axiom-coding-agent-setup 1.1.0 → 1.2.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/.agents/CONTEXT-MANAGEMENT.md +155 -155
- package/.agents/DEBUGGING.md +124 -124
- package/.agents/ENGINEERING.md +180 -180
- package/.agents/PERFORMANCE.md +164 -164
- package/.agents/SECURITY.md +109 -109
- package/.agents/WORKFLOW.md +143 -143
- package/.agents/skills/agent-browser/SKILL.md +55 -55
- package/.agents/skills/huggingface-deployment/SKILL.md +230 -0
- package/.agents/skills/project-design/SKILL.md +207 -207
- package/.agents/skills/project-design/references/ARCHITECTURE.md +641 -641
- package/.agents/skills/project-design/references/PROJECT_PLAN.md +315 -315
- package/.env.axiom +8 -8
- package/AGENTS.md +104 -104
- package/README.md +146 -145
- package/bin/cli.js +1 -1
- package/opencode.json +64 -64
- package/package.json +1 -1
- package/plugin/oh-my-openagent.json +198 -198
- package/skills-lock.json +57 -57
package/opencode.json
CHANGED
|
@@ -1,64 +1,64 @@
|
|
|
1
|
-
{
|
|
2
|
-
"$schema": "https://opencode.ai/config.json",
|
|
3
|
-
"plugin": ["superpowers@git+https://github.com/obra/superpowers.git"],
|
|
4
|
-
"mcp": {
|
|
5
|
-
"gradio": {
|
|
6
|
-
"type": "remote",
|
|
7
|
-
"url": "https://gradio-docs-mcp.hf.space/gradio_api/mcp/",
|
|
8
|
-
"enabled": true,
|
|
9
|
-
"headers": {}
|
|
10
|
-
},
|
|
11
|
-
"apify": {
|
|
12
|
-
"type": "remote",
|
|
13
|
-
"url": "https://mcp.apify.com/?tools=docs",
|
|
14
|
-
"enabled": true,
|
|
15
|
-
"headers": {}
|
|
16
|
-
},
|
|
17
|
-
"Agno": {
|
|
18
|
-
"type": "remote",
|
|
19
|
-
"url": "https://docs.agno.com/mcp",
|
|
20
|
-
"enabled": true,
|
|
21
|
-
"headers": {}
|
|
22
|
-
},
|
|
23
|
-
"Docs by LangChain": {
|
|
24
|
-
"type": "remote",
|
|
25
|
-
"url": "https://docs.langchain.com/mcp",
|
|
26
|
-
"enabled": true,
|
|
27
|
-
"headers": {}
|
|
28
|
-
},
|
|
29
|
-
"langfuse-docs": {
|
|
30
|
-
"type": "remote",
|
|
31
|
-
"url": "https://langfuse.com/api/mcp",
|
|
32
|
-
"enabled": true,
|
|
33
|
-
"headers": {}
|
|
34
|
-
},
|
|
35
|
-
"n8n-mcp": {
|
|
36
|
-
"type": "remote",
|
|
37
|
-
"url": "http://localhost:5678/mcp-server/http",
|
|
38
|
-
"enabled": true,
|
|
39
|
-
"headers": {
|
|
40
|
-
"Authorization": "Bearer {env:N8N_MCP_TOKEN}"
|
|
41
|
-
},
|
|
42
|
-
"oauth": false
|
|
43
|
-
},
|
|
44
|
-
"exa": {
|
|
45
|
-
"type": "remote",
|
|
46
|
-
"url": "https://mcp.exa.ai/mcp",
|
|
47
|
-
"enabled": true,
|
|
48
|
-
"headers": {}
|
|
49
|
-
},
|
|
50
|
-
"Neon": {
|
|
51
|
-
"type": "remote",
|
|
52
|
-
"url": "https://mcp.neon.tech/mcp",
|
|
53
|
-
"enabled": true,
|
|
54
|
-
"headers": {
|
|
55
|
-
"Authorization": "Bearer {env:NEON_API_KEY}"
|
|
56
|
-
}
|
|
57
|
-
},
|
|
58
|
-
"eraser": {
|
|
59
|
-
"type": "remote",
|
|
60
|
-
"url": "https://app.eraser.io/api/mcp",
|
|
61
|
-
"enabled": true
|
|
62
|
-
}
|
|
63
|
-
}
|
|
64
|
-
}
|
|
1
|
+
{
|
|
2
|
+
"$schema": "https://opencode.ai/config.json",
|
|
3
|
+
"plugin": ["superpowers@git+https://github.com/obra/superpowers.git"],
|
|
4
|
+
"mcp": {
|
|
5
|
+
"gradio": {
|
|
6
|
+
"type": "remote",
|
|
7
|
+
"url": "https://gradio-docs-mcp.hf.space/gradio_api/mcp/",
|
|
8
|
+
"enabled": true,
|
|
9
|
+
"headers": {}
|
|
10
|
+
},
|
|
11
|
+
"apify": {
|
|
12
|
+
"type": "remote",
|
|
13
|
+
"url": "https://mcp.apify.com/?tools=docs",
|
|
14
|
+
"enabled": true,
|
|
15
|
+
"headers": {}
|
|
16
|
+
},
|
|
17
|
+
"Agno": {
|
|
18
|
+
"type": "remote",
|
|
