_review: 'false'
-          pr_reviewer.require_estimate_effort_to_review: 'false'
-          pr_reviewer.require_estimate_contribution_time_cost: 'false'
-          pr_reviewer.require_can_be_split_review: 'false'
-          pr_reviewer.require_todo_scan: 'false'
-          pr_reviewer.require_ticket_analysis_review: 'false'
-          pr_reviewer.enable_review_labels_effort: 'false'
-          pr_reviewer.enable_review_labels_security: 'false'
-          pr_reviewer.extra_instructions: |
-            Return key_issues_to_review only for concrete behavior defects introduced by this pull request.
-            Each finding must identify the affected behavior, a reachable trigger, and the existing contract or invariant it violates.
-            Use issue_content to state the smallest correction boundary, not a code patch.
-            Do not report style preferences, comments, refactors, architecture alternatives, speculative races, extra hardening, optional tests, or hypothetical concerns.
-            Return no findings when the evidence is incomplete.
-
-      - name: Publish review comments and summary
-        if: >-
-          steps.pr_context.outputs.skip_pr_agent != 'true' &&
-          steps.review_analysis.outcome == 'success' &&
-          (
-            github.event_name == 'pull_request_target' ||
-            github.event.comment.body == '/review' ||
-            startsWith(github.event.comment.body, '/review ')
-          )
-        env:
-          GH_TOKEN: ${{ github.token }}
-          REPO: ${{ github.repository }}
-          PR_NUMBER: ${{ github.event.pull_request.number || github.event.issue.number }}
-          REVIEWED_HEAD_SHA: ${{ steps.pr_context.outputs.head_sha }}
-          CHANGED_FILES: ${{ steps.pr_context.outputs.changed_files }}
-          RESPONSE_LANGUAGE: ${{ steps.pr_context.outputs.response_language }}
-          REVIEW_JSON: ${{ steps.review_analysis.outputs.review }}
-        run: |
-          set -euo pipefail
-          current_head_sha="$(gh api "repos/${REPO}/pulls/${PR_NUMBER}" --jq '.head.sha')"
-          if [ "${current_head_sha}" != "${REVIEWED_HEAD_SHA}" ]; then
-            echo "PR head changed during analysis; skip stale review publication."
-            exit 0
-          fi
-
-          review_payload="$(mktemp)"
-          python3 - "${review_payload}" <<'PY'
-          import hashlib
-          import json
-          import os
-          import re
-          import subprocess
-          import sys
-          from urllib.parse import quote
-
-          review_raw = os.environ.get("REVIEW_JSON") or "{}"
-          review = json.loads(review_raw)
-          if not review_raw.strip() or review == {}:
-              if int(os.environ.get("CHANGED_FILES") or 0):
-                  raise SystemExit("Review analysis produced no structured output for a non-empty PR.")
-              review = {}
-
-          repo = os.environ["REPO"]
-          number = os.environ["PR_NUMBER"]
-          head_sha = os.environ["REVIEWED_HEAD_SHA"]
-          language = os.environ.get("RESPONSE_LANGUAGE") or "zh-CN"
-
-          def paged(endpoint):
-              result = json.loads(subprocess.check_output(
-                  ["gh", "api", "--paginate", "--slurp", endpoint], text=True
-              ))
-              if result and all(isinstance(page, list) for page in result):
-                  return [item for page in result for item in page]
-              return result
-
-          files = paged(f"repos/{repo}/pulls/{number}/files?per_page=100")
-          comments = paged(f"repos/{repo}/pulls/{number}/comments?per_page=100")
-          reviews = paged(f"repos/{repo}/pulls/{number}/reviews?per_page=100")
-
-          hunk_pattern = re.compile(r"^@@ -\d+(?:,\d+)? \+(\d+)(?:,\d+)? @@")
-          changed_lines = {}
-          for file_data in files:
-              path = str(file_data.get("filename") or "")
-              line = None
-              lines = set()
-              for patch_line in (file_data.get("patch") or "").splitlines():
-                  hunk = hunk_pattern.match(patch_line)
-                  if hunk:
-                      line = int(hunk.group(1))
-                      continue
-                  if line is None or patch_line.startswith("\\"):
-                      continue
-                  if patch_line.startswith("+") and not patch_line.startswith("+++"):
-                      lines.add(line)
-                      line += 1
-                  elif patch_line.startswith("-") and not patch_line.startswith("---"):
-                      continue
-                  else:
-                      line += 1
-              changed_lines[path] = lines
-
-          def fingerprint(path, line):
-              normalized = "\n".join((path, str(line)))
-              return hashlib.sha256(normalized.encode("utf-8")).hexdigest()[:16]
-
-          current_fingerprints = set()
-          current_locations = set()
-          marker_pattern = re.compile(r"<!-- pr-agent-review:([0-9a-f]{16}) -->")
-          for comment in comments:
-              if comment.get("user", {}).get("login") != "github-actions[bot]":
-                  continue
-              if comment.get("line") is None:
-                  continue
-              match = marker_pattern.search(str(comment.get("body") or ""))
-              if match:
-                  current_fingerprints.add(match.group(1))
-                  path = str(comment.get("path") or "")
-                  try:
-                      line = int(comment.get("line") or 0)
-                  except (TypeError, ValueError):
-                      line = 0
-                  if path and line > 0:
-                      current_locations.add((path, line))
-
-          def code_url(path, line):
-              return f"https://github.com/{repo}/blob/{head_sha}/{quote(path, safe='/')}#L{line}"
-
-          issues = review.get("key_issues_to_review") or []
-          findings = []
-          seen = set()
-          for issue in issues:
-              if not isinstance(issue, dict):
-                  continue
-              path = str(issue.get("relevant_file") or "").strip()
-              header = str(issue.get("issue_header") or "").strip()
-              content = str(issue.get("issue_content") or "").strip()
-              try:
-                  line = int(issue.get("start_line") or 0)
-              except (TypeError, ValueError):
-                  line = 0
-              if not path or not header or not content or line < 1:
-                  continue
-              finding_key = (path, line, header.lower(), " ".join(content.split()).lower())
-              if finding_key in seen:
-                  continue
-              seen.add(finding_key)
-              findings.append({
-                  "path": path,
-                  "line": line,
-                  "header": header,
-                  "content": content,
-                  "fingerprint": fingerprint(path, line),
-              })
-
-          new_comments = []
-          for finding in findings:
-              if finding["line"] not in changed_lines.get(finding["path"], set()):
-                  continue
-              if finding["fingerprint"] in current_fingerprints or (finding["path"], finding["line"]) in current_locations:
-                  continue
-              new_comments.append({
-                  "path": finding["path"],
-                  "line": finding["line"],
-                  "side": "RIGHT",
-                  "body": "\n".join([
-                      f"<!-- pr-agent-review:{finding['fingerprint']} -->",
-                      f"**{finding['header']}**",
-                      "",
-                      finding["content"],
-                  ]),
-              })
-
-          marker = "<!-- pr-agent-review-summary -->"
-          short_sha = head_sha[:7]
-          commit_url = f"https://github.com/{repo}/commit/{head_sha}"
-          chinese = language == "zh-CN"
-          lines = [marker, "## PR-Agent Code Review", ""]
-          if findings:
-              for finding in findings:
-                  location = f"{finding['path']}:{finding['line']}"
-                  concise = " ".join(finding["content"].split())[:360]
-                  separator = "：" if chinese else ":"
-                  lines.append(f"- [{location}]({code_url(finding['path'], finding['line'])}){separator} **{finding['header']}** - {concise}")
-          elif chinese:
-              lines.append("本次变更无需提出审查意见，暂无其他反馈。")
-          else:
-              lines.append("There are no review comments for the current changes. I have no additional feedback to provide.")
-          lines.extend([
-              "",
-              f"审查提交：[{short_sha}]({commit_url})" if chinese else f"Reviewed commit: [{short_sha}]({commit_url})",
-              "",
-          ])
-          payload = {
-              "body": "\n".join(lines),
-              "commit_id": head_sha,
-              "event": "COMMENT",
-          }
-          if new_comments:
-              payload["comments"] = new_comments
-          has_matching_summary = not new_comments and any(
-              existing.get("user", {}).get("login") == "github-actions[bot]"
-              and existing.get("commit_id") == head_sha
-              and str(existing.get("body") or "") == payload["body"]
-              for existing in reviews
-          )
-          # 同一提交的手工重审仍会完成分析；完全相同的结果不重复发布 Review。
-          if not has_matching_summary:
-              with open(sys.argv[1], "w", encoding="utf-8") as handle:
-                  json.dump(payload, handle, ensure_ascii=False)
-          PY
-
-          latest_head_sha="$(gh api "repos/${REPO}/pulls/${PR_NUMBER}" --jq '.head.sha')"
-          if [ "${latest_head_sha}" != "${REVIEWED_HEAD_SHA}" ]; then
-            echo "PR head changed while rendering review; skip stale review publication."
-            exit 0
-          fi
-
-          if [ -s "${review_payload}" ]; then
-            gh api --method POST "repos/${REPO}/pulls/${PR_NUMBER}/reviews" --input "${review_payload}" >/dev/null
-          fi
-          old_summary_ids="$(gh api --paginate "repos/${REPO}/issues/${PR_NUMBER}/comments?per_page=100" --jq ".[] | select(.user.login == \"github-actions[bot]\" and ((.body | startswith(\"<!-- pr-agent-update-notification -->\")) or (.body | startswith(\"<!-- pr-agent-code-review-summary -->\")) or (.body | startswith(\"<!-- pr-agent-review-summary -->\")) or (.body | startswith(\"<!-- pr-agent-lab:review -->\")))) | .id")"
-          while IFS= read -r comment_id; do
-            [ -z "${comment_id}" ] && continue
-            gh api --method DELETE "repos/${REPO}/issues/comments/${comment_id}" >/dev/null
-          done <<< "${old_summary_ids}"
-
-      - name: Answer PR question
-        if: >-
-          steps.pr_context.outputs.skip_pr_agent != 'true' &&
-          github.event_name == 'issue_comment' &&
-          (
-            github.event.comment.body == '/ask' ||
-            startsWith(github.event.comment.body, '/ask ')
-          )
-        uses: docker://pragent/pr-agent:0.39.0-github_action@sha256:b253845caa8c7ff5ce8be78f32996647982bdd4890826a962b78eff2e385a825
-        env:
-          GITHUB_TOKEN: ${{ github.token }}
-          OPENAI_KEY: ${{ secrets.OPENAI_KEY }}
-          OPENAI.API_BASE: ${{ secrets.OPENAI_API_BASE }}
-          config.model: 'gpt-5.6-terra'
-          config.fallback_models: '["gpt-5.5", "gpt-5.4"]'
-          config.custom_model_max_tokens: '1050000'
-          config.reasoning_effort: 'high'
-          config.ai_timeout: '900'
-          config.response_language: ${{ steps.pr_context.outputs.response_language }}
-          config.large_patch_policy: 'clip'
-          config.ignore_pr_title: '["^\\[Auto\\]", "^Auto"]'
-          config.ignore_pr_labels: '["skip pr-agent"]'
-          github_action_config.auto_review: 'false'
-          github_action_config.auto_describe: 'false'
-          github_action_config.auto_improve: 'false'
+          PRR_AUTO_REVIEW_SCOPE: all
+          PRR_ALLOWED_ASSOCIATIONS: '["OWNER", "MEMBER", "COLLABORATOR", "CONTRIBUTOR", "FIRST_TIME_CONTRIBUTOR"]'
+          PRR_DISABLED_COMMANDS: '["/improve"]'
+          PRR_SKIP_LABEL: skip pr-agent
+          PRR_SKIP_TITLE_PATTERN: '^(?:\[Auto\]|Auto)'
+          config.response_language: zh-CN
diff --git a/app/agent/llm/helper.py b/app/agent/llm/helper.py
index 292708de..04accebc 100644
--- a/app/agent/llm/helper.py
+++ b/app/agent/llm/helper.py
@@ -1058,6 +1058,29 @@ class LLMHelper:
                 http_async_client=_build_httpx_client(llm_proxy, async_client=True),
                 **thinking_kwargs,
             )
+        elif runtime["runtime"] == "bedrock":
+            from langchain_aws import ChatBedrockConverse
+
+            from app.agent.llm.provider import LLMProviderManager
+
+            aws_region = runtime.get("aws_region") or "us-east-1"
+            aws_auth = runtime.get("aws_auth") or {}
+            # Bearer 认证需要跳过 SigV4 签名并注入 Authorization 头，SigV4 认证
+            # 直接以 AK/SK 签名；两种方式统一由 provider 管理器构造 boto3 客户端。
+            bedrock_client = LLMProviderManager().create_bedrock_client(
+                "bedrock-runtime",
+                region=aws_region,
+                credentials=aws_auth,
+                base_url=runtime.get("base_url"),
+                use_proxy=use_proxy,
+                read_timeout=settings.LLM_TOOL_TIMEOUT,
+            )
+            model = ChatBedrockConverse(
+                model_id=model_name,
+                client=bedrock_client,
+                temperature=temperature_value,
+                disable_streaming=not streaming,
+            )
         elif runtime["runtime"] in {"anthropic_compatible", "copilot_anthropic"}:
             from langchain_anthropic import ChatAnthropic
 
