comfyui-mcp 0.52.53 → 0.52.55

package/README.md

@@ -29,7 +29,7 @@ Works on **macOS**, **Linux**, and **Windows**. Auto-detects your ComfyUI instal
 
 **Stuck or have a question? [Join the Discord](https://discord.gg/cW9arBhzCu)** — help, model tips, and release announcements.
 
-**37 MCP tools** | **41 AI skills** (Flux · WAN · LTX 2.3 video · MiniMax H3 · Qwen · Z-Image · Ideogram 4 · ERNIE · ANIMA · model registry · Civitai · node authoring · launch/perf flags) | **56 installer packs** | **11 slash commands** | **4 autonomous agents** | **3 hooks**
+**38 MCP tools** | **41 AI skills** (Flux · WAN · LTX 2.3 video · MiniMax H3 · Qwen · Z-Image · Ideogram 4 · ERNIE · ANIMA · model registry · Civitai · node authoring · launch/perf flags) | **56 installer packs** | **11 slash commands** | **4 autonomous agents** | **3 hooks**
 
 The plugin ships **expert skills that grow with every release** — model-specific generation guides with curated download URLs, workflow recipes, troubleshooting, and custom-node authoring — so Claude knows the right sampler, CFG, resolution, and model files for each architecture without trial and error.
 
@@ -263,7 +263,7 @@ for the port, the capability matrix, and the per-provider "clink" points, and th
 
 ## MCP Tools
 
-37 tools across workflow execution, generation, iteration, composition, models, and more:
+38 tools across workflow execution, generation, iteration, composition, models, and more:
 
 ### Image Generation (high-level)
 
@@ -370,6 +370,7 @@ Install [comfy-cli](https://docs.comfy.org/comfy-cli/getting-started#install-cli
 | Tool | Description |
 |------|-------------|
 | `get_system_stats (action:"logs")` | Get ComfyUI server logs with optional keyword filter (e.g., `error`, `warning`, a node name) |
+| `kitchen` | See what comfy-kitchen can do on this GPU (`status`), find where a graph leaves it on the table (`assess`), and apply the faster path (`apply`) — a failed probe is unknown, never a no |
 | `get_history` `action: "list"` | Get execution history with full error details, Python tracebacks, timing, and cached node info |
 | `get_history` `action: "diagnose"` | Explain a FAILED run in one call — the failed node and traceback PLUS the missing models (file + widget) and missing node types |
 

