Hermes Agent
Hermes video proposal makes automatic analysis an explicit routing choice
An open Hermes proposal gives video its own auxiliary-model configuration and optional automatic descriptions in chat gateways. The inspected code makes the default auto mode conditional on an explicitly configured video provider, model or endpoint, rather than forwarding every attachment automatically.
A Hermes contributor's screen-recording workflow exposed a gap between receiving an attachment and understanding it: video arrived as a local path, leaving the agent to decide whether to invoke analysis. The proposed repair gives video a first-class auxiliary task and a gateway enrichment path analogous to image descriptions. The important product decision is its default routing boundary. In the inspected implementation, auto mode enriches a video only when the video auxiliary configuration names a provider, model or base URL; otherwise the path remains available for an intentional tool call.
The facts
- PR #112464 changes video_analyze from the vision auxiliary task to a distinct video task. - Its configuration path reads auxiliary.video rather than silently resolving every video through auxiliary.vision. - The gateway's auto mode checks for an explicit video provider, model or base_url before automatic analysis. - Text mode always requests automatic analysis; off mode leaves the path note while keeping the analysis tool available. - Successful enrichment prepends a chronological description and retains the local path for closer examination; failures leave a retry note rather than fabricated observations. - The proposed documentation distinguishes gateway enrichment from WebUI delivery, where file paths remain available and the agent chooses whether to call the tool.
Why it matters
A screen recording is often the clearest report of a UI failure, but it can also contain credentials, customer data and transient on-screen text. Automatically describing it changes both convenience and the data path. A separate video configuration lets an operator choose a capable endpoint without accidentally assuming that image routing handles the same modality. The default gate matters more than the feature label: being able to analyze a video and automatically sending every attached video for analysis are different promises with different privacy implications.
Current
Inspected on 2026-09-16. The Hermes Agent stable-channel baseline is v2026.9.14 published 2026-09-14T16:04:14Z. The primary source was open (not merged) when captured. Mainline merges, open proposals and packaged releases are distinct availability states; the release baseline does not establish that a proposal has shipped.
Evidence
Primary evidence: NousResearch/hermes-agent PR #112464 (https://github.com/NousResearch/hermes-agent/pull/112464). Supporting context: Hermes video proposal — routing implementation diff (https://api.github.com/repos/NousResearch/hermes-agent/pulls/112464/files?per_page=100); Nous Research — Hermes vision documentation (https://hermes-agent.nousresearch.com/docs/user-guide/features/vision/). The source bodies and linked context were inspected. Test results quoted from contributors remain attributed reports, not independently reproduced experiments.
Operator take
Review the proposed modes against the intended use of each chat channel. On a test branch, first confirm the unconfigured default leaves only the path, then verify an explicitly selected endpoint receives the intended clip and failure handling preserves the original question. Use synthetic footage rather than a real customer screen recording. The source diff is decisive for the auto condition; broad prose about sharing the vision fallback chain must not be read as unconditional gateway forwarding. Before rollout, test provider compatibility with native video content and establish retention expectations for that endpoint.
Caveat
This is an open proposal, not a shipped-default announcement. The implementation and documentation were inspected, but no video was uploaded and no model output was independently evaluated. The source's fallback descriptions concern client resolution; automatic gateway enrichment still has the explicit configuration gate described above.
This is an open proposal, not a shipped-default announcement. The implementation and documentation were inspected, but no video was uploaded and no model output was independently evaluated. The source's fallback descriptions concern client resolution; automatic gateway enrichment still has the explicit configuration gate described above.