@@ -1107,7 +1130,11 @@ class LLMHelper:
         # 优先使用 provider / models.dev 目录中的上下文上限，减少用户手填成本。
         model_profile = getattr(model, "profile", None)
         if model_profile:
-            logger.debug(f"使用LLM模型: {model.model}，Profile: {model.profile}")
+            # ChatBedrockConverse 等模型类没有 model 属性，模型名存放在 model_id。
+            logged_model_name = getattr(model, "model", None) or getattr(
+                model, "model_id", model_name
+            )
+            logger.debug(f"使用LLM模型: {logged_model_name}，Profile: {model_profile}")
         else:
             model_record = runtime.get("model_record") or {}
             model_metadata = runtime.get("model_metadata") or {}
diff --git a/app/agent/llm/provider.py b/app/agent/llm/provider.py
index 70bf98b6..d6afcd7f 100644
--- a/app/agent/llm/provider.py
+++ b/app/agent/llm/provider.py
@@ -7,13 +7,14 @@ import base64
 import copy
 import hashlib
 import json
+import re
 import secrets
 import threading
 import time
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any, Dict, Optional, Tuple
-from urllib.parse import urlencode
+from urllib.parse import urlencode, urlsplit
 
 import aiofiles
 import httpx
@@ -106,6 +107,90 @@ class LLMProviderManager(metaclass=Singleton):
     _MODELS_DEV_BUNDLED_PATH = Path(__file__).with_name("models.json")
     _MODELS_DEV_CACHE_TTL = 7 * 24 * 60 * 60
     _AUTH_SESSION_DONE_RETENTION = 300
+    _BEDROCK_DEFAULT_REGION = "us-east-1"
+    _BEDROCK_API_KEY_PREFIX = "bedrock-api-key-"
+    _BEDROCK_GPT_OSS_BASE_REGIONS = (
+        "ap-northeast-1",
+        "ap-south-1",
+        "ap-southeast-2",
+        "eu-central-1",
+        "eu-north-1",
+        "eu-west-1",
+        "eu-west-2",
+        "sa-east-1",
+        "us-east-1",
+        "us-east-2",
+        "us-west-2",
+    )
+    _BEDROCK_GPT_OSS_SAFEGUARD_REGIONS = (
+        "ap-northeast-1",
+        "ap-south-1",
+        "ap-southeast-2",
+        "eu-west-1",
+        "eu-west-2",
+        "sa-east-1",
+        "us-east-1",
+        "us-east-2",
+        "us-west-2",
+    )
+    _BEDROCK_ON_DEMAND_MODEL_REGIONS = {
+        "openai.gpt-oss-120b-1:0": _BEDROCK_GPT_OSS_BASE_REGIONS,
+        "openai.gpt-oss-20b-1:0": _BEDROCK_GPT_OSS_BASE_REGIONS,
+        "openai.gpt-oss-safeguard-120b": _BEDROCK_GPT_OSS_SAFEGUARD_REGIONS,
+        "openai.gpt-oss-safeguard-20b": _BEDROCK_GPT_OSS_SAFEGUARD_REGIONS,
+        "amazon.nova-lite-v1:0": (
+            "ap-northeast-1",
+            "ap-southeast-2",
+            "eu-west-2",
+            "us-east-1",
+            "us-gov-west-1",
+        ),
+        "amazon.nova-micro-v1:0": (
+            "ap-southeast-2",
+            "eu-west-2",
+            "us-east-1",
+            "us-gov-west-1",
+        ),
+        "amazon.nova-pro-v1:0": (
+            "ap-southeast-2",
+            "eu-west-2",
+            "us-east-1",
+            "us-gov-west-1",
+        ),
+        "anthropic.claude-3-5-haiku-20241022-v1:0": (
+            "us-west-2",
+        ),
+        "anthropic.claude-3-5-sonnet-20240620-v1:0": (
+            "ap-northeast-1",
+            "ap-northeast-2",
+            "ap-southeast-1",
+            "eu-central-1",
+            "eu-central-2",
+            "us-east-1",
+            "us-gov-west-1",
+            "us-west-2",
+        ),
+        "anthropic.claude-3-5-sonnet-20241022-v2:0": (
+            "ap-southeast-2",
+            "us-west-2",
+        ),
+        "anthropic.claude-3-7-sonnet-20250219-v1:0": (
+            "eu-west-2",
+            "us-gov-west-1",
+        ),
+        "anthropic.claude-3-haiku-20240307-v1:0": (
+            "ap-northeast-1",
+            "ap-northeast-2",
+            "ap-south-1",
+            "ap-southeast-2",
+            "eu-central-1",
+            "eu-west-1",
+            "eu-west-3",
+            "us-east-1",
+            "us-gov-west-1",
+            "us-west-2",
+        ),
+    }
     _CHATGPT_CLIENT_ID = "app_EMoamEEZ73f0CkXaXp7hrann"
     _CHATGPT_ISSUER = "https://auth.openai.com"
     _CHATGPT_CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex"
@@ -367,6 +452,50 @@ class LLMProviderManager(metaclass=Singleton):
                 api_key_hint="填写 Anthropic API Key。",
                 description="Anthropic Claude 官方端点。",
             ),
+            ProviderSpec(
+                id="amazon-bedrock",
+                name="Amazon Bedrock",
+                runtime="bedrock",
+                models_dev_provider_id="amazon-bedrock",
+                default_base_url="https://bedrock-runtime.us-east-1.amazonaws.com",
+                base_url_presets=(
+                    url_preset(
+                        id="bedrock-us-east-1",
+                        label="美东（弗吉尼亚北部）us-east-1",
+                        value="https://bedrock-runtime.us-east-1.amazonaws.com",
+                    ),
+                    url_preset(
+                        id="bedrock-us-west-2",
+                        label="美西（俄勒冈）us-west-2",
+                        value="https://bedrock-runtime.us-west-2.amazonaws.com",
+                    ),
+                    url_preset(
+                        id="bedrock-eu-central-1",
+                        label="欧洲（法兰克福）eu-central-1",
+                        value="https://bedrock-runtime.eu-central-1.amazonaws.com",
+                    ),
+                    url_preset(
+                        id="bedrock-ap-northeast-1",
+                        label="亚太（东京）ap-northeast-1",
+                        value="https://bedrock-runtime.ap-northeast-1.amazonaws.com",
+                    ),
+                    url_preset(
+                        id="bedrock-ap-southeast-1",
+                        label="亚太（新加坡）ap-southeast-1",
+                        value="https://bedrock-runtime.ap-southeast-1.amazonaws.com",
+                    ),
+                ),
+                base_url_editable=True,
+                api_key_label="Bedrock API Key / AK:SK",
+                api_key_hint=(
+                    "支持两种认证方式：填写 Amazon Bedrock API Key（bedrock-api-key- 开头，"
+                    "Bearer 认证）；或填写 Access Key ID:Secret Access Key（可选追加 :Session Token，"
+                    "SigV4 认证）。Base URL 决定 AWS Region。"
+                ),
+                model_list_strategy="bedrock",
+                description="Amazon Bedrock 托管模型服务，支持 Bedrock API Key 与 AK/SK 双认证。",
+                sort_order=35,
+            ),
             ProviderSpec(
                 id="deepseek",
                 name="DeepSeek",
@@ -1743,6 +1872,112 @@ class LLMProviderManager(metaclass=Singleton):
             return normalized[:-3]
         return normalized
 
+    @classmethod
+    def _extract_bedrock_region(cls, base_url: Optional[str]) -> str:
+        """
+        从 Bedrock 运行时端点 URL 中提取 AWS Region
+
+        兼容标准端点、FIPS 端点与 PrivateLink（VPCE）端点等主机名形态，
+        从中识别 Region 段。
+
+        :param base_url: 形如 https://bedrock-runtime.us-east-1.amazonaws.com 的端点地址
+        :return: 提取到的 Region，无法识别时回退 us-east-1
+        """
+        hostname = urlsplit((base_url or "").strip().lower()).hostname or ""
+        match = re.search(
+            r"(?:^|\.)(?:bedrock(?:-runtime)?(?:-fips)?)"
+            r"\.([a-z0-9-]+-\d+)(?:\.|$)",
+            hostname,
+        )
+        if match:
+            return match.group(1)
+        return cls._BEDROCK_DEFAULT_REGION
+
+    # Inference Profile 的地理前缀与可用 Region 的对应关系，用于降级目录按
+    # 当前 Region 过滤掉不可调用的 Profile 条目。
+    _BEDROCK_GEO_PREFIXES: dict[str, tuple[str, ...]] = {
+        "us": ("us-east-", "us-west-"),
+        "eu": ("eu-",),
+        "apac": ("ap-",),
+        "au": ("ap-southeast-2", "ap-southeast-4"),
+        "jp": ("ap-northeast-1", "ap-northeast-3"),
+        "ca": ("ca-",),
+    }
+    _BEDROCK_NON_COMMERCIAL_REGION_PREFIXES = (
+        "cn-",
+        "eu-isoe-",
+        "us-gov-",
+        "us-iso-",
+        "us-isob-",
+        "us-isof-",
+    )
+
+    @classmethod
+    def _bedrock_model_matches_region(cls, model_id: str, region: str) -> bool:
+        """
+        判断目录中的模型 ID 在指定 Region 是否可调用
+
+        models.dev 目录同时收录裸模型 ID（直连调用）与带地理前缀的
+        Inference Profile ID（us./eu./apac./global. 等）。带前缀的条目只在
+        对应地理分区和 AWS 分区的 Region 可用；global Profile 仅允许商业
+        AWS 分区。裸 ID 仅在明确记录的 ON_DEMAND Region 可用，未知条目
+        按不可直连处理。
+
+        :param model_id: 目录中的模型 ID
+        :param region: 当前 Base URL 对应的 AWS Region
+        :return: 该模型在当前 Region 可调用时返回 True
+        """
+        prefix = model_id.split(".", 1)[0]
+        if prefix == "global":
+            return not region.startswith(cls._BEDROCK_NON_COMMERCIAL_REGION_PREFIXES)
+        region_prefixes = cls._BEDROCK_GEO_PREFIXES.get(prefix)
+        if region_prefixes is not None:
+            return (
+                    not region.startswith(cls._BEDROCK_NON_COMMERCIAL_REGION_PREFIXES)
+                    and region.startswith(region_prefixes)
+            )
+        on_demand_regions = cls._BEDROCK_ON_DEMAND_MODEL_REGIONS.get(model_id)
+        return on_demand_regions is not None and region in on_demand_regions
+
+    @classmethod
+    def _parse_bedrock_credentials(cls, api_key: Optional[str]) -> dict[str, Any]:
+        """
+        解析 Bedrock 凭证字符串，识别 Bearer 与 SigV4 两种认证方式
+
+        - Bedrock API Key（bedrock-api-key- 开头的长期 Key，或控制台生成的短期
+          Token）走 Bearer 认证；
+        - `AccessKeyId:SecretAccessKey` 或 `AccessKeyId:SecretAccessKey:SessionToken`
+          走 SigV4 认证，AWS Access Key ID 均以 "AKIA"/"ASIA" 开头。
+
+        :param api_key: 用户在 API Key 输入框填写的凭证内容
+        :return: 含 auth_scheme 及对应凭证字段的字典
+        """
+        normalized = str(api_key or "").strip()
+        if not normalized:
+            raise LLMProviderAuthError(
+                "Amazon Bedrock 需要填写 Bedrock API Key 或 Access Key ID:Secret Access Key"
+            )
+
+        if not normalized.startswith(cls._BEDROCK_API_KEY_PREFIX):
+            parts = [part.strip() for part in normalized.split(":")]
+            if len(parts) in {2, 3} and all(parts):
+                credentials = {
+                    "auth_scheme": "sigv4",
+                    "access_key_id": parts[0],
+                    "secret_access_key": parts[1],
+                }
+                if len(parts) == 3:
+                    credentials["session_token"] = parts[2]
+                return credentials
+            if ":" in normalized:
+                raise LLMProviderAuthError(
+                    "Amazon Bedrock AK/SK 凭证格式不正确，"
+                    "请按 AccessKeyId:SecretAccessKey 或 "
+                    "AccessKeyId:SecretAccessKey:SessionToken 填写"
+                )
+
+        return {"auth_scheme": "bearer", "bearer_token": normalized}
+
     async def _list_models_from_google(
             self,
             api_key: str,
@@ -1857,6 +2092,235 @@ class LLMProviderManager(metaclass=Singleton):
             )
         return sorted(results, key=lambda item: item["name"].lower())
 
