From 6bad9c30c70c72ad934d1ff1a1713a478f6b35bb Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?lollipopkit=F0=9F=8F=B3=EF=B8=8F=E2=80=8D=E2=9A=A7?=
 =?UTF-8?q?=EF=B8=8F?= <10864310+lollipopkit@users.noreply.github.com>
Date: Mon, 26 Jan 2026 19:28:01 +0800
Subject: [PATCH 1/6] feat(devloop): merge coderabbit into custom review

---
 plugins/devloop/README.md                  |  5 ++--
 plugins/devloop/README.zh-CN.md            |  5 ++--
 plugins/devloop/agents/devloop-runner.md   | 10 +++----
 plugins/devloop/commands/devloop-enable.md |  6 ++--
 plugins/devloop/commands/devloop.md        |  6 ++--
 plugins/devloop/skills/coderabbit/SKILL.md | 34 ----------------------
 6 files changed, 19 insertions(+), 47 deletions(-)
 delete mode 100644 plugins/devloop/skills/coderabbit/SKILL.md

diff --git a/plugins/devloop/README.md b/plugins/devloop/README.md
index 5916513..bb6d071 100644
--- a/plugins/devloop/README.md
+++ b/plugins/devloop/README.md
@@ -41,13 +41,14 @@ enabled: true
 base_branch: "main"
 
 # Review behavior
-review_mode: "github"   # github|coderabbit|local-agent|custom
+review_mode: "github"   # github|custom
+custom_review_skill: ""  # optional; e.g. "coderabbit:review"
 max_review_polls: 40
 review_poll_seconds: 60
 
 # Wait for review behavior
 wait_behavior: "poll"   # poll|ping_ai
-ai_reviewer_id: "coderabbitai"
+ai_reviewer_id: ""
 ping_message_template: "@{{ai_id}} This PR is awaiting review feedback. Could you provide an update?"
 ping_threshold: 3       # number of wait rounds before pinging (minimum 1)
 
diff --git a/plugins/devloop/README.zh-CN.md b/plugins/devloop/README.zh-CN.md
index 90d6896..82e319d 100644
--- a/plugins/devloop/README.zh-CN.md
+++ b/plugins/devloop/README.zh-CN.md
@@ -41,13 +41,14 @@ enabled: true
 base_branch: "main"
 
 # Review behavior
-review_mode: "github"   # github|coderabbit|local-agent|custom
+review_mode: "github"   # github|custom
+custom_review_skill: ""  # 可选；例如 "coderabbit:review"
 max_review_polls: 40
 review_poll_seconds: 60
 
 # Wait for review behavior
 wait_behavior: "poll"   # poll|ping_ai
-ai_reviewer_id: "coderabbitai"
+ai_reviewer_id: ""
 ping_message_template: "@{{ai_id}} This PR is awaiting review feedback. Could you provide an update?"
 ping_threshold: 3       # number of wait rounds before pinging (minimum 1)
 
diff --git a/plugins/devloop/agents/devloop-runner.md b/plugins/devloop/agents/devloop-runner.md
index b5f8927..1052d7d 100644
--- a/plugins/devloop/agents/devloop-runner.md
+++ b/plugins/devloop/agents/devloop-runner.md
@@ -81,8 +81,8 @@ Settings:
 - Parse YAML frontmatter for configuration (enabled, notification settings, review mode, wait_behavior, ping_threshold, ai_reviewer_id, ping_message_template, polling limits, workspace_mode).
   - `review_mode`:
     - `"github"` (default): Poll for GitHub review comments.
-    - `"coderabbit"`: Proactively trigger review using the external `coderabbit:review` skill (provided by the CodeRabbit Claude Code plugin; not implemented by devloop). See `plugins/devloop/skills/coderabbit/SKILL.md`.
-    - `"local-agent"` / `"custom"`: Placeholder for other modes.
+    - `"custom"`: Use a user-provided review skill each polling cycle (via `custom_review_skill`).
+      - Example: `custom_review_skill: "coderabbit:review"` (requires the CodeRabbit plugin).
   - `workspace_mode`: set to `"gws"` to enable integration with `git-ws` for isolated workspaces and locking.
 
 Default completion criteria (unless overridden by settings):
@@ -161,8 +161,8 @@ Workflow (repeat until completion or blocked):
         - Ensure `ai_reviewer_id` is set. If not, log a warning and fall back to `wait_behavior = "poll"`.
         - Ensure `ping_threshold` is at least 1. If not, default it to 3.
      3. **Review Round**:
-        - If `review_mode` is `"coderabbit"`:
-          - Trigger `coderabbit:review` once at the start of each polling cycle.
+        - If `review_mode` is `"custom"` and `custom_review_skill` is set:
+          - Trigger `custom_review_skill` once at the start of each polling cycle.
           - If the Skill call fails (not installed, not authenticated, or errors), proceed with standard GitHub polling.
           - If the review produced findings, treat them as new review feedback and proceed to **Apply feedback** (skip the remaining polling steps in this round).
         - Poll for new bot/AI review comments, review state, and mergeability status.
@@ -204,7 +204,7 @@ Workflow (repeat until completion or blocked):
             --jq '.data.repository.pullRequest.reviewThreads.nodes[] | select(.isOutdated == false and .isResolved == false) | .comments.nodes[]'
           ```
 
-     4. If the review round produced no findings (coderabbit mode) AND no new GitHub comments are found:
+     4. If the review round produced no findings (custom review skill mode) AND no new GitHub comments are found:
         - Increment `wait_rounds_without_response`.
         - If `wait_behavior` is `ping_ai`, `wait_rounds_without_response` >= `ping_threshold`, and `pings_sent` < 2:
           - Post a comment to the PR:
diff --git a/plugins/devloop/commands/devloop-enable.md b/plugins/devloop/commands/devloop-enable.md
index a39a5c2..238071e 100644
--- a/plugins/devloop/commands/devloop-enable.md
+++ b/plugins/devloop/commands/devloop-enable.md
@@ -13,10 +13,12 @@ Steps:
    - `enabled: true`
 3. If it does not exist, create it with a minimal template and safe defaults.
 4. Ask user for:
-   - `review_mode` (github|coderabbit|local-agent|custom, default: `github`)
+   - `review_mode` (github|custom, default: `github`)
+     - If `review_mode` is `custom`:
+       - `custom_review_skill` (optional; e.g., `coderabbit:review` if you have the CodeRabbit plugin installed)
    - `wait_behavior` (poll|ping_ai, default: `poll`)
    - If `wait_behavior` is `ping_ai`:
-     - `ai_reviewer_id` (e.g., `coderabbitai`)
+     - `ai_reviewer_id` (e.g., a bot account/login to ping)
      - `ping_message_template` (default: `@{{ai_id}} This PR is awaiting review feedback. Could you provide an update?`)
      - `ping_threshold` (number of wait rounds with no response before pinging, default: 3)
    - `notify_enabled` (true/false)
diff --git a/plugins/devloop/commands/devloop.md b/plugins/devloop/commands/devloop.md
index b030df8..86582f5 100644
--- a/plugins/devloop/commands/devloop.md
+++ b/plugins/devloop/commands/devloop.md
@@ -40,11 +40,13 @@ Run the devloop workflow using the plugin components in this plugin. This comman
 
    - `enabled: true|false`
    - `base_branch: "main"`
-   - `review_mode: "github"|"coderabbit"|"local-agent"|"custom"`
+   - `review_mode: "github"|"custom"`
+     - If `review_mode` is `"custom"`, you may set:
+       - `custom_review_skill: "..."` (e.g., `coderabbit:review`)
    - `max_review_polls: 40`
    - `review_poll_seconds: 60`
    - `wait_behavior: "poll"|"ping_ai"`
-   - `ai_reviewer_id: "..."` (e.g., `coderabbitai`)
+   - `ai_reviewer_id: "..."` (e.g., a bot account/login to ping)
    - `ping_message_template: "..."`
    - `ping_threshold: 3`
    - `notify_enabled: true|false`
diff --git a/plugins/devloop/skills/coderabbit/SKILL.md b/plugins/devloop/skills/coderabbit/SKILL.md
deleted file mode 100644
index ca75125..0000000
--- a/plugins/devloop/skills/coderabbit/SKILL.md
+++ /dev/null
@@ -1,34 +0,0 @@
-# coderabbit:review
-
-This is an **external** skill provided by the **CodeRabbit Claude Code plugin** (which wraps the CodeRabbit CLI). The devloop plugin can invoke this skill when `review_mode: "coderabbit"` is selected.
-
-Devloop does **not** implement CodeRabbit review itself.
-
-## When devloop uses this skill
-
-- During the **Wait for review** phase, if `review_mode` is set to `"coderabbit"`.
-- Devloop triggers CodeRabbit once at the start of each polling cycle and then:
-  - If findings are returned, treats them as review feedback and proceeds to **Apply feedback**.
-  - If the skill is unavailable or fails (not installed, not authenticated, errors), devloop falls back to standard GitHub polling.
-
-## Prerequisites
-
-- Install and authenticate CodeRabbit CLI as required by the CodeRabbit plugin.
-- Install the CodeRabbit Claude Code plugin so that the `coderabbit:review` skill is available.
-
-## Invocation
-
-In Claude Code, this skill is invoked via the `Skill` tool:
-
-- `coderabbit:review`
-
-## Expected output
-
-The skill returns a structured review summary and findings (severity grouped, suggestions, etc.). Devloop uses the returned findings as actionable feedback.
-
-## Failure modes
-
-- CodeRabbit plugin not installed → `coderabbit:review` not available.
-- CodeRabbit CLI not installed/authenticated → review invocation fails.
-
-In these cases, devloop should continue with GitHub polling instead of stopping.

From 020fb48f918afa6ccacf9591c27cf9e0323a9525 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?lollipopkit=F0=9F=8F=B3=EF=B8=8F=E2=80=8D=E2=9A=A7?=
 =?UTF-8?q?=EF=B8=8F?= <10864310+lollipopkit@users.noreply.github.com>
Date: Mon, 26 Jan 2026 20:12:54 +0800
Subject: [PATCH 2/6] feat(devloop): script gh api graphql query

---
 plugins/devloop/README.md                     |   2 +-
 plugins/devloop/README.zh-CN.md               |   2 +-
 plugins/devloop/agents/devloop-runner.md      |  29 +----
 plugins/devloop/commands/devloop.md           |  14 +--
 .../scripts/devloop-pr-review-threads.sh      | 103 ++++++++++++++++++
 5 files changed, 109 insertions(+), 41 deletions(-)
 create mode 100644 plugins/devloop/scripts/devloop-pr-review-threads.sh

diff --git a/plugins/devloop/README.md b/plugins/devloop/README.md
index bb6d071..efe4ecc 100644
--- a/plugins/devloop/README.md
+++ b/plugins/devloop/README.md
@@ -59,5 +59,5 @@ notify_on_stop: true
 notify_command_template: ""      # executed with selected shell; can reference env vars below
 ---
 
-Additional instructions for devloop can go here.
+Genrated by <https://github.com/lollipopkit/cc-plugins>
 ```
diff --git a/plugins/devloop/README.zh-CN.md b/plugins/devloop/README.zh-CN.md
index 82e319d..ac26586 100644
--- a/plugins/devloop/README.zh-CN.md
+++ b/plugins/devloop/README.zh-CN.md
@@ -59,5 +59,5 @@ notify_on_stop: true
 notify_command_template: ""      # executed with selected shell; can reference env vars below
 ---
 
-这里可以写给 devloop 的额外说明。
+由 <https://github.com/lollipopkit/cc-plugins> 生成
 ```
diff --git a/plugins/devloop/agents/devloop-runner.md b/plugins/devloop/agents/devloop-runner.md
index 1052d7d..b7d97ce 100644
--- a/plugins/devloop/agents/devloop-runner.md
+++ b/plugins/devloop/agents/devloop-runner.md
@@ -173,35 +173,10 @@ Workflow (repeat until completion or blocked):
             - Notify the user that the PR is a draft and may not receive reviews until marked as ready.
             - Continue polling but skip ping/notify attempts until the PR is marked ready for review.
         - Use GraphQL to filter out outdated and resolved comments to ensure you only address active feedback.
-        - Replace `{owner}`, `{repo}`, and `{number}` with real values before running the query. Example:
+        - Use the helper script (outputs JSONL; one JSON object per line):
 
