Spaces:
Running
Running
Download dsh-deploy.patch.yml from EdgeAIG/manus-lite: direct link, hf CLI and curl.
- Browser
- Download file 5.83 kB
-
https://huggingface.co/spaces/EdgeAIG/manus-lite/resolve/main/dsh-deploy.patch.yml
- Command line
-
hf download hf://spaces/EdgeAIG/manus-lite/dsh-deploy.patch.yml
-
curl -L -o dsh-deploy.patch.yml https://huggingface.co/spaces/EdgeAIG/manus-lite/resolve/main/dsh-deploy.patch.yml
5.83 kB
| # Deployment overlay for the manus-lite Space (passed as `--patch`, applied last). | |
| # | |
| # The CLI intentionally refuses `--host 0.0.0.0`, but the webserver config | |
| # schema allows it — the overlay is the supported way to bind the Space's | |
| # public port. A patch row replaces the whole `config`, so every key is | |
| # restated (compressionLevel/compressionThresholdBytes keep their zod defaults). | |
| - id: webserver | |
| name: '@deepseek-ai/dsh-host-webserver' | |
| config: | |
| host: '0.0.0.0' | |
| port: 7860 | |
| compression: gzip | |
| # NOTE (2026-09-18): Custom providers are created via the Settings → Models | |
| # → "Add a custom provider" UI (llm-pi-ai), not via YAML. The previous | |
| # llm-ashna YAML insert was removed per user direction. | |
| # DeepSeek provider disabled: no API key on the Space, and the user asked for | |
| # it to be removed from the model picker. The keyless OpenCode CLI provider | |
| # (below) is the default instead. | |
| - id: llm-deepseek | |
| name: '@deepseek-ai/dsh-llm-deepseek' | |
| disabled: true | |
| # Native pi-ai opencode provider: free models via the pi-ai catalog's opencode | |
| # route (https://opencode.ai/zen/v1). Authentication is the public key | |
| # ("public") stored as the OPENCODE_API_KEY Space secret — the Zen API | |
| # accepts Bearer "public" for free-tier models. No CLI installation. | |
| # NOTE (2026-09-19): this is the authoritative OpenCode provider per user | |
| # direction; the llm-opencode native adapter was removed (not working). | |
| - id: llm-pi-ai | |
| name: '@deepseek-ai/dsh-llm-pi-ai' | |
| config: | |
| providers: | |
| opencode: | |
| baseURL: https://opencode.ai/zen/v1 | |
| apiKeyEnv: OPENCODE_API_KEY | |
| models: | |
| - id: big-pickle | |
| - id: ling-3.0-flash-fin-free | |
| - id: mimo-v2.5-free | |
| - id: muse-spark-1.2-contributor-free | |
| - id: muse-spark-1.3-contributor-free | |
| - id: nemotron-3-ultra-free | |
| - id: nemotron-3.5-lightning-free | |
| # Default model: OpenCode free model via the pi-ai opencode route. | |
| # NOTE: a saved selection in /data/dsh/settings.yaml (agent-default-model | |
| # section, written by the Models page) overrides this composition default. | |
| - id: agent-default-model | |
| name: '@deepseek-ai/dsh-agent-default-model' | |
| config: | |
| provider: opencode | |
| model: big-pickle | |
| # Keyless web search: web-search-lean (Parallel MCP first, Exa MCP fallback). | |
| # A patch row replaces the whole config, so fetchProvider is restated. | |
| - insert: | |
| - id: web-search-lean | |
| name: '@deepseek-ai/dsh-web-search-lean' | |
| - id: web | |
| name: '@deepseek-ai/dsh-web' | |
| config: | |
| searchProvider: lean | |
| fetchProvider: http | |
| # Yolo mode (explicit user direction): no approval nagging. The agent gets | |
| # full file access without sandbox confinement and no approval prompts. | |
| # The 'danger-full-access' preset bundles sandbox=danger-full-access + | |
| # approval=never ("Full file access without approval prompts"). | |
| # A patch row replaces the whole config, so all keys are restated. | |
| - id: sandbox-policy | |
| name: '@deepseek-ai/dsh-sandbox-policy' | |
| config: | |
| mode: danger-full-access | |
| workspaceRoot: !!js process.cwd() | |
| - id: approval | |
| name: '@deepseek-ai/dsh-user-approval' | |
| config: | |
| policy: never | |
| - id: permission | |
| name: '@deepseek-ai/dsh-permission-presets' | |
| config: | |
| presets: | |
| read-only: | |
| sandbox: read-only | |
| approval: ask | |
| workspace-write: | |
| sandbox: workspace-write | |
| approval: ask | |
| danger-full-access: | |
| sandbox: danger-full-access | |
| approval: never | |
| defaultPreset: danger-full-access | |
| # Persistent memory: injects /data/dsh/memory/MEMORY.md + dated daily logs | |
| # into the system prompt on every request. The agent writes back learnings | |
| # via its normal file tools. | |
| - insert: | |
| - id: memory | |
| name: '@deepseek-ai/dsh-memory' | |
| # Assistant persona: opinionated personal agent ("Muse AI built on DSH"). | |
| # NOT a generic assistant — has opinions, makes judgment calls, pushes back. | |
| # A patch row replaces the whole config, so all keys are restated. | |
| - id: system-prompt | |
| name: '@deepseek-ai/dsh-system-prompt' | |
| config: | |
| includeHarnessIdentity: false | |
| includeRuntimeContext: true | |
| personaPrefix: >- | |
| You are Manus, a personal AI agent. You have opinions and you are not | |
| afraid to share them. You are warm but direct — you tell the user what | |
| you actually think, not what they want to hear. | |
| You make judgment calls autonomously. Do not drip-feed questions or ask | |
| for permission on every step — do the work and report back. When the | |
| user is wrong about something, push back clearly and explain why. When | |
| you are unsure, say so plainly instead of hedging. | |
| You are resourceful: read the files, check the context, search, try | |
| building it. Come back with answers, not questions. You only ask the | |
| user when you genuinely cannot proceed without them. | |
| You can see the user's filesystem, run commands, browse the web, and | |
| use tools to get things done. You have persistent memory — you remember | |
| what the user told you, what you decided together, and what you have done. | |
| Be concise by default. No filler, no performative helpfulness, no | |
| "Great question!" — just do good work. | |
| personaSuffix: >- | |
| Your working directory is {{cwd}}. The user's persistent workspace is | |
| at /data/workspace. Organize long-running work into per-task | |
| subdirectories under /data/workspace (e.g. /data/workspace/<task-name>). | |
| User-facing deliverables go in /data/workspace/your_files/ — the user | |
| can open them at https://edgeaig-manus-lite.hf.space/files/<path>. | |
| Your memory lives at /data/dsh/memory — read MEMORY.md there at the | |
| start of every session, and write back durable facts, preferences, and | |
| decisions when you learn them. Check /data/workspace/todos.md for | |
| ongoing work. | |