+    def _build_bedrock_boto3_config(
+            self,
+            use_proxy: Optional[bool] = None,
+    ) -> Any:
+        """
+        构造 Bedrock boto3 客户端配置，统一超时、重试与代理策略
+
+        :param use_proxy: 是否使用系统代理，None 时读取 LLM_USE_PROXY 配置
+        :return: botocore Config 实例
+        """
+        from botocore.config import Config
+
+        should_use_proxy = settings.LLM_USE_PROXY if use_proxy is None else use_proxy
+        proxies = None
+        if should_use_proxy and settings.PROXY_HOST:
+            proxies = {"http": settings.PROXY_HOST, "https": settings.PROXY_HOST}
+        return Config(
+            connect_timeout=10,
+            read_timeout=60,
+            retries={"max_attempts": 3, "mode": "standard"},
+            proxies=proxies,
+        )
+
+    @staticmethod
+    def _bedrock_endpoint_url(
+            service_name: str, base_url: Optional[str]
+    ) -> Optional[str]:
+        """
+        解析应传给 boto3 客户端的自定义端点 URL
+
+        标准公有端点交由 boto3 按 Region 自行推导；用户填写 PrivateLink、
+        FIPS 等非标准端点时才显式透传，保证所选网络路径实际生效。
+
+        :param service_name: boto3 服务名（bedrock 或 bedrock-runtime）
+        :param base_url: 用户配置的 Base URL
+        :return: 需要显式指定端点时返回 URL，否则返回 None
+        """
+        normalized = (base_url or "").strip().rstrip("/")
+        if not normalized:
+            return None
+        if re.fullmatch(
+                rf"https://{service_name}\.[a-z0-9-]+\.amazonaws\.com",
+                normalized,
+        ):
+            return None
+        return normalized
+
+    def create_bedrock_client(
+            self,
+            service_name: str,
+            region: str,
+            credentials: dict[str, Any],
+            base_url: Optional[str] = None,
+            use_proxy: Optional[bool] = None,
+            read_timeout: Optional[int] = None,
+    ) -> Any:
+        """
+        按解析后的凭证创建 Bedrock boto3 客户端，Bearer 方式注入 Authorization 头
+
+        :param service_name: boto3 服务名（bedrock 或 bedrock-runtime）
+        :param region: AWS Region
+        :param credentials: `_parse_bedrock_credentials` 的解析结果
+        :param base_url: 用户配置的 Base URL，非标准端点（PrivateLink/FIPS 等）时透传给 boto3
+        :param use_proxy: 是否使用系统代理
+        :param read_timeout: 读取超时秒数，None 时使用默认值
+        :return: boto3 客户端实例
+        """
+        import boto3
+        from botocore import UNSIGNED
+
+        config = self._build_bedrock_boto3_config(use_proxy)
+        if read_timeout:
+            config = config.merge(type(config)(read_timeout=read_timeout))
+        endpoint_kwargs: dict[str, Any] = {}
+        endpoint_url = self._bedrock_endpoint_url(service_name, base_url)
+        if endpoint_url:
+            endpoint_kwargs["endpoint_url"] = endpoint_url
+
+        if credentials["auth_scheme"] == "sigv4":
+            return boto3.client(
+                service_name,
+                region_name=region,
+                aws_access_key_id=credentials["access_key_id"],
+                aws_secret_access_key=credentials["secret_access_key"],
+                aws_session_token=credentials.get("session_token"),
+                config=config,
+                **endpoint_kwargs,
+            )
+
+        # Bearer 认证：以 UNSIGNED 跳过 SigV4 签名，再把 API Key 注入 Authorization 头。
+        bearer_token = credentials["bearer_token"]
+        config = config.merge(type(config)(signature_version=UNSIGNED))
+        client = boto3.client(
+            service_name,
+            region_name=region,
+            aws_access_key_id="unsigned",
+            aws_secret_access_key="unsigned",
+            config=config,
+            **endpoint_kwargs,
+        )
+
+        def _inject_bearer(request: Any, **_kwargs: Any) -> None:
+            request.headers["Authorization"] = f"Bearer {bearer_token}"
+
+        client.meta.events.register(
+            f"request-created.{service_name}",
+            _inject_bearer,
+        )
+        return client
+
+    async def _list_models_from_bedrock_fallback(
+            self,
+            region: str,
+            use_proxy: Optional[bool] = None,
+    ) -> list[dict[str, Any]]:
+        """
+        从 models.dev 目录筛选当前 Region 可调用的 Bedrock 模型
+
+        :param region: 当前 Base URL 对应的 AWS Region
+        :param use_proxy: 是否使用系统代理
+        :return: 过滤后的标准化模型记录列表
+        """
+        models = await self._list_models_from_models_dev_only(
+            provider_id="amazon-bedrock",
+            use_proxy=use_proxy,
+        )
+        return [
+            model
+            for model in models
+            if self._bedrock_model_matches_region(model["id"], region)
+        ]
+
+    async def _list_models_from_bedrock(
+            self,
+            api_key: str,
+            base_url: Optional[str],
+            use_proxy: Optional[bool] = None,
+    ) -> list[dict[str, Any]]:
+        """
+        从 Bedrock 控制面拉取模型目录，聚合跨区 Inference Profile 与直连模型
+
+        Bedrock 多数新模型仅允许通过 Inference Profile（us./eu./apac./global. 前缀）
+        调用，因此优先列出 Profile，再补充支持 ON_DEMAND 直连的基础模型。
+
+        :param api_key: 用户填写的凭证内容（Bedrock API Key 或 AK/SK）
+        :param base_url: Bedrock 运行时端点，决定 Region
+        :param use_proxy: 是否使用系统代理
+        :return: 标准化后的模型记录列表
+        """
+        credentials = self._parse_bedrock_credentials(api_key)
+        region = self._extract_bedrock_region(base_url)
+        # runtime VPCE 无法安全推导对应的控制面 VPCE；FIPS 端点也不能绕回
+        # 公有非 FIPS 控制面，因此直接使用本地目录。
+        if self._bedrock_endpoint_url("bedrock-runtime", base_url):
+            return await self._list_models_from_bedrock_fallback(region, use_proxy)
+        client = self.create_bedrock_client(
+            "bedrock",
+            region=region,
+            credentials=credentials,
+            use_proxy=use_proxy,
+        )
+
+        def _fetch() -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:
+            profiles: list[dict[str, Any]] = []
+            paginator = client.get_paginator("list_inference_profiles")
+            for page in paginator.paginate(typeEquals="SYSTEM_DEFINED"):
+                profiles.extend(page.get("inferenceProfileSummaries") or [])
+            foundation = client.list_foundation_models(
+                byOutputModality="TEXT",
+                byInferenceType="ON_DEMAND",
+            ).get("modelSummaries") or []
+            return profiles, foundation
+
+        try:
+            profile_summaries, foundation_summaries = await asyncio.to_thread(_fetch)
+        except Exception as err:
+            # 部分 Bedrock API Key 的授权范围仅覆盖 bedrock-runtime 推理接口，
+            # 控制面查询被拒时降级到 models.dev 目录，保证仍能选择模型。
+            logger.warning(
+                f"获取 Amazon Bedrock 控制面模型列表失败，降级 models.dev 目录: {err}"
+            )
+            return await self._list_models_from_bedrock_fallback(region, use_proxy)
+        finally:
+            await asyncio.to_thread(client.close)
+
+        results: list[dict[str, Any]] = []
+        seen_ids: set[str] = set()
+
+        def _append_record(model_id: str, display_name: Optional[str]) -> None:
+            if not model_id or model_id in seen_ids:
+                return
+            seen_ids.add(model_id)
+            # Inference Profile 带区域前缀，models.dev 目录按基础模型 ID 收录，
+            # 去掉首个前缀段再查一次元数据。
+            metadata = self._cached_models_dev_model("amazon-bedrock", model_id)
+            if not metadata and "." in model_id:
+                metadata = self._cached_models_dev_model(
+                    "amazon-bedrock",
+                    model_id.split(".", 1)[1],
+                )
+            results.append(
+                self._normalize_model_record(
+                    model_id=model_id,
+                    display_name=display_name or (metadata or {}).get("name") or model_id,
+                    metadata=metadata or {},
+                    source="provider",
+                )
+            )
+
+        for profile in profile_summaries:
+            if (profile.get("status") or "ACTIVE") != "ACTIVE":
+                continue
+            _append_record(
+                str(profile.get("inferenceProfileId") or "").strip(),
+                profile.get("inferenceProfileName"),
+            )
+        # 控制面已按当前 Region 和 ON_DEMAND 筛选，不能复用仅面向
+        # models.dev 降级目录的静态白名单，否则 AWS 新增模型会被遗漏。
+        for summary in foundation_summaries:
+            lifecycle = (summary.get("modelLifecycle") or {}).get("status") or "ACTIVE"
+            if lifecycle != "ACTIVE":
+                continue
+            _append_record(
+                str(summary.get("modelId") or "").strip(),
+                summary.get("modelName"),
+            )
+
+        return sorted(results, key=lambda item: item["name"].lower())
+
     @staticmethod
     def _copilot_headers(
             token: Optional[str] = None, include_auth: bool = True
@@ -2064,6 +2528,13 @@ class LLMProviderManager(metaclass=Singleton):
                 use_proxy=use_proxy,
             )
 
+        if resolved_model_list_strategy == "bedrock":
+            return await self._list_models_from_bedrock(
+                api_key=runtime["api_key"],
+                base_url=runtime.get("base_url"),
+                use_proxy=use_proxy,
+            )
+
         if resolved_model_list_strategy == "anthropic_compatible":
             return await self._list_models_from_models_dev_only(
                 provider_id=provider_id,
@@ -2731,6 +3202,22 @@ class LLMProviderManager(metaclass=Singleton):
             )
             return result
 
+        if resolved_runtime == "bedrock":
+            effective_base_url = normalized_base_url or self._default_base_url_for_provider(
+                spec
+            )
+            credentials = self._parse_bedrock_credentials(normalized_api_key)
+            result.update(
+                {
+                    "api_key": normalized_api_key,
+                    "base_url": effective_base_url,
+                    "aws_region": self._extract_bedrock_region(effective_base_url),
+                    "aws_auth": credentials,
+                    "auth_mode": "api_key",
+                }
+            )
+            return result
+
         if resolved_runtime == "anthropic_compatible":
             effective_base_url = normalized_base_url or self._default_base_url_for_provider(
                 spec
diff --git a/app/api/endpoints/system.py b/app/api/endpoints/system.py
index 0fd70b3c..98b44244 100644
--- a/app/api/endpoints/system.py
+++ b/app/api/endpoints/system.py
@@ -1339,12 +1339,7 @@ def restart_system(_: User = Depends(get_current_active_superuser)):
     """
     if not SystemHelper.can_restart():
         return schemas.Response(success=False, message="当前运行环境不支持重启操作！")
-    # 标识停止事件
-    global_vars.stop_system()
-    # 执行重启
     ret, msg = SystemHelper.restart()
-    if not ret:
-        global_vars.resume_system()
     return schemas.Response(success=ret, message=msg)
 
 
@@ -1362,11 +1357,7 @@ def upgrade_system(
     if not SystemHelper.can_restart():
         return schemas.Response(success=False, message="当前运行环境不支持升级操作！")
 
-    # 标识停止事件
-    global_vars.stop_system()
     ret, msg = SystemHelper.upgrade(mode=mode or "release")
-    if not ret:
-        global_vars.resume_system()
     return schemas.Response(success=ret, message=msg)
 
 
diff --git a/app/chain/mediaserver.py b/app/chain/mediaserver.py
index cb0d5f82..8f3dcf7d 100644
--- a/app/chain/mediaserver.py
+++ b/app/chain/mediaserver.py
@@ -1,6 +1,6 @@
 import threading
 from datetime import datetime
-from typing import Callable, List, Union, Optional, Generator, Any
+from typing import Callable, Dict, List, Union, Optional, Generator, Any
 
 from app.chain import ChainBase
 from app.core.config import global_vars
@@ -210,6 +210,24 @@ class MediaServerChain(ChainBase):
         """
         return self.run_module("mediaserver_play_url", server=server, item_id=item_id)
 
+    def get_season_episode_ids(self, server: str, item_id: Union[str, int],
+                               season: int) -> Dict[int, str]:
+        """
+        获取指定季的集号到媒体服务器条目 ID 映射
+
+        :param server: 媒体服务器名称
+        :param item_id: 剧集在媒体服务器中的条目 ID
+        :param season: 季号
+        :return: 集号到条目 ID 的映射，无数据时返回空字典
+        """
+        result = self.run_module(
+            "mediaserver_season_episode_ids",
+            server=server,
+            item_id=item_id,
+            season=season,
+        )
+        return result or {}
+
     def get_image_cookies(
         self, server: Optional[str], image_url: str
     ) -> Optional[str | dict]:
diff --git a/app/chain/subscribe.py b/app/chain/subscribe.py
index dc60f313..7b4b7d3c 100644
--- a/app/chain/subscribe.py
+++ b/app/chain/subscribe.py
@@ -11,6 +11,7 @@ from app import schemas
 from app.chain import ChainBase
 from app.chain.download import DownloadChain
 from app.chain.media import MediaChain
+from app.chain.mediaserver import MediaServerChain
 from app.chain.search import SearchChain
 from app.chain.tmdb import TmdbChain
 from app.chain.torrents import TorrentsChain
@@ -34,6 +35,7 @@ from app.helper.interaction import (
     supports_markdown,
     update_or_post_message,
 )
+from app.helper.mediaserver import MediaServerHelper
 from app.helper.server import MoviePilotServerHelper
 from app.helper.torrent import TorrentHelper
 from app.log import logger
@@ -3642,6 +3644,84 @@ class SubscribeChain(ChainBase):
                 else:
                     episodes[0].library.append(file_info)
 