package/dist/orchestrator/download-done-guard.js

@@ -1,34 +1,41 @@
 /**
- * #1574 — when the completion event and `download_model action:"status"` disagree, say so.
- *
- * The tray raised `transfer completed` for an 11.46GB download while
- * `download_model action:"status"` reported it still streaming, `list_local_models` showed the
- * file absent, and the category count only rose minutes later. The file landed AFTER the
- * event.
+ * #1574 / #2057 — when the tray event and `download_model action:"status"` disagree, say so.
  *
  * The completion event is built from a PROGRESS ROW read off disk each tick. `status` answers
  * from the JOB RECORD. Two stores, and the event consults only one — the same split #1545
  * documented from the other side.
  *
+ * #1574: the tray raised `transfer completed` for an 11.46GB download while status reported
+ * it still streaming. The file landed AFTER the event.
+ *
+ * #2057: the tray raised `Model download FAILED` (and claimed nothing transferred) for a
+ * 19.53GB HuggingFace fetch while status — same id — still said **downloading** with
+ * advancing bytes (0.14 → 0.22 GB). An agent that trusts FAILED re-issues into a live
+ * writer, which is the corruption hazard status itself warns about. #1150 only hedges
+ * when a LIVE TRAY ROW still shows that filename downloading; here the tray row itself
+ * was the error, so that hedge never fired. The job record is the same authority status
+ * reads, so a failure has to be checked against it the same way a completion is.
+ *
  * ## Why this ANNOTATES and never suppresses
  *
- * The first version of this dropped the contradicted event. Review killed it, correctly: a
- * terminal record may legitimately still read `downloading` until the ~15s persistence
- * heartbeat retries (see `persistDownloadJob`'s return value, #1545). The debounce bucket is
- * deleted before filtering and never requeued, so suppressing on a lagging record would
- * PERMANENTLY lose the completion notification for a download that genuinely finished.
+ * The first version of the completion guard dropped the contradicted event. Review killed
+ * it, correctly: a terminal record may legitimately still read `downloading` until the ~15s
+ * persistence heartbeat retries (see `persistDownloadJob`'s return value, #1545). The
+ * debounce bucket is deleted before filtering and never requeued, so suppressing on a
+ * lagging record would PERMANENTLY lose the notification.
  *
  * That trades a confusing message for a missing one, which is worse: the user is waiting on
- * that event. So the event always fires, and a disagreement is disclosed on it. The harm in
- * the report is a CONFIDENT false completion — "the natural next action is to use the file" —
- * and a hedge is precisely what removes that.
+ * that event. So the event always fires, and a disagreement is disclosed on it. For a
+ * completion the harm is a CONFIDENT false success; for a failure it is a CONFIDENT false
+ * FAILED — both removed by hedging, not by silence.
  *
  * ## Identity
  *
  * The two stores do not share an id. A progress row's `id` is the progress/tray identity;
- * the job's `id` is its public status handle (`6226e26ba97f8527` in the report, against tray
- * `93015fbfa0fa9933`). The row is matched against the job's `progressId ?? trayId`, which is
- * what writes those rows. An earlier version compared row id to job id and was therefore
+ * the job's `id` is its public status handle (`6226e26ba97f8527` in the #1574 report,
+ * against tray `93015fbfa0fa9933`; `07d14bb0c73bc007` against tray `31bba4a9cdb04e28` in
+ * #2057). The row is matched against the job's `progressId ?? trayId`, which is what
+ * writes those rows. An earlier version compared row id to job id and was therefore
  * INERT — it never matched anything, and its unit tests missed that because they fed
  * synthetic rows carrying whatever id the assertion wanted.
  */
@@ -38,6 +45,44 @@ function progressIdentityOf(job) {
     const tray = typeof job.trayId === "string" ? job.trayId : null;
     return progress ?? tray;
 }
+/**
+ * The job record for this progress row, or null when nothing matches.
+ *
+ * (id, target) — the SAME key the supersession logic uses, and for the same reason: a
+ * concurrent LOCAL and POD transfer of one URL shares an id but is two transfers with two
+ * outcomes. Matching on id alone could annotate the wrong terminal, or miss a real
+ * disagreement by finding the other one first (review).
+ *
+ * A target is compared only when BOTH sides carry one. Rows and records that predate the
+ * field, or a route that never sets it, must not silently stop matching — that would make
+ * the check inert again, which is exactly how the first version shipped.
+ *
+ * PREFER THE EXACT (id, target) MATCH (review, round 3). Taking the first id match in
+ * array order let a TARGETLESS record shadow the exact one: a targetless "downloading"
+ * sitting before an exact-target terminal reported a disagreement that does not exist.
