Skip to content

Media envelope parsing corrections from a live Hermes run

Date: 2026-08-30

Scope: packaged Hermes media skill documentation (1.0.51.0.6), CHANGELOG; no source, gateway, workflow, or deployment change

Deployment state: not-deployed; no route, model assignment, promotion, container, controller, or fleet state changed

What happened

A production media run on the fleet produced one real image job end to end: discovery through the packaged skill, submission with a caller-generated idempotency key, queued → completed, artifact retrieval with a verified SHA-256 digest, and an independent visual review that confirmed prompt adherence (a fish suspended mid-air against a clouded sky). The run exercised the standard 768×768 four-step profile of the qualified image workflow, and the gateway routed the job to the managed worker without any approval gate.

While driving the same gateway with a direct stdio JSON-RPC client alongside the packaged skill, three response-contract facts surfaced that the skill did not yet document. None is a server defect: the server behaved per its tests. They are client-parsing traps that cost one lost job id and one lost image before being diagnosed.

Corrections folded into the skill

Observation Correction (skill 1.0.6, new steps 11–14)
media_workflow_run and media_job_status envelopes nest the job object under data.job / {"job": ...}; artifacts live at job.artifacts[] Parse the nested shapes; do not assume flat job_id/artifacts keys
Native image content arrives as a second MCP content block (content[1], type image) beside the JSON text block; a client reading only text blocks silently drops the generated image Enumerate the whole content[] array and render every image block
MCP resources/read answers Method not found even though artifact metadata advertises a resource path Retrieve bytes through the authenticated artifact origin and verify sha256 before presenting

The corrections are transport-shaped, not client-specific: the same nesting and two-block content array appear through Hermes' MCP client and a raw stdio JSON-RPC client alike, because both meet the same gateway protocol layer.

Measured run

Item Value
Workflow qualified image workflow, standard profile (768×768, four steps)
Job latency 10.53 s end to end; 0.115 s generation; 10.296 s queue
Artifact 666,126-byte PNG; digest verified byte-for-byte against metadata
Visual review prompt-adherent; independent vision pass confirmed subject and scene
Skill versions production ran 1.0.5; corrections recorded in 1.0.6

Caveats

  • One image job is not a quality corpus; the existing bounded perceptual review evidence for the image workflow stands unchanged.
  • The findings above describe the current protocol contract as implemented and tested; if the envelope shape changes in a later release, this section and the skill steps must move with it (the skill test suite pins the skill version and the synchronized copies).