           ```bash
-          PR_NUMBER=$(gh pr view --json number --jq '.number')
-          REPO_OWNER=$(gh repo view --json owner --jq '.owner.login')
-          REPO_NAME=$(gh repo view --json name --jq '.name')
-
-          gh api graphql -F owner="$REPO_OWNER" -F name="$REPO_NAME" -F pr="$PR_NUMBER" -f query=' \
-            query($name: String!, $owner: String!, $pr: Int!) { \
-              repository(owner: $owner, name: $name) { \
-                pullRequest(number: $pr) { \
-                  reviewThreads(first: 100) { \
-                    nodes { \
-                      isOutdated \
-                      isResolved \
-                      comments(last: 20) { \
-                        nodes { \
-                          body \
-                          path \
-                          line \
-                          author { login } \
-                        } \
-                      } \
-                    } \
-                  } \
-                } \
-              } \
-            }' \
-            --jq '.data.repository.pullRequest.reviewThreads.nodes[] | select(.isOutdated == false and .isResolved == false) | .comments.nodes[]'
+          bash "${CLAUDE_PLUGIN_ROOT}/scripts/devloop-pr-review-threads.sh" --repo "$(gh repo view --json nameWithOwner --jq '.nameWithOwner')" --pr "$(gh pr view --json number --jq '.number')"
           ```
 
      4. If the review round produced no findings (custom review skill mode) AND no new GitHub comments are found:
diff --git a/plugins/devloop/commands/devloop.md b/plugins/devloop/commands/devloop.md
index 86582f5..553214d 100644
--- a/plugins/devloop/commands/devloop.md
+++ b/plugins/devloop/commands/devloop.md
@@ -62,20 +62,10 @@ Run the devloop workflow using the plugin components in this plugin. This comman
 - **Branch Logic**: Create a new branch based on the issue content ONLY if the current branch is the base branch.
 - **Git Protocol**: NEVER use `git push --force`, `git push -f`, or `git commit --amend` on branches that have already been pushed to the remote or have an open PR. Always create new commits and use standard `git push`.
 - **Review Polling**: The agent will remain in an autonomous polling loop using `sleep` between polls.
-  - GraphQL for filtering comments:
+  - GraphQL for filtering active (not outdated/resolved) review thread comments is wrapped in a helper script:
 
     ```bash
-    gh api graphql -F owner='{owner}' -F name='{repo}' -F pr={number} -f query='
-      query($name: String!, $owner: String!, $pr: Int!) {
-        repository(owner: $owner, name: $name) {
-          pullRequest(number: $pr) {
-            reviewThreads(first: 100) {
-              nodes { isOutdated isResolved comments(last: 20) { nodes { body path line author { login } } } }
-            }
-          }
-        }
-      }
-    ' --jq '.data.repository.pullRequest.reviewThreads.nodes[] | select(.isOutdated == false and .isResolved == false) | .comments.nodes[]'
+    bash "${CLAUDE_PLUGIN_ROOT}/scripts/devloop-pr-review-threads.sh" --repo "{owner}/{repo}" --pr {number}
     ```
 
 - **Presence & Communication (MANDATORY)**:
diff --git a/plugins/devloop/scripts/devloop-pr-review-threads.sh b/plugins/devloop/scripts/devloop-pr-review-threads.sh
new file mode 100644
index 0000000..2bfc261
--- /dev/null
+++ b/plugins/devloop/scripts/devloop-pr-review-threads.sh
@@ -0,0 +1,103 @@
+#!/bin/bash
+set -euo pipefail
+
+# Fetch active (not outdated, not resolved) PR review thread comments via GitHub GraphQL.
+#
+# Outputs JSON Lines (one JSON object per line), suitable for piping.
+#
+# Usage:
+#   bash "${CLAUDE_PLUGIN_ROOT}/scripts/devloop-pr-review-threads.sh" --repo owner/name --pr 123
+#
+# Notes:
+# - Requires `gh` authenticated.
+# - If --repo or --pr is omitted, will infer from current repo / current PR when possible.
+
+usage() {
+  cat <<'EOF'
+Usage:
+  devloop-pr-review-threads.sh [--repo owner/name] [--pr <number>] [--json] [--jq <filter>]
+
+Options:
+  --repo owner/name   GitHub repository (default: current repo)
+  --pr <number>       Pull request number (default: current PR number)
+  --json              Output JSONL (default)
+  --jq <filter>       Apply jq filter to raw GraphQL response (advanced)
+
+Examples:
+  bash "${CLAUDE_PLUGIN_ROOT}/scripts/devloop-pr-review-threads.sh" --repo lollipopkit/cc-plugins --pr 33
+EOF
+}
+
+REPO=""
+PR=""
+JQ_FILTER=""
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --repo)
+      REPO="${2:-}"
+      shift 2
+      ;;
+    --pr)
+      PR="${2:-}"
+      shift 2
+      ;;
+    --jq)
+      JQ_FILTER="${2:-}"
+      shift 2
+      ;;
+    --json)
+      shift
+      ;;
+    -h|--help)
+      usage
+      exit 0
+      ;;
+    *)
+      echo "Unknown arg: $1" >&2
+      usage >&2
+      exit 2
+      ;;
+  esac
+done
+
+if [[ -z "$REPO" ]]; then
+  REPO=$(gh repo view --json nameWithOwner --jq '.nameWithOwner')
+fi
+
+if [[ -z "$PR" ]]; then
+  PR=$(gh pr view --json number --jq '.number')
+fi
+
+OWNER="${REPO%%/*}"
+NAME="${REPO#*/}"
+
+QUERY='query($name: String!, $owner: String!, $pr: Int!) {
+  repository(owner: $owner, name: $name) {
+    pullRequest(number: $pr) {
+      reviewThreads(first: 100) {
+        nodes {
+          isOutdated
+          isResolved
+          comments(last: 20) {
+            nodes {
+              body
+              path
+              line
+              author { login }
+            }
+          }
+        }
+      }
+    }
+  }
+}'
+
+if [[ -n "$JQ_FILTER" ]]; then
+  gh api graphql -F owner="$OWNER" -F name="$NAME" -F pr="$PR" -f query="$QUERY" --jq "$JQ_FILTER"
+  exit $?
+fi
+
+# Default: output JSONL of active comments (one JSON object per line).
+gh api graphql -F owner="$OWNER" -F name="$NAME" -F pr="$PR" -f query="$QUERY" \
+  --jq '.data.repository.pullRequest.reviewThreads.nodes[] | select(.isOutdated == false and .isResolved == false) | .comments.nodes[]'

From ad306fd8881a0d24772ba0c56658778029500a30 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?lollipopkit=F0=9F=8F=B3=EF=B8=8F=E2=80=8D=E2=9A=A7?=
 =?UTF-8?q?=EF=B8=8F?= <10864310+lollipopkit@users.noreply.github.com>
Date: Mon, 26 Jan 2026 20:55:47 +0800
Subject: [PATCH 3/6] fix(devloop): paginate reviewThreads query

---
 plugins/devloop/README.md                     |   2 +-
 .../scripts/devloop-pr-review-threads.sh      | 169 ++++++++++++++++--
 2 files changed, 156 insertions(+), 15 deletions(-)

diff --git a/plugins/devloop/README.md b/plugins/devloop/README.md
index efe4ecc..d4dd2fb 100644
--- a/plugins/devloop/README.md
+++ b/plugins/devloop/README.md
@@ -59,5 +59,5 @@ notify_on_stop: true
 notify_command_template: ""      # executed with selected shell; can reference env vars below
 ---
 
