Runbook: PAD-12 rollout record (keys, ping, smoke latencies, rollback)

Runbook: service, endpoint, signature, logs, secret rotation, disabling,
heartbeat skipped reasons, new failure rows. Design: webhook trigger design,
notify hook, Plane SSRF guard (WEBHOOK_ALLOWED_IPS). Deploy: install steps.
Record: synthetic-delivery acceptance and measured latencies.

README.md

@@ -2,11 +2,12 @@
 
 Plane 看板驱动的通用 agent 调度服务：在 Plane 建工作项并指派给某个终端，agentd 在该终端起对应的 CLI agent（Codex / Claude Code / 脚本）执行，把结果贴回工作项并流转状态。调度是确定性代码；模型只出现在执行器和总指挥里。
 
+- 使用说明（在 Plane 上怎么派活、回答、取消，标签和默认值）：`docs/usage.md`
 - 设计：`docs/design.md`（架构、契约、失败策略、配置、阶段）
 - 首个租户：GWENT 项目（看板 https://plane.kappiee.com/kappiee/projects/79cb5a60-aa4f-4138-9864-1dd70426e665/issues/ ）
 - 凭据全部经 Agent Vault 注入内存，仓库和配置里只有键名。
 
 ## 使用
-- 契约与扩展：`docs/executor-contract.md`；安装、cron、日志、故障：`docs/runbook.md`
-- 入口：`cli/agentd tick|new|show|comment|state|assign|smoke|check-config`（在 `agent-access.py exec --secret-env …` 下运行）
+- 契约与扩展：`docs/executor-contract.md`；安装、cron、日志、故障：`docs/runbook.md`；LXC 部署（调度器 2026-10-21 起跑在 `agentd-01`）：`docs/deploy-lxc.md`
+- 入口：`cli/agentd tick|new|show|comment|state|assign|smoke|ping|check-config`（在 `agent-access.py exec --secret-env …` 下运行；`ping` 验证到终端的 direct-ssh / warpgate-ssh 路径）
 - 测试：`python3 -m unittest`（离线，不访问 Plane/win10）

agentd/cli.py

@@ -9,6 +9,7 @@ from . import config as config_mod
 from .core import Agentd
 from .plane import Plane, brief_to_html, html_to_text, text_to_html
 from .store import Busy
+from .transports import CallLog, TransportError, direct_ssh, make_transport
 
 
 def identity(cfg, wanted=None):
@@ -38,6 +39,14 @@ def cmd_tick(cfg, a):
         d.tick(project=a.project)
     except Busy as exc:
         print("skip: %s" % exc)
+        sys.exit(75)   # same code as the cron script's flock conflict: the webhook receiver retries later
+
+
+def cmd_webhook_serve(cfg, a):
+    if not cfg.webhook:
+        sys.exit("no webhook section in %s" % cfg.path)
+    from .webhook import serve
+    serve(cfg, a.bind, a.port)
 
 
 def cmd_new(cfg, a, pl=None, quiet=False):
@@ -116,17 +125,22 @@ def cmd_smoke(cfg, a):
     body = {"name": "agentd smoke %s %s" % (a.terminal, time.strftime("%Y-%m-%d %H:%M")),
             "description_html": SMOKE_BRIEF, "state": pl.states(p["id"])[p["states"]["todo"]],
             "assignees": [pl.members()[user]]}
+    if a.labels:
+        labels = pl.labels(p["id"])
+        body["labels"] = [labels[x] for x in a.labels.split(",")]
     issue = pl.create_issue(p["id"], body)
     key = "%s-%s" % (p.get("key_prefix", a.project), issue["sequence_id"])
-    print("created %s; ticking every %ss (timeout %s min)" % (key, a.interval, a.timeout))
-    d = Agentd(cfg)
+    print("created %s; %s every %ss (timeout %s min)" % (
+        key, "watching (the scheduled dispatcher runs it)" if a.watch else "ticking", a.interval, a.timeout), flush=True)
+    d = None if a.watch else Agentd(cfg)
     states = pl.states(p["id"])
     done_states = {states[p["states"][k]] for k in ("review", "waiting")}
     deadline = time.time() + a.timeout * 60
     state = None
     while time.time() < deadline:
         try:
-            d.tick(project=a.project, only_issue=issue["id"])
+            if d:
+                d.tick(project=a.project, only_issue=issue["id"])
         except Busy as exc:
             print("skip: %s" % exc)
         state = pl.get_issue(p["id"], issue["id"]).get("state")
@@ -146,6 +160,31 @@ def cmd_smoke(cfg, a):
     sys.exit(0 if ok else 1)
 
 
+def cmd_ping(cfg, a):
+    """Trivial command on a terminal, timed per call (verifies direct-ssh / warpgate-ssh after a deployment). The
+    direct-ssh key comes from the environment like the Plane tokens (agent-access.py exec --secret-env)."""
+    if a.terminal not in cfg.terminals:
+        sys.exit("unknown terminal %s" % a.terminal)
+    t = cfg.terminals[a.terminal]
+    if (a.transport or t["transport"]) == "local":
+        sys.exit("%s is a local terminal: nothing to ping" % a.terminal)
+    if a.transport == "direct-ssh" and "ssh" not in t:
+        sys.exit("terminals.%s has no ssh section" % a.terminal)
+    log = CallLog(event=lambda m: print("event: " + m, flush=True), echo=True)
+    tr = make_transport(cfg, a.terminal, log=log, force=a.transport)
+    ok = True
+    for i in range(a.count):
+        t0 = time.monotonic()
+        try:
+            pong = tr.ping()
+        except TransportError as exc:
+            ok = False
+            print("ping %d: FAILED after %d ms: %s" % (i + 1, (time.monotonic() - t0) * 1000, exc), flush=True)
+            continue
+        print("ping %d: %d ms %s" % (i + 1, (time.monotonic() - t0) * 1000, pong), flush=True)
+    sys.exit(0 if ok else 1)
+
+
 def cmd_check_config(cfg, a):
     print("config %s is valid" % cfg.path)
     print("projects: %s" % ", ".join(cfg.projects))
@@ -153,7 +192,17 @@ def cmd_check_config(cfg, a):
         env = cfg.token_env(name)
         print("terminal %-10s %-13s %-8s executors=%s token %s: %s" % (
             name, t["transport"], t["os"], ",".join(t["executors"]), env, "set" if os.environ.get(env) else "MISSING"))
+        if t["transport"] == "direct-ssh":
+            s = t["ssh"]
+            print("  direct-ssh %s@%s:%s, fallback %s" % (s["user"], s["host"], s.get("port", 22), t.get("fallback", "none")))
+    if cfg.uses_direct_ssh():
+        print("direct-ssh key %s: %s; ssh_dir %s" % (
+            cfg.ssh_key_env, "set" if direct_ssh.key_present(cfg.ssh_key_env) else "MISSING (direct terminals use their fallback)",
+            cfg.ssh_dir))
     print("cron/exec secret args: %s" % " ".join(cfg.secret_env_args()))
+    if cfg.webhook:
+        k = cfg.webhook["secret_key"]
+        print("webhook-serve secret arg: --secret-env %s=%s (%s)" % (k, k, "set" if os.environ.get(k) else "not in this env"))
 
 
 def parser():
@@ -195,9 +244,22 @@ def parser():
     s.add_argument("terminal")
     s.add_argument("--interval", type=int, default=30)
     s.add_argument("--timeout", type=int, default=30, help="minutes")
+    s.add_argument("--labels", default="", help="comma-separated Plane labels, e.g. model:sol")
+    s.add_argument("--watch", action="store_true",
+                   help="do not tick here; only wait for the scheduled dispatcher (use when it runs on another host)")
     s.add_argument("--as", dest="as_")
     s.set_defaults(fn=cmd_smoke)
+    s = sub.add_parser("webhook-serve", help="Plane webhook receiver: each delivery triggers a tick (agentd/webhook.py)")
+    s.add_argument("--bind", help="address (default: webhook.bind)")
+    s.add_argument("--port", type=int, help="port (default: webhook.port)")
+    s.set_defaults(fn=cmd_webhook_serve)
     sub.add_parser("check-config", help="validate the config and list token env vars").set_defaults(fn=cmd_check_config)
+    s = sub.add_parser("ping", help="time a trivial command on a terminal (direct-ssh needs the key env var, see check-config)")
+    s.add_argument("terminal")
+    s.add_argument("--transport", choices=["direct-ssh", "warpgate-ssh"],
+                   help="force one channel, no fallback (default: the terminal's transport with its fallback)")
+    s.add_argument("-n", "--count", type=int, default=3, help="calls (default 3: the 2nd+ reuse the SSH connection)")
+    s.set_defaults(fn=cmd_ping)
     return ap
 
 
@@ -207,4 +269,5 @@ def main(argv=None):
         cfg = config_mod.load(a.config)
     except config_mod.ConfigError as exc:
         sys.exit(str(exc))
+    direct_ssh.take_key(cfg.ssh_key_env)   # out of os.environ before any child process starts
     a.fn(cfg, a)

agentd/config.py

@@ -2,8 +2,10 @@
 
 The config only ever holds Vault *key names* (`token_key`). The token itself is read from the environment
 variable of the same name, which `agent-access.py exec --secret-env KEY=KEY` injects into the process."""
+import base64
 import json
 import os
+import struct
 from pathlib import Path
 
 import yaml
@@ -11,6 +13,7 @@ import yaml
 from . import ROOT
 
 SCHEMA_PATH = ROOT / "config" / "schema.json"
+DEFAULT_SSH_KEY = "AGENTD_SSH_KEY"     # env var (== Vault key name) carrying the direct-ssh private key
 DEFAULT_PATHS = (ROOT / "config" / "agentd.local.yaml", ROOT / "config" / "agentd.yaml")
 
 
@@ -83,6 +86,48 @@ def schema_errors(data, schema):
             for e in validator.iter_errors(data)]
 
 
+def host_key_error(value):
+    """None if `value` is a usable pinned host key ('ssh-ed25519 <base64>' whose blob names the same type)."""
+    parts = str(value).split()
+    if len(parts) < 2 or parts[0] != "ssh-ed25519":
+        return "must be 'ssh-ed25519 <base64>' (the terminal's /etc/ssh/ssh_host_ed25519_key.pub, verified out of band)"
+    try:
+        blob = base64.b64decode(parts[1], validate=True)
+        n = struct.unpack(">I", blob[:4])[0]
+        ok = blob[4:4 + n] == b"ssh-ed25519" and len(blob) == 4 + n + 4 + 32
+    except (ValueError, struct.error):
+        ok = False
+    return None if ok else "the base64 part is not an ssh-ed25519 public key (copy/paste damage?)"
+
+
+def ssh_dir_of(data):
+    d = (data.get("agentd") or {}).get("ssh_dir")
+    if d:
+        p = Path(os.path.expanduser(d))
+        return p if p.is_absolute() else ROOT / p
+    s = Path(os.path.expanduser((data.get("agentd") or {}).get("state_dir", "state")))
+    return (s if s.is_absolute() else ROOT / s) / "ssh"
+
+
+def remote_errors(data, tname, t, kind):
+    """Rules every remote transport (warpgate-ssh, direct-ssh, and the fallback channel) shares."""
+    errs = []
+    if t.get("os") == "windows":
+        if t.get("launch") != "schtasks":
+            errs.append("terminals.%s: %s on windows needs launch=schtasks" % (tname, kind))
+        if kind == "warpgate-ssh" and t.get("target", tname) != "win10":
+            errs.append("terminals.%s: the Windows helper scripts only reach target 'win10' for now" % tname)
+    else:
+        if t.get("launch") != "nohup":
+            errs.append("terminals.%s: %s on %s needs launch=nohup" % (tname, kind, t.get("os")))
+        for key in ("repo", "run_root"):
+            if key in t and not str(t[key]).startswith("/"):
+                errs.append("terminals.%s.%s: must be an absolute path on the remote (no ~)" % (tname, key))
+        if kind == "warpgate-ssh" and not data.get("vault", {}).get("config"):
+            errs.append("terminals.%s: remote POSIX terminals need vault.config (relay identity)" % tname)
+    return errs
+
+
 def semantic_errors(data):
     """Cross-references the schema cannot express."""
     errs = []
@@ -91,13 +136,36 @@ def semantic_errors(data):
         for ex in t.get("executors", []):
             if ex not in execs:
                 errs.append("terminals.%s: unknown executor '%s'" % (tname, ex))
-        if t.get("transport") == "warpgate-ssh":
-            if t.get("os") != "windows" or t.get("launch") != "schtasks":
-                errs.append("terminals.%s: warpgate-ssh currently supports only os=windows, launch=schtasks" % tname)
-            if t.get("target") != "win10":
-                errs.append("terminals.%s: warpgate-ssh helper scripts only reach target 'win10' for now" % tname)
-        if t.get("transport") == "local" and t.get("os") != "linux":
+        kind = t.get("transport")
+        if kind == "warpgate-ssh":
+            errs += remote_errors(data, tname, t, kind)
+        elif kind == "direct-ssh":
+            errs += remote_errors(data, tname, t, kind)
+            s = t.get("ssh")
+            if not s:
+                errs.append("terminals.%s: direct-ssh needs ssh: {host, user, port, host_key}" % tname)
+            else:
+                if not s.get("host") or str(s["host"]).startswith("-") or not s.get("user"):
+                    errs.append("terminals.%s.ssh: host and user must be non-empty (host not starting with '-')" % tname)
+                if s.get("port", 22) > 65535:
+                    errs.append("terminals.%s.ssh.port: %r > 65535" % (tname, s["port"]))
+                why = host_key_error(s.get("host_key", ""))
+                if why:
+                    errs.append("terminals.%s.ssh.host_key: %s" % (tname, why))
+            if t.get("fallback") == "warpgate-ssh":
+                if "target" not in t:
+                    errs.append("terminals.%s: fallback warpgate-ssh needs the Warpgate 'target'" % tname)
+                elif t.get("os") == "windows" and t["target"] != "win10":
+                    errs.append("terminals.%s: the Windows helper scripts (fallback) only reach target 'win10' for now" % tname)
+                if t.get("os") != "windows" and not data.get("vault", {}).get("config"):
+                    errs.append("terminals.%s: fallback warpgate-ssh needs vault.config (relay identity)" % tname)
+        elif t.get("fallback"):
+            errs.append("terminals.%s: 'fallback' only applies to transport direct-ssh" % tname)
+        if kind == "local" and t.get("os") != "linux":
             errs.append("terminals.%s: local transport supports only os=linux" % tname)
+    ctl = ssh_dir_of(data)   # "/<pin8>-<40-char %C>", plus ".<16 chars>" while ssh sets the socket up
+    if any(t.get("transport") == "direct-ssh" for t in terms.values()) and len(str(ctl)) + 1 + 9 + 40 + 17 > 104:
+        errs.append("agentd.ssh_dir: %s is too long for the ControlMaster socket path (UNIX limit 104)" % ctl)
     for ename, e in execs.items():
         adapter = e.get("adapter", ename)
         if adapter not in ("codex", "claude-code", "shell"):
@@ -140,15 +208,40 @@ class Config:
         self.executors = data["executors"]
         self.projects = data["projects"]
         self.agentd = data.get("agentd", {})
+        self.webhook = data.get("webhook", {})
+        self.notify = data.get("notify") or {"kind": "none"}
 
     @property
     def state_dir(self):
         p = Path(os.path.expanduser(self.agentd.get("state_dir", "state")))
         return p if p.is_absolute() else ROOT / p
 
+    @property
+    def ssh_dir(self):
+        """direct-ssh ControlMaster sockets + generated known_hosts (default <state_dir>/ssh)."""
+        return ssh_dir_of(self.data)
+
+    @property
+    def ssh_key_env(self):
+        """Env var (== Vault key name) holding the direct-ssh private key."""
+        return self.agentd.get("ssh_key") or DEFAULT_SSH_KEY
+
+    def uses_direct_ssh(self):
+        return any(t["transport"] == "direct-ssh" for t in self.terminals.values())
+
     def helpers_dir(self, terminal):
-        d = self.terminals[terminal].get("helpers_dir") or self.agentd.get("helpers_dir") or str(ROOT / "tools")
-        return Path(os.path.expanduser(d))
+        d = self.terminals[terminal].get("helpers_dir") or self.agentd.get("helpers_dir") or "tools/win10"
+        p = Path(os.path.expanduser(d))
+        return p if p.is_absolute() else ROOT / p
+
+    @property
+    def log_dir(self):
+        """agentd.log lives here (default: state_dir)."""
+        d = self.agentd.get("log_dir")
+        if not d:
+            return self.state_dir
+        p = Path(os.path.expanduser(d))
+        return p if p.is_absolute() else ROOT / p
 
     def token_env(self, terminal):
         """Environment variable that carries the terminal's Plane token (== the Vault key name)."""
@@ -158,19 +251,23 @@ class Config:
         return os.environ.get(self.token_env(terminal))
 
     def secret_keys(self):
-        """Every env var name that carries a secret (Plane tokens, webhook secrets): never handed to executors."""
+        """Every env var name that carries a secret (Plane tokens, the webhook secret): never handed to children."""
         keys = {t["token_key"] for t in self.terminals.values()}
-        keys |= {p["webhook_secret_key"] for p in self.projects.values() if p.get("webhook_secret_key")}
+        if self.webhook.get("secret_key"):
+            keys.add(self.webhook["secret_key"])
         return keys
 
     def scrubbed_env(self):
-        """os.environ without the secrets above, for every child process agentd starts."""
-        drop = self.secret_keys()
+        """os.environ without the secrets above, for every child process agentd starts. The direct-ssh key variable
+        is always dropped, even with no direct-ssh terminal (a rollback may leave the cron script injecting it)."""
+        drop = self.secret_keys() | {self.ssh_key_env}
         return {k: v for k, v in os.environ.items() if k not in drop}
 
     def secret_env_args(self):
-        """`--secret-env` arguments for agent-access.py exec, one per distinct token key."""
+        """`--secret-env` arguments for agent-access.py exec: one per distinct token key, plus the direct-ssh key."""
         keys = sorted({t["token_key"] for t in self.terminals.values()})
+        if self.uses_direct_ssh():
+            keys.append(self.ssh_key_env)
         return ["--secret-env %s=%s" % (k, k) for k in keys]
 
     def project(self, name):

agentd/core.py

@@ -3,7 +3,13 @@
       on failure post a structured note and retry (transient) or hand over (Waiting); kill on timeout;
   (b) start the lowest-sequence Todo items whose terminal has a free slot and whose serialize group is idle;
   (c) write the heartbeat.
-Everything is deterministic; models only run inside executors."""
+Everything is deterministic; models only run inside executors.
+
+Double-start guard: before the launch call a `launching` marker (the planned run and its handle) is saved. If the
+launch call fails or the process dies before the run is recorded, the next tick probes the run directory: launch
+evidence (exit.txt / launch.pid / pid.txt) -> adopt the run; nothing -> relaunch with the same run id in a fresh
+directory (`<run_id>-r2`, -r3 ...); probe unreachable -> keep the marker (it holds the terminal slot).
+A terminal whose relay fails is skipped for the rest of the tick; other terminals carry on."""
 import os
 import re
 import time
@@ -15,7 +21,7 @@ from .executors import ExecResult, Run, make_executor
 from .executors.base import fill
 from .plane import Plane, html_to_text, text_to_html
 from .store import Store
-from .transports import make_transport
+from .transports import CallLog, TransportError, make_transport
 
 TAGS = ("review", "question", "blocked", "progress")
 WORKER = "[worker]"
@@ -143,17 +149,27 @@ class Ctx:
 class Agentd:
     def __init__(self, cfg, store=None, plane_factory=None, executor_factory=None, transport_factory=None, now=time.time):
         self.cfg = cfg
-        self.store = store or Store(cfg.state_dir)
+        self.store = store or Store(cfg.state_dir, log_dir=cfg.log_dir)
         self.now = now
         self.log = self.store.log
+        # one line per remote call (terminal, channel, operation, ms, rc) -> remote.log; fallback/host-key -> agentd.log
+        self.calllog = CallLog(self.store.log_dir / "remote.log", self.log)
         self.plane_factory = plane_factory or (lambda term: Plane(
             cfg.plane["base_url"], cfg.plane["workspace"], cfg.token(term), cfg.plane.get("timeout", 60)))
-        self.transport_factory = transport_factory or (lambda term: make_transport(cfg, term))
+        self.transport_factory = transport_factory or (lambda term: make_transport(cfg, term, log=self.calllog))
         self.executor_factory = executor_factory or (lambda term, name: make_executor(cfg, term, name, self.transport(term)))
         self._reset_caches()
 
     def _reset_caches(self):
         self._clients, self._transports, self._executors, self._ready = {}, {}, {}, {}
+        self._down = set()   # terminals whose relay failed during this tick
+        self._busy = {}      # terminal -> why terminal_ready() said no
+        self._skipped = []   # Todo items not dispatched this tick, with the reason (heartbeat "skipped")
+
+    def mark_down(self, term, exc):
+        if term not in self._down:
+            self.log("%s unreachable, skipping it for the rest of this tick: %s" % (term, exc))
+        self._down.add(term)
 
     # ---------------------------------------------------------------- caches
     def client(self, terminal):
@@ -195,11 +211,33 @@ class Agentd:
                     self.log("%s tick error: %r" % (name, exc))
                 finally:
                     self.store.save(data)
-            self.store.heartbeat({"errors": errors, "inflight": [
+            if not only_issue and self.note_skips(data, [project] if project else list(self.cfg.projects)):
+                self.store.save(data)
+            self.store.heartbeat({"errors": errors, "down": sorted(self._down), "skipped": self._skipped, "inflight": [
                 {"project": p, "run": r["inflight"]["run_id"], "terminal": r["inflight"]["terminal"]}
-                for p, ps in data["projects"].items() for r in ps["issues"].values() if r.get("inflight")]})
+                for p, ps in data["projects"].items() for r in ps["issues"].values() if r.get("inflight")],
+                "launching": [{"project": p, "run": r["launching"]["run_id"], "terminal": r["launching"]["terminal"]}
+                              for p, ps in data["projects"].items() for r in ps["issues"].values() if r.get("launching")]})
             return data
 
+    def note_skips(self, data, projects):
+        """Log a skip reason only when it is new for that item (the cron log stays quiet on repeat ticks).
+        Persisted as data["skips"] {item: {project, terminal, reason}}; returns True when that changed."""
+        prev = data.get("skips", {})
+        now = {k: v for k, v in prev.items() if v.get("project") not in projects}
+        for s in self._skipped:
+            entry = {"project": s["project"], "terminal": s["terminal"], "reason": s["reason"]}
+            if prev.get(s["item"]) != entry:
+                self.log("%s not dispatched (%s): %s" % (s["item"], s["terminal"] or "-", s["reason"]))
+            now[s["item"]] = entry
+        changed = now != prev
+        data["skips"] = now
+        return changed
+
+    def skip(self, ctx, issue, term, reason):
+        self._skipped.append({"project": ctx.name, "item": ctx.key(issue.get("sequence_id")), "terminal": term,
+                              "reason": reason})
+
     def loop(self, interval):
         from .store import Busy
         while True:
@@ -216,14 +254,43 @@ class Agentd:
         pstate = data["projects"].setdefault(name, {"issues": {}})
         ctx = Ctx(name, pcfg, self.reader(pcfg))
         for iid, rec in list(pstate["issues"].items()):
-            if rec.get("inflight"):
-                try:
+            run = rec.get("inflight") or rec.get("launching")
+            if not run or run["terminal"] in self._down:
+                continue
+            try:
+                if rec.get("inflight"):
                     self.poll(ctx, data, iid, rec)
-                except Exception as exc:
-                    self.log("%s poll error: %r" % (ctx.key(rec.get("seq")), exc))
-                self.store.save(data)  # persist each round's outcome (thread id, retry) before touching the next
+                else:
+                    self.resolve_launch(ctx, iid, rec)
+            except TransportError as exc:
+                self.mark_down(run["terminal"], exc)
+            except Exception as exc:
+                self.log("%s poll error: %r" % (ctx.key(rec.get("seq")), exc))
+            self.store.save(data)  # persist each round's outcome (thread id, retry) before touching the next
         self.pick(ctx, data, pstate, only_issue)
 
+    def resolve_launch(self, ctx, iid, rec):
+        """A launch that was never confirmed (the launch call failed or agentd died mid-start): probe the run
+        directory and adopt the run if it exists; otherwise relaunch in a fresh directory, never wiping one."""
+        m = rec["launching"]
+        key = ctx.key(rec["seq"])
+        present = self.executor(m["terminal"], m["executor"]).probe(m["handle"])  # TransportError -> marker kept
+        if present:
+            self.log("%s run %s: unconfirmed launch found on %s (%s); adopting it" % (
+                key, m["run_id"], m["terminal"], ",".join(present)))
+            return self.launched(ctx, iid, rec, m, m["handle"], adopted=True)
+        rec.pop("launching")
+        issue = ctx.issues.get(iid)
+        state = issue and issue.get("state")
+        if state == ctx.state_id("todo") or (state == ctx.state_id("running") and rec.get("retry")):
+            rec["relaunch"] = {"run_id": m["run_id"], "attempt": m.get("attempt", 1) + 1}
+            self.log("%s run %s: nothing at %s; relaunching in %s-r%d" % (
+                key, m["run_id"], m["handle"].get("run_dir"), m["run_id"], rec["relaunch"]["attempt"]))
+        else:
+            rec.pop("relaunch", None)
+            self.log("%s run %s: nothing at %s and the item is no longer dispatchable; marker dropped" % (
+                key, m["run_id"], m["handle"].get("run_dir")))
+
     # ---------------------------------------------------------------- (a) poll / finish
     def poll(self, ctx, data, iid, rec):
         fl = rec["inflight"]
@@ -393,6 +460,13 @@ class Agentd:
         client.set_state(ctx.pid, iid, state, users)
         if iid in ctx.issues:  # keep this tick's view consistent
             ctx.issues[iid].update(state=state, assignees=users)
+        self.notify(ctx, iid, state_key, user)
+
+    def notify(self, ctx, iid, event, user):
+        """Notification hook (design §7 phase 3), called on every state move (event = todo/waiting/review/...).
+        Config `notify: {kind: none}` is the only kind today: a no-op. An ntfy sender would go here."""
+        if self.cfg.notify.get("kind", "none") == "none":
+            return
 
     def post(self, ctx, client, iid, rec, text, pre=None):
         c = client.post_comment(ctx.pid, iid, text_to_html(text, pre))
@@ -401,28 +475,31 @@ class Agentd:
 
     # ---------------------------------------------------------------- (b) pick / start
     def pick(self, ctx, data, pstate, only_issue=None):
-        load, groups = Counter(), set()
-        for p in data["projects"].values():
+        load, groups, holders = Counter(), {}, {}   # groups / holders: group / terminal -> keys of the runs holding it
+        for pname, p in data["projects"].items():
+            prefix = self.cfg.projects.get(pname, {}).get("key_prefix", pname)
             for r in p["issues"].values():
-                fl = r.get("inflight")
+                fl = r.get("inflight") or r.get("launching")   # an unconfirmed launch holds its slot
                 if fl:
                     load[fl["terminal"]] += 1
+                    holders.setdefault(fl["terminal"], []).append("%s-%s" % (prefix, r.get("seq")))
                     if fl.get("group"):
-                        groups.add(fl["group"])
+                        groups.setdefault(fl["group"], []).append("%s-%s" % (prefix, r.get("seq")))
         for g, info in list(data["paused_groups"].items()):
             if info["project"] == ctx.name:
                 issue = ctx.issues.get(info["issue"])
                 if issue is None or issue.get("state") != ctx.state_id("waiting"):
                     del data["paused_groups"][g]
                     self.log("serialize group %s resumed (%s left Waiting)" % (g, info.get("key")))
-        groups |= set(data["paused_groups"])
+        for g, info in data["paused_groups"].items():
+            groups.setdefault(g, []).append("%s in %s (paused)" % (info.get("key"), ctx.cfg["states"]["waiting"]))
         todo, running, now = ctx.state_id("todo"), ctx.state_id("running"), self.now()
         cands = []
         for iid, issue in ctx.issues.items():
             if only_issue and iid != only_issue:
                 continue
             rec = pstate["issues"].get(iid, {})
-            if rec.get("inflight"):
+            if rec.get("inflight") or rec.get("launching"):
                 continue
             if issue.get("state") == todo:
                 cands.append(issue)
@@ -434,36 +511,54 @@ class Agentd:
         for issue in sorted(cands, key=lambda i: i.get("sequence_id", 0)):
             route = self.route(ctx, issue)
             if isinstance(route, str):
-                if route:
-                    self.log("%s not dispatched: %s" % (ctx.key(issue.get("sequence_id")), route))
+                users = [ctx.member_names.get(a, a) for a in issue.get("assignees") or []]
+                self.skip(ctx, issue, None, route or "no route for assignee %s" % (",".join(users) or "(none)"))
                 continue
             term, group = route["terminal"], route["group"]
-            if load[term] >= self.cfg.terminals[term].get("concurrency", 1) or (group and group in groups):
+            cap = self.cfg.terminals[term].get("concurrency", 1)
+            if term in self._down:
+                self.skip(ctx, issue, term, "relay down")
+                continue
+            if load[term] >= cap:
+                self.skip(ctx, issue, term, "concurrency %d/%d (%s)" % (load[term], cap, ", ".join(holders.get(term, [])) or "-"))
+                continue
+            if group and group in groups:
+                self.skip(ctx, issue, term, "serialize_group %s held by %s" % (group, ", ".join(groups[group])))
                 continue
             if not self.terminal_ready(term, load[term]):
+                self.skip(ctx, issue, term, self._busy.get(term) or "relay down")
                 continue
             try:
-                self.start(ctx, pstate, issue, route)
+                self.start(ctx, data, pstate, issue, route)
             except Exception as exc:
                 self.log("%s start failed on %s: %r" % (ctx.key(issue.get("sequence_id")), term, exc))
-                continue
+                self.skip(ctx, issue, term, "start failed: %s" % (str(exc)[:200] or type(exc).__name__))
+                if isinstance(exc, TransportError):
+                    self.mark_down(term, exc)
+                if not pstate["issues"].get(issue["id"], {}).get("launching"):
+                    continue
+                # the launch call failed but may have started the run: its marker keeps the slot until probed
             self.store.save(data)  # a launched run must be on disk before anything else can fail
             load[term] += 1
+            holders.setdefault(term, []).append(ctx.key(issue.get("sequence_id")))
             if group:
-                groups.add(group)
+                groups.setdefault(group, []).append(ctx.key(issue.get("sequence_id")))
 
     def terminal_ready(self, term, our_load):
         """Optional guard against processes we did not start (e.g. a manual codex session)."""
+        if term in self._down:
+            return False
         if term not in self._ready:
             names = self.cfg.terminals[term].get("busy_processes", [])
             ready = True
             for name in names:
                 count = self.transport(term).process_count(name)
                 if count is None:
-                    self.log("%s unreachable; skipping" % term)
+                    self.mark_down(term, "busy-process probe failed")
                     ready = False
+                    break
                 elif count > 0 and our_load == 0:
-                    self.log("%s busy (%d %s running that agentd did not start)" % (term, count, name))
+                    self._busy[term] = "busy_process %s (%d running, not started by agentd)" % (name, count)
                     ready = False
             self._ready[term] = ready
         return self._ready[term]
@@ -491,6 +586,8 @@ class Agentd:
                 r["model"] = label[6:]
             elif label.startswith("effort:"):
                 r["effort"] = label[7:]
+            elif label.startswith("tier:"):
+                r["tier"] = label[5:]
             elif label.startswith("cap:"):
                 caps.append(label[4:])
             elif parse_budget(label) is not None:
@@ -503,6 +600,10 @@ class Agentd:
         ecfg = self.cfg.executors[r["executor"]]
         r["model"] = r["model"] or (ecfg.get("models") or [None])[0]
         r["effort"] = r["effort"] or ecfg.get("default_effort")
+        if r["tier"] is None:  # per-model default unless a label set one; `tier:none` = no tier argument
+            r["tier"] = (ecfg.get("tier_defaults") or {}).get(r["model"])
+        if r["tier"] in ("none", ""):
+            r["tier"] = None
         return r
 
     def protocol(self, pcfg):
@@ -516,7 +617,7 @@ class Agentd:
             return ""
         return p.read_text(encoding="utf-8")
 
-    def start(self, ctx, pstate, issue, route):
+    def start(self, ctx, data, pstate, issue, route):
         iid, seq = issue["id"], issue["sequence_id"]
         key = ctx.key(seq)
         rec = pstate["issues"].setdefault(iid, new_rec(seq))
@@ -546,25 +647,11 @@ class Agentd:
         intro = fill(ctx.cfg.get("intro", DEFAULT_INTRO),
                      dict(terminal=term, executor=exname, repo=tcfg["repo"], prefix=ctx.prefix, key=key))
         prompt = build_prompt(key, issue, comments, seen, bool(thread), self.protocol(pcfg), intro, handoff, retry_note)
-        run = Run(run_id, tcfg["repo"], prompt, route["model"], route["effort"], route["tier"], thread, route["script"])
-        handle = ex.start(run)
-        rec.pop("retry", None)
-        rec.pop("rebuild", None)
-        if not is_retry:
-            rec["retries"] = 0
-        now = self.now()
-        # transient retries continue the same logical round: they do not use up max_rounds_per_thread
-        rounds = (rec.get("thread_rounds", 0) + (0 if is_retry else 1)) if thread else 1
-        rec.update(runs=k, thread_rounds=rounds, inflight={
-            "run_id": run_id, "terminal": term, "executor": exname, "handle": handle, "round": k,
-            "started": now, "deadline": now + route["timeout_hours"] * 3600, "resumed": bool(thread),
-            "group": route["group"], "model": route["model"], "effort": route["effort"]})
-        if not thread:
-            rec.pop("thread_id", None)
-            rec.pop("thread_on", None)
-        rec["seen_comments"] = sorted(seen | {c["id"] for c in comments if c.get("id")})
-        rec["last_comment_at"] = max([c.get("created_at", "") for c in comments] + [rec.get("last_comment_at", "")])
-        client = self.client(term)
+        relaunch = rec.get("relaunch") or {}
+        attempt = relaunch.get("attempt", 1) if relaunch.get("run_id") == run_id else 1
+        dir_name = run_id if attempt == 1 else "%s-r%d" % (run_id, attempt)
+        run = Run(run_id, tcfg["repo"], prompt, route["model"], route["effort"], route["tier"], thread, route["script"],
+                  dir_name=dir_name)
         if thread:
             mode = "续线程 %s" % thread[:12]
         elif reason:
@@ -573,10 +660,53 @@ class Agentd:
             mode = "新线程（附交接摘要）" if handoff else "新线程"
         if is_retry:
             mode += "，重试 %d" % rec.get("retries", 0)
-        detail = " / ".join(x for x in (exname, route["model"], route["effort"], route["tier"]) if x)
+        if attempt > 1:
+            mode += "，重新启动（目录 %s）" % dir_name
+        # Everything the bookkeeping needs once the run is known to exist; saved before the launch call so that a
+        # launch whose outcome is unknown (relay timeout, crash) is probed and adopted instead of started twice.
+        plan = {"run_id": run_id, "terminal": term, "executor": exname, "attempt": attempt, "at": self.now(),
+                "handle": ex.handle(run), "round": k, "budget_s": route["timeout_hours"] * 3600,
+                "resumed": bool(thread), "group": route["group"], "model": route["model"], "effort": route["effort"],
+                "is_retry": is_retry,
+                # transient retries continue the same logical round: they do not use up max_rounds_per_thread
+                "thread_rounds": (rec.get("thread_rounds", 0) + (0 if is_retry else 1)) if thread else 1,
+                "seen": sorted(seen | {c["id"] for c in comments if c.get("id")}),
+                "last_comment_at": max([c.get("created_at", "") for c in comments] + [rec.get("last_comment_at", "")]),
+                "mode": mode,
+                "detail": " / ".join(x for x in (exname, route["model"], route["effort"], route["tier"]) if x)}
+        rec["launching"] = plan
+        self.store.save(data)
+        handle = ex.start(run)   # raises -> the marker stays and the next tick probes the run directory
+        self.launched(ctx, iid, rec, plan, handle)
+
+    def launched(self, ctx, iid, rec, plan, handle, adopted=False):
+        """The run exists on the terminal: record it as in flight, then move the item and post the start note."""
+        key, term = ctx.key(rec["seq"]), plan["terminal"]
+        rec.pop("launching", None)
+        rec.pop("relaunch", None)
+        rec.pop("retry", None)
+        rec.pop("rebuild", None)
+        if not plan["is_retry"]:
+            rec["retries"] = 0
+        rec.update(runs=plan["round"], thread_rounds=plan["thread_rounds"], inflight={
+            "run_id": plan["run_id"], "terminal": term, "executor": plan["executor"], "handle": handle,
+            "round": plan["round"], "started": plan["at"], "deadline": plan["at"] + plan["budget_s"],
+            "resumed": plan["resumed"], "group": plan["group"], "model": plan["model"], "effort": plan["effort"]})
+        if not plan["resumed"]:
+            rec.pop("thread_id", None)
+            rec.pop("thread_on", None)
+        rec["seen_comments"] = plan["seen"]
+        rec["last_comment_at"] = plan["last_comment_at"]
+        mode = plan["mode"] + ("，启动未确认、已接管" if adopted else "")
+        issue = ctx.issues.get(iid)
+        client = self.client(term)
         try:
-            self.move(ctx, client, iid, "running", tcfg["plane_user"])
-            self.post(ctx, client, iid, rec, "%s 第 %d 轮开始：%s @ %s，%s，run=%s。" % (WORKER, k, detail, term, mode, run_id))
+            # an adopted run whose item a person moved meanwhile is left alone: poll cancels it next tick
+            if issue is not None and issue.get("state") in (ctx.state_id("todo"), ctx.state_id("running")):
+                self.move(ctx, client, iid, "running", self.cfg.terminals[term]["plane_user"])
+            self.post(ctx, client, iid, rec, "%s 第 %d 轮开始：%s @ %s，%s，run=%s。" % (
+                WORKER, plan["round"], plan["detail"], term, mode, plan["run_id"]))
         except Exception as exc:
             self.log("%s bookkeeping after start failed: %r" % (key, exc))
-        self.log("%s run %s started on %s (%s, %s)" % (key, run_id, term, detail, mode))
+        self.log("%s run %s %s on %s (%s, %s)" % (key, plan["run_id"], "adopted" if adopted else "started", term,
+                                                  plan["detail"], mode))

agentd/executors/base.py

@@ -6,6 +6,8 @@ import re
 from dataclasses import dataclass, field
 from typing import List, Optional
 
+from ..transports import fetch_each
+
 
 @dataclass
 class Run:
@@ -17,6 +19,7 @@ class Run:
     tier: Optional[str] = None
     thread_id: Optional[str] = None   # set => resume this thread
     script: Optional[str] = None      # shell executor
+    dir_name: Optional[str] = None    # run directory / task name; default run_id (`<run_id>-r2` on a relaunch)
 
 
 @dataclass
@@ -35,9 +38,15 @@ class Executor:
     def __init__(self, name, cfg, terminal, tcfg, transport):
         self.name, self.cfg, self.terminal, self.tcfg, self.transport = name, cfg, terminal, tcfg, transport
 
+    def handle(self, run):           # -> the handle `start(run)` will return, computed without touching the terminal
+        raise NotImplementedError
+
     def start(self, run):            # -> handle (dict)
         raise NotImplementedError
 
+    def probe(self, handle):         # -> names of launch evidence present on the terminal ([] = never launched)
+        raise NotImplementedError
+
     def poll(self, handle):          # -> "running" | "finished"
         raise NotImplementedError
 
@@ -48,6 +57,11 @@ class Executor:
         raise NotImplementedError
 
 
+def parse_exit(text):
+    m = re.search(r"exit=(-?\d+)", text or "")
+    return int(m.group(1)) if m else None
+
+
 def fill(template, params):
     """Replace {name} placeholders that exist in params; leave anything else untouched."""
     return re.sub(r"\{(\w+)\}", lambda m: str(params[m.group(1)]) if m.group(1) in params else m.group(0), template)
@@ -84,26 +98,44 @@ class RunnerExecutor(Executor):
             program = "& " + program
         return (program + " " + rest).strip()
 
+    def handle(self, run):
+        name = run.dir_name or run.run_id
+        return {"run_dir": self.transport.join(self.run_root(), name), "task": "agentd-" + name, "thread_id": run.thread_id}
+
     def start(self, run):
         t = self.transport
-        run_dir = t.join(self.run_root(), run.run_id)
-        t.mkdir(run_dir, clean=True)
-        t.write_text(t.join(run_dir, "prompt.md"), run.prompt)
+        h = self.handle(run)
+        run_dir, task = h["run_dir"], h["task"]
         script_name, script = t.runner(run_dir, run.repo, self.command(run, run_dir), self.stdout_name)
-        script_path = t.join(run_dir, script_name)
-        t.write_text(script_path, script)
-        task = "agentd-" + run.run_id
-        t.launch(task, script_path, run_dir)
-        return {"run_dir": run_dir, "task": task, "thread_id": run.thread_id}
+        files = {"prompt.md": run.prompt, script_name: script}
+        if hasattr(t, "start_run"):  # remote: stage + launch in one call (direct-ssh; POSIX also on the relay)
+            t.start_run(task, run_dir, files, script_name)
+        else:
+            t.mkdir(run_dir, clean=True)
+            for name, text in files.items():
+                t.write_text(t.join(run_dir, name), text)
+            t.launch(task, t.join(run_dir, script_name), run_dir)
+        return h
+
+    PROBE_FILES = ("exit.txt", "launch.pid", "pid.txt")
+
+    def probe(self, handle):
+        """Launch evidence in the run directory (one transport call; raises TransportError when unreachable)."""
+        return self.transport.probe(handle["run_dir"], self.PROBE_FILES)
 
     def poll(self, handle):
         text = self.transport.read_text(self.transport.join(handle["run_dir"], "exit.txt"))
         return "finished" if text and "exit=" in text else "running"
 
     def exit_code(self, handle):
-        text = self.transport.read_text(self.transport.join(handle["run_dir"], "exit.txt")) or ""
-        m = re.search(r"exit=(-?\d+)", text)
-        return int(m.group(1)) if m else None
+        return parse_exit(self.transport.read_text(self.transport.join(handle["run_dir"], "exit.txt")))
+
+    def fetch(self, handle, ops):
+        """Reads of the run directory for `collect`: one call on remote transports (see transports.fetch)."""
+        t = self.transport
+        if hasattr(t, "fetch"):
+            return t.fetch(handle["run_dir"], ops)
+        return fetch_each(t, handle["run_dir"], ops)
 
     def kill(self, handle):
         self.transport.kill(handle["task"], handle["run_dir"])

agentd/executors/claude_code.py

@@ -2,7 +2,7 @@
 resume with `--resume <session_id>`. The JSON result carries `result` (last message) and `session_id`."""
 import json
 
-from .base import ExecResult, RunnerExecutor
+from .base import ExecResult, RunnerExecutor, parse_exit
 
 
 def parse_output(text):
@@ -26,13 +26,14 @@ def parse_output(text):
 class ClaudeCodeExecutor(RunnerExecutor):
     stdout_name = "out.json"
 
+    COLLECT = [("read", "exit.txt"), ("read", "out.json"), ("tail", "stderr.txt", 5)]
+
     def collect(self, handle):
-        t, d = self.transport, handle["run_dir"]
-        obj = parse_output(t.read_text(t.join(d, "out.json")))
+        exit_text, out, tail = self.fetch(handle, self.COLLECT)   # one remote call
+        obj = parse_output(out)
         err = bool(obj.get("is_error")) or obj.get("subtype", "success") != "success"
         last = "" if err else (obj.get("result") or "")
-        tail = self.stderr_tail(handle)
         if err:
             tail = ("%s %s" % (obj.get("subtype", ""), str(obj.get("result", ""))[:400]) + "\n" + tail).strip()
-        return ExecResult(last.strip(), self.exit_code(handle), obj.get("session_id") or handle.get("thread_id"),
+        return ExecResult(last.strip(), parse_exit(exit_text), obj.get("session_id") or handle.get("thread_id"),
                           self.logs(handle, "out.json", "stderr.txt"), tail)

agentd/executors/codex.py

@@ -2,22 +2,22 @@
 Thread id comes from the `thread.started` event in events.jsonl; the last message from last.md."""
 import json
 
