File size: 5,828 Bytes
6381aad
 
 
 
 
 
 
 
 
 
 
 
 
982a049
34c9a69
 
 
9d91fdf
 
 
 
 
 
 
 
1b73c6d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9d91fdf
1b73c6d
9d91fdf
 
 
 
 
1b73c6d
3bf3d37
9d91fdf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
# 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.