fix: 将 agent 定义移至 .claude/agents/ 实现 session 自动加载 - Closes #108

1. 创建 .claude/agents/dev-agent.md / qe-agent.md — agent 定义文件
2. _common.sh: launch_agent 改为接收绝对路径的 agent 定义文件
3. start_dev_agent.sh / start_qe_agent.sh: 传递 .claude/agents/ 下的文件路径

Claude Code 启动时通过 --agent .claude/agents/<name>.md 自动加载
frontmatter + body 作为系统指令。

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

scripts/_common.sh

@@ -73,23 +73,27 @@ banner() {
 }
 
 # ── Launch agent in selected mode ──────────────────────────────────────────────
-# Usage: launch_agent <agent-file> <agent-name> <single-shot-task> <polling-instruction>
+# Usage: launch_agent <agent-name> <agent-file> <display-name> <single-shot-task> <polling-instruction>
 #
-# agent-name is the persona name (e.g. "Dev-Agent", "QE-Agent"). It is used to
-# prefix prompts so the model adopts the correct identity.
+# agent-name is the agent config name (e.g. "dev-agent", "qe-agent") used with
+# --agent flag. The agent file lives in .claude/agents/<agent-name>.md (with
+# frontmatter + body loaded as system prompt at session start).
+#
+# display-name is the persona name (e.g. "Dev-Agent", "QE-Agent") used to prefix
+# prompts so the model adopts the correct identity.
 #
 # Mode 1 (single-shot): claude -p, runs once and exits.
 #   --dangerously-skip-permissions avoids blocking in non-interactive mode.
-#   The project .claude/settings.json already sets permissionMode: bypass.
 #
 # Mode 2 (interactive polling): claude --agent, opens Claude Code TUI.
-#   The agent file defines startup behavior (e.g. /loop 10m) and the
-#   user can observe or interact at any time.
+#   The agent config is loaded from .claude/agents/<agent-name>.md,
+#   its body becomes the system prompt.
 launch_agent() {
-    local agent_file="$1"
-    local agent_name="$2"
-    local single_shot_task="$3"
-    local polling_instruction="${4:-}"
+    local agent_name="$1"
+    local agent_file="$2"
+    local display_name="$3"
+    local single_shot_task="$4"
+    local polling_instruction="${5:-}"
 
     echo "模式选择:"
     echo "  [1] 单次任务 — 检查 Issue 并处理，完成后自动退出 (automode)"
@@ -106,16 +110,16 @@ launch_agent() {
             claude -p \
                 --agent "$agent_file" \
                 --dangerously-skip-permissions \
-                "你是 ${agent_name}。${single_shot_task}"
+                "你是 ${display_name}。${single_shot_task}"
             ;;
         2)
             echo "启动互动轮询模式..."
-            echo "${agent_name} 进入 Claude Code 界面后将自动开始轮询"
+            echo "${display_name} 进入 Claude Code 界面后将自动开始轮询"
             echo "你可以随时输入指令与 Agent 互动，按 Ctrl+C 停止"
             echo ""
             cd "$PROJECT_DIR"
             claude --agent "$agent_file" \
-                "你是 ${agent_name}。${polling_instruction}"
+                "你是 ${display_name}。${polling_instruction}"
             ;;
         *)
             echo "无效选择，请输入 1 或 2。"

scripts/start_dev_agent.sh

@@ -24,8 +24,10 @@ trap 'cleanup_worktree "$GITEA_USER"' EXIT
 banner "Dev"
 require_token
 
+AGENT_CONF="$_MAIN_REPO_DIR/.claude/agents/dev-agent.md"
 launch_agent \
-    "agents/DEV_AGENT.md" \
+    "dev-agent" \
+    "$AGENT_CONF" \
     "Dev-Agent" \
     "执行一次 Issue 巡检（单次任务，不要用 /loop）：
 1. python scripts/agent_poller.py --action list 列出所有打开的 Issue

scripts/start_qe_agent.sh

@@ -24,8 +24,10 @@ trap 'cleanup_worktree "$GITEA_USER"' EXIT
 banner "QE"
 require_token
 
+AGENT_CONF="$_MAIN_REPO_DIR/.claude/agents/qe-agent.md"
 launch_agent \
-    "agents/QE_AGENT.md" \
+    "qe-agent" \
+    "$AGENT_CONF" \
     "QE-Agent" \
     "执行一次 Issue 巡检（单次任务，不要用 /loop）：
 1. python scripts/agent_poller.py --action list --labels test-code 检查 test-code Issue