-from .base import ExecResult, RunnerExecutor
+from .base import ExecResult, RunnerExecutor, parse_exit
 
 
 class CodexExecutor(RunnerExecutor):
     stdout_name = "events.jsonl"
+    COLLECT = [("read", "exit.txt"), ("read", "last.md"), ("grep", "events.jsonl", "thread.started"),
+               ("tail", "stderr.txt", 5), ("tail", "events.jsonl", 1)]
 
     def collect(self, handle):
-        t, d = self.transport, handle["run_dir"]
-        last = t.read_text(t.join(d, "last.md")) or ""
+        exit_text, last, line, err_tail, last_event = self.fetch(handle, self.COLLECT)   # one remote call
         thread = handle.get("thread_id")
-        line = t.grep_first(t.join(d, "events.jsonl"), "thread.started")
         if line:
             try:
                 thread = json.loads(line).get("thread_id") or thread
             except ValueError:
                 pass
-        tail = (self.stderr_tail(handle) + "\n" + t.tail(t.join(d, "events.jsonl"), 1)[-600:]).strip()
-        return ExecResult(last.strip(), self.exit_code(handle), thread,
+        tail = (err_tail + "\n" + last_event[-600:]).strip()
+        return ExecResult((last or "").strip(), parse_exit(exit_text), thread,
                           self.logs(handle, "events.jsonl", "stderr.txt", "last.md"), tail)

agentd/executors/shell.py

@@ -1,6 +1,6 @@
 """Shell/script executor: runs `{script}` with the prompt on stdin. The last message is `last.md` in the run
 directory if the script wrote it (its path is passed as {out}), else the tail of stdout. No threads."""
-from .base import ExecResult, RunnerExecutor
+from .base import ExecResult, RunnerExecutor, parse_exit
 
 
 class ShellExecutor(RunnerExecutor):
@@ -10,10 +10,11 @@ class ShellExecutor(RunnerExecutor):
     def resume(self):
         return False
 
+    COLLECT = [("read", "exit.txt"), ("read", "last.md"), ("read", "stdout.txt"), ("tail", "stderr.txt", 5)]
+
     def collect(self, handle):
-        t, d = self.transport, handle["run_dir"]
-        last = t.read_text(t.join(d, "last.md"))
+        exit_text, last, stdout, tail = self.fetch(handle, self.COLLECT)   # one remote call
         if last is None:
-            last = (t.read_text(t.join(d, "stdout.txt")) or "")[-6000:]
-        return ExecResult(last.strip(), self.exit_code(handle), None,
-                          self.logs(handle, "stdout.txt", "stderr.txt"), self.stderr_tail(handle))
+            last = (stdout or "")[-6000:]
+        return ExecResult(last.strip(), parse_exit(exit_text), None,
+                          self.logs(handle, "stdout.txt", "stderr.txt"), tail)

agentd/store.py

@@ -17,8 +17,9 @@ class Busy(Exception):
 
 
 class Store:
-    def __init__(self, state_dir, echo=True):
+    def __init__(self, state_dir, echo=True, log_dir=None):
         self.dir = Path(state_dir)
+        self.log_dir = Path(log_dir) if log_dir else self.dir
         self.echo = echo
         self.path = self.dir / "agentd-state.json"
 
@@ -59,9 +60,9 @@ class Store:
         self._atomic_write(self.dir / "heartbeat.json", info)
 
     def log(self, msg):
-        self.dir.mkdir(parents=True, exist_ok=True)
+        self.log_dir.mkdir(parents=True, exist_ok=True)
         line = time.strftime("%Y-%m-%dT%H:%M:%S") + " " + msg
         if self.echo:
             print(line, flush=True)
-        with open(self.dir / "agentd.log", "a", encoding="utf-8") as fh:
+        with open(self.log_dir / "agentd.log", "a", encoding="utf-8") as fh:
             fh.write(line + "\n")

agentd/transports/__init__.py

@@ -4,21 +4,162 @@ Interface (duck-typed, see docs/executor-contract.md):
   os, join(*parts), quote(s), default_run_root()
   mkdir(path, clean=False), write_text(path, text), read_text(path) -> str|None,
   grep_first(path, needle) -> str|None, tail(path, n) -> str
+  fetch(run_dir, ops) -> [result per op]   (ops: ("read", name) | ("grep", name, needle) | ("tail", name, n))
   runner(run_dir, cwd, command, stdout_name) -> (script_name, script_text)
-  launch(task, script_path, run_dir), kill(task, run_dir), forget(task)
-  process_count(name) -> int|None (None = unreachable)"""
+  start_run(task, run_dir, files, script_name) (remote only), launch(task, script_path, run_dir), kill(task, run_dir),
+  forget(task), process_count(name) -> int|None (None = unreachable)
+
+Remote terminals are reached through a channel (design §9): `warpgate-ssh` (agent-access.py run through the
+Warpgate gateway) or `direct-ssh` (OpenSSH over Tailscale with a dedicated key, optional warpgate fallback)."""
+import base64
+import functools
+import re
+import time
 
 
 class TransportError(Exception):
     pass
 
 
-def make_transport(cfg, terminal):
-    kind = cfg.terminals[terminal]["transport"]
+class Unreached(TransportError):
+    """The remote command certainly did not run (connect / auth / agent failure, direct path unavailable): the
+    only kind of failure a direct-ssh terminal may repeat on its fallback channel."""
+
+
+class HostKeyError(TransportError):
+    """The terminal presented a host key other than the pinned one: a security error, never retried elsewhere."""
+
+
+class CallLog:
+    """One line per remote call -> `path` (remote.log); notable events (fallback, host key) also -> `event`."""
+
+    def __init__(self, path=None, event=None, echo=False):
+        self.path, self.event_fn, self.echo = path, event, echo
+        self._once = set()
+
+    def call(self, terminal, via, op, ms, rc, note=""):
+        line = "term=%s via=%s op=%s ms=%d rc=%s%s" % (terminal, via, op or "-", ms, rc, " " + note if note else "")
+        if self.echo:
+            print(line, flush=True)
+        if self.path:
+            try:
+                self.path.parent.mkdir(parents=True, exist_ok=True)
+                with open(self.path, "a", encoding="utf-8") as fh:
+                    fh.write(time.strftime("%Y-%m-%dT%H:%M:%S") + " " + line + "\n")
+            except OSError:
+                pass
+
+    def event(self, msg, once=None):
+        """`once` = a key: the message is logged only the first time per process."""
+        if once is not None:
+            if once in self._once:
+                return
+            self._once.add(once)
+        if self.event_fn:
+            self.event_fn(msg)
+
+
+NULL_LOG = CallLog()
+
+
+def operation(name):
+    """Method decorator: names the remote operation for the call log (innermost wins, restored on return)."""
+    def deco(fn):
+        @functools.wraps(fn)
+        def wrapper(self, *args, **kwargs):
+            prev, self._op = getattr(self, "_op", None), name
+            try:
+                return fn(self, *args, **kwargs)
+            finally:
+                self._op = prev
+        return wrapper
+    return deco
+
+
+def fetch_each(t, run_dir, ops):
+    """`fetch` built from the single-file primitives (one call per op): local transport and fallbacks."""
+    out = []
+    for o in ops:
+        path = t.join(run_dir, o[1])
+        if o[0] == "read":
+            out.append(t.read_text(path))
+        elif o[0] == "grep":
+            out.append(t.grep_first(path, o[2]))
+        elif o[0] == "tail":
+            out.append(t.tail(path, o[2]))
+        else:
+            raise ValueError("unknown fetch op %r" % (o,))
+    return out
+
+
+def parse_fetch(out, ops, what=""):
+    """Decode the `F<i>=<base64>` / `M<i>` (missing) lines of a remote fetch script into results per op:
+    read -> text|None, grep -> line|None, tail -> text ("" when missing)."""
+    res = []
+    for i, o in enumerate(ops):
+        m = re.search(r"^F%d=([A-Za-z0-9+/=]*)$" % i, out, re.M)
+        if m:
+            text = base64.b64decode(m.group(1)).decode("utf-8", "replace") if m.group(1) else ""
+        elif re.search(r"^M%d$" % i, out, re.M):
+            text = None
+        else:
+            raise TransportError("fetch %s: unexpected output: %s" % (what, out.strip()[-200:]))
+        if o[0] == "grep":
+            text = (text or "").rstrip("\n") or None
+        elif o[0] == "tail":
+            text = (text or "").rstrip("\n")
+        res.append(text)
+    return res
+
+
+class RoutedTransport:
+    """direct-ssh terminal: every operation runs on the primary (direct) transport; when it fails before the remote
+    command ran (`Unreached`) the operation is repeated on the fallback (warpgate-ssh) transport, which then serves
+    the rest of this transport's life (one tick). Each direct operation is a single remote call, so a repeat never
+    doubles a side effect. Host-key failures and failures after the command may have run propagate unchanged."""
+    PURE = ("join", "quote", "default_run_root", "runner")
+
+    def __init__(self, terminal, primary, fallback=None, log=None):
+        self.terminal, self.primary, self.fallback = terminal, primary, fallback
+        self.log = log or NULL_LOG
+        self.on_fallback = False
+
+    def __getattr__(self, name):
+        attr = getattr(self.primary, name)
+        if name in self.PURE or not callable(attr):
+            return attr
+
+        def routed(*args, **kwargs):
+            if self.on_fallback:
+                return getattr(self.fallback, name)(*args, **kwargs)
+            try:
+                return attr(*args, **kwargs)
+            except Unreached as exc:
+                if self.fallback is None:
+                    raise
+                self.on_fallback = True
+                self.log.event("%s: direct-ssh failed before the command ran (%s); using the warpgate-ssh fallback "
+                               "for the rest of this tick" % (self.terminal, exc))
+                return getattr(self.fallback, name)(*args, **kwargs)
+        return routed
+
+
+def make_transport(cfg, terminal, log=None, force=None):
+    """`force` = direct-ssh | warpgate-ssh: that channel only, no fallback (agentd ping)."""
+    t = cfg.terminals[terminal]
+    kind = force or t["transport"]
     if kind == "local":
         from .local import LocalTransport
         return LocalTransport(cfg, terminal)
     if kind == "warpgate-ssh":
         from .warpgate_ssh import WarpgateSshTransport
-        return WarpgateSshTransport(cfg, terminal)
+        return WarpgateSshTransport(cfg, terminal, log=log)
+    if kind == "direct-ssh":
+        from .direct_ssh import DirectChannel
+        from .warpgate_ssh import WarpgateSshTransport
+        primary = WarpgateSshTransport(cfg, terminal, channel=DirectChannel(cfg, terminal, log))
+        fallback = None
+        if not force and t.get("fallback") == "warpgate-ssh":
+            fallback = WarpgateSshTransport(cfg, terminal, log=log, label="fallback")
+        return RoutedTransport(terminal, primary, fallback, log)
     raise TransportError("unknown transport %s" % kind)

agentd/transports/direct_ssh.py

@@ -0,0 +1,313 @@
+"""direct-ssh channel (PAD-12): OpenSSH from the dispatcher straight to the terminal over Tailscale, no Warpgate.
+
+- Auth: one dedicated agentd key. Its private text arrives in the env var named by `agentd.ssh_key` (default
+  AGENTD_SSH_KEY, injected by `agent-access.py exec --secret-env`), is popped from os.environ (children never see
+  it) and loaded with `ssh-add -` into a per-process ssh-agent on a socket in a private 0700 temp dir. The key is
+  never written to disk, argv, logs or exceptions. No key / agent failure -> the channel is unavailable
+  (`Unreached`), so a terminal with `fallback: warpgate-ssh` uses Warpgate instead.
+- Host key: pinned per terminal (`ssh.host_key`); a known_hosts file is generated from it (public data) and
+  checked strictly. A mismatch raises HostKeyError: never a fallback.
+- Reuse: ControlMaster=auto / ControlPersist=60 under `agentd.ssh_dir` (root-only 0700), so the second call of a
+  tick (and the next tick within a minute) skips the handshake. The socket name carries a fingerprint of the pinned
+  host key: a changed pin never rides an open master that was verified against the old one.
+- Payload on stdin: Windows runs a fixed PowerShell bootstrap (-EncodedCommand) that reads the base64 of the real
+  script from stdin; POSIX runs a fixed bash loader that reads the script from stdin. One call per operation, no
+  argv size limit."""
+import atexit
+import base64
+import hashlib
+import os
+import re
+import shutil
+import signal
+import subprocess
+import tempfile
+import time
+
+from . import NULL_LOG, HostKeyError, TransportError, Unreached
+from .local import sh_quote
+
+CONNECT_TIMEOUT = 8
+CONTROL_PERSIST = 60
+SOCKET_MAX = 104          # sun_path; ssh adds ".<16 random chars>" while it sets up the master socket
+CTL_NAME = 1 + 9 + 40 + 17   # "/<pin8>-<%C: 40 hex>" + that suffix
+
+# Fixed PowerShell bootstrap: the operation's script arrives on stdin as base64 of its UTF-8 text (ASCII on the
+# wire, so the console input code page does not matter).
+BOOTSTRAP = ("$ProgressPreference='SilentlyContinue'\n"
+             "$ErrorActionPreference='Continue'\n"
+             "[Console]::OutputEncoding=[Text.Encoding]::UTF8\n"
+             "$s=[Text.Encoding]::UTF8.GetString([Convert]::FromBase64String([Console]::In.ReadToEnd().Trim()))\n"
+             "& ([ScriptBlock]::Create($s))\n")
+
+
+def encoded_command(script):
+    """PowerShell -EncodedCommand value: base64 of the UTF-16LE text."""
+    return base64.b64encode(script.encode("utf-16-le")).decode("ascii")
+
+
+PS_COMMAND = "powershell.exe -NoProfile -NonInteractive -EncodedCommand " + encoded_command(BOOTSTRAP)
+# POSIX loader: the script is read from stdin into a variable (no MAX_ARG_STRLEN), then runs in a subshell with
+# stdin = /dev/null. Its exit 255 is reported as 254: 255 stays ssh's own "no session" code. (The PowerShell
+# bootstrap exits 0/1 by $?; none of our PowerShell scripts calls `exit`.)
+SH_COMMAND = "bash -c " + sh_quote('exec 3<&0 </dev/null; s=$(cat <&3); exec 3<&-; (eval "$s"); r=$?; '
+                                   '[ $r = 255 ] && r=254; exit $r')
+
+
+def ps_payload(script):
+    return (base64.b64encode(script.encode("utf-8")).decode("ascii") + "\n").encode("ascii")
+
+
+# ---------------------------------------------------------------- failure classification
+# A host offering no ssh-ed25519 key at all counts as a host-key failure too (a downgrade is a security signal).
+HOSTKEY = re.compile(r"Host key verification failed|REMOTE HOST IDENTIFICATION HAS CHANGED|"
+                     r"No \S+ host key is known for|host key for \S+ has changed|no matching host key type", re.I)
+# ssh exit 255 with one of these = no remote session ran the command. Only ssh's own messages, anchored at the start
+# of a line (stderr also carries the remote command's stderr).
+UNREACHED = re.compile(r"^(?:ssh: connect to host |ssh: Could not resolve hostname |banner exchange: |"
+                       r"Connection timed out during banner exchange|kex_exchange_identification: |"
+                       r"Connection (?:closed|reset) by \S+ port \d+\s*$|"
+                       r"\S+: Permission denied \((?:publickey|password|keyboard-interactive|hostbased|gssapi)|"
+                       r"Received disconnect from .*Too many authentication failures|"
+                       r"sign_and_send_pubkey: |error fetching identities|Error connecting to agent|"
+                       r"Unable to negotiate with |Session open refused by peer)", re.I | re.M)
+
+
+def classify(rc, err, out=""):
+    """-> 'ok' (the remote command ran, whatever its exit code), 'hostkey', 'unreached' (certainly not run) or
+    'unknown' (ssh exit 255 without a recognisable pre-session error: the command may have run). Any stdout proves
+    that a session ran the command, so it is never 'unreached'."""
+    if rc != 255:
+        return "ok"
+    if (out or "").strip():
+        return "unknown"
+    if HOSTKEY.search(err or ""):
+        return "hostkey"
+    if UNREACHED.search(err or ""):
+        return "unreached"
+    return "unknown"
+
+
+# ---------------------------------------------------------------- per-process key agent
+def _die_with_parent():
+    """preexec: the agent gets SIGTERM when agentd dies, even by SIGKILL (Linux PR_SET_PDEATHSIG)."""
+    try:
+        import ctypes
+        ctypes.CDLL("libc.so.6", use_errno=True).prctl(1, signal.SIGTERM)
+    except Exception:
+        pass
+
+
+def normalise_key(text):
+    """Vault may hand the key with literal `\\n` escapes or without the final newline."""
+    if "\n" not in text.strip() and "\\n" in text:
+        text = text.replace("\\n", "\n")
+    return text.strip() + "\n"
+
+
+_KEYS = {}   # env var name -> key text, taken out of os.environ at CLI start until the agent loads it
+
+
+def take_key(env_name):
+    """Move the key out of os.environ right away (called by the CLI before anything else runs), so no child -- not
+    even one started before the first direct call -- can inherit it."""
+    if env_name not in _KEYS:
+        _KEYS[env_name] = os.environ.pop(env_name, None)
+    else:
+        os.environ.pop(env_name, None)
+
+
+def key_present(env_name):
+    return bool((_KEYS.get(env_name) or os.environ.get(env_name) or "").strip())
+
+
+class KeyAgent:
+    """ssh-agent of this process holding the agentd key. `error` set = direct-ssh unavailable here."""
+
+    def __init__(self, env_name):
+        self.env_name = env_name
+        self.dir = self.sock = self.pub = self.proc = None
+        self.error = None
+
+    def start(self):
+        key = os.environ.pop(self.env_name, None)   # popped even if loading fails: children never inherit it
+        key = _KEYS.pop(self.env_name, None) or key
+        if not key or not key.strip():
+            self.error = "%s is not set (agent-access.py exec --secret-env %s=<vault key>)" % (self.env_name, self.env_name)
+            return self
+        agent_bin, add_bin = shutil.which("ssh-agent"), shutil.which("ssh-add")
+        if not (agent_bin and add_bin and shutil.which("ssh")):
+            self.error = "ssh / ssh-agent / ssh-add not installed"
+            return self
+        env = {"PATH": os.environ.get("PATH", "/usr/bin:/bin"), "LANG": "C"}
+        try:
+            self.dir = tempfile.mkdtemp(prefix="agentd-agent-")      # 0700
+            self.sock = os.path.join(self.dir, "agent.sock")
+            atexit.register(self.close)
+            self.proc = subprocess.Popen([agent_bin, "-D", "-a", self.sock], stdin=subprocess.DEVNULL,
+                                         stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, env=env,
+                                         close_fds=True, preexec_fn=_die_with_parent)
+            end = time.monotonic() + 5
+            while not os.path.exists(self.sock):
+                if self.proc.poll() is not None or time.monotonic() > end:
+                    self.error = "ssh-agent did not start"
+                    return self
+                time.sleep(0.02)
+            env["SSH_AUTH_SOCK"] = self.sock
+            add = subprocess.run([add_bin, "-q", "-"], input=normalise_key(key).encode("utf-8"), env=env,
+                                 stdout=subprocess.DEVNULL, stderr=subprocess.PIPE, timeout=15)
+            key = None
+            if add.returncode != 0:
+                why = add.stderr.decode("utf-8", "replace").strip().splitlines()[:1]
+                why = why[0][:120] if why and "-----" not in why[0] else ""
+                self.error = "ssh-add failed (rc %d) %s" % (add.returncode, why)
+                return self
+            listed = subprocess.run([add_bin, "-L"], env=env, capture_output=True, timeout=15)
+            pub = next((l for l in listed.stdout.decode("ascii", "replace").splitlines() if l.startswith("ssh-")), None)
+            if not pub:
+                self.error = "ssh-agent holds no key after ssh-add"
+                return self
+            self.pub = os.path.join(self.dir, "agentd.pub")   # public half: selects the agent key (IdentitiesOnly)
+            with open(self.pub, "w", encoding="ascii") as fh:
+                fh.write(pub + "\n")
+        except (OSError, subprocess.SubprocessError) as exc:
+            self.error = "ssh-agent setup failed: %s" % type(exc).__name__
+        finally:
+            key = None
+            if self.error:
+                self.close()
+        return self
+
+    def close(self):
+        if self.proc and self.proc.poll() is None:
+            self.proc.terminate()
+            try:
+                self.proc.wait(3)
+            except subprocess.TimeoutExpired:
+                self.proc.kill()
+        if self.dir:
+            shutil.rmtree(self.dir, ignore_errors=True)
+
+
+_AGENTS = {}
+
+
+def key_agent(cfg, log=NULL_LOG):
+    """The process-wide agent for cfg's key env var, started on first use; raises Unreached when unavailable."""
+    name = cfg.ssh_key_env
+    agent = _AGENTS.get(name)
+    if agent is None:
+        agent = _AGENTS[name] = KeyAgent(name).start()
+    if agent.error:
+        log.event("direct-ssh unavailable in this process: %s" % agent.error, once="agent:" + name)
+        raise Unreached("direct-ssh unavailable: %s" % agent.error)
+    return agent
+
+
+def control_dir(cfg):
+    """agentd.ssh_dir: ControlMaster sockets + generated known_hosts; created 0700, must be ours."""
+    d = cfg.ssh_dir
+    try:
+        d.mkdir(parents=True, exist_ok=True)
+        os.chmod(d, 0o700)
+        if os.stat(d).st_uid != os.geteuid():
+            raise OSError("not owned by this user")
+    except OSError as exc:
+        raise Unreached("ssh_dir %s unusable: %s" % (d, exc)) from None
+    return d
+
+
+# ---------------------------------------------------------------- channel
+class DirectChannel:
+    """One terminal's direct OpenSSH path. ps() / sh() mirror the warpgate channel; payloads go on stdin."""
+    stdin = True
+
+    def __init__(self, cfg, terminal, log=None, label="direct"):
+        self.cfg, self.terminal, self.log, self.via = cfg, terminal, log or NULL_LOG, label
+        s = cfg.terminals[terminal]["ssh"]
+        self.host, self.user, self.port = s["host"], s["user"], int(s.get("port", 22))
+        self.host_key = " ".join(s["host_key"].split()[:2])
+        self.pin = hashlib.sha256(self.host_key.encode("ascii")).hexdigest()[:8]
+        self.alias = "agentd-" + terminal
+        self.hostkey_failed = None
+        self.name = "ssh %s@%s" % (self.user, self.host)
+
+    def known_hosts(self, d):
+        path = d / ("known_hosts." + self.terminal)
+        line = "%s %s\n" % (self.alias, self.host_key)
+        try:
+            if not path.exists() or path.read_text(encoding="ascii") != line:
+                tmp = path.with_name("%s.%d.tmp" % (path.name, os.getpid()))   # per terminal and process
+                tmp.write_text(line, encoding="ascii")
+                os.replace(tmp, path)
+        except OSError as exc:
+            raise Unreached("known_hosts %s: %s" % (path, exc)) from None
+        return path
+
+    def argv(self, agent, remote_command):
+        d = control_dir(self.cfg)
+        ctl = str(d / (self.pin + "-%C"))    # pin fingerprint: a new pin never reuses a master verified on the old one
+        if len(str(d)) + CTL_NAME > SOCKET_MAX:
+            self.log.event("agentd.ssh_dir %s is too long for a control socket; connection reuse off" % d, once="ctl")
+            mux = ["-o", "ControlMaster=no", "-o", "ControlPath=none"]
+        else:
+            mux = ["-o", "ControlMaster=auto", "-o", "ControlPersist=%d" % CONTROL_PERSIST, "-o", "ControlPath=" + ctl]
+        return (["ssh", "-F", "/dev/null", "-T", "-p", str(self.port), "-l", self.user,
+                 "-o", "BatchMode=yes", "-o", "IdentitiesOnly=yes",
+                 "-o", "IdentityAgent=" + agent.sock, "-o", "IdentityFile=" + agent.pub,
+                 "-o", "PreferredAuthentications=publickey", "-o", "PasswordAuthentication=no",
+                 "-o", "KbdInteractiveAuthentication=no",
+                 "-o", "StrictHostKeyChecking=yes", "-o", "UserKnownHostsFile=" + str(self.known_hosts(d)),
+                 "-o", "GlobalKnownHostsFile=/dev/null", "-o", "HostKeyAlias=" + self.alias,
+                 "-o", "HostKeyAlgorithms=" + self.host_key.split()[0], "-o", "UpdateHostKeys=no",
+                 "-o", "ConnectTimeout=%d" % CONNECT_TIMEOUT, "-o", "ServerAliveInterval=15",
+                 "-o", "ServerAliveCountMax=3", "-o", "ForwardAgent=no", "-o", "ForwardX11=no",
+                 "-o", "ClearAllForwardings=yes", "-o", "LogLevel=ERROR", "-o", "EscapeChar=none"]
+                + mux + ["--", self.host, remote_command])
+
+    def run(self, remote_command, payload, timeout, op):
+        """-> (rc, stdout, stderr) of a call whose command ran; raises HostKeyError / Unreached / TransportError."""
+        if self.hostkey_failed:
+            raise HostKeyError(self.hostkey_failed)
+        agent = key_agent(self.cfg, self.log)
+        argv = self.argv(agent, remote_command)
+        env = self.cfg.scrubbed_env()
+        env.pop("SSH_AUTH_SOCK", None)
+        t0 = time.monotonic()
+        ms = lambda: (time.monotonic() - t0) * 1000  # noqa: E731
+        # output goes to files, not pipes: a ControlPersist master forked by this call must never hold our pipes open
+        with tempfile.TemporaryFile() as fo, tempfile.TemporaryFile() as fe:
+            try:
+                p = subprocess.run(argv, input=payload, stdout=fo, stderr=fe, env=env, timeout=timeout)
+            except subprocess.TimeoutExpired:
+                self.log.call(self.terminal, self.via, op, ms(), "timeout")
+                raise TransportError("%s timeout after %ss" % (self.name, timeout)) from None   # may have run
+            except OSError as exc:
+                self.log.call(self.terminal, self.via, op, ms(), "launch-failed")
+                raise Unreached("%s: ssh launch failed: %s" % (self.name, type(exc).__name__)) from None
+            fo.seek(0)
+            fe.seek(0)
+            dec = lambda b: b.decode("utf-8", "replace").replace("\r", "")  # noqa: E731
+            out, err = dec(fo.read()), dec(fe.read())
+        kind = classify(p.returncode, err, out)
+        self.log.call(self.terminal, self.via, op, ms(), p.returncode, "" if kind == "ok" else kind)
+        if kind == "hostkey":
+            self.hostkey_failed = ("%s:%d host key verification FAILED (pinned %s...; other key or no such key type "
+                                   "offered): not falling back; check terminals.%s.ssh.host_key out of band"
+                                   % (self.host, self.port, self.host_key[:40], self.terminal))
+            self.log.event("SECURITY %s: %s" % (self.terminal, self.hostkey_failed), once="hostkey:" + self.terminal)
+            raise HostKeyError(self.hostkey_failed)
+        if kind == "unreached":
+            raise Unreached("%s exit 255: %s" % (self.name, err.strip()[-300:]))
+        if kind == "unknown":
+            raise TransportError("%s exit 255: %s" % (self.name, err.strip()[-300:] or "(no message)"))
+        return p.returncode, out, err
+
+    def ps(self, script, timeout, op=None):
+        rc, out, err = self.run(PS_COMMAND, ps_payload(script), timeout, op)
+        if rc != 0 and not out.strip():
+            raise TransportError("%s exit %s: %s" % (self.name, rc, err.strip()[-300:]))
+        return out
+
+    def sh(self, script, timeout, op=None):
+        return self.run(SH_COMMAND, script.encode("utf-8"), timeout, op)

agentd/transports/local.py

@@ -4,17 +4,27 @@ import signal
 import subprocess
 from pathlib import Path
 
+from . import TransportError
+
+PID_FILES = ("pid.txt", "launch.pid")   # a run directory holding one of these is never cleaned
+
+# Shared with the remote POSIX transport (warpgate_ssh): plain sh + `date +FMT`, so it also runs on macOS.
 RUNNER = """#!/usr/bin/env bash
-cd {cwd} || {{ echo "exit=97 at $(date -Is) (cd failed)" > {d}/exit.txt; exit 97; }}
+cd {cwd} || {{ echo "exit=97 at $(date +%Y-%m-%dT%H:%M:%S%z) (cd failed)" > {d}/exit.txt; exit 97; }}
 echo $$ > {d}/pid.txt
-date -Is > {d}/started.txt
+date +%Y-%m-%dT%H:%M:%S%z > {d}/started.txt
 (
 {command}
 ) < {d}/prompt.md > {d}/{stdout} 2> {d}/stderr.txt
-echo "exit=$? at $(date -Is)" > {d}/exit.txt
+echo "exit=$? at $(date +%Y-%m-%dT%H:%M:%S%z)" > {d}/exit.txt
 """
 
 
+def sh_quote(s):
+    """POSIX shell single-quote literal (always quoted, `'` -> `'\\''`)."""
+    return "'" + str(s).replace("'", "'\\''") + "'"
+
+
 class LocalTransport:
     os = "linux"
 
@@ -28,12 +38,14 @@ class LocalTransport:
         return os.path.join(*parts)
 
     def quote(self, s):
-        return "'" + str(s).replace("'", "'\\''") + "'"
+        return sh_quote(s)
 
     def mkdir(self, path, clean=False):
         p = Path(path)
         p.mkdir(parents=True, exist_ok=True)
         if clean:
+            if any((p / n).exists() for n in PID_FILES):
+                raise TransportError("refusing to clean %s: it holds a pid file (HASPID)" % p)
             for f in p.iterdir():
                 if f.is_file():
                     f.unlink()
@@ -45,6 +57,9 @@ class LocalTransport:
         p = Path(path)
         return p.read_text(encoding="utf-8", errors="replace") if p.exists() else None
 
+    def probe(self, run_dir, names):
+        return [n for n in names if os.path.exists(os.path.join(run_dir, n))]
+
     def grep_first(self, path, needle):
         text = self.read_text(path) or ""
         return next((line for line in text.splitlines() if needle in line), None)
@@ -76,5 +91,5 @@ class LocalTransport:
         pass
 
     def process_count(self, name):
-        out = subprocess.run(["pgrep", "-c", "-x", name], capture_output=True, text=True)
+        out = subprocess.run(["pgrep", "-c", "-x", name], capture_output=True, text=True, env=self.cfg.scrubbed_env())
         return int(out.stdout.strip() or 0)

agentd/transports/warpgate_ssh.py

@@ -1,16 +1,117 @@
-"""Windows terminal behind Warpgate, reached through the server-access relay helper scripts:
-  win10-ps.sh  <script>          run PowerShell (sent as -EncodedCommand)
-  win10-put.sh <local> <remote>  upload a file (gzip+base64 chunks, SHA-256 verified)
-Runs are launched as one-shot scheduled tasks (schtasks) so they outlive the SSH session.
-Note: the helper scripts currently hard-code `--target win10`; `target` is checked against that."""
+"""Remote terminals (Windows and Linux/macOS) and the Warpgate channel.
+
+The transport classes build the scripts; a *channel* carries them (design §9):
+- WarpgateChannel (`transport: warpgate-ssh`, also the direct-ssh fallback): the server-access relay client
+
+    agent-access.py --config <identity> run --target <target> --timeout N -- <argv...>
+
+  The client shlex-joins argv and the target account's login shell runs it (cmd.exe on Windows); stdin is
+  /dev/null (agent-access.py run_remote uses stdin=DEVNULL), so everything travels in argv. `relay()` is the single
+  place that builds that call.
+- DirectChannel (`transport: direct-ssh`, agentd/transports/direct_ssh.py): OpenSSH over Tailscale; the script goes
+  on stdin, so every operation (including a launch with its prompt) is one call.
+
+- os linux / darwin -> PosixWarpgateTransport: every operation is one `bash -c '<script>'` call. On the relay files go
+  up as base64 chunks (`printf %s '<b64>' | base64 -d`), on the direct channel in the same single call. The runner
+  starts detached with `setsid nohup bash run.sh` (its own session: pid == pgid == sid), `poll` reads exit.txt,
+  `fetch` returns everything `collect` needs in one call, `kill` signals the process group and session.
+- os windows -> WindowsWarpgateTransport: PowerShell scripts; runs are one-shot scheduled tasks (schtasks) so they
+  outlive the SSH session. On the relay it still drives the helper scripts win10-ps.sh (PowerShell -EncodedCommand,
+  argv-size bound) and win10-put.sh (chunked upload, SHA-256 verified), which hard-code `--target win10`; on the
+  direct channel mkdir + clean + prompt.md + run.ps1 + schtasks is one call.
+The class names are historical: the channel, not the class, decides the path."""
 import base64
 import hashlib
 import os
+import posixpath
 import re
 import subprocess
+import sys
 import tempfile
+import time
+
+from . import NULL_LOG, TransportError, Unreached, fetch_each, operation, parse_fetch
+from .local import RUNNER as POSIX_RUNNER, sh_quote
+
+DEFAULT_CLIENT = "~/.local/share/server-access-client/agent-access.py"
+
+
+def relay_argv(cfg, target, argv, timeout):
+    """argv for one `agent-access.py run` call (the vault client and identity come from the `vault` config)."""
+    v = cfg.data.get("vault") or {}
+    cmd = [sys.executable or "python3", "-I", os.path.expanduser(v.get("client") or DEFAULT_CLIENT)]
+    if v.get("config"):
+        cmd += ["--config", v["config"]]
+    return cmd + ["run", "--target", target, "--timeout", str(int(timeout)), "--"] + list(argv)
+
+
+def call(cfg, argv, timeout, extra_env=None):
+    """Run a local relay process (client or helper script) with Plane secrets scrubbed from its environment.
+    -> (returncode, stdout, stderr); raises TransportError on timeout / launch failure."""
+    env = dict(cfg.scrubbed_env(), **(extra_env or {}))
+    try:
+        out = subprocess.run(argv, capture_output=True, timeout=timeout + 60, env=env, stdin=subprocess.DEVNULL)
+    except subprocess.TimeoutExpired:
+        raise TransportError("relay timeout after %ss" % timeout) from None
+    except OSError as exc:
+        raise TransportError("relay launch failed: %s" % exc) from None
+    dec = lambda b: (b or b"").decode("utf-8", "replace").replace("\r", "")  # noqa: E731
+    return out.returncode, dec(out.stdout), dec(out.stderr)
+
+
+def relay(cfg, target, argv, timeout=150):
+    """Run argv on a Warpgate target. -> (returncode, stdout, stderr)."""
+    return call(cfg, relay_argv(cfg, target, argv, timeout), timeout)
+
+
+PS_ARGV_MAX = 7600   # win10-ps.sh refuses longer -EncodedCommand values (cmd.exe line limit 8191)
+PS_PRELUDE = 120     # characters win10-ps.sh puts in front of every script
+
+
+class WarpgateChannel:
+    """Calls through agent-access.py run (argv only). Each call is logged with its duration."""
+    stdin = False
+
+    def __init__(self, cfg, terminal, log=None, label="warpgate"):
+        self.cfg, self.terminal, self.log, self.via = cfg, terminal, log or NULL_LOG, label
+        t = cfg.terminals[terminal]
+        self.target = t.get("target", terminal)
+        self.name = "relay %s" % self.target
+        if t["os"] == "windows":
+            helpers = cfg.helpers_dir(terminal)
+            self.ps_script, self.put_script = str(helpers / "win10-ps.sh"), str(helpers / "win10-put.sh")
+
+    def timed(self, op, fn, *args):
+        t0 = time.monotonic()
+        try:
+            res = fn(*args)
+        except TransportError:
+            self.log.call(self.terminal, self.via, op, (time.monotonic() - t0) * 1000, "error")
+            raise
+        self.log.call(self.terminal, self.via, op, (time.monotonic() - t0) * 1000, res[0])
+        return res
+
+    def fits(self, script):
+        """Whether win10-ps.sh accepts the script (its -EncodedCommand stays under PS_ARGV_MAX)."""
+        units = len(script.encode("utf-16-le")) // 2 + PS_PRELUDE
+        return (2 * units + 2) // 3 * 4 <= PS_ARGV_MAX
+
+    def ps(self, script, timeout, op=None):
+        rc, out, err = self.timed(op, call, self.cfg, [self.ps_script, script], timeout, {"WIN10_TIMEOUT": str(timeout)})
+        if rc != 0 and not out.strip():
+            raise TransportError("relay exit %s: %s" % (rc, err.strip()[-300:]))
+        return out
+
+    def put(self, local, remote, op=None):
+        with open(local, "rb") as fh:
+            sha = hashlib.sha256(fh.read()).hexdigest()
+        rc, out, err = self.timed(op, call, self.cfg, [self.put_script, local, remote], 540)
+        if ("remote sha256 " + sha) not in out:
+            raise TransportError("upload of %s failed: %s" % (remote, (out + err).strip()[-300:]))
+
+    def sh(self, script, timeout, op=None):
+        return self.timed(op, relay, self.cfg, self.target, ["bash", "-c", script], timeout)
 
-from . import TransportError
 
 RUNNER = r"""$ErrorActionPreference = 'Continue'
 [Console]::OutputEncoding = [Text.Encoding]::UTF8
@@ -25,6 +126,15 @@ Set-Content -Path '{d}\exit.txt' -Value ("exit=" + $LASTEXITCODE + " at " + (Get
 """
 
 
+# Windows fetch helpers: F<i>=<base64> per result, M<i> = file missing (parse_fetch). Names must not collide with
+# PowerShell's built-in aliases (aliases win over functions: `r` is Invoke-History).
+PS_FETCH = r"""function AdRead($i,$n){$p=Join-Path $d $n; if(Test-Path -LiteralPath $p){"F$i="+[Convert]::ToBase64String([IO.File]::ReadAllBytes($p))}else{"M$i"}}
+function AdOut($i,$t){"F$i="+[Convert]::ToBase64String([Text.Encoding]::UTF8.GetBytes([string]$t))}
+function AdGrep($i,$n,$x){$p=Join-Path $d $n; $m=$null; if(Test-Path -LiteralPath $p){$m=Select-String -LiteralPath $p -SimpleMatch -Pattern $x|Select-Object -First 1}; if($m){AdOut $i $m.Line}else{"F$i="}}
+function AdTail($i,$n,$k){$p=Join-Path $d $n; $t=''; if(Test-Path -LiteralPath $p){$t=(Get-Content -LiteralPath $p -Encoding UTF8 -Tail $k) -join "`n"}; AdOut $i $t}
+"""
+
+
 def _b64_payload(out, marker):
     m = re.search(marker + r"=([A-Za-z0-9+/=]*)", out)
     if not m:
@@ -32,15 +142,14 @@ def _b64_payload(out, marker):
     return base64.b64decode(m.group(1)).decode("utf-8", "replace") if m.group(1) else ""
 
 
-class WarpgateSshTransport:
+class WindowsWarpgateTransport:
     os = "windows"
 
-    def __init__(self, cfg, terminal):
+    def __init__(self, cfg, terminal, channel=None, log=None, label="warpgate"):
         self.cfg, self.terminal = cfg, terminal
         t = cfg.terminals[terminal]
         self.target = t.get("target", terminal)
-        helpers = cfg.helpers_dir(terminal)
-        self.ps_script, self.put_script = str(helpers / "win10-ps.sh"), str(helpers / "win10-put.sh")
+        self.channel = channel or WarpgateChannel(cfg, terminal, log, label)
         self.repo = t["repo"]
 
     def default_run_root(self):
@@ -52,33 +161,60 @@ class WarpgateSshTransport:
     def quote(self, s):
         return "'" + str(s).replace("'", "''") + "'"
 
-    # ---------------------------------------------------------------- relay primitives
+    # ---------------------------------------------------------------- channel primitives
     def ps(self, script, timeout=150):
-        env = dict(self.cfg.scrubbed_env(), WIN10_TIMEOUT=str(timeout))
-        try:
-            out = subprocess.run([self.ps_script, script], capture_output=True, text=True, timeout=timeout + 60, env=env)
-        except subprocess.TimeoutExpired:
-            raise TransportError("relay timeout after %ss" % timeout) from None
-        if out.returncode != 0 and not out.stdout.strip():
-            raise TransportError("relay exit %s: %s" % (out.returncode, out.stderr.strip()[-300:]))
-        return out.stdout.replace("\r", "")
+        return self.channel.ps(script, timeout, getattr(self, "_op", None))
 
     def put(self, local, remote):
-        sha = hashlib.sha256(open(local, "rb").read()).hexdigest()
-        out = subprocess.run([self.put_script, local, remote], capture_output=True, text=True, timeout=600,
-                             env=self.cfg.scrubbed_env())
-        if ("remote sha256 " + sha) not in out.stdout:
-            raise TransportError("upload of %s failed: %s" % (remote, (out.stdout + out.stderr).strip()[-300:]))
+        self.channel.put(local, remote, getattr(self, "_op", None))
+
+    @operation("ping")
+    def ping(self):
+        out = self.ps("'PONG=' + $PSVersionTable.PSVersion.ToString() + ' ' + $env:COMPUTERNAME", timeout=60)
+        m = re.search(r"PONG=[^\n]*", out)
+        if not m:
+            raise TransportError("ping %s: unexpected output: %s" % (self.terminal, out.strip()[-200:]))
+        return m.group(0)
 
     # ---------------------------------------------------------------- file ops
-    def mkdir(self, path, clean=False):
+    def mkdir_script(self, path, clean=False):
         q = self.quote(path)
         script = "New-Item -ItemType Directory -Force -Path %s | Out-Null\n" % q
-        if clean:
+        if clean:  # never wipe a directory a runner has already started in
+            script += ("if ((Test-Path -LiteralPath (Join-Path %s 'pid.txt')) -or (Test-Path -LiteralPath (Join-Path %s 'launch.pid'))) "
+                       "{ 'HASPID'; return }\n" % (q, q))
             script += "Get-ChildItem -LiteralPath %s -File | Remove-Item -Force\n" % q
-        self.ps(script + "'MKDIR=ok'")
+        return script
 
+    @operation("mkdir")
+    def mkdir(self, path, clean=False):
+        out = self.ps(self.mkdir_script(path, clean) + "'MKDIR=ok'")
+        if "MKDIR=ok" not in out:
+            raise TransportError("mkdir %s failed: %s" % (path, out.strip()[-300:]))
+
+    @operation("probe")
+    def probe(self, run_dir, names):
+        out = self.ps("$d = %s\n'PROBE=' + ((@(%s) | Where-Object { Test-Path -LiteralPath (Join-Path $d $_) }) -join ',')"
+                      % (self.quote(run_dir), ",".join(self.quote(n) for n in names)))
+        m = re.search(r"PROBE=([^\n]*)", out)
+        if not m:
+            raise TransportError("probe %s: unexpected relay output: %s" % (run_dir, out.strip()[-200:]))
+        return [n for n in m.group(1).strip().split(",") if n]
+
+    def write_script(self, path, data):
+        """Write bytes exactly as given (UTF-8 without BOM for our texts), echo their SHA-256."""
+        return ("$p = %s\n[IO.File]::WriteAllBytes($p, [Convert]::FromBase64String('%s'))\n"
+                "'WROTE=' + (Get-FileHash -LiteralPath $p -Algorithm SHA256).Hash.ToLower()\n"
+                % (self.quote(path), base64.b64encode(data).decode("ascii")))
+
+    @operation("write")
     def write_text(self, path, text):
+        if self.channel.stdin:
+            data = text.encode("utf-8")
+            out = self.ps(self.write_script(path, data))
+            if re.findall(r"WROTE=([0-9a-f]{64})", out) != [hashlib.sha256(data).hexdigest()]:
+                raise TransportError("upload of %s failed: %s" % (path, out.strip()[-300:]))
+            return
         fd, tmp = tempfile.mkstemp(prefix="agentd-", suffix=".txt")
         try:
             with os.fdopen(fd, "w", encoding="utf-8", newline="") as fh:
@@ -87,6 +223,7 @@ class WarpgateSshTransport:
         finally:
             os.unlink(tmp)
 
+    @operation("read")
     def read_text(self, path):
         q = self.quote(path)
         out = self.ps("if (Test-Path -LiteralPath %s) { 'B64=' + [Convert]::ToBase64String([IO.File]::ReadAllBytes(%s)) }"
@@ -96,8 +233,9 @@ class WarpgateSshTransport:
         text = _b64_payload(out, "B64")
         if text is None:
             raise TransportError("read %s: unexpected relay output" % path)
-        return text.lstrip("")
+        return text.lstrip("\ufeff")
 
+    @operation("grep")
     def grep_first(self, path, needle):
         q = self.quote(path)
         out = self.ps("$m = if (Test-Path -LiteralPath %s) { Select-String -LiteralPath %s -SimpleMatch -Pattern %s | Select-Object -First 1 }\n"
@@ -105,12 +243,30 @@ class WarpgateSshTransport:
                       % (q, q, self.quote(needle)))
         return _b64_payload(out, "B64") or None
 
+    @operation("tail")
     def tail(self, path, n):
         q = self.quote(path)
         out = self.ps("$t = if (Test-Path -LiteralPath %s) { (Get-Content -LiteralPath %s -Encoding UTF8 -Tail %d) -join \"`n\" } else { '' }\n"
                       "'B64=' + [Convert]::ToBase64String([Text.Encoding]::UTF8.GetBytes([string]$t))" % (q, q, n))
         return _b64_payload(out, "B64") or ""
 
