chaoliu615/dsh-agnes

dsh-agnes is a native DSH (DeepSeek Harness) Cordis plugin distributed as a standard bundle, and the repository is named dsh-agnes with a dsh.bundle manifest plus a cordis.patch.yml configuration layer that targets the DSH runtime exclusively. It registers two tools inside DSH sessions: agnes_image_generate, which calls Agnes at POST /v1/images/generations synchronously and supports text-to-image, image-to-image (reference images) and multi-image composition; and agnes_video_generate, which calls Agnes video endpoints for text-to-video, image-to-video and keyframe animation (≥2 frames, image and keyframes mutually exclusive) using an async task flow that posts to /v1/videos, polls every five seconds with a 900-second total budget, and reports completion, failure or timeout with the video_id. The video tool adapts between two parameter regimes based on the configured model name: V2.0 takes width/height/num_frames/frame_rate and uses presets (81/121/241/441 frames at 24fps); the 2.5/2.5-flash line uses OpenAI Videos-style integer seconds 4–12 with size of 720P/960P/2K (flash locked to 720P) and aspect_ratio matched from a whitelist. Installation is via dsh plugin --profile web add dsh-agnes (npm, recommended, pre-built), dsh plugin --profile web add github:chaoliu615/dsh-agnes, or a packed tarball, and the package depends on DSH-provided peer modules @deepseek-ai/cordis, @deepseek-ai/dsh-credentials, @deepseek-ai/dsh-llm, @deepseek-ai/dsh-settings, @deepseek-ai/dsh-shell, and @deepseek-ai/dsh-tools; Node.js ≥ 20 is required.

The typical workflow starts with configuring AGNES_API_KEY inside DSH credentials — sources include startup environment variables, $DSH_HOME/.credentials.yaml, project or user .env files, or the in-app input on the Settings page — after which the key is forwarded to curl via the process environment so it never appears on the command line, and missing keys raise an explanatory error instead of issuing any request. Once loaded, the plugin contributes an "Agnes" tab to DSH's web settings UI served at /plugins/dsh-agnes/client.js, where users pick models from a dropdown aligned with docs/agnes-ai/ (agnes-image-2.1-flash, agnes-video-v2.0, agnes-video-2.5-flash) and configure defaults such as image size tier (1K–4K), aspect ratio (eight presets with scenario labels), video canvas/clarity, frame count (8n+1 rule, ≤441), frame rate (1–60, presets 24/30), and integer seconds for 2.5 models. Switching the video model automatically reshapes the form: V2.0 exposes width/height/num_frames/frame_rate; 2.5 hides frame rate and replaces frame counts with a 4–12 seconds selector; flash models collapse size to 720P. Persisted changes are atomically written to ~/.dsh/settings.yaml under the agnes: key, validated against a schema with revision gating (e.g., invalid num_frames errors at save time), and every tool call re-reads the resolved values, so updates take effect immediately without restarting the host; old or unknown model identifiers remain usable and surface as "(current)" in the dropdown. The product is aimed at users and teams who want Agnes image and video generation inside DSH without hand-rolling API plumbing or settings UI.

Dependencies and first-run caveats: Node.js ≥ 20 and the DSH-internal peer packages listed above; the npm tarball ships a pre-built lib/ plus cordis.patch.yml, while the GitHub source path triggers pnpm install to run prepare via tsdown. With pnpm ≥ 10 the default behavior is to block build scripts, so a first-time GitHub install must add allowBuilds: dsh-agnes: true to the target profile's pnpm-workspace.yaml; pinning a commit SHA (github:chaoliu615/dsh-agnes#<sha>) is recommended to prevent silent code changes. Practical limits include: V2.0 frames must satisfy 8n+1 and stay ≤441, 2.5-flash is locked to 720P with a five-keyframe ceiling, image requests accept only public HTTPS URLs or data:image/*;base64,… URIs, and very large images returned as b64_json may exceed the capture cap (URLs are preferred). Browser-half registration requires dsh.client to be detected at host start, so users upgrading an older install to add the UI must restart the web profile once; fresh installs need no extra step. The client UI uses only DSH theme tokens (Theme.listTokens aliases plus color-mix), declares color-scheme: light dark, and ships a built-in zh/en copy dictionary. License is MIT, and publishing to npm ships only lib/ and cordis.patch.yml via the files whitelist, with the upstream docs/agnes-ai/ reference copy git-ignored.