Comfy Debug

When something fails on the comfy CLI, read the JSON envelope first, act on the hint, and only go deeper if needed. This skill is the bottom-up playbook.

The envelope is the source of truth

Every command emits the standard JSON envelope (see core skill). When error is present, map error.code to a fix below.

Three rules:

1. Map error.code to a fix below. Most codes have a one-line resolution.
2. hint is authoritative. If it says "run: comfy launch", that's the fix. Don't paraphrase, don't guess.
3. details carries the why. Look here before scrolling logs.

Decision tree for common failures

server_not_running (local)

comfy launch              # start the local server in the foreground
# or, for background mode:
comfy launch --background

If comfy launch itself fails: check comfy --json env for the resolved interpreter and workspace, and inspect whatever stdout comfy launch was attached to.

cloud_not_configured / cloud_unauthorized (cloud)

comfy cloud login                   # opens browser, OAuth + PKCE
# or, if you already have an API key:
comfy cloud set-key --key sk-…      # alternative path; API key is a fallback, not an override
comfy --json cloud whoami           # confirm signed_in: true (auth_method is "oauth" or "api_key")

Auth precedence is OAuth-first: a live comfy cloud login session outranks COMFY_CLOUD_API_KEY env, which outranks a stored key. comfy cloud whoami shows which credential is active. If a stale session is causing 401s, comfy cloud logout && comfy cloud login — setting an API key will NOT override a live session. If you're on a custom env (PR preview), set the base URL first:

comfy cloud set-base-url https://fe-pr-NNNNN.testenvs.comfy.org

Do not confuse this with transient_auth (below). cloud_unauthorized happens before submission (your CLI session); an "Unauthorized" that appears as a job failure is the server-side token expiry and re-login does nothing.

transient_auth (cloud job failed with "Unauthorized: Please login first to use this node")

An API node's server-side session token expired mid-execution. It is transient and has nothing to do with your local credentials — comfy cloud login will NOT fix it, and your auth was fine at submit time (often it strikes minutes into a long render, sometimes a whole batch at once).

comfy --json run --workflow same_workflow.json --where cloud   # just resubmit — succeeds on retry

If several jobs from one batch died together with this code, resubmit them all; one retry each is normally enough.

workflow_not_api_format

UI-format workflows are converted to API format client-side using object_info (no server conversion endpoint exists). If conversion fails with conversion_error, re-export via File > Export (API) in ComfyUI; if object_info can't be fetched (cql_no_graph), retry (it's fetched live — check cloud sign-in/connection) or start a local server.

workflow_invalid_json

The file isn't valid JSON. Inspect the first/last 100 bytes — often it's an HTML error page that was saved with .json.

prompt_rejected (server returned node_errors)

The server validated the workflow and rejected nodes. The full per-node error map is in details.node_errors. The two most common causes:

• Unknown class_type → the cloud or local server lacks the custom node. Run:

  comfy --json nodes search MissingNodeName
  

  If data.count is 0 — or data.close_match is true, meaning the search found nothing and fell back to similarly-named classes — the node is genuinely absent. Then either install via comfy node install <pkg> (local) or pick a different workflow (cloud).

• Missing model file (ckpt_name, lora_name, vae_name not found) → confirm the loader node exists and see its choices:

  comfy --json nodes show CheckpointLoaderSimple   # inputs[].choices lists available checkpoints
  comfy --json env                                  # see workspace/models path
  

cloud_http_error (HTTP 4xx/5xx from cloud)

details.status is the HTTP code. details.body is the truncated server response. Most common:

• 401 invalid auth token → token audience mismatch. Decode the token at jwt.io or via python3 -c 'import base64,json,sys; ...' and check aud. The aud field must match what /api/prompt expects.
• 404 returning XML <AuthenticationRequired> → wrong path. Real ComfyUI endpoints on cloud live under /api/*; everything else hits the CDN catch-all.
• 503 → check the Comfy Cloud dashboard or server logs for deployment health.

cloud_timeout

cloud_timeout — the run went silent for --timeout seconds (default 120). comfy run --timeout is a per-event-silence deadline on both local and cloud: it resets whenever the job reports progress, so a workflow streaming progress can run indefinitely. Wall-clock limits exist only on comfy jobs watch --max-wait (default 600s, cloud). Recovery: re-run with a larger --timeout, or submit async and comfy jobs watch <id>.

RES=$(comfy --json run --workflow X.json --where cloud)   # async by default
PID=$(echo "$RES" | jq -r .data.prompt_id)
comfy --json-stream jobs watch "$PID" --where cloud --max-wait 1800

ws_timeout / ws_disconnected (local)

The WebSocket dropped mid-stream. Reconnect:

comfy jobs status <prompt_id>      # one-shot snapshot
comfy --json-stream jobs watch <prompt_id>   # re-attach

The local server keeps the job; the CLI just lost its tail.

cql_query_invalid

cql_query_invalid — raised when a legacy --query string is passed to comfy templates ls. There is no query grammar; use flag-based filtering instead: --type image|video|audio, --tag <t>, --model <m> (templates) and --produces/--accepts/--category/--pack (comfy nodes ls).

cql_no_graph

The CLI needs an object_info.json to query against. Two options:

comfy launch                                        # then re-run the nodes command
comfy --json nodes ls --produces IMAGE --input /path/to/object_info.json

comfy generate is partially machine-readable

comfy generate is partially machine-readable: generate <model> --json and generate resume <id> --json print the raw API response as JSON; generate upload <file> --json emits structured {url, expires_at, ...}; generate <model> --emit-workflow out.json goes through the full renderer envelope (command: "generate emit-workflow", error code emit_workflow_failed). The default proxy path (generate <model> without those flags) is still pretty-only — don't parse it. Known sharp edge: generate resume for BFL can raise a raw traceback on an expired/unknown id (HTTP 404 is unwrapped).

Job-stuck triage

If a prompt seems to hang:

# 1. Is it still tracked?
comfy --json jobs ls

# 2. What does the server think?
comfy --json jobs status <prompt_id>

# 3. Watch live events (NDJSON, one per line)
comfy --json-stream jobs watch <prompt_id>

Three possible answers:

Interrupting / cancelling

comfy --json run --workflow X.json    # async by default, returns a prompt_id
# later, if it's running too long:
comfy jobs cancel <prompt_id>         # works on both local and cloud

When --json makes diagnosis worse, not better

Pretty mode hides the envelope behind a rendered panel. If the rendered output is opaque, always re-run with --json — that's the contract. Conversely, if --json floods you with NDJSON, drop -stream to get a single envelope.

What to NOT do

• Don't chmod files in the model dirs to "fix" load failures — the path is wrong, not the permissions.
• Don't delete the OAuth session to "reset" — that loses state. Use comfy cloud logout and comfy skills uninstall instead.
• Don't paste raw error messages into prompts without the details block — the details are usually what disambiguates the fix.

Where to look when nothing in the envelope helps

1. comfy --json env → interpreter, workspace, server reachability, conda env
2. comfy --json which → resolved workspace path + how it was picked
3. comfy --json cloud whoami → cloud session state + base URL
4. Server logs:
   • Local: whatever stdout comfy launch was attached to (foreground), or the journal/console for comfy launch --background
   • Cloud: check the Comfy Cloud dashboard or server logs
5. The bundled comfy skill — re-read it; the right command for what you're trying to do is almost certainly already documented there.