+    def fetch_script(self, run_dir, ops):
+        lines = ["$d = %s" % self.quote(run_dir), PS_FETCH.rstrip("\n")]
+        for i, o in enumerate(ops):
+            fn = {"read": "AdRead", "grep": "AdGrep", "tail": "AdTail"}[o[0]]
+            arg = "" if o[0] == "read" else " " + (self.quote(o[2]) if o[0] == "grep" else str(int(o[2])))
+            lines.append("%s %d %s%s" % (fn, i, self.quote(o[1]), arg))
+        return "\n".join(lines) + "\n"
+
+    @operation("fetch")
+    def fetch(self, run_dir, ops):
+        """Several reads of one run directory in one call (collect)."""
+        script = self.fetch_script(run_dir, ops)
+        if not self.channel.stdin and not self.channel.fits(script):
+            return fetch_each(self, run_dir, ops)
+        res = parse_fetch(self.ps(script, timeout=200), ops, run_dir)
+        return [r.lstrip("\ufeff") if o[0] == "read" and r else r for o, r in zip(ops, res)]
+
     # ---------------------------------------------------------------- processes
     def runner(self, run_dir, cwd, command, stdout_name):
         return "run.ps1", RUNNER.format(cwd=self.quote(cwd), d=run_dir.replace("'", "''"), command=command, stdout=stdout_name)
@@ -125,24 +281,279 @@ class WarpgateSshTransport:
                 "schtasks /Run /TN '%(t)s' | Out-Null\n"
                 "'LAUNCHED=' + $LASTEXITCODE" % dict(t=task.replace("'", "''"), s=script_path.replace("'", "''")))
 
+    @operation("launch")
     def launch(self, task, script_path, run_dir):
         out = self.ps(self.launch_script(task, script_path))
         if "LAUNCHED=0" not in out:
             raise TransportError("schtasks launch of %s failed: %s" % (task, out.strip()[-300:]))
 
+    def start_script(self, task, run_dir, files, script_name):
+        """mkdir + clean (HASPID refusal) + every file + schtasks, as one script. -> (script, [sha256 per file])"""
+        script, shas = "$ErrorActionPreference = 'Stop'\n" + self.mkdir_script(run_dir, clean=True), []
+        for name, text in files.items():
+            data = text.encode("utf-8")
+            script += self.write_script(self.join(run_dir, name), data)
+            shas.append(hashlib.sha256(data).hexdigest())
+        script += "$ErrorActionPreference = 'Continue'\n" + self.launch_script(task, self.join(run_dir, script_name))
+        return script, shas
+
+    @operation("start")
+    def start_run(self, task, run_dir, files, script_name):
+        """Stage the run directory and start the task: one call on the direct channel; on the relay (argv only)
+        mkdir, chunked uploads and the launch go one by one."""
+        if not self.channel.stdin:
+            self.mkdir(run_dir, clean=True)
+            for name, text in files.items():
+                self.write_text(self.join(run_dir, name), text)
+            return self.launch(task, self.join(run_dir, script_name), run_dir)
+        script, shas = self.start_script(task, run_dir, files, script_name)
+        out = self.ps(script)
+        if re.search(r"^HASPID$", out, re.M):
+            raise TransportError("refusing to clean %s: it holds a pid file (HASPID)" % run_dir)
+        wrote = re.findall(r"WROTE=([0-9a-f]{64})", out)
+        if wrote != shas:
+            raise TransportError("staging %s: hash mismatch (%d/%d files): %s" % (run_dir, len(wrote), len(shas),
+                                                                                 out.strip()[-300:]))
+        if "LAUNCHED=0" not in out:
+            raise TransportError("schtasks launch of %s failed: %s" % (task, out.strip()[-300:]))
+
+    @operation("kill")
     def kill(self, task, run_dir):
+        t = task.replace("'", "''")
         self.ps("$p = Join-Path %s 'pid.txt'\n"
                 "if (Test-Path $p) { taskkill /T /F /PID (Get-Content $p | Select-Object -First 1) 2>&1 | Out-Null }\n"
                 "schtasks /End /TN '%s' 2>&1 | Out-Null\n"
-                "schtasks /Delete /F /TN '%s' 2>&1 | Out-Null\n'KILLED=ok'" % (self.quote(run_dir), task, task))
+                "schtasks /Delete /F /TN '%s' 2>&1 | Out-Null\n'KILLED=ok'" % (self.quote(run_dir), t, t))
 
+    @operation("forget")
     def forget(self, task):
-        self.ps("schtasks /Delete /F /TN '%s' 2>&1 | Out-Null; 'FORGOT=ok'" % task)
+        self.ps("schtasks /Delete /F /TN '%s' 2>&1 | Out-Null; 'FORGOT=ok'" % task.replace("'", "''"))
 
+    @operation("busy")
     def process_count(self, name):
         try:
             out = self.ps("'COUNT=' + @(Get-Process -Name %s -ErrorAction SilentlyContinue).Count" % self.quote(name))
+        except Unreached:
+            raise   # a direct-ssh terminal retries this on its fallback
         except TransportError:
             return None
         m = re.search(r"COUNT=(\d+)", out)
         return int(m.group(1)) if m else None
+
+
+# ==================================================================== Linux / macOS
+DETACH = {
+    # bash puts no background job in its own process group (no job control), so setsid(1) does not fork:
+    # $! is the runner itself and becomes pid == pgid == sid.
+    "linux": "setsid nohup bash %s",
+    # macOS has no setsid(1); perl's POSIX::setsid does the same in-process before exec.
+    "darwin": "nohup perl -MPOSIX -e 'POSIX::setsid() or die \"setsid: $!\"; exec @ARGV' bash %s",
+}
+
+
+class PosixWarpgateTransport:
+    """Linux/macOS terminal. Each primitive is one `bash -c` round trip (~10 s through Vault + Warpgate, well under
+    a second on a reused direct connection), so `start_run` stages the run directory and launches it in a single
+    call and `fetch` reads everything `collect` needs in one."""
+    CHUNK = 60000    # relay only: base64 chars per call; one argv element must stay below 128 KiB (MAX_ARG_STRLEN)
+
+    def __init__(self, cfg, terminal, channel=None, log=None, label="warpgate"):
+        self.cfg, self.terminal = cfg, terminal
+        t = cfg.terminals[terminal]
+        self.os = t["os"]
+        self.target = t.get("target", terminal)
+        self.channel = channel or WarpgateChannel(cfg, terminal, log, label)
+        self.repo = t["repo"]
+
+    def default_run_root(self):
+        return posixpath.join(self.repo, "runtime", "agentd")
+
+    def join(self, *parts):
+        return posixpath.join(*parts)
+
+    def quote(self, s):
+        return sh_quote(s)
+
+    # ---------------------------------------------------------------- channel primitive
+    def sh(self, script, timeout=150, marker=None):
+        """Run a bash script on the target. With `marker`, its absence in stdout is an error."""
+        rc, out, err = self.channel.sh(script, timeout, getattr(self, "_op", None))
+        if (marker and marker not in out) or (not marker and rc != 0 and not out.strip()):
+            raise TransportError("%s exit %s: %s" % (self.channel.name, rc, (out + err).strip()[-300:]))
+        return out
+
+    @operation("ping")
+    def ping(self):
+        out = self.sh("echo \"PONG=$(uname -sr) $(uname -n)\"", timeout=60, marker="PONG=")
+        return re.search(r"PONG=[^\n]*", out).group(0)
+
+    # ---------------------------------------------------------------- file ops
+    def mkdir_script(self, path, clean=False):
+        q = self.quote(path)
+        s = "mkdir -p %s || exit 1\n" % q
+        if clean:  # never wipe a directory a runner has already started in
+            s += "if [ -e %s ] || [ -e %s ]; then echo HASPID; exit 1; fi\n" % (
+                self.quote(posixpath.join(path, "pid.txt")), self.quote(posixpath.join(path, "launch.pid")))
+            s += "find %s -maxdepth 1 -type f -exec rm -f {} + || exit 1\n" % q
+        return s
+
+    @operation("mkdir")
+    def mkdir(self, path, clean=False):
+        self.sh(self.mkdir_script(path, clean) + "echo MKDIR=ok\n", marker="MKDIR=ok")
+
+    @staticmethod
+    def b64(text):
+        return base64.b64encode(text.encode("utf-8")).decode("ascii")
+
+    def put_script(self, path, b64):
+        """Write a base64 payload (alphabet has no quote characters) to path atomically, echo its byte size."""
+        q, part = self.quote(path), self.quote(path + ".part")
+        return ("printf %%s '%s' | base64 -d > %s && mv -f %s %s || exit 1\n"
+                "echo \"WROTE=$(wc -c < %s | tr -d ' ')\"\n" % (b64, part, part, q, q))
+
+    @operation("write")
+    def write_text(self, path, text):
+        data = self.b64(text)
+        size = len(text.encode("utf-8"))
+        if len(data) <= self.CHUNK or self.channel.stdin:   # stdin carries the script: no argv limit
+            out = self.sh(self.put_script(path, data), marker="WROTE=")
+        else:  # append chunks to <path>.b64, decode on the last call
+            q, qb = self.quote(path), self.quote(path + ".b64")
+            chunks = [data[i:i + self.CHUNK] for i in range(0, len(data), self.CHUNK)]
+            for i, c in enumerate(chunks):
+                s = "printf %%s '%s' %s %s || exit 1\n" % (c, ">" if i == 0 else ">>", qb)
+                if i < len(chunks) - 1:
+                    self.sh(s + "echo CHUNK=ok\n", marker="CHUNK=ok")
+                else:
+                    out = self.sh(s + "base64 -d < %s > %s && rm -f %s || exit 1\necho \"WROTE=$(wc -c < %s | tr -d ' ')\"\n"
+                                  % (qb, q, qb, q), marker="WROTE=")
+        if re.findall(r"WROTE=(\d+)", out) != [str(size)]:
+            raise TransportError("upload of %s: size mismatch (%s)" % (path, out.strip()[-200:]))
+
+    @operation("read")
+    def read_text(self, path):
+        q = self.quote(path)
+        out = self.sh("if [ -f %s ]; then printf B64=; base64 < %s | tr -d '\\n'; echo; else echo MISSING; fi" % (q, q),
+                      timeout=200)
+        if "B64=" not in out:
+            if "MISSING" in out:
+                return None
+            raise TransportError("read %s: unexpected relay output" % path)
+        return _b64_payload(out, "B64")
+
+    @operation("probe")
+    def probe(self, run_dir, names):
+        d = self.quote(run_dir)
+        out = self.sh("printf PROBE=; for f in %s; do [ -e %s/\"$f\" ] && printf '%%s,' \"$f\"; done; echo"
+                      % (" ".join(self.quote(n) for n in names), d), timeout=60, marker="PROBE=")
+        return [n for n in re.search(r"PROBE=([^\n]*)", out).group(1).split(",") if n]
+
+    def _b64_cmd(self, pipeline):
+        out = self.sh("printf B64=; { %s; } 2>/dev/null | base64 | tr -d '\\n'; echo" % pipeline)
+        text = _b64_payload(out, "B64")
+        if text is None:
+            raise TransportError("relay %s: unexpected output" % self.target)
+        return text
+
+    @operation("grep")
+    def grep_first(self, path, needle):
+        line = self._b64_cmd("grep -F -m1 -e %s %s" % (self.quote(needle), self.quote(path)))
+        return line.rstrip("\n") or None
+
+    @operation("tail")
+    def tail(self, path, n):
+        return self._b64_cmd("tail -n %d %s" % (int(n), self.quote(path))).rstrip("\n")
+
+    def fetch_script(self, run_dir, ops):
+        s = "d=%s\n" % self.quote(run_dir)
+        for i, o in enumerate(ops):
+            f = "\"$d\"/%s" % self.quote(o[1])
+            if o[0] == "read":
+                s += ("if [ -f %s ]; then printf F%d=; base64 < %s | tr -d '\\n'; echo; else echo M%d; fi\n"
+                      % (f, i, f, i))
+            else:
+                cmd = ("grep -F -m1 -e %s %s" % (self.quote(o[2]), f) if o[0] == "grep"
+                       else "tail -n %d %s" % (int(o[2]), f))
+                s += "printf F%d=; { %s; } 2>/dev/null | base64 | tr -d '\\n'; echo\n" % (i, cmd)
+        return s + "echo FETCH=ok\n"
+
+    @operation("fetch")
+    def fetch(self, run_dir, ops):
+        """Several reads of one run directory in one call (collect)."""
+        return parse_fetch(self.sh(self.fetch_script(run_dir, ops), timeout=200, marker="FETCH=ok"), ops, run_dir)
+
+    # ---------------------------------------------------------------- processes
+    def runner(self, run_dir, cwd, command, stdout_name):
+        return "run.sh", POSIX_RUNNER.format(cwd=self.quote(cwd), d=self.quote(run_dir), command=command,
+                                             stdout=stdout_name)
+
+    def launch_script(self, run_dir, script_name):
+        """Detach the runner from the SSH session; stdio closed so the channel ends at once. $! -> launch.pid
+        (kill falls back to it until run.sh has written pid.txt)."""
+        d = self.quote(run_dir)
+        return ("cd %s || exit 1\n"
+                "%s </dev/null >/dev/null 2>&1 &\n"
+                "echo $! > launch.pid\n"
+                "echo LAUNCHED=$!\n" % (d, DETACH[self.os] % self.quote(script_name)))
+
+    @operation("launch")
+    def launch(self, task, script_path, run_dir):
+        self.sh(self.launch_script(run_dir, posixpath.basename(script_path)), marker="LAUNCHED=")
+
+    @operation("start")
+    def start_run(self, task, run_dir, files, script_name):
+        """mkdir + clean, write `files` {name: text}, launch `script_name` -- one call; on the relay a file larger
+        than CHUNK goes up first through write_text."""
+        big = {} if self.channel.stdin else {n: t for n, t in files.items() if len(self.b64(t)) > self.CHUNK}
+        script = self.mkdir_script(run_dir, clean=True)
+        if big:
+            self.mkdir(run_dir, clean=True)
+            for name, text in big.items():
+                self.write_text(self.join(run_dir, name), text)
+            script = ""
+        sizes = {}
+        for name, text in files.items():
+            if name not in big:
+                script += self.put_script(self.join(run_dir, name), self.b64(text))
+                sizes[name] = len(text.encode("utf-8"))
+        out = self.sh(script + self.launch_script(run_dir, script_name), marker="LAUNCHED=")
+        wrote = re.findall(r"WROTE=(\d+)", out)
+        if wrote != [str(v) for v in sizes.values()]:
+            raise TransportError("staging %s: size mismatch %s != %s" % (run_dir, wrote, list(sizes.values())))
+
+    def kill_script(self, run_dir):
+        """TERM then KILL the runner's process group and session. Skipped once exit.txt exists (the runner is
+        done and its pid may be reused)."""
+        d = self.quote(run_dir)
+        return ("d=%s\n"
+                "if [ -f \"$d/exit.txt\" ]; then echo KILLED=finished; exit 0; fi\n"
+                "p=$(head -n1 \"$d/pid.txt\" 2>/dev/null); [ -n \"$p\" ] || p=$(head -n1 \"$d/launch.pid\" 2>/dev/null)\n"
+                "case \"$p\" in ''|*[!0-9]*) echo KILLED=nopid; exit 0;; esac\n"
+                "kill -TERM -- \"-$p\" 2>/dev/null; pkill -TERM -s \"$p\" 2>/dev/null\n"
+                "for i in 1 2 3 4 5; do pgrep -s \"$p\" >/dev/null 2>&1 || break; sleep 1; done\n"
+                "kill -KILL -- \"-$p\" 2>/dev/null; pkill -KILL -s \"$p\" 2>/dev/null\n"
+                "sleep 1; echo \"KILLED=ok left=$(pgrep -s \"$p\" 2>/dev/null | wc -l | tr -d ' ')\"\n" % d)
+
+    @operation("kill")
+    def kill(self, task, run_dir):
+        self.sh(self.kill_script(run_dir), marker="KILLED=")
+
+    def forget(self, task):
+        pass
+
+    @operation("busy")
+    def process_count(self, name):
+        try:
+            out = self.sh("echo COUNT=$(pgrep -x %s | wc -l | tr -d ' ')" % self.quote(name), timeout=60)
+        except Unreached:
+            raise   # a direct-ssh terminal retries this on its fallback
+        except TransportError:
+            return None
+        m = re.search(r"COUNT=(\d+)", out)
+        return int(m.group(1)) if m else None
+
+
+def WarpgateSshTransport(cfg, terminal, channel=None, log=None, label="warpgate"):
+    """Factory: the transport class for a remote terminal follows its `os`; the channel defaults to Warpgate."""
+    cls = WindowsWarpgateTransport if cfg.terminals[terminal]["os"] == "windows" else PosixWarpgateTransport
+    return cls(cfg, terminal, channel=channel, log=log, label=label)

agentd/webhook.py

@@ -0,0 +1,238 @@
+"""Plane webhook receiver (PAD-7): a delivery for a tenant project triggers a dispatcher tick right away; the
+2-minute cron tick stays as the fallback. Stdlib only.
+
+POST <path> (default /plane):
+  - X-Plane-Signature = hex HMAC-SHA256 of the request body with the webhook secret (Plane signs
+    json.dumps(payload) and sends the same bytes); missing or wrong -> 401, nothing runs.
+  - body {"event": "issue"|"issue_comment", "action": ..., "data": {..., "project": <uuid>}, "activity": ...};
+    data.project not a configured tenant (or absent, e.g. on delete) -> 202, nothing runs.
+  - otherwise -> 200 and the Trigger schedules a tick.
+GET /healthz -> 200 JSON (pending / running / last tick), no secrets.
+
+Trigger: deliveries within `debounce_seconds` coalesce into one tick; deliveries that arrive while a tick runs cause
+one more tick after it (the running tick may have read Plane before the change). The tick command exits 75 when the
+cron lock (or agentd's own state lock) is held: then it is retried every `retry_seconds` for up to
+`retry_window_seconds`; after that the next cron tick picks the change up. Every delivery and tick is logged to
+<log_dir>/webhook.log."""
+import hashlib
+import hmac
+import json
+import os
+import subprocess
+import threading
+import time
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+from . import ROOT
+
+BUSY = 75             # tick command exit code: lock held by another tick (flock -E 75, cli tick on Busy)
+MAX_BODY = 2 << 20    # Plane payloads are a few KB
+
+
+def sign(secret, body):
+    return hmac.new(secret.encode("utf-8"), body, hashlib.sha256).hexdigest()
+
+
+def verify(secret, body, signature):
+    """Constant-time check of the hex signature against the raw body; falls back to Plane's own serialisation
+    (json.dumps of the parsed payload) in case a proxy re-encoded the body."""
+    if not secret or not signature:
+        return False
+    sig = signature.strip().lower().encode("ascii", "replace")
+    if hmac.compare_digest(sign(secret, body).encode(), sig):
+        return True
+    try:
+        canonical = json.dumps(json.loads(body)).encode("utf-8")
+    except ValueError:
+        return False
+    return canonical != body and hmac.compare_digest(sign(secret, canonical).encode(), sig)
+
+
+def project_of(payload):
+    data = payload.get("data") if isinstance(payload, dict) else None
+    pid = (data or {}).get("project") or (data or {}).get("project_id")
+    return pid.get("id") if isinstance(pid, dict) else pid
+
+
+def route(payload, projects):
+    """Configured project name for the delivery, or None."""
+    pid = project_of(payload)
+    return next((name for name, p in projects.items() if pid and p["id"] == pid), None)
+
+
+class Logger:
+    def __init__(self, path):
+        self.path = path
+        self.lock = threading.Lock()
+
+    def __call__(self, msg):
+        line = time.strftime("%Y-%m-%dT%H:%M:%S") + " " + msg + "\n"
+        with self.lock:
+            if self.path:
+                os.makedirs(os.path.dirname(self.path), exist_ok=True)
+                with open(self.path, "a", encoding="utf-8") as fh:
+                    fh.write(line)
+            else:
+                print(line, end="", flush=True)
+
+
+class Trigger:
+    """Debounced tick trigger. `run()` executes one tick and returns its exit code (BUSY = lock held)."""
+
+    def __init__(self, run, debounce=3.0, retry=15.0, retry_window=120.0, log=print, clock=time.monotonic):
+        self.run, self.debounce, self.retry, self.retry_window = run, debounce, retry, retry_window
+        self.log, self.clock = log, clock
+        self.cv = threading.Condition()
+        self.due = None           # clock() when the next tick starts; None = nothing pending
+        self.reasons = []         # deliveries folded into the pending tick
+        self.busy_since = None    # first BUSY of the current retry chain
+        self.running = False
+        self.last = {}
+        self.ticks = 0
+
+    def kick(self, reason):
+        with self.cv:
+            if self.due is None:
+                self.due = self.clock() + self.debounce
+            self.reasons.append(reason)
+            self.cv.notify()
+
+    def status(self):
+        with self.cv:
+            return {"pending": self.due is not None, "queued": len(self.reasons), "running": self.running,
+                    "retrying_since": self.busy_since and round(self.clock() - self.busy_since, 1), "ticks": self.ticks,
+                    "last": dict(self.last)}
+
+    def start(self):
+        threading.Thread(target=self.serve_forever, name="tick-trigger", daemon=True).start()
+        return self
+
+    def serve_forever(self):
+        while True:
+            self.step()
+
+    def step(self, wait=None):
+        """Wait until a tick is due (or `wait` seconds pass), run it, handle the result. True if a tick ran."""
+        with self.cv:
+            end = None if wait is None else self.clock() + wait
+            while self.due is None or self.due > self.clock():
+                timeout = None if self.due is None else self.due - self.clock()
+                if end is not None:
+                    left = end - self.clock()
+                    if left <= 0:
+                        return False
+                    timeout = left if timeout is None else min(timeout, left)
+                self.cv.wait(timeout)
+            reasons, self.reasons, self.due, self.running = self.reasons, [], None, True
+        start = self.clock()
+        self.log("tick start (%d deliveries: %s)" % (len(reasons), ", ".join(reasons[:5]) + (" …" if len(reasons) > 5 else "")))
+        try:
+            rc = self.run()
+        except Exception as exc:   # a broken runner must not kill the trigger thread
+            rc = "error %r" % (exc,)
+        took = round(self.clock() - start, 1)
+        with self.cv:
+            self.running = False
+            self.ticks += 1
+            self.last = {"rc": rc, "seconds": took, "deliveries": len(reasons), "at": time.strftime("%Y-%m-%dT%H:%M:%S")}
+            if rc != BUSY:
+                self.busy_since = None
+                self.log("tick done rc=%s in %.1fs" % (rc, took))
+                return True
+            now = self.clock()
+            self.busy_since = self.busy_since or start
+            if now - self.busy_since + self.retry > self.retry_window:
+                self.log("tick busy for %.0fs, giving up; the next cron tick covers it" % (now - self.busy_since))
+                self.busy_since = None
+                return True
+            self.reasons = reasons + self.reasons
+            self.due = min(self.due, now + self.retry) if self.due is not None else now + self.retry
+            self.log("tick busy (another tick holds the lock), retry in %gs" % self.retry)
+            return True
+
+
+def command_runner(cmd, env, timeout=1000):
+    """Run the tick command (it logs to cron.log itself); returns its exit code."""
+    def run():
+        try:
+            return subprocess.run(cmd, env=env, stdin=subprocess.DEVNULL, stdout=subprocess.DEVNULL,
+                                  stderr=subprocess.DEVNULL, timeout=timeout).returncode
+        except subprocess.TimeoutExpired:
+            return "timeout"
+    return run
+
+
+def make_handler(secret, projects, trigger, log, path="/plane"):
+    class Handler(BaseHTTPRequestHandler):
+        server_version = "plane-agentd-webhook"
+
+        def log_message(self, fmt, *args):   # deliveries are logged by us, one line each
+            pass
+
+        def reply(self, code, obj):
+            body = json.dumps(obj).encode()
+            self.send_response(code)
+            self.send_header("Content-Type", "application/json")
+            self.send_header("Content-Length", str(len(body)))
+            self.end_headers()
+            self.wfile.write(body)
+
+        def do_GET(self):
+            if self.path.split("?")[0] == "/healthz":
+                return self.reply(200, dict(ok=True, **trigger.status()))
+            self.reply(404, {"error": "not found"})
+
+        def do_POST(self):
+            if self.path.split("?")[0] != path:
+                return self.reply(404, {"error": "not found"})
+            delivery = self.headers.get("X-Plane-Delivery", "-")
+            try:
+                n = int(self.headers.get("Content-Length", ""))
+            except ValueError:
+                n = -1
+            if not 0 <= n <= MAX_BODY:
+                log("delivery=%s from=%s decision=400 bad length" % (delivery, self.client_address[0]))
+                return self.reply(400, {"error": "bad length"})
+            body = self.rfile.read(n)
+            sig = self.headers.get("X-Plane-Signature")
+            if not verify(secret, body, sig):
+                log("delivery=%s from=%s decision=401 %s signature" % (
+                    delivery, self.client_address[0], "bad" if sig else "missing"))
+                return self.reply(401, {"error": "signature"})
+            try:
+                payload = json.loads(body)
+            except ValueError:
+                log("delivery=%s decision=400 bad json" % delivery)
+                return self.reply(400, {"error": "json"})
+            event = "%s/%s" % (payload.get("event") or self.headers.get("X-Plane-Event"), payload.get("action"))
+            name = route(payload, projects)
+            what = "delivery=%s event=%s project=%s item=%s" % (
+                delivery, event, name or project_of(payload) or "-", (payload.get("data") or {}).get("id", "-"))
+            if not name:
+                log(what + " decision=202 ignored (not a tenant project)")
+                return self.reply(202, {"ignored": "project"})
+            trigger.kick(delivery)
+            log(what + " decision=tick")
+            self.reply(200, {"tick": "scheduled"})
+
+    return Handler
+
+
+def serve(cfg, bind=None, port=None):
+    """`agentd webhook-serve`: the secret comes from the environment (agent-access.py exec --secret-env)."""
+    w = cfg.webhook
+    secret = os.environ.get(w["secret_key"])
+    if not secret:
+        raise SystemExit("webhook secret %s not in the environment; run under agent-access.py exec --secret-env %s=%s"
+                         % (w["secret_key"], w["secret_key"], w["secret_key"]))
+    log = Logger(str(cfg.log_dir / "webhook.log"))
+    cmd = list(w["tick_command"])
+    if "/" in cmd[0] and not os.path.isabs(cmd[0]):
+        cmd[0] = str(ROOT / cmd[0])   # relative script path = repo
+    run = command_runner(cmd, cfg.scrubbed_env(), w.get("tick_timeout_seconds", 1000))
+    trigger = Trigger(run, w.get("debounce_seconds", 3), w.get("retry_seconds", 15), w.get("retry_window_seconds", 120),
+                      log=log).start()
+    bind, port = bind or w.get("bind", "127.0.0.1"), port or w.get("port", 8790)
+    httpd = ThreadingHTTPServer((bind, port), make_handler(secret, cfg.projects, trigger, log, w.get("path", "/plane")))
+    log("listening on http://%s:%s%s (tenants: %s)" % (bind, port, w.get("path", "/plane"), ", ".join(cfg.projects)))
+    httpd.serve_forever()

cli/agentd

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""Entry point: agentd tick|new|show|comment|state|assign|smoke|check-config (see agentd/cli.py)."""
+"""Entry point: agentd tick|new|show|comment|state|assign|smoke|webhook-serve|check-config (see agentd/cli.py)."""
 import sys
 from pathlib import Path
 

config/agentd.example.yaml

@@ -7,15 +7,32 @@ vault: {client: ~/.local/share/server-access-client/agent-access.py, config: htt
 
 agentd:
   state_dir: state                      # relative to the repo root; JSON state, logs, heartbeat, local run dirs
-  helpers_dir: /home/ubuntu/Documents/wjj/project/gwent-design/tools   # win10-ps.sh / win10-put.sh
+  helpers_dir: tools/win10             # win10-ps.sh / win10-put.sh (Windows terminals only); relative = repo
   interval: 120                         # seconds between ticks in --loop mode
   max_rounds_per_thread: 8              # rebuild the executor thread (with a handoff summary) after this many rounds
   comment_max_chars: 6000
   cli_terminal: oregon                  # identity used by `agentd new/comment/state/assign`
+  # ssh_key: AGENTD_SSH_KEY             # direct-ssh: Vault key / env var of the dedicated private key (default)
+  # ssh_dir: /run/plane-agentd/ssh      # direct-ssh: ControlMaster sockets + generated known_hosts (default state_dir/ssh)
+
+# Plane webhook receiver (`agentd webhook-serve`, docs/runbook.md "Webhook"); leave the section out to run on cron only.
+webhook:
+  bind: 127.0.0.1
+  port: 8790
+  path: /plane
+  secret_key: PLANE_WEBHOOK_SECRET    # Vault key name of the Plane webhook secret (workspace-level webhook)
+  tick_command: [cron/agentd-tick.sh] # exits 75 when another tick holds the lock (flock -E 75)
+  debounce_seconds: 3
+  retry_seconds: 15
+  retry_window_seconds: 120
+notify: {kind: none}                  # notification hook: only `none` (no-op) exists today
 
 terminals:
   win10:
-    transport: warpgate-ssh
+    transport: warpgate-ssh             # or direct-ssh over Tailscale (design §9, docs/deploy-lxc.md "直连 SSH"):
+    # transport: direct-ssh
+    # ssh: {host: 100.123.209.19, user: administrator, port: 22, host_key: "ssh-ed25519 AAAA..."}   # pinned host key
+    # fallback: warpgate-ssh            # used only when the direct path fails before the command ran
     target: win10
     os: windows
     launch: schtasks
@@ -29,14 +46,19 @@ terminals:
     capabilities: [android, game]
     executors: [codex]
   oregon:
-    transport: local
+    # remote even though agentd runs on oregon today: the dispatcher is moving to an LXC without agent CLIs
+    # (design §9), so this goes through Warpgate like any other host. `transport: local` + `launch: subprocess`
+    # still works for a dispatcher co-located with its executors.
+    transport: warpgate-ssh
+    target: oregon                      # agent-access.py target name
     os: linux
-    launch: subprocess
+    launch: nohup
     repo: /home/ubuntu/Documents/wjj/project/gwent
+    run_root: /home/ubuntu/.gwent-board/agentd-runs   # absolute: the remote shell does not expand ~ inside quotes
     plane_user: oregon
     token_key: PLANE_OREGON_API_TOKEN
     concurrency: 2
-    bins: {claude: /home/ubuntu/.local/bin/claude}   # cron has a minimal PATH
+    bins: {claude: /home/ubuntu/.local/bin/claude}   # the SSH shell has the system PATH only
     executors: [claude-code, codex, shell]
 
 executors:
@@ -45,6 +67,7 @@ executors:
     run: "codex exec --json -C '{repo}' -s workspace-write -m {model} -c 'model_reasoning_effort=\"{effort}\"' {tier} -o '{out}' -"
     resume: "codex exec resume {thread} --json -m {model} -c 'model_reasoning_effort=\"{effort}\"' {tier} -o '{out}' -"
     tier_arg: "-c 'service_tier=\"{tier}\"'"
+    tier_defaults: {gpt-6.1-sol: fast}   # tier when no label sets one (`tier:<x>`, `tier:none` = no tier)
     thread_from: events.jsonl:thread.started.thread_id
     last_message: '{out}'
     models: [gpt-6.1-sol, gpt-6-astra]
@@ -66,7 +89,6 @@ projects:
   GWENT:
     id: 79cb5a60-aa4f-4138-9864-1dd70426e665
     key_prefix: GW
-    webhook_secret_key: PLANE_WEBHOOK_GWENT
     states: {backlog: Backlog, todo: Todo, running: In Progress, waiting: Waiting, review: Review, done: Done}
     reviewer: oregon                    # Plane user that gets [review] items
     answerer: oregon                    # Plane user that gets [question]/[blocked] and failures (default: reviewer)

config/agentd.oregon.yaml

@@ -0,0 +1,89 @@
+# plane-agentd dispatcher configuration for oregon -- the ROLLBACK target since the dispatcher moved to the LXC
+# agentd-01 (PAD-11; live config there: config/agentd.yaml). Used by cron/agentd-tick.sh (not in oregon's crontab
+# while the LXC dispatches). No secrets: token_key values are Vault key names.
+plane: {base_url: https://plane.kappiee.com, workspace: kappiee}
+vault: {client: ~/.local/share/server-access-client/agent-access.py, config: https://nas.tail9531b.ts.net:8443/identity/agent.json}
+
+agentd:
+  state_dir: ~/.gwent-board/agentd     # JSON state, agentd.log, heartbeat.json, local run dirs (runs/)
+  helpers_dir: tools/win10             # win10-ps.sh / win10-put.sh (Windows terminals only); relative = repo
+  interval: 120
+  max_rounds_per_thread: 8
+  comment_max_chars: 6000
+  cli_terminal: oregon
+
+terminals:
+  win10:
+    transport: warpgate-ssh
+    target: win10
+    os: windows
+    launch: schtasks
+    repo: 'C:\Users\Administrator\Documents\GwentAssistantProject\gwent'
+    run_root: 'C:\Users\Administrator\Documents\GwentAssistantProject\gwent\runtime\agentd'
+    bins: {codex: 'C:\Users\Administrator\AppData\Local\Programs\OpenAI\Codex\bin\codex.exe'}
+    plane_user: window10
+    token_key: PLANE_WINDOW10_API_TOKEN
+    concurrency: 1
+    busy_processes: [codex]             # do not start while a codex.exe agentd did not start is running
+    capabilities: [android, game]
+    executors: [codex]
+  oregon:
+    # remote even though agentd runs on oregon today: the dispatcher is moving to an LXC without agent CLIs
+    # (design §9), so this goes through Warpgate like any other host. `transport: local` + `launch: subprocess`
+    # still works for a dispatcher co-located with its executors.
+    transport: warpgate-ssh
+    target: oregon                      # agent-access.py target name
+    os: linux
+    launch: nohup
+    repo: /home/ubuntu/Documents/wjj/project/gwent
+    run_root: /home/ubuntu/.gwent-board/agentd-runs   # absolute: the remote shell does not expand ~ inside quotes
+    plane_user: oregon
+    token_key: PLANE_OREGON_API_TOKEN
+    concurrency: 1                      # one shared worktree: no parallel edits
+    bins: {claude: /home/ubuntu/.local/bin/claude}   # the SSH non-login shell has the system PATH only; no codex on oregon
+    executors: [claude-code]
+
+executors:
+  codex:
+    adapter: codex
+    run: "codex exec --json -C '{repo}' -s workspace-write -m {model} -c 'model_reasoning_effort=\"{effort}\"' {tier} -o '{out}' -"
+    # `exec resume` takes no -C/-s: the runner already cd's into the repo, the sandbox goes in as a -c override
+    resume: "codex exec resume {thread} --json -m {model} -c 'model_reasoning_effort=\"{effort}\"' -c 'sandbox_mode=\"workspace-write\"' {tier} -o '{out}' -"
+    tier_arg: "-c 'service_tier=\"{tier}\"'"
+    tier_defaults: {gpt-6.1-sol: fast}   # tier when no label sets one (`tier:<x>`, `tier:none` = no tier)
+    thread_from: events.jsonl:thread.started.thread_id
+    last_message: '{out}'
+    models: [gpt-6.1-sol, gpt-6-astra]
+    default_effort: medium
+  claude-code:
+    adapter: claude-code
+    run: "claude -p --model {model} --output-format json --permission-mode {perm}"
+    resume: "claude -p --resume {thread} --model {model} --output-format json --permission-mode {perm}"
+    permission_mode: acceptEdits
+    thread_from: json:session_id
+    last_message: json:result
+    models: [sonnet, opus, haiku]
+
+projects:
+  GWENT:
+    id: 79cb5a60-aa4f-4138-9864-1dd70426e665
+    key_prefix: GW
+    states: {backlog: Backlog, todo: Todo, running: In Progress, waiting: Waiting, review: Review, done: Done}
+    reviewer: oregon
+    answerer: oregon
+    protocol_file: config/protocols/gwent-board-protocol.md   # relative = repo; refresh from gwent/design/board-protocol.md
+    routing:
+      by_assignee: {window10: win10, oregon: oregon}
+      by_label:
+        'model:astra': {model: gpt-6-astra, effort: high}
+        'model:sol': {model: gpt-6.1-sol, tier: fast}
+        'effort:high': {effort: high}
+        'needs-game': {serialize_group: game}
+      default_executor: {win10: codex, oregon: claude-code}
+    retry:
+      transient_max: 3
+      backoff_minutes: [2, 5, 10]
+    timeout_hours: 4
+    max_rounds_per_thread: 8
+    max_progress_rounds: 12
+    # notify (ntfy) is phase 3: left out on purpose

config/agentd.yaml

@@ -0,0 +1,114 @@
+# plane-agentd live configuration on the LXC agentd-01 (/opt/plane-agentd, runs as root; docs/deploy-lxc.md).
+# No secrets: token_key / ssh_key values are Vault key names; the values are injected by
+# `agent-access.py exec --secret-env KEY=KEY` (cron/agentd-tick-lxc.sh). Rollback to oregon: config/agentd.oregon.yaml.
+plane: {base_url: https://plane.kappiee.com, workspace: kappiee}
+vault: {client: ~/.local/share/server-access-client/agent-access.py, config: https://nas.tail9531b.ts.net:8443/identity/agent.json}
+
+agentd:
+  state_dir: /var/lib/plane-agentd     # JSON state, heartbeat.json, lock
+  log_dir: /var/log/plane-agentd       # agentd.log (cron output: cron.log, same dir)
+  helpers_dir: tools/win10             # win10-ps.sh / win10-put.sh (Windows terminals only); relative = repo
+  interval: 120
+  max_rounds_per_thread: 8
+  comment_max_chars: 6000
+  cli_terminal: oregon
+  # direct-ssh (PAD-12, docs/deploy-lxc.md "直连 SSH"): the dedicated key comes from Vault AGENTD_SSH_KEY via
+  # --secret-env; ControlMaster sockets and the known_hosts generated from the pinned host keys live in ssh_dir
+  ssh_key: AGENTD_SSH_KEY
+  ssh_dir: /run/plane-agentd/ssh
+
+# Plane webhook receiver (`agentd webhook-serve`, systemd plane-agentd-webhook.service; docs/runbook.md "Webhook").
+# One workspace-level Plane webhook (work items + comments) -> POST http://10.77.10.249:8790/plane -> one tick.
+webhook:
+  bind: 10.77.10.249                  # LAN only (Plane VM 506 10.77.10.69 is on the same /24)
+  port: 8790
+  path: /plane
+  secret_key: PLANE_WEBHOOK_SECRET    # Vault key name; the value is injected by agent-access.py exec --secret-env
+  # the cron script holds `flock -n -E 75`; --scope moves the tick out of the receiver's cgroup, so restarting the
+  # receiver never kills a running tick
+  tick_command: [systemd-run, --scope, --quiet, --collect, /opt/plane-agentd/cron/agentd-tick-lxc.sh]
+  debounce_seconds: 3                 # deliveries within this window -> one tick
+  retry_seconds: 15                   # lock busy -> retry every 15 s ...
+  retry_window_seconds: 120           # ... for up to 2 min; then the cron tick covers it
+notify: {kind: none}                  # notification hook, no-op (no ntfy service); core calls it on every state move
+
+terminals:
+  win10:
+    # direct over Tailscale (PAD-12); Warpgate only when the direct path fails before a command ran.
+    # Rollback: transport: warpgate-ssh (ssh/fallback may stay, they are ignored then)
+    transport: direct-ssh
+    ssh: {host: 100.123.209.19, user: administrator, port: 22,
+          host_key: "ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIFzLX7QNHDFYwTgY4wNl0ihZryIyKgpp9HPpPDWbrTX3"}
+    fallback: warpgate-ssh
+    target: win10                       # Warpgate target of the fallback
+    os: windows
+    launch: schtasks
+    repo: 'C:\Users\Administrator\Documents\GwentAssistantProject\gwent'
+    run_root: 'C:\Users\Administrator\Documents\GwentAssistantProject\gwent\runtime\agentd'
+    bins: {codex: 'C:\Users\Administrator\AppData\Local\Programs\OpenAI\Codex\bin\codex.exe'}
+    plane_user: window10
+    token_key: PLANE_WINDOW10_API_TOKEN
+    concurrency: 1
+    busy_processes: [codex]             # do not start while a codex.exe agentd did not start is running
+    capabilities: [android, game]
+    executors: [codex]
+  oregon:
+    # the dispatcher (LXC, no agent CLIs) reaches oregon directly over Tailscale, Warpgate as fallback (design §9)
+    transport: direct-ssh
+    ssh: {host: 100.127.58.78, user: ubuntu, port: 22,
+          host_key: "ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAICUoI6W7npX+h/kHcXX3iy2XAWbNusrOjg28Xed75wWr"}
+    fallback: warpgate-ssh
+    target: oregon                      # agent-access.py target name (fallback)
+    os: linux
+    launch: nohup
+    repo: /home/ubuntu/Documents/wjj/project/gwent
+    run_root: /home/ubuntu/.gwent-board/agentd-runs   # absolute: the remote shell does not expand ~ inside quotes
+    plane_user: oregon
+    token_key: PLANE_OREGON_API_TOKEN
+    concurrency: 1                      # one shared worktree: no parallel edits
+    bins: {claude: /home/ubuntu/.local/bin/claude}   # the SSH non-login shell has the system PATH only; no codex on oregon
+    executors: [claude-code]
+
+executors:
+  codex:
+    adapter: codex
+    run: "codex exec --json -C '{repo}' -s workspace-write -m {model} -c 'model_reasoning_effort=\"{effort}\"' {tier} -o '{out}' -"
+    # `exec resume` takes no -C/-s: the runner already cd's into the repo, the sandbox goes in as a -c override
+    resume: "codex exec resume {thread} --json -m {model} -c 'model_reasoning_effort=\"{effort}\"' -c 'sandbox_mode=\"workspace-write\"' {tier} -o '{out}' -"
+    tier_arg: "-c 'service_tier=\"{tier}\"'"
+    tier_defaults: {gpt-6.1-sol: fast}   # tier when no label sets one (`tier:<x>`, `tier:none` = no tier)
+    thread_from: events.jsonl:thread.started.thread_id
+    last_message: '{out}'
+    models: [gpt-6.1-sol, gpt-6-astra]
+    default_effort: medium
+  claude-code:
+    adapter: claude-code
+    run: "claude -p --model {model} --output-format json --permission-mode {perm}"
+    resume: "claude -p --resume {thread} --model {model} --output-format json --permission-mode {perm}"
+    permission_mode: acceptEdits
+    thread_from: json:session_id
+    last_message: json:result
+    models: [sonnet, opus, haiku]
+
+projects:
+  GWENT:
+    id: 79cb5a60-aa4f-4138-9864-1dd70426e665
+    key_prefix: GW
+    states: {backlog: Backlog, todo: Todo, running: In Progress, waiting: Waiting, review: Review, done: Done}
+    reviewer: oregon
+    answerer: oregon
+    protocol_file: config/protocols/gwent-board-protocol.md   # relative = repo; refresh from gwent/design/board-protocol.md
+    routing:
+      by_assignee: {window10: win10, oregon: oregon}
+      by_label:
+        'model:astra': {model: gpt-6-astra, effort: high}
+        'model:sol': {model: gpt-6.1-sol, tier: fast}
+        'effort:high': {effort: high}
+        'needs-game': {serialize_group: game}
+      default_executor: {win10: codex, oregon: claude-code}
+    retry:
+      transient_max: 3
+      backoff_minutes: [2, 5, 10]
+    timeout_hours: 4
+    max_rounds_per_thread: 8
+    max_progress_rounds: 12