+        # 合并所有媒体服务器已存在条目（逐台查询，不只取第一个命中）
+        mediaserver_chain = MediaServerChain()
+        server_names = list(MediaServerHelper().get_services().keys())
+
+        def _has_server_entry(library_list: List[schemas.SubscribeLibraryFileInfo],
+                              server_name: Optional[str],
+                              server_type: Optional[str]) -> bool:
+            for info in library_list or []:
+                if info.server and server_name and info.server == server_name:
+                    return True
+                if info.server_type and server_type and info.server_type == server_type \
+                        and info.server == server_name \
+                        and (not info.file_path or str(info.file_path).startswith(("http://", "https://"))):
+                    return True
+            return False
+
+        for server_name in server_names:
+            exists_media = self.media_exists(mediainfo=mediainfo, server=server_name)
+            # 仅合并真实媒体服务器结果，跳过本地 FileManager 兜底（已由 media_files 覆盖）
+            if not exists_media or not (exists_media.server or exists_media.server_type):
+                continue
+
+            resolved_server = exists_media.server or server_name
+            server_storage = exists_media.server_type or resolved_server
+            server_itemid = str(exists_media.itemid) if exists_media.itemid is not None else None
+            series_detail_url = None
+            if resolved_server and exists_media.itemid is not None:
+                series_detail_url = mediaserver_chain.get_play_url(
+                    server=resolved_server,
+                    item_id=exists_media.itemid,
+                )
+
+            if subscribe.type == MediaType.TV.value:
+                season_number = subscribe.season if subscribe.season is not None else 1
+                exist_episodes = (exists_media.seasons or {}).get(season_number) or []
+                episode_item_ids: Dict[int, str] = {}
+                if resolved_server and exists_media.itemid is not None:
+                    episode_item_ids = mediaserver_chain.get_season_episode_ids(
+                        server=resolved_server,
+                        item_id=exists_media.itemid,
+                        season=season_number,
+                    )
+                for episode_number in exist_episodes:
+                    episode_info = episodes.get(episode_number)
+                    if not episode_info:
+                        continue
+                    if _has_server_entry(episode_info.library, resolved_server, exists_media.server_type):
+                        continue
+                    episode_itemid = episode_item_ids.get(episode_number) or server_itemid
+                    detail_url = series_detail_url
+                    if resolved_server and episode_item_ids.get(episode_number):
+                        detail_url = mediaserver_chain.get_play_url(
+                            server=resolved_server,
+                            item_id=episode_itemid,
+                        ) or series_detail_url
+                    episode_info.library.append(
+                        schemas.SubscribeLibraryFileInfo(
+                            storage=server_storage,
+                            file_path=detail_url,
+                            server=resolved_server,
+                            server_type=exists_media.server_type,
+                            itemid=str(episode_itemid) if episode_itemid is not None else None,
+                        )
+                    )
+            else:
+                episode_info = episodes.get(0)
+                if episode_info and not _has_server_entry(
+                        episode_info.library, resolved_server, exists_media.server_type):
+                    episode_info.library.append(
+                        schemas.SubscribeLibraryFileInfo(
+                            storage=server_storage,
+                            file_path=series_detail_url,
+                            server=resolved_server,
+                            server_type=exists_media.server_type,
+                            itemid=server_itemid,
+                        )
+                    )
+
         # 更新订阅信息
         subscribe_info.subscribe = Subscribe(**subscribe.to_dict())
         subscribe_info.episodes = episodes
diff --git a/app/chain/system.py b/app/chain/system.py
index c3bc5cf6..59abaab9 100644
--- a/app/chain/system.py
+++ b/app/chain/system.py
@@ -38,8 +38,6 @@ class SystemChain(ChainBase):
         """
         重启系统
         """
-        from app.core.config import global_vars
-
         if channel and userid:
             self.post_message(Notification(
                 channel=channel,
@@ -54,8 +52,6 @@ class SystemChain(ChainBase):
             }, self._restart_file)
         # 主动备份一次插件
         self.backup_plugins()
-        # 设置停止标志，通知所有模块准备停止
-        global_vars.stop_system()
         # 重启
         SystemHelper.restart()
 
diff --git a/app/core/config.py b/app/core/config.py
index 3d7a5f6d..06f52260 100644
--- a/app/core/config.py
+++ b/app/core/config.py
@@ -1211,12 +1211,6 @@ class GlobalVar(object):
         """
         self.STOP_EVENT.set()
 
-    def resume_system(self):
-        """
-        恢复系统运行标记。
-        """
-        self.STOP_EVENT.clear()
-
     @property
     def is_system_stopped(self):
         """
diff --git a/app/main.py b/app/main.py
index 9cc53f29..400cc75e 100644
--- a/app/main.py
+++ b/app/main.py
@@ -30,16 +30,31 @@ elif SystemUtils.is_frozen():
     sys.stderr = open(os.devnull, 'w')
 
 from app.factory import app
-from app.core.config import settings
+from app.core.config import global_vars, settings
 from app.db.init import init_db, update_db
 
 # 设置进程名
 setproctitle.setproctitle(settings.PROJECT_NAME)
 
+
+class MoviePilotServer(uvicorn.Server):
+    """在 Uvicorn 开始优雅退出前发布应用协作停止标志"""
+
+    def handle_exit(self, sig, frame) -> None:
+        global_vars.stop_system()
+        super().handle_exit(sig, frame)
+
+
 # uvicorn服务
-Server = uvicorn.Server(Config(app, host=settings.HOST, port=settings.PORT,
-                               reload=settings.DEV, workers=multiprocessing.cpu_count() * 2 + 1,
-                               timeout_graceful_shutdown=60))
+Server = MoviePilotServer(Config(app, host=settings.HOST, port=settings.PORT,
+                                 reload=settings.DEV, workers=multiprocessing.cpu_count() * 2 + 1,
+                                 timeout_graceful_shutdown=60))
+
+
+def request_shutdown() -> None:
+    """发布协作停止标志并请求 Uvicorn 退出"""
+    global_vars.stop_system()
+    Server.should_exit = True
 
 
 def start_tray():
@@ -64,8 +79,8 @@ def start_tray():
         """
         退出程序
         """
+        request_shutdown()
         TrayIcon.stop()
-        Server.should_exit = True
 
     import pystray
 
@@ -93,10 +108,11 @@ def signal_handler(signum, frame):
     信号处理函数，用于优雅停止服务
     """
     print(f"收到信号 {signum}，开始优雅停止服务...")
-    Server.should_exit = True
+    request_shutdown()
 
 
-if __name__ == '__main__':
+def run_application() -> None:
+    """初始化进程并启动 API 服务"""
     # 注册信号处理器
     signal.signal(signal.SIGTERM, signal_handler)
     signal.signal(signal.SIGINT, signal_handler)
@@ -109,3 +125,7 @@ if __name__ == '__main__':
     update_db()
     # 启动API服务
     Server.run()
+
+
+if __name__ == '__main__':
+    run_application()
diff --git a/app/modules/emby/__init__.py b/app/modules/emby/__init__.py
index 665a4d6e..1aa36104 100644
--- a/app/modules/emby/__init__.py
+++ b/app/modules/emby/__init__.py
@@ -1,4 +1,4 @@
-from typing import Any, Generator, List, Optional, Tuple, Union
+from typing import Any, Dict, Generator, List, Optional, Tuple, Union
 
 from app import schemas
 from app.core.context import MediaInfo
@@ -300,6 +300,21 @@ class EmbyModule(_ModuleBase, _MediaServerBase[Emby]):
             return None
         return server_obj.get_play_url(item_id)
 
+    def mediaserver_season_episode_ids(self, server: str, item_id: Union[str, int],
+                                       season: int) -> Optional[Dict[int, str]]:
+        """
+        获取指定季的集号到条目 ID 映射
+
+        :param server: Emby 媒体服务器名称
+        :param item_id: 剧集在 Emby 中的条目 ID
+        :param season: 季号
+        :return: 集号到条目 ID 的映射，服务器不可用或无数据时返回 None
+        """
+        server_obj: Emby = self.get_instance(server)
+        if not server_obj:
+            return None
+        return server_obj.get_season_episode_ids(str(item_id), season)
+
     def mediaserver_latest(self, server: Optional[str] = None, count: Optional[int] = 20,
                            username: Optional[str] = None) -> List[schemas.MediaServerPlayItem]:
         """
diff --git a/app/modules/emby/emby.py b/app/modules/emby/emby.py
index 68bba8b7..5b25f4a7 100644
--- a/app/modules/emby/emby.py
+++ b/app/modules/emby/emby.py
@@ -474,6 +474,37 @@ class Emby:
             return None, None
         return None, {}
 
+    def get_season_episode_ids(self, item_id: str, season: int) -> Dict[int, str]:
+        """
+        获取指定季的集号到媒体服务器条目 ID 映射
+        :param item_id: 剧集在 Emby 中的 ID
+        :param season: 季号
+        :return: {集号: episode_item_id}
+        """
+        if not item_id or not self._host or not self._apikey:
+            return {}
+        try:
+            url = f"{self._host}emby/Shows/{item_id}/Episodes"
+            params = {
+                "Season": season,
+                "IsMissing": "false",
+                "api_key": self._apikey
+            }
+            res_json = RequestUtils().get_res(url, params)
+            if not res_json:
+                return {}
+            episode_ids: Dict[int, str] = {}
+            for res_item in res_json.json().get("Items") or []:
+                episode_index = res_item.get("IndexNumber")
+                episode_id = res_item.get("Id")
+                if episode_index is None or not episode_id:
+                    continue
+                episode_ids[int(episode_index)] = str(episode_id)
+            return episode_ids
+        except Exception as e:
+            logger.error(f"获取 Emby 季集条目 ID 出错：{str(e)}")
+            return {}
+
     def get_remote_image_by_id(self, item_id: str, image_type: str) -> Optional[str]:
         """
         根据ItemId从Emby查询TMDB的图片地址
diff --git a/app/modules/filemanager/__init__.py b/app/modules/filemanager/__init__.py
index b6e7b76c..078bc7a3 100644
--- a/app/modules/filemanager/__init__.py
+++ b/app/modules/filemanager/__init__.py
@@ -606,8 +606,12 @@ class FileManagerModule(_ModuleBase):
         """
         判断媒体文件是否存在于文件系统（网盘或本地文件），只支持标准媒体库结构
         :param mediainfo:  识别的媒体信息
+        :param server:  指定媒体服务器名称时跳过本地文件系统检查
         :return: 如不存在返回None，存在时返回信息，包括每季已存在所有集{type: movie/tv, seasons: {season: [episodes]}}
         """
+        if kwargs.get("server"):
+            return None
+
         if not settings.LOCAL_EXISTS_SEARCH:
             return None
 
diff --git a/app/modules/jellyfin/__init__.py b/app/modules/jellyfin/__init__.py
index 5b10f384..b0da28b3 100644
--- a/app/modules/jellyfin/__init__.py
+++ b/app/modules/jellyfin/__init__.py
@@ -1,4 +1,4 @@
-from typing import Any, Generator, List, Optional, Tuple, Union
+from typing import Any, Dict, Generator, List, Optional, Tuple, Union
 
 from app import schemas
 from app.core.context import MediaInfo
@@ -299,6 +299,21 @@ class JellyfinModule(_ModuleBase, _MediaServerBase[Jellyfin]):
             return None
         return server_obj.get_play_url(item_id)
 
+    def mediaserver_season_episode_ids(self, server: str, item_id: Union[str, int],
+                                       season: int) -> Optional[Dict[int, str]]:
+        """
+        获取指定季的集号到条目 ID 映射
+
+        :param server: Jellyfin 媒体服务器名称
+        :param item_id: 剧集在 Jellyfin 中的条目 ID
+        :param season: 季号
+        :return: 集号到条目 ID 的映射，服务器不可用或无数据时返回 None
+        """
+        server_obj: Jellyfin = self.get_instance(server)
+        if not server_obj:
+            return None
+        return server_obj.get_season_episode_ids(str(item_id), season)
+
     def mediaserver_latest(self, server: Optional[str] = None, count: Optional[int] = 20,
                            username: Optional[str] = None) -> List[schemas.MediaServerPlayItem]:
         """
diff --git a/app/modules/jellyfin/jellyfin.py b/app/modules/jellyfin/jellyfin.py
index 3a6f324f..005e0244 100644
--- a/app/modules/jellyfin/jellyfin.py
+++ b/app/modules/jellyfin/jellyfin.py
@@ -523,6 +523,38 @@ class Jellyfin:
             return None, None
         return None, {}
 
+    def get_season_episode_ids(self, item_id: str, season: int) -> Dict[int, str]:
+        """
+        获取指定季的集号到媒体服务器条目 ID 映射
+        :param item_id: 剧集在 Jellyfin 中的 ID
+        :param season: 季号
+        :return: {集号: episode_item_id}
+        """
+        if not item_id or not self._host or not self._apikey or not self.user:
+            return {}
+        try:
+            url = f"{self._host}Shows/{item_id}/Episodes"
+            params = {
+                "season": season,
+                "userId": self.user,
+                "isMissing": "false",
+                "api_key": self._apikey
+            }
+            res_json = RequestUtils().get_res(url, params)
+            if not res_json:
+                return {}
+            episode_ids: Dict[int, str] = {}
+            for res_item in res_json.json().get("Items") or []:
+                episode_index = res_item.get("IndexNumber")
+                episode_id = res_item.get("Id")
+                if episode_index is None or not episode_id:
+                    continue
+                episode_ids[int(episode_index)] = str(episode_id)
+            return episode_ids
+        except Exception as e:
+            logger.error(f"获取 Jellyfin 季集条目 ID 出错：{str(e)}")
+            return {}
+
     def get_remote_image_by_id(self, item_id: str, image_type: str) -> Optional[str]:
         """
         根据ItemId从Jellyfin查询TMDB图片地址
diff --git a/app/modules/plex/__init__.py b/app/modules/plex/__init__.py
index f386658f..46cf2ebe 100644
--- a/app/modules/plex/__init__.py
+++ b/app/modules/plex/__init__.py
@@ -1,4 +1,4 @@
-from typing import Optional, Tuple, Union, Any, List, Generator
+from typing import Optional, Tuple, Union, Any, List, Generator, Dict
 
 from app import schemas
 from app.core.context import MediaInfo
@@ -349,3 +349,18 @@ class PlexModule(_ModuleBase, _MediaServerBase[Plex]):
         if not server_obj:
             return None
         return server_obj.get_play_url(item_id)
+
+    def mediaserver_season_episode_ids(self, server: str, item_id: Union[str, int],
+                                       season: int) -> Optional[Dict[int, str]]:
+        """
+        获取指定季的集号到条目 ID 映射
+
+        :param server: Plex 媒体服务器名称
+        :param item_id: 剧集在 Plex 中的条目 ID / key
+        :param season: 季号
+        :return: 集号到条目 ID 的映射，服务器不可用或无数据时返回 None
+        """
+        server_obj: Plex = self.get_instance(server)
+        if not server_obj:
+            return None
+        return server_obj.get_season_episode_ids(str(item_id), season)
diff --git a/app/modules/plex/plex.py b/app/modules/plex/plex.py
index c051ee95..f0a38cc5 100644
--- a/app/modules/plex/plex.py
+++ b/app/modules/plex/plex.py
@@ -293,6 +293,31 @@ class Plex:
             season_episodes[episode.seasonNumber].append(episode.index)
         return videos.key, season_episodes
 
