fix(agents): a claude purge keeps ownership evidence until last (Issue #31, review)

A session file is what ties a pid's key file to the session, and a job's
state.json what ties the job dir: removing them first left the rest
unattributable after a partial failure (residual: none, and a retry could
not find it). Keys now go before their session file, which stays while any
key does; a job dir is emptied with state.json last and kept if anything
stays. residual lists a session file's key files (masked). A startedAt of 0
or less cannot tell a reused pid and counts as live; the sessions/ listdir
no longer raises.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

.agents/skills/lib_py/agents/adapters/claude.py

@@ -70,6 +70,8 @@ def _session_file_live(data: dict, pid) -> bool:
         started = float(data["startedAt"]) / 1000.0
     except Exception:
         return True
+    if started <= 0:
+        return True
     # startedAt is written 1.7-3.2 s after the process starts; 2 s covers
     # lstart's one-second resolution.
     return lstart <= started + 2
@@ -85,6 +87,41 @@ def _remove(path: str) -> None:
         pass
 
 
+def _key_files(session_file: str) -> list:
+    """The key files of the pid a sessions/<pid>.json is named after,
+    matched by name only; their content is never read."""
+    d, name = os.path.split(session_file)
+    stem = name[:-len(".json")]
+    if not stem.isdigit():
+        return []
+    try:
+        names = sorted(os.listdir(d))
+    except OSError:
+        return []
+    return [os.path.join(d, n) for n in names if re.fullmatch(rf"{stem}\.[0-9a-f]{{64}}\.key", n)]
+
+
+def _remove_job(job: str) -> None:
+    """state.json goes last: it is the only proof that the dir is this
+    session's, so a dir that could not be emptied keeps it."""
+    if os.path.islink(job):
+        _remove(job)
+        return
+    try:
+        names = os.listdir(job)
+    except OSError:
+        return
+    for n in names:
+        if n != "state.json":
+            _remove(os.path.join(job, n))
+    try:
+        left = os.listdir(job)
+    except OSError:
+        return
+    if left in ([], ["state.json"]):
+        _remove(job)
+
+
 def _label(path: str) -> str:
     """How a path is printed: a session key file is named by pid only, its
     hex part is not shown."""
@@ -235,15 +272,10 @@ class ClaudeAgentAdapter(BaseAgentAdapter):
         if job:
             paths.append(job)
         for path, pid, data in self._session_files(uuid, ctx):
-            if _session_file_live(data, pid):
-                continue
-            paths.append(path)
-            stem = os.path.basename(path)[:-len(".json")]
-            sessions = os.path.dirname(path)
-            if stem.isdigit():
-                # The key file is matched by name only; its content is never read.
-                paths.extend(os.path.join(sessions, n) for n in sorted(os.listdir(sessions))
-                             if re.fullmatch(rf"{stem}\.[0-9a-f]{{64}}\.key", n))
+            if not _session_file_live(data, pid):
+                # Keys before the session file: the file is what ties them to
+                # this session, so it must outlive them.
+                paths += _key_files(path) + [path]
         for root in _tmp_roots():
             try:
                 keys = sorted(os.listdir(root))
@@ -265,8 +297,16 @@ class ClaudeAgentAdapter(BaseAgentAdapter):
         # Best effort: never raises (stop runs this inside its registry
         # transaction). What could not be removed is not listed here and
         # shows up in residual_artifacts instead.
+        # Ownership evidence (a job's state.json, a session file) is removed
+        # only after what it vouches for, so a retry can still find the rest.
         purged = []
+        job = self._owned_job(uuid, ctx)
         for path in self._owned_paths(uuid, ctx):
+            if path == job:
+                _remove_job(path)
+            elif path.endswith(".json") and any(os.path.lexists(k) for k in _key_files(path)):
+                continue
+            else:
                 _remove(path)
             if not os.path.lexists(path) and _label(path) not in purged:
                 purged.append(_label(path))
@@ -276,7 +316,8 @@ class ClaudeAgentAdapter(BaseAgentAdapter):
         """What still carries this session: names holding the uuid under the
         config root, in projects (two levels: key, then entry) and in the tmp
         roots (two levels), plus the content-matched jobs dir and session
-        files. projects is not walked deeper; it can be very large."""
+        files with their (masked) key files. projects is not walked deeper;
+        it can be very large."""
         if not _plausible_id(uuid):
             return []
         cfg = _config_dir(ctx)
@@ -299,7 +340,9 @@ class ClaudeAgentAdapter(BaseAgentAdapter):
         job = self._owned_job(uuid, ctx)
         if job:
             found.add(job)
-        found.update(path for path, _, _ in self._session_files(uuid, ctx))
+        for path, _, _ in self._session_files(uuid, ctx):
+            found.add(path)
+            found.update(_label(k) for k in _key_files(path))
         for root in _tmp_roots():
             try:
                 tkeys = os.listdir(root)