config/protocols/gwent-board-protocol.md

@@ -0,0 +1,48 @@
+# 看板协作协议（Plane 作为任务板与信箱）
+
+看板 = https://plane.kappiee.com/kappiee/projects/79cb5a60-aa4f-4138-9864-1dd70426e665/issues/ ，工作项编号 `GW-k`。Gitea 只管代码（仓库、提交、PR），不再放任务。
+
+## 账号 = 终端
+- `kappiee` 用户（所有者，做决定）；`oregon` Claude（设计、派活、验收）；`window10` win10 Codex（执行）；`mac` 之后接入的 Mac 终端。
+- 凭据键（Agent Vault 注入内存，不落盘、不进 win10、不进评论）：`PLANE_OREGON_API_TOKEN`、`PLANE_WINDOW10_API_TOKEN`（worker 以 window10 发帖）。API 头 `X-API-Key`。
+
+## 状态机
+`Backlog → Todo → In Progress → Waiting → Review → Done / Cancelled`
+- Backlog：想法/未排期。Todo + assignee `window10`：交给 Codex，worker 下一个 tick 取走。
+- In Progress：worker 已启动 codex 运行（worker 自动设置）。
+- Waiting + assignee `oregon`：Codex 有待回答的问题或被阻塞（`[question]`/`[blocked]`），等 Claude 回答或升级给 kappiee。
+- Review + assignee `oregon`：产出就绪（`[review]`），Claude 验收。通过 → Done；不通过 → 写意见，改回 Todo + `window10`。
+- `[progress]`：worker 自动改回 Todo + `window10`，下一轮续跑。
+- 需要用户决定时 Claude 把 assignee 改为 `kappiee`（state 保持 Waiting）。
+
+## 工作项
+- 标题 `T-xxx 简述`；描述 = 任务书（目标、允许的工具、不做的事、产出、验收线）。
+- 评论 = 消息，作者由账号区分；正文首行保留角色标记 `[claude]` / `[codex]` / `[worker]`，worker 构造 prompt 时跳过 `[worker]` 评论。
+- 结果标签（Codex 最后一条消息首行必须是其一）：
+  - `[review]` 产出完成，附：改动落点、产出路径、SHA、自检结果、如何重跑。
+  - `[question]` 需要回答才能继续，附"若无人回答，我默认会怎么做"。
+  - `[blocked]` 无法继续，附原因与已尝试的办法。
+  - `[progress]` 做了一部分，需要下一轮继续。
+  - 未标注首行按 `[review]` 处理。
+
+## 标签、模块、关系、视图
+- 标签：`model:astra`（gpt-6-astra，high）/ `model:sol`（gpt-6.1-sol fast，medium，默认）/ `effort:high` / `needs-game`（需要游戏在线，别与其他占用游戏的任务并行）。
+- 模块（Modules）= 任务模块（T-007-0 …）；一个模块完成后 Claude 亲自验收再关闭。
+- 依赖用关系 "blocked by"，被阻塞的项留在 Backlog，前置完成后再转 Todo。
+- 视图：`等我决定`（assignee=kappiee 或 Waiting）、`Codex 执行中`（In Progress / Todo + window10）、`待验收`（Review）。
+
+## Codex 侧（worker 自动化）
+worker 跑在 Claude 所在的 Linux 机（`gwent-design/tools/board-worker-plane.py`，cron 每 2 分钟一次，token 只由 Vault 注入内存）。
+1. 取 state=Todo 且 assignee=`window10` 的工作项，按 `GW-k` 从小到大取一个；win10 上已有 codex 进程则本轮跳过（避免抢游戏/ADB）。
+2. 名称 + 描述 + 评论转成文本拼成 prompt 跑 `codex exec`；同一工作项第二次起 `codex exec resume <thread>` 只追加新评论。
+3. 结束后把最后一条消息贴回评论，按结果标签改状态与 assignee（见状态机）。
+
+## Claude 侧
+- 工具：`gwent-design/tools/plane-claude.sh`（`show` / `comment` / `state` / `assign` / `new` / `done`）。
+- 轮询 Waiting 与 Review（assignee=`oregon`）：回答问题、验收、改任务书或升级给 kappiee。新任务一律建工作项，不再用临时 brief 文件（紧急修复除外）。
+
+## 通用规则
+- 消息要短、要有结论；路径用仓库相对路径；大图只贴缩略拼图（≤1600px）。
+- 不在评论里贴 token、密码、完整日志。
+- 不在 win10 的 checkout 上执行任何 git 写操作（commit / checkout / reset / pull / stash）；产物留在 live 或 stage，入库由 Claude 侧完成。
+- 有疑问优先按任务书里的默认值继续并在评论里说明，而不是停下等人。

config/schema.json

@@ -28,11 +28,38 @@
       "additionalProperties": false,
       "properties": {
         "state_dir": {"type": "string"},
+        "log_dir": {"type": "string"},
         "helpers_dir": {"type": "string"},
         "interval": {"type": "integer", "minimum": 10},
         "max_rounds_per_thread": {"type": "integer", "minimum": 1},
         "comment_max_chars": {"type": "integer", "minimum": 500},
-        "cli_terminal": {"type": "string"}
+        "cli_terminal": {"type": "string"},
+        "ssh_dir": {"type": "string"},
+        "ssh_key": {"type": "string"}
+      }
+    },
+    "webhook": {
+      "type": "object",
+      "required": ["secret_key", "tick_command"],
+      "additionalProperties": false,
+      "properties": {
+        "bind": {"type": "string"},
+        "port": {"type": "integer", "minimum": 1},
+        "path": {"type": "string"},
+        "secret_key": {"type": "string"},
+        "tick_command": {"type": "array", "minItems": 1, "items": {"type": "string"}},
+        "tick_timeout_seconds": {"type": "integer", "minimum": 10},
+        "debounce_seconds": {"type": "number", "minimum": 0},
+        "retry_seconds": {"type": "number", "minimum": 1},
+        "retry_window_seconds": {"type": "number", "minimum": 0}
+      }
+    },
+    "notify": {
+      "type": "object",
+      "additionalProperties": false,
+      "properties": {
+        "kind": {"enum": ["none"]},
+        "on": {"$ref": "#/definitions/strings"}
       }
     },
     "terminals": {"type": "object", "minProperties": 1, "additionalProperties": {"$ref": "#/definitions/terminal"}},
@@ -46,14 +73,26 @@
       "required": ["transport", "os", "launch", "repo", "plane_user", "token_key", "executors"],
       "additionalProperties": false,
       "properties": {
-        "transport": {"enum": ["local", "warpgate-ssh"]},
+        "transport": {"enum": ["local", "warpgate-ssh", "direct-ssh"]},
         "target": {"type": "string"},
-        "os": {"enum": ["linux", "windows"]},
+        "ssh": {
+          "type": "object",
+          "required": ["host", "user", "host_key"],
+          "additionalProperties": false,
+          "properties": {
+            "host": {"type": "string"},
+            "user": {"type": "string"},
+            "port": {"type": "integer", "minimum": 1},
+            "host_key": {"type": "string"}
+          }
+        },
+        "fallback": {"enum": ["warpgate-ssh"]},
+        "os": {"enum": ["linux", "darwin", "windows"]},
         "launch": {"enum": ["subprocess", "nohup", "schtasks"]},
         "repo": {"type": "string"},
         "plane_user": {"type": "string"},
         "token_key": {"type": "string"},
-        "concurrency": {"type": "integer", "minimum": 1},
+        "concurrency": {"type": "integer", "minimum": 0},
         "capabilities": {"$ref": "#/definitions/strings"},
         "executors": {"type": "array", "items": {"type": "string"}, "minItems": 1},
         "bins": {"type": "object", "additionalProperties": {"type": "string"}},
@@ -76,6 +115,7 @@
         "default_effort": {"type": "string"},
         "permission_mode": {"type": "string"},
         "tier_arg": {"type": "string"},
+        "tier_defaults": {"type": "object", "additionalProperties": {"type": "string"}},
         "script": {"type": "string"}
       }
     },
@@ -99,7 +139,6 @@
       "properties": {
         "id": {"type": "string"},
         "key_prefix": {"type": "string"},
-        "webhook_secret_key": {"type": "string"},
         "states": {
           "type": "object",
           "required": ["todo", "running", "waiting", "review", "done"],

cron/agentd-tick-lxc.sh

@@ -0,0 +1,37 @@
+#!/usr/bin/env bash
+# LXC agentd-01 cron entry (*/2, root): one agentd tick under server-access exec (Plane tokens only in the
+# process env), with a lock. The flock skips a tick while the previous one still runs; agentd also holds
+# its own state lock. Logs: /var/log/plane-agentd/{agentd.log,cron.log}; state: /var/lib/plane-agentd.
+# Also the webhook receiver's tick command (agentd/webhook.py): exit 75 = lock held (flock -E 75, or agentd's own
+# state lock), the receiver then retries.
+# AGENTD_SSH_KEY = the direct-ssh key (PAD-12). agent-access.py has no optional secrets: when Vault cannot hand it
+# out (credential_unavailable / credential_invalid, nothing ran) the tick is repeated once without it, so the
+# direct-ssh terminals use their warpgate fallback instead of every tick failing. The tick's stderr is appended to
+# cron.log after its stdout (it has to be inspected for that error first).
+LOCK=/var/lib/plane-agentd/cron.lock
+LOG=/var/log/plane-agentd/cron.log
+CLIENT=(python3 -I /root/.local/share/server-access-client/agent-access.py --config https://nas.tail9531b.ts.net:8443/identity/agent.json)
+
+exec 9>>"$LOCK"
+flock -n -E 75 9 || exit $?
+
+tick() {   # "$@" = extra exec arguments; fd 9 (the lock) is not handed to the tick's children
+  timeout 900 "${CLIENT[@]}" exec \
+    --secret-env PLANE_OREGON_API_TOKEN=PLANE_OREGON_API_TOKEN \
+    --secret-env PLANE_WINDOW10_API_TOKEN=PLANE_WINDOW10_API_TOKEN \
+    "$@" --timeout 850 -- \
+    python3 /opt/plane-agentd/cli/agentd --config /opt/plane-agentd/config/agentd.yaml tick --once 9>&-
+}
+
+err=$(mktemp) || exit 1
+trap 'rm -f "$err"' EXIT
+tick --secret-env AGENTD_SSH_KEY=AGENTD_SSH_KEY >>"$LOG" 2>"$err"
+rc=$?
+cat "$err" >>"$LOG"
+case "$rc:$(cat "$err")" in
+  "1:access-hub: credential_unavailable" | "1:access-hub: credential_invalid")
+    echo "$(date -Is) AGENTD_SSH_KEY not available from Vault: tick repeated without it (direct-ssh terminals use their warpgate fallback)" >>"$LOG"
+    tick >>"$LOG" 2>&1
+    rc=$? ;;
+esac
+exit $rc

cron/agentd-tick.sh

@@ -0,0 +1,12 @@
+#!/usr/bin/env bash
+# oregon cron entry (*/2) -- ROLLBACK ONLY since PAD-11: the dispatcher runs on the LXC agentd-01
+# (cron/agentd-tick-lxc.sh). Never run both: two dispatchers double-dispatch. See docs/runbook.md "回滚到 oregon".
+# One agentd tick under server-access exec (Plane tokens in the process env only), with a lock.
+exec flock -n -E 75 /home/ubuntu/.gwent-board/agentd.lock \
+  timeout 900 python3 -I /home/ubuntu/.local/share/server-access-client/agent-access.py \
+  --config https://nas.tail9531b.ts.net:8443/identity/agent.json exec \
+  --secret-env PLANE_OREGON_API_TOKEN=PLANE_OREGON_API_TOKEN \
+  --secret-env PLANE_WINDOW10_API_TOKEN=PLANE_WINDOW10_API_TOKEN --timeout 850 -- \
+  python3 /home/ubuntu/Documents/wjj/project/plane-agentd/cli/agentd \
+  --config /home/ubuntu/Documents/wjj/project/plane-agentd/config/agentd.oregon.yaml \
+  tick --once >> /home/ubuntu/.gwent-board/agentd-cron.log 2>&1

docs/deploy-lxc.md

@@ -0,0 +1,175 @@
+# 在精简 LXC 上部署 agentd（design §9 阶段 3）
+
+调度器跑在一台**不装任何 agent CLI** 的 Debian 12 LXC（`agentd-01`）上；所有终端（win10、oregon、mac…）都是远程启动：经 Warpgate SSH（`transport: warpgate-ssh`），或经 Tailscale 直连并以 Warpgate 兜底（`transport: direct-ssh`，PAD-12，§9）。LXC 只做调度、回帖、通知。
+
+实际部署（2026-10-21，PAD-11）：`agentd-01` = PVE CT 600，Debian 12，tailnet `100.69.78.124`，Warpgate target `agentd-01`，登录 **root**（Vault skill `ssh-agentd-01`；没有另建部署用户，agentd 以 root 运行）。下面是实际做过的步骤；从 oregon 发命令都经 `agent-access.py … run --target agentd-01 --timeout N -- bash -c '…'`。
+
+## 1. 系统包
+LXC 交付时已有 `python3`（3.11）、`python3-yaml`、`openssh-client`、`git`、`cron`（active）；win10 helper 需要的 `iconv gzip base64 sha256sum` 也都在。从零装：
+```bash
+apt-get update && apt-get install -y python3 python3-yaml openssh-client git cron ca-certificates
+```
+`jsonschema` 可选，没有就用内置子集校验器。`direct-ssh` 要用的 `ssh`、`ssh-agent`、`ssh-add`、`ssh-keygen` 都在 `openssh-client` 包里（Debian 12：OpenSSH 9.2）。
+
+## 2. 网络
+| 去向 | 用途 |
+|---|---|
+| `nas.tail9531b.ts.net:8443`（tailnet） | Agent Vault 身份配置 `identity/agent.json`（含 vault token 与 Warpgate 主机公钥）→ LXC 已在 tailnet |
+| Vault（身份配置里的地址） | 取 Plane token、Warpgate 密码 |
+| Warpgate SSH `115.28.210.101:2222` | `warpgate-ssh` 终端的中继，以及 `direct-ssh` 的兜底（LXC → win10 / oregon） |
+| tailnet：win10 `100.123.209.19:22`、oregon `100.127.58.78:22` | `direct-ssh` 直连（PAD-12，§9） |
+| `https://plane.kappiee.com` | Plane API |
+| `https://git.kappiee.com` | `git pull`（公开仓库，无需凭据） |
+
+## 2.1 到 NAS 的直连路径（OPNsense 静态端口 NAT，2026-10-21）
+PVE 站点是双层 NAT（OPNsense → 192.168.1.1），默认的随机源端口会让 tailscale 打洞失败、只能走 DERP（此前取 vault 身份端点 30/30 超时）。修法：
+- **OPNsense 出站 NAT 规则**（用户在 GUI 里添加，因为 opnsense 中继账号 kappiee 无权限，我们只读）：描述 `Tailscale agentd-01 static-port`，接口 WAN，源 `10.77.10.249/32`，目标 any，协议 UDP，勾选"静态端口"（`staticnatport 1`）。从 kpe-edge-01 的 `Tailscale edge static-port`（10.77.10.110/32）规则克隆而来。加规则后要清掉旧状态（Firewall → Diagnostics → States，按源 IP）。
+- **LXC tailscaled 用 UDP 41642**：`/etc/default/tailscaled` 里 `PORT="41642"`（备份 `.bak-20261021144838`），必须与 edge 的 41641 不同，否则两台机器的静态端口映射会在同一 NAS 端点上撞元组。
+- **验证**（LXC 上）：`tailscale ping 100.89.188.28` 应得 `pong from nas … via 223.102.82.206:1549`（不是 `via DERP`）；`tailscale netcheck` 应 `UDP: true`、`MappingVariesByDestIP: false`；身份端点循环（注意 `agent.json` 含 bearer token，只看状态码和耗时）：`for i in $(seq 30); do curl -s -o /dev/null -w '%{http_code} %{time_total}\n' --max-time 15 https://nas.tail9531b.ts.net:8443/identity/agent.json; done`。2026-10-21 实测 30/30 成功、中位数 0.089 s，ping 5 ms。
+- **回滚**：在 OPNsense GUI 删掉该规则（Firewall → NAT → Outbound）并清状态；LXC 上还原 `/etc/default/tailscaled.bak-20261021144838` 后 `systemctl restart tailscaled`。回滚后路径退回 DERP，调度器应同时切回 oregon（runbook）。
+
+## 3. Agent Vault 客户端
+只复制程序文件（客户端没有落盘的 token 缓存；oregon 目录里的 `agent-service-check.py` 是运维工具，不需要）：
+```bash
+# 在 oregon 上：两个文件 base64 后经中继写到 LXC（一次 run 调用，sha256 与 oregon 一致）
+/root/.local/share/server-access-client/agent-access.py          # 700 root
+/root/.local/share/server-access-client/agent-access-askpass.py  # 700 root，必须与 agent-access.py 同目录且可执行
+# 在 LXC 上验证：
+cd /root/.local/share/server-access-client
+python3 -I agent-access.py --config https://nas.tail9531b.ts.net:8443/identity/agent.json status   # "authenticated": true
+python3 -I agent-access.py --config https://nas.tail9531b.ts.net:8443/identity/agent.json list     # 含 win10、oregon、agentd-01
+```
+- 身份是 tailnet 上的 URL（`--config https://…/identity/agent.json`），不落盘。
+- **Warpgate known_hosts 不用配**：客户端每次用身份配置里的 `warpgate.host_key` 生成临时 known_hosts。LXC → oregon 一次中继约 9 s。
+
+## 4. 代码与配置
+```bash
+cd /opt/plane-agentd && git pull          # 仓库已 clone 在 /opt/plane-agentd（https://git.kappiee.com/kappiee/plane-agentd.git）
+mkdir -p /var/lib/plane-agentd /var/log/plane-agentd && chmod 700 /var/lib/plane-agentd
+python3 -m unittest && python3 cli/agentd check-config
+```
+不需要本地配置文件，全部在仓库里（部署可复现）：
+- `config/agentd.yaml` = agentd-01 的实时配置：`state_dir: /var/lib/plane-agentd`（状态、心跳、锁），`log_dir: /var/log/plane-agentd`（`agentd.log`），`helpers_dir: tools/win10`，`vault.client: ~/.local/share/server-access-client/agent-access.py`（root 下即 `/root/...`）。终端 win10（direct-ssh + 兜底 warpgate-ssh/windows/schtasks）和 oregon（direct-ssh + 兜底 warpgate-ssh/linux/nohup，`run_root: /home/ubuntu/.gwent-board/agentd-runs`，`bins.claude`）；没有 local 终端。`agentd.ssh_key: AGENTD_SSH_KEY`，`agentd.ssh_dir: /run/plane-agentd/ssh`（ControlMaster 套接字和生成的 known_hosts，agentd 自己建，权限 0700）。
+- `tools/win10/win10-ps.sh`、`win10-put.sh`：win10 helper（vault 客户端目录取 `AGENT_ACCESS_HOME`，默认 `~/.local/share/server-access-client`；身份取 `AGENT_ACCESS_CONFIG`）。相对 `helpers_dir` 按仓库根解析。
+- `config/protocols/gwent-board-protocol.md`：看板协议（拼进每轮 prompt，LXC 本机读取）。来源是 gwent 仓库 `design/board-protocol.md`，**协议更新时复制过来提交**，然后 LXC `git pull`。
+- `config/agentd.oregon.yaml`：oregon 作调度器时的配置（回滚用，见 runbook）。
+- 相对路径（`state_dir`/`log_dir`/`helpers_dir`/`protocol_file`）都按仓库根解析；远程终端的 `repo`/`run_root`/`bins` 必须是远程绝对路径。
+
+## 5. 冒烟与切换（实际顺序）
+1. LXC 上手工跑一次（oregon cron 仍在跑，LXC 未排程；先确认 Plane 上没有 Todo 项可派，否则两个调度器会同时派单）：
+   ```bash
+   cd /root/.local/share/server-access-client && python3 -I agent-access.py --config https://nas.tail9531b.ts.net:8443/identity/agent.json exec \
+     --secret-env PLANE_OREGON_API_TOKEN=PLANE_OREGON_API_TOKEN --secret-env PLANE_WINDOW10_API_TOKEN=PLANE_WINDOW10_API_TOKEN \
+     --timeout 250 -- python3 /opt/plane-agentd/cli/agentd tick --once      # 空闲：无输出，heartbeat errors=[]
+   ```
+2. oregon：备份 crontab（`~/.gwent-board/crontab.backup-20261021-pad11`），删掉 `agentd-tick.sh` 那一行；`flock ~/.gwent-board/agentd.lock true` 等正在跑的 tick 结束。
+3. 状态：`~/.gwent-board/agentd/agentd-state.json`（oregon）→ `/var/lib/plane-agentd/agentd-state.json`（LXC，600；base64 经中继写入，比对 sha256）。在途运行的 handle 是终端上的绝对路径，新调度器照常 poll/collect。
+4. LXC 上 root 的 crontab（只有这一行）：
+   ```
+   */2 * * * * /opt/plane-agentd/cron/agentd-tick-lxc.sh
+   ```
+   `cron/agentd-tick-lxc.sh` = 拿 `/var/lib/plane-agentd/cron.lock`（`flock -n -E 75`）后执行 `timeout 900 python3 -I /root/.local/share/server-access-client/agent-access.py --config <identity> exec --secret-env PLANE_OREGON_API_TOKEN=… --secret-env PLANE_WINDOW10_API_TOKEN=… --secret-env AGENTD_SSH_KEY=AGENTD_SSH_KEY --timeout 850 -- python3 /opt/plane-agentd/cli/agentd --config /opt/plane-agentd/config/agentd.yaml tick --once`，输出追加到 `/var/log/plane-agentd/cron.log`；Vault 取不到 `AGENTD_SSH_KEY` 时不带它重跑一次（§9）。
+   注意：经中继装 crontab 时别在 `set -e` 下写 `(crontab -l | grep -v …; echo …) | crontab -`——没有旧 crontab 时 grep 返回 1，子 shell 提前退出，装进去的是空 crontab。直接 `printf '%s\n' '<行>' | crontab -`。
+5. 验证：`/var/lib/plane-agentd/heartbeat.json` 的 `at` 每 2 分钟更新。
+6. 冒烟经 LXC 调度器：在 oregon 上 `cli/agentd smoke GWENT oregon --watch`（`--watch` = 只建项并等待，不在本机 tick；oregon 上 `cli/agentd` 默认读 `config/agentd.yaml`，CLI 只需要 Plane token）。
+
+- token 只经 `exec --secret-env` 进 agentd 进程内存；agentd 起的中继子进程环境里去掉了所有 `token_key`（`Config.scrubbed_env()`）。
+- `--secret-env` 的完整列表：`cli/agentd check-config` 最后一行。
+- **任何时候只能有一个调度器排程**（oregon 的 crontab 里不能有 `agentd-tick.sh`）。
+
+## 6. 日志与心跳（LXC）
+| 路径 | 内容 |
+|---|---|
+| `/var/log/plane-agentd/agentd.log` | 每 tick 的动作、错误 |
+| `/var/log/plane-agentd/cron.log` | cron 输出（空闲时为空；`skip: another tick holds …` 无害） |
+| `/var/lib/plane-agentd/heartbeat.json` | `at` 超过 10 分钟不更新 = 调度器没在跑；`down` = 本 tick 中继不通而跳过的终端；`launching` = 启动未确认的运行 |
+| `/var/lib/plane-agentd/agentd-state.json` | 全部工作项状态 |
+| `/var/lib/plane-agentd/agentd-state.json.lxc-smoke` | 切换前手工冒烟 tick 产生的空状态（可删） |
+| 终端上 `<run_root>/<run_id>/` | 每轮的 `prompt.md run.sh/run.ps1 pid.txt launch.pid exit.txt stderr.txt` + 执行器输出；重新启动的运行在 `<run_id>-r2/` |
+
+## 7. 加一个终端（例：mac）
+1. Warpgate 加 SSH target（如 `mac`），给 `agent` 身份授权；`agent-access.py … run --target mac --timeout 30 -- bash -c 'echo ok; uname'` 能通。
+2. 在终端上装好 agent CLI 并登录（Claude Code / Codex），clone 工作仓库；记下二进制绝对路径。
+3. 配置：
+   ```yaml
+   mac:
+     transport: warpgate-ssh
+     target: mac
+     os: darwin            # linux | darwin；darwin 用 perl POSIX::setsid 脱离会话（无 setsid 命令）
+     launch: nohup
+     repo: /Users/me/src/gwent
+     run_root: /Users/me/.agentd-runs
+     bins: {claude: /Users/me/.local/bin/claude}
+     plane_user: mac
+     token_key: PLANE_MAC_API_TOKEN
+     concurrency: 1
+     executors: [claude-code]
+   ```
+   远程 shell 是目标账户的登录 shell（bash/zsh 均可，agentd 发的是 `bash -c '<script>'`）；终端需有 `bash base64 nohup pgrep pkill`（Linux 另需 `setsid`，macOS 需 `perl`）。
+4. Plane：建用户 `mac`，token 存 Vault 键 `PLANE_MAC_API_TOKEN`；项目 `routing.by_assignee` 加 `mac: mac`，`default_executor` 加 `mac: claude-code`；`cron/agentd-tick-lxc.sh` 加 `--secret-env PLANE_MAC_API_TOKEN=PLANE_MAC_API_TOKEN`。
+5. 提交、推送、LXC `git pull`；`cli/agentd check-config`，然后 `cli/agentd smoke GWENT mac --watch`。
+
+## 8. Webhook 接收端（PAD-7）
+1. 网络：Plane VM（PVE VM 506 plane-02，10.77.10.69，prime-cli 安装；没有 VM 505）与 LXC 10.77.10.249 同在 10.77.10.0/24，接收端只听 `10.77.10.249:8790`（LXC 内无防火墙）。**Plane 侧前提**：Plane 默认拒绝内网 webhook 地址（"The webhook URL is not allowed. Provide a publicly reachable HTTP or HTTPS URL."），要在 Plane 的 api（创建时检查）与 worker（每次投递检查）容器环境里设 `WEBHOOK_ALLOWED_IPS=10.77.10.249/32`（逗号分隔 IP/CIDR）并重建相关容器（design §10）。Plane 3.3.3 没有 compose override 文件：prime-cli 由 `/opt/plane/plane.env` 渲染 `/var/plane/docker-compose.yml`，两处都已有空的 `WEBHOOK_ALLOWED_IPS`（compose 里每个服务一行 `WEBHOOK_ALLOWED_IPS: ""`）。做法：备份两个文件为 `.bak-pad7-<时间戳>` → 把 `plane.env` 的 `WEBHOOK_ALLOWED_IPS=` 与渲染 compose 里的空值改为 `10.77.10.249/32` → `docker compose -p plane -f /var/plane/docker-compose.yml up -d --no-deps api worker webhook-consumer outbox-poller automation-consumer beat-worker`（先 `--dry-run`，确认只重建这 6 个；商业版的 v2 投递走 webhook-consumer/worker，创建时校验在 api；不要用 prime-cli，它要 TTY。compose 里 `WEBHOOK_ALLOWED_HOSTS: ${WEB_URL}` 不带 env 文件时为空，prime-cli 起的容器也是空值，属现状）。**状态（2026-10-22）：已应用**（备份 `*.bak-pad7-20261022-011453`，6 个容器 `WEBHOOK_ALLOWED_IPS=10.77.10.249/32`），webhook `c039c471-e78a-49c9-b009-5cf77a389ad6` 已建并启用（v2；签名与接收端一致，见 design §10），接收端 active/enabled。纯 http 本身是允许的（http/https 都行）。
+2. 密钥：Plane 建 webhook 时生成、只显示一次 → `tools/plane-webhook.py create http://10.77.10.249:8790/plane` 直接写进 Vault `PLANE_WEBHOOK_SECRET`（建出来是停用状态；在 oregon 上以 kappiee 管理员会话运行，见 runbook "Webhook"）。核对长度：`agent-access.py … exec --secret-env S=PLANE_WEBHOOK_SECRET -- python3 -c 'import os;print(len(os.environ["S"]))'`。
+3. 代码与服务（LXC）：
+   ```bash
+   cd /opt/plane-agentd && flock /var/lib/plane-agentd/cron.lock git pull
+   cp systemd/plane-agentd-webhook.service /etc/systemd/system/ && systemctl daemon-reload
+   systemctl enable --now plane-agentd-webhook
+   curl -s http://10.77.10.249:8790/healthz       # {"ok": true, ...}
+   tail /var/log/plane-agentd/webhook.log           # listening on http://10.77.10.249:8790/plane (tenants: GWENT)
+   ```
+   单元文件复用 cron 脚本的 vault 客户端调用方式：`python3 -I /root/.local/share/server-access-client/agent-access.py --config <identity> exec --secret-env PLANE_WEBHOOK_SECRET=PLANE_WEBHOOK_SECRET --timeout 315360000 -- python3 /opt/plane-agentd/cli/agentd … webhook-serve --bind 10.77.10.249 --port 8790`，`Restart=always`。tick 由接收端经 `systemd-run --scope … cron/agentd-tick-lxc.sh` 拉起（同一把 `cron.lock`，`flock -n -E 75`）。
+4. 启用 Plane 侧：`tools/plane-webhook.py patch <id> is_active=true`，然后 `systemctl restart plane-agentd-webhook`（若第 2 步之后密钥变过）。
+5. 验证：随便改一个 GWENT 工作项 → `webhook.log` 出现 `decision=tick` 和 `tick start`/`tick done`；错签名请求得 401 且没有 tick；`journalctl -u cron` 里 cron 仍每 2 分钟一次；`systemctl restart plane-agentd-webhook` 后 `/healthz` 恢复。
+
+## 9. 直连 SSH（PAD-12）
+LXC 经 Tailscale 直连终端的 sshd，不经过 Warpgate（design §9"终端通道"）。每个终端单独配置，Warpgate 留作兜底。前提：终端在 tailnet 上，LXC（`100.69.78.124`）能连到它的 22 端口。实测：win10 `100.123.209.19`（Windows OpenSSH，账号 administrator，直连 RTT 27 ms），oregon `100.127.58.78`（账号 ubuntu，RTT 233 ms）。
+
+**上线顺序**：`cron/agentd-tick-lxc.sh` 带着 `--secret-env AGENTD_SSH_KEY=AGENTD_SSH_KEY`。agent-access.py 没有可选密钥，取不到键时报 `credential_unavailable`，整个 exec 失败；cron 脚本认出这种情况后会去掉这个参数再跑一次（cron.log 记一行 `AGENTD_SSH_KEY not available from Vault …`），所以先 pull 也不会停派单，只是直连终端一直走兜底。仍然建议第 1–3 步做完再 pull，免得每个 tick 多取一次 Vault。
+
+1. **生成专用密钥**（在 LXC 上；私钥只进 Vault，不留在磁盘上）：
+   ```bash
+   umask 077; d=$(mktemp -d -p /dev/shm)   # tmpfs：私钥不落到磁盘（磁盘上 shred 在日志式/写时复制文件系统上不可靠）
+   ssh-keygen -q -t ed25519 -N '' -C plane-agentd@agentd-01 -f "$d/agentd"
+   cat "$d/agentd.pub"            # 公钥，下面两步要用
+   ```
+   私钥 `$d/agentd` 原样（多行文本）写进 Vault 键 **`AGENTD_SSH_KEY`**：在 Vault UI 里粘贴，或者用与 `tools/plane-webhook.py` 的 `vault_store()` 相同的接口（`POST /v1/credentials`，agent 身份），读文件直接写，不打印。值里的换行被转义成 `\n` 也可以，agentd 会还原。写完删掉：`rm -rf "$d"`（在 tmpfs 上，删掉即可）。核对时只看长度：
+   `agent-access.py --config <identity> exec --secret-env K=AGENTD_SSH_KEY --timeout 60 -- python3 -c 'import os;print(len(os.environ["K"]))'`（ed25519 私钥约 400 字符）。
+2. **oregon**（经 Warpgate 操作）：在 `~ubuntu/.ssh/authorized_keys` 追加一行（目录 700、文件 600）：
+   ```
+   restrict,from="100.69.78.124" ssh-ed25519 AAAA…（上一步的公钥） plane-agentd@agentd-01
+   ```
+   `restrict` 关掉所有转发、pty 和 `~/.ssh/rc`（agentd 用 `-T`，不需要 pty），不必再单独写 `no-port-forwarding` 等；`from=` 只放行 LXC 的 tailnet 地址。sshd 不认识 `restrict`（OpenSSH 7.2 之前）时会把整行判为无效：认证失败、终端走兜底，不会放宽权限；第 5 步的 ping 能看出来（`sshd -V` 或 Windows 上 `sshd.exe -V` 看版本）。可选的再一层：Tailscale ACL 只允许 LXC（`100.69.78.124` 或它的 tag）访问 win10/oregon 的 `:22`，这样密钥泄露了也没用。
+3. **win10**（经 Warpgate，`tools/win10/win10-ps.sh`）：Administrator 属于 Administrators 组，所以 sshd 读的是 `C:\ProgramData\ssh\administrators_authorized_keys`，不是用户目录下的文件。用 ASCII 追加同样的一行（别覆盖已有的行），再把 ACL 收紧到只有 Administrators 和 SYSTEM（用 SID，不受系统语言影响），否则 sshd 会忽略这个文件：
+   ```powershell
+   $f = 'C:\ProgramData\ssh\administrators_authorized_keys'
+   $line = 'restrict,from="100.69.78.124" ssh-ed25519 AAAA… plane-agentd@agentd-01'
+   if ((Test-Path $f) -and (Get-Content -Raw $f) -and -not (Get-Content -Raw $f).EndsWith("`n")) { $line = "`r`n" + $line }   # 原文件末尾没换行时，别把新行粘到最后一把钥匙后面
+   Add-Content -Path $f -Encoding ascii -Value $line
+   icacls $f /inheritance:r /grant '*S-1-5-32-544:F' /grant '*S-1-5-18:F'
+   ```
+   顺便确认 sshd 的默认 shell 是 cmd.exe（`HKLM:\SOFTWARE\OpenSSH` 下没有 `DefaultShell`，或它指向 cmd.exe）。agentd 发的远程命令是 `powershell.exe -NoProfile -NonInteractive -EncodedCommand <引导>`，真正的脚本经 stdin 传过去。外层如果被改成 PowerShell，stdin 可能转不到内层；第 5 步的 ping 能直接发现这个问题。
+4. **钉住的主机公钥**已写在 `config/agentd.yaml` 的 `terminals.<t>.ssh.host_key`，经两个通道核对过：Warpgate 读终端上的 `ssh_host_ed25519_key.pub`，与 tailnet 上 `ssh-keyscan -t ed25519` 的结果一致。新加终端照此办理：
+   ```bash
+   ssh-keyscan -t ed25519 <tailnet ip>                                          # LXC 上，直连
+   agent-access.py … run --target <t> -- cat /etc/ssh/ssh_host_ed25519_key.pub   # 第二通道（Windows：type C:\ProgramData\ssh\ssh_host_ed25519_key.pub）
+   ```
+   两边一致才能写进配置。agentd 每次调用都用它生成 known_hosts（`<ssh_dir>/known_hosts.<终端>`，别名 `agentd-<终端>`），全局和用户的 known_hosts 都不读。
+5. **上线与验证**（在 LXC 上；拿着 cron 锁做，期间 cron 和 webhook 的 tick 会跳过，webhook 会自动重试）：
+   ```bash
+   cd /root/.local/share/server-access-client
+   flock /var/lib/plane-agentd/cron.lock bash -c '
+     git -C /opt/plane-agentd pull &&
+     for t in win10 oregon; do
+       python3 -I agent-access.py --config https://nas.tail9531b.ts.net:8443/identity/agent.json exec \
+         --secret-env AGENTD_SSH_KEY=AGENTD_SSH_KEY --timeout 120 -- \
+         python3 /opt/plane-agentd/cli/agentd ping $t --transport direct-ssh -n 3
+     done'
+   ```
+   预期每个终端打印 3 行 `term=<t> via=direct op=ping ms=… rc=0` 和 `ping n: … ms PONG=…`。第 1 次要握手（win10 约 0.5–1.5 s，oregon 约 2–3 s），第 2、3 次复用连接，应明显更快。win10 若是 `timeout`，多半是 stdin 没到内层 PowerShell，见第 3 步。`--transport warpgate-ssh` 可顺带确认兜底仍然通（约 9 s）。
+   `cli/agentd check-config` 会多出 `direct-ssh …` 行，`cron/exec secret args` 末尾多了 `--secret-env AGENTD_SSH_KEY=AGENTD_SSH_KEY`。
+6. **观察**：下一个 tick 后看 `/var/log/plane-agentd/remote.log`，应该只有 `via=direct` 行；出现 `via=fallback` 说明直连失败了，同时 `agentd.log` 会有一行原因。在 oregon 上 `cli/agentd smoke GWENT win10 --watch`，量"建项 → `[worker] 第 1 轮开始`"，目标 ≤ 15 s（原来 91 s）。
+7. **回滚**：把终端的 `transport` 改回 `warpgate-ssh`（提交、推送、LXC 上 `git pull`；`ssh`/`fallback` 可以留着）。cron 脚本里的 `--secret-env AGENTD_SSH_KEY=…` 可以保留：agentd 总会把它从子进程环境里去掉。Vault 键被删掉也不会停 tick（脚本会不带它重跑），但每个 tick 会多取一次 Vault，届时最好也删掉这一行。紧急情况下可在 LXC 上 `git -C /opt/plane-agentd checkout 83bc0f2 -- config/agentd.yaml`（旧配置，全部走 Warpgate，新代码照常能用）。要彻底撤销直连授权：删掉两台终端 authorized_keys 里那一行，再在 LXC 上关掉已经建好的 ControlMaster 连接（删 authorized_keys 不会断开它们，紧挨着的 webhook tick 能让它们活过 60 s）：`for s in /run/plane-agentd/ssh/*-*; do [ -S "$s" ] && ssh -O exit -o ControlPath="$s" x; done`。
+8. **轮换密钥**：照第 1 步生成新钥，先把新公钥加到两台终端，再更新 Vault `AGENTD_SSH_KEY`，然后删掉旧公钥的行，再照第 7 步关掉已有的 ControlMaster 连接（否则它们空闲 60 s 后才退出）。之后新连接就用新钥。改了某个终端钉住的 `host_key` 不需要这一步：套接字名里带着 pin 的指纹，新 pin 自动用新连接。

docs/design.md

@@ -3,22 +3,22 @@
 Plane 看板驱动的通用 agent 调度服务：任何人/agent 在 Plane 建工作项并指派给一个"终端"，agentd 在该终端起对应的 CLI agent（Codex、Claude Code、脚本…）执行，把结果贴回工作项并流转状态。调度是确定性代码，模型只出现在执行器和总指挥里。
 
 ## 1. 目标与非目标
-- 目标：多项目隔离（一份配置一段）、多终端（本机/经 Warpgate 的远程机）、多 CLI（执行器适配器）、多轮（线程持久、进程不常驻）、失败可解释可重试、凭据只经 Vault 注入内存、任何会话可接手。
+- 目标：多项目隔离（一份配置一段）、多终端（本机/经 Warpgate 或 Tailscale 直连的远程机）、多 CLI（执行器适配器）、多轮（线程持久、进程不常驻）、失败可解释可重试、凭据只经 Vault 注入内存、任何会话可接手。
 - 非目标：不是工作流引擎（无 DAG/分支，顺序交给 Plane "blocked by"——**TODO：阶段 1 尚未实现**，调度不读 blocked-by 关系，被阻塞的 Todo 项照样会被拾起）；不替代 Plane/Temporal 的界面；不做账号系统；不存密钥。
 
 ## 2. 架构
 ```
-Plane ──webhook v2（HMAC）/ 轮询兜底──► agentd（接收端 + 调度核心 + 执行器适配器）
+Plane ──webhook（HMAC）即时触发 / cron 轮询兜底──► agentd（接收端 + 调度核心 + 执行器适配器）
                                         │ 按 assignee→终端、label→执行器/模型，串行组，重试，超时，心跳
                         ┌───────────────┼───────────────┐
                   codex（远程/本机）  claude-code（本机/远程）   shell
                         └───────────────┴───────────────┘
-                 执行器只产出"最后消息 + 产物路径"；agentd 统一回帖、改状态、通知（ntfy）
+                 执行器只产出"最后消息 + 产物路径"；agentd 统一回帖、改状态、通知（notify 钩子，暂为空操作）
 ```
 阶段 3 之后调度核心搬到 Temporal（workflow = 一个工作项的生命周期，activity = 起执行器/等结果/回帖），用 Temporal UI 看运行与重试。
 
 ## 3. 概念与契约
-- **终端（terminal）**：怎么到达一台机器 — `transport`（local | warpgate-ssh）、`os`、`launch`（subprocess | schtasks | nohup）、`repo`、`plane_user`（回帖身份，对应 Vault 键）、`concurrency`、`capabilities`、可用 `executors`。
+- **终端（terminal）**：怎么到达一台机器 — `transport`（local | warpgate-ssh | direct-ssh，§9）、`os`、`launch`（subprocess | schtasks | nohup）、`repo`、`plane_user`（回帖身份，对应 Vault 键）、`concurrency`、`capabilities`、可用 `executors`。
 - **执行器（executor）**：某个 CLI 的适配器 — 起一轮的命令模板、续线程命令模板、`resume: true|false`、可用模型、默认推理强度、沙箱参数、最后消息取法、线程 ID 取法。
 - **路由（routing）**：`by_assignee`（Plane 用户 → 终端）、`by_label`（`exec:*`、`model:*`、`effort:*`、`needs-*` 串行组、`cap:*` 能力、`budget:*` 时限）、每终端默认执行器。
 - **工作项契约**：正文 = 任务书；评论 = 消息；执行器最后消息首行 ∈ `[review] | [question] | [blocked] | [progress]`；agentd 回帖格式固定（轮次、run id、退出码、日志路径）。
@@ -34,7 +34,7 @@ Plane ──webhook v2（HMAC）/ 轮询兜底──► agentd（接收端 + 调
 | 环境类（焦点丢失/弹窗/断网） | 当 blocked 处理，暂停同串行组的派发 | Waiting |
 | 运行中工作项被移出 In Progress | 结束任务进程，回帖"运行已取消：工作项状态被改为 X"，释放终端槽位/串行组，保留线程 | 不改 |
 | 连续 `max_progress_rounds`（默认 12）轮 `[progress]` | 回帖说明，不再续跑 | Waiting → 应答者 |
-| agentd 自身 | cron 拉起；心跳 10 min 无更新 → ntfy 报警 | — |
+| agentd 自身 | cron 拉起（webhook 只是提前触发）；心跳 10 min 无更新 → 报警（notify 钩子，未实现） | — |
 回到 Todo 的工作项无条件重新拾起；逻辑性失败永不自动重试。
 