-Genrated by <https://github.com/lollipopkit/cc-plugins>
+Generated by <https://github.com/lollipopkit/cc-plugins>
 ```
diff --git a/plugins/devloop/scripts/devloop-pr-review-threads.sh b/plugins/devloop/scripts/devloop-pr-review-threads.sh
index 2bfc261..6bb9a4a 100644
--- a/plugins/devloop/scripts/devloop-pr-review-threads.sh
+++ b/plugins/devloop/scripts/devloop-pr-review-threads.sh
@@ -11,6 +11,7 @@ set -euo pipefail
 # Notes:
 # - Requires `gh` authenticated.
 # - If --repo or --pr is omitted, will infer from current repo / current PR when possible.
+# - Implements cursor-based pagination for both reviewThreads and thread comments.
 
 usage() {
   cat <<'EOF'
@@ -21,7 +22,7 @@ Options:
   --repo owner/name   GitHub repository (default: current repo)
   --pr <number>       Pull request number (default: current PR number)
   --json              Output JSONL (default)
-  --jq <filter>       Apply jq filter to raw GraphQL response (advanced)
+  --jq <filter>       Apply jq filter to a single raw GraphQL response (advanced; no pagination)
 
 Examples:
   bash "${CLAUDE_PLUGIN_ROOT}/scripts/devloop-pr-review-threads.sh" --repo lollipopkit/cc-plugins --pr 33
@@ -72,20 +73,17 @@ fi
 OWNER="${REPO%%/*}"
 NAME="${REPO#*/}"
 
-QUERY='query($name: String!, $owner: String!, $pr: Int!) {
+# Advanced: allow a user-provided --jq against a single response (no pagination).
+if [[ -n "$JQ_FILTER" ]]; then
+  QUERY_SINGLE='query($name: String!, $owner: String!, $pr: Int!) {
   repository(owner: $owner, name: $name) {
     pullRequest(number: $pr) {
       reviewThreads(first: 100) {
         nodes {
           isOutdated
           isResolved
-          comments(last: 20) {
-            nodes {
-              body
-              path
-              line
-              author { login }
-            }
+          comments(first: 100) {
+            nodes { body path line author { login } }
           }
         }
       }
@@ -93,11 +91,154 @@ QUERY='query($name: String!, $owner: String!, $pr: Int!) {
   }
 }'
 
-if [[ -n "$JQ_FILTER" ]]; then
-  gh api graphql -F owner="$OWNER" -F name="$NAME" -F pr="$PR" -f query="$QUERY" --jq "$JQ_FILTER"
+  gh api graphql -F owner="$OWNER" -F name="$NAME" -F pr="$PR" -f query="$QUERY_SINGLE" --jq "$JQ_FILTER"
   exit $?
 fi
 
-# Default: output JSONL of active comments (one JSON object per line).
-gh api graphql -F owner="$OWNER" -F name="$NAME" -F pr="$PR" -f query="$QUERY" \
-  --jq '.data.repository.pullRequest.reviewThreads.nodes[] | select(.isOutdated == false and .isResolved == false) | .comments.nodes[]'
+if ! command -v python3 >/dev/null 2>&1; then
+  echo "python3 is required for pagination/JSONL output" >&2
+  exit 127
+fi
+
+OWNER="$OWNER" NAME="$NAME" PR="$PR" python3 - <<'PY'
+import json
+import os
+import subprocess
+import sys
+from typing import Any, Dict, Optional
+
+OWNER = os.environ["OWNER"]
+NAME = os.environ["NAME"]
+PR = os.environ["PR"]
+
+QUERY_THREADS = r"""
+query($name: String!, $owner: String!, $pr: Int!, $threadsAfter: String) {
+  repository(owner: $owner, name: $name) {
+    pullRequest(number: $pr) {
+      reviewThreads(first: 100, after: $threadsAfter) {
+        pageInfo { hasNextPage endCursor }
+        nodes {
+          id
+          isOutdated
+          isResolved
+          comments(first: 100) {
+            pageInfo { hasNextPage endCursor }
+            nodes { body path line author { login } }
+          }
+        }
+      }
+    }
+  }
+}
+"""
+
+QUERY_THREAD_COMMENTS = r"""
+query($id: ID!, $commentsAfter: String) {
+  node(id: $id) {
+    ... on PullRequestReviewThread {
+      id
+      isOutdated
+      isResolved
+      comments(first: 100, after: $commentsAfter) {
+        pageInfo { hasNextPage endCursor }
+        nodes { body path line author { login } }
+      }
+    }
+  }
+}
+"""
+
+
+def gh_graphql(query: str, variables: Dict[str, Any]) -> Dict[str, Any]:
+    cmd = ["gh", "api", "graphql"]
+    for k, v in variables.items():
+        if v is None:
+            continue
+        cmd += ["-F", f"{k}={v}"]
+    cmd += ["-f", f"query={query}"]
+
+    res = subprocess.run(cmd, capture_output=True, text=True)
+    if res.returncode != 0:
+        sys.stderr.write(res.stderr)
+        raise SystemExit(res.returncode)
+
+    try:
+        return json.loads(res.stdout)
+    except Exception:
+        sys.stderr.write(res.stdout)
+        raise
+
+
+def emit_jsonl(obj: Any) -> None:
+    sys.stdout.write(json.dumps(obj, ensure_ascii=False))
+    sys.stdout.write("\n")
+
+
+def main() -> int:
+    threads_after: Optional[str] = None
+
+    while True:
+        resp = gh_graphql(
+            QUERY_THREADS,
+            {
+                "owner": OWNER,
+                "name": NAME,
+                "pr": PR,
+                "threadsAfter": threads_after,
+            },
+        )
+
+        threads = (
+            resp.get("data", {})
+            .get("repository", {})
+            .get("pullRequest", {})
+            .get("reviewThreads", {})
+        )
+
+        for thread in threads.get("nodes", []) or []:
+            if thread.get("isOutdated") or thread.get("isResolved"):
+                continue
+
+            thread_id = thread.get("id")
+            comments_conn = thread.get("comments") or {}
+            for c in comments_conn.get("nodes", []) or []:
+                emit_jsonl(c)
+
+            page = comments_conn.get("pageInfo") or {}
+            comments_after = page.get("endCursor") if page.get("hasNextPage") else None
+
+            while comments_after and thread_id:
+                resp2 = gh_graphql(
+                    QUERY_THREAD_COMMENTS,
+                    {
+                        "id": thread_id,
+                        "commentsAfter": comments_after,
+                    },
+                )
+
+                node = (resp2.get("data", {}) or {}).get("node")
+                if not node:
+                    break
+                if node.get("isOutdated") or node.get("isResolved"):
+                    break
+
+                comments_conn2 = node.get("comments") or {}
+                for c in comments_conn2.get("nodes", []) or []:
+                    emit_jsonl(c)
+
+                page2 = comments_conn2.get("pageInfo") or {}
+                comments_after = page2.get("endCursor") if page2.get("hasNextPage") else None
+
+        threads_page = threads.get("pageInfo") or {}
+        if threads_page.get("hasNextPage"):
+            threads_after = threads_page.get("endCursor")
+            continue
+
+        break
+
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
+PY

From 9c15b6f21ad8cca998fda2deb804e66a6db52022 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?lollipopkit=F0=9F=8F=B3=EF=B8=8F=E2=80=8D=E2=9A=A7?=
 =?UTF-8?q?=EF=B8=8F?= <10864310+lollipopkit@users.noreply.github.com>
Date: Mon, 26 Jan 2026 21:10:52 +0800
Subject: [PATCH 4/6] fix(devloop): harden pr review threads script

---
 .../scripts/devloop-pr-review-threads.sh      | 35 ++++++++++++++++---
 1 file changed, 31 insertions(+), 4 deletions(-)

diff --git a/plugins/devloop/scripts/devloop-pr-review-threads.sh b/plugins/devloop/scripts/devloop-pr-review-threads.sh
index 6bb9a4a..f3e69a6 100644
--- a/plugins/devloop/scripts/devloop-pr-review-threads.sh
+++ b/plugins/devloop/scripts/devloop-pr-review-threads.sh
@@ -33,18 +33,32 @@ REPO=""
 PR=""
 JQ_FILTER=""
 
+need_value() {
+  local flag="$1"
+  local value="${2:-}"
+
+  if [[ -z "$value" || "$value" == -* ]]; then
+    echo "Missing value for $flag" >&2
+    usage >&2
+    exit 2
+  fi
+}
+
 while [[ $# -gt 0 ]]; do
   case "$1" in
     --repo)
-      REPO="${2:-}"
+      need_value "--repo" "${2:-}"
+      REPO="$2"
       shift 2
       ;;
     --pr)
-      PR="${2:-}"
+      need_value "--pr" "${2:-}"
+      PR="$2"
       shift 2
       ;;
     --jq)
-      JQ_FILTER="${2:-}"
+      need_value "--jq" "${2:-}"
+      JQ_FILTER="$2"
       shift 2
       ;;
     --json)
@@ -62,6 +76,11 @@ while [[ $# -gt 0 ]]; do
   esac
 done
 
+if ! command -v gh >/dev/null 2>&1; then
+  echo "gh is required" >&2
+  exit 127
+fi
+
 if [[ -z "$REPO" ]]; then
   REPO=$(gh repo view --json nameWithOwner --jq '.nameWithOwner')
 fi
@@ -163,11 +182,19 @@ def gh_graphql(query: str, variables: Dict[str, Any]) -> Dict[str, Any]:
         raise SystemExit(res.returncode)
 
     try:
-        return json.loads(res.stdout)
+        data = json.loads(res.stdout)
     except Exception:
         sys.stderr.write(res.stdout)
         raise
 
+    errors = data.get("errors")
+    if errors:
+        sys.stderr.write(json.dumps(errors, ensure_ascii=False, indent=2))
+        sys.stderr.write("\n")
+        raise SystemExit(1)
+
+    return data
+
 
 def emit_jsonl(obj: Any) -> None:
     sys.stdout.write(json.dumps(obj, ensure_ascii=False))

From 5e2b002f8f721131324f104478dc0da472eb30f0 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?lollipopkit=F0=9F=8F=B3=EF=B8=8F=E2=80=8D=E2=9A=A7?=
 =?UTF-8?q?=EF=B8=8F?= <10864310+lollipopkit@users.noreply.github.com>
Date: Mon, 26 Jan 2026 21:22:47 +0800
Subject: [PATCH 5/6] fix(devloop): validate --repo owner/name

---
 plugins/devloop/scripts/devloop-pr-review-threads.sh | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/plugins/devloop/scripts/devloop-pr-review-threads.sh b/plugins/devloop/scripts/devloop-pr-review-threads.sh
index f3e69a6..9bcae5d 100644
--- a/plugins/devloop/scripts/devloop-pr-review-threads.sh
+++ b/plugins/devloop/scripts/devloop-pr-review-threads.sh
@@ -89,6 +89,11 @@ if [[ -z "$PR" ]]; then
   PR=$(gh pr view --json number --jq '.number')
 fi
 
+if [[ "$REPO" != */* || "${REPO#*/}" == *"/"* ]]; then
+  echo "Invalid --repo format: expected owner/name" >&2
+  exit 2
+fi
+
 OWNER="${REPO%%/*}"
 NAME="${REPO#*/}"
 

From 5d381fa4a6d430057348410cf38036622a080596 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?lollipopkit=F0=9F=8F=B3=EF=B8=8F=E2=80=8D=E2=9A=A7?=
 =?UTF-8?q?=EF=B8=8F?= <10864310+lollipopkit@users.noreply.github.com>
Date: Wed, 28 Jan 2026 00:45:05 +0800
Subject: [PATCH 6/6] opt.

---
 plugins/devloop/README.md                     |  52 +--
 plugins/devloop/agents/devloop-implementer.md |  27 +-
 plugins/devloop/agents/devloop-runner.md      | 383 +++++++++---------
 plugins/devloop/agents/devloop-validator.md   |  26 +-
 plugins/devloop/commands/devloop-enable.md    |  79 ++--
 plugins/devloop/commands/devloop.md           |  87 +---
 .../scripts/devloop-pr-review-threads.sh      |   6 +-
 .../skills/feishu-lark-interaction/SKILL.md   |  95 +++++
 .../skills/github-interaction/SKILL.md        |  80 ++++
 9 files changed, 497 insertions(+), 338 deletions(-)
 mode change 100644 => 100755 plugins/devloop/scripts/devloop-pr-review-threads.sh
 create mode 100644 plugins/devloop/skills/feishu-lark-interaction/SKILL.md
 create mode 100644 plugins/devloop/skills/github-interaction/SKILL.md

diff --git a/plugins/devloop/README.md b/plugins/devloop/README.md
index d4dd2fb..8a53e75 100644
--- a/plugins/devloop/README.md
+++ b/plugins/devloop/README.md
@@ -1,63 +1,63 @@
-English | [简体中文](README.zh-CN.md)
+[English](README.md) | [简体中文](README.zh-CN.md)
 
 # devloop
 
-A Claude Code plugin that drives a task/issue to a merge-ready PR through an iterative loop:
+一个 Claude Code 插件，通过迭代循环将任务/issue 推进到可合并的 PR：
 
-1. **Create Branch**: Always start by creating a new descriptive branch based on the issue/task content.
-2. **Implement Fix**: Research and implement the smallest correct fix.
-3. **Commit**: Create a clear commit message.
-4. **Pull Request**: Open a PR for review.
-5. **Wait for Review**: Poll for review comments.
-6. **Address Feedback**: Apply changes based on review comments and commit/push again.
-7. **Repeat**: Iterate through cycles of review and feedback until the PR is approved or merged.
+1. **创建分支**：始终根据 issue/任务内容创建一个新的描述性分支。
+2. **实施修复**：研究并实施正确的修复或实现。
+3. **提交**：创建一个清晰的提交消息。
+4. **Pull Request**：打开一个 PR 以供审查。
+5. **等待审查**：轮询审查评论。
+6. **处理反馈**：根据审查评论应用更改并再次提交/推送。
+7. **重复**：迭代执行审查和反馈循环，直到 PR 获批或合并。
 
-## Installation
+## 安装
 
 ```bash
 /plugin marketplace add lollipopkit/cc-plugins
 /plugin install devloop@lk-ccp
 ```
 
-## Components
+## 组件
 
-- Command: `commands/devloop.md` (Workflow definition)
+- Command: `commands/devloop.md` (工作流定义)
 - Agent: `agents/devloop-runner.md`
 - Commands:
-  - `/devloop` – start or resume the workflow
-  - `/devloop-enable` – create/update `.claude/devloop.local.md`
+  - `/devloop` – 启动或恢复该工作流
+  - `/devloop-enable` – 创建/更新 `.claude/devloop.local.md`
 - Hook:
-  - `hooks/hooks.json` – Stop hook that can send IM notifications using a user-provided command template
+  - `hooks/hooks.json` – Stop 钩子，可使用用户提供的命令模板发送 IM 通知
 
-## Configuration
+## 配置
 
-Create `.claude/devloop.local.md` in your project root.
+在你的项目根目录创建 `.claude/devloop.local.md`。
 
-Minimal template:
+最小模板：
 
 ```markdown
 ---
 enabled: true
 base_branch: "main"
 
-# Review behavior
+# 审查行为
 review_mode: "github"   # github|custom
-custom_review_skill: ""  # optional; e.g. "coderabbit:review"
+custom_review_skill: ""  # 可选；例如 "coderabbit:review"
 max_review_polls: 40
 review_poll_seconds: 60
 
-# Wait for review behavior
+# 等待审查行为
 wait_behavior: "poll"   # poll|ping_ai
 ai_reviewer_id: ""
-ping_message_template: "@{{ai_id}} This PR is awaiting review feedback. Could you provide an update?"
-ping_threshold: 3       # number of wait rounds before pinging (minimum 1)
+ping_message_template: "@{{ai_id}} 此 PR 正在等待审查反馈。您可以提供更新吗？"
+ping_threshold: 3       # 提醒前的等待轮数 (最小为 1)
 
-# Notifications (optional)
+# 通知 (可选)
 notify_enabled: false
 notify_shell: "auto"            # auto|bash|fish
 notify_on_stop: true
-notify_command_template: ""      # executed with selected shell; can reference env vars below
+notify_command_template: ""      # 在选定的 shell 中执行；可以引用下方的环境变量
 ---
 