19
|
+
"url": "https://docs.agno.com/mcp",
|
|
20
|
+
"enabled": true,
|
|
21
|
+
"headers": {}
|
|
22
|
+
},
|
|
23
|
+
"Docs by LangChain": {
|
|
24
|
+
"type": "remote",
|
|
25
|
+
"url": "https://docs.langchain.com/mcp",
|
|
26
|
+
"enabled": true,
|
|
27
|
+
"headers": {}
|
|
28
|
+
},
|
|
29
|
+
"langfuse-docs": {
|
|
30
|
+
"type": "remote",
|
|
31
|
+
"url": "https://langfuse.com/api/mcp",
|
|
32
|
+
"enabled": true,
|
|
33
|
+
"headers": {}
|
|
34
|
+
},
|
|
35
|
+
"n8n-mcp": {
|
|
36
|
+
"type": "remote",
|
|
37
|
+
"url": "http://localhost:5678/mcp-server/http",
|
|
38
|
+
"enabled": true,
|
|
39
|
+
"headers": {
|
|
40
|
+
"Authorization": "Bearer {env:N8N_MCP_TOKEN}"
|
|
41
|
+
},
|
|
42
|
+
"oauth": false
|
|
43
|
+
},
|
|
44
|
+
"exa": {
|
|
45
|
+
"type": "remote",
|
|
46
|
+
"url": "https://mcp.exa.ai/mcp",
|
|
47
|
+
"enabled": true,
|
|
48
|
+
"headers": {}
|
|
49
|
+
},
|
|
50
|
+
"Neon": {
|
|
51
|
+
"type": "remote",
|
|
52
|
+
"url": "https://mcp.neon.tech/mcp",
|
|
53
|
+
"enabled": true,
|
|
54
|
+
"headers": {
|
|
55
|
+
"Authorization": "Bearer {env:NEON_API_KEY}"
|
|
56
|
+
}
|
|
57
|
+
},
|
|
58
|
+
"eraser": {
|
|
59
|
+
"type": "remote",
|
|
60
|
+
"url": "https://app.eraser.io/api/mcp",
|
|
61
|
+
"enabled": true
|
|
62
|
+
}
|
|
63
|
+
}
|
|
64
|
+
}
|
package/package.json
CHANGED
|
@@ -1,198 +1,198 @@
|
|
|
1
|
-
{
|
|
2
|
-
"$schema": "https://raw.githubusercontent.com/code-yeongyu/oh-my-openagent/dev/assets/oh-my-opencode.schema.json",
|
|
3
|
-
"agents": {
|
|
4
|
-
// === ORCHESTRATION CORE ===
|
|
5
|
-
// Kimi K2.6: #1 open-weight for long-horizon agentic work, 300-subagent swarm,
|
|
6
|
-
// 12h autonomous sessions, officially omo's top non-Claude fallback for Sisyphus.
|
|
7
|
-
// Fallback: DeepSeek V4 Pro for its raw SWE-bench muscle when K2.6 is unavailable.
|
|
8
|
-
"sisyphus": {
|
|
9
|
-
"model": "opencode-go/kimi-k2.6",
|
|
10
|
-
"fallback_models": [
|
|
11
|
-
{ "model": "opencode-go/deepseek-v4-pro" }
|
|
12
|
-
]
|
|
13
|
-
},
|
|
14
|
-
|
|
15
|
-
// === PLANNING & STRATEGY ===
|
|
16
|
-
// Prometheus auto-detects model family and switches prompts — Kimi K2.6 fits the
|
|
17
|
-
// Claude-like instruction-following prompt that omo uses for strategic planning.
|
|
18
|
-
// Fallback: GLM-5.1 for its independently verified long-horizon execution loop.
|
|
19
|
-
"prometheus": {
|
|
20
|
-
"model": "opencode-go/kimi-k2.6",
|
|
21
|
-
"fallback_models": [
|
|
22
|
-
{ "model": "opencode-go/glm-5.1" }
|
|
23
|
-
]
|
|
24
|
-
},
|
|
25
|
-
|
|
26
|
-
// === PLAN REVIEW ===
|
|
27
|
-
// Metis is the plan reviewer — needs strong reasoning + precision.
|
|
28
|
-
// Kimi K2.6 excels at catching non-obvious bugs and maintaining architectural integrity
|
|
29
|
-
// over extended review sessions per enterprise beta feedback.
|
|
30
|
-
// Fallback: GLM-5.1 which ranked #1 on NL2Repo for codebase structure comprehension.
|
|
31
|
-
"metis": {
|
|
32
|
-
"model": "opencode-go/kimi-k2.6",
|
|
33
|
-
"fallback_models": [
|
|
34
|
-
{ "model": "opencode-go/glm-5.1" }
|
|
35
|
-
]
|
|
36
|
-
},
|
|
37
|
-
|
|
38
|
-
// === ARCHITECTURE & DEBUGGING ===
|
|
39
|
-
// Oracle needs surgical precision in large codebases — Kimi K2.6 won benchmarks
|
|
40
|
-
// specifically on async/TypeVar edge-case bugs that require multi-cycle inference state.
|
|
41
|
-
// Fallback: DeepSeek V4 Pro, strongest on SWE-bench Verified (80.6%) and Codeforces (3206).