 阶段 1 的具体取舍（均有单测）：
@@ -50,7 +50,7 @@ plane: {base_url: https://plane.kappiee.com, workspace: kappiee}
 vault: {client: ~/.local/share/server-access-client/agent-access.py, config: https://nas.tail9531b.ts.net:8443/identity/agent.json}
 terminals:
   win10:  {transport: warpgate-ssh, target: win10, os: windows, launch: schtasks, repo: 'C:\Users\Administrator\Documents\GwentAssistantProject\gwent', plane_user: window10, token_key: PLANE_WINDOW10_API_TOKEN, concurrency: 1, capabilities: [android, game], executors: [codex, powershell]}
-  oregon: {transport: local, os: linux, launch: subprocess, repo: /home/ubuntu/Documents/wjj/project/gwent, plane_user: oregon, token_key: PLANE_OREGON_API_TOKEN, concurrency: 2, executors: [claude-code, codex, shell]}
+  oregon: {transport: warpgate-ssh, target: oregon, os: linux, launch: nohup, repo: /home/ubuntu/Documents/wjj/project/gwent, run_root: /home/ubuntu/.gwent-board/agentd-runs, plane_user: oregon, token_key: PLANE_OREGON_API_TOKEN, concurrency: 1, executors: [claude-code]}
 executors:
   codex:       {run: 'codex exec --json -C {repo} -s workspace-write -m {model} -c model_reasoning_effort="{effort}" {tier} -o {out} -', resume: 'codex exec resume {thread} --json -m {model} -c model_reasoning_effort="{effort}" {tier} -o {out} -', thread_from: events.jsonl:thread.started.thread_id, last_message: '{out}', models: [gpt-6-astra, gpt-6.1-sol], default_effort: medium}
   claude-code: {run: 'claude -p --model {model} --output-format json --permission-mode {perm} -', resume: 'claude -p --resume {thread} --model {model} --output-format json -', thread_from: json:session_id, models: [opus, sonnet, haiku]}
@@ -58,7 +58,6 @@ executors:
 projects:
   GWENT:
     id: 79cb5a60-aa4f-4138-9864-1dd70426e665
-    webhook_secret_key: PLANE_WEBHOOK_GWENT
     states: {todo: Todo, running: In Progress, waiting: Waiting, review: Review, done: Done}
     reviewer: oregon
     routing:
@@ -67,16 +66,20 @@ projects:
       default_executor: {win10: codex, oregon: claude-code}
     retry: {transient_max: 3, backoff_minutes: [2, 5, 10]}
     timeout_hours: 4
-    notify: {ntfy_topic: gwent, on: [waiting, review, failed]}
+webhook: {bind: 10.77.10.249, port: 8790, path: /plane, secret_key: PLANE_WEBHOOK_SECRET,
+          tick_command: [systemd-run, --scope, --quiet, --collect, /opt/plane-agentd/cron/agentd-tick-lxc.sh],
+          debounce_seconds: 3, retry_seconds: 15, retry_window_seconds: 120}
+notify: {kind: none}     # 钩子：核心每次改状态调用 Agentd.notify()，none = 空操作
 ```
 
 ## 6. 仓库布局
 ```
 agentd/            调度核心、plane 客户端、vault 注入、状态存储（json/sqlite）、心跳
 agentd/executors/  codex.py claude_code.py shell.py（适配器契约见 docs/executor-contract.md）
-agentd/transports/ local.py warpgate_ssh.py（含 Windows schtasks 启动）
+agentd/transports/ local.py warpgate_ssh.py（远程 Linux/macOS：bash -c；Windows：PowerShell + schtasks；Warpgate 通道）direct_ssh.py（Tailscale 直连通道，PAD-12）
 agentd/web/        配置页面（阶段 2）：读写 YAML + schema 校验 + 连通/冒烟测试 + 最近运行
-cli/               agentd new|show|comment|state|assign|smoke
+cli/               agentd tick|new|show|comment|state|assign|smoke|webhook-serve|check-config
+systemd/           plane-agentd-webhook.service（LXC 上的 webhook 接收端）
 config/            agentd.yaml（示例 + 实际）、schema.json
 docs/              design.md executor-contract.md runbook.md
 ```
@@ -84,14 +87,34 @@ docs/              design.md executor-contract.md runbook.md
 ## 7. 阶段
 1. **核心**：配置驱动调度（轮询）、codex-remote(win10) + claude-local(oregon) 两个执行器、线程续接与重建、失败说明与重试、心跳、CLI；gwent 项目作为首个租户切换过来（替换现有 board-worker-plane）。
 2. **配置页面**：单页，三个标签（终端/执行器/路由）+ 状态条；保存即校验、提交 Git、热加载；不显示密钥。
-3. **Webhook + ntfy**：Plane webhook v2 签名校验、即时触发；ntfy 通知 Waiting/Review/失败；轮询保留兜底。前提：Plane VM 进 tailnet 或接收端放在可达位置。
+3. **Webhook + 通知**：Plane webhook 签名校验、即时触发，轮询保留兜底——**PAD-7 已实现**（§10）。通知 Waiting/Review/失败：只留了空操作钩子 `notify: {kind: none}`（没有 ntfy 服务）。
 4. **Temporal**：调度核心迁为 workflow/activity，获得重试/超时/历史/UI；agentd 变为 worker + webhook 入口。
 5. **Gitea 联动**：执行器产物以分支/PR 提交并回帖链接；可选 Gitea webhook 回写 Plane。
 
 ## 8. 安全
-配置里只有 Vault 键名；token 由 `agent-access.py exec --secret-env` 注入 agentd 进程；执行器不持有 Plane 凭据、不能直接改工作项；webhook 按项目签名校验；页面仅 tailnet 可达，登录走既有网关。
+配置里只有 Vault 键名；token 由 `agent-access.py exec --secret-env` 注入 agentd 进程；执行器不持有 Plane 凭据、不能直接改工作项；webhook 是工作区级的一个密钥（Vault 键 `PLANE_WEBHOOK_SECRET`，只经 `--secret-env` 进接收端进程，不传给子进程），HMAC 校验失败 401，只认配置里的项目；接收端只听内网地址；页面仅 tailnet 可达，登录走既有网关。
 
 ## 9. 部署位置
-- Plane VM（plane-01）只跑 Plane：它是设备化节点（固定版本、每日备份停写路径），不放 agentd。
+- 网络（2026-10-21）：LXC 到 NAS 的 tailnet 路径依赖 OPNsense 的静态端口出站 NAT 规则 + LXC tailscaled 固定 UDP 41642，已直连（详见 docs/deploy-lxc.md §2.1）；曾因 DERP 导致调度器暂回 oregon，已于同日切回 LXC。
+- Plane VM（Plane 3.3.3，PVE VM 506 plane-02）只跑 Plane：它是设备化节点（固定版本、每日备份停写路径），不放 agentd。
 - 阶段 1–2：agentd 跑在 oregon（已有 Vault 客户端、Warpgate 中继脚本、Claude Code、现有 worker 状态），cron 拉起；配置页面同进程，只监听 tailnet 地址。
-- 阶段 3 起：迁到 PVE 上的专用小节点 `agentd-01`（1–2 vCPU / 2 GB，tailnet + 内网 10.77.10.0/24，与 Plane/Temporal 同网段，webhook 走内网），装 Vault 客户端、中继脚本、Claude Code 与 Codex CLI，使其同时也是一个执行终端（跑核对/验收类任务）。对外访问经 Warpgate HTTP target，复用既有登录。
+- 阶段 3 起（**2026-10-21 PAD-11 已切换**）：迁到 PVE 上的精简 LXC `agentd-01`（CT 600，root，仓库 `/opt/plane-agentd`，状态 `/var/lib/plane-agentd`，日志 `/var/log/plane-agentd`；1 vCPU / 1–2 GB，tailnet + 内网 10.77.10.0/24，与 Plane/Temporal 同网段，webhook 走内网）。**LXC 上不装任何 agent CLI**，只有 Python 3.10+、PyYAML、Agent Vault 客户端、ssh 客户端；所有终端（win10、oregon、mac…）一律远程启动（`transport: warpgate-ssh` 或 `direct-ssh`，见下文"终端通道"），LXC 只负责调度、回帖、通知。对外访问经 Warpgate HTTP target，复用既有登录。
+- 远程启动的通用实现（P1-6，已实现）：Linux/macOS 经 Warpgate SSH（`agent-access.py run --target <终端 target> -- bash -c …`）上传 `prompt.md`/`run.sh` 并一次调用内启动，`setsid nohup` 脱离会话（macOS 用 perl `POSIX::setsid`），轮询 `exit.txt`，从终端上的输出文件取回最后消息/线程 ID，kill 按进程组 + 会话；中继目标名由终端配置给出。oregon 已切到这条路径（调度器暂仍在 oregon，启动经 Warpgate 回到本机）。部署步骤：`docs/deploy-lxc.md`。
+- 防重复启动（PAD-11）：启动调用之前先把计划中的运行（run id、终端、handle、记账字段）作为 `launching` 标记落盘；启动调用失败/超时或 agentd 中途退出时，下一 tick 探测运行目录（`exit.txt`/`launch.pid`/`pid.txt`）：有 → 接管；无 → 同一 run id 换新目录 `<run_id>-r2` 重启（不清空旧目录）；探测不通 → 保留标记并占住终端槽位。transport 拒绝清空含 pid 文件的运行目录。某终端中继失败时本 tick 余下操作跳过该终端，其他终端照常。
+- Windows 走 Warpgate 时仍用 helper 脚本 `win10-ps.sh` / `win10-put.sh`（仓库 `tools/win10/`，写死 `--target win10`）+ 计划任务：`agent-access.py run` 把 ssh 的 stdin 接到 /dev/null（`run_remote` 里 `stdin=subprocess.DEVNULL`），cmd.exe 命令行又限 8191 字符，文件只能按 2400 字符分块上传。走 `direct-ssh` 时不用 helper 脚本。收尾读取在 Warpgate 上也已合成一次调用（脚本足够短，放得进 `-EncodedCommand`）。
+- **终端通道（PAD-12，2026-10-22）**：LXC 到每个远程终端走哪条路，按终端配置 `transport`：
+  - `warpgate-ssh`（原有，缺省）：`agent-access.py run` 经 Warpgate SSH 网关（`115.28.210.101:2222`）。每次调用都要取 Vault 凭据、与网关握手、密码认证、网关再连终端，LXC 上实测中位数约 9 s（客户端 + vault 0.8 s、网关 banner/kex 1.8 s、密码认证 2.2 s、二跳 + 命令 2–5 s）。客户端写死 `ControlMaster=no`，连接不能复用，数据只能放 argv。一次 win10 启动约 10 次串行调用（busy 检查、mkdir/清空、prompt.md 分块上传 + 重置 + 解码、run.ps1、schtasks）：GW-26-1 用了 91 s；收尾约 7 次，约 60 s。
+  - `direct-ssh`（新）：LXC 经 Tailscale 直连终端的 sshd（`ssh: {host, user, port, host_key}`）。认证用一把专用 ed25519 密钥：Vault 键 `AGENTD_SSH_KEY` 经 `--secret-env` 进 tick 进程，载入本进程私有的 ssh-agent（0700 临时目录里的套接字，`ssh-add -` 从 stdin 读）后立即从环境变量删掉，子进程拿不到；不落盘，不进 argv、日志、Plane 评论或异常信息。主机公钥钉死在配置里，`StrictHostKeyChecking=yes`，known_hosts 由钉住的公钥生成，只含这一把。`ControlMaster=auto`，`ControlPersist=60`，套接字放在只有 root 可进的 `agentd.ssh_dir`（`/run/plane-agentd/ssh`），名字是 `<host_key 指纹 8 位>-%C`：pin 改了就换新连接，不会沿用按旧 pin 校验过的主连接。密钥在 CLI 启动时就从环境变量里拿走（第一次直连之前起的子进程也继承不到）。脚本走 stdin：Windows 是一段固定的 PowerShell 引导（`-EncodedCommand`），从 stdin 读真正脚本的 base64；Linux 是固定的 bash 加载器。所以每个操作恰好一次调用，不受 argv 长度限制。win10 启动（建目录、清空并遵守 HASPID 拒绝、写 UTF-8 无 BOM 的 prompt.md/run.ps1、注册并启动计划任务）一次；收尾一次（exit、最后消息、线程 ID、stderr 尾、事件尾）加删计划任务一次。oregon 启动一次，收尾一次。busy 检查和 poll 仍各一次。实测直连路径：win10 RTT 27 ms（握手到认证 0.24 s），oregon RTT 233 ms（握手 2.2 s）。目标：派单延迟 ≤ 15 s。
+  - `fallback: warpgate-ssh`（可选，需要 `target`）：直连**在远程命令运行之前**失败时（ssh 退出 255、没有任何 stdout，且 stderr 里有 ssh 自己在行首打印的连接超时/拒绝/无路由/解析失败/认证失败/agent 错误，或者本进程没拿到密钥。远程加载器把脚本自己的 255 改报 254，所以 255 只可能来自 ssh；有 stdout 就说明命令已经跑过），这个操作改走 Warpgate 重做，本 tick 里该终端余下的操作也都走 Warpgate，`agentd.log` 记一行。**主机公钥不符（包括终端不提供 ssh-ed25519 主机密钥）绝不回退**：按安全事件处理，`agentd.log` 记 `SECURITY …` 行，该终端本 tick 标为 down。命令可能已经运行过的失败（超时、会话中途断开、认不出的 255）不会换路径重试。启动这种非幂等操作仍靠 `launching` 标记加下一 tick 的探测/接管（`resolve_launch`），不存在双启动的路径。直连的每个操作都只有一次远程调用，所以"整个操作换路径重做"不会重复任何副作用。
+  - **取舍**：直连绕过了 Warpgate 的会话录像和集中授权。访问控制改由终端自己负责：authorized_keys 只认这把专用密钥，并限定来源 `from="100.69.78.124"`（LXC 的 tailnet 地址）。用户已逐终端批准（win10、oregon，2026-10-22）。回滚：把该终端的 `transport` 改回 `warpgate-ssh`（`ssh`/`fallback` 可以留着，此时被忽略）。
+  - 观测：每次远程调用在 `<log_dir>/remote.log` 记一行（`term= via=direct|warpgate|fallback op= ms= rc=`）。部署后用 `agentd ping <终端> [--transport direct-ssh|warpgate-ssh]` 验证。步骤见 `docs/deploy-lxc.md` §9。
+
+## 10. Webhook 触发（PAD-7）
+- **形态**：`agentd webhook-serve`（`agentd/webhook.py`，只用标准库）常驻在 LXC 上（systemd `plane-agentd-webhook`），Plane 工作区级 webhook（事件 `issue` + `issue_comment`）投到 `http://10.77.10.249:8790/plane`。它不自己调度，只是**提前跑一次 cron 的 tick 脚本**：调度仍然只有一份代码、一把锁、一个状态文件；cron 每 2 分钟照跑，webhook 丢了/接收端挂了只是退回到轮询延迟。
+- **校验**：`X-Plane-Signature` = HMAC-SHA256(密钥, 请求体) 的十六进制（Plane 签 `json.dumps(payload)` 并原样发送）；常量时间比较，先比原始字节、再比重新序列化的结果。缺失/不符 → 401，不触发。
+- **路由**：`data.project` ∈ 配置的项目 → 200 并触发；否则（别的项目、删除事件没有项目）→ 202 忽略。不按事件类型或操作者过滤：agentd 自己回帖/改状态也会产生投递，结果是每次有动作的 tick 之后多一次 tick（空闲 tick 不写 Plane，不会循环）。
+- **合并与补跑**：第一条投递后等 3 s，期间的投递合并成一次 tick；tick 运行中又来的投递 → 结束后再跑一次（运行中的 tick 可能在变化之前读过 Plane）。锁被占（cron/手工 tick；脚本 `flock -n -E 75`，`cli tick` 撞上 agentd 状态锁也退出 75）→ 每 15 s 重试，最长 2 min，之后交给 cron。
+- **进程隔离**：tick 用 `systemd-run --scope` 跑在独立 cgroup 里，重启/停止接收端不会杀掉正在跑的 tick（tick 可能正在启动远程运行）。
+- **延迟构成**：投递 → tick 开始 ≈ 3 s 合并窗口；tick 本身 = vault 取 token（1–2 s）+ 读 Plane + 先 poll 每个在途运行（每个中继调用约 10 s）+ 启动（中继约 12 s；win10 上传更久）。`[worker] 第 k 轮开始` 评论在启动调用返回之后才发，所以"建项 → 开始评论"受中继成本限制，而不是触发方式。PAD-12 之后，直连终端每次调用不再是约 10 s，而是亚秒到约 2 s（§9 终端通道）。
+- **Plane 的 SSRF 防护**：Plane（≥1.4 / 商业版 3.x）拒绝解析到私有/CGNAT 地址的 webhook URL（创建时和每次投递时都查），除非 Plane 的环境里有 `WEBHOOK_ALLOWED_IPS`（CIDR 列表）或 `WEBHOOK_ALLOWED_HOSTS`。内网接收端需要在 Plane 的 api 与 worker 容器环境里设 `WEBHOOK_ALLOWED_IPS=10.77.10.249/32`（改 `/opt/plane/plane.env` 与渲染的 `/var/plane/docker-compose.yml`，重建 api/worker，见 deploy-lxc §8）。2026-10-22 已应用（同时重建了 webhook-consumer、outbox-poller、automation-consumer、beat-worker，见 deploy-lxc §8），webhook 已启用，cron 保留为兜底。Plane 建出的是 v2 webhook（`plane/webhook/bgtasks/webhook_task.py` 的 `send_webhook`，负载 `{version: v2, event, entity_type, data, …}`），签名同为 HMAC-SHA256(secret, json.dumps(payload))，`data.project` 路由照常，接收端无需改动。
+- **心跳里的跳过原因**：每个 tick 把没派出去的 Todo 项和原因写进 `heartbeat.json` 的 `skipped`（`concurrency 1/1 (GW-15)`、`serialize_group game held by GW-4`、`busy_process codex (1 running, not started by agentd)`、`relay down`、`no route for assignee kappiee`、`start failed: …`），同一项的原因变化时才写一行日志。

docs/executor-contract.md

@@ -6,13 +6,15 @@ agentd 不常驻：一轮运行比起它的那次 tick 活得久，所以执行
 | 成员 | 说明 |
 |---|---|
 | `resume: bool` | 能否续线程。`False` 时每轮都是新线程（有前序轮次时附交接摘要）。 |
-| `start(run: Run) -> handle` | 起一轮，立即返回；不得阻塞到运行结束。 |
+| `handle(run) -> handle` | 不接触终端，算出 `start(run)` 将返回的 handle（core 在启动前把它写进 `launching` 标记）。 |
+| `start(run: Run) -> handle` | 起一轮，立即返回；不得阻塞到运行结束。抛异常不代表一定没启动：core 下一 tick 用 `probe` 判断。 |
+| `probe(handle) -> [names]` | 运行目录里的启动证据（`exit.txt`/`launch.pid`/`pid.txt` 中存在的）；空 = 没启动；终端不可达抛 `TransportError`。 |
 | `poll(handle) -> "running" \| "finished"` | 便宜的完成检测。 |
 | `collect(handle) -> ExecResult` | 只在 `finished` 后调用，可重复调用（幂等）：Plane 回帖失败时下一 tick 会再 collect。 |
 | `kill(handle)` | 超时时结束整个进程树。 |
 | `forget(handle)`（可选） | collect 之后清理（如删除计划任务）。 |
 
-`Run(run_id, repo, prompt, model, effort, tier, thread_id, script)`：`thread_id` 非空即续该线程。
+`Run(run_id, repo, prompt, model, effort, tier, thread_id, script, dir_name)`：`thread_id` 非空即续该线程；`dir_name` 是运行目录名/计划任务名（缺省 = `run_id`，重新启动时为 `<run_id>-r2` 等）。
 `ExecResult(last_message, exit_code, thread_id, log_paths, tail)`：`last_message` 原样回帖，首行应为 `[review]/[question]/[blocked]/[progress]`；`thread_id` 供下一轮续接；`tail` 是失败说明里的"最后一帧"（stderr 尾部/最后事件）。
 
 执行器**不接触 Plane**：不持有 token、不改工作项。回帖、状态流转、重试、超时、线程重建都由 core 统一处理。transport 启动 runner/中继脚本时传入去掉所有 `token_key`/`webhook_secret_key` 变量的环境（`Config.scrubbed_env()`）；新 transport 必须同样处理。
@@ -24,11 +26,11 @@ agentd 不常驻：一轮运行比起它的那次 tick 活得久，所以执行
 - `kill(handle)` 既用于超时，也用于取消（工作项在运行中被移出 In Progress）；必须能在运行未结束时调用。取消后 core 会尝试 `collect(handle)` 取线程 ID，所以 `collect` 在未结束的运行上也不能抛出致命错误（抛异常会被忽略）。
 
 ## RunnerExecutor（现有三个适配器的共同形态）
-1. 在终端上建运行目录 `<run_root>/<run_id>`（清空旧文件），写 `prompt.md`。
-2. 用配置里的命令模板生成命令：`run`（新线程）或 `resume`（续线程）。占位符：`{repo} {model} {effort} {tier} {out} {thread} {perm} {script} {run_dir}`；未知占位符原样保留。`{tier}` 展开为 `tier_arg`（如 `-c 'service_tier="fast"'`），无 tier 时为空。模板首词若在终端 `bins` 里，替换为绝对路径（Windows 前加 `&`）。参数请用单引号，bash 和 PowerShell 都认。
+1. 在终端上建运行目录 `<run_root>/<dir_name>`（清空旧文件；目录里已有 `pid.txt`/`launch.pid` 时拒绝并报 `HASPID`），写 `prompt.md`。transport 有 `start_run` 时（远程终端），建目录、写 `prompt.md` 与 runner、启动由 transport 一起做：直连（direct-ssh）和远程 POSIX 是一次调用，Windows 经 Warpgate 时仍逐步上传。
+2. 用配置里的命令模板生成命令：`run`（新线程）或 `resume`（续线程）。占位符：`{repo} {model} {effort} {tier} {out} {thread} {perm} {script} {run_dir}`；未知占位符原样保留。`{tier}` 展开为 `tier_arg`（如 `-c 'service_tier="fast"'`），无 tier 时为空。tier 来源：`tier:<x>` 标签或 `routing.by_label` 的 `tier`，否则执行器的 `tier_defaults[<model>]`；`tier:none` 表示不加档位参数。模板首词若在终端 `bins` 里，替换为绝对路径（Windows 前加 `&`）。参数请用单引号，bash 和 PowerShell 都认。
 3. transport 生成 runner 脚本（Linux `run.sh`，Windows `run.ps1`）：进入 repo，写 `pid.txt`/`started.txt`，把 `prompt.md` 从 stdin 喂给命令，stdout → `stdout_name`，stderr → `stderr.txt`，结束写 `exit.txt`（`exit=N at …`）。
-4. transport 分离启动（Linux `setsid -f bash run.sh`，Windows 一次性计划任务 `agentd-<run_id>`）。
-5. `poll` = `exit.txt` 是否存在；`collect` 由各适配器解析：
+4. transport 分离启动（本机 `setsid -f bash run.sh`；远程 Linux `setsid nohup bash run.sh &`，macOS 用 `perl POSIX::setsid`，`$!` 记入 `launch.pid`；Windows 一次性计划任务 `agentd-<run_id>`）。
+5. `poll` = `exit.txt` 是否存在；`collect` 由各适配器从终端上保存的文件解析（远程时读回，不依赖调度器本机有任何 CLI）。适配器把要读的东西列成 `COLLECT`（`("read", 文件)`、`("grep", 文件, 子串)`、`("tail", 文件, n)`），经 `transport.fetch` 一次调用取回；transport 没有 `fetch` 时逐个读：
 
 | 适配器 | stdout | 最后消息 | 线程 ID |
 |---|---|---|---|
@@ -39,9 +41,10 @@ agentd 不常驻：一轮运行比起它的那次 tick 活得久，所以执行
 配置里的 `thread_from` / `last_message` 目前只是说明性字段，解析逻辑在适配器里。
 
 ## Transport 接口（`agentd/transports/`）
-`os`、`join`、`quote`、`default_run_root`、`mkdir(path, clean)`、`write_text`、`read_text`（不存在返回 None）、`grep_first(path, needle)`、`tail(path, n)`、`runner(run_dir, cwd, command, stdout_name)`、`launch(task, script, run_dir)`、`kill(task, run_dir)`、`forget(task)`、`process_count(name)`（None = 不可达）。
-- `local`：Linux 本机文件 + `setsid -f`；kill = 对 `pid.txt` 的进程组发 SIGTERM。
-- `warpgate-ssh`：Windows，经 `win10-ps.sh`（PowerShell `-EncodedCommand`）和 `win10-put.sh`（分块上传 + SHA-256 校验）；文件读取用 base64 标记回传；kill = `taskkill /T /F /PID <pid.txt>` + `schtasks /End /Delete`。脚本目录由 `agentd.helpers_dir`（或终端 `helpers_dir`）配置；脚本目前写死 `--target win10`。路径都放进 PowerShell 单引号字面量（`'` 双写）；`schtasks /TR` 里脚本路径用 `\"…\"` 包裹，依赖 Windows PowerShell 5.1 的旧式原生参数传递（PowerShell 7.3+ 的 Standard 模式下需改写）。runner 进不了 repo 时写 `exit=97` 结束。
+`os`、`join`、`quote`、`default_run_root`、`mkdir(path, clean)`（clean 遇到 pid 文件必须拒绝）、`probe(run_dir, names)`（存在的文件名列表，一次调用）、`write_text`、`read_text`（不存在返回 None）、`grep_first(path, needle)`、`tail(path, n)`、`runner(run_dir, cwd, command, stdout_name)`、`launch(task, script, run_dir)`、`kill(task, run_dir)`、`forget(task)`、`process_count(name)`（None = 不可达）；可选 `start_run(task, run_dir, files, script_name)`（建目录 + 写文件 + 启动）和 `fetch(run_dir, ops)`（一次读回多个结果，`collect` 用）。远程 transport 的脚本由一个"通道"送达：`warpgate-ssh`（`agent-access.py run`，只能走 argv）或 `direct-ssh`（Tailscale 直连 OpenSSH，脚本走 stdin，每个操作一次调用；可配 `fallback: warpgate-ssh`，见 design §9）。直连在命令运行前失败时抛 `Unreached`（`RoutedTransport` 据此换到兜底通道），主机公钥不符时抛 `HostKeyError`（不回退）；`process_count` 不能吞掉 `Unreached`。
+- `local`：Linux 本机文件 + `setsid -f`；kill = 对 `pid.txt` 的进程组发 SIGTERM。用于调度器与执行器同机的部署和测试。
+- `warpgate-ssh` + `os: linux|darwin`（`PosixWarpgateTransport`）：每个操作是一次 `agent-access.py --config <vault.config> run --target <target> -- bash -c '<script>'`（`relay()`，目标名取终端 `target`，缺省为终端名；客户端把 argv `shlex.join` 后交给目标账户的登录 shell，stdin 为 /dev/null）。文件：base64 放进 `printf %s '<b64>' | base64 -d`，超过 60000 字符分块追加再解码，回显字节数校验；读文件/`grep_first`/`tail` 的结果以 `B64=` 标记回传。runner 与本机相同（`run.sh`），`setsid nohup` 脱离 SSH 会话（pid = pgid = sid）；kill = 对该进程组 `kill -TERM -- -<pid>` + `pkill -TERM -s <pid>`，最多等 5 s 后 KILL，`exit.txt` 已存在时不发信号（防 pid 复用）；`process_count` = `pgrep -x`。`repo`/`run_root` 必须是远程绝对路径。每次调用约 10–14 s，所以启动只用一次调用。
+- `warpgate-ssh` + `os: windows`（`WindowsWarpgateTransport`）：经 `win10-ps.sh`（PowerShell `-EncodedCommand`）和 `win10-put.sh`（分块上传 + SHA-256 校验）；文件读取用 base64 标记回传；kill = `taskkill /T /F /PID <pid.txt>` + `schtasks /End /Delete`。脚本目录由 `agentd.helpers_dir`（或终端 `helpers_dir`）配置；脚本目前写死 `--target win10`（TODO：改走同一个 `relay()`，argv 为 `powershell.exe -NoProfile -NonInteractive -EncodedCommand <b64>`，去掉 helper 脚本）。路径都放进 PowerShell 单引号字面量（`'` 双写）；`schtasks /TR` 里脚本路径用 `\"…\"` 包裹，依赖 Windows PowerShell 5.1 的旧式原生参数传递（PowerShell 7.3+ 的 Standard 模式下需改写）。runner 进不了 repo 时写 `exit=97` 结束。
 - Codex 续线程：`codex exec resume <thread>` 不接受 `-C`/`-s`，`resume` 模板里不要写；工作目录由 runner 先进入 repo 保证，沙箱沿用 Codex 对该会话/全局配置的默认值（若需强制，用 `-c 'sandbox_mode="workspace-write"'`，上线前在 win10 上验证）。
 
 ## 新增一个执行器
@@ -55,4 +58,4 @@ agentd 不常驻：一轮运行比起它的那次 tick 活得久，所以执行
 2. 项目 `routing.by_assignee` 加 `<plane_user>: <name>`，`default_executor` 加默认执行器。
 3. cron 的 `agent-access.py exec` 加 `--secret-env <token_key>=<token_key>`（`agentd check-config` 会打印完整参数）。
 4. `agentd smoke <PROJECT> <name>` 验证一轮往返。
-新的 transport 类型（如 Linux 远程机）需在 `agentd/transports/` 实现上面的接口并在 `make_transport` 登记。
+远程 Linux/macOS 机器不需要新代码：`transport: warpgate-ssh, target: <Warpgate target>, os: linux|darwin, launch: nohup`，`repo`/`run_root`/`bins` 写远程绝对路径（步骤见 `docs/deploy-lxc.md` §7）。全新的 transport 类型需在 `agentd/transports/` 实现上面的接口并在 `make_transport` 登记。

docs/runbook.md

@@ -8,27 +8,36 @@ cp config/agentd.example.yaml config/agentd.local.yaml   # git-ignored；或用
 cli/agentd check-config                       # 校验 + 列出各终端 token 环境变量是否已注入 + cron 所需 --secret-env 参数
 python3 -m unittest                           # 离线单测
 ```
+调度器迁到不装 agent CLI 的 LXC：见 `docs/deploy-lxc.md`。
 配置里只写 Vault 键名（`token_key`）。agentd 读同名环境变量，token 只由 `agent-access.py exec --secret-env KEY=KEY` 注入进程内存；不要 export 到 shell、不要写文件。agentd 起的所有子进程（本机 runner/执行器、win10 中继脚本）拿到的环境都去掉了全部 `token_key` 与 `webhook_secret_key` 变量。
 
 ## 运行
-单次 tick（cron 推荐）：
+**调度器 2026-10-21（PAD-11）起跑在 LXC `agentd-01`**（root，仓库 `/opt/plane-agentd`），部署细节见 `docs/deploy-lxc.md`。任何时候只能有一个调度器排程。
+单次 tick（在 LXC 上）：
 ```bash
-cd ~/.local/share/server-access-client && python3 -I agent-access.py \
+cd /root/.local/share/server-access-client && python3 -I agent-access.py \
   --config https://nas.tail9531b.ts.net:8443/identity/agent.json exec \
   --secret-env PLANE_OREGON_API_TOKEN=PLANE_OREGON_API_TOKEN \
   --secret-env PLANE_WINDOW10_API_TOKEN=PLANE_WINDOW10_API_TOKEN \
-  --timeout 900 -- /home/ubuntu/Documents/wjj/project/plane-agentd/cli/agentd tick --once
+  --secret-env AGENTD_SSH_KEY=AGENTD_SSH_KEY \
+  --timeout 900 -- python3 /opt/plane-agentd/cli/agentd tick --once
 ```
-crontab（每 2 分钟；tick 有文件锁，重叠的 tick 直接跳过）：
+`AGENTD_SSH_KEY` 是 direct-ssh 终端的专用密钥（PAD-12，design §9）。手工运行时不加这个参数不会出错，只是直连终端整次 tick 都走 Warpgate 兜底（慢），`agentd.log` 里会有一行 `direct-ssh unavailable in this process`。但加了参数而 Vault 里没有这个键，agent-access.py 会以 `access-hub: credential_unavailable` 直接失败，agentd 根本不运行（它没有可选密钥）；cron 脚本遇到这种情况会去掉这个参数再跑一次，手工运行则要自己去掉。
+crontab（LXC root，每 2 分钟）：
 ```
-*/2 * * * * cd $HOME/.local/share/server-access-client && python3 -I agent-access.py --config https://nas.tail9531b.ts.net:8443/identity/agent.json exec --secret-env PLANE_OREGON_API_TOKEN=PLANE_OREGON_API_TOKEN --secret-env PLANE_WINDOW10_API_TOKEN=PLANE_WINDOW10_API_TOKEN --timeout 900 -- $HOME/Documents/wjj/project/plane-agentd/cli/agentd tick --once >> $HOME/Documents/wjj/project/plane-agentd/state/cron.log 2>&1
+*/2 * * * * /opt/plane-agentd/cron/agentd-tick-lxc.sh
 ```
+`cron/agentd-tick-lxc.sh` = 拿 `/var/lib/plane-agentd/cron.lock`（`flock -n -E 75`，撞锁退出 75）后执行 `timeout 900 python3 -I agent-access.py --config <identity> exec --secret-env PLANE_OREGON_API_TOKEN=… --secret-env PLANE_WINDOW10_API_TOKEN=… --secret-env AGENTD_SSH_KEY=AGENTD_SSH_KEY --timeout 850 -- python3 /opt/plane-agentd/cli/agentd --config /opt/plane-agentd/config/agentd.yaml tick --once`，输出追加到 `/var/log/plane-agentd/cron.log`（stderr 排在 stdout 之后）。如果这次 exec 只报了 `access-hub: credential_unavailable`/`credential_invalid`（退出 1，agentd 没运行），脚本在 cron.log 记一行 `AGENTD_SSH_KEY not available from Vault: tick repeated without it …`，去掉 `AGENTD_SSH_KEY` 再跑一次：直连终端走兜底，派单不中断。空闲 tick 不打印任何东西（cron 日志为空是正常的），活着与否看 `heartbeat.json` 的 `at`；`skip: another tick holds …` 表示与手工 tick 撞锁，无害。
+实际配置 `config/agentd.yaml`（已提交，不含密钥）：`state_dir: /var/lib/plane-agentd`、`log_dir: /var/log/plane-agentd`，所以下表的 `state/…` 在 LXC 上是 `/var/lib/plane-agentd/…`，`agentd.log` 在 `/var/log/plane-agentd/`。改配置/代码 = 在 oregon 提交并推送，然后 LXC 上 `cd /opt/plane-agentd && git pull`（下一个 tick 生效）。
+oregon 上的 `cli/agentd` 只当 CLI 用（new/show/comment/state/assign，`smoke … --watch`）；**不要在 oregon 上跑 `tick` 或不带 `--watch` 的 `smoke`**（那是第二个调度器）。
+暂停一个终端（不再派新运行、也不探测，在途运行照常收尾）：把它的 `concurrency` 设为 0。
+
 常驻替代：`... exec ... -- cli/agentd tick --loop --interval 120`（vault 会话超时要够长，cron 更稳）。
-cron 的 PATH 很短：本机执行器的二进制用终端 `bins` 写绝对路径（如 `claude: /home/ubuntu/.local/bin/claude`）。
+执行器的二进制一律用终端 `bins` 写绝对路径（如 `claude: /home/ubuntu/.local/bin/claude`）：cron 和 SSH 非登录 shell 的 PATH 都只有系统目录。
 
 ### 执行器二进制的解析（`bins` 或 PATH）
 命令模板的首词（`codex`、`claude`…）若在该终端的 `bins` 里，替换为那里的绝对路径；否则**原样交给终端的 shell 按 PATH 查找**。没有任何其他探测或回退：
-- oregon（local）：runner 是 `bash run.sh`，环境继承 agentd 进程（cron 下 PATH 只有 `/usr/bin:/bin`），所以 `claude` 必须写进 `bins`。**示例配置的 oregon 列了 `codex` 执行器但没有 `bins.codex`，即假定 `codex` 在 PATH 上——目前 oregon 没装 codex**（`which codex` 为空），带 `exec:codex` 的 oregon 工作项会以 exit 127（`command not found`，类型=crash，非临时性）转 Waiting。要用就装 codex 并加 `bins: {codex: <绝对路径>}`，否则从 oregon 的 `executors` 去掉 codex。
+- oregon（2026-10-21 起 `warpgate-ssh`，os linux）：agentd 经中继 `agent-access.py run --target oregon -- bash -c …` 在 oregon 上 `setsid nohup bash run.sh`，runner 的环境是 sshd 给 ubuntu 用户的非登录环境（PATH = 系统默认，没有 `~/.local/bin`），所以 `claude` 必须写进 `bins`。`transport: local`（runner 继承 agentd 进程环境，cron 下 PATH 只有 `/usr/bin:/bin`）仍可用于调度器与执行器同机的部署和测试。**示例配置的 oregon 列了 `codex` 执行器但没有 `bins.codex`，即假定 `codex` 在 PATH 上——目前 oregon 没装 codex**（`which codex` 为空），带 `exec:codex` 的 oregon 工作项会以 exit 127（`command not found`，类型=crash，非临时性）转 Waiting。要用就装 codex 并加 `bins: {codex: <绝对路径>}`，否则从 oregon 的 `executors` 去掉 codex。
 - win10（warpgate-ssh）：runner 是计划任务里的 `powershell.exe -File run.ps1`，以创建任务的用户（Administrator，交互登录）运行，PATH 取该用户的环境。示例配置用 `bins.codex` 写死 `codex.exe` 绝对路径；去掉它就等于假定 Administrator 的 PATH 里有 `codex`。
 - `check-config` 不检查二进制是否存在；首次接入一个终端用 `agentd smoke <PROJECT> <terminal>` 验证。
 