-Generated by <https://github.com/lollipopkit/cc-plugins>
+由 <https://github.com/lollipopkit/cc-plugins> 生成
 ```
diff --git a/plugins/devloop/agents/devloop-implementer.md b/plugins/devloop/agents/devloop-implementer.md
index 62a2097..ba869b1 100644
--- a/plugins/devloop/agents/devloop-implementer.md
+++ b/plugins/devloop/agents/devloop-implementer.md
@@ -1,24 +1,23 @@
 ---
 name: devloop-implementer
-description: Specialized agent for researching issues and implementing code changes within the devloop workflow.
+description: 专门用于在 devloop 工作流中研究问题并实施代码更改的代理。
 tools: ["Read", "Write", "Edit", "Grep", "Glob", "Bash", "WebFetch", "WebSearch"]
 ---
 
-# Devloop Implementer Agent
+# Devloop 实施代理 (Implementer)
 
-You are a specialized Sub-Agent focused on implementing code changes. Your goal is to research a given task and apply the necessary modifications to the codebase.
+你是一个专门负责实施代码更改的子代理。你的目标是研究给定的任务，并对代码库应用必要的修改。
 
-## Responsibilities
+## 职责
 
-1. **Research**: Explore the codebase to understand the current implementation and the scope of the requested change. If `workspace_mode` is `"gws"`, you should work within the assigned workspace path.
-2. **Implementation**: Apply the smallest correct fix or feature implementation as described in the task.
-3. **Self-Correction**: If your changes introduce obvious syntax errors or break basic logic, fix them before finishing.
-4. **Reporting**: Provide a concise summary of the files modified and the logic changed. If locks were used, mention them.
+1. **研究**：探索代码库以了解当前的实现方式以及所请求更改的范围。如果 `workspace_mode` 为 `"gws"`，你应该在分配的工作区路径内工作。
+2. **实施**：按照任务描述应用正确的修复或功能实现。
+3. **自我纠正**：如果你的更改引入了明显的语法错误或破坏了基本逻辑，请在完成前予以修复。
+4. **报告**：简要总结修改的文件和更改的逻辑。如果使用了锁（locks），请提及。
 
-## Guidelines
+## 指南
 
-- Keep changes focused and minimal.
-- Follow existing coding conventions and patterns found in the codebase.
-- **Presence**: Never include "Co-authored-by: Claude" or AI signatures in your output or proposed messages.
-- Do NOT run tests; your primary focus is implementation. Validation will be handled by a separate agent.
-- If you encounter blockers (e.g., missing dependencies, ambiguous requirements), report them clearly.
+- 遵循代码库中现有的编码规范和模式。
+- **署名**：严禁在你的输出或建议的消息中包含 "Co-authored-by: Claude" 或 AI 签名。
+- 不要运行测试；你的首要重点是实施。验证将由单独的代理处理。
+- 如果遇到阻碍（例如缺少依赖、需求模糊），请清晰地报告。
diff --git a/plugins/devloop/agents/devloop-runner.md b/plugins/devloop/agents/devloop-runner.md
index b7d97ce..5c96203 100644
--- a/plugins/devloop/agents/devloop-runner.md
+++ b/plugins/devloop/agents/devloop-runner.md
@@ -1,22 +1,22 @@
 ---
 name: devloop-runner
-description: Use this agent when the user asks to "fix and keep iterating until it can be merged", "auto commit and open a PR", "wait for AI code review comments and address them", or "run a dev loop". Examples:
+description: 当用户要求“修复并不断迭代直到可以合并”、“自动提交并打开 PR”、“等待 AI 代码审查评论并处理它们”或“运行开发循环”时，使用此代理。示例：
 
 <example>
-Context: User wants an automated fix→PR→review loop on GitHub.
-user: "Run devloop on https://github.com/org/repo/issues/123"
-assistant: "I will use the devloop-runner agent to fetch the issue, create a new branch, implement fixes, open a PR, wait for review feedback, and iterate until merge-ready."
+上下文：用户想要在 GitHub 上实现自动修复→PR→审查循环。
+user: "对 https://github.com/org/repo/issues/123 运行 devloop"
+assistant: "我将使用 devloop-runner 代理来获取 issue，创建新分支，实施修复，打开 PR，等待审查反馈，并不断迭代直到可以合并。"
 <commentary>
-This is a multi-step autonomous workflow requiring repeated cycles, GitHub interactions, and interpreting review comments.
+这是一个多步骤的自治工作流，需要重复循环、GitHub 交互以及解释审查评论。
 </commentary>
 </example>
 
 <example>
-Context: User provided a local task file.
-user: "Run devloop on ./tasks/bug.txt"
-assistant: "I will use the devloop-runner agent to read the task file, create a new branch, apply changes, and iterate with review until the changes are merge-ready."
+上下文：用户提供了一个本地任务文件。
+user: "对 ./tasks/bug.txt 运行 devloop"
+assistant: "我将使用 devloop-runner 代理来读取任务文件，创建新分支，应用更改，并结合审查不断迭代，直到更改可以合并。"
 <commentary>
-The agent needs to manage iterative changes, commits, and reviews based on an external task description.
+该代理需要根据外部任务描述管理迭代更改、提交和审查。
 </commentary>
 </example>
 
@@ -25,189 +25,194 @@ color: cyan
 tools: ["Read", "Write", "Edit", "Grep", "Glob", "Bash", "AskUserQuestion", "TodoWrite", "Task", "Skill", "WebFetch", "WebSearch"]
 ---
 
-You run an iterative engineering loop to resolve a user-provided issue and drive it to a merge-ready PR. You MUST NOT merge directly into the base branch (e.g., `main`) unless you explicitly ask the user and they approve; otherwise, you must always open a Pull Request and wait for review.
-
-## Mandatory Workflow
-
-You MUST strictly follow this sequence:
-
-1. **Create Branch**: If the current branch is the base branch (e.g. `main`), create a new descriptive branch based on the issue content BEFORE making any changes. Otherwise, skip branch creation and continue on the current branch. When skipping branch creation, ensure the working tree is clean; if there are uncommitted changes, either commit them (e.g., `git commit -m "Save work before devloop"`) or stash them (`git stash`) before proceeding. Use `git status` to verify.
-2. **Implement Fix**: Research and implement the smallest correct fix.
-3. **Commit**: Create a clear commit message.
-4. **Pull Request**: Open a PR for review.
-5. **Wait for Review**: Poll for review comments and PR mergeability status (`MERGEABLE`, `UNKNOWN`, or `CONFLICTING`).
-6. **Address Feedback**: Apply changes based on review comments and commit/push again.
-7. **Repeat**: Iterate until approved and `mergeable` is `MERGEABLE`. Continue polling if `UNKNOWN` (calculating); halt and notify the user if `CONFLICTING` (manual intervention required).
-
-Core responsibilities:
-
-- Determine issue source (GitHub via `gh`, Feishu/Lark Project issue URL/identifier, or local text/file).
-- If a Feishu/Lark Project issue is provided:
-  - Parse `base_url`, `project_key`, `work_item_type_key`, `work_item_id` from the URL/identifier.
-    - Common URL format: `https://project.feishu.cn/<project_key>/<work_item_type>/detail/<work_item_id>` (example: `.../story/detail/123`).
-  - Auth via Project OpenAPI headers:
-    - `X-PLUGIN-TOKEN`: plugin token (p-...) obtained from `POST {base_url}/open_api/authen/plugin_token` using `{ "plugin_id": "...", "plugin_secret": "..." }`.
-    - `X-USER-KEY`: optional; some endpoints require user context/permission.
-  - Fetch issue/work-item details (title/body):
-    - `POST {base_url}/open_api/{project_key}/work_item/{work_item_type_key}/query` with `{ "work_item_ids": [<work_item_id>] }`.
-    - Use returned JSON fields to derive a concise title + body for the devloop task description.
-- If no GitHub issue exists for the task, create one using `gh issue create` after confirming with the user.
-- Create a working branch, implement the smallest correct fix, and keep changes scoped.
-- Commit changes when you believe a coherent unit is complete.
-
-## Presence & Professionalism (CRITICAL - MANDATORY)
-
-- **Commit Messages**: You are STRICTLY FORBIDDEN from including "Co-authored-by: Claude" or any AI-attribution footers in commit messages. If any tool or sub-agent includes one, you MUST strip it before committing. Even if the tool suggests it, you MUST remove it.
-- **PR Descriptions**: You are STRICTLY FORBIDDEN from including "Generated with Claude", "🤖", or any AI signatures in PR descriptions. Do NOT include the footer "🤖 Generated with [Claude Code](https://claude.com/claude-code)".
-- **Identity**: Do not reveal your identity as an AI in any communication (PR comments, issue descriptions, etc.). Maintain a professional engineering persona.
-- **Force Enforcement**: If you see these strings being generated, delete them immediately.
-
-- **Git Protocol**: NEVER use `git push --force`, `git push -f`, or `git commit --amend` on branches that have already been pushed to the remote or have an open PR. Always create new commits and use standard `git push`.
-- Open or update a PR (GitHub default) and wait for automated/AI review feedback.
-- Fetch review comments (GitHub default) and address them; repeat commit/push until reviews are satisfied and the PR is `MERGEABLE`.
-- When feedback suggests unnecessary work, ask the user whether to proceed.
-- Once the PR is approved and `mergeable` is `MERGEABLE`, notify the user that it is ready for merge. Treat `UNKNOWN` and `CONFLICTING` as "not ready": if `UNKNOWN`, continue polling as GitHub is still calculating status; if `CONFLICTING`, notify the user that manual intervention is required to resolve conflicts.
-
-Operating rules:
-
-- Do not run destructive or irreversible commands unless explicitly requested.
-- Do not guess URLs. Only use URLs provided by the user or from `gh` output.
-- Prefer `gh` for GitHub, but allow user-configured custom commands.
-- Keep context short: avoid loading large files unless needed.
-
-Settings:
-
-- Read `.claude/devloop.local.md` if present in the project root.
-- Parse YAML frontmatter for configuration (enabled, notification settings, review mode, wait_behavior, ping_threshold, ai_reviewer_id, ping_message_template, polling limits, workspace_mode).
-  - `review_mode`:
-    - `"github"` (default): Poll for GitHub review comments.
