Media gateway live validation through Hermes, MCP, and A2A

Date: 2026-08-28

Scope: exact merged Anvil Serving wheel, isolated authenticated gateway on Primary Node, managed ComfyUI worker on one RTX 5090, disposable Hermes profile on Companion Node, c1 image/video execution

Decision: exact-build live acceptance passed for the gateway and image path; the exact Wan2.2 workflow failed bounded perceptual review; both workflows remain unavailable and no-promotion

Result card

The exact merged media gateway completed real Hermes image and video jobs on one RTX 5090, passed MCP/A2A/artifact/lifecycle controls, produced two prompt-adherent FLUX.2 images, and produced one decodable but visibly failed Wan2.2 video.

Setup	Qualified value
Models	FLUX.2 Klein 4B FP8 image workflow; Wan2.2 TI2V 5B video workflow; exact revisions remain in the bundle lock
Hardware	one NVIDIA RTX 5090, 32,607 MiB, sm_120; Primary Node gateway, Fakoli Mid Mod resource owner, Companion Node Hermes client
Runtime	Anvil Serving 0.35.1 from merge 19320de6; ComfyUI v0.33.4 at 7a131a3a; CUDA 13.0; PyTorch 2.13.0+cu130
Recipe	managed serves.comfyui.toml; immutable workflow graphs 991b63b8...2e4f and bd12b2de...9572
Measurement path	authenticated cross-host gateway; modern MCP, bundled stdio bridge, real Hermes, A2A 1.0, and authenticated artifact origin
Contract	image 512×512/four steps; video 512×288/17 frames/eight steps/16 fps; c1; cold approval required
Evidence	functional, routed/client acceptance, bounded independent quality; teardown complete
Decision	image path live-accepted but not enabled; video workflow quality-failed; no production cutover or promotion

Headline measurement	Local result	Conditions
Prompt-adherent FLUX.2 artifacts	2/2	one gateway approval-path image and one real-Hermes image; independent visual review
Real-Hermes Wan2.2 transport/decode	1/1	117,738-byte H.264 MP4; 512×288; 17 frames; 16 fps; 1.0625 seconds
Real-Hermes Wan2.2 perceptual result	fail	severe spatial/frame artifacts; requested subject and clear-sky scene not reliably present
Cold-backend negative control	0 new jobs	Hermes stopped after backend_unavailable; worker remained absent
Final reservation state	0 MiB committed	worker absent; 28,511 MiB free inside the protected media envelope

Why it matters: An agent can discover and execute the bounded image/video API through the same Anvil gateway using the packaged Hermes skill, while the gateway still fails closed when the backend is cold and preserves one durable job across MCP and A2A.

Important caveat: This was an isolated exact-build deployment, not a production-router cutover. Two image samples and one very short video are not a broad quality corpus, and the failed video contact sheet cannot assess temporal smoothness.

Evidence manifest: README.md · Publication summary: publication-summary.md

Outcome and decision

The exact wheel from merge 19320de6b4fa83814ebffc0ac3b6a6ce1234ae7a was installed into isolated gateway and lifecycle environments on Primary Node and the resource controller on Fakoli Mid Mod. Companion Node used the matching Anvil Serving package and its bundled MCP bridge through a disposable Hermes profile. The authenticated gateway exposed the expected eight media tools, modern MCP 2026-07-28, A2A 1.0, Agent Card, health, and artifact routes. Wrong-token access returned 401 before dispatch.

The live path passed:

• cold-job persistence and explicit approval before the managed worker start;
• one-use lifecycle approval and identical idempotent replay;
• real Hermes image and video submission through only the packaged media skill;
• full artifact download, byte-range download, missing-artifact 404, signature, declared-size, and SHA-256 checks;
• cross-protocol A2A replay of the Hermes video job without a second execution;
• A2A task projection and fail-closed cancellation of an already completed job;
• cold-backend validation that returned backend_unavailable and created no job; and
• managed teardown to an absent worker, idle GPU baseline, zero reservation commitment, and no owner.

The independent adversarial visual review passed both FLUX.2 images for prompt adherence. It failed the Wan2.2 contact sheet because the requested kite and minimal clear-sky scene were replaced by severe smearing, chromatic separation, repeated distorted structures, and a strong horizontal streak. The exact video workflow is therefore not a promotion candidate in its current form. The image workflow also remains unavailable: this bounded review does not authorize a production cutover or route exposure.

Exact configuration

The deployed wheel was anvil_serving-0.35.1-py3-none-any.whl, 1,395,378 bytes, SHA-256 43643570eae62510f70cadb82895543347a0a729df687025bf355fb380407a73. It includes the controller idempotency fix merged at 7920752d and the cold-proxied-backend classification fix merged at 19320de6.

The worker retained the pinned configuration from the prior direct qualification:

• ComfyUI v0.33.4 at 7a131a3afadc8200120f67f9236311a2c48b7445;
• FLUX.2 graph digest 991b63b8c61ff4322d72b8ae81ef43656f4905ddf4b0709c1989b84cfb8f2e4f;
• Wan2.2 graph digest bd12b2de2a33bbedc91d7ad6120714f3c3adbd174694ac74d6f0213ecef9572e;
• immutable model and custom-node revisions from the packaged bundle lock; and
• one 32,607 MiB RTX 5090 with a 4,096 MiB protected reserve and a 28,511 MiB on-demand media envelope.