@@ -39,18 +48,45 @@ cli/agentd show GWENT GW-12 3        # 状态/指派/标签 + 最近 3 条评论
 cli/agentd comment GWENT 12 reply.md # 或 - 读 stdin
 cli/agentd state GWENT 12 Todo       # Plane 状态名或配置键（todo/review…）
 cli/agentd assign GWENT 12 window10
-cli/agentd smoke GWENT oregon        # 建一个要求回复 "[review] ok" 的工作项，tick 到结束，PASS 则移到 Done
+cli/agentd smoke GWENT oregon --watch   # 建一个要求回复 "[review] ok" 的工作项，等排程中的调度器跑完，PASS 则移到 Done
+cli/agentd smoke GWENT win10 --labels model:sol --watch   # --labels 给冒烟项加标签（路由到模型/档位）
+cli/agentd ping win10 [--transport direct-ssh|warpgate-ssh] [-n 3]   # 在终端上跑一条空命令并计时，每次调用一行
+                                     # （direct-ssh 需要 --secret-env AGENTD_SSH_KEY=AGENTD_SSH_KEY，不需要 Plane token）
+                                     # 不带 --watch = 本机自己 tick（只在调度器所在机器上、且它的 cron 停着时用）
 ```
 CLI 以 `--as <terminal>` / `agentd.cli_terminal` / 第一个有 token 的终端的身份操作。
 
+## Webhook（PAD-7：Plane 事件即时触发 tick）
+cron 每 2 分钟一次的 tick 保留作兜底；另有常驻接收端，Plane 一有工作项/评论变化就立刻触发一次 tick。
+- **服务**：LXC `plane-agentd-webhook.service`（单元文件在仓库 `systemd/`，`Restart=always`）= `agent-access.py … exec --secret-env PLANE_WEBHOOK_SECRET=PLANE_WEBHOOK_SECRET --timeout 315360000 -- python3 /opt/plane-agentd/cli/agentd webhook-serve --bind 10.77.10.249 --port 8790`。只听内网地址；密钥只在进程环境里，agentd 起的子进程（tick 脚本）环境里去掉了它。
+- **端点**：`POST http://10.77.10.249:8790/plane`；`GET /healthz` → `{"ok": true, "pending", "queued", "running", "retrying_since", "ticks", "last": {rc, seconds, deliveries, at}}`。
+- **Plane 侧**：工作区级 webhook（kappiee 工作区设置 → Webhooks，或 `tools/plane-webhook.py`），订阅"工作项"（`issue`）和"评论"（`issue_comment`）。Plane 会把所有项目的事件都发来，接收端按 `data.project` 只认配置里的项目。
+- **签名**：头 `X-Plane-Signature` = 用 webhook 密钥对请求体做 HMAC-SHA256 的十六进制摘要（Plane 签的是 `json.dumps(payload)`，发出的正是这串字节；接收端先比原始请求体，再比重新序列化的结果，均为常量时间比较）。缺失/不对 → 401，什么都不触发。其他头：`X-Plane-Event`（issue / issue_comment）、`X-Plane-Delivery`（每次投递的 UUID）。请求体：`{"event","action"(created/updated/deleted),"webhook_id","workspace_id","workspace_slug","data":{…,"project":<uuid>},"activity":{field,new_value,old_value,actor…}}`；删除事件的 `data` 只有 `id`（没有项目 → 202 忽略，由 cron 兜底）。
+- **触发逻辑**：配置项目的投递 → 200，3 s（`debounce_seconds`）内的投递合并成一次 tick；tick 运行期间又来的投递 → 它结束后再跑一次（正在跑的 tick 可能在变化之前就读过 Plane）。tick 命令就是 cron 脚本（`systemd-run --scope … cron/agentd-tick-lxc.sh`，在独立 scope 里跑，重启接收端不会杀掉正在跑的 tick）。脚本持 `flock -n -E 75`：锁被 cron/手工 tick 占着 → 退出码 75 → 每 15 s 重试，最长 2 min，之后交给下一个 cron tick。tick 本身的输出照旧进 `cron.log`/`agentd.log`。
+- **日志** `/var/log/plane-agentd/webhook.log`：每次投递一行（`delivery=<X-Plane-Delivery> event=issue/updated project=GWENT item=<uuid> decision=tick|202 ignored|401 bad signature|401 missing signature`），每次 tick 两行（`tick start (n deliveries: …)`、`tick done rc=0 in 12.3s` / `tick busy …, retry in 15s` / `giving up`）。Plane 侧的投递记录：`tools/plane-webhook.py logs <webhook id>`。
+- **查看**：`systemctl status plane-agentd-webhook`、`journalctl -u plane-agentd-webhook`、`curl -s http://10.77.10.249:8790/healthz`、`tail -f /var/log/plane-agentd/webhook.log`。
+- **轮换密钥**（在 oregon 上，kappiee 管理员会话，密钥不出进程、直接写进 Vault）：
+  ```bash
+  cd ~/.local/share/server-access-client && python3 -I agent-access.py --config https://nas.tail9531b.ts.net:8443/identity/agent.json exec \
+    --secret-env PLANE_KAPPIEE_PASSWORD=PLANE_KAPPIEE_PASSWORD --timeout 90 -- \
+    python3 -I /home/ubuntu/Documents/wjj/project/plane-agentd/tools/plane-webhook.py regenerate <webhook id>   # id: … plane-webhook.py list
+  ```
+  然后 LXC 上 `systemctl restart plane-agentd-webhook`（接收端只在启动时取一次密钥；轮换到重启之间的投递会 401，cron 兜底）。核对：`agent-access.py … exec --secret-env S=PLANE_WEBHOOK_SECRET -- python3 -c 'import os;print(len(os.environ["S"]))'`。
+- **停用**：只停接收端 `systemctl disable --now plane-agentd-webhook`（Plane 投递会失败并重试，多次失败后 Plane 自动停用该 webhook 并给管理员发邮件）；连 Plane 侧一起停：`plane-webhook.py patch <id> is_active=false`。cron 不受影响。
+- **通知**：配置里的 `notify: {kind: none}` 是钩子，核心每次改状态都调用 `Agentd.notify()`，目前什么都不做（没有 ntfy 服务）。
+
 ## 文件
 | 路径 | 内容 |
 |---|---|
 | `state/agentd-state.json` | 每项目每工作项：轮次、线程 ID 与所在终端/执行器、线程轮数、各轮标签、已见评论、重试计数/下次重试、在途运行（handle、deadline） |
-| `state/agentd.log` | 每次 tick 的动作与错误 |
-| `state/heartbeat.json` | 最近一次 tick 的时间、在途运行、错误；超过 10 分钟不更新 = agentd 没在跑 |
+| `<log_dir>/agentd.log`（默认 `state/`） | 每次 tick 的动作与错误 |
+| `state/heartbeat.json` | 最近一次 tick 的时间、在途运行、启动未确认的运行（`launching`）、本 tick 跳过的不可达终端（`down`）、没派出去的 Todo 项及原因（`skipped`：`[{project, item, terminal, reason}]`）、错误；超过 10 分钟不更新 = agentd 没在跑 |
+| `<log_dir>/webhook.log` | webhook 接收端：每次投递的决定、每次触发的 tick（见上节） |
+| `<log_dir>/remote.log` | 每次远程调用一行：`term=win10 via=direct\|warpgate\|fallback op=start\|fetch\|read\|busy\|probe\|kill\|forget… ms=… rc=…`，失败时末尾带 `unreached`/`hostkey`/`unknown`（PAD-12）。不回显到 cron.log；没有轮转，需要时手工截断 |
+| `agentd.ssh_dir`（LXC `/run/plane-agentd/ssh`） | direct-ssh 的 ControlMaster 套接字（`<host_key 指纹 8 位>-<%C 哈希>`，空闲 60 s 后自动退出；手工关闭：`for s in /run/plane-agentd/ssh/*-*; do [ -S "$s" ] && ssh -O exit -o ControlPath="$s" x; done`）和由钉住公钥生成的 `known_hosts.<终端>`；root 0700，agentd 自己建。ssh-agent 在 `/tmp/agentd-agent-*`（每个 tick 进程一个，退出时删除） |
 | `state/agentd.lock` | tick 互斥锁 |
-| `state/runs/<run_id>/` | 本机终端的运行目录：`prompt.md run.sh pid.txt started.txt exit.txt stderr.txt` + 执行器输出 |
+| `state/runs/<run_id>/` | `transport: local` 终端的运行目录：`prompt.md run.sh pid.txt started.txt exit.txt stderr.txt` + 执行器输出 |
+| oregon `/home/ubuntu/.gwent-board/agentd-runs/<run_id>/` | 远程 Linux 终端（`run_root`）：同上，另有 `launch.pid`（启动时的 `$!`，与 `pid.txt` 相同）；claude 输出 `out.json` |
 | `<repo>\runtime\agentd\<run_id>\`（win10） | 同上，`run.ps1`；codex 输出 `events.jsonl last.md` |
 
 ## 工作项生命周期
@@ -66,14 +102,96 @@ Todo + 指派给终端用户 → agentd 选最小编号、终端有空位、串
 | 现象 | 原因 / 处理 |
 |---|---|
 | `no Plane token in the environment` / tick 报 `no Plane token` | 没在 `agent-access.py exec --secret-env …` 下运行，或键名不对；`check-config` 看 MISSING |
-| `skip: another tick holds state/agentd.lock` | 上一 tick 还在跑（win10 上传慢）；正常。持续出现则查是否有卡死的 agentd 进程 |
+| `skip: another tick holds state/agentd.lock` | 上一 tick 还在跑（win10 上传慢）；正常（`tick` 此时退出码 75）。持续出现则查是否有卡死的 agentd 进程 |
+| Todo 项一直不动、心跳看着空闲 | 看 `heartbeat.json` 的 `skipped`：`concurrency n/m (GW-x)` 槽位被谁占着；`serialize_group … held by …`；`busy_process codex (… not started by agentd)` = win10 上有不是 agentd 起的（或上一轮还没退干净的）codex.exe；`relay down`；`no route for assignee …` = 指派的用户不是终端。原因第一次出现时 `agentd.log`/`cron.log` 有一行 `GW-n not dispatched (<终端>): <原因>` |
+| `webhook.log` 里 `decision=401` | 签名不对：Plane 侧密钥轮换后接收端没重启，或不是 Plane 发的。`systemctl restart plane-agentd-webhook` |
+| Plane 建 webhook 报 "The webhook URL is not allowed…" | Plane 的 SSRF 防护拒绝内网地址：Plane api/worker 环境需 `WEBHOOK_ALLOWED_IPS=10.77.10.249/32`（design §10） |
 | `[worker] 第 k 轮失败：类型=crash（临时性）…重试 n/N` | 网络/中继/429/5xx：按 `backoff_minutes` 自动重试，续原线程；超过 `transient_max` 转 Waiting |
 | `类型=timeout` | 超过 `timeout_hours`（或 `budget:2h` 标签）；进程树已结束，转 Waiting。看运行目录日志决定拆分任务或加预算 |
 | `类型=crash`（非临时）且是续线程 | 自动在新线程重跑一次（附交接摘要）；再失败转 Waiting |
 | `类型=env` | 匹配 `retry.env_patterns`（焦点丢失/弹窗/断网等），转 Waiting，同串行组暂停到该项离开 Waiting |
 | `win10 unreachable; skipping` / `busy (N codex running…)` | 中继不通，或 win10 上有不是 agentd 起的 codex（`busy_processes`）；等它结束 |
 | `schtasks launch … failed` / `upload … failed` | 中继或 Windows 计划任务问题；手工 `win10-ps.sh "schtasks /Query /TN agentd-<run>"` 查看 |
+| `relay oregon exit 255: …` / `relay timeout` / `<terminal> unreachable, skipping it for the rest of this tick` | 中继（Vault、Warpgate、目标 sshd）不通；该终端本 tick 余下的 poll/启动全部跳过（`heartbeat.down`），其他终端照常。poll 失败下一 tick 再试。手工：`agent-access.py --config … run --target oregon --timeout 30 -- bash -c 'echo ok'` |
+| `run GW-n-k: unconfirmed launch found …; adopting it` | 启动调用失败/超时（或 agentd 在启动中途退出），但终端上已有 `exit.txt`/`launch.pid`/`pid.txt`：接管该运行，不重复启动；开始评论注明"启动未确认、已接管" |
+| `run GW-n-k: nothing at …; relaunching in GW-n-k-r2` | 启动没有落地：同一 run id、新目录 `<run_id>-r2`（再失败 -r3…）重新启动，旧目录不清空。运行目录里已有 pid 文件时 transport 拒绝清空（`HASPID`），下一 tick 探测后接管 |
+| heartbeat `launching` 一直不消失 | 终端探测不通（中继坏），标记保留并占着该终端的槽位；中继恢复后自动接管或重启。人工处理：停 cron，在终端上看 `<run_root>/<run_id>/`，再从状态文件删掉该项的 `launching` |
+| `upload of … size mismatch` / `staging … size mismatch` | 远程写文件不完整（磁盘满、`run_root` 不可写）；看远程 `run_root` |
+| 远程终端每步都慢 | `warpgate-ssh`（或 `via=fallback`）每次调用约 9–14 s（取 Vault 凭据 + 网关握手 + 密码认证 + 二跳）。`direct-ssh` 首次握手 win10 约 0.5 s、oregon 约 2 s，之后 60 s 内复用连接。看 `remote.log` 的 `ms=`。一轮的调用次数：启动 1 次（Warpgate 上 prompt 超过约 45 KB 时多几次分块上传；win10 经 Warpgate 约 8–10 次中继调用），每 tick poll 1 次，收尾 1 次（win10 另加删计划任务 1 次） |
+| `SECURITY <终端>: … host key verification FAILED …` | 终端出示的主机公钥和配置里钉住的不一致：可能是重装、sshd 重新生成了主机密钥，也可能是中间人。**不会回退到 Warpgate**，该终端本 tick 标为 down（`heartbeat.down`），每个 tick 都会再报一次。先经第二通道核对（Warpgate：`agent-access.py … run --target oregon -- cat /etc/ssh/ssh_host_ed25519_key.pub`，win10 `win10-ps.sh "Get-Content C:\ProgramData\ssh\ssh_host_ed25519_key.pub"`），确认是合法变更后再更新 `terminals.<t>.ssh.host_key`（deploy-lxc §9 第 4 步）。新的 pin 会换一个 ControlMaster 套接字名（`<pin 指纹>-%C`），不会复用按旧 pin 校验过的连接。终端根本不提供 ssh-ed25519 主机密钥（`no matching host key type`）也按这一行处理。等核对期间要继续派单：把 `transport` 改回 `warpgate-ssh` |
+| `<终端>: direct-ssh failed before the command ran (…); using the warpgate-ssh fallback for the rest of this tick` | 直连在命令运行前就失败了（tailscale 断、sshd 停、authorized_keys 缺或 ACL 不对、密钥被拒）。这个操作已在 Warpgate 上重做，本 tick 余下的也走 Warpgate，下个 tick 再试直连。`remote.log` 里会出现 `via=fallback`。查原因：`cli/agentd ping <终端> --transport direct-ssh -n 1`，括号里是 ssh 的原话（如 `Permission denied (publickey)`、`connect to host … Connection timed out`） |
+| `direct-ssh unavailable in this process: AGENTD_SSH_KEY is not set …` / `ssh-add failed …` | 进程没拿到密钥（手工运行没加 `--secret-env AGENTD_SSH_KEY=AGENTD_SSH_KEY`，或 cron 脚本因为 Vault 取不到键而去掉了它），或 Vault 里的值不是无口令的 OpenSSH 私钥。直连终端整次都走兜底；没配 `fallback` 的终端本 tick down |
+| cron.log：`AGENTD_SSH_KEY not available from Vault: tick repeated without it …` | Vault 键 `AGENTD_SSH_KEY` 被删、改名或不可读，agent-access.py 拒绝了整个 exec。tick 已不带密钥重跑，直连终端每个 tick 都走 Warpgate 兜底（慢，但不停派单）。修复：按 deploy-lxc §9 第 1 步恢复这个键；`agent-access.py … list` 能看到键名是否还在 |
+| 手工 `exec … --secret-env AGENTD_SSH_KEY=AGENTD_SSH_KEY` 只打印 `access-hub: credential_unavailable` | 同上，Vault 里没有这个键（也可能是某个 Plane token 键）。手工运行时去掉 `AGENTD_SSH_KEY` 那个参数即可 |
+| `ssh administrator@… timeout after 150s` / `ssh … exit 255: …`（无 fallback 行） | 直连的命令**可能已经运行**（超时、会话中途断开、认不出的 255），按规则不换路径重试：终端本 tick down；启动的话保留 `launching` 标记，下一 tick 探测后接管或在 `-r2` 重启。win10 上每次都超时：多半是 stdin 没到 PowerShell（sshd 默认 shell 被改成了 PowerShell），见 deploy-lxc §9 第 3 步 |
 | 工作项在运行中被人改了状态 | 运行还没结束：下一 tick 取消（kill + `[worker] 运行已取消…` 评论），不改状态。已经结束（`exit.txt` 已写）：结果照常回帖，但不再改状态/重试 |
 | Plane 回帖失败 | 在途记录保留，下一 tick 重新 collect 并回帖（可能重复一条评论） |
-| 想手工终止一轮 | 把工作项移出 In Progress（如 Waiting）即可，下一 tick agentd 自己 kill 并释放槽位。中继不通、kill 一直失败时：在终端上手工结束（本机 `kill -TERM -<pid.txt>`，win10 `taskkill /T /F /PID <pid.txt>`），停 cron 后从 `state/agentd-state.json` 删掉该项的 `inflight` |
+| 想手工终止一轮 | 把工作项移出 In Progress（如 Waiting）即可，下一 tick agentd 自己 kill 并释放槽位。中继不通、kill 一直失败时：在终端上手工结束（Linux 终端 `kill -TERM -- -$(cat pid.txt); pkill -TERM -s $(cat pid.txt)`，win10 `taskkill /T /F /PID <pid.txt>`），停 cron 后从 `state/agentd-state.json` 删掉该项的 `inflight` |
 | `已连续 N 轮 [progress]` | 达到 `max_progress_rounds`：看最近几轮是否在原地打转；拆分任务或补充说明后移回 Todo |
+
+## 回滚到 oregon（LXC 调度器出问题时）
+1. 停 LXC：`agent-access.py … run --target agentd-01 -- bash -c 'crontab -r; flock /var/lib/plane-agentd/cron.lock true'`（删 crontab，等正在跑的 tick 结束）。
+2. 状态拷回 oregon：LXC `/var/lib/plane-agentd/agentd-state.json` → oregon `~/.gwent-board/agentd/agentd-state.json`（base64 经中继读回，比对 sha256）。在途运行照常被 oregon 接着 poll。
+3. oregon 恢复 cron：`crontab ~/.gwent-board/crontab.backup-20261021-pad11`（即 `*/2 * * * * /home/ubuntu/Documents/wjj/project/plane-agentd/cron/agentd-tick.sh`）。`cron/agentd-tick.sh` 用 `config/agentd.oregon.yaml`（`state_dir: ~/.gwent-board/agentd`，日志 `~/.gwent-board/agentd/agentd.log`、`~/.gwent-board/agentd-cron.log`），与 LXC 配置只差路径。
+4. 看 oregon 的 `~/.gwent-board/agentd/heartbeat.json` 更新。再切回 LXC = 反过来做一遍（docs/deploy-lxc.md §5）。
+
+## 上线记录（2026-10-22，PAD-12：直连 SSH + 单次调用派单）
+- 代码 5192e04：direct-ssh 传输（会话建立前失败才回落 Warpgate，主机公钥不符绝不回落）、主机公钥固定、Vault 密钥载入 ssh-agent、复用连接、win10 一次调用启动和一次调用收尾、`remote.log`、`agentd ping`。148 个单测通过；经正确性、安全两轮 opus 审查后修正。
+- 密钥：在 LXC 的 `/dev/shm` 生成 ed25519，私钥只存在 Vault `AGENTD_SSH_KEY`（指纹 `SHA256:+6uooRFFB3tX4fQ+Sfl+3FmLm9KJpPYpSPzoTcCq/2M`），临时文件已删。公钥行 `restrict,from="100.69.78.124" ssh-ed25519 … plane-agentd@agentd-01`：
+  - oregon：`~ubuntu/.ssh/authorized_keys`，备份 `.bak-pad12-20261022-025955`；
+  - win10：`C:\ProgramData\ssh\administrators_authorized_keys`，备份 `.bak-pad12-20261022-110041`，ACL 未变（SYSTEM 和 Administrators 完全控制）。win10 的 sshd 是 OpenSSH_9.5p1，DefaultShell 未设（即 cmd.exe）。
+  - 两台的主机公钥经 ssh-keyscan 和经 Warpgate 读公钥文件两路核对一致后，固定写进配置。
+- 部署 03:01（持 `cron.lock` 执行 git pull 和 ping）。
+  - ping：oregon 首次 3.9 s（含建立主连接），复用 0.77 s；win10 首次 0.52 s，复用 0.17 s。
+  - 部署时 win10 的第一次调用有一次 60 s 超时。随后 3 轮新建主连接的首次调用都是 0.52 s，复现不出，推测是 tailnet 线路空闲后刚恢复。这类超时不回落，该终端本 tick 算不可用、下个 tick 重试；在 `remote.log` 里继续观察。
+- 冒烟（真实 webhook）：
+  - GW-27 win10：建项 03:06:17.9 → 开工评论 03:06:28.4，**10.5 s**（之前 95 s）。其中 busy 0.22 s、start 0.32 s；收尾 read、fetch、forget 三次调用共 0.9 s。
+  - GW-28 oregon：建项 03:10:09.5 → 开工 03:10:23.6，**14.1 s**（start 4.1 s，含新建主连接）；03:10:32.6 结果已贴回。
+- 剩下的延迟在收尾：运行结束后要等下一次 tick（cron 2 分钟，或别的投递触发）才收。迁 Temporal（PAD-8）时用定时器轮询解决。
+- 回滚：把该终端的 `transport` 改回 `warpgate-ssh` 再 pull。撤销访问：删掉两台 authorized_keys 里的那一行，再对 `/run/plane-agentd/ssh/` 下的主连接执行 `ssh -O exit`（见 deploy-lxc §9）。
+
+## 上线记录（2026-10-22，PAD-7：Plane 侧 webhook 启用）
+- 01:14 Plane VM 506：`plane.env` 与渲染 compose 设 `WEBHOOK_ALLOWED_IPS=10.77.10.249/32`（备份 `*.bak-pad7-20261022-011453`），`--dry-run` 确认后只重建 api、worker、webhook-consumer、outbox-poller、automation-consumer、beat-worker（01:15:43–01:16:33）；01:18:26 实测 API 已返回 200（之前未测，停机上限约 2.5 分钟），api 容器 4 分钟内转 healthy。
+- 01:19 `tools/plane-webhook.py create http://10.77.10.249:8790/plane` → 201（白名单生效），密钥写入 Vault `PLANE_WEBHOOK_SECRET`（替换临时值）；01:22:30 `systemctl enable --now plane-agentd-webhook`（LXC 代码 1b4d965，单元文件与仓库一致）；01:22:57 `patch c039c471-… is_active=true`。
+- Plane 建出的是 **v2** webhook：负载 `{version: v2, event, entity_type, data, …}`，由 `plane/webhook/bgtasks/webhook_task.py` 的 `send_webhook` 发出，签名 `HMAC-SHA256(secret, json.dumps(payload))`，与接收端一致；`data.project` 路由到 GWENT 正常。
+- 实测（真实投递）：评论 GW-25 → 01:24:34 投递到达 → 01:24:37 tick → 4.6 s 完成；GW-25 改 Done 同样 3 s 后 tick。建项 GW-26（01:25:14）→ 01:25:15 投递 → 01:25:18 tick 开始 → 01:26:49 `[worker] 第 1 轮开始`（win10 codex/astra）：**建项 → 开工 95 s，其中 tick 本身 91 s**（中继、上传任务书、计划任务启动）；只靠 cron 时在这之前还要平均多等约 60 s。剩下的成本都在中继调用上（PAD-12）。
+- cron `*/2` 保留为兜底（同一把 `cron.lock`）；oregon crontab 仍为空，只有一个调度器。
+
+## 上线记录（2026-10-21，PAD-7：webhook 接收端）
+- 13:25 LXC 上 `plane-agentd-webhook` 启用（`6dc7e72`），cron 不变。**Plane 侧 webhook 未建成**：Plane 3.3.3 对 `http://10.77.10.249:8790/plane` 返回 400 "The webhook URL is not allowed. Provide a publicly reachable HTTP or HTTPS URL."（SSRF 防护拒绝内网地址，见 design §10）；需要在 Plane 的 api/worker 环境加 `WEBHOOK_ALLOWED_IPS=10.77.10.249/32`，之后按 deploy-lxc §8 第 2–5 步建 webhook。在此之前 Vault 里的 `PLANE_WEBHOOK_SECRET` 是临时随机值（`plane_wh_interim_…`），`plane-webhook.py create` 会用 Plane 生成的密钥覆盖它。
+- 验收用的是在 LXC 上按 Plane 格式签名的合成投递（`/root/lxc_send.sh good|bad|missing|other`）：错签名、缺签名 → 401，不触发；别的项目 → 202；正确签名 → 200，3 s 后 tick 开始。tick 运行中又来一条 → 结束后补跑一次（13:27 两条投递：224 s 的 tick——收尾 GW-15、在 win10 启动 GW-5——之后接着 24 s 的补跑）；那段时间 cron 的 tick 撞锁立即退出，cron 本身照常每 2 分钟触发。
+- 冒烟 GW-17（oregon claude-code）：建项 13:43:15.8 → 合成投递 13:43:41（经 oregon→LXC 中继发出，人为多了约 25 s）→ tick 开始 13:43:44 → `[worker] 第 1 轮开始` 13:44:56 → `[review] ok` 13:47:54，PASS 并移到 Done。**投递 → tick 开始 = 3 s；tick 开始 → 开始评论 = 72 s**（这一 tick 先 poll win10 上在途的 GW-5，再经中继启动 oregon；同期 cron 的启动 tick 也要 70–77 s）。真实 Plane 投递约 1 s 到达时，建项 → 开始评论预计约 75 s，而只靠 cron 是平均约 60 s 等待 + 同样的 tick 时长。≤10 s 的目标在"开始评论"这个口径上达不到：瓶颈是中继调用（每次约 10 s，vault + Warpgate SSH）和启动流程，不是触发方式。
+- GW-16（第一次冒烟）：webhook 触发的 tick 里 LXC→中继报 `access-hub: config_private_transport_failed`（LXC 经 tailnet 取 `nas…/identity/agent.json` 15 s 超时；13:38 的 cron tick 也撞上一次），启动失败；下一 cron tick 探测无运行目录 → `GW-16-1-r2` 重新启动，PASS（防重复启动机制按设计工作）。
+- 重启：tick 运行中 `systemctl restart plane-agentd-webhook`，tick（在 `systemd-run --scope` 里）没被杀、照常结束（heartbeat 13:50:10，errors=[]）；接收端约 19 s 后重新监听（取 vault 密钥 + 启动）。
+
+## 切换记录（2026-10-21，PAD-11：调度器迁到 LXC agentd-01）
+- 代码：防重复启动（`launching` 标记 + 探测接管/换目录重启）与"中继不通的终端本 tick 跳过"随本次一起上线（4f45d36）；LXC 用仓库里的 `config/agentd.yaml`，oregon 回滚用 `config/agentd.oregon.yaml`。
+- 切换前 LXC 手工 `tick --once`（oregon cron 仍在跑，Plane 上无 Todo 项）：空闲、无错误；LXC 经 helper 探测 win10（`codex` 计数）与经中继探测 oregon 运行目录均正常。
+- 切换时 oregon 刚收尾 GW-11-1（11:53 由 oregon 派出、11:57 `[review]`），无在途运行。11:59 删除 oregon crontab 行（备份 `~/.gwent-board/crontab.backup-20261021-pad11`，oregon crontab 现为空），状态文件（sha256 `106c2bce…`）复制到 LXC，12:00 装 LXC crontab，12:02:09 第一个 cron tick（heartbeat 更新，errors=[]）。
+- 冒烟（oregon 上 `smoke … --watch`，由 LXC cron 派发）：GW-12 oregon claude-code/sonnet 12:04:20 启动 → 12:06:47 `[review] ok` PASS；GW-13 win10 codex/gpt-6.1-sol/fast 12:05:58 启动 → 12:07:45 `[review] ok` PASS。续线程：GW-12 评论"回复 `[review] ok 2`"后移回 Todo，12:08:20 第 2 轮续 session `2348d52f…`（两轮 `out.json` 的 `session_id` 相同），12:10:44 `[review] ok 2`。两项均移到 Done。
+
+## 切换记录（2026-10-21，P1-5）
+- 旧 worker `gwent-design/tools/board-worker-plane.py`（cron `board-worker-plane-cron.sh`）已从 crontab 移除；脚本与 `~/.gwent-board/plane-state.json`、`plane-worker.log` 保留作历史，不要再启用（两者会抢同一批 window10 的 Todo 项）。
+- 切换时旧 worker 无在途运行（GW-6 的 board-6-2 已于 06:30 以 `[blocked]` 结束，项在 Waiting）。GW-6 的 codex 线程 `01a15ce3…` 已迁入 agentd 状态（`runs=2`、`thread_on=[win10, codex]`、旧评论标为已见），移回 Todo 并指派 window10 时 agentd 续这个线程，run id 从 `GW-6-3` 起。
+- 运行目录变了：旧 worker 用 win10 `runtime\board\board-<n>-<k>`，agentd 用 `runtime\agentd\GW-<n>-<k>`。
+- 冒烟：oregon（GW-7，claude-code/sonnet，约 30 s）与 win10（GW-8，codex/gpt-6.1-sol/fast，约 3.5 min）均 PASS。win10 上验证过：`codex exec resume` 也会输出 `thread.started`（与原线程同 ID）；续线程时 `-c sandbox_mode="workspace-write"` 生效，能在仓库内写文件；把运行中的项移出 In Progress 后，`taskkill /T /F /PID <pid.txt>` 结束了整棵树（runner powershell → codex.exe → code-mode-host / command-runner → powershell），`Get-Process codex` 归零，计划任务被删除。
+- 差异（P1-5 时）：旧 worker 对没有 model 标签的项用 sol + `service_tier="fast"`；agentd 当时只在有 `model:sol` 标签时加 fast。P1-6 起 codex 执行器有 `tier_defaults: {gpt-6.1-sol: fast}`，无标签项也是 sol + fast；`tier:<x>` 标签改档位，`tier:none` 去掉档位参数。
+- win10 上 codex 的 stderr 常见 websocket 连接被拒（`os error 10061`，`wss://chatgpt.com/...`）后回退成功；只有运行失败时它才出现在失败说明的尾部，不影响分类（不匹配临时性模式）。
+
+## 切换记录（2026-10-21，P1-6：oregon 改为远程启动）
+- `config/agentd.yaml` 的 oregon 从 `transport: local` 改为 `warpgate-ssh`（`target: oregon, os: linux, launch: nohup, run_root: /home/ubuntu/.gwent-board/agentd-runs`）。agentd 仍在 oregon 上跑，但启动执行器走 Warpgate 回到本机，与将来 LXC 上的路径完全相同。切换时 oregon 无在途运行；GW-7/GW-10 的 claude 线程（`thread_on=[oregon, claude-code]`）不受影响——会话存在 oregon 的 `~/.claude`，与 transport 无关。
+- win10 不变（helper 脚本 + schtasks），只是中继子进程改由共用的 `call()` 启动；切换后 GW-9 的 poll/collect 已经走新代码并正常结束。
+- 冒烟（经 Warpgate 远程启动）：GW-10 第 1 轮 claude-code/sonnet PASS（`[review] ok`，启动 08:08:54 → 结束 08:10:17，session `fa3be0f7…`）；评论"请回复 `[review] ok 2`"并移回 Todo 后第 2 轮续同一 session（`out.json` 的 `session_id` 不变），回 `[review] ok 2` 转 Review。启动只用 1 次中继调用（约 12 s）；72 KB 的 prompt 走分块上传（约 75 s）。
+- kill：直接经中继对 oregon 起 `sleep 600 & sleep 500` 的运行再 kill，两个 sleep 都被结束（进程组 + 会话），`pid.txt` 与 `launch.pid` 相同（setsid 没有再 fork）。经 Plane 的取消冒烟（运行中移到 Waiting）因 Plane 当时 502（08:14 起）未做，待补。
+
+
+## 回滚记录（2026-10-21：调度器回到 oregon）
+- 原因：LXC agentd-01 经 tailnet 到 NAS 的路径走 DERP 中继，取 vault 身份端点（`nas…/identity/agent.json`）TCP 连接 30/30 超时，tick 里的中继调用不稳定。网络修好后再切回 LXC（docs/deploy-lxc.md §5）。
+- 做法：LXC 删 cron 行、停用 `plane-agentd-webhook`、等 `agentd.lock`/`cron.lock` 空闲；`agentd-state.json`（sha256 7201077b…eb87，两端一致）拷到 oregon `~/.gwent-board/agentd/`（旧文件备份为 `.bak-<时间戳>`）；oregon 恢复 `*/2` 的 `cron/agentd-tick.sh`。LXC 上的文件、服务单元均未删除。`config/agentd.oregon.yaml` 无需改动（`check-config` 通过）。
+- 验证：oregon 首个 tick 14:20:01 接管，在途的 GW-5-1（win10 Codex）已结束，结果回帖并进入 Review；heartbeat `errors=[]`、`inflight=[]`，下一 tick（14:22）同样。
+- 重新切回 LXC 之前：先 `crontab -r`（oregon）并等 tick 结束，再按 deploy-lxc §5 把状态拷回 LXC，并重新 enable webhook 服务；任何时刻只能有一个调度器。
+
+## 切换记录（2026-10-21，网络修复后重新切回 LXC）
+- 网络：OPNsense 增加出站 NAT 静态端口规则 `Tailscale agentd-01 static-port`（10.77.10.249/32，UDP，WAN），LXC tailscaled 改听 UDP 41642；LXC 到 NAS 直连（`tailscale ping` 5 ms via 223.102.82.206:1549，身份端点 30/30 成功、中位 0.089 s）。详见 deploy-lxc §2.1。
+- 切换：oregon crontab 清空（备份 `~/.gwent-board/crontab.backup-20261021-152423-recut`），状态文件（sha256 `54b747db…7155`，两端一致；LXC 旧文件备份 `agentd-state.json.bak-20261021-152504`）拷到 LXC，装 LXC crontab `*/2 … cron/agentd-tick-lxc.sh`。在途的 GW-18 带过去（LXC 上为 GW-18-2），之后的 run-1 结果已回帖。heartbeat errors=[]、down=[]、skipped=[]，只有一个调度器。
+- Webhook（PAD-7）仍未启用：`plane-agentd-webhook` inactive/disabled；Plane 的 `WEBHOOK_ALLOWED_IPS` 尚未设置（见 deploy-lxc §8 的具体改法，Plane 实为 VM 506 / 10.77.10.69）。PLANE_WEBHOOK_SECRET 仍是临时值。待办：应用该改动后做 §8 第 2–5 步；考虑轮换 Plane 的 AES_SECRET_KEY（排查时被打印进了一次会话记录）。

docs/usage.md

@@ -0,0 +1,105 @@
+# 使用说明：在 Plane 上派活给各机器上的 agent
+
+核对依据：`agentd/core.py`、`agentd/webhook.py`、`config/agentd.yaml`、LXC 上的实际状态、Plane GWENT 项目的标签和成员（2026-10-22）。代码或配置改了要同步改这里。
+
+## 1. 派单需要的字段
+- **必填**：状态设为 `Todo`，指派人里至少有一个终端账号（`window10` 或 `oregon`）。标题加描述就是任务书，原样进入 prompt。标签可选，模块与调度无关。
+- 有多个指派人时，按 Plane 返回的顺序，第一个能对上终端的账号生效（`core.py` 的 `route()`）。所以"kappiee + window10"会派到 win10。
+- **只指派 `kappiee`、指派 `mac`（暂无终端）或不指派**：`route()` 返回空，这一项被静默跳过。Plane 上不会有任何评论，只在 `heartbeat.json` 的 `skipped` 里记一条 `no route for assignee kappiee`，`agentd.log` 里在原因第一次出现时记一行。
+- 只有 `Todo` 状态的项会被拾起（`pick()`）。Backlog、Waiting、Review 永远不会自动执行。唯一例外是处于 In Progress、等待临时性重试的项。
+
+## 2. 终端、执行器和默认值
+
+| Plane 指派人 | 终端 | 执行器 | 默认模型 / effort / 档位 | 运行目录（日志和输出） |
+|---|---|---|---|---|
+| `window10` | win10（Tailscale 直连 SSH，Warpgate 兜底；Windows 计划任务） | codex（`codex exec`，`workspace-write`） | `gpt-6.1-sol` / `medium` / `fast` | `C:\…\gwent\runtime\agentd\<run_id>\`（`prompt.md run.ps1 events.jsonl last.md stderr.txt exit.txt`） |
+| `oregon` | oregon（Tailscale 直连 SSH，Warpgate 兜底；`setsid nohup`） | claude-code（`claude -p --permission-mode acceptEdits`） | `sonnet` / （不传 effort）/ — | `/home/ubuntu/.gwent-board/agentd-runs/<run_id>/`（`out.json` 等） |
+| `kappiee` / `mac` / 无人 | 不派 | — | — | — |
+
+- 两个终端的并发都是 1，按终端计数，跨项目共享。
+- win10 设了 `busy_processes: [codex]`：agentd 自己在 win10 上没有在途运行、却发现有 `codex.exe` 在跑时，不派单（`busy_process codex (… not started by agentd)`）。
+- oregon 的运行直接在 `/home/ubuntu/Documents/wjj/project/gwent` 里改文件。这正是指挥所用的 checkout。
+
+## 3. 代码里实际生效的标签
+Plane 里目前只建了 4 个标签：`model:astra`、`model:sol`、`effort:high`、`needs-game`。其余前缀的标签要先在 Plane 里建出来才能挂上。
+
+| 标签 | 效果（`route()`） |
+|---|---|
+| `model:astra` | 改用 gpt-6-astra，effort high（来自 by_label 配置） |
+| `model:sol` | 改用 gpt-6.1-sol，tier fast |
+| `model:<x>`（其他任意） | 模型名原样改成 x，例如 oregon 上的 `model:opus`。模型名与执行器相关：`model:astra` 挂在 oregon 的项上会让 claude 失败 |
+| `effort:<x>` | 只对 codex 生效。claude 的命令模板里没有 `{effort}`，所以被忽略 |
+| `tier:<x>` / `tier:none` | 只对 codex 生效，改档位或去掉档位参数 |
+| `needs-game` | 串行组 `game`：同组同时只跑一项；env 类失败会暂停整组，直到该项离开 Waiting。**`needs-*` 不是通用前缀**，只有配置里写了的才生效 |
+| `exec:<executor>` | 要求同名执行器存在且终端支持。现在每个终端只有一个执行器，例如 `exec:codex` 挂在 oregon 上只会得到 `executor codex not available` 跳过，等于没用 |
+| `cap:<x>` | 终端的 `capabilities` 里必须有 x。win10 有 `[android, game]`，oregon 没有，不满足就跳过 |
+| `budget:2h` / `budget:90m` | 覆盖超时时间，默认 4 h |
+
+多个标签按顺序处理，后面的覆盖前面的。
+
+## 4. 一轮的完整流程
+1. **触发**：Plane 的工作区级 webhook（`issue` 和 `issue_comment` 事件）约 1 s 内送到 `http://10.77.10.249:8790/plane`。
+   - 接收端用 HMAC 校验签名，只认 GWENT 项目，其他项目返回 202 忽略。
+   - 等 3 s，把这期间的投递合并成一次 tick。tick 跑的就是 cron 脚本。
+   - tick 运行中又来投递时，结束后再补跑一次。锁被占用（退出码 75）时每 15 s 重试，最多 2 min。
+   - 兜底：cron 每 2 分钟一次 tick，同一把 `cron.lock`。
+   - agentd 自己回帖、改状态也会产生投递，所以每个有动作的 tick 后面都会多一次空 tick（实测 11.1 s）。
+2. **tick 的步骤**：先 poll 所有在途运行（每个一次远程调用：直连亚秒到约 2 s，走 Warpgate 时约 10 s），再按编号从小到大挑 Todo 项。依次检查：终端是否可达、并发槽位、串行组、busy 进程。
+   - 启动前先把 `launching` 标记落盘，防止重复启动。
+   - 启动后把项改为 In Progress 并指派给终端账号，以终端账号发评论 `[worker] 第 k 轮开始：codex / gpt-6-astra / high @ win10，新线程，run=GW-26-1。`
+3. **实测耗时**（runbook 记录和 LXC 的 webhook.log）：
+   - 建项到投递约 1 s，再加 3 s 合并窗口。
+   - **win10 启动的那次 tick 91.4 s**（GW-26：建项 01:25:14，开工评论 01:26:49）。
+   - oregon 启动的 tick 约 72 s。
+   - 空 tick 4.6–4.8 s。
+   - 以上是 Warpgate 时期的数字，瓶颈在中继调用，与触发方式无关。
+   - PAD-12 起两个终端都走 Tailscale 直连：win10 启动由约 10 次中继调用变成 busy 检查 + 1 次启动调用，收尾由约 7 次变成 2 次。目标是建项到开工评论 ≤ 15 s，上线后在这里补实测。直连不通时这一轮自动走 Warpgate，只是慢，派单行为不变（design §9）。
+4. **结束时**：以终端账号回帖。
+   - 第一行是 `[codex] 第 k 轮结束（exit=…，run=…，终端 win10，日志 <运行目录>）`，下面是 agent 的最后消息原文（`<pre>` 包裹，超过 6000 字截断，并注明全文所在路径）。
+   - 再按最后消息首行的标签流转（`transition()`）：
+
+| 首行 | 状态 / 指派人 |
+|---|---|
+| `[review]` | Review，指派 `oregon`（reviewer） |
+| `[question]` / `[blocked]` | Waiting，指派 `oregon`（answerer） |
+| `[progress]` | 回到 Todo，留在原终端，下一 tick 续同一线程。连续 12 轮后转 Waiting |
+| 没有标签但 exit=0 | 按 review 处理，并注明"未按协议标注首行" |
+
+   注意：Waiting 和 Review 都指派给 **oregon，不是 kappiee**。没有任何东西会自动处理这两个状态，由指挥（或你）介入。
+5. **失败**：回帖 `[worker] 第 k 轮失败：类型=crash|timeout|env|logic，run，exit，重试 n/3。…` 并附上最后一帧输出。
+   - 临时性错误（stderr 尾部匹配超时、429、5xx、连接被拒等）自动重试，退避 2/5/10 min，续原线程，状态保持 In Progress。
+   - 续线程那一轮出现非临时性失败时，立即在新线程重跑一次，并附交接摘要。
+   - 其余情况转 Waiting，指派 oregon。超时会先 kill 进程树。逻辑类失败永不自动重试。
+
+## 5. 回答问题和续跑
+1. 在项上发评论回答，开头不要写 `[worker]`，这类评论永远会被跳过。
+2. **把状态改回 Todo，并把指派人改回原终端**（例如 `window10`）。这一步是坑：如果指派人还是 oregon 就改 Todo，会派给 oregon 的 Claude。因为终端或执行器变了，会**重建线程**并附交接摘要。
+3. 同一终端、同一执行器、线程续了不到 8 轮时，走 `codex exec resume` 或 `claude --resume`，只发送上次之后的新评论（按评论 ID 判断）。
+4. 单发评论只会触发一次空 tick，**不会**让 agent 动起来，必须改回 Todo。运行期间发的评论，跑着的那一轮看不到，下一轮才带上。
+
+## 6. 取消和停止
+- **把运行中的项移出 In Progress**（改成 Waiting、Backlog、Cancelled 都可以）。下一 tick 会 kill 进程树，回帖 `[worker] 运行已取消：工作项状态被改为 X（run=…）`，释放槽位和串行组，保留线程 ID（`cancel()`）。
+- 如果移到的是 **Todo**：同一 tick 里先取消，再立刻按新一轮重新派发（续原线程），相当于"带上新评论重启"。
+- 运行已经结束时（`exit.txt` 已写）再移动：结果照常回帖，但不再改状态，也不重试。
+- 暂停整个终端：把它的 `concurrency` 设为 0（需要改配置，不能在 Plane 里操作）。
+
+## 7. 日志和状态的位置（LXC）
+- `/var/log/plane-agentd/{agentd.log, cron.log, webhook.log}`
+- `/var/lib/plane-agentd/{agentd-state.json, heartbeat.json}`：heartbeat 里有 `skipped` 及原因、`inflight`、`down`。
+- 健康检查：`GET http://10.77.10.249:8790/healthz`
+- 任务产出本身在终端的 repo 里，gwent 的惯例是 win10 的 `runtime/edits/GW-k/`。
+
+
+## 8. 用手机试一下（冒烟测试，5 步）
+1. GWENT 新建工作项，标题 `冒烟 oregon 往返`。描述：`只回复一行 [review] ok；不要读写或修改任何文件，不要运行命令。` oregon 的 Claude 是 acceptEdits，并且在指挥用的 checkout 里运行，所以这句必须写。
+2. 指派人**只选 oregon**，不加标签（默认 sonnet），状态设为 **Todo**。
+3. 约 1–1.5 分钟后：状态变成 In Progress，出现 oregon 发的评论 `[worker] 第 1 轮开始：claude-code / sonnet @ oregon，新线程，run=GW-n-1。`
+4. 再过约 1–3 分钟（历次冒烟 1.5–3 min）：出现评论 `[claude-code] 第 1 轮结束（exit=0 …）` 加 `[review] ok`，状态变为 Review、指派 oregon。
+5. 可选，测续线程：评论 `请只回复 [review] ok 2`，状态改回 Todo（指派保持 oregon）。开始评论会写"续线程 xxxx"，结束后回 `[review] ok 2`。最后手动改成 Done。
+   - 超过 3 分钟没动静：看 heartbeat 的 `skipped`，常见原因是 oregon 槽位被占，或指派人不对。
+
+## 9. 通知和手机 App（2026-10-22 核实）
+- Plane 官方手机 App 可以登录自建的 Commercial 版（选 "Self hosted sign in"，填 `https://plane.kappiee.com`）。plane.kappiee.com 只对内网和 tailnet 开放，手机要开 Tailscale。
+- **自建实例没有手机推送**：Plane 文档原文 "Push notifications are currently available only for Plane Cloud users."（docs.plane.so/devices/mobile）。App 里只能在收件箱看到通知。
+- 收件箱只收你被指派、被 @、创建或订阅的项。Waiting 和 Review 现在都指派给 oregon，不会进你的收件箱。需要你处理的项，由指挥改成指派 kappiee 或 @你。
+- 真正推到手机要走外部渠道：给 Plane 配 SMTP 发邮件，或用 ntfy 之类。计划放在最后（PAD-13）。agentd 里现在的 `notify: {kind: none}` 是空钩子。

systemd/plane-agentd-webhook.service

@@ -0,0 +1,24 @@
+# Plane webhook receiver on the LXC agentd-01 (docs/runbook.md "Webhook", docs/deploy-lxc.md §8).
+# The webhook secret is injected by agent-access.py exec (Vault key PLANE_WEBHOOK_SECRET) into the process env only.
+# Install: cp systemd/plane-agentd-webhook.service /etc/systemd/system/ && systemctl daemon-reload &&
+#          systemctl enable --now plane-agentd-webhook
+[Unit]
+Description=plane-agentd Plane webhook receiver (triggers dispatcher ticks)
+Wants=network-online.target
+After=network-online.target
+StartLimitIntervalSec=0
+
+[Service]
+Type=simple
+Environment=HOME=/root
+WorkingDirectory=/opt/plane-agentd
+ExecStart=/usr/bin/python3 -I /root/.local/share/server-access-client/agent-access.py \
+  --config https://nas.tail9531b.ts.net:8443/identity/agent.json exec \
+  --secret-env PLANE_WEBHOOK_SECRET=PLANE_WEBHOOK_SECRET --timeout 315360000 -- \
+  /usr/bin/python3 /opt/plane-agentd/cli/agentd --config /opt/plane-agentd/config/agentd.yaml \
+  webhook-serve --bind 10.77.10.249 --port 8790
+Restart=always
+RestartSec=5
+
+[Install]
+WantedBy=multi-user.target

tests/fakes.py

@@ -6,10 +6,12 @@ from agentd.config import Config
 from agentd.core import Agentd
 from agentd.executors import ExecResult
 from agentd.store import Store
+from agentd.transports import TransportError
 
 STATES = {"Backlog": "s-bl", "Todo": "s-todo", "In Progress": "s-run", "Waiting": "s-wait", "Review": "s-rev", "Done": "s-done"}
 MEMBERS = {"window10": "u-w10", "oregon": "u-or", "kappiee": "u-k"}
-LABELS = {"model:astra": "l-astra", "needs-game": "l-game", "exec:claude": "l-claude", "budget:30m": "l-b30", "cap:gpu": "l-gpu"}
+LABELS = {"model:astra": "l-astra", "needs-game": "l-game", "exec:claude": "l-claude", "budget:30m": "l-b30", "cap:gpu": "l-gpu",
+          "tier:none": "l-tnone", "tier:flex": "l-tflex"}
 
 
 def config_data():
@@ -88,15 +90,39 @@ class FakeExecutor:
         self.resume = resume
         self.started, self.killed, self.done = [], [], {}
         self.fail_starts, self.fail_kills = 0, 0
+        self.lost_starts = 0           # the run launches but the relay reports a failure (e.g. timeout)
+        self.fail_probes = self.fail_polls = 0
+        self.dirs = {}                 # run_dir -> files on the "terminal"
+        self.polls = []
+
+    def handle(self, run):
+        return {"run_dir": "/runs/" + (run.dir_name or run.run_id), "run_id": run.run_id, "thread_id": run.thread_id}
 
     def start(self, run):
         if self.fail_starts:
             self.fail_starts -= 1
-            raise RuntimeError("relay down")
+            raise TransportError("relay down")
+        h = self.handle(run)
+        if "pid.txt" in self.dirs.get(h["run_dir"], ()):
+            raise TransportError("HASPID")
         self.started.append(run)
-        return {"run_dir": "/runs/" + run.run_id, "run_id": run.run_id, "thread_id": run.thread_id}
+        self.dirs[h["run_dir"]] = {"launch.pid", "pid.txt"}
+        if self.lost_starts:
+            self.lost_starts -= 1
+            raise TransportError("relay timeout after 150s")
+        return h
+
+    def probe(self, handle):
+        if self.fail_probes:
+            self.fail_probes -= 1
+            raise TransportError("relay down")
+        return sorted(self.dirs.get(handle["run_dir"], ()))
 
     def poll(self, handle):
+        self.polls.append(handle["run_id"])
+        if self.fail_polls:
+            self.fail_polls -= 1
+            raise TransportError("relay down")
         return "finished" if handle["run_id"] in self.done else "running"
 
     def collect(self, handle):

tests/test_config.py

@@ -45,6 +45,12 @@ class ConfigTest(unittest.TestCase):
         d["terminals"]["oregon"]["concurrency"] = "two"
         self.check_both(d, "concurrency")
 
+    def test_concurrency_zero_is_valid(self):  # 0 = terminal paused (nothing new is dispatched)
+        d = config_data()
+        d["terminals"]["win10"]["concurrency"] = 0
+        self.assertEqual(config.validate(d), [])
+        self.assertEqual(config.mini_validate(d, SCHEMA), [])
+
     def test_semantic_routing_to_unknown_terminal(self):
         d = config_data()
         d["projects"]["GWENT"]["routing"]["by_assignee"]["mac"] = "mac"

tests/test_core.py

@@ -1,3 +1,4 @@
+import json
 import unittest
 
 from tests.fakes import Harness, config_data
@@ -160,6 +161,14 @@ class CoreTest(unittest.TestCase):
         self.assertEqual(started, ["GW-1-1", "GW-3-1"])  # 2 slots; GW-2 waits for the game group
         self.assertEqual([self.p.state_of(x) for x in (a, b, c, d)], ["In Progress", "Todo", "In Progress", "Todo"])
 
+    def test_concurrency_zero_pauses_terminal(self):
+        self.h.cfg.terminals["win10"]["concurrency"] = 0
+        w = self.p.add_issue(1)
+        o = self.p.add_issue(2, assignee="oregon")
+        self.h.tick()
+        self.assertNotIn(("win10", "codex"), self.h.execs)  # not even probed (busy_processes)
+        self.assertEqual((self.p.state_of(w), self.p.state_of(o)), ("Todo", "In Progress"))
+
     def test_label_routing(self):
         self.p.add_issue(1, labels=["model:astra"])
         self.p.add_issue(2, assignee="oregon", labels=["cap:gpu"])  # oregon lacks the capability
@@ -329,10 +338,13 @@ class CoreTest(unittest.TestCase):
         saves = []
         orig = self.h.d.store.save
         self.h.d.store.save = lambda data: (saves.append(sorted(
-            r["inflight"]["run_id"] for r in data["projects"]["GWENT"]["issues"].values() if r.get("inflight"))), orig(data))
+            ("L" if r.get("launching") else "") + (r.get("inflight") or r.get("launching"))["run_id"]
+            for r in data["projects"]["GWENT"]["issues"].values() if r.get("inflight") or r.get("launching"))), orig(data))
         self.h.tick()
-        # one save per launched run (a crash mid-tick cannot orphan it), then the end-of-tick save
-        self.assertEqual(saves, [["GW-1-1"], ["GW-1-1", "GW-2-1"], ["GW-1-1", "GW-2-1"]])
+        # the launching marker is on disk before each launch call, the run right after it (a crash mid-tick cannot
+        # orphan it or start it twice), then the end-of-tick save
+        self.assertEqual(saves, [["LGW-1-1"], ["GW-1-1"], ["GW-1-1", "LGW-2-1"], ["GW-1-1", "GW-2-1"],
+                                 ["GW-1-1", "GW-2-1"]])
 
 class EnvFailureTest(unittest.TestCase):
     def test_env_failure_pauses_group_until_item_leaves_waiting(self):
@@ -354,5 +366,142 @@ class EnvFailureTest(unittest.TestCase):
         h.close()
 
 
+
+class TierDefaultTest(unittest.TestCase):
+    """executors.<x>.tier_defaults: tier per model unless a label (`tier:<x>`, `tier:none`) sets one."""
+    def run_with(self, labels):
+        data = config_data()
+        data["executors"]["codex"]["tier_defaults"] = {"sol": "fast"}
+        h = Harness(data)
+        try:
+            iid = h.plane.add_issue(1, labels=labels)
+            h.tick()
+            return h.execs[("win10", "codex")].started[0].tier, h.plane.texts(iid)[-1]
+        finally:
+            h.close()
+
+    def test_default_model_gets_its_tier(self):
+        tier, text = self.run_with(())
+        self.assertEqual(tier, "fast")
+        self.assertIn("codex / sol / medium / fast @ win10", text)
+
+    def test_other_model_has_no_default(self):
+        self.assertIsNone(self.run_with(("model:astra",))[0])
+
+    def test_labels_override(self):
+        self.assertEqual(self.run_with(("tier:flex",))[0], "flex")
+        self.assertIsNone(self.run_with(("tier:none",))[0])
+
+
 if __name__ == "__main__":
     unittest.main()
+
+
+class DoubleStartGuardTest(unittest.TestCase):
+    """A launch whose outcome is unknown is probed on the next tick: adopted if it exists, relaunched in a fresh
+    directory if not; never started twice in the same directory."""
+    def setUp(self):
+        self.h = Harness()
+        self.p = self.h.plane
+
+    def tearDown(self):
+        self.h.close()
+
+    def codex(self):
+        return self.h.execs[("win10", "codex")]
+
+    def test_lost_launch_is_adopted_not_restarted(self):
+        iid = self.p.add_issue(1)
+        self.h.d.executor("win10", "codex").lost_starts = 1
+        self.h.tick()                                   # run starts, but the relay reports a timeout
+        rec = self.h.rec(iid)
+        self.assertNotIn("inflight", rec)
+        self.assertEqual((rec["launching"]["run_id"], rec["launching"]["handle"]["run_dir"]), ("GW-1-1", "/runs/GW-1-1"))
+        self.assertEqual(self.p.state_of(iid), "Todo")
+        self.h.tick()                                   # probe finds launch.pid/pid.txt -> adopt
+        self.assertEqual(len(self.codex().started), 1)
+        rec = self.h.rec(iid)
+        self.assertNotIn("launching", rec)
+        self.assertEqual((rec["inflight"]["run_id"], rec["runs"]), ("GW-1-1", 1))
+        self.assertEqual(self.p.state_of(iid), "In Progress")
+        self.assertIn("启动未确认、已接管", self.p.texts(iid)[-1])
+        self.codex().finish("[review] ok")
+        self.h.tick()
+        self.assertEqual(self.p.state_of(iid), "Review")
+
+    def test_failed_launch_relaunches_same_run_id_in_fresh_dir(self):
+        iid = self.p.add_issue(1)
+        self.codex_ex = self.h.d.executor("win10", "codex")
+        self.codex_ex.fail_starts = 1
+        self.h.tick()                                   # nothing launched
+        self.assertIn("launching", self.h.rec(iid))
+        self.h.tick()                                   # probe: nothing -> relaunch as GW-1-1 in GW-1-1-r2
+        run = self.codex().started[-1]
+        self.assertEqual((run.run_id, run.dir_name), ("GW-1-1", "GW-1-1-r2"))
+        rec = self.h.rec(iid)
+        self.assertEqual(rec["inflight"]["handle"]["run_dir"], "/runs/GW-1-1-r2")
+        self.assertNotIn("relaunch", rec)
+        self.assertIn("重新启动（目录 GW-1-1-r2）", self.p.texts(iid)[-1])
+
+    def test_unprobed_marker_holds_the_terminal_slot(self):
+        a = self.p.add_issue(1)
+        b = self.p.add_issue(2)
+        ex = self.h.d.executor("win10", "codex")
+        ex.lost_starts = 1
+        self.h.tick()
+        self.assertEqual([r.run_id for r in ex.started], ["GW-1-1"])   # GW-2 not started next to a maybe-live run
+        ex.fail_probes = 1
+        self.h.tick()                                   # relay down: marker kept, win10 skipped, GW-2 still waits
+        self.assertIn("launching", self.h.rec(a))
+        self.assertEqual(len(ex.started), 1)
+        self.assertEqual(self.h.d.store.load()["projects"]["GWENT"]["issues"].get(b), None)
+        self.h.tick()                                   # relay back: adopt GW-1-1; concurrency 1 keeps GW-2 waiting
+        self.assertEqual(self.h.rec(a)["inflight"]["run_id"], "GW-1-1")
+        self.assertEqual(len(ex.started), 1)
+
+    def test_marker_dropped_when_item_moved_away(self):
+        iid = self.p.add_issue(1)
+        self.h.d.executor("win10", "codex").fail_starts = 1
+        self.h.tick()
+        self.p.issues[iid]["state"] = "s-bl"            # a person parks it
+        self.h.tick()
+        rec = self.h.rec(iid)
+        self.assertNotIn("launching", rec)
+        self.assertNotIn("relaunch", rec)
+        self.assertEqual(self.codex().started, [])
+
+    def test_never_restarts_into_a_dir_with_a_pid_file(self):
+        iid = self.p.add_issue(1)
+        ex = self.h.d.executor("win10", "codex")
+        ex.dirs["/runs/GW-1-1"] = {"pid.txt"}         # leftover of a launch the state does not know about
+        self.h.tick()                                   # HASPID -> marker
+        self.assertIn("launching", self.h.rec(iid))
+        self.h.tick()                                   # probe finds pid.txt -> adopt instead of wiping
+        self.assertEqual(self.h.rec(iid)["inflight"]["handle"]["run_dir"], "/runs/GW-1-1")
+        self.assertEqual(ex.started, [])
+
+
+class RelayDownTest(unittest.TestCase):