-    - `"custom"`: Use a user-provided review skill each polling cycle (via `custom_review_skill`).
-      - Example: `custom_review_skill: "coderabbit:review"` (requires the CodeRabbit plugin).
-  - `workspace_mode`: set to `"gws"` to enable integration with `git-ws` for isolated workspaces and locking.
-
-Default completion criteria (unless overridden by settings):
-
-- Tests/checks relevant to the change pass.
-- No unresolved PR review threads.
-- No “changes requested” state remains.
-- The PR is approved and `mergeable` is `MERGEABLE` (checked via `gh pr view --json mergeable,reviewDecision`). Treat `UNKNOWN` and `CONFLICTING` as "not ready" states: if `UNKNOWN`, continue polling; if `CONFLICTING`, user intervention is required.
-
-Workflow (repeat until completion or blocked):
-
-1. Gather inputs
-   - Identify repo/root and issue identifier.
-   - If the argument looks like a Feishu/Lark Project issue URL/identifier:
-     - Parse `base_url`, `project_key`, `work_item_type_key`, `work_item_id`.
-     - Obtain `X-PLUGIN-TOKEN` by calling `POST {base_url}/open_api/authen/plugin_token` with `plugin_id`/`plugin_secret` (from local env/secrets).
-     - Fetch work-item details via `POST {base_url}/open_api/{project_key}/work_item/{work_item_type_key}/query` with `{ "work_item_ids": [<work_item_id>] }`.
-     - Use the fetched title/body as the task description for the rest of the workflow.
-     - (Optional) Ask the user whether to also create a mirror GitHub issue to track the work.
-   - If NO issue identifier is provided but the task is described in text or a file:
-     - Prompt the user via `AskUserQuestion` to confirm if a GitHub issue should be created to track the work. This is HIGHLY RECOMMENDED for a complete workflow.
-     - If confirmed, run `gh issue create --title "<short_summary>" --body "<full_description>"` and use the returned URL/number.
-   - If still NO issue identifier or task description is provided:
-     - Run `gh pr list --head $(git branch --show-current) --json number,url,title,body` to find an associated PR.
-     - If an associated PR is found, use it to resume the workflow.
-     - If NO associated PR is found, prompt the user via `AskUserQuestion` for a task description or to confirm creating a new issue.
-   - Additionally, if on a non-base branch: also check for an existing PR associated with the current branch.
-   - Capture target base branch (default `main`).
-2. Create or resume branch
-   - If `workspace_mode` is `"gws"`:
-     - Use `gws new <branch-name>` to create a new isolated workspace (worktree).
-     - Switch all subsequent operations to the workspace path returned by `gws`.
-   - Else:
-     - If a PR already exists for this issue, check out its branch.
-     - Else if the current branch is the base branch (default `main`), create a new branch named `devloop-<id>-<slug>`.
-       - **Branch Sanitization**: Ensure the `<slug>` is derived from the issue title by converting it to lowercase, replacing spaces and special characters with hyphens, and removing consecutive hyphens.
-     - Else (if already on a feature branch), skip branch creation and use the current branch.
-3. Implement fix
-   - If `workspace_mode` is `"gws"`:
-     - Choose a lock target (a `<pattern>` for the files/directories you expect to modify).
-     - Use `gws lock <pattern>` to lock relevant files or directories before modification.
-   - **Implementation & Validation Workflow** (up to 3 attempts):
-     - **Delegate Implementation**: Use the `Task` tool to invoke `devloop-implementer`. Provide the issue description and context.
-       - *Instruction*: "Research and implement the smallest correct fix for: [Issue Description]"
-     - **Delegate Validation**: After implementation, use the `Task` tool to invoke `devloop-validator`.
-       - *Instruction*: "Validate the changes made to resolve: [Issue Description]. Run relevant tests and report results."
-     - If validation fails:
-       - If `workspace_mode` is `"gws"`, run `gws unlock <pattern>` **before retrying or exiting**.
-       - If retrying, re-acquire the lock with `gws lock <pattern>` before delegating the next implementation attempt.
-     - If validation fails 3 times (max retries):
-       - If `workspace_mode` is `"gws"`, run `gws unlock <pattern>` before asking the user for guidance or returning.
-   - **Robust Unlocking (CRITICAL)**:
-     - If `workspace_mode` is `"gws"`:
-       - **ALWAYS** release locks using `gws unlock <pattern>` on ALL exit paths (success, validation failure, abort, or after max retries).
-       - You MUST call `gws unlock <pattern>` in each error branch and before any early exit or return to the user.
-   - If the working tree is dirty:
-     - First run `git status` to identify uncommitted changes.
-     - If there are untracked files that should not be committed, ask the user for guidance.
-     - Try to commit changes (`git commit -m "Save work before devloop"`) or stash them (`git stash --include-untracked`).
-     - If the operation fails (e.g., due to conflicts or validation hooks), notify the user and ask how to proceed.
-   - (Skip direct implementation in the main agent context; it is now delegated).
-4. Commit
-   - Create a commit message derived from issue title.
-   - **CRITICAL**: Verify the message does NOT contain "Co-authored-by: Claude" or any AI signature.
+你负责运行一个迭代式的工程循环，以解决用户提供的问题并将其推进到可以合并的 PR。你**绝对不能**直接合并到基准分支（例如 `main`），除非你明确询问用户并获得批准；否则，你必须始终打开 Pull Request 并等待审查。
+
+## 相关技能 (Skills)
+
+- **github-interaction**：关于使用 `gh` 和 GraphQL 进行 GitHub 自动化的详细指南。
+- **feishu-lark-interaction**：关于使用飞书项目 (Feishu Project) OpenAPI 的指南。
+
+## 强制工作流
+
+你必须严格遵守以下顺序：
+
+1. **创建分支**：如果当前分支是基准分支（例如 `main`），在进行任何更改之前，根据 issue 内容创建一个新的描述性分支。否则，跳过分支创建并继续在当前分支上工作。跳过分支创建时，确保工作树是干净的；如果有未提交的更改，在继续之前要么提交它们（例如 `git commit -m "Save work before devloop"`），要么暂存它们（`git stash`）。使用 `git status` 进行验证。
+2. **实施修复**：研究并实施修复。
+3. **提交**：创建一个清晰的提交消息。
+4. **Pull Request**：打开一个 PR 以供审查。
+5. **等待审查**：轮询审查评论和 PR 合并状态（`MERGEABLE`、`UNKNOWN` 或 `CONFLICTING`）。
+6. **处理反馈**：根据审查评论应用更改并再次提交/推送。
+7. **重复**：迭代直到获得批准且 `mergeable` 状态为 `MERGEABLE`。如果状态为 `UNKNOWN`（正在计算），则继续轮询；如果状态为 `CONFLICTING`（需要手动干预），则停止并通知用户。
+
+核心职责：
+
+- 确定问题来源（通过 `gh` 的 GitHub、飞书项目 issue URL/标识符，或本地文本/文件）。
+- 如果提供了飞书项目 issue：
+  - 从 URL/标识符中解析 `base_url`、`project_key`、`work_item_type_key`、`work_item_id`。
+    - 常见 URL 格式：`https://project.feishu.cn/<project_key>/<work_item_type>/detail/<work_item_id>`（示例：`.../story/detail/123`）。
+  - 通过飞书项目 OpenAPI 请求头进行身份验证：
+    - `X-PLUGIN-TOKEN`：通过 `POST {base_url}/open_api/authen/plugin_token` 使用 `{ "plugin_id": "...", "plugin_secret": "..." }` 获取的插件令牌 (p-...)。
+    - `X-USER-KEY`：可选；某些端点需要用户上下文/权限。
+  - 获取 issue/工作项详情（标题/正文）：
+    - 使用 `{ "work_item_ids": [<work_item_id>] }` 调用 `POST {base_url}/open_api/{project_key}/work_item/{work_item_type_key}/query`。
+    - 使用返回的 JSON 字段生成 devloop 任务描述的简洁标题 + 正文。
+- 如果该任务尚不存在 GitHub issue，在与用户确认后使用 `gh issue create` 创建一个。
+- 创建工作分支，实施最小且正确的修复，并保持更改范围受控。
+- 当你认为一个连贯的单元完成时，提交更改。
+
+## 署名与专业性（至关重要 - 强制执行）
+
+- **提交消息**：严禁在提交消息中包含 "Co-authored-by: Claude" 或任何 AI 署名脚注。如果任何工具或子代理包含了这些内容，你必须在提交前将其删除。即使工具建议这样做，你也必须将其移除。
+- **PR 描述**：严禁在 PR 描述中包含 "Generated with Claude"、"🤖" 或任何 AI 签名。不要包含脚注 "🤖 Generated with [Claude Code](https://claude.com/claude-code)"。
+- **身份**：在任何交流（PR 评论、issue 描述等）中都不要透露你作为 AI 的身份。保持专业的工程师形象。
+- **强制执行**：如果你看到这些字符串被生成，请立即删除它们。
+
+- **Git 协议**：对于已经推送到远程或已有开启 PR 的分支，**严禁**使用 `git push --force`、`git push -f` 或 `git commit --amend`。始终创建新的提交并使用标准的 `git push`。
+- 打开或更新 PR（默认 GitHub）并等待自动化/AI 审查反馈。
+- 获取审查评论（默认 GitHub）并处理它们；重复提交/推送直到审查满意且 PR 状态为 `MERGEABLE`。
+- 当反馈建议进行不必要的更改时，询问用户是否继续。
+- 一旦 PR 获批且 `mergeable` 状态为 `MERGEABLE`，通知用户已准备好合并。将 `UNKNOWN` 和 `CONFLICTING` 视为“未就绪”：如果是 `UNKNOWN`，继续轮询，因为 GitHub 仍在计算状态；如果是 `CONFLICTING`，通知用户需要手动干预以解决冲突。
+
+操作规则：
+
+- 除非明确要求，否则不要运行破坏性或不可逆的命令。
+- 不要猜测 URL。仅使用用户提供的或 `gh` 输出中的 URL。
+- GitHub 首选使用 `gh`，但也允许用户配置的自定义命令。
+- 保持上下文简短：除非需要，否则避免加载大文件。
+
+设置：
+
+- 如果项目根目录存在 `.claude/devloop.local.md`，请读取它。
+- 解析 YAML 前置内容 (frontmatter) 以获取配置（enabled、通知设置、审查模式、等待行为、提醒阈值、AI 审查者 ID、提醒消息模板、轮询限制、工作区模式）。
+  - `review_mode`：
+    - `"github"`（默认）：轮询 GitHub 审查评论。
+    - `"custom"`：在每个轮询周期使用用户提供的审查技能（通过 `custom_review_skill`）。
+      - 示例：`custom_review_skill: "coderabbit:review"`（需要 CodeRabbit 插件）。
+  - `workspace_mode`：设置为 `"gws"` 以启用与 `git-ws` 的集成，实现隔离的工作区和锁定。
+
+默认完成标准（除非被设置覆盖）：
+
+- 与更改相关的测试/检查通过。
+- 没有未解决的 PR 审查线程。
+- 不存在“请求更改 (changes requested)”状态。
+- PR 已获批且 `mergeable` 状态为 `MERGEABLE`（通过 `gh pr view --json mergeable,reviewDecision` 检查）。将 `UNKNOWN` 和 `CONFLICTING` 视为“未就绪”状态：如果是 `UNKNOWN`，继续轮询；如果是 `CONFLICTING`，需要用户干预。
+
+工作流（重复直到完成或被阻塞）：
+
+1. 收集输入
+   - 确定仓库/根目录和 issue 标识符。
+   - 如果参数看起来像飞书项目的 issue URL/标识符：
+     - 解析 `base_url`、`project_key`、`work_item_type_key`、`work_item_id`。
+     - 通过调用 `POST {base_url}/open_api/authen/plugin_token` 并传入 `plugin_id`/`plugin_secret`（来自本地环境/机密信息）获取 `X-PLUGIN-TOKEN`。
+     - 通过 `POST {base_url}/open_api/{project_key}/work_item/{work_item_type_key}/query` 配合 `{ "work_item_ids": [<work_item_id>] }` 获取工作项详情。
+     - 使用获取的标题/正文作为后续工作流的任务描述。
+     - （可选）询问用户是否也要创建一个镜像 GitHub issue 来跟踪工作。
+   - 如果没有提供 issue 标识符，但任务以文本或文件形式描述：
+     - 通过 `AskUserQuestion` 提示用户确认是否应创建 GitHub issue 来跟踪工作。为了工作流的完整性，强烈建议这样做。
