Double-start guard: launching marker, probe-and-adopt, fresh-dir relaunch; skip unreachable terminals per tick

Before the launch call the planned run (run id, terminal, handle, bookkeeping) is saved as a launching marker. If the launch call fails or agentd dies before recording the run, the next tick probes the run directory (exit.txt / launch.pid / pid.txt): evidence -> adopt the run; nothing -> relaunch with the same run id in <run_id>-r2 (-r3 ...); probe unreachable -> keep the marker, which holds the terminal slot. Transports refuse to clean a run directory holding a pid file. A terminal whose relay fails is skipped for the rest of the tick (heartbeat.down) while other terminals carry on.
2026-10-21 11:54:40 +00:00
parent dbe305103c
commit 4f45d368ad
8 changed files with 357 additions and 50 deletions
@@ -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 TransportError, make_transport

 TAGS = ("review", "question", "blocked", "progress")
 WORKER = "[worker]"
@@ -143,7 +149,7 @@ 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
        self.plane_factory = plane_factory or (lambda term: Plane(
@@ -154,6 +160,12 @@ class Agentd:

    def _reset_caches(self):
        self._clients, self._transports, self._executors, self._ready = {}, {}, {}, {}
+        self._down = set()   # terminals whose relay failed during this tick
+
+    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):
@@ -195,9 +207,11 @@ class Agentd:
                    self.log("%s tick error: %r" % (name, exc))
                finally:
                    self.store.save(data)
-            self.store.heartbeat({"errors": errors, "inflight": [
+            self.store.heartbeat({"errors": errors, "down": sorted(self._down), "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 loop(self, interval):
@@ -216,14 +230,43 @@ 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"]
@@ -404,7 +447,7 @@ class Agentd:
        load, groups = Counter(), set()
        for p in data["projects"].values():
            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
                    if fl.get("group"):
@@ -422,7 +465,7 @@ class Agentd:
            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)
@@ -438,15 +481,21 @@ class Agentd:
                    self.log("%s not dispatched: %s" % (ctx.key(issue.get("sequence_id")), route))
                continue
            term, group = route["terminal"], route["group"]
+            if term in self._down:
+                continue
            if load[term] >= self.cfg.terminals[term].get("concurrency", 1) or (group and group in groups):
                continue
            if not self.terminal_ready(term, load[term]):
                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
+                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
            if group:
@@ -454,14 +503,17 @@ class Agentd:

    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))
                    ready = False
@@ -522,7 +574,7 @@ 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))
@@ -552,25 +604,11 @@ 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:
@@ -579,10 +617,53 @@ 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))
@@ -17,6 +17,7 @@ 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 +36,15 @@ 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

@@ -84,11 +91,15 @@ 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)
+        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)
-        task = "agentd-" + run.run_id
        files = {"prompt.md": run.prompt, script_name: script}
        if hasattr(t, "start_run"):  # remote POSIX: stage + launch in one relay round trip
            t.start_run(task, run_dir, files, script_name)
@@ -97,7 +108,13 @@ class RunnerExecutor(Executor):
            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 {"run_dir": run_dir, "task": task, "thread_id": run.thread_id}
+        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"))
@@ -4,6 +4,10 @@ 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 +%Y-%m-%dT%H:%M:%S%z) (cd failed)" > {d}/exit.txt; exit 97; }}
@@ -40,6 +44,8 @@ class LocalTransport:
        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()
@@ -51,6 +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)
@@ -113,9 +113,21 @@ class WindowsWarpgateTransport:
    def mkdir(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'")
+        out = self.ps(script + "'MKDIR=ok'")
+        if "MKDIR=ok" not in out:
+            raise TransportError("mkdir %s failed: %s" % (path, out.strip()[-300:]))
+
+    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_text(self, path, text):
        fd, tmp = tempfile.mkstemp(prefix="agentd-", suffix=".txt")
@@ -230,7 +242,9 @@ class PosixWarpgateTransport:
    def mkdir_script(self, path, clean=False):
        q = self.quote(path)
        s = "mkdir -p %s || exit 1\n" % q
-        if clean:
+        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

@@ -275,6 +289,12 @@ class PosixWarpgateTransport:
            raise TransportError("read %s: unexpected relay output" % path)
        return _b64_payload(out, "B64")

+    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")
@@ -6,6 +6,7 @@ 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"}
@@ -89,15 +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):
@@ -1,3 +1,4 @@
+import json
 import unittest

 from tests.fakes import Harness, config_data
@@ -337,10 +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):
@@ -391,3 +395,113 @@ class TierDefaultTest(unittest.TestCase):

 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()
@@ -203,3 +203,17 @@ class PosixRemoteTest(unittest.TestCase):

 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
@@ -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])