|
|
42
|
-
"oracle": {
|
|
43
|
-
"model": "opencode-go/kimi-k2.6",
|
|
44
|
-
"fallback_models": [
|
|
45
|
-
{ "model": "opencode-go/deepseek-v4-pro" }
|
|
46
|
-
]
|
|
47
|
-
},
|
|
48
|
-
|
|
49
|
-
// === HIGH-ACCURACY REVIEW ===
|
|
50
|
-
// Momus is the strictness reviewer — prompt is tuned for Claude-like models.
|
|
51
|
-
// GLM-5.1 promoted to primary: independently verified Code Arena Elo 1530 (#3 globally),
|
|
52
|
-
// spontaneously applied composition patterns in head-to-head tests vs K2.6.
|
|
53
|
-
// Fallback: Kimi K2.6 for its 12% improvement in code generation accuracy over K2.5.
|
|
54
|
-
"momus": {
|
|
55
|
-
"model": "opencode-go/glm-5.1",
|
|
56
|
-
"fallback_models": [
|
|
57
|
-
{ "model": "opencode-go/kimi-k2.6" }
|
|
58
|
-
]
|
|
59
|
-
},
|
|
60
|
-
|
|
61
|
-
// === TODO ORCHESTRATION ===
|
|
62
|
-
// Atlas is the task/todo orchestrator — auto-detects model family.
|
|
63
|
-
// Kimi K2.6 is the best fit for long multi-step coordination (300 sub-agents, 4000 steps).
|
|
64
|
-
// Fallback: DeepSeek V4 Pro for its $3.48/M cost efficiency at scale.
|
|
65
|
-
"atlas": {
|
|
66
|
-
"model": "opencode-go/kimi-k2.6",
|
|
67
|
-
"fallback_models": [
|
|
68
|
-
{ "model": "opencode-go/deepseek-v4-pro" }
|
|
69
|
-
]
|
|
70
|
-
},
|
|
71
|
-
|
|
72
|
-
// === DOCS & CODE SEARCH ===
|
|
73
|
-
// Librarian does documentation lookup and contextual code search — benefits most from
|
|
74
|
-
// Qwen3.6 Plus's 1M token context (vs K2.6's 256K ceiling). Handles full repo ingestion
|
|
75
|
-
// in a single pass. 4-6x cheaper than Claude-class for large-context read-heavy work.
|
|
76
|
-
// Fallback: DeepSeek V4 Pro — also 1M context, strong on LiveCodeBench (93.5%).
|
|
77
|
-
"librarian": {
|
|
78
|
-
"model": "opencode-go/qwen3.6-plus",
|
|
79
|
-
"fallback_models": [
|
|
80
|
-
{ "model": "opencode-go/deepseek-v4-pro" }
|
|
81
|
-
]
|
|
82
|
-
},
|
|
83
|
-
|
|
84
|
-
// === FAST CODEBASE GREP ===
|
|
85
|
-
// Explore is a fast grep/search agent — speed matters more than raw intelligence here.
|
|
86
|
-
// Qwen3.6 Plus runs 2-3x faster TPS than Claude Opus and has 1M context for large repos.
|
|
87
|
-
// Fallback: GLM-5.1 for its 55+ tokens/sec generation speed.
|
|
88
|
-
"explore": {
|
|
89
|
-
"model": "opencode-go/qwen3.6-plus",
|
|
90
|
-
"fallback_models": [
|
|
91
|
-
{ "model": "opencode-go/glm-5.1" }
|
|
92
|
-
]
|
|
93
|
-
},
|
|
94
|
-
|
|
95
|
-
// === VISION / SCREENSHOTS ===
|
|
96
|
-
// Multimodal-looker needs a model with solid vision. Kimi K2.6 has native multimodality
|
|
97
|
-
// and specifically handles visual-to-code workflows (UI designs → working code).
|
|
98
|
-
// No fallback: this is the only model in the stack with reliable multimodal support.
|
|
99
|
-
"multimodal-looker": {
|
|
100
|
-
"model": "opencode-go/kimi-k2.6"
|
|
101
|
-
},
|
|
102
|
-
|
|
103
|
-
// === JUNIOR WORKER ===
|
|
104
|
-
// Sisyphus-Junior handles delegated subtasks — Kimi K2.6 maintains session stability
|
|
105
|
-
// for parallel spawned workers (tool invocation success rate 96.60% per CodeBuddy eval).
|
|
106
|
-
// Fallback: DeepSeek V4 Pro for cost efficiency on high-volume parallel calls.
|
|
107
|
-
"sisyphus-junior": {
|
|
108
|
-
"model": "opencode-go/kimi-k2.6",
|
|
109
|
-
"fallback_models": [
|
|
110
|
-
{ "model": "opencode-go/deepseek-v4-pro" }
|
|
111
|
-
]
|
|
112
|
-
}
|
|
113
|
-
},
|
|
114
|
-
|
|
115
|
-
"categories": {
|
|
116
|
-
// === VISUAL / FRONTEND ENGINEERING ===
|
|
117
|
-
// GLM-5.1 promoted to primary: #3 globally on agentic webdev (Arena.ai Elo 1530),
|
|
118
|
-
// produced correct Tailwind + TypeScript components on first pass in head-to-head tests.
|
|
119
|
-
// Kimi K2.6 as fallback — strong on visual-to-code via native multimodal training.
|
|
120
|
-
"visual-engineering": {
|
|
121
|
-
"model": "opencode-go/glm-5.1",
|
|
122
|
-
"fallback_models": [
|
|
123
|
-
{ "model": "opencode-go/kimi-k2.6" }
|
|
124
|
-
]
|
|
125
|
-
},
|
|
126
|
-
|
|
127
|
-
// === MAXIMUM REASONING ===
|
|
128
|
-
// Ultrabrain is the highest-stakes category — Kimi K2.6 leads open-weight AI Index (54).
|
|
129
|
-
// Fallback: DeepSeek V4 Pro for its Codeforces 3206 and LiveCodeBench 93.5% supremacy.
|
|
130
|
-
"ultrabrain": {
|
|
131
|
-
"model": "opencode-go/kimi-k2.6",
|
|
132
|
-
"fallback_models": [
|
|
133
|
-
{ "model": "opencode-go/deepseek-v4-pro" }
|
|
134
|
-
]
|
|
135
|
-
},
|
|
136
|
-
|
|
137
|
-
// === DEEP / COMPLEX WORK ===
|
|
138
|
-
// Kimi K2.6: best open-source for long-horizon, sustained multi-step execution.
|
|
139
|
-
// Fallback: GLM-5.1 — demonstrated 655-iteration autonomous optimization loop,
|
|
140
|
-
// 8h uninterrupted task execution, strongest open-weight for backend deep dives.
|
|
141
|
-
"deep": {
|
|
142
|
-
"model": "opencode-go/kimi-k2.6",
|
|
143
|
-
"fallback_models": [
|
|
144
|
-
{ "model": "opencode-go/glm-5.1" }
|
|
145
|
-
]
|
|
146
|
-
},
|
|
147
|
-
|
|
148
|
-
// === CREATIVE / UI ARTISTRY ===
|
|
149
|
-
// GLM-5.1 as primary: spontaneously applies composition patterns, correct JSX on first
|
|
150
|
-
// pass, Code Arena voters prefer it for frontend aesthetics in head-to-head evals.
|
|
151
|
-
// Kimi K2.6 as fallback: Moonshot claims Awwwards-level frontend from single prompts.
|
|
152
|
-
"artistry": {
|
|
153
|
-
"model": "opencode-go/glm-5.1",
|
|
154
|
-
"fallback_models": [
|
|
155
|
-
{ "model": "opencode-go/kimi-k2.6" }
|
|
156
|
-
]
|
|
157
|
-
},
|
|
158
|
-
|
|
159
|
-
// === QUICK / TRIVIAL TASKS ===
|
|
160
|
-
// GLM-5.1: 55+ tokens/sec, HN devs rate it as "actually usable" for piecemeal tasks,
|
|
161
|
-
// compares well to GPT-5.4 for scoped, well-defined subtasks.
|
|
162
|
-
// No fallback needed — quick tasks should just pass or skip to the next agent.
|
|
163
|
-
"quick": {
|
|
164
|
-
"model": "opencode-go/glm-5.1"
|
|
165
|
-
},
|
|
166
|
-
|
|
167
|
-
// === UNSPECIFIED / MODERATE TASKS ===
|
|
168
|
-
// Kimi K2.6 is the safest general-purpose choice for ambiguous category routing.
|
|
169
|
-
// Fallback: Qwen3.6 Plus at $0.50/M input — 30x cheaper than Claude, close in SWE-bench.
|
|
170
|
-
"unspecified-low": {
|
|
171
|
-
"model": "opencode-go/kimi-k2.6",
|
|
172
|
-
"fallback_models": [
|
|
173
|
-
{ "model": "opencode-go/qwen3.6-plus" }
|
|
174
|
-
]
|
|
175
|
-
},
|
|
176
|
-
|
|
177
|
-
// === UNSPECIFIED / COMPLEX TASKS ===
|
|
178
|
-
// Same primary as unspecified-low — Kimi K2.6 handles both well.
|
|
179
|
-
// Fallback: GLM-5.1 rather than Qwen, for stronger reasoning depth on hard unknowns.
|
|
180
|
-
"unspecified-high": {
|
|
181
|
-
"model": "opencode-go/kimi-k2.6",
|
|
182
|
-
"fallback_models": [
|
|
183
|
-
{ "model": "opencode-go/glm-5.1" }
|
|
184
|
-
]
|
|
185
|
-
},
|
|
186
|
-
|
|
187
|
-
// === WRITING / DOCUMENTATION ===
|
|
188
|
-
// Kimi K2.6: strong instruction-following and consistent output quality.
|
|
189
|
-
// Fallback: GLM-5.1 — bilingual training (EN/CN) useful for mixed-language codebases
|
|
190
|
-
// and documentation; Qwen3.6 Plus also strong here but GLM edges it on coherence.
|
|
191
|
-
"writing": {
|
|
192
|
-
"model": "opencode-go/kimi-k2.6",
|
|
193
|
-
"fallback_models": [
|
|
194
|
-
{ "model": "opencode-go/glm-5.1" }
|
|
195
|
-
]
|
|
196
|
-
}
|
|
197
|
-
}
|
|
198
|
-
}
|
|
1
|
+
{
|
|
2
|
+
"$schema": "https://raw.githubusercontent.com/code-yeongyu/oh-my-openagent/dev/assets/oh-my-opencode.schema.json",
|
|
3
|
+
"agents": {
|
|
4
|
+
// === ORCHESTRATION CORE ===
|
|
5
|
+
// Kimi K2.6: #1 open-weight for long-horizon agentic work, 300-subagent swarm,
|
|
6
|
+
// 12h autonomous sessions, officially omo's top non-Claude fallback for Sisyphus.
|
|
7
|
+
// Fallback: DeepSeek V4 Pro for its raw SWE-bench muscle when K2.6 is unavailable.
|
|
8
|
+
"sisyphus": {
|
|
9
|
+
"model": "opencode-go/kimi-k2.6",
|
|
10
|
+
"fallback_models": [
|
|
11
|
+
{ "model": "opencode-go/deepseek-v4-pro" }
|
|
12
|
+
]
|
|
13
|
+
},
|
|
14
|
+
|
|
15
|
+
// === PLANNING & STRATEGY ===
|
|
16
|
+
// Prometheus auto-detects model family and switches prompts — Kimi K2.6 fits the
|
|
17
|
+
// Claude-like instruction-following prompt that omo uses for strategic planning.
|
|
18
|
+
// Fallback: GLM-5.1 for its independently verified long-horizon execution loop.
|
|
19
|
+
"prometheus": {
|
|
20
|
+
"model": "opencode-go/kimi-k2.6",
|
|
21
|
+
"fallback_models": [
|
|
22
|
+
{ "model": "opencode-go/glm-5.1" }
|
|
23
|
+
]
|
|
24
|
+
},
|
|
25
|
+
|
|
26
|
+
// === PLAN REVIEW ===
|
|
27
|
+
// Metis is the plan reviewer — needs strong reasoning + precision.
|
|
28
|
+
// Kimi K2.6 excels at catching non-obvious bugs and maintaining architectural integrity
|
|
29
|
+
// over extended review sessions per enterprise beta feedback.
|
|
30
|
+
// Fallback: GLM-5.1 which ranked #1 on NL2Repo for codebase structure comprehension.
|
|
31
|
+
"metis": {
|
|
32
|
+
"model": "opencode-go/kimi-k2.6",
|
|
33
|
+
"fallback_models": [
|
|
34
|
+
{ "model": "opencode-go/glm-5.1" }
|
|
35
|
+
]
|
|
36
|
+
},
|
|
37
|
+
|
|
38
|
+
// === ARCHITECTURE & DEBUGGING ===
|
|
39
|
+
// Oracle needs surgical precision in large codebases — Kimi K2.6 won benchmarks
|
|
40
|
+
// specifically on async/TypeVar edge-case bugs that require multi-cycle inference state.
|
|
41
|
+
// Fallback: DeepSeek V4 Pro, strongest on SWE-bench Verified (80.6%) and Codeforces (3206).
|
|
42
|
+
"oracle": {
|
|
43
|
+
"model": "opencode-go/kimi-k2.6",
|
|
44
|
+
"fallback_models": [
|
|
45
|
+
{ "model": "opencode-go/deepseek-v4-pro" }
|
|
46
|
+
]
|
|
47
|
+
},
|
|
48
|
+
|
|
49
|
+
// === HIGH-ACCURACY REVIEW ===
|
|
50
|
+
// Momus is the strictness reviewer — prompt is tuned for Claude-like models.
|
|
51
|
+
// GLM-5.1 promoted to primary: independently verified Code Arena Elo 1530 (#3 globally),
|
|
52
|
+
// spontaneously applied composition patterns in head-to-head tests vs K2.6.
|
|
53
|
+
// Fallback: Kimi K2.6 for its 12% improvement in code generation accuracy over K2.5.
|
|
54
|
+
"momus": {
|
|
55
|
+
"model": "opencode-go/glm-5.1",
|
|
56
|
+
"fallback_models": [
|
|
57
|
+
{ "model": "opencode-go/kimi-k2.6" }
|
|
58
|
+
]
|
|
59
|
+
},
|
|
60
|
+
|
|
61
|
+
// === TODO ORCHESTRATION ===
|
|
62
|
+
// Atlas is the task/todo orchestrator — auto-detects model family.
|
|
63
|
+
// Kimi K2.6 is the best fit for long multi-step coordination (300 sub-agents, 4000 steps).
|
|
64
|
+
// Fallback: DeepSeek V4 Pro for its $3.48/M cost efficiency at scale.
|
|
65
|
+
"atlas": {
|
|
66
|
+
"model": "opencode-go/kimi-k2.6",
|
|
67
|
+
"fallback_models": [
|
|
68
|
+
{ "model": "opencode-go/deepseek-v4-pro" }
|
|
69
|
+
]
|
|
70
|
+
},
|
|
71
|
+
|
|
72
|
+
// === DOCS & CODE SEARCH ===
|
|
73
|
+
// Librarian does documentation lookup and contextual code search — benefits most from
|
|
74
|
+
// Qwen3.6 Plus's 1M token context (vs K2.6's 256K ceiling). Handles full repo ingestion
|
|
75
|
+
// in a single pass. 4-6x cheaper than Claude-class for large-context read-heavy work.
|
|
76
|
+
// Fallback: DeepSeek V4 Pro — also 1M context, strong on LiveCodeBench (93.5%).
|
|
77
|
+
"librarian": {
|
|
78
|
+
"model": "opencode-go/qwen3.6-plus",
|
|
79
|
+
"fallback_models": [
|
|
80
|
+
{ "model": "opencode-go/deepseek-v4-pro" }
|
|
81
|
+
]
|
|
82
|
+
},
|
|
83
|
+
|
|
84
|
+
// === FAST CODEBASE GREP ===
|
|
85
|
+
// Explore is a fast grep/search agent — speed matters more than raw intelligence here.
|
|
86
|
+
// Qwen3.6 Plus runs 2-3x faster TPS than Claude Opus and has 1M context for large repos.
|
|
87
|
+
// Fallback: GLM-5.1 for its 55+ tokens/sec generation speed.
|
|
88
|
+
"explore": {
|
|
89
|
+
"model": "opencode-go/qwen3.6-plus",
|
|
90
|
+
"fallback_models": [
|
|
91
|
+
{ "model": "opencode-go/glm-5.1" }
|
|
92
|
+
]
|
|
93
|
+
},
|
|
94
|
+
|
|
95
|
+
// === VISION / SCREENSHOTS ===
|
|
96
|
+
// Multimodal-looker needs a model with solid vision. Kimi K2.6 has native multimodality
|
|
97
|
+
// and specifically handles visual-to-code workflows (UI designs → working code).
|
|
98
|
+
// No fallback: this is the only model in the stack with reliable multimodal support.
|
|
99
|
+
"multimodal-looker": {
|
|
100
|
+
"model": "opencode-go/kimi-k2.6"
|
|
101
|
+
},
|
|
102
|
+
|
|
103
|
+
// === JUNIOR WORKER ===
|
|
104
|
+
// Sisyphus-Junior handles delegated subtasks — Kimi K2.6 maintains session stability
|
|
105
|
+
// for parallel spawned workers (tool invocation success rate 96.60% per CodeBuddy eval).
|
|
106
|
+
// Fallback: DeepSeek V4 Pro for cost efficiency on high-volume parallel calls.
|
|
107
|
+
"sisyphus-junior": {
|
|
108
|
+
"model": "opencode-go/kimi-k2.6",
|
|
109
|
+
"fallback_models": [
|
|
110
|
+
{ "model": "opencode-go/deepseek-v4-pro" }
|
|
111
|
+
]
|
|
112
|
+
}
|
|
113
|
+
},
|
|
114
|
+
|
|
115
|
+
"categories": {
|
|
116
|
+
// === VISUAL / FRONTEND ENGINEERING ===
|
|
117
|
+
// GLM-5.1 promoted to primary: #3 globally on agentic webdev (Arena.ai Elo 1530),
|
|
118
|
+
// produced correct Tailwind + TypeScript components on first pass in head-to-head tests.
|
|
119
|
+
// Kimi K2.6 as fallback — strong on visual-to-code via native multimodal training.
|
|
120
|
+
"visual-engineering": {
|
|
121
|
+
"model": "opencode-go/glm-5.1",
|
|
122
|
+
"fallback_models": [
|
|
123
|
+
{ "model": "opencode-go/kimi-k2.6" }
|
|
124
|
+
]
|
|
125
|
+
},
|
|
126
|
+
|
|
127
|
+
// === MAXIMUM REASONING ===
|
|
128
|
+
// Ultrabrain is the highest-stakes category — Kimi K2.6 leads open-weight AI Index (54).
|
|
129
|
+
// Fallback: DeepSeek V4 Pro for its Codeforces 3206 and LiveCodeBench 93.5% supremacy.
|
|
130
|
+
"ultrabrain": {
|
|
131
|
+
"model": "opencode-go/kimi-k2.6",
|
|
132
|
+
"fallback_models": [
|
|
133
|
+
{ "model": "opencode-go/deepseek-v4-pro" }
|
|
134
|
+
]
|
|
135
|
+
},
|
|
136
|
+
|
|
137
|
+
// === DEEP / COMPLEX WORK ===
|
|
138
|
+
// Kimi K2.6: best open-source for long-horizon, sustained multi-step execution.
|
|
139
|
+
// Fallback: GLM-5.1 — demonstrated 655-iteration autonomous optimization loop,
|
|
140
|
+
// 8h uninterrupted task execution, strongest open-weight for backend deep dives.
|
|
141
|
+
"deep": {
|
|
142
|
+
"model": "opencode-go/kimi-k2.6",
|
|
143
|
+
"fallback_models": [
|
|
144
|
+
{ "model": "opencode-go/glm-5.1" }
|
|
145
|
+
]
|
|
146
|
+
},
|
|
147
|
+
|
|
148
|
+
// === CREATIVE / UI ARTISTRY ===
|
|
149
|
+
// GLM-5.1 as primary: spontaneously applies composition patterns, correct JSX on first
|
|
150
|
+
// pass, Code Arena voters prefer it for frontend aesthetics in head-to-head evals.
|
|
151
|
+
// Kimi K2.6 as fallback: Moonshot claims Awwwards-level frontend from single prompts.