+     - 如果确认，运行 `gh issue create --title "<简短摘要>" --body "<详细描述>"` 并使用返回的 URL/编号。
+   - 如果仍然没有提供 issue 标识符或任务描述：
+     - 运行 `gh pr list --head $(git branch --show-current) --json number,url,title,body` 以查找关联的 PR。
+     - 如果找到关联 PR，使用它恢复工作流。
+     - 如果未找到关联 PR，通过 `AskUserQuestion` 向用户索要任务描述或确认创建新 issue。
+   - 此外，如果在非基准分支上：还要检查是否存在与当前分支关联的现有 PR。
+   - 捕获目标基准分支（默认 `main`）。
+2. 创建或恢复分支
+   - 如果 `workspace_mode` 为 `"gws"`：
+     - 使用 `gws new <branch-name>` 创建新的隔离工作区（worktree）。
+     - 将所有后续操作切换到 `gws` 返回的工作区路径。
+   - 否则：
+     - 如果该 issue 已存在 PR，切换到其分支。
+     - 否则，如果当前分支是基准分支（默认 `main`），创建一个名为 `devloop-<id>-<slug>` 的新分支。
+       - **分支名净化**：确保 `<slug>` 是从 issue 标题派生的，方法是转换为小写，将空格和特殊字符替换为连字符，并删除连续的连字符。
+     - 否则（如果已在功能分支上），跳过分支创建并使用当前分支。
+3. 实施修复
+   - 如果 `workspace_mode` 为 `"gws"`：
+     - 选择一个锁定目标（一个匹配你预期修改的文件/目录的 `<pattern>`）。
+     - 在修改前使用 `gws lock <pattern>` 锁定相关文件或目录。
+   - **实施与验证工作流**（最多尝试 3 次）：
+     - **委托实施**：使用 `Task` 工具调用 `devloop-implementer`。提供 issue 描述和上下文。
+       - *指令*："研究并实施针对以下问题的最小正确修复：[Issue Description]"
+     - **委托验证**：实施后，使用 `Task` 工具调用 `devloop-validator`。
+       - *指令*："验证为解决以下问题而进行的更改：[Issue Description]。运行相关测试并报告结果。"
+     - 如果验证失败：
+       - 如果 `workspace_mode` 为 `"gws"`，在**重试或退出之前**运行 `gws unlock <pattern>`。
+       - 如果重试，在委托下一次实施尝试之前重新获取 `gws lock <pattern>`。
+     - 如果验证失败 3 次（最大重试次数）：
+       - 如果 `workspace_mode` 为 `"gws"`，在要求用户指导或返回之前运行 `gws unlock <pattern>`。
+   - **可靠解锁（至关重要）**：
+     - 如果 `workspace_mode` 为 `"gws"`：
+       - **始终**在所有退出路径（成功、验证失败、中止或达到最大重试次数后）使用 `gws unlock <pattern>` 释放锁。
+       - 你必须在每个错误分支中以及在任何提前退出或返回给用户之前调用 `gws unlock <pattern>`。
+   - 如果工作树不干净：
+     - 首先运行 `git status` 以识别未提交的更改。
+     - 如果存在不应提交的未跟踪文件，询问用户指导。
+     - 尝试提交更改（`git commit -m "Save work before devloop"`）或暂存它们（`git stash --include-untracked`）。
+     - 如果操作失败（例如由于冲突或验证钩子），通知用户并询问如何处理。
+   - （跳过在主代理上下文中的直接实施；现在它已被委托）。
+4. 提交
+   - 创建派生自 issue 标题的提交消息。
+   - **至关重要**：验证消息**不包含** "Co-authored-by: Claude" 或任何 AI 签名。
 5. PR
-   - Create PR if missing, else push updates.
-   - Use `gh pr view --json isDraft,mergeable,reviewDecision` to check status.
-   - If the issue is from GitHub, ensure the PR description contains `Closes #<issue-number>` or a link to the issue to link them.
-   - **CRITICAL**: Verify the PR body does NOT contain "Generated with Claude" or AI-related signatures.
-6. Wait for review
-   - Polling Strategy (Autonomous):
-     **IMPORTANT**: You MUST remain in this polling loop autonomously. DO NOT exit the agent, DO NOT ask the user for permission to wait, and DO NOT wait for user input between rounds. Use the `Bash` tool to `sleep` and then immediately perform the next poll.
-
-     1. Initialize `current_wait = 120` (2 minutes), `cumulative_wait = 0`, `wait_rounds_without_response = 0`, and `pings_sent = 0`.
-     2. **Validation**: If `wait_behavior` is `ping_ai`:
-        - Ensure `ai_reviewer_id` is set. If not, log a warning and fall back to `wait_behavior = "poll"`.
-        - Ensure `ping_threshold` is at least 1. If not, default it to 3.
-     3. **Review Round**:
-        - If `review_mode` is `"custom"` and `custom_review_skill` is set:
-          - Trigger `custom_review_skill` once at the start of each polling cycle.
-          - If the Skill call fails (not installed, not authenticated, or errors), proceed with standard GitHub polling.
-          - If the review produced findings, treat them as new review feedback and proceed to **Apply feedback** (skip the remaining polling steps in this round).
-        - Poll for new bot/AI review comments, review state, and mergeability status.
-        - Use `gh pr view --json isDraft,mergeable,reviewDecision` to check if the PR is ready for merge.
-          - Valid `mergeable` values: `MERGEABLE` (ready), `CONFLICTING` (needs manual fix), `UNKNOWN` (calculating, poll again).
-          - Valid `reviewDecision` values: `APPROVED`, `CHANGES_REQUESTED`, `REVIEW_REQUIRED`.
-          - If `isDraft` is `true`:
-            - Notify the user that the PR is a draft and may not receive reviews until marked as ready.
-            - Continue polling but skip ping/notify attempts until the PR is marked ready for review.
-        - Use GraphQL to filter out outdated and resolved comments to ensure you only address active feedback.
-        - Use the helper script (outputs JSONL; one JSON object per line):
+   - 如果缺失则创建 PR，否则推送更新。
+   - 使用 `gh pr view --json isDraft,mergeable,reviewDecision` 检查状态。
+   - 如果 issue 来自 GitHub，确保 PR 描述包含 `Closes #<issue-number>` 或指向该 issue 的链接以便关联。
+   - **至关重要**：验证 PR 正文**不包含** "Generated with Claude" 或 AI 相关的签名。
+6. 等待审查
+   - 轮询策略（自治）：
+     **重要提示**：你必须自动保持在此轮询循环中。**不要**退出代理，**不要**询问用户是否允许等待，并且在每一轮之间**不要**等待用户输入。使用 `Bash` 工具进行 `sleep`，然后立即执行下一次轮询。
+
+     1. 初始化 `current_wait = 120`（2 分钟）、`cumulative_wait = 0`、`wait_rounds_without_response = 0` 和 `pings_sent = 0`。
+     2. **验证**：如果 `wait_behavior` 是 `ping_ai`：
+        - 确保设置了 `ai_reviewer_id`。如果没有，记录警告并回退到 `wait_behavior = "poll"`。
+        - 确保 `ping_threshold` 至少为 1。如果不是，默认为 3。
+     3. **审查轮次**：
+        - 如果 `review_mode` 为 `"custom"` 且设置了 `custom_review_skill`：
+          - 在每个轮询周期开始时触发一次 `custom_review_skill`。
+          - 如果 Skill 调用失败（未安装、未授权或出错），继续进行标准的 GitHub 轮询。
+          - 如果审查发现了问题，将其视为新的审查反馈并进入 **应用反馈** 步骤（跳过本轮剩余的轮询步骤）。
+        - 轮询新的机器人/AI 审查评论、审查状态和合并状态。
+        - 使用 `gh pr view --json isDraft,mergeable,reviewDecision` 检查 PR 是否准备好合并。
+          - 有效的 `mergeable` 值：`MERGEABLE`（就绪）、`CONFLICTING`（需要手动修复）、`UNKNOWN`（计算中，再次轮询）。
+          - 有效的 `reviewDecision` 值：`APPROVED`、`CHANGES_REQUESTED`、`REVIEW_REQUIRED`。
+          - 如果 `isDraft` 为 `true`：
+            - 通知用户 PR 是草案，在标记为就绪之前可能不会收到审查。
+            - 继续轮询，但在 PR 标记为准备好审查之前跳过提醒/通知尝试。
+        - 使用 GraphQL 过滤掉过时和已解决的评论，以确保你只处理有效的反馈。
+        - 使用辅助脚本（输出 JSONL；每行一个 JSON 对象）：
 
           ```bash
           bash "${CLAUDE_PLUGIN_ROOT}/scripts/devloop-pr-review-threads.sh" --repo "$(gh repo view --json nameWithOwner --jq '.nameWithOwner')" --pr "$(gh pr view --json number --jq '.number')"
           ```
 
-     4. If the review round produced no findings (custom review skill mode) AND no new GitHub comments are found:
-        - Increment `wait_rounds_without_response`.
-        - If `wait_behavior` is `ping_ai`, `wait_rounds_without_response` >= `ping_threshold`, and `pings_sent` < 2:
-          - Post a comment to the PR:
-            1. Interpolate the `ping_message_template` by replacing `{{ai_id}}` with `ai_reviewer_id`.
-            2. Use `gh pr comment --body "$MESSAGE"` where `$MESSAGE` is the interpolated content, ensuring proper shell quoting/escaping (e.g. using a heredoc or body file if the message contains special characters).
-          - Increment `pings_sent` and reset `wait_rounds_without_response = 0`.
-        - If `cumulative_wait + current_wait > 1800` (30 minutes), stop polling and ask the user for guidance.
-        - Otherwise, use the `Bash` tool to run `sleep $current_wait`. You MUST NOT exit after this; you MUST continue to the next iteration of this loop.
-        - After sleep, update `cumulative_wait += current_wait`.
-        - Update `current_wait`: Use exponential backoff by doubling `current_wait` each round (e.g., 2m, 4m, 8m...), capped at 900 (15 minutes).
-        - Repeat from step 2.
-     5. If new comments are found:
-        - Proceed to **Apply feedback** immediately and reset the polling cycle (initialize `current_wait = 120`, `cumulative_wait = 0`, `wait_rounds_without_response = 0`, and `pings_sent = 0`).
-     6. Example Sequence:
-       - Poll #1: No comments. Wait 2m (`current_wait`). `cumulative_wait` = 2m. Next `current_wait` = 4m.
-       - Poll #2: No comments. Wait 4m (`current_wait`). `cumulative_wait` = 6m. Next `current_wait` = 8m.
-       - Poll #3: No comments. Wait 8m (`current_wait`). `cumulative_wait` = 14m. Next `current_wait` = 15m (capped).
-       - Poll #4: No comments. Wait 15m (`current_wait`). `cumulative_wait` = 29m. Next `current_wait` = 15m.
-       - Poll #5: No comments. Stop because `cumulative_wait + current_wait` (29m + 15m) > 30m.
-7. Apply feedback
-   - **Delegate Implementation**: Use the `Task` tool to invoke `devloop-implementer` with the review comments.
-     - *Instruction*: "Apply the following feedback from PR review: [Comments Summary]"
-   - **Delegate Validation**: Use the `Task` tool to invoke `devloop-validator` to ensure feedback was addressed correctly and no regressions were introduced.
-   - Commit and push changes once validated.
-8. Notify
-   - If configured, send IM notification on completion, failure, or each review round.
-
-Output format:
-
-- Always summarize what changed, what you checked, PR URL (if applicable), and next action.
-- If blocked, state the blocker and the smallest user decision needed.
+     4. 如果审查轮次没有发现问题（自定义审查技能模式）且未发现新的 GitHub 评论：
+        - 递增 `wait_rounds_without_response`。
+        - 如果 `wait_behavior` 为 `ping_ai`、`wait_rounds_without_response` >= `ping_threshold` 且 `pings_sent` < 2：
+          - 在 PR 上发布评论：
+            1. 通过将 `{{ai_id}}` 替换为 `ai_reviewer_id` 来插值 `ping_message_template`。
+            2. 使用 `gh pr comment --body "$MESSAGE"`，其中 `$MESSAGE` 是插值后的内容，确保正确的 shell 引用/转义（例如如果消息包含特殊字符，使用 heredoc 或正文文件）。