+    def test_down_terminal_is_skipped_and_others_continue(self):
+        data = config_data()
+        data["terminals"]["win10"]["concurrency"] = 2
+        h = Harness(data)
+        w1, w2 = h.plane.add_issue(1), h.plane.add_issue(2)
+        o = h.plane.add_issue(3, assignee="oregon")
+        h.tick()
+        win, ore = h.execs[("win10", "codex")], h.execs[("oregon", "claude-code")]
+        self.assertEqual(len(win.started), 2)
+        ore.finish("[review] done")
+        w3 = h.plane.add_issue(4)
+        win.fail_polls, win.polls[:] = 1, []
+        h.tick()
+        self.assertEqual(len(win.polls), 1)             # second win10 run not polled once the relay failed
+        self.assertEqual(h.plane.state_of(o), "Review")  # oregon carried on
+        self.assertEqual(h.plane.state_of(w3), "Todo")  # nothing new started on win10
+        hb = json.loads((h.d.store.dir / "heartbeat.json").read_text())
+        self.assertEqual((hb["down"], hb["errors"]), (["win10"], []))
+        self.assertTrue(all("inflight" in h.rec(i) for i in (w1, w2)))
+        win.finish("[review] w")
+        h.tick()                                        # relay back: polled again
+        self.assertIn("win10", {r["terminal"] for r in json.loads((h.d.store.dir / "heartbeat.json").read_text())["inflight"]})
+        h.close()

tests/test_direct_ssh.py

@@ -0,0 +1,827 @@
+"""direct-ssh (PAD-12): config validation, channel selection and fallback rules, the ssh command line, the Windows
+bootstrap / payload, single-call launch and collect. Nothing here opens a network connection: the POSIX direct
+channel is replaced by a local `bash -c <loader>` fed on stdin, so the real loader and scripts run."""
+import base64
+import hashlib
+import io
+import json
+import os
+import shutil
+import subprocess
+import tempfile
+import unittest
+from contextlib import redirect_stdout
+from pathlib import Path
+from unittest import mock
+
+from agentd import ROOT, cli, config
+from agentd.config import Config
+from agentd.executors import Run, make_executor
+from agentd.transports import (CallLog, HostKeyError, RoutedTransport, TransportError, Unreached, make_transport,
+                               parse_fetch)
+from agentd.transports import direct_ssh
+from agentd.transports.direct_ssh import (BOOTSTRAP, PS_COMMAND, SH_COMMAND, DirectChannel, KeyAgent, classify,
+                                          ps_payload)
+from agentd.transports.warpgate_ssh import PosixWarpgateTransport, WarpgateSshTransport, WindowsWarpgateTransport
+from tests.fakes import config_data
+
+SCHEMA = json.loads((ROOT / "config" / "schema.json").read_text())
+WIN_KEY = "ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIFzLX7QNHDFYwTgY4wNl0ihZryIyKgpp9HPpPDWbrTX3"
+LIN_KEY = "ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAICUoI6W7npX+h/kHcXX3iy2XAWbNusrOjg28Xed75wWr"
+SECRET = "-----BEGIN OPENSSH PRIVATE KEY-----\nNOT-A-REAL-KEY-but-must-never-leak\n-----END OPENSSH PRIVATE KEY-----\n"
+
+
+def direct_data(root="/srv/x"):
+    d = config_data()
+    d["vault"] = {"client": "/opt/vault/agent-access.py", "config": "https://vault.example/identity.json"}
+    d["agentd"]["ssh_dir"] = str(Path(root) / "ssh")
+    d["terminals"]["win10"].update(transport="direct-ssh", fallback="warpgate-ssh",
+                                   ssh={"host": "100.123.209.19", "user": "administrator", "port": 22, "host_key": WIN_KEY})
+    d["terminals"]["lin"] = {"transport": "direct-ssh", "target": "lin", "os": "linux", "launch": "nohup",
+                             "repo": str(Path(root) / "repo"), "run_root": str(Path(root) / "runs"), "plane_user": "oregon",
+                             "token_key": "TOK_LIN", "concurrency": 1, "executors": ["claude-code", "shell"],
+                             "ssh": {"host": "100.127.58.78", "user": "ubuntu", "host_key": LIN_KEY},
+                             "fallback": "warpgate-ssh"}
+    return d
+
+
+class FakeAgent:
+    sock, pub, error = "/tmp/agentd-agent-x/agent.sock", "/tmp/agentd-agent-x/agentd.pub", None
+
+
+class ConfigTest(unittest.TestCase):
+    def errs(self, d):
+        return config.validate(d)
+
+    def test_direct_config_is_valid_and_live_config_uses_it(self):
+        self.assertEqual(self.errs(direct_data()), [])
+        self.assertEqual(config.mini_validate(direct_data(), SCHEMA), [])
+        cfg = config.load(ROOT / "config" / "agentd.yaml")
+        for term in ("win10", "oregon"):
+            t = cfg.terminals[term]
+            self.assertEqual((t["transport"], t["fallback"], t["target"]), ("direct-ssh", "warpgate-ssh", term))
+            self.assertIsNone(config.host_key_error(t["ssh"]["host_key"]))
+        self.assertIn("--secret-env AGENTD_SSH_KEY=AGENTD_SSH_KEY", cfg.secret_env_args())
+        config.load(ROOT / "config" / "agentd.oregon.yaml")   # rollback config still valid
+
+    def test_bad_direct_settings_have_clear_errors(self):
+        cases = [
+            (lambda d: d["terminals"]["lin"].pop("ssh"), "direct-ssh needs ssh:"),
+            (lambda d: d["terminals"]["lin"]["ssh"].update(host_key="ssh-rsa AAAAB3Nza"), "must be 'ssh-ed25519 <base64>'"),
+            (lambda d: d["terminals"]["lin"]["ssh"].update(host_key=WIN_KEY[:-4]), "not an ssh-ed25519 public key"),
+            (lambda d: d["terminals"]["lin"]["ssh"].update(host_key="ssh-ed25519 !!!"), "not an ssh-ed25519 public key"),
+            (lambda d: d["terminals"]["lin"].pop("target"), "needs the Warpgate 'target'"),
+            (lambda d: d["terminals"]["lin"]["ssh"].update(host="-oProxyCommand=x"), "host and user must be non-empty"),
+            (lambda d: d["terminals"]["lin"]["ssh"].update(port=70000), "> 65535"),
+            (lambda d: d["terminals"]["lin"].update(run_root="~/runs"), "must be an absolute path"),
+            (lambda d: d["terminals"]["win10"].update(launch="nohup"), "direct-ssh on windows needs launch=schtasks"),
+            (lambda d: d["terminals"]["win10"].update(target="win11"), "(fallback) only reach target 'win10'"),
+            (lambda d: d["vault"].pop("config"), "fallback warpgate-ssh needs vault.config"),
+            (lambda d: d["terminals"]["oregon"].update(fallback="warpgate-ssh"), "only applies to transport direct-ssh"),
+            (lambda d: d["agentd"].update(ssh_dir="/" + "x" * 60), "too long for the ControlMaster socket"),
+        ]
+        for mutate, fragment in cases:
+            d = direct_data()
+            mutate(d)
+            errs = self.errs(d)
+            self.assertTrue(any(fragment in e for e in errs), (fragment, errs))
+
+    def test_schema_rejects_unknown_ssh_key_and_fallback_value(self):
+        d = direct_data()
+        d["terminals"]["lin"]["ssh"]["password"] = "x"
+        self.assertTrue(any("password" in e for e in config.mini_validate(d, SCHEMA)))
+        d = direct_data()
+        d["terminals"]["lin"]["fallback"] = "local"
+        self.assertTrue(config.mini_validate(d, SCHEMA))
+
+    def test_key_env_is_always_scrubbed_and_only_requested_when_used(self):
+        cfg = Config(config_data())   # no direct-ssh terminal
+        self.assertNotIn("--secret-env AGENTD_SSH_KEY=AGENTD_SSH_KEY", cfg.secret_env_args())
+        with mock.patch.dict(os.environ, {"AGENTD_SSH_KEY": SECRET}):
+            self.assertNotIn("AGENTD_SSH_KEY", cfg.scrubbed_env())
+        d = direct_data()
+        d["agentd"]["ssh_key"] = "OTHER_KEY"
+        self.assertEqual(Config(d).secret_env_args()[-1], "--secret-env OTHER_KEY=OTHER_KEY")
+
+
+class SelectionTest(unittest.TestCase):
+    def setUp(self):
+        self.cfg = Config(direct_data())
+
+    def test_direct_terminal_gets_direct_primary_and_warpgate_fallback(self):
+        t = make_transport(self.cfg, "win10")
+        self.assertIsInstance(t, RoutedTransport)
+        self.assertIsInstance(t.primary, WindowsWarpgateTransport)
+        self.assertIsInstance(t.primary.channel, DirectChannel)
+        self.assertEqual((t.fallback.channel.via, t.fallback.channel.stdin, t.fallback.target), ("fallback", False, "win10"))
+        self.assertEqual(t.os, "windows")
+        self.assertTrue(hasattr(t, "start_run"))
+
+    def test_forced_channel_has_no_fallback(self):
+        self.assertIsNone(make_transport(self.cfg, "lin", force="direct-ssh").fallback)
+        t = make_transport(self.cfg, "lin", force="warpgate-ssh")
+        self.assertIsInstance(t, PosixWarpgateTransport)
+        self.assertEqual(t.channel.via, "warpgate")
+
+    def test_rollback_to_warpgate_keeps_ssh_section(self):
+        d = direct_data()
+        d["terminals"]["lin"]["transport"] = "warpgate-ssh"   # ssh/fallback stay, ignored
+        self.assertEqual(config.validate(d), [])
+        self.assertIsInstance(make_transport(Config(d), "lin"), PosixWarpgateTransport)
+
+
+class ClassifyTest(unittest.TestCase):
+    def test_classification(self):
+        self.assertEqual(classify(0, ""), "ok")
+        self.assertEqual(classify(1, "ssh: connect to host x port 22: Connection timed out"), "ok")   # command ran
+        for err in ("ssh: connect to host 100.1.2.3 port 22: Connection timed out",
+                    "ssh: connect to host 100.1.2.3 port 22: Connection refused",
+                    "ssh: connect to host 100.1.2.3 port 22: No route to host",
+                    "ssh: Could not resolve hostname x: Name or service not known",
+                    "Connection timed out during banner exchange",
+                    "kex_exchange_identification: read: Connection reset by peer",
+                    "Connection closed by 100.1.2.3 port 22",
+                    "ubuntu@100.1.2.3: Permission denied (publickey).",
+                    "sign_and_send_pubkey: signing failed for ED25519 \"x\" from agent: agent refused operation"):
+            self.assertEqual(classify(255, err), "unreached", err)
+        for err in ("@@@@@@@@@@@\n@    WARNING: REMOTE HOST IDENTIFICATION HAS CHANGED!     @\nHost key verification failed.",
+                    "No ED25519 host key is known for agentd-win10 and you have requested strict checking.\n"
+                    "Host key verification failed."):
+            self.assertEqual(classify(255, err), "hostkey")
+        for err in ("Connection to 100.1.2.3 closed by remote host.", "Timeout, server 100.1.2.3 not responding.",
+                    "client_loop: send disconnect: Connection reset by peer", ""):
+            self.assertEqual(classify(255, err), "unknown", err)   # the command may have run: no fallback
+
+
+class Prim:
+    """Stand-in transport recording calls; `fail` = exception raised by every call."""
+    os = "linux"
+
+    def __init__(self, fail=None):
+        self.fail, self.calls = fail, []
+
+    def join(self, *p):
+        return "/".join(p)
+
+    def start_run(self, *a):
+        self.calls.append(("start_run", a))
+        if self.fail:
+            raise self.fail
+        return "ok"
+
+    def process_count(self, name):
+        self.calls.append(("process_count", name))
+        if self.fail:
+            raise self.fail
+        return 0
+
+
+class RoutedTest(unittest.TestCase):
+    def setUp(self):
+        self.events = []
+        self.log = CallLog(event=self.events.append)
+
+    def test_unreached_falls_back_once_and_sticks(self):
+        p, f = Prim(Unreached("ssh: connect to host x: Connection timed out")), Prim()
+        t = RoutedTransport("lin", p, f, self.log)
+        self.assertEqual(t.start_run("task", "/r", {}, "run.sh"), "ok")
+        self.assertEqual(t.process_count("codex"), 0)
+        self.assertEqual(len(p.calls), 1)              # direct not retried within the tick
+        self.assertEqual([c[0] for c in f.calls], ["start_run", "process_count"])
+        self.assertEqual(len(self.events), 1)
+        self.assertIn("fallback", self.events[0])
+        self.assertEqual(t.join("a", "b"), "a/b")       # pure helpers come from the primary
+
+    def test_host_key_and_after_command_failures_never_fall_back(self):
+        for exc in (HostKeyError("host key verification FAILED"), TransportError("ssh timeout after 150s")):
+            p, f = Prim(exc), Prim()
+            t = RoutedTransport("lin", p, f, self.log)
+            with self.assertRaises(type(exc)):
+                t.start_run("task", "/r", {}, "run.sh")
+            self.assertEqual(f.calls, [])
+
+    def test_no_fallback_configured_raises(self):
+        with self.assertRaises(Unreached):
+            RoutedTransport("lin", Prim(Unreached("down")), None, self.log).start_run("t", "/r", {}, "run.sh")
+
+    def test_busy_check_unreached_is_not_swallowed(self):
+        cfg = Config(direct_data())
+        for cls, term in ((PosixWarpgateTransport, "lin"), (WindowsWarpgateTransport, "win10")):
+            ch = mock.Mock(stdin=True)
+            ch.sh.side_effect = ch.ps.side_effect = Unreached("down")
+            with self.assertRaises(Unreached):
+                cls(cfg, term, channel=ch).process_count("codex")
+            ch.sh.side_effect = ch.ps.side_effect = TransportError("exit 1")
+            self.assertIsNone(cls(cfg, term, channel=ch).process_count("codex"))
+
+
+class MissingKeyTest(unittest.TestCase):
+    def setUp(self):
+        direct_ssh._AGENTS.clear()
+        direct_ssh._KEYS.clear()
+        self.tmp = tempfile.TemporaryDirectory()
+        self.cfg = Config(direct_data(self.tmp.name))
+
+    def tearDown(self):
+        direct_ssh._AGENTS.clear()
+        self.tmp.cleanup()
+
+    def test_missing_key_makes_direct_unavailable_and_uses_the_fallback(self):
+        events = []
+        log = CallLog(event=events.append)
+        env = {k: v for k, v in os.environ.items() if k != "AGENTD_SSH_KEY"}
+        with mock.patch.dict(os.environ, env, clear=True), \
+                mock.patch("agentd.transports.direct_ssh.subprocess.run") as run:
+            t = make_transport(self.cfg, "lin", log=log)
+            t.fallback = Prim()
+            self.assertEqual(t.process_count("claude"), 0)
+            self.assertEqual(t.process_count("claude"), 0)
+            run.assert_not_called()                     # no ssh attempted at all
+        self.assertTrue(any("AGENTD_SSH_KEY is not set" in e for e in events))
+        self.assertEqual(sum("unavailable in this process" in e for e in events), 1)   # once per process
+        self.assertEqual(sum("using the warpgate-ssh fallback" in e for e in events), 1)  # once per tick
+
+    def test_empty_key_is_popped_and_reported(self):
+        with mock.patch.dict(os.environ, {"AGENTD_SSH_KEY": "  "}):
+            a = KeyAgent("AGENTD_SSH_KEY").start()
+            self.assertNotIn("AGENTD_SSH_KEY", os.environ)
+        self.assertIn("not set", a.error)
+
+    @unittest.skipUnless(all(shutil.which(b) for b in ("ssh-agent", "ssh-add", "ssh-keygen", "ssh")), "needs OpenSSH")
+    def test_real_agent_loads_key_from_env_and_cleans_up(self):
+        keyfile = Path(self.tmp.name) / "k"
+        subprocess.run(["ssh-keygen", "-q", "-t", "ed25519", "-N", "", "-C", "agentd-test", "-f", str(keyfile)], check=True)
+        key = keyfile.read_text()
+        with mock.patch.dict(os.environ, {"AGENTD_SSH_KEY": key.replace("\n", "\\n")}):   # Vault-style escapes
+            a = KeyAgent("AGENTD_SSH_KEY").start()
+            self.assertNotIn("AGENTD_SSH_KEY", os.environ)
+        try:
+            self.assertIsNone(a.error)
+            self.assertEqual(oct(os.stat(a.dir).st_mode & 0o777), "0o700")
+            self.assertEqual(Path(a.pub).read_text().split()[:2], (keyfile.with_suffix(".pub")).read_text().split()[:2])
+            for f in Path(a.dir).iterdir():                          # nothing on disk holds the private key
+                if f.is_file():
+                    self.assertNotIn("PRIVATE KEY", f.read_text())
+        finally:
+            a.close()
+        self.assertFalse(os.path.exists(a.dir))
+        self.assertIsNotNone(a.proc.poll())
+
+    def test_bad_key_reports_without_leaking(self):
+        if not all(shutil.which(b) for b in ("ssh-agent", "ssh-add", "ssh")):
+            self.skipTest("needs OpenSSH")
+        with mock.patch.dict(os.environ, {"AGENTD_SSH_KEY": SECRET}):
+            a = KeyAgent("AGENTD_SSH_KEY").start()
+        self.assertIn("ssh-add failed", a.error)
+        self.assertNotIn("NOT-A-REAL-KEY", a.error)
+        self.assertFalse(os.path.exists(a.dir))   # agent stopped, socket dir gone
+
+
+class SshCommandTest(unittest.TestCase):
+    def setUp(self):
+        self.tmp = tempfile.TemporaryDirectory()
+        self.cfg = Config(direct_data(self.tmp.name))
+        self.events, self.lines = [], []
+        self.log = CallLog(event=self.events.append)
+        self.log.call = lambda *a, **k: self.lines.append(a)
+
+    def tearDown(self):
+        self.tmp.cleanup()
+
+    def run_with(self, rc, err=b"", out=b"PONG=x\n"):
+        seen = {}
+
+        def fake_run(argv, input=None, stdout=None, stderr=None, env=None, timeout=None):
+            seen.update(argv=argv, input=input, env=env, timeout=timeout)
+            stdout.write(out)
+            stderr.write(err)
+            return mock.Mock(returncode=rc)
+        return seen, fake_run
+
+    def test_command_line_options_and_no_secret(self):
+        ch = DirectChannel(self.cfg, "lin", self.log)
+        seen, fake = self.run_with(0)
+        with mock.patch.dict(os.environ, {"AGENTD_SSH_KEY": SECRET, "TOK_LIN": "plane-token", "SSH_AUTH_SOCK": "/x"}), \
+                mock.patch("agentd.transports.direct_ssh.key_agent", return_value=FakeAgent), \
+                mock.patch("agentd.transports.direct_ssh.subprocess.run", fake):
+            rc, out, _ = ch.sh("echo hi", 30, "ping")
+        argv = seen["argv"]
+        opts = {argv[i + 1] for i, a in enumerate(argv) if a == "-o"}
+        ssh_dir = Path(self.tmp.name) / "ssh"
+        for o in ("BatchMode=yes", "IdentitiesOnly=yes", "IdentityAgent=" + FakeAgent.sock, "IdentityFile=" + FakeAgent.pub,
+                  "StrictHostKeyChecking=yes", "UserKnownHostsFile=%s/known_hosts.lin" % ssh_dir,
+                  "GlobalKnownHostsFile=/dev/null", "HostKeyAlias=agentd-lin", "HostKeyAlgorithms=ssh-ed25519",
+                  "ConnectTimeout=8", "ServerAliveInterval=15", "ServerAliveCountMax=3", "ForwardAgent=no",
+                  "ForwardX11=no", "LogLevel=ERROR", "ControlMaster=auto", "ControlPersist=60",
+                  "ControlPath=%s/%s-%%C" % (ssh_dir, ch.pin), "PasswordAuthentication=no"):
+            self.assertIn(o, opts)
+        self.assertEqual(argv[:3], ["ssh", "-F", "/dev/null"])
+        self.assertEqual(argv[-3:], ["--", "100.127.58.78", SH_COMMAND])
+        self.assertEqual(argv[argv.index("-l") + 1], "ubuntu")
+        self.assertEqual(seen["input"], b"echo hi")                 # the script travels on stdin
+        self.assertFalse(any("NOT-A-REAL-KEY" in a or "plane-token" in a for a in argv))
+        self.assertNotIn("AGENTD_SSH_KEY", seen["env"])
+        self.assertNotIn("TOK_LIN", seen["env"])
+        self.assertNotIn("SSH_AUTH_SOCK", seen["env"])
+        self.assertEqual((Path(ssh_dir) / "known_hosts.lin").read_text(), "agentd-lin %s\n" % LIN_KEY)
+        self.assertEqual(oct(os.stat(ssh_dir).st_mode & 0o777), "0o700")
+        self.assertEqual((rc, out), (0, "PONG=x\n"))
+        self.assertEqual(self.lines[-1][:3], ("lin", "direct", "ping"))
+        self.assertEqual(self.lines[-1][4], 0)
+
+    def test_host_key_failure_is_loud_sticky_and_not_unreached(self):
+        ch = DirectChannel(self.cfg, "lin", self.log)
+        seen, fake = self.run_with(255, b"No ED25519 host key is known for agentd-lin and you have requested strict "
+                                        b"checking.\r\nHost key verification failed.\r\n", b"")
+        with mock.patch("agentd.transports.direct_ssh.key_agent", return_value=FakeAgent), \
+                mock.patch("agentd.transports.direct_ssh.subprocess.run", side_effect=fake) as run:
+            with self.assertRaises(HostKeyError) as cm:
+                ch.sh("true", 30)
+            self.assertNotIsInstance(cm.exception, Unreached)
+            with self.assertRaises(HostKeyError):
+                ch.sh("true", 30)
+            self.assertEqual(run.call_count, 1)                    # no second attempt in this tick
+        self.assertTrue(self.events and self.events[0].startswith("SECURITY lin:"))
+
+    def test_connect_failure_is_unreached_and_timeout_is_not(self):
+        ch = DirectChannel(self.cfg, "lin", self.log)
+        _, fake = self.run_with(255, b"ssh: connect to host 100.127.58.78 port 22: Connection timed out\n", b"")
+        with mock.patch("agentd.transports.direct_ssh.key_agent", return_value=FakeAgent), \
+                mock.patch("agentd.transports.direct_ssh.subprocess.run", fake):
+            with self.assertRaises(Unreached):
+                ch.sh("true", 30)
+        with mock.patch("agentd.transports.direct_ssh.key_agent", return_value=FakeAgent), \
+                mock.patch("agentd.transports.direct_ssh.subprocess.run",
+                           side_effect=subprocess.TimeoutExpired("ssh", 30)):
+            with self.assertRaises(TransportError) as cm:
+                ch.sh("true", 30)
+            self.assertNotIsInstance(cm.exception, Unreached)    # may have run: resolve_launch probes instead
+
+    def test_windows_ps_uses_bootstrap_and_stdin(self):
+        ch = DirectChannel(self.cfg, "win10", self.log)
+        seen, fake = self.run_with(0, out=b"X=ok\r\n")
+        with mock.patch("agentd.transports.direct_ssh.key_agent", return_value=FakeAgent), \
+                mock.patch("agentd.transports.direct_ssh.subprocess.run", fake):
+            self.assertEqual(ch.ps("'X=ok'", 30), "X=ok\n")
+        self.assertEqual(seen["argv"][-2:], ["100.123.209.19", PS_COMMAND])
+        self.assertEqual(base64.b64decode(seen["input"]).decode("utf-8"), "'X=ok'")
+
+
+class BootstrapTest(unittest.TestCase):
+    def test_encoded_command_round_trip_and_length(self):
+        value = PS_COMMAND.split()[-1]
+        self.assertEqual(base64.b64decode(value).decode("utf-16-le"), BOOTSTRAP)
+        self.assertLess(len(PS_COMMAND), 2000)          # cmd.exe limit is 8191
+        self.assertIn("[Console]::In.ReadToEnd()", BOOTSTRAP)
+        script = "'é 中文 ''q'''\n"
+        self.assertTrue(ps_payload(script).isascii())
+        self.assertEqual(base64.b64decode(ps_payload(script)).decode("utf-8"), script)
+
+
+class FakeWinChannel:
+    """Direct-like Windows channel: answers the start / fetch scripts the way PowerShell would."""
+    stdin, name, via = True, "ssh fake", "direct"
+
+    def __init__(self, files=None, haspid=False):
+        self.calls, self.files, self.haspid = [], files or {}, haspid
+
+    def ps(self, script, timeout, op=None):
+        self.calls.append((op, script))
+        if op == "start":
+            if self.haspid:
+                return "HASPID\n"
+            out = ["WROTE=" + hashlib.sha256(base64.b64decode(b)).hexdigest()
+                   for b in __import__("re").findall(r"FromBase64String\('([A-Za-z0-9+/=]*)'\)", script)]
+            return "\n".join(out + ["LAUNCHED=0"]) + "\n"
+        if op == "fetch":
+            ops = [l.split()[:4] for l in script.splitlines() if l.split(" ")[0] in ("AdRead", "AdGrep", "AdTail")]
+            out = []
+            for fn, i, name, *rest in ops:
+                text = self.files.get(name.strip("'"))
+                if fn == "AdRead":
+                    out.append("M%s" % i if text is None else "F%s=%s" % (i, base64.b64encode(text.encode()).decode()))
+                else:
+                    if fn == "AdGrep":
+                        text = next((l for l in (text or "").splitlines() if "thread.started" in l), "")
+                    else:
+                        text = "\n".join((text or "").splitlines()[-int(rest[0]):])
+                    out.append("F%s=%s" % (i, base64.b64encode(text.encode()).decode()))
+            return "\n".join(out) + "\n"
+        return "FORGOT=ok\n"
+
+
+class WindowsDirectTest(unittest.TestCase):
+    def setUp(self):
+        self.cfg = config.load(ROOT / "config" / "agentd.yaml")
+
+    def transport(self, ch):
+        return WarpgateSshTransport(self.cfg, "win10", channel=ch)
+
+    def test_launch_is_one_call_with_exact_utf8_files(self):
+        ch = FakeWinChannel()
+        t = self.transport(ch)
+        ex = make_executor(self.cfg, "win10", "codex", t)
+        prompt = "中文 'quoted' \"dq\" $env:X `tick`\r\n" * 2000      # ~70 KB: far beyond any argv limit
+        h = ex.start(Run("GW-26-1", t.repo, prompt, "gpt-6-astra", "high"))
+        self.assertEqual([c[0] for c in ch.calls], ["start"])
+        script = ch.calls[0][1]
+        blobs = __import__("re").findall(r"FromBase64String\('([A-Za-z0-9+/=]*)'\)", script)
+        self.assertEqual(base64.b64decode(blobs[0]), prompt.encode("utf-8"))       # UTF-8, no BOM, bytes exact
+        self.assertFalse(base64.b64decode(blobs[1]).startswith(b"\xef\xbb\xbf"))
+        self.assertIn("Set-Location -LiteralPath", base64.b64decode(blobs[1]).decode("utf-8"))
+        self.assertLess(script.index("'HASPID'; return"), script.index("WriteAllBytes"))
+        self.assertLess(script.index("WriteAllBytes"), script.index("schtasks /Run /TN 'agentd-GW-26-1'"))
+        self.assertTrue(h["run_dir"].endswith("\\runtime\\agentd\\GW-26-1"))
+
+    def test_launch_refuses_a_started_directory(self):
+        t = self.transport(FakeWinChannel(haspid=True))
+        with self.assertRaises(TransportError) as cm:
+            t.start_run("agentd-GW-1-1", "C:\\r\\GW-1-1", {"prompt.md": "p", "run.ps1": "r"}, "run.ps1")
+        self.assertIn("HASPID", str(cm.exception))
+
+    def test_collect_is_one_call_plus_forget(self):
+        files = {"exit.txt": "\ufeffexit=0 at 2026-10-22T10:00:00", "last.md": "[review] done\n",
+                 "events.jsonl": '{"type":"x"}\n{"type":"thread.started","thread_id":"th-42"}\n{"type":"turn.completed"}',
+                 "stderr.txt": "warn 1\nwarn 2"}
+        ch = FakeWinChannel(files)
+        ex = make_executor(self.cfg, "win10", "codex", self.transport(ch))
+        h = {"run_dir": "C:\\r\\GW-1-1", "task": "agentd-GW-1-1", "thread_id": None}
+        res = ex.collect(h)
+        ex.forget(h)
+        self.assertEqual([c[0] for c in ch.calls], ["fetch", "forget"])
+        self.assertEqual((res.last_message, res.exit_code, res.thread_id), ("[review] done", 0, "th-42"))
+        self.assertEqual(res.tail, 'warn 1\nwarn 2\n{"type":"turn.completed"}')
+
+    def test_warpgate_fetch_is_one_call_when_it_fits(self):
+        cfg = config.load(ROOT / "config" / "agentd.example.yaml")
+        t = WindowsWarpgateTransport(cfg, "win10")
+        self.assertFalse(t.channel.stdin)
+        ops = make_executor(cfg, "win10", "codex", t).COLLECT
+        script = t.fetch_script(t.join(t.default_run_root(), "GW-26-1-r2"), ops)
+        self.assertTrue(t.channel.fits(script))
+        calls = []
+        t.ps = lambda s, timeout=150: (calls.append(s), "\n".join("M%d" % i for i in range(len(ops))))[1]
+        self.assertEqual(t.fetch("C:\\r", ops), [None, None, None, "", ""])
+        self.assertEqual(len(calls), 1)
+        self.assertFalse(t.channel.fits("x" * 3000))
+
+
+class LocalDirect:
+    """POSIX direct channel stand-in: runs the real loader (SH_COMMAND) through a local login-style shell with the
+    script on stdin, exactly what sshd does with the remote command."""
+    stdin, name, via = True, "ssh local", "direct"
+
+    def __init__(self):
+        self.calls = []
+
+    def sh(self, script, timeout, op=None):
+        self.calls.append(op)
+        p = subprocess.run(["sh", "-c", SH_COMMAND], input=script.encode("utf-8"),   # as sshd: $SHELL -c "<command>"
+                           capture_output=True, timeout=timeout)
+        return p.returncode, p.stdout.decode("utf-8", "replace"), p.stderr.decode("utf-8", "replace")
+
+
+class PosixDirectTest(unittest.TestCase):
+    def setUp(self):
+        self.tmp = tempfile.TemporaryDirectory()
+        root = Path(self.tmp.name)
+        (root / "repo").mkdir()
+        self.cfg = Config(direct_data(root))
+        self.ch = LocalDirect()
+        self.t = PosixWarpgateTransport(self.cfg, "lin", channel=self.ch)
+
+    def tearDown(self):
+        self.tmp.cleanup()
+
+    def test_loader_quoting_matches_sh_quote(self):
+        self.assertEqual(SH_COMMAND, "bash -c 'exec 3<&0 </dev/null; s=$(cat <&3); exec 3<&-; (eval \"$s\"); r=$?; "
+                                     "[ $r = 255 ] && r=254; exit $r'")
+
+    def test_loader_never_exits_255(self):
+        self.assertEqual(self.ch.sh("echo x >&2; exit 255", 10)[0], 254)   # 255 = ssh's own "no session" code
+        self.assertEqual(self.ch.sh("exit 3", 10)[0], 3)
+        self.assertEqual(self.ch.sh("true", 10)[0], 0)
+
+    def test_big_prompt_launch_and_collect_are_one_call_each(self):
+        ex = make_executor(self.cfg, "lin", "shell", self.t)
+        prompt = ("ünïcode 'q' \"d\" $(nope) `x`\n" * 6000)          # ~200 KB > CHUNK and > MAX_ARG_STRLEN
+        h = ex.start(Run("GW-7-1", self.cfg.terminals["lin"]["repo"], prompt, script="wc -c"))
+        self.assertEqual(self.ch.calls, ["start"])
+        self.assertEqual(Path(h["run_dir"], "prompt.md").read_text(encoding="utf-8"), prompt)
+        for _ in range(100):
+            if ex.poll(h) == "finished":
+                break
+            __import__("time").sleep(0.1)
+        n = len(self.ch.calls)
+        res = ex.collect(h)
+        self.assertEqual(self.ch.calls[n:], ["fetch"])
+        self.assertEqual((res.last_message, res.exit_code), (str(len(prompt.encode("utf-8"))), 0))
+        with self.assertRaises(TransportError):          # pid.txt present: never wiped, never relaunched there
+            ex.start(Run("GW-7-1", self.cfg.terminals["lin"]["repo"], "again", script="true"))
+        self.assertEqual(Path(h["run_dir"], "prompt.md").read_text(encoding="utf-8"), prompt)
+
+    def test_fetch_missing_files_and_stdin_is_not_the_scripts(self):
+        d = Path(self.tmp.name) / "runs" / "x"
+        d.mkdir(parents=True)
+        (d / "a.txt").write_text("l1\nl2 needle\nl3\n")
+        ops = [("read", "a.txt"), ("read", "none"), ("grep", "a.txt", "needle"), ("grep", "none", "x"),
+               ("tail", "a.txt", 2), ("tail", "none", 3)]
+        self.assertEqual(self.t.fetch(str(d), ops), ["l1\nl2 needle\nl3\n", None, "l2 needle", None, "l2 needle\nl3", ""])
+        rc, out, _ = self.ch.sh("cat; echo done", 10)    # the script's own stdin is /dev/null
+        self.assertEqual((rc, out), (0, "done\n"))
+
+    def test_parse_fetch_rejects_garbage(self):
+        with self.assertRaises(TransportError):
+            parse_fetch("nothing here", [("read", "x")])
+
+
+class CallLogTest(unittest.TestCase):
+    def test_line_format_and_once(self):
+        with tempfile.TemporaryDirectory() as tmp:
+            events = []
+            log = CallLog(Path(tmp) / "sub" / "remote.log", events.append)
+            log.call("win10", "direct", "start", 812.4, 0)
+            log.call("win10", "fallback", "busy", 9100, 255, "unreached")
+            lines = (Path(tmp) / "sub" / "remote.log").read_text().splitlines()
+            self.assertTrue(lines[0].endswith("term=win10 via=direct op=start ms=812 rc=0"))
+            self.assertTrue(lines[1].endswith("via=fallback op=busy ms=9100 rc=255 unreached"))
+            log.event("a", once="k")
+            log.event("a", once="k")
+            self.assertEqual(events, ["a"])
+
+
+class PingCliTest(unittest.TestCase):
+    def test_ping_prints_timing_and_forces_channel(self):
+        cfg = Config(direct_data())
+        seen = {}
+
+        def fake_make(cfg_, term, log=None, force=None):
+            seen.update(term=term, force=force)
+            return mock.Mock(ping=mock.Mock(return_value="PONG=Linux 6.8 oregon"))
+        buf = io.StringIO()
+        with mock.patch("agentd.cli.make_transport", fake_make), redirect_stdout(buf), self.assertRaises(SystemExit) as cm:
+            cli.cmd_ping(cfg, cli.parser().parse_args(["ping", "lin", "--transport", "warpgate-ssh", "-n", "2"]))
+        self.assertEqual(cm.exception.code, 0)
+        self.assertEqual(seen, {"term": "lin", "force": "warpgate-ssh"})
+        self.assertEqual(buf.getvalue().count("PONG=Linux 6.8 oregon"), 2)
+
+
+class ReviewFixesTest(unittest.TestCase):
+    """Follow-ups of the PAD-12 review: classification limits, socket and known_hosts names, key handling."""
+    def setUp(self):
+        direct_ssh._KEYS.clear()
+        self.tmp = tempfile.TemporaryDirectory()
+        self.cfg = Config(direct_data(self.tmp.name))
+
+    def tearDown(self):
+        direct_ssh._KEYS.clear()
+        self.tmp.cleanup()
+
+    def test_output_or_remote_stderr_never_counts_as_unreached(self):
+        self.assertEqual(classify(255, "Connection reset by 100.1.2.3 port 22", "WROTE=ab\n"), "unknown")
+        self.assertEqual(classify(255, "git: ssh: connect to host github.com port 22: Connection timed out"), "unknown")
+        self.assertEqual(classify(255, "fatal: Could not read from remote repository"), "unknown")
+
+    def test_no_ed25519_offered_is_a_host_key_failure(self):
+        self.assertEqual(classify(255, "Unable to negotiate with 100.1.2.3 port 22: no matching host key type found. "
+                                       "Their offer: ssh-rsa"), "hostkey")
+        self.assertEqual(classify(255, "Unable to negotiate with 100.1.2.3 port 22: no matching cipher found. "
+                                       "Their offer: aes128-cbc"), "unreached")
+
+    def test_reset_after_output_is_not_retried_elsewhere(self):
+        ch = DirectChannel(self.cfg, "lin")
+
+        def fake_run(argv, input=None, stdout=None, stderr=None, env=None, timeout=None):
+            stdout.write(b"WROTE=1\n")
+            stderr.write(b"Connection reset by 100.127.58.78 port 22\r\n")
+            return mock.Mock(returncode=255)
+        with mock.patch("agentd.transports.direct_ssh.key_agent", return_value=FakeAgent), \
+                mock.patch("agentd.transports.direct_ssh.subprocess.run", fake_run):
+            with self.assertRaises(TransportError) as cm:
+                ch.sh("true", 30)
+        self.assertNotIsInstance(cm.exception, Unreached)
+
+    def test_control_path_follows_the_pinned_key(self):
+        a = DirectChannel(self.cfg, "lin")
+        d = direct_data(self.tmp.name)
+        d["terminals"]["lin"]["ssh"]["host_key"] = WIN_KEY
+        b = DirectChannel(Config(d), "lin")
+        path = lambda ch: [x for x in ch.argv(FakeAgent, "true") if x.startswith("ControlPath=")][0]  # noqa: E731
+        self.assertNotEqual(path(a), path(b))
+        self.assertEqual(path(a), "ControlPath=%s/%s-%%C" % (self.cfg.ssh_dir, a.pin))
+        self.assertEqual(len(a.pin), 8)
+        long_dir = direct_data("/" + "x" * 40)
+        self.assertTrue(any("too long" in e for e in config.validate(long_dir)))   # 41 + 67 > 104
+
+    def test_known_hosts_temp_files_are_per_terminal_and_process(self):
+        d = direct_ssh.control_dir(self.cfg)
+        tmps = []
+        real = os.replace
+        with mock.patch("agentd.transports.direct_ssh.os.replace",
+                        side_effect=lambda a, b: (tmps.append(str(a)), real(a, b))):
+            DirectChannel(self.cfg, "win10").known_hosts(d)
+            DirectChannel(self.cfg, "lin").known_hosts(d)
+        self.assertEqual(len(set(tmps)), 2)
+        self.assertTrue(all(str(os.getpid()) in t for t in tmps))
+        self.assertEqual((d / "known_hosts.win10").read_text(), "agentd-win10 %s\n" % WIN_KEY)
+        self.assertEqual((d / "known_hosts.lin").read_text(), "agentd-lin %s\n" % LIN_KEY)
+        self.assertEqual(sorted(f.name for f in d.iterdir()), ["known_hosts.lin", "known_hosts.win10"])
+
+    def test_key_leaves_the_environment_at_cli_start(self):
+        buf = io.StringIO()
+        with mock.patch.dict(os.environ, {"AGENTD_SSH_KEY": SECRET}), redirect_stdout(buf):
+            cli.main(["--config", str(ROOT / "config" / "agentd.yaml"), "check-config"])
+            self.assertNotIn("AGENTD_SSH_KEY", os.environ)
+        self.assertIn("direct-ssh key AGENTD_SSH_KEY: set", buf.getvalue())
+        self.assertNotIn("NOT-A-REAL-KEY", buf.getvalue())
+        self.assertTrue(direct_ssh.key_present("AGENTD_SSH_KEY"))
+        with mock.patch("agentd.transports.direct_ssh.shutil.which", return_value=None):
+            a = KeyAgent("AGENTD_SSH_KEY").start()      # the agent takes the key over from the CLI
+        self.assertIn("not installed", a.error)
+        self.assertFalse(direct_ssh.key_present("AGENTD_SSH_KEY"))
+
+    def test_local_busy_check_gets_the_scrubbed_env(self):
+        cfg = Config(config_data())
+        from agentd.transports.local import LocalTransport
+        with mock.patch.dict(os.environ, {"AGENTD_SSH_KEY": SECRET, "TOK_OR": "plane-token"}), \
+                mock.patch("agentd.transports.local.subprocess.run", return_value=mock.Mock(stdout="2\n")) as run:
+            self.assertEqual(LocalTransport(cfg, "oregon").process_count("codex"), 2)
+        env = run.call_args.kwargs["env"]
+        self.assertNotIn("AGENTD_SSH_KEY", env)
+        self.assertNotIn("TOK_OR", env)
+
+    def test_task_names_are_quoted_in_kill_and_forget(self):
+        ch = mock.Mock(stdin=True)
+        ch.ps.return_value = "KILLED=ok"
+        t = WindowsWarpgateTransport(self.cfg, "win10", channel=ch)
+        t.kill("agentd-a'b", "C:\\runs\\x")
+        t.forget("agentd-a'b")
+        for c in ch.ps.call_args_list:
+            self.assertIn("'agentd-a''b'", c.args[0])
+            self.assertNotIn("'agentd-a'b'", c.args[0])
+
+
+class FakeTerm:
+    """Remote transport stand-in for core-level routing tests: pure helpers from the real Windows transport,
+    start_run / probe on a dict of run directories shared by the direct and fallback channels. `fail` = exceptions
+    for the next calls: Unreached and HostKeyError strike before the command ran, any other after it (the run
+    exists). A host-key failure sticks, like DirectChannel's."""
+
+    def __init__(self, cfg, dirs):
+        self.real = WindowsWarpgateTransport(cfg, "win10", channel=mock.Mock(stdin=True))
+        self.dirs, self.fail, self.calls, self.hostkey = dirs, [], [], None
+
+    def __getattr__(self, name):
+        return getattr(self.real, name)
+
+    def _next(self, name):
+        self.calls.append(name)
+        if self.hostkey:
+            raise self.hostkey
+        exc = self.fail.pop(0) if self.fail else None
+        if isinstance(exc, HostKeyError):
+            self.hostkey = exc
+        return exc
+
+    def start_run(self, task, run_dir, files, script_name):
+        exc = self._next("start_run")
+        if isinstance(exc, (Unreached, HostKeyError)):
+            raise exc
+        if run_dir in self.dirs:
+            raise TransportError("HASPID")
+        self.dirs[run_dir] = ["launch.pid"]
+        if exc:
+            raise exc
+
+    def probe(self, run_dir, names):
+        exc = self._next("probe")
