feat: agent definition v2 (registry, persistence, developerInstructions channel)

Registry: POST /tasks accepts agent_name, loading <name>.json
from repo-root agents/ (traversal-guarded, skill-pack re-validated);
agent and agent_name together is a 400. Persistence: tasks gains
agent_name/agent_version/agent_instructions (migrate + PG schema),
returned by GET /tasks; stored prompt stays clean. Channel: local
and docker dispatch pass --developer-instructions to host.py
dispatch -> create_thread(developerInstructions=...); run/fork do
not re-send. Cloud executor has no such channel, so the worker
falls back to the v1 marked header in the in-memory dispatch
prompt only (DB prompt stays clean). Goal/coordinator tasks
still reject instructions.

Tests: test_agent_def_v2.py (7 tests) + updated v1 suite.

host.py

@@ -1803,7 +1803,8 @@ def _seed_worktree(cwd, seed_dir):
 
 def dispatch(server, state, name, prompt, repo, model, effort, sandbox,
              approval_policy, use_worktree=True, parent_task=None,
-             depends_on=None, priority=0, max_retries=0, seed_dir=None):
+             depends_on=None, priority=0, max_retries=0, seed_dir=None,
+             developer_instructions=None):
     if name in state["tasks"]:
         raise RuntimeError(f"task 已存在: {name}，先换名或 discard 它")
     if repo and not os.path.isabs(repo) and not os.path.exists(repo):
@@ -1823,8 +1824,10 @@ def dispatch(server, state, name, prompt, repo, model, effort, sandbox,
             raise RuntimeError(
                 "seed-dir 需要 worktree，当前派发没有建 worktree")
         _seed_worktree(cwd, seed_dir)
-    thread_id = server.create_thread(cwd, model=model, sandbox=sandbox,
-                                     approval_policy=approval_policy)
+    thread_id = server.create_thread(
+        cwd, model=model, sandbox=sandbox,
+        approval_policy=approval_policy,
+        instructions=developer_instructions)
     task = {"name": name, "thread_id": thread_id, "cwd": cwd,
             "repo": os.path.abspath(repo) if repo else None,
             "branch": branch, "sandbox": sandbox, "status": "running",
@@ -2247,6 +2250,10 @@ def main():
     ap.add_argument("--sandbox", default=None)
     ap.add_argument("--approval-policy", default=None)
     ap.add_argument("--approval", default=None, choices=["auto", "deny"])
+    ap.add_argument("--developer-instructions", default=None,
+                    help="agent 定义 v2：派发时传给 codex 的"
+                         " developerInstructions（只在 dispatch 建"
+                         " 线程时用，run/goal/fork 不用它）")
     ap.add_argument("--turn-timeout", type=int, default=None,
                     help="单轮等待上限（秒）；0 表示不限时、对齐 Codex App "
                          "等 turn/completed 的语义；不传时用配置 "
@@ -2709,7 +2716,9 @@ def main():
                             use_worktree=not args.no_worktree,
                             depends_on=deps, priority=args.priority,
                             max_retries=args.max_retries,
-                            seed_dir=args.seed_dir)
+                            seed_dir=args.seed_dir,
+                            developer_instructions=(
+                                args.developer_instructions))
             # 生产教训：先落盘任务登记再跑 turn，中途被杀/超时后
             # 任务还在 host-state 里，可用 run 续跑；否则 turn 中断
             # 就等于任务丢了（2026-10-25 委派评审 300 秒超时实测）。