+ * The targetless record still stands in when nothing matches on target.
+ */
+function matchingRecord(row, jobs) {
+    if (!row || typeof row !== "object")
+        return null;
+    const id = typeof row.id === "string" ? row.id : null;
+    if (!id)
+        return null;
+    if (!Array.isArray(jobs))
+        return null;
+    const target = typeof row.target === "string" ? row.target : null;
+    const sameId = jobs.filter((j) => j && typeof j === "object" && progressIdentityOf(j) === id);
+    if (!sameId.length)
+        return null;
+    return ((target ? sameId.find((j) => j.target === target) : undefined) ??
+        sameId.find((j) => typeof j.target !== "string") ??
+        (target ? undefined : sameId[0]) ??
+        null);
+}
+function recordStillDownloading(row, jobs) {
+    const record = matchingRecord(row, jobs);
+    return record != null && record.status === "downloading";
+}
 /**
  * Does the job record disagree with announcing this row as completed?
  *
@@ -48,41 +93,26 @@ function progressIdentityOf(job) {
 export function completionDisagreesWithRecord(row, jobs) {
     if (!row || typeof row !== "object")
         return false;
-    // Only COMPLETIONS. A failure event carries its own hedged wording (#1150) and must be
-    // left entirely alone.
+    // Completions only. Failures have their own check (#2057); #1150's live-tray hedge
+    // stays a separate question (a different id writing the same filename).
     if (row.status !== "done")
         return false;
-    const id = typeof row.id === "string" ? row.id : null;
-    if (!id)
-        return false;
-    if (!Array.isArray(jobs))
-        return false;
-    // (id, target) — the SAME key the supersession logic uses, and for the same reason: a
-    // concurrent LOCAL and POD transfer of one URL shares an id but is two transfers with two
-    // outcomes. Matching on id alone could annotate the wrong completion, or miss a real
-    // disagreement by finding the other one first (review).
-    //
-    // A target is compared only when BOTH sides carry one. Rows and records that predate the
-    // field, or a route that never sets it, must not silently stop matching — that would make
-    // the check inert again, which is exactly how the first version shipped.
-    const target = typeof row.target === "string" ? row.target : null;
-    const sameId = jobs.filter((j) => j && typeof j === "object" && progressIdentityOf(j) === id);
-    if (!sameId.length)
+    return recordStillDownloading(row, jobs);
+}
+/**
+ * Does the job record disagree with announcing this row as FAILED?
+ *
+ * Same identity and same positive-contradiction rule as {@link completionDisagreesWithRecord}.
+ * #1150 only looks at live TRAY rows of the same filename; this looks at the job RECORD
+ * `download_model action:"status"` reads. The reported case had no live tray row — the tray
+ * itself was the error — while status still showed the same id streaming and advancing.
+ */
+export function failureDisagreesWithRecord(row, jobs) {
+    if (!row || typeof row !== "object")
         return false;
-    // PREFER THE EXACT (id, target) MATCH (review, round 3). Taking the first id match in
-    // array order let a TARGETLESS record shadow the exact one: a targetless "downloading"
-    // sitting before an exact-target "done" reported a disagreement that does not exist, and
-    // would have hedged a completion that was perfectly fine.
-    //
-    // The targetless record still stands in when nothing matches on target — that is what
-    // keeps rows and records predating the field from silently going unmatched, which would
-    // make the whole check inert again.
-    const record = (target ? sameId.find((j) => j.target === target) : undefined) ??
-        sameId.find((j) => typeof j.target !== "string") ??
-        (target ? undefined : sameId[0]);
-    if (!record)
+    if (row.status !== "error")
         return false;
-    return record.status === "downloading";
+    return recordStillDownloading(row, jobs);
 }
 /** The disclosure appended to a completion the record disagrees with. Deliberately states
  *  BOTH readings and what to do, rather than picking a winner this cannot establish. */
@@ -91,4 +121,8 @@ export const COMPLETION_DISAGREEMENT_NOTE = "CAVEAT: `download_model action:\"st
     "seconds, so this may simply be that — but a completion event has also been observed to " +
     "arrive minutes before the file existed (#1574). Check `download_model action:\"status\"` " +
     "and that the file is present before loading it.";
+/** Instruction for a FAILED tray row whose job record is still downloading (#2057).
+ *  Must not contain the word FAILED: that headline is what made the agent re-issue. */
+export const FAILURE_DISAGREEMENT_NOTE = "`download_model action:\"status\"` still reports this transfer as downloading — " +
+    "do NOT report it as failed";
 //# sourceMappingURL=download-done-guard.js.map