+    def get_season_episode_ids(self, item_id: str, season: int) -> Dict[int, str]:
+        """
+        获取指定季的集号到媒体服务器条目 ID 映射
+        :param item_id: 剧集在 Plex 中的 ID / key
+        :param season: 季号
+        :return: {集号: episode_item_key}
+        """
+        if not self._plex or not item_id:
+            return {}
+        try:
+            videos = self.__fetch_item(item_id)
+            if not videos:
+                return {}
+            episode_ids: Dict[int, str] = {}
+            for episode in videos.episodes():
+                if episode.seasonNumber != int(season):
+                    continue
+                if episode.index is None or not episode.key:
+                    continue
+                episode_ids[int(episode.index)] = str(episode.key)
+            return episode_ids
+        except Exception as e:
+            logger.error(f"获取 Plex 季集条目 ID 出错：{str(e)}")
+            return {}
+
     def __search_show(self,
                       title: Optional[str] = None,
                       original_title: Optional[str] = None,
diff --git a/app/modules/thetvdb/__init__.py b/app/modules/thetvdb/__init__.py
index de4e2711..d7d9baa5 100644
--- a/app/modules/thetvdb/__init__.py
+++ b/app/modules/thetvdb/__init__.py
@@ -114,7 +114,6 @@ class TheTvDbModule(_ModuleBase):
         return 4
 
     def stop(self):
-        logger.info("TheTvDbModule 停止。正在清除 TVDB 会话。")
         with self.__auth_lock:
             self.tvdb = None
 
diff --git a/app/schemas/subscribe.py b/app/schemas/subscribe.py
index a16338ed..68dcc354 100644
--- a/app/schemas/subscribe.py
+++ b/app/schemas/subscribe.py
@@ -217,6 +217,12 @@ class SubscribeLibraryFileInfo(BaseModel):
     storage: Optional[str] = "local"
     # 文件路径
     file_path: Optional[str] = None
+    # 媒体服务器名称
+    server: Optional[str] = None
+    # 媒体服务器类型：emby、jellyfin、plex 等
+    server_type: Optional[str] = None
+    # 媒体服务器条目 ID
+    itemid: Optional[str] = None
 
 
 class SubscribeEpisodeInfo(BaseModel):
diff --git a/app/startup/lifecycle.py b/app/startup/lifecycle.py
index a40d1525..3d114271 100644
--- a/app/startup/lifecycle.py
+++ b/app/startup/lifecycle.py
@@ -1,5 +1,7 @@
 import asyncio
+import inspect
 from contextlib import asynccontextmanager
+from typing import Callable
 
 from fastapi import FastAPI
 
@@ -20,7 +22,7 @@ from app.chain.system import SystemChain
 from app.core.config import global_vars, settings
 from app.helper.server import MoviePilotServerHelper
 from app.helper.system import SystemHelper
-from app.log import LoggerManager
+from app.log import logger, LoggerManager
 from app.startup.command_initializer import init_command, stop_command, restart_command
 from app.startup.modules_initializer import init_modules, stop_modules
 from app.startup.monitor_initializer import stop_monitor, init_monitor
@@ -56,6 +58,16 @@ async def init_extra():
     await MoviePilotServerHelper.async_report_usage()
 
 
+async def run_shutdown_step(name: str, callback: Callable[[], object]) -> None:
+    """隔离单个关闭阶段的异常，确保后续资源仍有机会释放"""
+    try:
+        result = callback()
+        if inspect.isawaitable(result):
+            await result
+    except Exception as err:
+        logger.error(f"关闭{name}失败：{err}")
+
+
 @asynccontextmanager
 async def lifespan(app: FastAPI):
     """
@@ -90,6 +102,7 @@ async def lifespan(app: FastAPI):
         yield
     finally:
         print("Shutting down...")
+        global_vars.stop_system()
         # 取消同步插件任务
         try:
             sync_plugins_task.cancel()
@@ -100,22 +113,19 @@ async def lifespan(app: FastAPI):
             print(str(e))
         try:
             if not settings.MOVIEPILOT_SAFE_MODE:
-                # 备份插件
-                SystemChain().backup_plugins()
-                # 停止工作流
-                stop_workflow()
-                # 停止命令
-                stop_command()
-                # 停止监控器
-                stop_monitor()
-                # 停止定时器
-                stop_scheduler()
-                # 停止插件
-                stop_plugins()
-            # 停止模块
-            await stop_modules()
-            # 关闭共享的异步 HTTP 连接池，释放底层连接资源
-            await aclose_shared_async_transports()
+                await run_shutdown_step(
+                    "插件备份", lambda: SystemChain().backup_plugins()
+                )
+                await run_shutdown_step("工作流", stop_workflow)
+                await run_shutdown_step("命令服务", stop_command)
+                await run_shutdown_step("监控器", stop_monitor)
+                await run_shutdown_step("定时器", stop_scheduler)
+                await run_shutdown_step("插件", stop_plugins)
+            await run_shutdown_step("模块服务", stop_modules)
+            await run_shutdown_step(
+                "共享异步 HTTP 连接池",
+                aclose_shared_async_transports,
+            )
         finally:
             # 日志最后关闭，确保其他组件的收尾信息已写入文件
             LoggerManager.shutdown()
diff --git a/app/startup/modules_initializer.py b/app/startup/modules_initializer.py
index 840cfd3d..e8a66f54 100644
--- a/app/startup/modules_initializer.py
+++ b/app/startup/modules_initializer.py
@@ -1,4 +1,6 @@
+import inspect
 import sys
+from typing import Callable
 
 from app.helper.redis import RedisHelper, AsyncRedisHelper
 
@@ -129,29 +131,27 @@ async def stop_modules():
     """
     服务关闭
     """
-    # 停止AI智能体
-    await stop_agent()
-    # 停止模块
-    ModuleManager().stop()
-    # 停止事件消费
-    EventManager().stop()
-    # 停止虚拟显示
-    DisplayHelper().stop()
-    # 停止 DoH 服务
-    DohHelper().shutdown()
-    # 停止线程池
-    ThreadHelper().shutdown()
-    # 停止消息服务
-    stop_message()
-    # 关闭Redis缓存连接
-    RedisHelper().close()
-    await AsyncRedisHelper().close()
-    # 停止数据库连接
-    await close_database()
-    # 停止前端服务
-    stop_frontend()
-    # 清理临时文件
-    clear_temp()
+    async def run_step(name: str, callback: Callable[[], object]) -> None:
+        """单个模块资源关闭失败时继续执行后续阶段"""
+        try:
+            result = callback()
+            if inspect.isawaitable(result):
+                await result
+        except Exception as err:
+            logger.error(f"关闭{name}失败：{err}")
+
+    await run_step("AI智能体", stop_agent)
+    await run_step("模块", lambda: ModuleManager().stop())
+    await run_step("事件消费", lambda: EventManager().stop())
+    await run_step("虚拟显示", lambda: DisplayHelper().stop())
+    await run_step("DoH服务", lambda: DohHelper().shutdown())
+    await run_step("线程池", lambda: ThreadHelper().shutdown())
+    await run_step("消息服务", stop_message)
+    await run_step("Redis缓存连接", lambda: RedisHelper().close())
+    await run_step("异步Redis缓存连接", lambda: AsyncRedisHelper().close())
+    await run_step("数据库连接", close_database)
+    await run_step("前端服务", stop_frontend)
+    await run_step("临时文件", clear_temp)
 
 
 def init_modules():
diff --git a/app/utils/http.py b/app/utils/http.py
index dbf8ddf2..12662d5e 100644
--- a/app/utils/http.py
+++ b/app/utils/http.py
@@ -76,6 +76,14 @@ _REQUESTS_RETRY_IDEMPOTENT_METHODS = ("GET", "HEAD", "OPTIONS")
 _pending_eviction_tasks: set[asyncio.Task] = set()
 
 
+def _discard_pending_eviction_task(task: asyncio.Task) -> None:
+    """从跨线程共享集合移除已完成的 transport 关闭任务"""
+    with _shared_async_transports_lock:
+        _pending_eviction_tasks.discard(task)
+    if not task.cancelled() and (error := task.exception()):
+        logger.debug(f"LRU 淘汰共享 transport 时关闭失败: {error!r}")
+
+
 def _get_shared_async_transport(
     proxy: Optional[str],
     verify: Union[bool, str],
@@ -140,8 +148,9 @@ def _get_shared_async_transport(
         try:
             task = loop.create_task(evicted_transport.aclose())
             # 强引用避免 task 仅被 loop 弱持有而触发 "Task was destroyed but pending"
-            _pending_eviction_tasks.add(task)
-            task.add_done_callback(_pending_eviction_tasks.discard)
+            with _shared_async_transports_lock:
+                _pending_eviction_tasks.add(task)
+            task.add_done_callback(_discard_pending_eviction_task)
         except Exception as e:  # pragma: no cover - 防御性
             logger.debug(f"LRU 淘汰共享 transport 时调度关闭失败: {e!r}")
 
@@ -160,17 +169,27 @@ async def aclose_shared_async_transports() -> None:
     # 弹出而非 get+clear，避免外层 dict 残留空 OrderedDict 占位
     with _shared_async_transports_lock:
         per_loop = _shared_async_transports.pop(loop, None)
-    if not per_loop:
+        pending_evictions = [
+            task
+            for task in _pending_eviction_tasks
+            if task.get_loop() is loop
+        ]
+    transports = list(per_loop.values()) if per_loop else []
+    if per_loop:
+        per_loop.clear()
+    if not transports and not pending_evictions:
         return
-    transports = list(per_loop.values())
-    per_loop.clear()
     # 并行关闭：每个 transport 的 TLS close_notify 各占一个 RTT，
     # 顺序等待会线性放大 shutdown 耗时；return_exceptions 让单点失败
     # 不影响其他 transport 的释放
     results = await asyncio.gather(
-        *(t.aclose() for t in transports), return_exceptions=True
+        *pending_evictions,
+        *(t.aclose() for t in transports),
+        return_exceptions=True,
     )
-    for result in results:
+    with _shared_async_transports_lock:
+        _pending_eviction_tasks.difference_update(pending_evictions)
+    for result in results[len(pending_evictions):]:
         if isinstance(result, BaseException):
             logger.debug(f"关闭共享 AsyncHTTPTransport 失败: {result!r}")
 
diff --git a/docs/pr-agent.md b/docs/pr-agent.md
index a13def77..be64ae36 100644
--- a/docs/pr-agent.md
+++ b/docs/pr-agent.md
@@ -1,6 +1,6 @@
 # PR-Agent 使用说明
 
-本仓库通过 GitHub Actions 运行 PR-Agent，帮助贡献者维护 PR 摘要、获取代码审查结果和提出 PR 相关问题。
+本仓库通过 PR Review Runner 运行 PR-Agent，帮助贡献者维护 PR 摘要、获取代码审查结果和提出 PR 相关问题。
 
