Hermes image-quality profiles and production enablement¶

Date: 2026-08-28

Scope: production Anvil media gateway on Primary Node, managed on-demand ComfyUI worker on one RTX 5090 in Fakoli Mid Mod, and the real Hermes client on Companion Node; c1 text-to-image execution

Decision: the exact FLUX.2 Klein workflow is production-enabled through three fixed quality profiles; available=true, promoted=false. The Wan2.2 video workflow remains unavailable with quality_failed; no fallback exists.

Result card¶

Real Hermes produced eight independently reviewed FLUX.2 images through the authenticated MCP gateway, including all three fixed quality profiles and five cold approval/resume paths. Six passed strict bounded prompt-adherence review; two additional draft samples completed technically but exposed material-style and exact-count limitations. Video remains fail-closed.

Setup	Qualified value
Model	FLUX.2 Klein 4B FP8 `5b4408e5`; Qwen3 4B encoder and FLUX.2 VAE `5f526678`; immutable graph `991b63b8...2e4f`
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 `0.36.0`, executable fix-forward line through `5ea1edc`; ComfyUI v0.33.4 at `7a131a3a`; CUDA 13.0; PyTorch 2.13.0+cu130
Managed recipe	`image.flux2-klein-4b-fp8-v1`; lifecycle-owned `serves.comfyui.toml`; worker absent when idle
Measurement path	real Hermes natural-language turn → exact eight-tool skill → authenticated MCP gateway → bounded controller approval → managed ComfyUI → authenticated artifact origin
Profiles	`draft` 512×512, `standard` 768×768, `high` 1024×1024; four steps each; prompt and optional seed are caller-controlled; c1
Evidence	`functional`, routed/client acceptance, bounded `capacity`, bounded independent `quality`; production cutover and teardown
Decision	image workflow available at the three exact profiles; arbitrary graph/size/model input rejected; video remains `quality_failed`

Headline measurement	Local result	Conditions
Independent image review	6/8 strict pass	one warm image per profile plus five cold `draft` samples; six prompts; every artifact technically valid
Warm gateway E2E	1.352 / 1.242 / 1.650 s	`draft` / `standard` / `high`; one natural Hermes request per profile
Warm generation phase	0.101 / 0.128 / 0.189 s	gateway phase telemetry for the same three requests; not caller wall time
Warm Hermes caller wall	38.354 / 24.420 / 41.962 s	complete agent turns, including model reasoning and tool use; one request per profile
Cold lifecycle E2E	2,045.626 s	one `draft`; includes repeated fix-forward failures, approval/build/resume, queue, and generation; not steady-state latency
Cold retry caller wall	47.788 s	identical Hermes request after approval/start completed; 363,074-byte PNG
Final hardened cold E2E	908.936 s	`b46f6ce` plus skill `1.0.4`; server-issued exact resume bundle, 335.751 s approval wait, 898.152 s accepted-to-queued, 10.696 s queue, 0.087 s generation
Final post-approval Hermes wall	43.091 s	exact `created=false` same-job resume, terminal polling, native image inspection, and English completion

Why it matters: Hermes can now satisfy an ordinary image request without learning ComfyUI graphs or infrastructure details. The user selects a small quality vocabulary, while the gateway retains workflow identity, lifecycle authority, authentication, idempotency, artifact ownership, and latency telemetry.

Important caveat: Eight images over six prompts establish only a bounded production smoke. Two strict draft reviews failed: an intended origami whale read as a metallic fish/whale, and an intended glass lighthouse rendered as an opaque tower with two rather than exactly three birds. The result does not establish broad image quality, text rendering, hands, counting, material fidelity, or monotonic visual improvement between tiers. The first cold E2E number intentionally includes multiple defects found and repaired during the same run and must not be presented as normal generation latency.

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

Product contract¶

The production gateway exposes the same eight bounded media tools through MCP and A2A. Hermes receives only those tools and an environment-reference-only configuration authenticated with the router credential, never either controller credential. Its media skill completes the generation turn, reports an approval requirement with the server-issued complete resume bundle when the worker is cold, and retries the exact workflow, version, quality profile, parameters, and idempotency key after approval. It requires created=false and the original job identity, continues through nonterminal states, and selects reply language from the current request rather than tool output or session history. It does not diagnose infrastructure, submit a raw ComfyUI graph, choose a model filename, or silently substitute another workflow.

The final deployed skill is 1.0.5. It makes the mandatory fail-closed cancellation for a missing or inexact server resume bundle an explicit exception to the normal user-request-only cancellation rule. It also names the five legal replay inputs separately from the job and approval correlation fields, and preserves an explicit empty profile for workflows that declare no quality profiles. These client and generic gateway hardenings do not relabel the measured image requests below, which used skill 1.0.4 on the b46f6ce runtime.