package/dist/orchestrator/download-done-guard.js.map

@@ -1 +1 @@
-{"version":3,"file":"download-done-guard.js","sourceRoot":"","sources":["../../src/orchestrator/download-done-guard.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAiCG;AAmBH,oDAAoD;AACpD,SAAS,kBAAkB,CAAC,GAAY;IACtC,MAAM,QAAQ,GAAG,OAAO,GAAG,CAAC,UAAU,KAAK,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,UAAU,CAAC,CAAC,CAAC,IAAI,CAAC;IAC5E,MAAM,IAAI,GAAG,OAAO,GAAG,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC;IAChE,OAAO,QAAQ,IAAI,IAAI,CAAC;AAC1B,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,6BAA6B,CAAC,GAAY,EAAE,IAAwB;IAClF,IAAI,CAAC,GAAG,IAAI,OAAO,GAAG,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAC;IAClD,uFAAuF;IACvF,uBAAuB;IACvB,IAAI,GAAG,CAAC,MAAM,KAAK,MAAM;QAAE,OAAO,KAAK,CAAC;IACxC,MAAM,EAAE,GAAG,OAAO,GAAG,CAAC,EAAE,KAAK,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC;IACtD,IAAI,CAAC,EAAE;QAAE,OAAO,KAAK,CAAC;IACtB,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC;QAAE,OAAO,KAAK,CAAC;IACvC,sFAAsF;IACtF,0FAA0F;IAC1F,qFAAqF;IACrF,wDAAwD;IACxD,EAAE;IACF,yFAAyF;IACzF,0FAA0F;IAC1F,yEAAyE;IACzE,MAAM,MAAM,GAAG,OAAO,GAAG,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC;IAClE,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,IAAI,OAAO,CAAC,KAAK,QAAQ,IAAI,kBAAkB,CAAC,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC;IAC9F,IAAI,CAAC,MAAM,CAAC,MAAM;QAAE,OAAO,KAAK,CAAC;IACjC,sFAAsF;IACtF,uFAAuF;IACvF,yFAAyF;IACzF,0DAA0D;IAC1D,EAAE;IACF,sFAAsF;IACtF,wFAAwF;IACxF,oCAAoC;IACpC,MAAM,MAAM,GACV,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,MAAM,KAAK,MAAM,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QAC9D,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,CAAC,MAAM,KAAK,QAAQ,CAAC;QAChD,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC;IACnC,IAAI,CAAC,MAAM;QAAE,OAAO,KAAK,CAAC;IAC1B,OAAO,MAAM,CAAC,MAAM,KAAK,aAAa,CAAC;AACzC,CAAC;AAED;wFACwF;AACxF,MAAM,CAAC,MAAM,4BAA4B,GACvC,yFAAyF;IACzF,2FAA2F;IAC3F,yFAAyF;IACzF,2FAA2F;IAC3F,iDAAiD,CAAC"}
+{"version":3,"file":"download-done-guard.js","sourceRoot":"","sources":["../../src/orchestrator/download-done-guard.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAwCG;AAmBH,oDAAoD;AACpD,SAAS,kBAAkB,CAAC,GAAY;IACtC,MAAM,QAAQ,GAAG,OAAO,GAAG,CAAC,UAAU,KAAK,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,UAAU,CAAC,CAAC,CAAC,IAAI,CAAC;IAC5E,MAAM,IAAI,GAAG,OAAO,GAAG,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC;IAChE,OAAO,QAAQ,IAAI,IAAI,CAAC;AAC1B,CAAC;AAED;;;;;;;;;;;;;;;;GAgBG;AACH,SAAS,cAAc,CAAC,GAAY,EAAE,IAAwB;IAC5D,IAAI,CAAC,GAAG,IAAI,OAAO,GAAG,KAAK,QAAQ;QAAE,OAAO,IAAI,CAAC;IACjD,MAAM,EAAE,GAAG,OAAO,GAAG,CAAC,EAAE,KAAK,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC;IACtD,IAAI,CAAC,EAAE;QAAE,OAAO,IAAI,CAAC;IACrB,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC;QAAE,OAAO,IAAI,CAAC;IACtC,MAAM,MAAM,GAAG,OAAO,GAAG,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC;IAClE,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,IAAI,OAAO,CAAC,KAAK,QAAQ,IAAI,kBAAkB,CAAC,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC;IAC9F,IAAI,CAAC,MAAM,CAAC,MAAM;QAAE,OAAO,IAAI,CAAC;IAChC,OAAO,CACL,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,MAAM,KAAK,MAAM,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QAC9D,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,CAAC,MAAM,KAAK,QAAQ,CAAC;QAChD,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;QAChC,IAAI,CACL,CAAC;AACJ,CAAC;AAED,SAAS,sBAAsB,CAAC,GAAY,EAAE,IAAwB;IACpE,MAAM,MAAM,GAAG,cAAc,CAAC,GAAG,EAAE,IAAI,CAAC,CAAC;IACzC,OAAO,MAAM,IAAI,IAAI,IAAI,MAAM,CAAC,MAAM,KAAK,aAAa,CAAC;AAC3D,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,6BAA6B,CAAC,GAAY,EAAE,IAAwB;IAClF,IAAI,CAAC,GAAG,IAAI,OAAO,GAAG,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAC;IAClD,mFAAmF;IACnF,wEAAwE;IACxE,IAAI,GAAG,CAAC,MAAM,KAAK,MAAM;QAAE,OAAO,KAAK,CAAC;IACxC,OAAO,sBAAsB,CAAC,GAAG,EAAE,IAAI,CAAC,CAAC;AAC3C,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,0BAA0B,CAAC,GAAY,EAAE,IAAwB;IAC/E,IAAI,CAAC,GAAG,IAAI,OAAO,GAAG,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAC;IAClD,IAAI,GAAG,CAAC,MAAM,KAAK,OAAO;QAAE,OAAO,KAAK,CAAC;IACzC,OAAO,sBAAsB,CAAC,GAAG,EAAE,IAAI,CAAC,CAAC;AAC3C,CAAC;AAED;wFACwF;AACxF,MAAM,CAAC,MAAM,4BAA4B,GACvC,yFAAyF;IACzF,2FAA2F;IAC3F,yFAAyF;IACzF,2FAA2F;IAC3F,iDAAiD,CAAC;AAEpD;uFACuF;AACvF,MAAM,CAAC,MAAM,yBAAyB,GACpC,kFAAkF;IAClF,4BAA4B,CAAC"}