No raw ComfyUI graph, node ID, model filename, backend path, alternate generator, or fallback was caller-selectable.

Method

1. Build and hash the wheel from the exact merged source revision, then install that artifact into isolated gateway, lifecycle, and resource-controller environments.
2. Start authenticated candidate services without replacing the production router or controller. Verify health, auth-before-dispatch, Agent Card, MCP negotiation, and the eight-tool media catalog.
3. Submit an image while ComfyUI is cold. Verify a durable awaiting_approval job and an exact media_worker_prepare action, execute the user-authorized action through the lifecycle controller, replay it with the same idempotency key, and then complete the job.
4. Configure a disposable Hermes profile with environment-variable credential references and the exact media-tool allowlist. Generate one image and one video through the packaged skill with fallback disabled.
5. Replay the exact video request through A2A, poll the projected task, and attempt cancellation after completion.
6. Retrieve every artifact with a separate standard-library client and verify full, range, missing, signature, size, and digest behavior. Decode the video independently with ffprobe. Have the separately administered adversarial reviewer inspect both images and a video contact sheet.
7. Tear the worker down through the managed lifecycle action. With the backend cold, ask Hermes to validate another image request and require it to stop before media_workflow_run. Confirm the durable job count is unchanged.
8. Remove the disposable Hermes profile and isolated candidate runtimes, then verify candidate ports are closed and the original Hermes default has no MCP server configured.

Results

Artifacts and independent review

Path	Artifact	SHA-256	Independent disposition
Cold approval image	248,804-byte PNG	2d7628b8...99f4c	pass; glossy red ceramic robot, one blue flower, watering action, and cream background were clear; minor nozzle-source ambiguity
Real-Hermes image	397,383-byte PNG	226cf4c2...b2ec6	pass with caveats; brass observatory robot, teal lantern, stars, and midnight-blue sky were clear; small signature-like mark and slightly ambiguous lantern geometry
Real-Hermes video	117,738-byte MP4	083d8a54...c370f	functional/decode pass, perceptual fail; subject and clean sky were not reliably identifiable and severe spatial artifacts dominated the frames

The video decoded as H.264 at 512×288, 17 frames, 16 fps, and 1.0625 seconds. The contact-sheet review cannot establish left-to-right motion, temporal smoothness, or a static camera; those remain unassessed. The visible spatial failure is sufficient to block this exact workflow version.

Protocol and lifecycle controls

Control	Result
Modern MCP / stdio bridge	pass; protocol 2026-07-28, eight scoped tools, real Hermes calls
A2A	pass; exact video request replayed the MCP-created job; task projection matched; completed-task cancellation returned TASK_NOT_CANCELABLE
Artifact origin	pass; full 200, range 206 with exact content range, missing 404, signatures, sizes, and hashes
Cold approval	pass; worker stayed absent before explicit approval; identical approval replay did not start a second worker
Cold backend	pass; backend_unavailable, no run call, durable job count stayed at three
Teardown	pass; worker absent, GPU 448/32,607 MiB, reservation committed 0 MiB/free 28,511 MiB, no owner
Client restoration	pass; disposable profile/alias removed, original default active, no default MCP server

Failures and caveats

• The exact Wan2.2 workflow failed bounded prompt adherence and spatial quality. Longer clips, other prompts, and temporal scoring were not attempted after that failure.
• A legacy Windows PowerShell artifact client stalled after a partial transfer. Gateway logs showed client connection resets while the service remained healthy; a separate standard-library client immediately completed the same authenticated transfer and all framing/hash checks. This is retained as a client-harness caveat, not hidden as a gateway pass.
• An initial administrative attempt showed that setting HERMES_PROFILE alone does not scope Hermes configuration. The transient default MCP entry was immediately removed and verified absent before testing resumed with explicit profile selection. Final client state matches the pre-test default.
• The validation did not rebuild or replace the production router/controller. Exact wheel parity and live behavior are proven only for the isolated candidate deployment.
• Generated binaries were transient validation artifacts. The public packet retains sizes, hashes, decode metadata, and the independent review record, but not the binary image/video files themselves.

What to test next

1. Replace or revise the Wan2.2 workflow as a new immutable version, then run a multi-prompt temporal and prompt-adherence corpus before another live gate.
2. Expand FLUX.2 review beyond two samples, including text rendering, hands, multi-object counting, negative prompts if supported by a new descriptor, and repeatability across seeds.
3. If image enablement is desired, perform a reviewed production router/controller cutover with exact endpoint identity, rollback, and a fresh real-Hermes smoke. Do not couple that image decision to the failed video candidate.
4. Add a product-owned cross-platform artifact download probe so legacy PowerShell client behavior cannot masquerade as a gateway framing defect.

Evidence boundary

This finding proves that the exact merged package can run the designed cross-host media control plane and real Hermes client path, maintain cross-protocol idempotency, serve authenticated artifacts correctly, fail closed while cold, and restore the managed resource state. It provides two bounded positive FLUX.2 perceptual samples and one bounded negative Wan2.2 sample.

It does not prove production-router deployment, broad image quality, temporal video quality, concurrency above one, long-video stability, or authorization to expose either workflow. Both public descriptors remain unavailable and no model, workflow, route, or client default was promoted.