 ## 自动执行
 
@@ -25,26 +25,24 @@ PR 带有 `skip pr-agent` 标签，或标题以 `[Auto]`、`Auto` 开头时，
 /ask 这次改动有没有遗漏权限校验？
 ```
 
-- `/describe`：更新 PR Body 内按语言显示的 `PR-Agent 摘要` 或 `PR-Agent Summary`，并保留贡献者原有的 PR 描述。
+- `/describe`：更新 PR Body 内的 `PR-Agent 摘要`，并保留贡献者原有的 PR 描述。
 - `/review`：发起一次代码审查。
 - `/ask ...`：就当前 PR 提问，回复会发布在普通 PR 评论中。
 
+本仓库禁用 `/improve` 及其等价别名。其他命令是否可用由 runner 所包含的 PR-Agent 能力决定。
+
 手工命令仅允许以下 GitHub 身份关联的用户使用：`OWNER`、`MEMBER`、`COLLABORATOR`、`CONTRIBUTOR`、`FIRST_TIME_CONTRIBUTOR`。
 
 新建的合法命令评论会触发执行；编辑后仍为合法命令的评论也会触发。编辑普通讨论评论不会调用模型。
 
 ## 审查结果
 
-`/describe` 的结果位于 PR Body 中按语言显示的 `PR-Agent 摘要` 或 `PR-Agent Summary` 区域，用于概览本次变更。
+本仓库固定使用中文生成 PR-Agent 内容。`/describe` 的结果位于 PR Body 的 `PR-Agent 摘要` 区域，用于概览本次变更。
 
-`/review` 和自动审查会通过原生 GitHub Review 发布，结果位于 Review 页签，标题固定为 `PR-Agent Code Review`。审查摘要包含可点击的 `文件:行号` 链接；可定位到本次变更的具体问题会在对应代码行以行内评论呈现，无法行内定位的问题仍通过摘要中的链接呈现。
+`/review` 和自动审查会通过原生 GitHub Review 发布，结果位于 Review 页签，标题固定为 `PR-Agent Code Review`。可定位到本次变更的问题会在对应代码行以行内评论呈现，并使用 high、medium 或 low 风险标识。
 
-审查不会额外创建专用的摘要评论。未发现需要处理的问题时，Review 会显示：
-
-> 本次变更无需提出审查意见，暂无其他反馈。
+Review 摘要会自然概括本次变更和整体审查结论，不会复制行内评论。未发现需要处理的问题时，摘要会概括变更并自然说明暂无其他反馈。审查不会额外创建专用的 issue comment 摘要。
 
 ## 安全边界
 
-workflow 使用固定 digest 的 PR-Agent 容器镜像，不使用浮动标签。自动审查通过 `pull_request_target` 在目标仓库上下文中读取 PR 信息，但不会 checkout 或执行 PR 分支代码。
-
-权限保持最小化：只授予读取仓库内容所需的 `contents: read`，以及更新 PR Body、发布 Review 和回复 PR 评论所需的写权限；不会向仓库推送代码或创建提交。
+自动审查通过 `pull_request_target` 在目标仓库上下文中读取 PR 信息，并使用共享 runner 的 `latest` 镜像，但不会 checkout 或执行 PR 分支代码。权限保持最小化：只授予读取仓库内容所需的 `contents: read`，以及更新 PR Body、发布 Review 和回复 PR 评论所需的写权限；不会向仓库推送代码或创建提交。
diff --git a/requirements.in b/requirements.in
index fbeaedc9..3f598b72 100644
--- a/requirements.in
+++ b/requirements.in
@@ -1,4 +1,4 @@
-moviepilot-rust~=0.2.2
+moviepilot-rust~=0.2.3
 pydantic>=2.13.4,<3.0.0
 pydantic-settings>=2.14.1,<3.0.0
 SQLAlchemy~=2.0.50
@@ -80,6 +80,8 @@ langchain~=1.3.9
 langchain-core~=1.4.7
 langchain-community~=0.4.2
 langchain-anthropic~=1.4.6
+langchain-aws~=1.6.2
+boto3~=1.42.42
 langchain-openai~=1.3.2
 langchain-google-genai~=4.2.5
 langchain-deepseek~=1.1.0
diff --git a/tests/test_lifecycle_shutdown.py b/tests/test_lifecycle_shutdown.py
index c984a3f5..34b9c7d4 100644
--- a/tests/test_lifecycle_shutdown.py
+++ b/tests/test_lifecycle_shutdown.py
@@ -1,16 +1,28 @@
 import asyncio
+import signal
+import threading
 from unittest.mock import AsyncMock, MagicMock
 
 import pytest
 from fastapi import FastAPI
 
-from app.startup import lifecycle
+from app.startup import lifecycle, modules_initializer
+from app.utils import http as http_utils
 
 
-def test_lifespan_closes_logger_when_early_shutdown_step_fails(monkeypatch):
-    """前置关闭步骤失败时仍应关闭 Logger"""
+def _assert_completed_once(mock: MagicMock) -> None:
+    if isinstance(mock, AsyncMock):
+        mock.assert_awaited_once_with()
+    else:
+        mock.assert_called_once_with()
+
+
+def _patch_lifespan(monkeypatch, *, failing_step: str | None = None) -> dict:
+    """隔离 lifespan 的外部依赖，并按名称注入一个关闭失败"""
     monkeypatch.setattr(lifecycle.settings, "MOVIEPILOT_SAFE_MODE", False)
     monkeypatch.setattr(lifecycle.global_vars, "set_loop", MagicMock())
+    monkeypatch.setattr(lifecycle.global_vars, "stop_system", MagicMock())
+
     for name in (
         "init_routers",
         "init_modules",
@@ -19,28 +31,375 @@ def test_lifespan_closes_logger_when_early_shutdown_step_fails(monkeypatch):
         "init_monitor",
         "init_command",
         "init_workflow",
+    ):
+        monkeypatch.setattr(lifecycle, name, MagicMock())
+
+    system_chain = MagicMock()
+    monkeypatch.setattr(lifecycle, "SystemChain", MagicMock(return_value=system_chain))
+    monkeypatch.setattr(lifecycle, "init_extra", AsyncMock())
+
+    shutdown_steps = {
+        "backup_plugins": system_chain.backup_plugins,
+        "stop_workflow": MagicMock(),
+        "stop_command": MagicMock(),
+        "stop_monitor": MagicMock(),
+        "stop_scheduler": MagicMock(),
+        "stop_plugins": MagicMock(),
+        "stop_modules": AsyncMock(),
+        "close_http": AsyncMock(),
+    }
+    for name in (
         "stop_workflow",
         "stop_command",
         "stop_monitor",
         "stop_scheduler",
         "stop_plugins",
     ):
-        monkeypatch.setattr(lifecycle, name, MagicMock())
+        monkeypatch.setattr(lifecycle, name, shutdown_steps[name])
+    monkeypatch.setattr(lifecycle, "stop_modules", shutdown_steps["stop_modules"])
+    monkeypatch.setattr(
+        lifecycle,
+        "aclose_shared_async_transports",
+        shutdown_steps["close_http"],
+    )
+
+    if failing_step:
+        shutdown_steps[failing_step].side_effect = RuntimeError(
+            f"{failing_step} failed"
+        )
 
-    system_chain = MagicMock()
-    system_chain.backup_plugins.side_effect = RuntimeError("backup failed")
-    monkeypatch.setattr(lifecycle, "SystemChain", MagicMock(return_value=system_chain))
-    monkeypatch.setattr(lifecycle, "init_extra", AsyncMock())
-    monkeypatch.setattr(lifecycle, "stop_modules", AsyncMock())
-    monkeypatch.setattr(lifecycle, "aclose_shared_async_transports", AsyncMock())
     logger_shutdown = MagicMock()
     monkeypatch.setattr(lifecycle.LoggerManager, "shutdown", logger_shutdown)
+    shutdown_steps["logger"] = logger_shutdown
+    return shutdown_steps
+
+
+@pytest.mark.parametrize(
+    "failing_step",
+    [
+        "backup_plugins",
+        "stop_workflow",
+        "stop_command",
+        "stop_monitor",
+        "stop_scheduler",
+        "stop_plugins",
+        "stop_modules",
+        "close_http",
+    ],
+)
+def test_lifespan_continues_after_each_shutdown_owner_failure(
+    monkeypatch,
+    failing_step,
+):
+    """任一关闭阶段失败都不能跳过后续资源所有者"""
+    shutdown_steps = _patch_lifespan(monkeypatch, failing_step=failing_step)
 
     async def run_lifespan():
-        with pytest.raises(RuntimeError, match="backup failed"):
-            async with lifecycle.lifespan(FastAPI()):
-                pass
+        async with lifecycle.lifespan(FastAPI()):
+            pass
 
     asyncio.run(run_lifespan())
 
-    logger_shutdown.assert_called_once_with()
+    lifecycle.global_vars.stop_system.assert_called_once_with()
+    for step in shutdown_steps.values():
+        _assert_completed_once(step)
+
+
+def test_uvicorn_signal_publishes_stop_before_server_exit(monkeypatch):
+    """Uvicorn 接管系统信号时必须先发布协作停止标志"""
+    from app import main
+
+    calls = []
+    monkeypatch.setattr(main.global_vars, "stop_system", lambda: calls.append("stop"))
+    monkeypatch.setattr(
+        main.uvicorn.Server,
+        "handle_exit",
+        lambda _self, _sig, _frame: calls.append("uvicorn"),
+    )
+
+    server = object.__new__(main.MoviePilotServer)
+    server.handle_exit(signal.SIGTERM, None)
+
+    assert calls == ["stop", "uvicorn"]
+
+
+def test_application_preserves_stop_requested_before_startup(monkeypatch):
+    """启动流程不能清除初始化前已经发布的退出请求"""
+    from app import main
+
+    stop_event = threading.Event()
+    stop_event.set()
+    monkeypatch.setattr(main.global_vars, "STOP_EVENT", stop_event)
+    calls = []
+    monkeypatch.setattr(
+        main.signal,
+        "signal",
+        lambda *_args: calls.append("signal"),
+    )
+    monkeypatch.setattr(main, "start_tray", lambda: calls.append("tray"))
+    monkeypatch.setattr(main, "init_db", lambda: calls.append("init_db"))
+    monkeypatch.setattr(main, "update_db", lambda: calls.append("update_db"))
+    monkeypatch.setattr(main.Server, "run", lambda: calls.append("server"))
+
+    main.run_application()
+
+    assert stop_event.is_set()
+    assert calls == [
+        "signal",
+        "signal",
+        "tray",
+        "init_db",
+        "update_db",
+        "server",
+    ]
+
+
+def test_uvicorn_preserves_stop_requested_before_serve(monkeypatch):
+    """Uvicorn 启动不能清除数据库初始化阶段已经发布的停止请求"""
+    from app import main
+
+    stop_event = threading.Event()
+    monkeypatch.setattr(main.global_vars, "STOP_EVENT", stop_event)
+    main.global_vars.stop_system()
+
+    async def serve(_self, sockets=None):
+        assert main.global_vars.is_system_stopped
+
+    monkeypatch.setattr(main.uvicorn.Server, "serve", serve)
+    server = object.__new__(main.MoviePilotServer)
+    asyncio.run(server.serve())
+
+
+@pytest.mark.parametrize("endpoint_name", ["restart_system", "upgrade_system"])
+@pytest.mark.parametrize(
+    "initially_stopped",
+    [False, True],
+    ids=["running", "stopping"],
+)
+def test_restart_endpoint_failure_preserves_stop_state(
+    monkeypatch,
+    endpoint_name,
+    initially_stopped,
+):
+    """重启或升级失败不能发布或撤销停止请求"""
+    from app.api.endpoints import system
+
+    stop_event = threading.Event()
+    if initially_stopped:
+        stop_event.set()
+    monkeypatch.setattr(system.global_vars, "STOP_EVENT", stop_event)
+    monkeypatch.setattr(system.SystemHelper, "can_restart", MagicMock(return_value=True))
+    monkeypatch.setattr(
+        system.SystemHelper,
+        "restart" if endpoint_name == "restart_system" else "upgrade",
+        MagicMock(return_value=(False, "restart failed")),
+    )
+
+    if endpoint_name == "restart_system":
+        response = system.restart_system(None)
+    else:
+        response = system.upgrade_system(None, None)
+
+    assert not response.success
+    assert stop_event.is_set() is initially_stopped
+
+
+def test_command_restart_failure_does_not_publish_stop_request(monkeypatch):
+    """命令重启失败时进程仍在运行，不能提前发布停止请求"""
+    from app.chain.system import SystemChain
+    from app.core.config import global_vars
+
+    stop_event = threading.Event()
+    monkeypatch.setattr(global_vars, "STOP_EVENT", stop_event)
+    monkeypatch.setattr(SystemChain, "backup_plugins", MagicMock())
+    restart = MagicMock(return_value=(False, "restart failed"))
+    monkeypatch.setattr("app.chain.system.SystemHelper.restart", restart)
+
+    chain = object.__new__(SystemChain)
+    chain.restart(channel=None, userid=None)
+
+    restart.assert_called_once_with()
+    assert not stop_event.is_set()
+
+
+def test_stop_modules_continues_after_internal_owner_failures(monkeypatch):
+    """模块关闭编排中的多个失败不能阻断其余清理"""
+    stop_agent = AsyncMock(side_effect=RuntimeError("agent failed"))
+    monkeypatch.setattr(modules_initializer, "stop_agent", stop_agent)
+    dependencies = _patch_module_shutdown_dependencies(monkeypatch)
+    dependencies["module"].side_effect = RuntimeError("module failed")
+
+    asyncio.run(modules_initializer.stop_modules())
+
+    stop_agent.assert_awaited_once_with()
+    for dependency in dependencies.values():
+        _assert_completed_once(dependency)
+
+
+def _patch_module_shutdown_dependencies(monkeypatch) -> dict:
+    """替换 stop_modules 的资源所有者，避免测试启动真实后台服务"""
+    dependencies = {}
+    for name, method_name in (
+        ("ModuleManager", "stop"),
+        ("EventManager", "stop"),
+        ("DisplayHelper", "stop"),
+        ("DohHelper", "shutdown"),
+        ("ThreadHelper", "shutdown"),
+        ("RedisHelper", "close"),
+    ):
+        instance = MagicMock()
+        setattr(instance, method_name, MagicMock())
+        monkeypatch.setattr(
+            modules_initializer,
+            name,
+            MagicMock(return_value=instance),
+        )
+        key = name.removesuffix("Helper").removesuffix("Manager").lower()
+        dependencies[key] = getattr(instance, method_name)
+
+    for name in ("stop_message", "stop_frontend", "clear_temp"):
+        dependency = MagicMock()
+        monkeypatch.setattr(modules_initializer, name, dependency)
+        dependencies[name] = dependency
+
+    async_redis = MagicMock()
+    async_redis.close = AsyncMock()
+    monkeypatch.setattr(
+        modules_initializer,
+        "AsyncRedisHelper",
+        MagicMock(return_value=async_redis),
+    )
+    dependencies["async_redis"] = async_redis.close
+    close_database = AsyncMock()
+    monkeypatch.setattr(modules_initializer, "close_database", close_database)
+    dependencies["close_database"] = close_database
+    return dependencies
+
+
+def test_shared_http_close_waits_for_real_lru_eviction(monkeypatch):
+    """最终 HTTP 关闭必须等待真实 LRU 淘汰任务并消费其异常"""
+
+    class FakeTransport:
+        created = []
+
+        def __init__(self, **_kwargs):
+            self.close_started = asyncio.Event()
+            self.release_close = asyncio.Event()
+            self.closed = False
+            self.fail_on_close = not self.created
+            if not self.fail_on_close:
+                self.release_close.set()
+            self.created.append(self)
+
+        async def aclose(self):
+            self.close_started.set()
+            await self.release_close.wait()
