> ## Documentation Index
> Fetch the complete documentation index at: https://comfyui-mcp.artokun.io/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# June 27: render-wedge watchdog, color analysis, provider onboarding

> A self-healing queue/render watchdog (backpressure + stall detection + escalating cancel), a new analyze_color tool, registry-first custom-node installs with a hardened clone fallback, and a wave of panel polish — provider onboarding, rich media metadata, code-block copy/wrap, and a background-tab streaming fix. Plus the Krea 2 ULTRA V2 re-slice.

*June 27, 2026*

A busy day across the MCP server (0.20.4 → 0.20.9) and the Agent Panel
(0.4.3 → 0.4.4): the headline is a **self-healing queue/render watchdog**, joined by
color analysis, much more reliable custom-node installs, and a round of panel polish.
The krea2 packs were also re-sliced from the new KREA 2 ULTRA **V2** workflow.

## Queue/render wedge watchdog

The nastiest failure mode: a wedged high-res sampler step lets the agent stack jobs
behind a zombie render it can't see or kill. Three best-effort, fail-safe guards close
it (if the watchdog WebSocket never opens, nothing changes):

* **`panel_run` backpressure** — appends a QUEUE WARNING to the tool result when a
  render is already running, so the agent stops stacking behind it.
* **Passive `QueueMonitor`** — a best-effort WS to ComfyUI tracks the running prompt /
  node / progress; a stuck step (the same progress value re-emitted) trips a one-line
  STALL/BACKLOG note prepended to the agent's next turn, deduped per episode. Threshold
  via `COMFYUI_MCP_STALL_S` (default 180s, clamped 15–3600), and **live-tunable** from
  the panel's **Render stall warning (seconds)** setting via a `set_config` frame — no
  reconnect needed.
* **`cancel_job` escalation** — interrupt → verify it actually stopped (within
  `COMFYUI_MCP_INTERRUPT_S`, default 30s) → escalate to `/free` → report WEDGED and
  suggest `restart_comfyui` if it still won't die. A new **`clear_pending`** drops all
  pending jobs in the same call.

See [How it works → the queue/render watchdog](../concepts#self-healing-the-queuerender-watchdog).

## `analyze_color` — color stats without a vision round-trip

A new **`analyze_color`** tool returns palette / contrast / color statistics for a
generated image — dominant colors, average + luminance stats, contrast checks — so the
agent can reason about an image's color without spending a vision turn on it.

## Custom-node installs: registry-first, with a hardened clone fallback

`install_custom_node` / `apply_manifest` used to pass the full git URL as the Manager's
`id`, but ComfyUI-Manager keys its node DB by repo-name / CNR id (never a URL), so the
lookup matched nothing and the queue reported "done" without cloning — a **false
success**. Installs are now:

* **Registry-first** — git URLs are resolved the way the Manager UI does (repo name,
  `nightly` version, `dev` channel, `cache` mode).
* **Verified** against `/v2/customnode/installed` (which reflects on-disk packs, so a
  freshly-installed node is seen before a reboot); a non-URL id that doesn't install is
  reported as a hard failure, not a false success. `apply_manifest` applies the same
  verification.
* **Clone fallback** — only when the Manager genuinely can't resolve the pack (an
  unregistered repo) does it fall back to a direct `git clone` (+ best-effort
  `pip install -r requirements.txt` via the ComfyUI venv). The fallback now **fails fast**
  instead of hanging for minutes on a credential prompt — git network ops run
  non-interactively (`GIT_TERMINAL_PROMPT=0` + `GIT_ASKPASS`) with a tightened 180s clone
  timeout.

The install path is also **hardened** against git-option injection (a URL starting with
`-`) and path traversal (a repo name resolving outside `custom_nodes`): the URL is
validated up front, every on-disk use is containment-checked, and `git clone`/`checkout`
use `--end-of-options`. Separately, `update_all` now applies its `mode` (the Manager
reads those params from the query string, where they're now sent).