package/dist/orchestrator/index.js

@@ -48,7 +48,7 @@ import { setConnectedPanelOrigins } from "../comfyui/fetch.js";
 import { publishConnectedPanelOrigins } from "../services/panel-origin-channel.js";
 import { logger } from "../utils/logger.js";
 import { listDownloadJobs } from "../services/download-jobs.js";
-import { completionDisagreesWithRecord } from "./download-done-guard.js";
+import { completionDisagreesWithRecord, failureDisagreesWithRecord } from "./download-done-guard.js";
 import { assembleVocabularyHash, describeVocabularySkew } from "../tools/vocabulary.js";
 import { buildPanelToolDefs } from "./panel-tools.js";
 /** The panel vocabulary hashes whose MISMATCH has already been reported (#236).
@@ -129,6 +129,7 @@ import { dedupeAudioRefs, splitAudioAttachments } from "./audio-attachment.js";
 import { startPanelConsoleHttpServer } from "./panel-console-http.js";
 import { readComfyuiCrashLog, formatCrashNote } from "../services/crash-log.js";
 import { QueueMonitor } from "../services/queue-monitor.js";
+import { formatQueueNote } from "./queue-note.js";
 import { RunCompletions, describe as describeCorrelation, } from "./run-completion-journal.js";
 import { createRunCompletionWatchdog, resolveHistoryCompletionImages, resolveHistoryCompletionStatus, } from "./run-completion-watchdog.js";
 import { AskAnswers, preview as previewQuestion } from "./ask-answer-journal.js";
@@ -170,7 +171,7 @@ const MCP_VERSION_RUNNING = (() => {
  *  silently failed to interpolate would type-check, build, and ship the placeholder. */
 export const PANEL_SYSTEM_APPEND = `You are the autonomous assistant embedded directly in a ComfyUI sidebar panel. The person is working in ComfyUI and talks to you through that panel: their messages arrive as your prompts, and everything you write is shown to them in the panel chat. Write for that reader — lead with the result, keep replies short and concrete, and don't narrate routine internal steps.
 
-You can SEE and EDIT the workflow the user currently has open, via the panel_* tools (panel_graph_outline, panel_query_graph, panel_add_node, panel_connect, panel_set_widget, panel_run, panel_get_errors, panel_save_workflow, …). STRONGLY PREFER building on their live canvas: read it first (panel_graph_outline, then panel_query_graph for specifics), add/wire/configure nodes with the panel_* tools, then panel_run to queue it — so the user watches the work happen and the result loads in their own workflow with full Ctrl+Z undo. Only fall back to the headless generate_image/enqueue_workflow tools when the user explicitly wants a one-off they don't need on their canvas, or when no panel tab is connected (a panel_* call will error if so). On a LARGE graph (a loaded pack/template with dozens of nodes), do NOT dump the whole thing and scan it — and NEVER shell out to grep/jq/python over a saved workflow file. To UNDERSTAND the graph, call panel_graph_outline FIRST: a compact, dependency-ordered TEXT map (nodes topologically sorted source→sink, each with its key widgets and ← inputs / → outputs wiring, plus a groups index) made for you to read top-to-bottom. To PINPOINT and INSPECT specific nodes, use panel_query_graph: filter by types/title/widget predicates ('cfg>7'), traverse upstream_of/downstream_of a node, aggregate with group_by:'type', and read ONE node's exact slot/widget detail with {ids:[id], fields:'detail'} — output is token-bounded so it can never flood your context. panel_find_nodes remains for free-text search across all fields.
+You can SEE and EDIT the workflow the user currently has open, via the panel_* tools (panel_graph_outline, panel_query_graph, panel_add_node, panel_connect, panel_set_widget, panel_run, panel_get_errors, panel_save_workflow, panel_kitchen, …). STRONGLY PREFER building on their live canvas: read it first (panel_graph_outline, then panel_query_graph for specifics), add/wire/configure nodes with the panel_* tools, then panel_run to queue it — so the user watches the work happen and the result loads in their own workflow with full Ctrl+Z undo. Only fall back to the headless generate_image/enqueue_workflow tools when the user explicitly wants a one-off they don't need on their canvas, or when no panel tab is connected (a panel_* call will error if so). On a LARGE graph (a loaded pack/template with dozens of nodes), do NOT dump the whole thing and scan it — and NEVER shell out to grep/jq/python over a saved workflow file. To UNDERSTAND the graph, call panel_graph_outline FIRST: a compact, dependency-ordered TEXT map (nodes topologically sorted source→sink, each with its key widgets and ← inputs / → outputs wiring, plus a groups index) made for you to read top-to-bottom. To PINPOINT and INSPECT specific nodes, use panel_query_graph: filter by types/title/widget predicates ('cfg>7'), traverse upstream_of/downstream_of a node, aggregate with group_by:'type', and read ONE node's exact slot/widget detail with {ids:[id], fields:'detail'} — output is token-bounded so it can never flood your context. panel_find_nodes remains for free-text search across all fields.
 