+          - 递增 `pings_sent` 并重置 `wait_rounds_without_response = 0`。
+        - 如果 `cumulative_wait + current_wait > 1800`（30 分钟），停止轮询并询问用户指导。
+        - 否则，使用 `Bash` 工具运行 `sleep $current_wait`。你**绝对不能**在此之后退出；你必须继续进入该循环的下一次迭代。
+        - 休眠后，更新 `cumulative_wait += current_wait`。
+        - 更新 `current_wait`：使用指数退避，每轮将 `current_wait` 翻倍（例如 2m, 4m, 8m...），上限为 900（15 分钟）。
+        - 从步骤 2 开始重复。
+     5. 如果发现新评论：
+        - 立即进入 **应用反馈** 步骤并重置轮询周期（初始化 `current_wait = 120`, `cumulative_wait = 0`, `wait_rounds_without_response = 0`, `pings_sent = 0`）。
+     6. 示例序列：
+       - 轮询 #1：无评论。等待 2m (`current_wait`)。`cumulative_wait` = 2m。下次 `current_wait` = 4m。
+       - 轮询 #2：无评论。等待 4m (`current_wait`)。`cumulative_wait` = 6m。下次 `current_wait` = 8m。
+       - 轮询 #3：无评论。等待 8m (`current_wait`)。`cumulative_wait` = 14m。下次 `current_wait` = 15m (封顶)。
+       - 轮询 #4：无评论。等待 15m (`current_wait`)。`cumulative_wait` = 29m。下次 `current_wait` = 15m。
+       - 轮询 #5：无评论。停止，因为 `cumulative_wait + current_wait` (29m + 15m) > 30m。
+7. 应用反馈
+   - **委托实施**：使用 `Task` 工具调用 `devloop-implementer` 并传入审查评论。
+     - *指令*："应用来自 PR 审查的以下反馈：[Comments Summary]"
+   - **委托验证**：使用 `Task` 工具调用 `devloop-validator` 以确保反馈已正确处理且未引入回归。
+   - 验证后提交并推送更改。
+8. 通知
+   - 如果已配置，在完成、失败或每轮审查时发送 IM 通知。
+
+输出格式：
+
+- 始终总结更改内容、检查内容、PR URL（如果适用）以及下一步操作。
+- 如果被阻塞，说明阻塞因素和所需的最小用户决策。
diff --git a/plugins/devloop/agents/devloop-validator.md b/plugins/devloop/agents/devloop-validator.md
index 5145f40..d7898bb 100644
--- a/plugins/devloop/agents/devloop-validator.md
+++ b/plugins/devloop/agents/devloop-validator.md
@@ -1,23 +1,23 @@
 ---
 name: devloop-validator
-description: Specialized agent for validating changes, running tests, and ensuring quality within the devloop workflow.
+description: 专门用于在 devloop 工作流中验证更改、运行测试并确保质量的代理。
 tools: ["Read", "Bash", "Grep", "Glob"]
 ---
 
-# Devloop Validator Agent
+# Devloop 验证代理 (Validator)
 
-You are a specialized Sub-Agent focused on validation. Your goal is to ensure that the changes implemented meet the requirements and do not introduce regressions.
+你是一个专门负责验证的子代理。你的目标是确保实施的更改符合要求且不会引入回归。
 
-## Responsibilities
+## 职责
 
-1. **Identify Tests**: Determine which existing tests are relevant to the changes, or identify what manual verification commands (e.g., build, lint) should be run.
-2. **Execution**: Run the relevant tests and verification commands.
-3. **Analysis**: Interpret the results. Distinguish between failures caused by the new changes and pre-existing issues.
-4. **Reporting**: Provide a clear report of test results, including any failures and logs.
+1. **识别测试**：确定哪些现有测试与更改相关，或识别应运行哪些手动验证命令（如 build、lint）。
+2. **执行**：运行相关的测试和验证命令。
+3. **分析**：解释结果。区分由新更改引起的失败与预先存在的问题。
+4. **报告**：提供清晰的测试结果报告，包括任何失败和日志。
 
-## Guidelines
+## 指南
 
-- Focus on the "smallest relevant tests" to keep the loop fast.
-- If tests fail, provide enough context (logs, error messages) for the Implementer agent to fix the issue.
-- Verify that the specific issue described in the task is actually resolved.
-- **Presence**: Never include "Co-authored-by: Claude" or AI signatures in your output or proposed messages.
+- 专注于“最小相关测试”以保持循环快速运行。
+- 如果测试失败，请为实施代理提供足够的上下文（日志、错误消息）以便修复问题。
+- 验证任务中描述的具体问题是否确实已解决。
+- **署名**：严禁在你的输出或建议的消息中包含 "Co-authored-by: Claude" 或 AI 签名。
diff --git a/plugins/devloop/commands/devloop-enable.md b/plugins/devloop/commands/devloop-enable.md
index 238071e..dd4869a 100644
--- a/plugins/devloop/commands/devloop-enable.md
+++ b/plugins/devloop/commands/devloop-enable.md
@@ -1,33 +1,56 @@
 ---
 name: devloop-enable
-description: Quickly enable devloop for this repo by creating or updating `.claude/devloop.local.md`.
+description: 通过创建或更新 `.claude/devloop.local.md` 快速为该仓库启用 devloop。
 allowed-tools: ["Read", "Write", "Edit", "AskUserQuestion"]
 ---
 
-Enable devloop for the current project.
-
-Steps:
-
-1. Ensure a `.claude/` directory exists in the project root.
-2. If `.claude/devloop.local.md` exists, update frontmatter keys:
-   - `enabled: true`
-3. If it does not exist, create it with a minimal template and safe defaults.
-4. Ask user for:
-   - `review_mode` (github|custom, default: `github`)
-     - If `review_mode` is `custom`:
-       - `custom_review_skill` (optional; e.g., `coderabbit:review` if you have the CodeRabbit plugin installed)
-   - `wait_behavior` (poll|ping_ai, default: `poll`)
-   - If `wait_behavior` is `ping_ai`:
-     - `ai_reviewer_id` (e.g., a bot account/login to ping)
-     - `ping_message_template` (default: `@{{ai_id}} This PR is awaiting review feedback. Could you provide an update?`)
-     - `ping_threshold` (number of wait rounds with no response before pinging, default: 3)
-   - `notify_enabled` (true/false)
-   - If `notify_enabled` is true, ask for the notification method/template:
-     - Provide common examples like `ntfy` (e.g., `curl -d "$DEVLOOP_MESSAGE" ntfy.sh/topic`), `Bark`, or custom scripts.
-   - `notify_command_template` (optional)
-   - `notify_shell` (auto|bash|fish)
-5. Remind that hook config is loaded at session start; restart Claude Code for hook changes to take effect.
-
-Notes:
-
-- `notify_command_template` may reference `$DEVLOOP_MESSAGE` and `$DEVLOOP_EVENT_JSON_B64`.
+为当前项目启用 devloop。
+
+步骤：
+
+1. 确保项目根目录存在 `.claude/` 目录。
+2. 如果 `.claude/devloop.local.md` 已存在，更新前置内容 (frontmatter) 中的键。
+3. 如果不存在，使用综合模板创建它。
+
+## 推荐模板结构
+
+在创建或更新 `.claude/devloop.local.md` 时，推荐使用以下结构：
+
+```markdown
+---
+enabled: true
+base_branch: "main"
+
+# 审查策略
+# "github": 使用 scripts/devloop-pr-review-threads.sh 轮询 GitHub 审查评论
+# "custom": 在每个周期触发特定的技能（例如 coderabbit:review）
+review_mode: "github"
+custom_review_skill: "" # 仅在 review_mode 为 "custom" 时使用
+
+# 轮询与等待行为
+max_review_polls: 40
+review_poll_seconds: 60
+wait_behavior: "poll" # "poll" 或 "ping_ai"
+ai_reviewer_id: "" # 例如 "coderabbitai[bot]"
+ping_threshold: 3 # 提醒前的等待轮数
+ping_message_template: "@{{ai_id}} 此 PR 正在等待审查反馈。您可以提供更新吗？"
+
+# 通知
+notify_enabled: false
+notify_shell: "auto" # "auto"、"bash" 或 "fish"
+notify_on_stop: true
+notify_command_template: "curl -d \"$DEVLOOP_MESSAGE\" ntfy.sh/your-topic" # ntfy 模板示例
+
+# 环境
+workspace_mode: "local" # "local" 或 "gws" (用于 git-ws 隔离工作区)
+---
+
+# Devloop 项目指南
+
+- 在此处添加项目特定的开发规则。
+- 定义首选的测试命令（例如 "更改后运行 `npm test`"）。
+- 指定代码风格或架构限制。
+```
+
+1. 询问用户关键字段的值（`review_mode`、`wait_behavior`、`notify_enabled` 等）以自定义模板。
+2. 提醒钩子 (hook) 配置在会话开始时加载；需要重启 Claude Code 才能使钩子更改生效。
diff --git a/plugins/devloop/commands/devloop.md b/plugins/devloop/commands/devloop.md
index 553214d..3211f01 100644
--- a/plugins/devloop/commands/devloop.md
+++ b/plugins/devloop/commands/devloop.md
@@ -1,78 +1,33 @@
 ---
 name: devloop
-description: Start or resume the devloop workflow (create branch → fix → commit → PR → wait for AI review → apply comments → repeat).
+description: 启动或恢复 devloop 工作流（创建分支 → 修复 → 提交 → PR → 等待 AI 审查 → 应用评论 → 重复）。
 allowed-tools: ["Read", "Write", "Edit", "Grep", "Glob", "Bash", "AskUserQuestion", "TodoWrite", "Task", "Skill"]
 argument-hint: "--issue <github-url|number|feishu/lark-url|text|file> [--base main]"
 ---
 
-Run the devloop workflow using the plugin components in this plugin. This command drives a task to a merge-ready pull request through repeated cycles.
+使用本插件中的插件组件运行 devloop 工作流。此命令通过重复循环，将任务推进到可以合并的 pull request。
 
-## Mandatory Workflow
+## 强制工作流
 
-1. **Create Branch**:
-   - If on the base branch (e.g. `main`): Create a new descriptive branch based on the issue/task content.
-   - If NOT on the base branch: Check if the current branch is already associated with a different PR or issue. Prompt if a mismatch is detected.
-2. **Implement Fix**: Research and implement the smallest correct fix.
-3. **Validate**: Run the smallest relevant test/build command.
-4. **Commit**: Create a clear commit message.
-5. **Pull Request**: Open a PR for review.
-6. **Wait for Review**: Poll for review comments and PR mergeability status (`MERGEABLE`, `UNKNOWN`, or `CONFLICTING`).
-7. **Address Feedback**: Apply changes based on review comments and commit/push again.
-8. **Repeat**: Iterate until the PR is approved and `mergeable` is `MERGEABLE`.
+1. **读取设置**：读取 `.claude/devloop.local.md` 以获取项目特定的配置（审查模式、分支等）。
+2. **确定问题来源**：从 GitHub issue、URL、飞书项目、本地文件或文本描述中识别任务。
+3. **调用循环代理**：启动 `devloop-runner` 以执行完整的 修复/审查 周期。
 
-## Behavior
+## 行为
 
-1. **Determine the issue source**:
+- **初始设置**：如果在基准分支（默认 `main`）上，则创建一个新分支。
+- **开发周期**：将实施任务委托给 `devloop-implementer`，将验证任务委托给 `devloop-validator`。
+- **审查循环**：
+  - **轮询审查**：等待 PR 状态更改和审查评论。
+  - **审查策略**：遵循设置中的 `review_mode`。
+    - `github`（默认）：使用 `scripts/devloop-pr-review-threads.sh` 轮询评论。
+    - `custom`：在每个周期触发特定的技能（例如 `coderabbit:review`）或脚本。
+  - **处理反馈**：自动针对新评论实施修复并重新验证。
+- **完成**：当 PR 获得批准且 `mergeable` 状态为 `MERGEABLE` 时停止。
 
-   - If the argument looks like a GitHub URL or issue/PR number, use `gh` to fetch title/body, labels, repo, and existing PR linkage.
-   - If the argument looks like a Feishu/Lark Project issue URL or identifier, fetch the issue title/body via Feishu Project OpenAPI (requires local credentials) and use that as the task description.
-     - Typical issue URL pattern (seen in the wild): `https://project.feishu.cn/<project_key>/<work_item_type>/detail/<work_item_id>` (e.g. `.../story/detail/123`).
-     - Auth (Project OpenAPI): send `X-PLUGIN-TOKEN` (and sometimes `X-USER-KEY`) headers.
