alissa-tools-github-revloop 0.16.15__tar.gz → 0.16.16__tar.gz

{alissa_tools_github_revloop-0.16.15/src/main/alissa_tools_github_revloop.egg-info → alissa_tools_github_revloop-0.16.16}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: alissa-tools-github-revloop
-Version: 0.16.15
+Version: 0.16.16
 Summary: ALISSA-TOOLS-GITHUB-REVLOOP
 Home-page: https://alissa.app
 Author: Fahera

alissa_tools_github_revloop-0.16.16/src/main/alissa/tools/github/revloop/version

@@ -0,0 +1 @@
+0.16.16

{alissa_tools_github_revloop-0.16.15 → alissa_tools_github_revloop-0.16.16}/src/main/alissa/tools/github/revloop/webui/__init__.py

@@ -15,14 +15,16 @@ pass writes one `poll_snapshots` row (UI-1, PR #35) carrying the pass timing,
 the candidate count, the decision-summary counts, and the compact per-item
 stage list. The sidecar reads that table through `State.read_snapshots`, plus
 the spawn ledger, the escalation table and the ping ledger (the operator
-inbox), all read-only. Its only live signals are local (`alissa tmux ls`, a
-`/proc` walk of each session's pane-PID tree) or cached (`gh api rate_limit`,
-60s; the PyPI version JSON, 10m) -- so a fleet of operators refreshing the
-dashboard never moves the daemon's rate budget.
+inbox), all read-only. Its only live signals are local (`alissa tmux ls`, one
+`/proc` walk serving both each session's pane-PID tree and the host-wide
+top-by-RSS list, and the container's own cgroup v2 memory charge) or cached
+(`gh api rate_limit`, 60s; the PyPI version JSON, 10m) -- so a fleet of
+operators refreshing the dashboard never moves the daemon's rate budget.
 
 Layout:
   auth.py     -- fail-closed passcode, HMAC-signed sessions, CSRF, login throttle
-  sysinfo.py  -- /proc process-tree CPU%/RSS (sample-free, vanished-PID tolerant)
+  sysinfo.py  -- /proc process-tree CPU%/RSS + host-wide top-by-RSS and the
+                 cgroup memory split (sample-free, vanished-PID tolerant)
   sources.py  -- the read-only data layer + the retry-now UPDATE, cached checks
   page.py     -- the single static HTML page (studio design system, both themes)
   server.py   -- ThreadingHTTPServer wiring, routing, auth/CSRF gating, actions

{alissa_tools_github_revloop-0.16.15 → alissa_tools_github_revloop-0.16.16}/src/main/alissa/tools/github/revloop/webui/page.py

@@ -437,14 +437,29 @@ _JS = r"""
     // The meter is the RESIDENT share of the charge, deliberately: it is the
     // only part of a rising charge that a limit can actually kill for, so the
     // shared warn/crit thresholds mean the same thing here as on the others.
+    // Gated on ANY of the three, not on `charged`: cgroup_memory reads
+    // memory.current and memory.stat through separate helpers so each degrades
+    // on its own, and a tile keyed to the headline alone would throw away a
+    // breakdown the reader deliberately preserved. The headline and the meter
+    // carry their own null handling, so a missing `charged` costs only itself.
     var mem = t.memory;
-    if (mem && mem.charged != null) {
+    if (mem && (mem.charged != null || mem.resident != null || mem.reclaimable != null)) {
       var msub = (mem.resident == null ? '--' : bytes(mem.resident)) + ' resident · ' +
         (mem.reclaimable == null ? '--' : bytes(mem.reclaimable)) + ' reclaimable';
+      // shmem is charged, is NOT droppable (swap-backed), and is excluded from
+      // reclaimable -- so it has to be visible, or a tmpfs-heavy container
+      // shows a charge that neither of the other two numbers accounts for.
+      if (mem.shmem) msub += ' · ' + bytes(mem.shmem) + ' shmem';
       var mpct = (mem.resident == null || !mem.charged) ? null :
         Math.round(100 * mem.resident / mem.charged);
       out += tile('Container Memory', bytes(mem.charged), msub, mpct);
-    } else { out += tile('Container Memory', '--', 'no cgroup v2'); }
+      // 'unavailable', not a cgroup-v2 verdict: this branch is also reached on a
+      // host-namespace deployment (the controller exists, its interface files
+      // are absent from the v2 root) and on a partial read. The console cannot
+      // tell those apart, and a wrong diagnosis sends an operator to check the
+      // wrong thing -- the neighbouring tiles say 'unavailable' for the same
+      // reason, and so does the origin task's acceptance detail.
+    } else { out += tile('Container Memory', '--', 'unavailable'); }
     out += tile('Review Queue', t.queue_depth, 'PRs awaiting me, last poll');
     el('tiles').innerHTML = out;
   }