+            self.closed = True
+            if self.fail_on_close:
+                raise RuntimeError("eviction close failed")
+
+    monkeypatch.setattr(http_utils, "_MAX_SHARED_TRANSPORTS_PER_LOOP", 1)
+    monkeypatch.setattr(http_utils.httpx, "AsyncHTTPTransport", FakeTransport)
+    debug = MagicMock()
+    monkeypatch.setattr(http_utils.logger, "debug", debug)
+
+    async def run_test():
+        transport_kwargs = {
+            "proxy": None,
+            "verify": True,
+            "http2": False,
+            "max_keepalive_connections": 1,
+            "max_connections": 1,
+        }
+        evicted_transport = http_utils._get_shared_async_transport(
+            **transport_kwargs,
+            keepalive_expiry=1,
+        )
+        active_transport = http_utils._get_shared_async_transport(
+            **transport_kwargs,
+            keepalive_expiry=2,
+        )
+        await asyncio.wait_for(evicted_transport.close_started.wait(), timeout=1)
+
+        loop = asyncio.get_running_loop()
+        with http_utils._shared_async_transports_lock:
+            eviction_tasks = [
+                task
+                for task in http_utils._pending_eviction_tasks
+                if task.get_loop() is loop
+            ]
+        assert len(eviction_tasks) == 1
+
+        close_task = asyncio.create_task(http_utils.aclose_shared_async_transports())
+        await asyncio.sleep(0)
+        try:
+            assert not close_task.done()
+            evicted_transport.release_close.set()
+            await close_task
+            await asyncio.sleep(0)
+            assert eviction_tasks[0].done()
+            assert evicted_transport.closed
+            assert active_transport.closed
+            with http_utils._shared_async_transports_lock:
+                assert not any(
+                    task.get_loop() is loop
+                    for task in http_utils._pending_eviction_tasks
+                )
+        finally:
+            evicted_transport.release_close.set()
+            active_transport.release_close.set()
+            await asyncio.gather(close_task, return_exceptions=True)
+            await http_utils.aclose_shared_async_transports()
+
+    asyncio.run(run_test())
+
+    debug.assert_any_call(
+        "LRU 淘汰共享 transport 时关闭失败: "
+        "RuntimeError('eviction close failed')"
+    )
+
+
+def test_shared_http_close_ignores_eviction_from_other_loop():
+    """当前事件循环关闭不能等待其他循环持有的淘汰任务"""
+    ready = threading.Event()
+    release = threading.Event()
+    failures = []
+    state = {}
+
+    def run_foreign_loop():
+        loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(loop)
+
+        async def delayed_close():
+            while not release.is_set():
+                await asyncio.sleep(0.01)
+
+        task = loop.create_task(delayed_close())
+        state["task"] = task
+        with http_utils._shared_async_transports_lock:
+            http_utils._pending_eviction_tasks.add(task)
+        task.add_done_callback(http_utils._discard_pending_eviction_task)
+        ready.set()
+        try:
+            loop.run_until_complete(task)
+            loop.run_until_complete(asyncio.sleep(0))
+        except BaseException as err:
+            failures.append(err)
+        finally:
+            with http_utils._shared_async_transports_lock:
+                http_utils._pending_eviction_tasks.discard(task)
+            loop.close()
+
+    thread = threading.Thread(target=run_foreign_loop)
+    thread.start()
+    try:
+        assert ready.wait(timeout=2)
+        asyncio.run(http_utils.aclose_shared_async_transports())
+        assert thread.is_alive()
+        assert not state["task"].done()
+    finally:
+        release.set()
+        thread.join(timeout=2)
+
+    assert not thread.is_alive()
+    assert not failures
diff --git a/tests/test_llm_provider_bedrock.py b/tests/test_llm_provider_bedrock.py
new file mode 100644
index 00000000..c3a6f265
--- /dev/null
+++ b/tests/test_llm_provider_bedrock.py
@@ -0,0 +1,392 @@
+"""Amazon Bedrock provider 的凭证解析、Region 提取与运行时解析测试"""
+
+import asyncio
+import time
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from app.agent.llm.provider import (
+    LLMProviderAuthError,
+    LLMProviderManager,
+)
+
+
+@pytest.fixture(autouse=True)
+def _reset_manager_singleton():
+    """每个用例前后清理 LLMProviderManager 单例，避免缓存互相污染"""
+    LLMProviderManager._instances.clear()
+    yield
+    LLMProviderManager._instances.clear()
+
+
+def test_bedrock_provider_registered():
+    manager = LLMProviderManager()
+    spec = manager.get_provider("amazon-bedrock")
+
+    assert spec.runtime == "bedrock"
+    assert spec.model_list_strategy == "bedrock"
+    assert spec.base_url_editable is True
+    assert spec.default_base_url == "https://bedrock-runtime.us-east-1.amazonaws.com"
+    preset_ids = {preset.id for preset in spec.base_url_presets}
+    assert "bedrock-us-east-1" in preset_ids
+    assert "bedrock-ap-northeast-1" in preset_ids
+
+
+def test_parse_bedrock_credentials_bearer_api_key():
+    credentials = LLMProviderManager._parse_bedrock_credentials(
+        "bedrock-api-key-abcdef123456"
+    )
+
+    assert credentials["auth_scheme"] == "bearer"
+    assert credentials["bearer_token"] == "bedrock-api-key-abcdef123456"
+
+
+def test_parse_bedrock_credentials_sigv4_ak_sk():
+    credentials = LLMProviderManager._parse_bedrock_credentials(
+        "AKIAIOSFODNN7EXAMPLE:wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY"
+    )
+
+    assert credentials["auth_scheme"] == "sigv4"
+    assert credentials["access_key_id"] == "AKIAIOSFODNN7EXAMPLE"
+    assert credentials["secret_access_key"] == "wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY"
+    assert "session_token" not in credentials
+
+
+def test_parse_bedrock_credentials_sigv4_with_session_token():
+    credentials = LLMProviderManager._parse_bedrock_credentials(
+        "ASIAIOSFODNN7EXAMPLE:secret/key:session-token-value"
+    )
+
+    assert credentials["auth_scheme"] == "sigv4"
+    assert credentials["session_token"] == "session-token-value"
+
+
+def test_parse_bedrock_credentials_empty_rejected():
+    with pytest.raises(LLMProviderAuthError):
+        LLMProviderManager._parse_bedrock_credentials("")
+
+
+def test_parse_bedrock_credentials_malformed_colon_rejected():
+    with pytest.raises(LLMProviderAuthError):
+        LLMProviderManager._parse_bedrock_credentials("AKIA123:")
+
+
+def test_extract_bedrock_region_from_base_url():
+    """应从标准、FIPS 与 PrivateLink Bedrock 端点提取 Region"""
+    extract = LLMProviderManager._extract_bedrock_region
+
+    assert extract("https://bedrock-runtime.us-east-1.amazonaws.com") == "us-east-1"
+    assert extract("https://bedrock-runtime.ap-northeast-1.amazonaws.com/") == "ap-northeast-1"
+    assert extract("https://bedrock-runtime.mx-central-1.amazonaws.com") == "mx-central-1"
+    assert extract("https://bedrock.eu-central-1.amazonaws.com") == "eu-central-1"
+    # FIPS 与 PrivateLink（VPCE）端点同样能识别 Region
+    assert extract("https://bedrock-runtime-fips.us-east-1.amazonaws.com") == "us-east-1"
+    assert (
+        extract("https://vpce-0abc123-xyz.bedrock-runtime.us-west-2.vpce.amazonaws.com")
+        == "us-west-2"
+    )
+    # 无法识别时回退默认 Region
+    assert extract("https://example.com/us-west-2") == "us-east-1"
+    assert extract("https://example.com?region=.us-west-2.") == "us-east-1"
+    assert extract("https://example.com") == "us-east-1"
+    assert extract(None) == "us-east-1"
+    assert extract("") == "us-east-1"
+
+
+def test_bedrock_endpoint_url_passthrough():
+    """自定义 Bedrock 端点应透传，标准端点交由 boto3 推导"""
+    resolve = LLMProviderManager._bedrock_endpoint_url
+
+    # 标准公有端点交由 boto3 推导，不显式透传
+    assert resolve("bedrock-runtime", "https://bedrock-runtime.us-east-1.amazonaws.com") is None
+    assert resolve("bedrock", "https://bedrock.eu-central-1.amazonaws.com") is None
+    assert resolve("bedrock-runtime", None) is None
+    assert resolve("bedrock-runtime", "") is None
+    # FIPS / PrivateLink 等非标准端点需要显式生效
+    assert (
+        resolve("bedrock-runtime", "https://bedrock-runtime-fips.us-east-1.amazonaws.com")
+        == "https://bedrock-runtime-fips.us-east-1.amazonaws.com"
+    )
+    assert (
+        resolve(
+            "bedrock-runtime",
+            "https://vpce-0abc123-xyz.bedrock-runtime.us-west-2.vpce.amazonaws.com/",
+        )
+        == "https://vpce-0abc123-xyz.bedrock-runtime.us-west-2.vpce.amazonaws.com"
+    )
+    # runtime 端点填给控制面服务名时不匹配标准形态，同样透传
+    assert (
+        resolve("bedrock", "https://bedrock-runtime.us-east-1.amazonaws.com")
+        == "https://bedrock-runtime.us-east-1.amazonaws.com"
+    )
+
+
+def test_create_bedrock_client_uses_custom_endpoint():
+    """创建 Bedrock 客户端时应把 PrivateLink 地址传给 boto3"""
+    manager = LLMProviderManager()
+    endpoint_url = (
+        "https://vpce-0abc123-xyz.bedrock-runtime.us-west-2.vpce.amazonaws.com"
+    )
+    client = MagicMock()
+
+    with patch("boto3.client", return_value=client) as create_client:
+        result = manager.create_bedrock_client(
+            service_name="bedrock-runtime",
+            region="us-west-2",
+            credentials={
+                "auth_scheme": "sigv4",
+                "access_key_id": "AKIAIOSFODNN7EXAMPLE",
+                "secret_access_key": "secret",
+            },
+            base_url=endpoint_url,
+            use_proxy=False,
+        )
+
+    assert result is client
+    assert create_client.call_args.kwargs["endpoint_url"] == endpoint_url
+
+
+def test_resolve_runtime_bedrock_bearer():
+    manager = LLMProviderManager()
+    runtime = asyncio.run(
+        manager.resolve_runtime(
+            provider_id="amazon-bedrock",
+            model="global.anthropic.claude-haiku-4-5-20251001-v1:0",
+            api_key="bedrock-api-key-abc123",
+            base_url="https://bedrock-runtime.ap-northeast-1.amazonaws.com",
+        )
+    )
+
+    assert runtime["runtime"] == "bedrock"
+    assert runtime["aws_region"] == "ap-northeast-1"
+    assert runtime["aws_auth"]["auth_scheme"] == "bearer"
+
+
+def test_resolve_runtime_bedrock_sigv4_default_region():
+    manager = LLMProviderManager()
+    runtime = asyncio.run(
+        manager.resolve_runtime(
+            provider_id="amazon-bedrock",
+            model="us.anthropic.claude-sonnet-4-5-20250929-v1:0",
+            api_key="AKIAIOSFODNN7EXAMPLE:wJalrXUtnFEMI/K7MDENG",
+        )
+    )
+
+    assert runtime["runtime"] == "bedrock"
+    assert runtime["aws_region"] == "us-east-1"
+    assert runtime["aws_auth"]["auth_scheme"] == "sigv4"
+    assert runtime["aws_auth"]["access_key_id"] == "AKIAIOSFODNN7EXAMPLE"
+
+
+def test_resolve_runtime_bedrock_missing_credentials_rejected():
+    manager = LLMProviderManager()
+    with pytest.raises(LLMProviderAuthError):
+        asyncio.run(
+            manager.resolve_runtime(
+                provider_id="amazon-bedrock",
+                model="us.anthropic.claude-sonnet-4-5-20250929-v1:0",
+                api_key=None,
+            )
+        )
+
+
+def test_bedrock_model_matches_region():
+    """目录模型应按 Profile 分区及裸模型 ON_DEMAND Region 过滤"""
+    matches = LLMProviderManager._bedrock_model_matches_region
+
+    # 已知裸模型 ID 仅在其支持 ON_DEMAND 的 Region 保留
+    assert matches("anthropic.claude-3-5-sonnet-20241022-v2:0", "us-west-2")
+    assert matches("anthropic.claude-3-5-sonnet-20241022-v2:0", "ap-southeast-2")
+    assert not matches("anthropic.claude-3-5-sonnet-20241022-v2:0", "ap-northeast-1")
+    assert not matches("anthropic.claude-sonnet-4-5-20250929-v1:0", "us-west-2")
+    assert not matches("amazon.nova-premier-v1:0", "ap-northeast-1")
+    assert not matches("meta.llama4-maverick-17b-instruct-v1:0", "ap-northeast-1")
+    # 已确认支持 ON_DEMAND 的裸模型与 global Profile 维持可用
+    assert matches("amazon.nova-lite-v1:0", "ap-northeast-1")
+    assert matches("openai.gpt-oss-20b-1:0", "ap-northeast-1")
+    assert not matches("openai.gpt-oss-20b-1:0", "ap-southeast-1")
+    assert matches("global.anthropic.claude-sonnet-4-5-20250929-v1:0", "ap-northeast-1")
+    assert not matches(
+        "global.anthropic.claude-sonnet-4-5-20250929-v1:0",
+        "us-gov-west-1",
+    )
+    # 地理前缀只在对应分区 Region 可调用
+    assert matches("us.anthropic.claude-haiku-4-5-20251001-v1:0", "us-west-2")
+    assert not matches("us.anthropic.claude-haiku-4-5-20251001-v1:0", "ap-northeast-1")
+    assert not matches("us.anthropic.claude-haiku-4-5-20251001-v1:0", "us-gov-west-1")
+    assert matches("apac.amazon.nova-micro-v1:0", "ap-southeast-1")
+    assert not matches("apac.amazon.nova-micro-v1:0", "eu-central-1")
