Merge pull request 'Rename a resumed session's husk with an 'X ' prefix' (#2) from rename-husk into main

README.md

@@ -62,6 +62,12 @@ runs `claude stop <old id>` after a successful resume. Without that the husk
 would be picked up again on a later tick and resumed into a *second* session
 working the same conversation — two agents racing on the same git worktree.
 
+A stopped husk is still listed, under the same name as the successor that
+replaced it, which is confusing to work next to. So the watchdog also prefixes
+it — `X rewrite larvitsmpp typescript esm` — by rewriting `name` in the husk's
+`state.json`, which is where `claude agents` reads it and which no command sets.
+`--superseded-prefix` changes it.
+
 ### Host headroom
 
 A resume starts a fresh process that reloads a whole conversation, so before
@@ -190,6 +196,7 @@ hardcoded. To make one permanent, add it to the crontab line.
 | `--max-swap-used-pct` | `WATCHDOG_MAX_SWAP_USED_PCT` | `50`         |
 | `--min-available-mb`  | `WATCHDOG_MIN_AVAILABLE_MB`  | `2048`       |
 | `--prompt`            | `WATCHDOG_PROMPT`            | see `--help` |
+| `--superseded-prefix` | `WATCHDOG_SUPERSEDED_PREFIX` | `X `         |
 | `--dry-run`           | —                            | off          |
 
 ## Files

resume-stalled-sessions.py

@@ -29,6 +29,8 @@ DEFAULT_PROMPT = (
     "transient failure (usage limit, API outage, or network), which has cleared."
 )
 
+DEFAULT_PREFIX = "X "
+
 # Every quota Claude Code can report as "You've hit your <limit> · <how to clear
 # it>", including the ones that cost money rather than time.
 QUOTA_BLOCK = re.compile(
@@ -287,11 +289,42 @@ def candidates(sessions, jobs_dir, now):
             "blocked_at": blocked_at,
             "name": session.get("name"),
             "session_id": session_id,
+            "state_file": state_file,
         })
     return sorted(found, key=lambda job: job["blocked_at"])
 
 
-def resume(job, claude_bin, prompt, dry_run):
+def mark_superseded(state_file, prefix):
+    """Prefix the husk's name so it reads apart from its successor, or say why not.
+
+    `claude agents` takes the name from state.json and offers no command to set
+    it, so the file is the only seam.
+    """
+    try:
+        state = json.loads(state_file.read_text())
+    except (OSError, json.JSONDecodeError) as err:
+        return str(err)
+    name = state.get("name")
+    if not isinstance(name, str) or name.startswith(prefix):
+        return None
+
+    state["name"] = prefix + name
+    flags = state.get("respawnFlags")
+    if isinstance(flags, list) and "--name" in flags:
+        at = flags.index("--name") + 1
+        if at < len(flags):
+            flags[at] = state["name"]
+
+    staged = state_file.with_name(state_file.name + ".new")
+    try:
+        staged.write_text(json.dumps(state, indent=2))
+        staged.replace(state_file)
+    except OSError as err:
+        return str(err)
+    return None
+
+
+def resume(job, claude_bin, prompt, prefix, dry_run):
     cwd = job["cwd"]
     if not cwd or not Path(cwd).is_dir():
         log(f"{job['id']}: skipped, working directory is gone ({cwd})")
@@ -326,6 +359,10 @@ def resume(job, claude_bin, prompt, dry_run):
         )
     except (OSError, subprocess.SubprocessError) as err:
         log(f"{job['id']}: could not retire stale entry: {err}")
+
+    problem = mark_superseded(job["state_file"], prefix)
+    if problem:
+        log(f"{job['id']}: could not rename stale entry: {problem}")
     return True
 
 
@@ -367,6 +404,12 @@ def main():
         help="memory a resume needs the host to have spare, in MB (default: 2048)",
     )
     parser.add_argument("--prompt", default=os.environ.get("WATCHDOG_PROMPT", DEFAULT_PROMPT))
+    parser.add_argument(
+        "--superseded-prefix",
+        default=os.environ.get("WATCHDOG_SUPERSEDED_PREFIX", DEFAULT_PREFIX),
+        help="prepended to the name of a husk once its successor runs "
+             f"(default: {DEFAULT_PREFIX!r})",
+    )
     args = parser.parse_args()
 
     claude_bin = resolve_claude(args.claude_bin)
@@ -417,7 +460,7 @@ def main():
         if strain:
             log(f"holding off: {strain}")
             break
-        if not resume(job, claude_bin, args.prompt, args.dry_run):
+        if not resume(job, claude_bin, args.prompt, args.superseded_prefix, args.dry_run):
             continue
         ledger["attempts"][job["session_id"]] = tried + 1
         ledger["resumes"] = recent + [now.timestamp()]

test-resume-stalled-sessions.py

@@ -172,11 +172,50 @@ class Candidates(unittest.TestCase):
         session = self.blocked_job(self.SPEND, timedelta(hours=2))
         found = watchdog.candidates([session], self.jobs, self.now)
         self.assertEqual([job["id"] for job in found], [session["id"]])
+        self.assertEqual(found[0]["state_file"], self.jobs / session["id"] / "state.json")
 
     def test_fresh_spend_block_waits(self):
         session = self.blocked_job(self.SPEND, timedelta(minutes=10))
         self.assertEqual(watchdog.candidates([session], self.jobs, self.now), [])
 
 
+class MarkSuperseded(unittest.TestCase):
+    """The husk keeps its conversation, so its name must not read as the live one."""
+
+    def setUp(self):
+        temp = tempfile.TemporaryDirectory()
+        self.addCleanup(temp.cleanup)
+        self.state = Path(temp.name) / "state.json"
+
+    def write(self, **fields):
+        self.state.write_text(json.dumps(fields))
+
+    def read(self):
+        return json.loads(self.state.read_text())
+
+    def test_name_gains_the_prefix(self):
+        self.write(name="rewrite larvitsmpp typescript esm", state="blocked")
+        self.assertIsNone(watchdog.mark_superseded(self.state, "X "))
+        self.assertEqual(self.read()["name"], "X rewrite larvitsmpp typescript esm")
+
+    def test_the_respawn_name_stays_in_step(self):
+        self.write(name="fejkdata", respawnFlags=["--agent", "claude", "--name", "fejkdata"])
+        watchdog.mark_superseded(self.state, "X ")
+        self.assertEqual(self.read()["respawnFlags"], ["--agent", "claude", "--name", "X fejkdata"])
+
+    def test_a_husk_is_prefixed_only_once(self):
+        self.write(name="X fejkdata")
+        self.assertIsNone(watchdog.mark_superseded(self.state, "X "))
+        self.assertEqual(self.read()["name"], "X fejkdata")
+
+    def test_an_unnamed_job_is_left_alone(self):
+        self.write(state="blocked")
+        self.assertIsNone(watchdog.mark_superseded(self.state, "X "))
+        self.assertNotIn("name", self.read())
+
+    def test_unwritable_state_is_reported_rather_than_raised(self):
+        self.assertIsNotNone(watchdog.mark_superseded(self.state.with_name("gone.json"), "X "))
+
+
 if __name__ == "__main__":
     unittest.main()