 ${DEFERRED_PANEL_TOOLS_STEERING}
 
@@ -397,31 +398,6 @@ function stallThresholdMs() {
     const s = Number(process.env.COMFYUI_MCP_STALL_S);
     return Number.isFinite(s) && s > 0 ? Math.round(s * 1000) : 180000;
 }
-/** Build a one-line agent note from a stall/backlog report, or null when the
- *  queue is healthy. Stall takes priority over a plain backlog. */
-function formatQueueNote(rep) {
-    if (rep.stalled) {
-        const secs = Math.round(rep.stalledForMs / 1000);
-        return (`⚠️ The current ComfyUI render appears STALLED: ` +
-            `${rep.currentNode ? `node ${rep.currentNode} ` : ""}${rep.progress ? `(progress ${rep.progress}) ` : ""}` +
-            `on prompt ${rep.runningPromptId ?? "?"} has not advanced for ~${secs}s. ComfyUI only checks interrupts ` +
-            `BETWEEN steps, so a stuck step can ignore a cancel. If it's wedged: call queue (action:"cancel") with ` +
-            `clear_pending:true; if it reports the job still wedged, restart_comfyui / panel_restart_comfyui. ` +
-            `Do NOT queue another run on top.`);
-    }
-    // A plain backlog (depth > 1) is NOT evidence of a wedge — deliberately queuing
-    // a batch is a normal workflow. Suppress the note entirely when the in-flight
-    // work is this session's own recent jobs, and when it isn't, report it NEUTRALLY
-    // (no false "you likely queued behind a stuck job" diagnosis, no destructive
-    // clear_pending as the headline). A genuinely stalled job is handled above (#559).
-    if (rep.backlog && !rep.selfAttributed) {
-        const pending = Math.max(0, rep.queueDepth - 1);
-        return (`ℹ️ ComfyUI queue: ${rep.queueDepth} tasks in flight (1 running + ${pending} pending) that this session ` +
-            `didn't queue. This is only a problem if the running one is stuck — inspect with queue (action:"list"). ` +
-            `queue (action:"cancel_queued") drops a single pending item; queue (action:"cancel") with clear_pending:true resets everything.`);
-    }
-    return null;
-}
 /**
  * Lockfile path for a given bridge port. The orchestrator self-registers its
  * REAL node pid here (not the npx shim's), plus the ComfyUI pid that launched
@@ -5574,6 +5550,18 @@ export async function runPanelOrchestrator() {
             if (disagreeing.length) {
                 logger.warn("[panel-orchestrator] a download completion disagrees with the job record; disclosing rather than suppressing (#1574)", { ids: disagreeing.map((d) => String(d.id ?? "")) });
             }
+            // #2057 — the SAME split in the other direction. A tray ERROR while the job record
+            // still says downloading is how "Model download FAILED" + "NOTHING transferred"
+            // fired for a 19.53 GB fetch whose status (same id) was still streaming and
+            // advancing. #1150 only sees a live TRAY row of that filename; here the tray
+            // row itself was the error, so that hedge never ran. Flag it so the formatter
+            // will not say FAILED.
+            const failedDisagreeing = settled.filter((d) => failureDisagreesWithRecord(d, records));
+            for (const d of failedDisagreeing)
+                d.recordDisagrees = true;
+            if (failedDisagreeing.length) {
+                logger.warn("[panel-orchestrator] a download failure disagrees with the job record; disclosing rather than announcing FAILED (#2057)", { ids: failedDisagreeing.map((d) => String(d.id ?? "")) });
+            }
             // #884 — a download has no originating TAB (its row names the owning
             // conversation), so its turn INHERITS the conversation's LAST
             // ESTABLISHED origin — never the active tab (confirming gate 2, P0 rule: