fleet

name	fleet
description	Adaptive parallel LLM router for Ollama cloud models. Use when the user asks to "/fleet <prompt>", wants to route a prompt to the best Ollama cloud model, or wants to compare/ensemble multiple models on a single prompt.

Fleet Router

When the user invokes /fleet <args>, you MUST execute the fleet binary via the Bash tool — do NOT answer the prompt yourself. The whole point of this skill is to route through Fleet → Ollama, which only happens by running the CLI.

How to invoke

1. Build the command. The binary is on PATH at ~/.local/bin/fleet; prefer fleet, fall back to /Users/bistrocloud/fleet-router/venv/bin/fleet.

2. Pass the user's prompt as a single quoted argument. Forward --parallel, --model <name>, and --config <path> if supplied.

3. Suppress sentence-transformers/HF Hub boilerplate so the user sees the answer, not loader noise. Pipe stderr through grep:

   fleet [flags] "<prompt>" 2> >(grep -v -E "huggingface_hub|Loading weights|Batches:|Warning: You are sending unauthenticated") 1>&1
   

   Or merge+filter both streams:

   fleet [flags] "<prompt>" 2>&1 | grep -v -E "huggingface_hub|Loading weights|Batches:|Warning: You are"
   

4. Bash timeout: pass timeout: 300000 (5 minutes). Max-quality defaults fan out to 3 models × N samples + verifier + judge + optional escalation/refinement — 60–120s is normal, 180s+ on cold start with sentence-transformers loading. The default 2-minute timeout will clip many runs; do not assume the call hung.

5. If the user wants a fast answer, suggest --model deepseek-v4-flash which skips parallel synthesis and returns in ~5–10s.

6. Surface stdout verbatim — fleet already formats parallel results with --- model --- headers. If exit code is non-zero, show stderr and stop. Do NOT re-answer the prompt yourself; that defeats the skill's purpose.

Examples

/fleet "write a login function"
   → fleet "write a login function"

/fleet --parallel "compare these approaches"
   → fleet --parallel "compare these approaches"

/fleet --model glm-5.1 "translate this paragraph: ..."
   → fleet --model glm-5.1 "translate this paragraph: ..."

Behavior (informational)

Fleet:

1. Classifies the prompt into a task tag (code, creative, math, reasoning, summarize, translate, general).
2. If classifier confidence ≥ thresholds.single_confidence, routes to the single best model.
3. Otherwise (low confidence, --parallel, or default max-quality config), runs up to 3 models in parallel and synthesizes the best answer via verifiers + judge.

Configuration lives in ~/.fleet/config.yaml (falls back to fleet/config.yaml shipped with the package). Cloud access requires base_url: https://ollama.com plus an api_key from ollama.com/settings/keys.

Models (default config)

Model	Best For
deepseek-v4-pro	Code, reasoning, math
glm-5.1	Creative writing, Chinese, long context
minimax-m2.7	Summarization, dialogue
deepseek-v4-flash	Fast drafts