Image requests accept a prompt, an optional seed, and one of three profiles:

Profile	Width × height	Steps	Intended use
`draft`	512×512	4	fastest bounded preview
`standard`	768×768	4	default balanced result
`high`	1024×1024	4	largest qualified output

The profile, dimensions, graph, runtime, model assets, node set, and output format are server-owned. A direct width override failed with quality_profile_parameter_override before execution. A video request failed with quality_failed and did not fall back to image or another generator.

Native MCP image results are bounded to 6 MiB and the framed bridge to 10 MiB. The largest observed result was 1,495,385 bytes. Video artifacts are resource-only even after a future workflow passes its independent quality gate.

Method¶

Rotate the production gateway credential, distribute only environment references, prove the current token returns 200, and prove retired and incorrect tokens return 401 before dispatch.
Deploy the image-enabled router and the bounded lifecycle/resource controllers without replacing or restarting the unrelated Qwen service.
Verify modern MCP discovery exposes exactly eight media tools and that the Hermes profile has the exact matching allowlist with fallback disabled.
Ask real Hermes for the same cinematic observatory scene once at each quality profile. Retain caller wall time, gateway phase telemetry, artifact size and digest, and an independent visual disposition.
Tear the worker down, submit a new draft request through Hermes, approve its declared lifecycle action, build/start the managed worker, and replay the exact request and idempotency key to completion.
Confirm teardown through the managed controller, then verify the worker is absent, the reservation has no owner or committed memory, and the resource controller has no orphan descendants.
Exercise negative controls for authentication, arbitrary dimensions, video quality state, duplicate approval, and idempotent replay.
Deploy the final executable revision, reconcile the Hermes skill, and run a second cold natural-language request through awaiting approval, exact same-job reattachment, terminal polling, native image return, teardown, and independent strict prompt-adherence review.

Warm profile results¶

All three warm requests used the prompt “A cinematic glass observatory on a mossy cliff, a copper telescope visible inside, blue hour, warm interior lights.” The images were distinct, decodable PNGs at their exact declared dimensions.

Profile	Dimensions	Bytes	SHA-256	Gateway E2E	Generation	Hermes wall	Review
`draft`	512×512	321,710	`67867fe5...cd56`	1.352 s	0.101 s	38.354 s	pass; complete scene, minor simplified geometry
`standard`	768×768	766,691	`6d88d55d...d063`	1.242 s	0.128 s	24.420 s	pass; strong adherence, minor roof/railing perspective
`high`	1024×1024	1,495,385	`7c3874ca...2a7e`	1.650 s	0.189 s	41.962 s	pass; detailed and coherent, mildly ambiguous glass extension

These are single observations. The non-monotonic caller wall times reflect a complete agent turn, not just GPU work, and do not rank the profiles. The very short generation phases are warm/cached gateway observations and do not replace the earlier cold direct qualification.

Cold approval, build, resume, and teardown¶

The cold request asked for “A small copper observatory dome emerging from a foggy pine forest at dawn” at draft with a fixed seed. It entered the durable approval state without starting the worker. The approved managed start built the pinned worker, passed its storage and /system_stats checks, and gave the job lifecycle ownership. Hermes then replayed the same request and idempotency key; the existing job advanced to completion instead of creating another execution.

The 363,074-byte, 512×512 PNG had SHA-256 ebb6f351c07c2109bea995ffb7dd03e48c40667b53d2b5e0aa8bc1b056c1b3ce. Independent review passed it: the copper-toned dome, layered pine forest, fog, and dawn light were clear and coherent. The dome was slightly pink from the sunrise and somewhat pavilion-like without an obvious telescope slit; distant trees repeated, but no major rendering defect was present.

Recorded job phases were 0.061 seconds submission, 2,035.526 seconds from acceptance to queued, 10.0 seconds queued, 0.1 seconds generation, and 2,045.626 seconds E2E. The approval-wait instrument recorded 151.78 seconds of the final approval segment. The much larger accepted-to-queued interval spans the complete fix-forward sequence below, so it is an incident/lifecycle measurement rather than a normal cold-start service objective.

Managed teardown then removed the worker. The final resource view showed the worker absent, zero media memory committed, 28,511 MiB free inside the media envelope, no lifecycle owner, and no orphan build/start process.

Final exact-revision regressions¶

Four later cold draft requests closed the immutable deployment and client behavior gates. A 305,978-byte sample completed in 719.3 seconds E2E and passed independent review. A 373,257-byte silver-whale sample completed in 695.344 seconds E2E but failed strict review because the subject read as a metallic fish/whale rather than folded origami, despite the salt flat, sunrise, and tiny blue balloon adhering. Its phase telemetry was 93.631 seconds approval wait, 684.225 seconds accepted-to-queued, 11.012 seconds queue, and 0.108 seconds generation.

The final bb798b0 regression used a complete resume bundle for a red glass lighthouse request. Hermes returned created=false for the original job, continued from queued to terminal in the same turn, inspected the native PNG, replied in English, and reported every latency field. The 337,286-byte PNG had SHA-256 4eb0d995...207c; recorded phases were 0.065 seconds submission, 169.714 seconds approval wait, 642.609 seconds accepted-to-queued, 9.888 seconds queue, 0.110 seconds generation, and 652.607 seconds E2E. Strict visual review failed because the tower was predominantly opaque rather than glass and only two of the requested three white birds were visible. This is retained negative quality evidence, not relabeled as a pass. Managed teardown again restored the worker-absent, zero-commit, 28,511-MiB-free baseline.

The final measured b46f6ce regression hardened that boundary further. The router assembled the complete seven-field resume bundle from validated current request fields plus durable job, selected-profile, and approval identity; skill 1.0.4 copied it literally instead of reconstructing it. After the approved 475.983-second managed prepare, a fresh Hermes one-shot reattached with created=false, polled queued to completed, inspected one native image, and replied in English in 43.091 seconds. The 250,151-byte, 512×512 PNG had SHA-256 78c147bb...0076; authenticated HTTP bytes, the MCP image block, and artifact metadata matched exactly. Recorded phases were 0.062 seconds submission, 335.751 seconds approval wait, 898.152 seconds accepted-to-queued, 10.696 seconds queue, 0.087 seconds generation, and 908.936 seconds E2E. Independent review passed: one cobalt-blue ceramic sphere was centered on a matte white pedestal against the requested plain pale-gray studio background. Teardown took 12.563 seconds and restored the worker-absent, zero-commit, 28,511-MiB-free baseline.

Defects fixed forward during the production gate¶

The full production gate exposed twenty-three actionable defects: seventeen on the live path, four in post-run gateway/skill/release review, and two in evidence review. Each was repaired in the product or evidence surface with regression coverage before the gate closed. Jobs whose replies could not preserve an exact bundle were canceled; later fresh requests proved the corrected path:

The parent controller had a 30-second child timeout while the child allowed 300 seconds. Child timeout plus a bounded margin is now propagated.
The controller image did not include Docker Buildx. The pinned plugin is now part of the managed controller image.
A resource-scoped serves up attempted to co-locate the router. Controller starts now use the explicit no-router path.
A failed approved start stranded the job without a fresh approval path. The same job/key may now mint a new one-use approval after failure, while an in-flight start cannot mint a competing approval.
Buildx tried to write its client state under a read-only home. The controller uses writable tmpfs-backed Docker configuration state.
A five-minute cold-build timeout was too short and killed only the direct child, leaving Compose/Buildx descendants. The bounded timeout is now 1,800 seconds, and timeout handling terminates the whole process tree on POSIX and Windows.
A successful remote apply nested applied=true, but the lifecycle receipt reported false. Apply status is now normalized from the controller envelope.
Fleet status probed runtime-relative upstreams from the command host and falsely reported an outage. Live probes now execute from the installed router runtime, report a safe perspective/kind, and retain a typed mismatch diagnostic for configured-file evidence.
The managed Hermes bridge defaulted to the controller credential even though its endpoint is the router MCP gateway. The bridge now defaults to the least-authority router credential; real discovery returned exactly eight tools and a second reconcile had no drift.
The cold approval reply omitted a durable resume bundle, while a resumed one-shot could select stale session evidence or stop after saying it would poll. Skill 1.0.1 now requires the complete bundle, rejects any job or created mismatch, forbids session reconstruction, and keeps nonterminal states inside the tool loop.
The first 1.0.1 approval-boundary reply switched an English request to Chinese. Skill 1.0.2 now derives language only from the current request; the exact same-job completion reply was English.
One generic controller credential crossed both the router-to-lifecycle and lifecycle-to-resource edges. Dedicated, independently rotated credentials now protect each edge, and every cross-use attempt returns 401.
Multi-cycle approval telemetry selected the first completed approval interval even though the public contract describes the final completed interval. Latency now selects the latest completed cycle, with a two-cycle regression test.
The artifact tool catalog said inspection returned no bytes while eligible bounded images already returned native MCP image content. The description now states the image/video and size boundary exactly.
Skill 1.0.2 said a cold resume bundle was preserved but omitted the literal mapping. The affected job was canceled, and the contract now refuses approval unless every field is emitted.
Skill 1.0.3 emitted an incomplete mapping: it omitted the job identity and abbreviated the idempotency key. That job was also canceled rather than resumed from reconstructed state.
Revision b46f6ce makes the gateway assemble and return the exact resumeBundle from validated request fields plus durable job/profile/ approval identity; skill 1.0.4 must copy that object literally and cancels the job if it cannot. A fresh cold live regression passed the complete same-job path.
Adversarial review found that skill 1.0.4 also said cancellation was allowed only on a user request, contradicting its mandatory fail-closed cancellation for a missing or inexact resume bundle. Skill 1.0.5 makes that safety action an explicit exception.
A valid profileless workflow could reserve a cold job and then lose its resume boundary because an empty profile was treated as invalid. Revision 5ea1edc preserves the validated empty profile, accepts it in the MCP schema, and proves cold replay and cancellation through real operations.
Skill 1.0.4 said to replay the complete seven-field bundle even though job_id and approval_transaction_id are not submission inputs. Skill 1.0.5 names the exact five submission fields and reserves the other two for same-job and operator-approval correlation; the protocol regression proves full-bundle rejection and five-field replay.
The evidence incorrectly said every fix preceded completion of the same request even though two malformed-bundle jobs were deliberately canceled. It now distinguishes those canceled jobs from the fresh passing requests.
The evidence called the complete resume payload persisted even though raw request parameters are not stored. It now accurately separates validated current request fields from durable job/profile/approval identity.
The Windows CI checkout converted both Hermes skill copies to CRLF while the fail-closed packaged-skill loader validated an LF frontmatter prefix. The repository now pins both skill trees to LF in .gitattributes, and a hermetic regression proves LF stability plus packaged/example byte identity.

