hybrid
Skill1 starsChanged 4 months agoMITView on GitHub
Unified hybrid orchestrator skill for Claude, Grok, and Codex. Offload bulk map-phase work to local ollamaq/Ollama; keep premium tokens for orchestration, reduce/clean, and high-reasoning only. Use when bulk/mechanical/gen_test work appears, or when starting a tandem-tokens session.
- Reads credentials
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 15
Never echo bulk worker output into the conversation.
- Very specific: line 49
Ensure daemon: ollamaq daemon (one per machine).
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Hybrid orchestrator (Claude / Grok / Codex)
You are the premium orchestrator. Local ollamaq (tandem-tokens) runs map work on Ollama. Energy and cloud $ are scarce — maximize local/offline.
Core contract
- Worker output stays on disk under
results/<job_id>/. - You read only
digest.md(viaollamaq digest/ MCPqueue_digest/hybrid_digest) — hard-capped ~8 KB. - Pull
merged.jsonlonly for a deliberate reduce on residue; prefer a slice. - Never echo bulk worker output into the conversation.
When to offload
Offload when ≥ ~20 similar independent items AND one of:
| Mode | Use for |
map | clean / normalize / extract / classify / summarize / boilerplate codegen |
gen_test | generate candidates + scriptable verifier |
pipeline | cheap triage → stronger model on flagged subset |
When NOT to offload
- Single hard reasoning problem
- Needs repo-wide context the local model lacks
- Wrong answers silently corrupt data with no verifier
- Fewer than ~20 items (overhead > savings)
Routing (config/models.yaml)
model: "auto" + task_tag:
clean/extract/code→ workhorse (qwen2.5-14b class)summarize→ long-context localclassify→ fast small localgenerate→ gen_test model
Optional residual escalate: job retry.escalate_to or config fallbacks.escalate_to —
local bigger model, or cloud:grok / cloud:grok/grok-3-mini (needs cloud.grok.enabled + XAI_API_KEY).
Workflow
- Offload — write a validated job spec →
ollamaq submit/ MCPqueue_submit/hybrid_offload. - Ensure daemon:
ollamaq daemon(one per machine). Don’t poll tightly. - Harvest — when
done, read digest only; reduce residue; write final artifact. - Score —
ollamaq score/hybrid_scorefor tokens, cloud $ avoided, GPU Wh.
MCP tools
| Standard | Grok alias | Purpose |
ollama_quick | hybrid_quick | tiny one-shot (capped) |
queue_submit | hybrid_offload | enqueue batch |
queue_status | hybrid_status | queue state |
queue_digest | hybrid_digest | finished digest |
recommend_models | hybrid_recommend | VRAM fit |
hybrid_score | same | scoreboard + energy |
hybrid_energy | same | live nvidia-smi power |
hybrid_policy | same | this policy text |
- Claude Code:
.mcp.json→python -m ollamaq.mcp_server(+CLAUDE.md/.claude/skills/) - Codex:
clients/codex.config.toml→~/.codex/config.toml; policy inAGENTS.md - Grok:
python -m ollamaq.mcp_grok(hybrid_* aliases) —clients/grok.mcp.json - Cursor / VS Code / Continue / Cline / Windsurf:
clients/cursor.mcp.jsonorclients/generic.mcp.json - No MCP: shell only —
ollamaq submit|run|status|digest|score(works from any agent that can run commands)
Client matrix: clients/README.md.
Example map spec
{
"name": "clean-contacts",
"mode": "map",
"model": "auto",
"task_tag": "clean",
"input": {
"kind": "csv",
"path": "data/contacts.csv",
"chunk": { "strategy": "rows", "size": 50 }
},
"prompt": {
"system": "You are a data cleaning engine. Output ONLY JSON.",
"template": "Normalize these rows... return {\"rows\": [...]}. Input:\\n{{chunk}}",
"schema": { "type": "object", "required": ["rows"] }
},
"params": { "temperature": 0.1 },
"retry": {
"max_attempts": 3,
"on_schema_fail": "repair_then_retry",
"escalate_to": "auto"
}
}Kickoff checklist
ollamaq doctor— config + Ollama OKollamaq recommend— models fit GPUhybrid_policyor re-read this skill- Offload eligible bulk work; harvest digests; report savings
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