+        if exc:
+            raise exc
+        return list(self.dirs.get(run_dir, ()))
+
+
+class CoreRoutingTest(unittest.TestCase):
+    """Agentd.start -> codex executor -> RoutedTransport: the double-launch guarantee end to end."""
+
+    def setUp(self):
+        from tests.fakes import Harness
+        self.h = Harness(direct_data())
+        self.dirs, self.events = {}, []
+        self.direct, self.fallback = FakeTerm(self.h.cfg, self.dirs), FakeTerm(self.h.cfg, self.dirs)
+        log = CallLog(event=self.events.append)
+        self.h.d.executor_factory = lambda term, name: make_executor(   # a fresh routed transport per tick
+            self.h.cfg, term, name, RoutedTransport(term, self.direct, self.fallback, log))
+
+    def tearDown(self):
+        self.h.close()
+
+    def test_unreached_launch_moves_to_the_fallback_once(self):
+        iid = self.h.plane.add_issue(1)
+        self.direct.fail = [Unreached("ssh: connect to host 100.123.209.19 port 22: Connection timed out")]
+        self.h.tick()
+        rec = self.h.rec(iid)
+        self.assertNotIn("launching", rec)
+        self.assertEqual(rec["inflight"]["run_id"], "GW-1-1")
+        self.assertEqual((self.direct.calls, self.fallback.calls), (["start_run"], ["start_run"]))
+        self.assertEqual(len(self.dirs), 1)
+        self.assertEqual(self.h.plane.state_of(iid), "In Progress")
+        self.assertTrue(any("warpgate-ssh fallback" in e for e in self.events))
+
+    def test_failure_after_the_command_ran_is_probed_and_adopted_not_relaunched(self):
+        iid = self.h.plane.add_issue(1)
+        self.direct.fail = [TransportError("ssh 100.123.209.19 timeout after 150s")]
+        self.h.tick()
+        rec = self.h.rec(iid)
+        self.assertIn("launching", rec)
+        self.assertNotIn("inflight", rec)
+        self.assertEqual(self.fallback.calls, [])
+        self.h.tick()                                   # probe finds launch.pid -> adopt
+        rec = self.h.rec(iid)
+        self.assertNotIn("launching", rec)
+        self.assertEqual(rec["inflight"]["handle"]["run_dir"], list(self.dirs)[0])
+        self.assertEqual(self.direct.calls, ["start_run", "probe"])
+        self.assertEqual((self.fallback.calls, len(self.dirs)), ([], 1))
+
+    def test_host_key_failure_keeps_the_marker_and_never_falls_back(self):
+        iid = self.h.plane.add_issue(1)
+        self.direct.fail = [HostKeyError("host key verification FAILED")]
+        self.h.tick()
+        self.assertIn("launching", self.h.rec(iid))
+        self.h.tick()                                   # the probe fails the same way: marker kept, slot held
+        rec = self.h.rec(iid)
+        self.assertIn("launching", rec)
+        self.assertNotIn("inflight", rec)
+        self.assertEqual(self.direct.calls, ["start_run", "probe"])
+        self.assertEqual((self.fallback.calls, self.dirs), ([], {}))
+        self.assertEqual(self.h.plane.state_of(iid), "Todo")
+
+
+FAKE_CLIENT = """import json, os, sys
+argv = sys.argv[1:]
+with open(os.environ["FAKE_CALLS"], "a") as fh:
+    fh.write(json.dumps(argv) + "\\n")
+if os.environ.get("FAKE_MISSING") and "AGENTD_SSH_KEY=AGENTD_SSH_KEY" in argv:
+    print("access-hub: credential_unavailable", file=sys.stderr)
+    sys.exit(1)
+print("ticked")
+print("agentd stderr line", file=sys.stderr)
+sys.exit(int(os.environ.get("FAKE_RC", "0")))
+"""
+
+
+@unittest.skipUnless(all(shutil.which(b) for b in ("bash", "flock", "timeout")), "needs bash, flock, timeout")
+class CronScriptTest(unittest.TestCase):
+    """cron/agentd-tick-lxc.sh with a fake agent-access.py: a missing AGENTD_SSH_KEY must not stop the tick."""
+
+    def setUp(self):
+        import re
+        self.tmp = tempfile.TemporaryDirectory()
+        d = Path(self.tmp.name)
+        (d / "client.py").write_text(FAKE_CLIENT)
+        text = (ROOT / "cron" / "agentd-tick-lxc.sh").read_text()
+        for var, value in (("LOCK", d / "cron.lock"), ("LOG", d / "cron.log"), ("CLIENT", "(python3 -I %s)" % (d / "client.py"))):
+            text, n = re.subn(r"(?m)^%s=.*$" % var, "%s=%s" % (var, value), text)
+            self.assertEqual(n, 1, var)
+        self.script, self.calls, self.log = d / "tick.sh", d / "calls", d / "cron.log"
+        self.script.write_text(text)
+
+    def tearDown(self):
+        self.tmp.cleanup()
+
+    def run_tick(self, **env):
+        e = dict(os.environ, FAKE_CALLS=str(self.calls), **env)
+        rc = subprocess.run(["bash", str(self.script)], env=e, timeout=60).returncode
+        calls = [json.loads(l) for l in self.calls.read_text().splitlines()] if self.calls.exists() else []
+        return rc, calls, self.log.read_text() if self.log.exists() else ""
+
+    def test_normal_tick_passes_the_key(self):
+        rc, calls, log = self.run_tick()
+        self.assertEqual((rc, len(calls)), (0, 1))
+        self.assertIn("--secret-env", calls[0])
+        self.assertIn("AGENTD_SSH_KEY=AGENTD_SSH_KEY", calls[0])
+        self.assertEqual(calls[0][-3:], ["/opt/plane-agentd/config/agentd.yaml", "tick", "--once"])
+        self.assertIn("ticked", log)
+        self.assertIn("agentd stderr line", log)
+
+    def test_missing_vault_key_repeats_the_tick_without_it(self):
+        rc, calls, log = self.run_tick(FAKE_MISSING="1")
+        self.assertEqual((rc, len(calls)), (0, 2))
+        self.assertNotIn("AGENTD_SSH_KEY=AGENTD_SSH_KEY", calls[1])
+        self.assertIn("PLANE_OREGON_API_TOKEN=PLANE_OREGON_API_TOKEN", calls[1])
+        self.assertIn("AGENTD_SSH_KEY not available from Vault", log)
+        self.assertIn("ticked", log)
+
+    def test_a_failing_tick_is_not_repeated(self):
+        rc, calls, _ = self.run_tick(FAKE_RC="1")
+        self.assertEqual((rc, len(calls)), (1, 1))
+
+    def test_lock_held_exits_75(self):
+        import fcntl
+        with open(Path(self.tmp.name) / "cron.lock", "a") as fh:
+            fcntl.flock(fh, fcntl.LOCK_EX | fcntl.LOCK_NB)
+            rc, calls, _ = self.run_tick()
+        self.assertEqual((rc, calls), (75, []))
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_posix_remote.py

@@ -0,0 +1,221 @@
+"""Remote POSIX transport (warpgate-ssh, os linux/darwin): the generated bash is executed by a local bash in
+place of the relay, so quoting, chunked uploads, detached launch and process-group kill are exercised for real."""
+import base64
+import os
+import shutil
+import subprocess
+import tempfile
+import time
+import unittest
+from pathlib import Path
+from unittest import mock
+
+from agentd import ROOT, config
+from agentd.config import Config
+from agentd.executors import Run, make_executor
+from agentd.transports import make_transport
+from agentd.transports.warpgate_ssh import PosixWarpgateTransport, relay_argv
+from tests.fakes import config_data
+
+FAKE_CLAUDE = """#!/usr/bin/env bash
+prompt="$(cat)"
+echo "args: $*" >&2
+printf '{"type":"result","subtype":"success","is_error":false,"result":"[review] got %s","session_id":"sess-7"}' "${#prompt}"
+"""
+
+
+def remote_data(root):
+    d = config_data()
+    d["vault"] = {"client": "/opt/vault/agent-access.py", "config": "https://vault.example/identity.json"}
+    d["terminals"]["mac"] = {"transport": "warpgate-ssh", "target": "mac-target", "os": "linux", "launch": "nohup",
+                             "repo": str(root / "my repo's"), "run_root": str(root / "run root"), "plane_user": "oregon",
+                             "token_key": "TOK_MAC", "concurrency": 1, "executors": ["claude-code", "shell"],
+                             "bins": {"claude": str(root / "bin dir" / "claude")}}
+    return d
+
+
+class LocalBash(PosixWarpgateTransport):
+    """The relay replaced by `bash -c` on this machine; records every script (one entry = one relay call)."""
+    def __init__(self, cfg, terminal):
+        super().__init__(cfg, terminal)
+        self.calls = []
+
+    def sh(self, script, timeout=150, marker=None):
+        self.calls.append(script)
+        out = subprocess.run(["bash", "-c", script], capture_output=True, text=True, timeout=timeout,
+                             env=self.cfg.scrubbed_env())
+        if marker and marker not in out.stdout:
+            raise AssertionError("marker %s missing: %s %s" % (marker, out.stdout, out.stderr))
+        return out.stdout
+
+
+def wait(ex, h, seconds=10):
+    end = time.time() + seconds
+    while time.time() < end:
+        if ex.poll(h) == "finished":
+            return True
+        time.sleep(0.1)
+    return False
+
+
+class PosixRemoteTest(unittest.TestCase):
+    def setUp(self):
+        self.tmp = tempfile.TemporaryDirectory()
+        self.root = Path(self.tmp.name)
+        (self.root / "my repo's").mkdir()
+        (self.root / "bin dir").mkdir()
+        fake = self.root / "bin dir" / "claude"
+        fake.write_text(FAKE_CLAUDE)
+        fake.chmod(0o755)
+        self.cfg = Config(remote_data(self.root))
+        self.t = LocalBash(self.cfg, "mac")
+
+    def tearDown(self):
+        self.tmp.cleanup()
+
+    def test_factory_picks_posix_for_linux_and_config_is_valid(self):
+        self.assertEqual(config.validate(remote_data(self.root)), [])
+        self.assertIsInstance(make_transport(self.cfg, "mac"), PosixWarpgateTransport)
+        self.assertEqual(make_transport(self.cfg, "win10").os, "windows")
+
+    def test_live_config_runs_oregon_remotely(self):
+        cfg = config.load(ROOT / "config" / "agentd.yaml")
+        t = make_transport(cfg, "oregon")   # PAD-12: direct-ssh over Tailscale, Warpgate relay as fallback
+        self.assertIsInstance(t.primary, PosixWarpgateTransport)
+        self.assertIsInstance(t.fallback, PosixWarpgateTransport)
+        self.assertEqual((t.primary.channel.stdin, t.fallback.channel.stdin), (True, False))
+        self.assertEqual((t.fallback.target, t.os), ("oregon", "linux"))
+        self.assertEqual(cfg.executors["codex"]["tier_defaults"], {"gpt-6.1-sol": "fast"})
+
+    def test_relay_argv_and_scrubbed_env(self):
+        argv = relay_argv(self.cfg, "mac-target", ["bash", "-c", "echo hi"], 30)
+        self.assertEqual(argv[1:], ["-I", "/opt/vault/agent-access.py", "--config", "https://vault.example/identity.json",
+                                    "run", "--target", "mac-target", "--timeout", "30", "--", "bash", "-c", "echo hi"])
+        seen = {}
+
+        def fake_run(args, **kw):
+            seen.update(args=args, env=kw["env"])
+            return mock.Mock(returncode=0, stdout=b"MKDIR=ok\n", stderr=b"")
+        with mock.patch.dict(os.environ, {"TOK_MAC": "s", "TOK_W10": "s"}), \
+                mock.patch("agentd.transports.warpgate_ssh.subprocess.run", fake_run):
+            PosixWarpgateTransport(self.cfg, "mac").mkdir("/x")
+        self.assertNotIn("TOK_MAC", seen["env"])
+        self.assertNotIn("TOK_W10", seen["env"])
+        self.assertEqual(seen["args"][-3:-1], ["bash", "-c"])
+
+    def test_claude_round_is_one_call_and_resume_uses_session_id(self):
+        ex = make_executor(self.cfg, "mac", "claude-code", self.t)
+        prompt = "it's \"quoted\" $HOME `x` \\ 中文\n"
+        h = ex.start(Run("GW-4-1", self.cfg.terminals["mac"]["repo"], prompt, model="sonnet"))
+        self.assertEqual(len(self.t.calls), 1)        # mkdir + prompt + run.sh + launch in one relay call
+        self.assertIn("setsid nohup bash 'run.sh' </dev/null >/dev/null 2>&1 &", self.t.calls[0])
+        self.assertTrue(wait(ex, h))
+        rd = Path(h["run_dir"])
+        self.assertEqual((rd / "prompt.md").read_text(encoding="utf-8"), prompt)
+        self.assertEqual((rd / "pid.txt").read_text().strip(), (rd / "launch.pid").read_text().strip())
+        res = ex.collect(h)
+        self.assertEqual((res.last_message, res.exit_code, res.thread_id),
+                         ("[review] got %d" % len(prompt.rstrip("\n")), 0, "sess-7"))
+        cmd = ex.command(Run("GW-4-2", "r", "p", model="sonnet", thread_id="sess-7"), h["run_dir"])
+        self.assertEqual(cmd, "'%s' -p --resume sess-7" % (self.root / "bin dir" / "claude"))
+
+    def test_runner_quotes_paths_and_cd_failure_writes_exit(self):
+        _, script = self.t.runner("/a b/O'Neil/run", "/no such/repo's", "cmd", "out.json")
+        self.assertIn("cd '/no such/repo'\\''s' ||", script)
+        self.assertIn("< '/a b/O'\\''Neil/run'/prompt.md > '/a b/O'\\''Neil/run'/out.json", script)
+        ex = make_executor(self.cfg, "mac", "shell", self.t)
+        h = ex.start(Run("GW-4-3", str(self.root / "missing"), "", script="true"))
+        self.assertTrue(wait(ex, h))
+        self.assertEqual(ex.exit_code(h), 97)
+
+    def test_chunked_upload_and_reads(self):
+        self.t.CHUNK = 64
+        path = str(self.root / "run root" / "x y" / "f'.txt")
+        text = "line1 'q' \"d\" $(nope)\n" * 20 + "last ünïcode"
+        self.t.mkdir(os.path.dirname(path))
+        n = len(self.t.calls)
+        self.t.write_text(path, text)
+        self.assertGreater(len(self.t.calls) - n, 2)
+        self.assertEqual(Path(path).read_text(encoding="utf-8"), text)
+        self.assertFalse(os.path.exists(path + ".b64"))
+        self.assertEqual(self.t.read_text(path), text)
+        self.assertIsNone(self.t.read_text(path + ".missing"))
+        self.assertEqual(self.t.grep_first(path, "$(nope)"), "line1 'q' \"d\" $(nope)")
+        self.assertIsNone(self.t.grep_first(path, "absent"))
+        self.assertEqual(self.t.tail(path, 2), "line1 'q' \"d\" $(nope)\nlast ünïcode")
+        self.assertEqual(self.t.tail(path + ".missing", 2), "")
+
+    def test_big_prompt_goes_up_before_the_launch_call(self):
+        self.t.CHUNK = 100
+        ex = make_executor(self.cfg, "mac", "shell", self.t)
+        h = ex.start(Run("GW-4-4", self.cfg.terminals["mac"]["repo"], "z" * 500, script="wc -c"))
+        self.assertGreater(len(self.t.calls), 2)
+        self.assertTrue(wait(ex, h))
+        self.assertEqual(ex.collect(h).last_message, "500")
+
+    @unittest.skipUnless(shutil.which("setsid") and shutil.which("pgrep"), "needs setsid/pgrep")
+    def test_kill_ends_the_whole_process_group(self):
+        ex = make_executor(self.cfg, "mac", "shell", self.t)
+        marker = "agentd-test-%d" % os.getpid()
+        h = ex.start(Run("GW-4-5", self.cfg.terminals["mac"]["repo"], "",
+                         script="bash -c 'sleep 300; : %s' & sleep 300; : %s" % (marker, marker)))
+        end = time.time() + 5
+        while time.time() < end and not subprocess.run(["pgrep", "-f", marker], capture_output=True).stdout:
+            time.sleep(0.1)
+        self.assertTrue(subprocess.run(["pgrep", "-f", marker], capture_output=True).stdout)
+        ex.kill(h)
+        self.assertIn("KILLED=ok left=0", subprocess.run(["bash", "-c", self.t.kill_script(h["run_dir"])],
+                                                         capture_output=True, text=True).stdout)
+        self.assertFalse(subprocess.run(["pgrep", "-f", marker], capture_output=True).stdout)
+        self.assertEqual(ex.poll(h), "running")       # killed runners write no exit.txt; core treats it as cancelled
+
+    def test_kill_after_finish_does_not_signal_a_reused_pid(self):
+        ex = make_executor(self.cfg, "mac", "shell", self.t)
+        h = ex.start(Run("GW-4-6", self.cfg.terminals["mac"]["repo"], "", script="true"))
+        self.assertTrue(wait(ex, h))
+        (Path(h["run_dir"]) / "pid.txt").write_text("1\n")
+        self.assertIn("KILLED=finished", subprocess.run(["bash", "-c", self.t.kill_script(h["run_dir"])],
+                                                        capture_output=True, text=True).stdout)
+
+    def test_darwin_detaches_without_setsid_binary(self):
+        d = remote_data(self.root)
+        d["terminals"]["mac"]["os"] = "darwin"
+        t = PosixWarpgateTransport(Config(d), "mac")
+        s = t.launch_script("/r", "run.sh")
+        self.assertIn("nohup perl -MPOSIX -e 'POSIX::setsid() or die", s)
+        self.assertNotIn("setsid nohup", s)
+        self.assertIn("echo $! > launch.pid", s)
+
+    def test_semantic_checks(self):
+        d = remote_data(self.root)
+        d["terminals"]["mac"]["launch"] = "schtasks"
+        d["terminals"]["mac"]["run_root"] = "~/runs"
+        del d["vault"]["config"]
+        errs = config.validate(d)
+        for frag in ("needs launch=nohup", "mac.run_root: must be an absolute path", "need vault.config"):
+            self.assertTrue(any(frag in e for e in errs), (frag, errs))
+        d = remote_data(self.root)
+        d["terminals"]["win10"]["target"] = "win11"
+        self.assertTrue(any("only reach target 'win10'" in e for e in config.validate(d)))
+
+    def test_upload_payload_is_plain_base64(self):
+        s = self.t.put_script("/r/p.md", base64.b64encode("x'y".encode()).decode())
+        self.assertIn("printf %s 'eCd5' | base64 -d > '/r/p.md.part' && mv -f '/r/p.md.part' '/r/p.md'", s)
+
+
+if __name__ == "__main__":
+    unittest.main()
+
+    def test_probe_dir_name_and_pid_guard(self):
+        ex = make_executor(self.cfg, "mac", "shell", self.t)
+        run = Run("GW-5-1", self.cfg.terminals["mac"]["repo"], "", script="sleep 0.2", dir_name="GW-5-1-r2")
+        self.assertEqual(ex.probe(ex.handle(run)), [])
+        h = ex.start(run)
+        self.assertEqual((os.path.basename(h["run_dir"]), h["task"]), ("GW-5-1-r2", "agentd-GW-5-1-r2"))
+        self.assertTrue(wait(ex, h))
+        self.assertEqual(ex.probe(h), ["exit.txt", "launch.pid", "pid.txt"])
+        n = len(self.t.calls)
+        with self.assertRaises(AssertionError):        # LocalBash: marker LAUNCHED= missing (HASPID)
+            ex.start(run)
+        self.assertIn("HASPID", self.t.calls[n])
+        self.assertTrue(os.path.exists(os.path.join(h["run_dir"], "exit.txt")))   # nothing wiped

tests/test_skips.py

@@ -0,0 +1,71 @@
+import json
+import unittest
+
+from tests.fakes import Harness
+
+
+class FakeProbe:
+    def __init__(self, counts):
+        self.counts = counts
+
+    def process_count(self, name):
+        return self.counts.get(name, 0)
+
+
+class SkipReasonTest(unittest.TestCase):
+    """Why a Todo item was not dispatched: heartbeat "skipped", logged once per new reason."""
+
+    def setUp(self):
+        self.h = Harness()
+        self.p = self.h.plane
+        self.lines = []
+        self.h.d.log = lambda msg: self.lines.append(msg)
+
+    def tearDown(self):
+        self.h.close()
+
+    def skipped(self):
+        hb = json.loads((self.h.d.store.dir / "heartbeat.json").read_text())
+        return {s["item"]: (s["terminal"], s["reason"]) for s in hb["skipped"]}
+
+    def not_dispatched(self):
+        return [l for l in self.lines if "not dispatched" in l]
+
+    def test_reasons_in_heartbeat_and_logged_once(self):
+        self.p.add_issue(1)                                            # starts on win10
+        self.p.add_issue(2)                                            # win10 full
+        self.p.add_issue(3, assignee="kappiee")                        # no terminal for kappiee
+        self.p.add_issue(4, assignee="oregon", labels=["needs-game"])  # starts on oregon, holds the game group
+        self.p.add_issue(5, assignee="oregon", labels=["needs-game"])  # group held
+        self.h.tick()
+        self.assertEqual(self.skipped(), {
+            "GW-2": ("win10", "concurrency 1/1 (GW-1)"),
+            "GW-3": (None, "no route for assignee kappiee"),
+            "GW-5": ("oregon", "serialize_group game held by GW-4"),
+        })
+        self.assertEqual(len(self.not_dispatched()), 3)
+        self.h.tick()   # same reasons: heartbeat still has them, the log does not repeat them
+        self.assertEqual(len(self.skipped()), 3)
+        self.assertEqual(len(self.not_dispatched()), 3)
+
+    def test_busy_process_and_relay_down(self):
+        self.h.cfg.terminals["win10"]["busy_processes"] = ["codex"]
+        probe = FakeProbe({"codex": 1})
+        self.h.d.transport_factory = lambda term: probe
+        self.p.add_issue(1)
+        self.h.tick()
+        self.assertEqual(self.skipped(), {"GW-1": ("win10", "busy_process codex (1 running, not started by agentd)")})
+        probe.counts = {}
+        self.h.tick()   # leftover codex.exe gone: dispatched, nothing skipped
+        self.assertEqual(self.skipped(), {})
+        self.assertEqual(self.p.state_of("i1"), "In Progress")
+        self.p.add_issue(2, assignee="oregon")
+        self.h.cfg.terminals["oregon"]["busy_processes"] = ["claude"]
+        probe.counts = {"claude": None}   # probe fails: relay down
+        probe.process_count = lambda name: None
+        self.h.tick()
+        self.assertEqual(self.skipped()["GW-2"], ("oregon", "relay down"))
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_transports.py

@@ -12,11 +12,11 @@ from agentd.config import Config
 from agentd.executors import Run, make_executor
 from agentd.store import Store
 from agentd.transports.local import LocalTransport
-from agentd.transports.warpgate_ssh import WarpgateSshTransport
+from agentd.transports.warpgate_ssh import PosixWarpgateTransport, WarpgateSshTransport, WindowsWarpgateTransport
 from tests.fakes import config_data
 
 
-class RecordingWin(WarpgateSshTransport):
+class RecordingWin(WindowsWarpgateTransport):
     """Windows transport with the relay replaced by an in-memory recorder."""
     def __init__(self, cfg, terminal):
         super().__init__(cfg, terminal)
@@ -38,7 +38,7 @@ class LocalSecretsTest(unittest.TestCase):
             data = config_data()
             data["agentd"]["state_dir"] = str(Path(tmp) / "state")
             data["terminals"]["oregon"]["repo"] = tmp
-            data["projects"]["GWENT"]["webhook_secret_key"] = "HOOK_GW"
+            data["webhook"] = {"secret_key": "HOOK_GW", "tick_command": ["true"]}
             cfg = Config(data)
             self.assertEqual(cfg.secret_keys(), {"TOK_OR", "TOK_W10", "HOOK_GW"})
             t = LocalTransport(cfg, "oregon")
@@ -95,7 +95,7 @@ class WindowsScriptTest(unittest.TestCase):
 
         def fake_run(args, **kw):
             seen.update(kw["env"])
-            return mock.Mock(returncode=0, stdout="X=ok", stderr="")
+            return mock.Mock(returncode=0, stdout=b"X=ok", stderr=b"")
         with mock.patch.dict(os.environ, {"PLANE_WINDOW10_API_TOKEN": "s"}), \
                 mock.patch("agentd.transports.warpgate_ssh.subprocess.run", fake_run):
             WarpgateSshTransport(self.cfg, "win10").ps("'X=ok'")
@@ -120,3 +120,30 @@ class StoreTest(unittest.TestCase):
 
 if __name__ == "__main__":
     unittest.main()
+
+
+class GuardTest(unittest.TestCase):
+    def test_local_clean_refuses_dir_with_pid_file(self):
+        from agentd.transports import TransportError
+        with tempfile.TemporaryDirectory() as tmp:
+            t = LocalTransport(Config(config_data()), "oregon")
+            (Path(tmp) / "prompt.md").write_text("x")
+            t.mkdir(tmp, clean=True)
+            self.assertEqual(t.probe(tmp, ("exit.txt", "pid.txt")), [])
+            (Path(tmp) / "pid.txt").write_text("1")
+            with self.assertRaises(TransportError):
+                t.mkdir(tmp, clean=True)
+            self.assertEqual(t.probe(tmp, ("exit.txt", "pid.txt")), ["pid.txt"])
+
+    def test_windows_probe_and_clean_guard(self):
+        from agentd.transports import TransportError
+        cfg = config.load(ROOT / "config" / "agentd.example.yaml")
+        t = WindowsWarpgateTransport(cfg, "win10")
+        scripts = []
+        t.ps = lambda s, timeout=150: (scripts.append(s), "PROBE=launch.pid,pid.txt\n")[1]
+        self.assertEqual(t.probe("C:\\r\\GW-1-1", ("exit.txt", "launch.pid", "pid.txt")), ["launch.pid", "pid.txt"])
+        self.assertIn("Join-Path $d $_", scripts[-1])
+        t.ps = lambda s, timeout=150: (scripts.append(s), "HASPID\n")[1]
+        with self.assertRaises(TransportError):
+            t.mkdir("C:\\r\\GW-1-1", clean=True)
+        self.assertIn("'pid.txt'", scripts[-1])

tests/test_webhook.py

@@ -0,0 +1,184 @@
+import json
+import threading
+import time
+import unittest
+import urllib.error
+import urllib.request
+from http.server import ThreadingHTTPServer
+
+from agentd.webhook import BUSY, Trigger, make_handler, route, sign, verify
+
+SECRET = "plane_wh_test"
+PROJECTS = {"GWENT": {"id": "proj-gw"}, "PAD": {"id": "proj-pad"}}
+
+
+def payload(project="proj-gw", event="issue", action="updated"):
+    return {"event": event, "action": action, "webhook_id": "w1", "workspace_id": "ws", "workspace_slug": "kappiee",
+            "data": {"id": "i1", "project": project}, "activity": {"field": "state"}}
+
+
+def plane_body(p):
+    return json.dumps(p).encode()   # what Plane signs (json.dumps(payload)) and sends
+
+
+class SignatureTest(unittest.TestCase):
+    def test_ok_bad_missing(self):
+        body = plane_body(payload())
+        self.assertTrue(verify(SECRET, body, sign(SECRET, body)))
+        self.assertTrue(verify(SECRET, body, sign(SECRET, body).upper()))
+        self.assertFalse(verify(SECRET, body, sign("other", body)))
+        self.assertFalse(verify(SECRET, body + b" ", sign(SECRET, body + b"x")))
+        self.assertFalse(verify(SECRET, body, None))
+        self.assertFalse(verify(SECRET, body, ""))
+        self.assertFalse(verify("", body, sign("", body)))       # no secret configured: reject everything
+        self.assertFalse(verify(SECRET, body, "zzé"))         # non-hex garbage
+
+    def test_reencoded_body_matches_plane_serialisation(self):
+        p = payload()
+        compact = json.dumps(p, separators=(",", ":")).encode()
+        self.assertTrue(verify(SECRET, compact, sign(SECRET, plane_body(p))))
+        self.assertFalse(verify(SECRET, b"not json", sign(SECRET, plane_body(p))))
+
+
+class RouteTest(unittest.TestCase):
+    def test_route(self):
+        self.assertEqual(route(payload(), PROJECTS), "GWENT")
+        self.assertEqual(route(payload("proj-pad", "issue_comment", "created"), PROJECTS), "PAD")
+        self.assertIsNone(route(payload("proj-other"), PROJECTS))
+        self.assertIsNone(route({"event": "issue", "action": "deleted", "data": {"id": "i1"}}, PROJECTS))
+        self.assertIsNone(route({"event": "issue"}, PROJECTS))
+        self.assertEqual(route({"data": {"project": {"id": "proj-gw"}}}, PROJECTS), "GWENT")
+
+
+class FakeTrigger:
+    def __init__(self):
+        self.kicks = []
+
+    def kick(self, reason):
+        self.kicks.append(reason)
+
+    def status(self):
+        return {"pending": bool(self.kicks)}
+
+
+class HandlerTest(unittest.TestCase):
+    def setUp(self):
+        self.trigger, self.lines = FakeTrigger(), []
+        self.httpd = ThreadingHTTPServer(("127.0.0.1", 0), make_handler(SECRET, PROJECTS, self.trigger, self.lines.append))
+        threading.Thread(target=self.httpd.serve_forever, daemon=True).start()
+        self.url = "http://127.0.0.1:%d" % self.httpd.server_address[1]
+
+    def tearDown(self):
+        self.httpd.shutdown()
+        self.httpd.server_close()
+
+    def post(self, body, sig, path="/plane"):
+        headers = {"Content-Type": "application/json", "X-Plane-Delivery": "d-1", "X-Plane-Event": "issue"}
+        if sig is not None:
+            headers["X-Plane-Signature"] = sig
+        req = urllib.request.Request(self.url + path, data=body, headers=headers, method="POST")
+        try:
+            with urllib.request.urlopen(req, timeout=5) as r:
+                return r.status
+        except urllib.error.HTTPError as e:
+            return e.code
+
+    def test_signed_tenant_delivery_triggers(self):
+        body = plane_body(payload())
+        self.assertEqual(self.post(body, sign(SECRET, body)), 200)
+        self.assertEqual(self.trigger.kicks, ["d-1"])
+        self.assertIn("project=GWENT", self.lines[-1])
+        self.assertIn("decision=tick", self.lines[-1])
+
+    def test_bad_or_missing_signature_is_401_and_triggers_nothing(self):
+        body = plane_body(payload())
+        self.assertEqual(self.post(body, sign("wrong", body)), 401)
+        self.assertEqual(self.post(body, None), 401)
+        self.assertEqual(self.trigger.kicks, [])
+        self.assertIn("401 missing", self.lines[-1])
+
+    def test_other_project_is_202(self):
+        body = plane_body(payload("proj-other"))
+        self.assertEqual(self.post(body, sign(SECRET, body)), 202)
+        self.assertEqual(self.trigger.kicks, [])
+
+    def test_wrong_path_and_healthz(self):
+        body = plane_body(payload())
+        self.assertEqual(self.post(body, sign(SECRET, body), "/other"), 404)
+        with urllib.request.urlopen(self.url + "/healthz", timeout=5) as r:
+            self.assertEqual((r.status, json.loads(r.read())["ok"]), (200, True))
+
+
+class FakeRunner:
+    def __init__(self, codes=(), hold=0.0):
+        self.codes, self.hold, self.calls = list(codes), hold, 0
+        self.started = threading.Event()
+
+    def __call__(self):
+        self.calls += 1
+        self.started.set()
+        time.sleep(self.hold)
+        return self.codes.pop(0) if self.codes else 0
+
+
+class TriggerTest(unittest.TestCase):
+    def make(self, runner, **kw):
+        kw = dict(dict(debounce=0.2, retry=0.1, retry_window=1.0), **kw)
+        self.lines = []
+        return Trigger(runner, log=self.lines.append, **kw)
+
+    def test_burst_coalesces_into_one_tick(self):
+        run = FakeRunner()
+        t = self.make(run)
+        for i in range(5):
+            t.kick("d%d" % i)
+        self.assertTrue(t.step(wait=2))
+        self.assertEqual(run.calls, 1)
+        self.assertIn("5 deliveries", self.lines[0])
+        self.assertFalse(t.step(wait=0.4))   # nothing left
+        self.assertEqual(run.calls, 1)
+
+    def test_debounce_waits_before_ticking(self):
+        run = FakeRunner()
+        t = self.make(run, debounce=0.3)
+        t0 = time.monotonic()
+        t.kick("d1")
+        t.step(wait=2)
+        self.assertGreaterEqual(time.monotonic() - t0, 0.29)
+
+    def test_delivery_during_a_tick_causes_one_follow_up(self):
+        run = FakeRunner(hold=0.3)
+        t = self.make(run, debounce=0.05).start()
+        t.kick("d1")
+        self.assertTrue(run.started.wait(2))
+        t.kick("d2")      # arrives while the first tick runs
+        t.kick("d3")
+        time.sleep(1.0)
+        self.assertEqual(run.calls, 2)
+        self.assertEqual(t.status()["pending"], False)
+
+    def test_busy_lock_is_retried_until_the_tick_runs(self):
+        run = FakeRunner(codes=[BUSY, BUSY, 0])
+        t = self.make(run, debounce=0.0)
+        t.kick("d1")
+        for _ in range(3):
+            self.assertTrue(t.step(wait=2))
+        self.assertEqual(run.calls, 3)
+        self.assertEqual(sum("retry in" in l for l in self.lines), 2)
+        self.assertIn("1 deliveries: d1", self.lines[-2])   # the delivery is carried into the retry
+        self.assertIsNone(t.status()["retrying_since"])
+        self.assertFalse(t.step(wait=0.3))
+
+    def test_busy_gives_up_after_the_window(self):
+        run = FakeRunner(codes=[BUSY] * 10)
+        t = self.make(run, debounce=0.0, retry=0.1, retry_window=0.35)
+        t.kick("d1")
+        while t.step(wait=1):
+            pass
+        self.assertTrue(any("giving up" in l for l in self.lines))
+        self.assertLessEqual(run.calls, 5)
+        self.assertFalse(t.status()["pending"])
+
+
+if __name__ == "__main__":
+    unittest.main()

tools/plane-webhook.py

@@ -0,0 +1,121 @@
+#!/usr/bin/env python3
+"""Plane workspace webhooks, as the workspace admin kappiee (session login; the internal /api/workspaces/<slug>/webhooks/
+endpoint needs admin). Never prints secrets: a new or regenerated secret goes straight into the Vault as
+PLANE_WEBHOOK_SECRET (then restart plane-agentd-webhook). Run under
+  agent-access.py --config <identity> exec --secret-env PLANE_KAPPIEE_PASSWORD=PLANE_KAPPIEE_PASSWORD -- python3 -I tools/plane-webhook.py ...
+usage: list | create URL (inactive; work item + comment events) | patch ID key=value... (e.g. is_active=true)
+       | regenerate ID | delete ID | logs ID [n]"""
+import http.cookiejar, json, os, ssl, sys, urllib.parse, urllib.request
+
+BASE = "https://plane.kappiee.com"
+WS = "kappiee"
+jar = http.cookiejar.CookieJar()
+op = urllib.request.build_opener(urllib.request.HTTPCookieProcessor(jar))
+HDR = {"Referer": BASE + "/", "Origin": BASE, "User-Agent": "plane-agentd-admin"}
+
+
+def csrf():
+    return next((c.value for c in jar if c.name == "csrftoken"), "")
+
+
+def call(method, path, body=None, form=None):
+    h = dict(HDR)
+    data = None
+    if form is not None:
+        data = urllib.parse.urlencode(form).encode()
+        h["Content-Type"] = "application/x-www-form-urlencoded"
+    elif body is not None:
+        data = json.dumps(body).encode()
+        h["Content-Type"] = "application/json"
+    if method != "GET":
+        h["X-CSRFToken"] = csrf()
+    req = urllib.request.Request(BASE + path, data=data, headers=h, method=method)
+    try:
+        with op.open(req, timeout=30) as r:
+            raw = r.read()
+            return r.status, raw, r.geturl()
+    except urllib.error.HTTPError as e:
+        return e.code, e.read(), path
+
+
+def login():
+    st, raw, _ = call("GET", "/auth/get-csrf-token/")
+    tok = json.loads(raw).get("csrf_token") or csrf()
+    st, raw, url = call("POST", "/auth/sign-in/", form={"csrfmiddlewaretoken": tok, "email": "kappiee@kappiee.com",
+                                                        "password": os.environ["PLANE_KAPPIEE_PASSWORD"]})
+    ok = any(c.name.startswith("session") for c in jar)
+    print("login", st, "session cookie" if ok else "NO SESSION", urllib.parse.urlparse(url).path)
+    if not ok:
+        sys.exit(1)
+
+
+def redact(d):
+    return {k: ("<%d chars>" % len(v) if k == "secret_key" and v else v) for k, v in d.items()}
+
+
+def vault_store(name, value):
+    ident = json.load(urllib.request.urlopen("https://nas.tail9531b.ts.net:8443/identity/agent.json", timeout=30))
+    v = ident["vault"]
+    ctx = ssl.create_default_context()
+    ctx.check_hostname = False
+    ctx.verify_mode = ssl.CERT_NONE
+    req = urllib.request.Request(v["url"].rstrip("/") + "/v1/credentials", method="POST",
+                                 data=json.dumps({"vault": "default", "credentials": {name: value}}).encode(),
+                                 headers={"Authorization": "Bearer " + v["token"], "X-Vault": "default",
+                                          "Content-Type": "application/json"})
+    with urllib.request.urlopen(req, context=ctx, timeout=30) as r:
+        print("vault store", name, r.status)
+
+
+def store_secret(w):
+    if w.get("secret_key"):
+        vault_store("PLANE_WEBHOOK_SECRET", w["secret_key"])
+
+
+login()
+P = "/api/workspaces/%s/webhooks/" % WS
+cmd = sys.argv[1]
+if cmd == "list":
+    st, raw, _ = call("GET", P)
+    print("list", st)
+    if st == 200:
+        for w in json.loads(raw):
+            print(json.dumps(redact(w)))
+    else:
+        print(raw[:300])
+elif cmd == "create":
+    body = {"url": sys.argv[2], "is_active": False, "project": False, "issue": True, "module": False, "cycle": False,
+            "issue_comment": True}
+    st, raw, _ = call("POST", P, body)
+    print("create", st)
+    if st in (200, 201):
+        w = json.loads(raw)
+        store_secret(w)
+        print(json.dumps(redact(w)))
+    else:
+        print(raw[:500])
+elif cmd == "patch":
+    body = {}
+    for kv in sys.argv[3:]:
+        k, v = kv.split("=", 1)
+        body[k] = {"true": True, "false": False}.get(v, v)
+    st, raw, _ = call("PATCH", P + sys.argv[2] + "/", body)
+    print("patch", st, json.dumps(redact(json.loads(raw))) if st == 200 else raw[:500])
+elif cmd == "regenerate":
+    st, raw, _ = call("POST", P + sys.argv[2] + "/regenerate/", {})
+    print("regenerate", st)
+    if st in (200, 201):
+        w = json.loads(raw)
+        store_secret(w)
+        print(json.dumps(redact(w)))
+    else:
+        print(raw[:500])
+elif cmd == "delete":
+    st, raw, _ = call("DELETE", P + sys.argv[2] + "/")
+    print("delete", st, raw[:200])
+elif cmd == "logs":
+    st, raw, _ = call("GET", "/api/workspaces/%s/webhook-logs/%s/" % (WS, sys.argv[2]))
+    print("logs", st)
+    if st == 200:
+        for l in json.loads(raw)[: int(sys.argv[3]) if len(sys.argv) > 3 else 10]:
+            print(l.get("created_at"), l.get("event_type"), l.get("request_method"), l.get("response_status"), (l.get("response_body") or "")[:120], l.get("retry_count"))

tools/win10/win10-ps.sh

@@ -0,0 +1,18 @@
+#!/usr/bin/env bash
+# Run a PowerShell script (passed as $1 text, or a file with -f) on win10 through server-access,
+# encoded as -EncodedCommand so quoting never crosses the relay. Output is UTF-8; progress noise off.
+# AGENT_ACCESS_HOME = directory of the vault client (default ~/.local/share/server-access-client);
+# AGENT_ACCESS_CONFIG = identity URL/file.
+set -euo pipefail
+CLIENT="${AGENT_ACCESS_HOME:-$HOME/.local/share/server-access-client}/agent-access.py"
+CFG="${AGENT_ACCESS_CONFIG:-https://nas.tail9531b.ts.net:8443/identity/agent.json}"
+TIMEOUT="${WIN10_TIMEOUT:-120}"
+if [ "${1:-}" = "-f" ]; then SCRIPT="$(cat "$2")"; else SCRIPT="$1"; fi
+SCRIPT="[Console]::OutputEncoding=[Text.Encoding]::UTF8
+\$ErrorActionPreference='Continue'
+\$ProgressPreference='SilentlyContinue'
+$SCRIPT"
+B64="$(printf '%s' "$SCRIPT" | iconv -f UTF-8 -t UTF-16LE | base64 -w0)"
+if [ "${#B64}" -gt 7600 ]; then echo "encoded command too long (${#B64} > 7600)" >&2; exit 2; fi
+exec python3 -I "$CLIENT" --config "$CFG" run --target win10 --timeout "$TIMEOUT" -- \
+  powershell.exe -NoProfile -NonInteractive -EncodedCommand "$B64" 2> >(grep -v -E '^#< CLIXML|^<Objs|正在准备首次使用模块' >&2)

tools/win10/win10-put.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+# Upload a local text/binary file to win10: gzip+base64, appended in chunks through the relay,
+# then decoded remotely and verified by SHA-256.
+# usage: win10-put.sh <local-file> <remote-windows-path>
+set -euo pipefail
+HERE="$(cd "$(dirname "$0")" && pwd)"
+LOCAL="$1"; REMOTE="$2"
+TMP="$REMOTE.b64"
+SHA="$(sha256sum "$LOCAL" | cut -d' ' -f1)"
+B64="$(gzip -9c "$LOCAL" | base64 -w0)"
+CHUNK=2400
+"$HERE/win10-ps.sh" "New-Item -ItemType Directory -Force -Path (Split-Path -Parent '$REMOTE') | Out-Null; Set-Content -Path '$TMP' -Value '' -NoNewline -Encoding Ascii; 'reset'" >/dev/null
+TOTAL=$(( (${#B64} + CHUNK - 1) / CHUNK ))
+for ((i=0; i<${#B64}; i+=CHUNK)); do
+  PART="${B64:i:CHUNK}"
+  n=$(( i / CHUNK + 1 ))
+  "$HERE/win10-ps.sh" "Add-Content -Path '$TMP' -Value '$PART' -NoNewline -Encoding Ascii; 'chunk $n/$TOTAL'" | tail -1
+done
+"$HERE/win10-ps.sh" "\$b=[Convert]::FromBase64String((Get-Content -Raw -Encoding Ascii '$TMP'))
+\$ms=New-Object IO.MemoryStream(,\$b)
+\$gz=New-Object IO.Compression.GZipStream(\$ms,[IO.Compression.CompressionMode]::Decompress)
+\$out=New-Object IO.MemoryStream; \$gz.CopyTo(\$out)
+[IO.File]::WriteAllBytes('$REMOTE',\$out.ToArray())
+Remove-Item '$TMP'
+'remote sha256 ' + (Get-FileHash '$REMOTE' -Algorithm SHA256).Hash.ToLower()
+'expected      $SHA'"