comfyui-mcp 0.52.178 → 0.52.180

package/dist/orchestrator/panel-tools.js

@@ -1338,7 +1338,7 @@ function isWorkerTransportRetrySafeCmd(cmd) {
 // This is an EXPLICIT ALLOWLIST, deliberately NOT "everything not read-only":
 // several genuine reads/probes/views that flow through ctx.call are absent from
 // BRIDGE_READONLY_CMDS (e.g. graph_list_subgraphs, training_get_state,
-// graph_canvas, graph_screenshot), so an exclusion rule would wrongly make THOSE
+// graph_canvas), so an exclusion rule would wrongly make THOSE
 // wait out the reconnect budget. Under-inclusion here is at worst an unfixed edge
 // (a command keeps today's behavior); over-inclusion would regress a read — so we
 // list only commands that unambiguously mutate the graph. Keep in sync when new
@@ -21432,8 +21432,39 @@ CHECKED FOR YOU: the graph read this message prescribes was just run, and it ` +
                 // never carries the token (a ledger answer for a read is a STALE
                 // outcome), and the withRetryToken wrapper can't reach this direct send.
                 { cmd: "graph_screenshot", padding: args.padding }, target ?? { mode: "current" });
+                // panel#2191 — the SAME bounded budget the sibling idempotent reads
+                // already take, for the same reason and against the same evidence.
+                //
+                // A screenshot's cost scales with the graph: the frontend handler forces
+                // one synchronous LiteGraph repaint of every node and group at the fitted
+                // transform, PNG-encodes the result, and repaints again on restore — all
+                // on the panel's single main thread. Dispatched with no `timeoutMs` it
+                // took the bare 20 000 ms default, and a reporter's 264-node / 63-group
+                // graph was declared "backgrounded or frozen" on a tab that, in the same
+                // session, answered panel_query_graph, panel_get_errors over all 264
+                // nodes, and panel_save_workflow. That is a busy-but-alive main thread,
+                // which is precisely what graph_get_errors' own budget comment describes
+                // — and graph_get_errors survived that graph because it already passes
+                // this constant while the screenshot alone was left on the default.
+                //
+                // Safe to widen because the read is idempotent (see the panel#2191 entry
+                // in BRIDGE_READONLY_CMDS): waiting longer costs a slow reply, never a
+                // double-applied write. Still BOUNDED, never Infinity, so a genuinely
+                // frozen tab fails — at 90 s instead of 20 s.
+                //
+                // ONE-SHOT ACROSS RECONNECTS, deliberately — and still so after #2761.
+                // Being a read makes this command eligible for the bridge's
+                // park-and-resume. That resume now requires the returning connection to
+                // PROVE it is the tab that issued the read (#2761), so the wrong-canvas
+                // outcome this originally guarded against is closed at the bridge for
+                // every command in the set, not just this one. The budget stays zero
+                // regardless: a capture is free to re-ask, so it gains nothing from
+                // resuming, and keeping it out of the branch entirely means this call
+                // site does not depend on that rule staying correct.
                 const res = (await ctx.bridge.send(cmd, {
                     tabId: ctx.tabId,
+                    timeoutMs: OBJECT_INFO_REFRESH_ACK_TIMEOUT_MS,
+                    maxReconnectRetries: 0,
                 }));
                 if (!res?.image)
                     return fail("screenshot returned no image");