+    assert matches("eu.anthropic.claude-haiku-4-5-20251001-v1:0", "eu-central-1")
+    assert not matches("eu.anthropic.claude-haiku-4-5-20251001-v1:0", "us-east-1")
+    assert not matches("eu.anthropic.claude-haiku-4-5-20251001-v1:0", "eu-isoe-west-1")
+
+
+def test_bedrock_au_profile_matches_melbourne_region():
+    """AU Inference Profile 应允许从悉尼和墨尔本 Region 调用"""
+    matches = LLMProviderManager._bedrock_model_matches_region
+
+    assert matches("au.amazon.nova-lite-v1:0", "ap-southeast-2")
+    assert matches("au.amazon.nova-lite-v1:0", "ap-southeast-4")
+
+
+def test_list_models_bedrock_custom_endpoint_skips_control_plane():
+    """自定义 runtime 端点刷新模型时应直接使用离线目录"""
+    manager = LLMProviderManager()
+    manager._models_dev_data = {
+        "amazon-bedrock": {
+            "id": "amazon-bedrock",
+            "name": "Amazon Bedrock",
+            "models": {
+                "global.anthropic.claude-sonnet-4-5-20250929-v1:0": {
+                    "name": "Claude Sonnet 4.5 (Global)",
+                    "limit": {"context": 200000, "output": 64000},
+                },
+            },
+        }
+    }
+    manager._models_dev_loaded_at = time.time()
+
+    with patch.object(
+        LLMProviderManager,
+        "create_bedrock_client",
+        side_effect=AssertionError("不应访问控制面"),
+    ):
+        models = asyncio.run(
+            manager._list_models_from_bedrock(
+                api_key="bedrock-api-key-runtime-only",
+                base_url="https://bedrock-runtime-fips.us-east-1.amazonaws.com",
+                use_proxy=False,
+            )
+        )
+
+    assert [model["id"] for model in models] == [
+        "global.anthropic.claude-sonnet-4-5-20250929-v1:0"
+    ]
+
+
+def test_list_models_bedrock_keeps_control_plane_on_demand_models():
+    """控制面返回的 ON_DEMAND 基础模型不应被静态降级规则遗漏"""
+    manager = LLMProviderManager()
+    client = MagicMock()
+    client.get_paginator.return_value.paginate.return_value = [
+        {
+            "inferenceProfileSummaries": [
+                {
+                    "inferenceProfileId": (
+                        "global.anthropic.claude-sonnet-4-5-20250929-v1:0"
+                    ),
+                    "inferenceProfileName": "Claude Sonnet 4.5 (Global)",
+                    "status": "ACTIVE",
+                }
+            ]
+        }
+    ]
+    client.list_foundation_models.return_value = {
+        "modelSummaries": [
+            {
+                "modelId": "openai.gpt-oss-20b-1:0",
+                "modelName": "GPT OSS 20B",
+                "modelLifecycle": {"status": "ACTIVE"},
+            },
+            {
+                "modelId": "amazon.nova-lite-v1:0",
+                "modelName": "Nova Lite",
+                "modelLifecycle": {"status": "ACTIVE"},
+            },
+        ]
+    }
+
+    with patch.object(
+        LLMProviderManager, "create_bedrock_client", return_value=client
+    ):
+        models = asyncio.run(
+            manager._list_models_from_bedrock(
+                api_key="bedrock-api-key-runtime-only",
+                base_url="https://bedrock-runtime.ap-northeast-1.amazonaws.com",
+                use_proxy=False,
+            )
+        )
+
+    assert {model["id"] for model in models} == {
+        "amazon.nova-lite-v1:0",
+        "global.anthropic.claude-sonnet-4-5-20250929-v1:0",
+        "openai.gpt-oss-20b-1:0",
+    }
+    client.close.assert_called_once()
+
+
+def test_list_models_bedrock_falls_back_to_models_dev_on_control_plane_denial():
+    """控制面被拒（如 API Key 仅授权 bedrock-runtime）时降级 models.dev 目录"""
+    manager = LLMProviderManager()
+    # 预填 models.dev 内存缓存，降级路径不触发真实网络请求
+    manager._models_dev_data = {
+        "amazon-bedrock": {
+            "id": "amazon-bedrock",
+            "name": "Amazon Bedrock",
+            "models": {
+                "anthropic.claude-3-5-sonnet-20241022-v2:0": {
+                    "name": "Claude Sonnet 3.5 v2",
+                    "limit": {"context": 200000, "output": 8192},
+                },
+                "amazon.nova-lite-v1:0": {
+                    "name": "Nova Lite",
+                    "limit": {"context": 300000, "output": 5000},
+                },
+                "openai.gpt-oss-20b-1:0": {
+                    "name": "GPT OSS 20B",
+                    "limit": {"context": 131072, "output": 16384},
+                },
+                "apac.amazon.nova-lite-v1:0": {
+                    "name": "Nova Lite (APAC)",
+                    "limit": {"context": 300000, "output": 5000},
+                },
+                "meta.llama4-maverick-17b-instruct-v1:0": {
+                    "name": "Llama 4 Maverick",
+                    "limit": {"context": 1000000, "output": 8192},
+                },
+                "anthropic.claude-sonnet-4-5-20250929-v1:0": {
+                    "name": "Claude Sonnet 4.5",
+                    "limit": {"context": 200000, "output": 64000},
+                },
+                "global.anthropic.claude-sonnet-4-5-20250929-v1:0": {
+                    "name": "Claude Sonnet 4.5 (Global)",
+                    "limit": {"context": 200000, "output": 64000},
+                },
+                "us.anthropic.claude-haiku-4-5-20251001-v1:0": {
+                    "name": "Claude Haiku 4.5 (US)",
+                    "limit": {"context": 200000, "output": 64000},
+                },
+            },
+        }
+    }
+    manager._models_dev_loaded_at = time.time()
+
+    denied_client = MagicMock()
+    denied_client.get_paginator.side_effect = Exception(
+        "AccessDeniedException: not authorized to perform bedrock:ListInferenceProfiles"
+    )
+
+    with patch.object(
+        LLMProviderManager, "create_bedrock_client", return_value=denied_client
+    ):
+        models = asyncio.run(
+            manager._list_models_from_bedrock(
+                api_key="bedrock-api-key-runtime-only",
+                base_url="https://bedrock-runtime.ap-northeast-1.amazonaws.com",
+                use_proxy=False,
+            )
+        )
+
+    # 降级后仅保留东京 Region 可调用的裸模型与 Profile
+    model_ids = {m["id"] for m in models}
+    assert "anthropic.claude-3-5-sonnet-20241022-v2:0" not in model_ids
+    assert "amazon.nova-lite-v1:0" in model_ids
+    assert "openai.gpt-oss-20b-1:0" in model_ids
+    assert "apac.amazon.nova-lite-v1:0" in model_ids
+    assert "meta.llama4-maverick-17b-instruct-v1:0" not in model_ids
+    assert "global.anthropic.claude-sonnet-4-5-20250929-v1:0" in model_ids
+    assert "us.anthropic.claude-haiku-4-5-20251001-v1:0" not in model_ids
+    assert "anthropic.claude-sonnet-4-5-20250929-v1:0" not in model_ids
+    assert all(m["source"] == "models.dev" for m in models)
+    denied_client.close.assert_called_once()
diff --git a/tests/test_subscribe_files_info.py b/tests/test_subscribe_files_info.py
new file mode 100644
index 00000000..d672f1db
--- /dev/null
+++ b/tests/test_subscribe_files_info.py
@@ -0,0 +1,143 @@
+"""订阅文件统计相关测试"""
+
+from types import SimpleNamespace
+from unittest.mock import MagicMock, patch
+
+from app.chain.subscribe import SubscribeChain
+from app.modules.filemanager import FileManagerModule
+from app.schemas.mediaserver import ExistMediaInfo
+from app.schemas.types import MediaType
+
+
+def _build_subscribe(**overrides):
+    data = {
+        "id": 1,
+        "name": "Test Show",
+        "year": "2026",
+        "type": MediaType.TV.value,
+        "season": 1,
+        "tmdbid": None,
+        "doubanid": None,
+        "imdbid": None,
+        "tvdbid": None,
+        "bangumiid": None,
+        "episode_group": None,
+        "start_episode": 1,
+        "total_episode": 2,
+    }
+    data.update(overrides)
+    subscribe = SimpleNamespace(**data)
+    subscribe.to_dict = lambda: dict(data)
+    return subscribe
+
+
+def _build_mediainfo():
+    return SimpleNamespace(
+        type=MediaType.TV,
+        title="Test Show",
+        title_year="Test Show (2026)",
+        year="2026",
+        tmdb_id=None,
+        douban_id=None,
+    )
+
+
+def test_filemanager_media_exists_skips_local_when_server_specified():
+    module = FileManagerModule()
+    mediainfo = _build_mediainfo()
+
+    with patch.object(module, "media_files", return_value=[SimpleNamespace(path="/media/test.mkv")]) as media_files:
+        result = module.media_exists(mediainfo, server="Emby1")
+
+    assert result is None
+    media_files.assert_not_called()
+
+
+def test_subscribe_files_info_merges_multiple_mediaservers():
+    subscribe = _build_subscribe(season=1, total_episode=2)
+    mediainfo = _build_mediainfo()
+
+    def _media_exists_side_effect(*, mediainfo, server=None, **kwargs):
+        if server == "Emby1":
+            return ExistMediaInfo(
+                type=MediaType.TV,
+                seasons={1: [1]},
+                server_type="emby",
+                server="Emby1",
+                itemid="emby-series",
+            )
+        if server == "Jellyfin1":
+            return ExistMediaInfo(
+                type=MediaType.TV,
+                seasons={1: [1]},
+                server_type="jellyfin",
+                server="Jellyfin1",
+                itemid="jf-series",
+            )
+        return None
+
+    helper = MagicMock()
+    helper.get_services.return_value = {"Emby1": object(), "Jellyfin1": object()}
+
+    mediaserver_chain = MagicMock()
+    mediaserver_chain.get_play_url.side_effect = lambda server, item_id: f"https://{server}/item/{item_id}"
+    mediaserver_chain.get_season_episode_ids.side_effect = lambda server, item_id, season: {1: f"{item_id}-ep1"}
+
+    chain = SubscribeChain()
+    with patch("app.chain.subscribe.DownloadHistoryOper") as download_oper, \
+            patch.object(chain, "recognize_media", return_value=mediainfo), \
+            patch.object(chain, "media_files", return_value=None), \
+            patch.object(chain, "media_exists", side_effect=_media_exists_side_effect), \
+            patch("app.chain.subscribe.MediaServerHelper", return_value=helper), \
+            patch("app.chain.subscribe.MediaServerChain", return_value=mediaserver_chain), \
+            patch("app.chain.subscribe.Subscribe", side_effect=lambda **kwargs: SimpleNamespace(**kwargs)):
+        download_oper.return_value.get_by_mediaid.return_value = []
+        result = chain.subscribe_files_info(subscribe)
+
+    library = result.episodes[1].library
+    servers = {item.server for item in library}
+    assert servers == {"Emby1", "Jellyfin1"}
+    assert all(str(item.file_path).startswith("https://") for item in library)
+
+
+def test_subscribe_files_info_uses_season_zero_for_tv():
+    subscribe = _build_subscribe(season=0, total_episode=1, start_episode=1)
+    mediainfo = _build_mediainfo()
+    captured_seasons = []
+
+    def _media_exists_side_effect(*, mediainfo, server=None, **kwargs):
+        if server == "Emby1":
+            return ExistMediaInfo(
+                type=MediaType.TV,
+                seasons={0: [1]},
+                server_type="emby",
+                server="Emby1",
+                itemid="emby-special",
+            )
+        return None
+
+    def _get_season_episode_ids(server, item_id, season):
+        captured_seasons.append(season)
+        return {1: f"{item_id}-ep1"}
+
+    helper = MagicMock()
+    helper.get_services.return_value = {"Emby1": object()}
+
+    mediaserver_chain = MagicMock()
+    mediaserver_chain.get_play_url.return_value = "https://emby/item/1"
+    mediaserver_chain.get_season_episode_ids.side_effect = _get_season_episode_ids
+
+    chain = SubscribeChain()
+    with patch("app.chain.subscribe.DownloadHistoryOper") as download_oper, \
+            patch.object(chain, "recognize_media", return_value=mediainfo), \
+            patch.object(chain, "media_files", return_value=None), \
+            patch.object(chain, "media_exists", side_effect=_media_exists_side_effect), \
+            patch("app.chain.subscribe.MediaServerHelper", return_value=helper), \
+            patch("app.chain.subscribe.MediaServerChain", return_value=mediaserver_chain), \
+            patch("app.chain.subscribe.Subscribe", side_effect=lambda **kwargs: SimpleNamespace(**kwargs)):
+        download_oper.return_value.get_by_mediaid.return_value = []
+        result = chain.subscribe_files_info(subscribe)
+
+    assert captured_seasons == [0]
+    assert len(result.episodes[1].library) == 1
+    assert result.episodes[1].library[0].server == "Emby1"
diff --git a/version.py b/version.py
index 35c1b00c..79b83442 100644
--- a/version.py
+++ b/version.py
@@ -1,2 +1,2 @@
-APP_VERSION = 'v2.14.3'
-FRONTEND_VERSION = 'v2.14.3'
+APP_VERSION = 'v2.14.4'
+FRONTEND_VERSION = 'v2.14.4'