The final behavior is a fix-forward result, not a waiver of those failures.

Safety and negative controls¶

Control	Result
Router credential	authenticated MCP discovery passed; both media-controller credentials and the generic controller credential returned 401
Lifecycle credential	authenticated health and prepare/teardown passed; resource and generic controller credentials returned 401
Resource credential	authenticated health and managed serve operations passed; lifecycle and generic controller credentials returned 401
Hermes tool scope	exactly eight media tools; environment references only
Caller width override	rejected before execution
Video request	`quality_failed`; no fallback or execution
Cold worker before approval	remained absent
Approval replay	one durable job and one owner; no competing start
Exact Hermes resume	server-issued complete bundle; `created=false`; same job; terminal polling in-turn
Hermes reply language	English request produced English terminal response under measured skill `1.0.4`; final skill `1.0.5` retains the rule
Final router/skill smoke	`5ea1edc` router healthy; skill `1.0.5`; exact seven-field cold reply; explicit Hermes cancellation reached `canceled`
Native artifact return	MCP image content, authenticated resource bytes, size, digest, and 512×512 PNG header matched exactly
Router fleet probe	installed router perspective; 9/9 configured routes returned HTTP 200
Post-run teardown	worker absent; reservation reclaimed; no orphan descendants
Existing Qwen service	retained throughout the router/controller cutover

Release security boundary¶

The final public candidate passed the semantic scanner over 2,007 tracked files with zero findings, its non-ignored untracked surface had zero findings, and the repository-pinned Gitleaks image reported zero current-candidate signatures. The private operator candidate also had zero Gitleaks signatures. Its semantic scan still reports the real topology identities and operator paths that the private repository exists to retain; those expected private-policy matches did not cross into the public candidate.

A separate verified scan of all public commits reachable at the release gate is deliberately reported as findings remain, not clean. It found 21 historical signatures: 14 synthetic fixture or non-secret environment/provenance matches, six derivative copies in historical generated search indexes, and one credential-shaped deviceToken field in a historical temporary-loopback OpenClaw smoke artifact. That field is redacted in the current artifact, its historical value is absent from both current repositories, and it is unrelated to the production router credential rotated for this deployment. No Git-history rewrite or force-push was performed; that destructive operation requires separate authorization and coordination.

Evidence boundary¶

This finding establishes that the exact production image contract works from real Hermes through MCP, the router gateway, lifecycle approval, an on-demand ComfyUI worker, and authenticated artifact return. It also establishes exact profile dimensions, telemetry fields, fail-closed overrides, idempotent cold resume, and teardown on one RTX 5090.

It does not establish concurrency above one, broad or comparative image quality, exact counting or material fidelity, monotonic improvement by profile, arbitrary aspect ratios, a cold-start latency objective, or a usable video generator. The two retained strict draft failures make those limitations explicit. The Wan2.2 workflow remains quality_failed and unavailable until a new immutable workflow version passes a separate temporal and perceptual corpus. Six strict passes and two retained failures remain a small bounded sample, not a general quality score.