## Output discovery + correctness fixes

* **`list_output_images` now recurses into subfolders** and lists **videos**, not just
  images. SaveVideo / VHS write to paths like `output/video/clip_00001.mp4`; those used
  to look "not found" on a flat scan. Each result carries its `subfolder` and a
  `kind: "image" | "video"`, so you can pass `{ filename, subfolder }` straight to
  `stage_output_as_input` / `get_image`. Verify a video render here, not via `/history`.
* **`get_history` (no `prompt_id`) returns the newest run.** It now selects by ComfyUI's
  monotonic queue number instead of object-iteration order, so it no longer lags one run
  behind — pass a `prompt_id` when naming a just-produced output.
* **"Send now" / interrupt no longer wedges the agent.** The aborted turn's `result`
  event now drives the turn-gate release at the right moment (with a bounded fallback),
  so an interrupt can never stop cold or run the gate ahead.

## Panel: onboarding, rich media, code blocks, stall warning

* **Provider onboarding.** Connect-time readiness detection per provider (CLI on PATH +
  a login on disk; macOS Keychain handled). An onboarding card shows only when neither
  provider is signed in; the panel auto-switches to a ready provider when the saved pick
  isn't usable (the saved preference is kept), and a not-ready row becomes a "set up"
  action. See [Backends → readiness & onboarding](../backends#connect-time-readiness--onboarding).
* **Rich media metadata in agent pushes.** A render's executed-event note (and a
  structured `metadata` field) now carry each output's path, file size, pixel dimensions,
  asset-set grouping ("output K of N from this run" + siblings, or "single output"),
  render duration, and completion time; video storyboards add format + real frame
  count/fps.
* **Code-block tools.** Rendered fenced code blocks get a Copy button + a persisted
  global line-wrap toggle (off by default); inline code gets Copy.
* **Render-stall warning setting** (General; default 180s, range 15–3600) — sent on
  connect and pushed **live**, so changing it applies without a reconnect.
* **Background-tab streaming fix.** Streamed replies now render even when the ComfyUI tab
  is hidden — the reply commit finalizes synchronously in a background tab (and a
  `visibilitychange` handler flushes on hide / resumes the typewriter on return), instead
  of leaving an empty bubble with a stuck cursor that looked like the agent was "stuck
  thinking."

## Krea 2 ULTRA V2 — detail boost + two-pass combo

Aitrepreneur shipped a **V2** of the KREA 2 ULTRA workflow. The krea2 packs are
re-sliced from it, with a new detail-boost node and a new two-pass combo pack.

V2 adds the **`Krea2T-Enhancer`** node (`capitan01R/ComfyUI-Krea2T-Enhancer`), a
MODEL→MODEL patch wired in the model path (Power Lora → Krea2T-Enhancer → sampler).
It ships **active** in all three packs — bypass it to A/B against the un-boosted
result. V2 also **drops** v1's `ConditioningKrea2Rebalance` node entirely.

A new pack, **`krea2-combo`**, runs a two-pass detail boost: a first pass (8 steps,
`er_sde`, denoise 1) → VAE roundtrip → a low-denoise **second pass** (4 steps, `euler`,
denoise **0.3**), with the **krea2 turbo LoRA** @0.2 on both passes; it saves **both**
passes so you can see the boost. The optional **IdeoKrea** LoRA
(`Aitrepreneur/IdeoKrea`) — an Ideogram-4-style style/prompting LoRA — is downloaded but
not wired by default; drop it into the Power Lora Loader's empty slot (start \~0.5–1.0)
for the turbo + IdeoKrea combo.

The three packs:

* **`krea2-txt2img-manual`** — prose prompt + detail boost.
* **`krea2-txt2img-json`** — Ideogram-style JSON / area prompting + detail boost.
* **`krea2-combo`** — two-pass detail-boost combo (turbo LoRA + optional IdeoKrea).

All are sliced from the V2 monolith and statically validated (clean slice +
structural lint); the V2 detail-boost + combo are pending a live render once the
`Krea2T-Enhancer` node and the turbo/IdeoKrea LoRAs are installed.