{alissa_tools_github_revloop-0.16.15 → alissa_tools_github_revloop-0.16.16}/src/main/alissa/tools/github/revloop/webui/sources.py

@@ -501,6 +501,12 @@ class Sources:
         # trees out of it and the top-process list ranks the same snapshot, so
         # the two panels can never disagree about a process that exited between
         # them -- and the host pays for one walk per poll, not two.
+        #
+        # The cost on the other side of that trade, so nobody reorders this
+        # thinking it is free: the snapshot is now taken BEFORE `sessions()`
+        # shells out to tmux, so a pane that starts inside that window is
+        # missing from `stats` and its row shows "--" for CPU%/RSS for one
+        # poll. Self-healing in ~10s, and cheaper than walking /proc twice.
         proc_index = sysinfo.build_index(self._proc_root)
         sessions = self.sessions(index=proc_index)
         rate = self.rate_limit()

{alissa_tools_github_revloop-0.16.15 → alissa_tools_github_revloop-0.16.16}/src/main/alissa/tools/github/revloop/webui/sysinfo.py

@@ -276,6 +276,15 @@ def _sum_or_none(*values: "int | None") -> "int | None":
     return sum(value for value in values if value is not None)
 
 
+def _drop_shmem(file_bytes: "int | None", shmem: "int | None") -> "int | None":
+    """The droppable part of `file`. `shmem` unknown means the droppable part
+    is unknown -- publishing `file` whole would be the reassuring-direction
+    error the split exists to prevent."""
+    if file_bytes is None or shmem is None:
+        return None
+    return max(0, file_bytes - shmem)
+
+
 def cgroup_memory(
     cgroup_root: "str | os.PathLike[str]" = "/sys/fs/cgroup",
 ) -> dict:
@@ -287,13 +296,25 @@ def cgroup_memory(
     So we also report:
 
     * `resident` -- `anon`, the process memory that is really in use.
-    * `reclaimable` -- `file` + `slab_reclaimable`, the page cache plus the
-      shrinkable kernel caches: the "not really in use" bucket.
+    * `reclaimable` -- `(file - shmem) + slab_reclaimable`: the page cache plus
+      the shrinkable kernel caches, the "not really in use" bucket.
+    * `shmem` -- broken out of `file` rather than left inside it, because it is
+      not droppable: tmpfs, shm segments and shared anonymous mmaps are
+      swap-backed, so with swap disabled or capped (the normal container case)
+      the kernel cannot reclaim them under pressure -- they are what the
+      container gets OOM-killed for. Counting them as reclaimable is the one
+      way this reader can be wrong in the REASSURING direction, telling an
+      operator "5.9 GB of cache, no leak" about pages nothing can free. Issue
+      #74's scope text prescribes `file + slab_reclaimable` verbatim; the same
+      sentence calls the result the "not really in use" bucket, and where the
+      formula and the words disagree the words are the requirement. Surfaced as
+      its own magnitude instead of folded into `resident`, because "5 GB of
+      tmpfs" and "5 GB of process heap" are different operator problems even
+      though neither is droppable.
 
     The raw keys ride along for a spot check against an in-container
-    `memory.stat`. Note the three summary numbers do NOT partition the charge
-    (kernel stacks, pagetables, sockets are charged too, and `shmem` is counted
-    inside `file` while behaving like anon memory) -- they are the three
+    `memory.stat`. Note the summary numbers do NOT partition the charge --
+    kernel stacks, pagetables and sockets are charged too -- they are the
     magnitudes an operator compares, not a balance sheet.
 
     Never raises. On a host without cgroup v2 -- a dev laptop, macOS, a v1
@@ -306,7 +327,14 @@ def cgroup_memory(
     out: "dict[str, int | None]" = {
         "charged": charged,
         "resident": stat["anon"],
-        "reclaimable": _sum_or_none(stat["file"], stat["slab_reclaimable"]),
+        # `shmem` is a SUBSET of `file` in cgroup v2, so it is subtracted, not
+        # added. Clamped at zero: the two keys are sampled from one read of one
+        # file and cannot legitimately cross, but a negative "cache" number
+        # would be a worse thing to render than a zero if a kernel ever
+        # disagreed with that.
+        "reclaimable": _sum_or_none(
+            _drop_shmem(stat["file"], stat["shmem"]), stat["slab_reclaimable"]
+        ),
     }
     out.update(stat)
     return out

{alissa_tools_github_revloop-0.16.15 → alissa_tools_github_revloop-0.16.16/src/main/alissa_tools_github_revloop.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: alissa-tools-github-revloop
-Version: 0.16.15
+Version: 0.16.16
 Summary: ALISSA-TOOLS-GITHUB-REVLOOP
 Home-page: https://alissa.app
 Author: Fahera

alissa_tools_github_revloop-0.16.15/src/main/alissa/tools/github/revloop/version

@@ -1 +0,0 @@
-0.16.15