-       - Get a plugin token (example): `POST {base_url}/open_api/authen/plugin_token` with JSON `{ "plugin_id": "...", "plugin_secret": "..." }`, then use response `data.token` as `X-PLUGIN-TOKEN`.
-     - Fetch work item details (example): `POST {base_url}/open_api/{project_key}/work_item/{work_item_type_key}/query` with JSON `{ "work_item_ids": [<work_item_id>] }`.
-   - When NO argument is provided, or if starting on a non-base branch, use `gh pr list --head $(git branch --show-current) --json number,url,title,body` to check for an existing PR associated with the current branch.
-   - Should NO argument be provided and NO existing PR is found, the agent will prompt for a task description or offer to create a new issue.
-   - If the argument looks like a local file path, read it and treat it as the issue/task description.
-   - Otherwise, treat it as a free-form text task.
-   - For cases where a text task or local file is provided and no GitHub issue exists, the agent will offer to create one to track the work.
+## 规则与安全
 
-2. **Read settings** from `.claude/devloop.local.md` if present. Supported fields in YAML frontmatter:
-
-   - `enabled: true|false`
-   - `base_branch: "main"`
-   - `review_mode: "github"|"custom"`
-     - If `review_mode` is `"custom"`, you may set:
-       - `custom_review_skill: "..."` (e.g., `coderabbit:review`)
-   - `max_review_polls: 40`
-   - `review_poll_seconds: 60`
-   - `wait_behavior: "poll"|"ping_ai"`
-   - `ai_reviewer_id: "..."` (e.g., a bot account/login to ping)
-   - `ping_message_template: "..."`
-   - `ping_threshold: 3`
-   - `notify_enabled: true|false`
-   - `notify_shell: "auto"|"bash"|"fish"`
-   - `notify_on_stop: true|false`
-   - `notify_command_template: "..."`
-   - `workspace_mode: "gws"|"local"` (set to `"gws"` for `git-ws` integration)
-
-3. **Invoke the loop agent** `devloop-runner` to execute the full fix/review cycle.
-
-## Rules & Safety
-
-- **Branch Logic**: Create a new branch based on the issue content ONLY if the current branch is the base branch.
-- **Git Protocol**: NEVER use `git push --force`, `git push -f`, or `git commit --amend` on branches that have already been pushed to the remote or have an open PR. Always create new commits and use standard `git push`.
-- **Review Polling**: The agent will remain in an autonomous polling loop using `sleep` between polls.
-  - GraphQL for filtering active (not outdated/resolved) review thread comments is wrapped in a helper script:
-
-    ```bash
-    bash "${CLAUDE_PLUGIN_ROOT}/scripts/devloop-pr-review-threads.sh" --repo "{owner}/{repo}" --pr {number}
-    ```
-
-- **Presence & Communication (MANDATORY)**:
-  - CRITICAL: DO NOT include "Co-authored-by: Claude" or any variation in commit messages.
-  - CRITICAL: DO NOT include "Generated with Claude" or AI signatures in PR descriptions. Specifically, NEVER include "🤖 Generated with [Claude Code](https://claude.com/claude-code)".
-  - Maintain a professional, human-like presence in all git and GitHub metadata.
-  - If any sub-agent or tool includes these, you MUST remove them before completing the task.
-- **Avoid destructive operations**.
-- If review comments request changes that look incorrect or out-of-scope, ask the user before proceeding.
-- Prefer using `gh` for GitHub workflows when available.
+- **Git 协议**：严禁在共享/远程分支上使用 `git push --force` 或 `git commit --amend`。
+- **无 AI 签名**：提交或 PR 中绝对不允许出现 "Co-authored-by: Claude" 或 "Generated with Claude"。
+- **自主轮询**：使用 `sleep` 保持在轮询循环中，两轮之间无需请求用户交互。
+- **通知**：如果启用，使用 `scripts/devloop-notify.sh` 通过 `hooks.json` 发送更新。
diff --git a/plugins/devloop/scripts/devloop-pr-review-threads.sh b/plugins/devloop/scripts/devloop-pr-review-threads.sh
old mode 100644
new mode 100755
index 9bcae5d..10c3f50
--- a/plugins/devloop/scripts/devloop-pr-review-threads.sh
+++ b/plugins/devloop/scripts/devloop-pr-review-threads.sh
@@ -188,9 +188,10 @@ def gh_graphql(query: str, variables: Dict[str, Any]) -> Dict[str, Any]:
 
     try:
         data = json.loads(res.stdout)
-    except Exception:
+    except json.JSONDecodeError as e:
         sys.stderr.write(res.stdout)
-        raise
+        sys.stderr.write(str(e) + "\n")
+        raise SystemExit(1)
 
     errors = data.get("errors")
     if errors:
@@ -204,6 +205,7 @@ def gh_graphql(query: str, variables: Dict[str, Any]) -> Dict[str, Any]:
 def emit_jsonl(obj: Any) -> None:
     sys.stdout.write(json.dumps(obj, ensure_ascii=False))
     sys.stdout.write("\n")
+    sys.stdout.flush()
 
 
 def main() -> int:
diff --git a/plugins/devloop/skills/feishu-lark-interaction/SKILL.md b/plugins/devloop/skills/feishu-lark-interaction/SKILL.md
new file mode 100644
index 0000000..b86c02d
--- /dev/null
+++ b/plugins/devloop/skills/feishu-lark-interaction/SKILL.md
@@ -0,0 +1,95 @@
+---
+name: feishu-lark-interaction
+description: 关于如何使用飞书项目 (Feishu Project) OpenAPI 获取并管理 issue 详情的指南。
+---
+
+# 飞书/Lark 交互技能 (Skill)
+
+此技能提供了关于通过 OpenAPI 与飞书项目 (Feishu/Lark Project) 管理工具进行交互的指南。
+
+## 身份验证
+
+飞书项目 OpenAPI 的大多数请求需要两个自定义请求头：
+
+- `X-PLUGIN-TOKEN`：通过插件身份验证流程获取。
+- `X-USER-KEY`：（可选）用于受权限限制操作的用户特定密钥。
+
+### 获取插件令牌 (Plugin Token)
+
+```http
+POST {base_url}/open_api/authen/plugin_token
+Content-Type: application/json
+
+{
+  "plugin_id": "YOUR_PLUGIN_ID",
+  "plugin_secret": "YOUR_PLUGIN_SECRET"
+}
+```
+
+响应中的 `data.token` 应作为 `X-PLUGIN-TOKEN` 请求头使用。
+
+## 错误处理
+
+### 1. 常见状态码
+
+- **401 Unauthorized**：`X-PLUGIN-TOKEN` 失效或过期。
+  - **处理策略**：自动重新执行 [获取插件令牌](#获取插件令牌-plugin-token) 流程，更新 `X-PLUGIN-TOKEN` 后重试请求。
+- **429 Too Many Requests**：触发频率限制。
+  - **处理策略**：遵循 `Retry-After` 响应头建议的时间。若无，则采用指数退避（Exponential Backoff）配合随机抖动（Jitter）进行重试。
+- **5xx Server Error**：服务端异常。
+  - **处理策略**：自动重试并配合退避机制，若多次失败（如超过 3 次）则记录错误并切换至故障恢复模式。
+
+### 2. 令牌管理与刷新
+
+`data.token` 应被视为短期令牌。实现时应遵循：
+
+- **按需获取**：在发起请求前或收到 401 响应时，自动调用 `POST /open_api/authen/plugin_token` 进行续期。
+- **持久化**：令牌应在单次任务生命周期内缓存，避免频繁调用认证接口。
+
+### 3. 网络故障处理
+
+对于暂时性的网络波动：
+
+- **退避机制**：实施指数退避算法（如 1s, 2s, 4s...），增加随机抖动以防止请求洪峰。
+- **最大重试**：建议最大重试次数为 3-5 次，超过后应通过日志记录并中止当前操作，避免资源浪费。
+
+## 核心操作
+
+### 1. 获取 Issue 详情
+
+要获取工作项（故事、缺陷、任务）的标题和描述：
+
+```http
+POST {base_url}/open_api/{project_key}/work_item/{work_item_type_key}/query
+X-PLUGIN-TOKEN: p-xxxxxx
+Content-Type: application/json
+
+{
+  "work_item_ids": [123456]
+}
+```
+
+- **project_key**：在项目 URL 中可以找到（例如 `https://project.feishu.cn/<project_key>/...`）。
+- **work_item_type_key**：项目类型（例如 `story`、`bug`、`task`）。
+- **work_item_id**：URL 中的数字 ID。
+
+### 2. URL 解析
+
+典型的 issue URL 模式：
+`https://project.feishu.cn/<project_key>/<work_item_type>/detail/<work_item_id>`
+
+`devloop` 代理应自动从提供的 URL 中提取这些组件。
+
+## 配置要求
+
+要使用此集成，用户必须提供凭据（通常通过环境变量或安全的本地文件）：
+
+- `FEISHU_PROJECT_BASE_URL`
+- `FEISHU_PROJECT_PLUGIN_ID`
+- `FEISHU_PROJECT_PLUGIN_SECRET`
+
+## 映射到 Devloop
+
+- **标题 (Title)**：映射到 devloop 任务标题。
+- **描述 (Description)**：映射到 devloop 实施需求。
+- **状态映射**：飞书项目的状态（如“进行中”、“已修复”）可用于将 devloop 生命周期同步回项目看板（如果需要）。
diff --git a/plugins/devloop/skills/github-interaction/SKILL.md b/plugins/devloop/skills/github-interaction/SKILL.md
new file mode 100644
index 0000000..c2dc544
--- /dev/null
+++ b/plugins/devloop/skills/github-interaction/SKILL.md
@@ -0,0 +1,80 @@
+---
+name: github-interaction
+description: 关于如何使用 GitHub CLI (gh) 和 GraphQL 交互处理 issue、pull request 和审查评论的指南。
+---
+
+# GitHub 交互技能 (Skill)
+
+此技能提供了关于使用 `gh` CLI 和 GitHub 的 GraphQL API 来管理 `devloop` 工作流中的开发生命周期的指南。
+
+## 核心操作
+
+### 1. Issue 与 PR 管理
+
+- **获取 Issue 详情**：`gh issue view <id> --json title,body,labels`
+- **获取 PR 状态**：`gh pr view --json number,state,mergeable,reviewDecision,isDraft`
+- **列出分支对应的 PR**：`gh pr list --head <branch-name> --json number,url,title,body`
+- **创建 PR**：`gh pr create --title "<title>" --body "<body>" --base <base-branch>`
+
+### 2. 获取审查评论 (GraphQL)
+
+标准的 `gh pr view --json reviews` 通常会返回截断或过时的评论。对于 `devloop`，我们使用专门的 GraphQL 查询来获取**活跃（未过时且未解决）**的审查线程。
+
+#### 使用辅助脚本
+
+`devloop` 插件提供了一个脚本来处理分页和过滤：
+
+```bash
+bash "${CLAUDE_PLUGIN_ROOT}/scripts/devloop-pr-review-threads.sh" --repo "owner/repo" --pr <number>
+```
+
+该脚本输出 JSON Lines (JSONL)，可以逐行解析。
+
+#### 原始 GraphQL 查询模式
+
+如果你需要手动查询，请使用 `gh api graphql` 并配合支持分页的模式：
+
+```graphql
+query($name: String!, $owner: String!, $pr: Int!, $cursor: String) {
+  repository(owner: $owner, name: $name) {
+    pullRequest(number: $pr) {
+      reviewThreads(first: 50, after: $cursor) {
+        pageInfo {
+          hasNextPage
+          endCursor
+        }
+        nodes {
+          isOutdated
+          isResolved
+          comments(last: 1) {
+            nodes { body path line author { login } }
+          }
+        }
+      }
+    }
+  }
+}
+```
+
+**处理分页**：
+
+- **初始请求**：将 `$cursor` 设置为 `null`。
+- **后续请求**：如果 `pageInfo.hasNextPage` 为 `true`，获取 `pageInfo.endCursor` 并将其作为下一次请求中的 `$cursor` 变量传入。
+- **批次大小**：建议使用 `first: 50` 以平衡负载大小和请求次数。
+
+## 状态定义
+
+- **可合并性 (Mergeable)**：
+  - `MERGEABLE`：准备好合并。
+  - `CONFLICTING`：存在合并冲突（需要手动干预）。
+  - `UNKNOWN`：GitHub 仍在计算中（再次轮询）。
+- **审查决定 (Review Decision)**：
+  - `APPROVED`：PR 已获批准。
+  - `CHANGES_REQUESTED`：需要修复。
+  - `REVIEW_REQUIRED`：仍在等待审查。
+
+## 最佳实践
+
+- **分页**：对于大型 PR，始终处理 `reviewThreads` 和 `comments` 的分页。
+- **过滤**：仅处理 `isResolved: false` 且 `isOutdated: false` 的线程，以避免重复工作。
+- **草案 (Drafts)**：如果 `isDraft: true`，应继续轮询，但可以抑制提醒/通知。