|
|
152
|
+
"artistry": {
|
|
153
|
+
"model": "opencode-go/glm-5.1",
|
|
154
|
+
"fallback_models": [
|
|
155
|
+
{ "model": "opencode-go/kimi-k2.6" }
|
|
156
|
+
]
|
|
157
|
+
},
|
|
158
|
+
|
|
159
|
+
// === QUICK / TRIVIAL TASKS ===
|
|
160
|
+
// GLM-5.1: 55+ tokens/sec, HN devs rate it as "actually usable" for piecemeal tasks,
|
|
161
|
+
// compares well to GPT-5.4 for scoped, well-defined subtasks.
|
|
162
|
+
// No fallback needed — quick tasks should just pass or skip to the next agent.
|
|
163
|
+
"quick": {
|
|
164
|
+
"model": "opencode-go/glm-5.1"
|
|
165
|
+
},
|
|
166
|
+
|
|
167
|
+
// === UNSPECIFIED / MODERATE TASKS ===
|
|
168
|
+
// Kimi K2.6 is the safest general-purpose choice for ambiguous category routing.
|
|
169
|
+
// Fallback: Qwen3.6 Plus at $0.50/M input — 30x cheaper than Claude, close in SWE-bench.
|
|
170
|
+
"unspecified-low": {
|
|
171
|
+
"model": "opencode-go/kimi-k2.6",
|
|
172
|
+
"fallback_models": [
|
|
173
|
+
{ "model": "opencode-go/qwen3.6-plus" }
|
|
174
|
+
]
|
|
175
|
+
},
|
|
176
|
+
|
|
177
|
+
// === UNSPECIFIED / COMPLEX TASKS ===
|
|
178
|
+
// Same primary as unspecified-low — Kimi K2.6 handles both well.
|
|
179
|
+
// Fallback: GLM-5.1 rather than Qwen, for stronger reasoning depth on hard unknowns.
|
|
180
|
+
"unspecified-high": {
|
|
181
|
+
"model": "opencode-go/kimi-k2.6",
|
|
182
|
+
"fallback_models": [
|
|
183
|
+
{ "model": "opencode-go/glm-5.1" }
|
|
184
|
+
]
|
|
185
|
+
},
|
|
186
|
+
|
|
187
|
+
// === WRITING / DOCUMENTATION ===
|
|
188
|
+
// Kimi K2.6: strong instruction-following and consistent output quality.
|
|
189
|
+
// Fallback: GLM-5.1 — bilingual training (EN/CN) useful for mixed-language codebases
|
|
190
|
+
// and documentation; Qwen3.6 Plus also strong here but GLM edges it on coherence.
|
|
191
|
+
"writing": {
|
|
192
|
+
"model": "opencode-go/kimi-k2.6",
|
|
193
|
+
"fallback_models": [
|
|
194
|
+
{ "model": "opencode-go/glm-5.1" }
|
|
195
|
+
]
|
|
196
|
+
}
|
|
197
|
+
}
|
|
198
|
+
}
|
package/skills-lock.json
CHANGED
|
@@ -1,57 +1,57 @@
|
|
|
1
|
-
{
|
|
2
|
-
"version": 1,
|
|
3
|
-
"skills": {
|
|
4
|
-
"agent-browser": {
|
|
5
|
-
"source": "vercel-labs/agent-browser",
|
|
6
|
-
"sourceType": "github",
|
|
7
|
-
"skillPath": "skills/agent-browser/SKILL.md",
|
|
8
|
-
"computedHash": "228f87d57035100d9dc6efcfc05aafd4b6e3962adacaa04b8217ab2fadb15dc8"
|
|
9
|
-
},
|
|
10
|
-
"developing-with-streamlit": {
|
|
11
|
-
"source": "streamlit/agent-skills",
|
|
12
|
-
"sourceType": "github",
|
|
13
|
-
"skillPath": "developing-with-streamlit/SKILL.md",
|
|
14
|
-
"computedHash": "51823384978f738f76ff8f601e11c1343aebea03d2ad782dc73fc22d3f129812"
|
|
15
|
-
},
|
|
16
|
-
"fastapi": {
|
|
17
|
-
"source": "fastapi/fastapi",
|
|
18
|
-
"sourceType": "github",
|
|
19
|
-
"skillPath": "fastapi/.agents/skills/fastapi/SKILL.md",
|
|
20
|
-
"computedHash": "9feb517089c269af4fdab5031dea1100a92fd50081e906ccfaf8a971858e39f9"
|
|
21
|
-
},
|
|
22
|
-
"fastapi-templates": {
|
|
23
|
-
"source": "wshobson/agents",
|
|
24
|
-
"sourceType": "github",
|
|
25
|
-
"computedHash": "1a5fe66bd2683afd1db9afdf37a25cd8b9195c369c98489333badf0406fc91b3"
|
|
26
|
-
},
|
|
27
|
-
"frontend-design": {
|
|
28
|
-
"source": "anthropics/skills",
|
|
29
|
-
"sourceType": "github",
|
|
30
|
-
"skillPath": "skills/frontend-design/SKILL.md",
|
|
31
|
-
"computedHash": "516bd2154eb843a8240e43d5b285229129853114ad7075a5e141e1c08e408c84"
|
|
32
|
-
},
|
|
33
|
-
"git-commit": {
|
|
34
|
-
"source": "github/awesome-copilot",
|
|
35
|
-
"sourceType": "github",
|
|
36
|
-
"computedHash": "2607fc60629b82b257136dd2a7a373f0a4466c0b49df7746d845d59313c99b21"
|
|
37
|
-
},
|
|
38
|
-
"gradio": {
|
|
39
|
-
"source": "gradio-app/gradio",
|
|
40
|
-
"sourceType": "github",
|
|
41
|
-
"skillPath": ".agents/skills/gradio/SKILL.md",
|
|
42
|
-
"computedHash": "5078d8a4cc2b562731620f36f5f0d3ad054bd51f7a98baa547b04b893744823b"
|
|
43
|
-
},
|
|
44
|
-
"skill-creator": {
|
|
45
|
-
"source": "anthropics/skills",
|
|
46
|
-
"sourceType": "github",
|
|
47
|
-
"skillPath": "skills/skill-creator/SKILL.md",
|
|
48
|
-
"computedHash": "7e3c9cd74e9e2b4828527a857170e86310f2dab5ea8030a9043df2c7e6c88857"
|
|
49
|
-
},
|
|
50
|
-
"ui-ux-pro-max": {
|
|
51
|
-
"source": "nextlevelbuilder/ui-ux-pro-max-skill",
|
|
52
|
-
"sourceType": "github",
|
|
53
|
-
"skillPath": ".claude/skills/ui-ux-pro-max/SKILL.md",
|
|
54
|
-
"computedHash": "6337038fe1fe6bbe1b9f252ab678ee575859190bab6f0f246f4061824eb40875"
|
|
55
|
-
}
|
|
56
|
-
}
|
|
57
|
-
}
|
|
1
|
+
{
|
|
2
|
+
"version": 1,
|
|
3
|
+
"skills": {
|
|
4
|
+
"agent-browser": {
|
|
5
|
+
"source": "vercel-labs/agent-browser",
|
|
6
|
+
"sourceType": "github",
|
|
7
|
+
"skillPath": "skills/agent-browser/SKILL.md",
|
|
8
|
+
"computedHash": "228f87d57035100d9dc6efcfc05aafd4b6e3962adacaa04b8217ab2fadb15dc8"
|
|
9
|
+
},
|
|
10
|
+
"developing-with-streamlit": {
|
|
11
|
+
"source": "streamlit/agent-skills",
|
|
12
|
+
"sourceType": "github",
|
|
13
|
+
"skillPath": "developing-with-streamlit/SKILL.md",
|
|
14
|
+
"computedHash": "51823384978f738f76ff8f601e11c1343aebea03d2ad782dc73fc22d3f129812"
|
|
15
|
+
},
|
|
16
|
+
"fastapi": {
|
|
17
|
+
"source": "fastapi/fastapi",
|
|
18
|
+
"sourceType": "github",
|
|
19
|
+
"skillPath": "fastapi/.agents/skills/fastapi/SKILL.md",
|
|
20
|
+
"computedHash": "9feb517089c269af4fdab5031dea1100a92fd50081e906ccfaf8a971858e39f9"
|
|
21
|
+
},
|
|
22
|
+
"fastapi-templates": {
|
|
23
|
+
"source": "wshobson/agents",
|
|
24
|
+
"sourceType": "github",
|
|
25
|
+
"computedHash": "1a5fe66bd2683afd1db9afdf37a25cd8b9195c369c98489333badf0406fc91b3"
|
|
26
|
+
},
|
|
27
|
+
"frontend-design": {
|
|
28
|
+
"source": "anthropics/skills",
|
|
29
|
+
"sourceType": "github",
|
|
30
|
+
"skillPath": "skills/frontend-design/SKILL.md",
|
|
31
|
+
"computedHash": "516bd2154eb843a8240e43d5b285229129853114ad7075a5e141e1c08e408c84"
|
|
32
|
+
},
|
|
33
|
+
"git-commit": {
|
|
34
|
+
"source": "github/awesome-copilot",
|
|
35
|
+
"sourceType": "github",
|
|
36
|
+
"computedHash": "2607fc60629b82b257136dd2a7a373f0a4466c0b49df7746d845d59313c99b21"
|
|
37
|
+
},
|
|
38
|
+
"gradio": {
|
|
39
|
+
"source": "gradio-app/gradio",
|
|
40
|
+
"sourceType": "github",
|
|
41
|
+
"skillPath": ".agents/skills/gradio/SKILL.md",
|
|
42
|
+
"computedHash": "5078d8a4cc2b562731620f36f5f0d3ad054bd51f7a98baa547b04b893744823b"
|
|
43
|
+
},
|
|
44
|
+
"skill-creator": {
|
|
45
|
+
"source": "anthropics/skills",
|
|
46
|
+
"sourceType": "github",
|
|
47
|
+
"skillPath": "skills/skill-creator/SKILL.md",
|
|
48
|
+
"computedHash": "7e3c9cd74e9e2b4828527a857170e86310f2dab5ea8030a9043df2c7e6c88857"
|
|
49
|
+
},
|
|
50
|
+
"ui-ux-pro-max": {
|
|
51
|
+
"source": "nextlevelbuilder/ui-ux-pro-max-skill",
|
|
52
|
+
"sourceType": "github",
|
|
53
|
+
"skillPath": ".claude/skills/ui-ux-pro-max/SKILL.md",
|
|
54
|
+
"computedHash": "6337038fe1fe6bbe1b9f252ab678ee575859190bab6f0f246f4061824eb40875"
|
|
55
|
+
}
|
|
56
|
+
}
|
|
57
|
+
}
|