consult-cloudflare-models

name

consult-cloudflare-models

description

Multi-model fan-out across Cloudflare Workers AI. Lists all available text-gen models (Llama 3.3 70B, GPT-OSS 120B, Nemotron 3 120B, Qwen 3, GLM 4.7, Mistral, Phi, Gemma, etc.) and fans a single prompt to N of them in parallel, then aggregates replies. Use when the user says "/consult-cloudflare-models", "/consult-cloudflaremodels", "ask cloudflare models", "cf fan-out", or wants triangulation across multiple CF-hosted models. Subcommands - list, search, run, fan-out.

consult-cloudflare-models

Fan a single prompt across multiple Cloudflare Workers AI text-generation models in one call. Complements consult-cloudflare (single-model with list/search/run subcommands) by being multi-model-by-default.

When to use

• User says /consult-cloudflare-models, /consult-cloudflaremodels, "ask cloudflare models", "cf fan-out", "cf second opinions"
• User names ≥2 CF-hosted models: llama-3.3-70b, gpt-oss-120b, nemotron-3-120b, qwen3-30b, glm-4.7-flash, mistral-small-24b, gemma-4-26b
• User wants a triangulated panel across families
• User wants the list of available models before picking
• User explicitly wants CF (not NIM) for cost / region / latency reasons

For single-model CF calls use consult-cloudflare.

Prerequisites

export CF_API_TOKEN='cfut_...'   # Cloudflare API token with Workers AI scope
export CF_ACCOUNT_ID='3c1f...'   # 32-char hex account id (dash.cloudflare.com right sidebar)

Verify the token:

curl -s "https://api.cloudflare.com/client/v4/user/tokens/verify" \
  -H "Authorization: Bearer $CF_API_TOKEN"

{"success":true,"result":{"status":"active"}} = good. If the token is scope-restricted (typical for cfut_ user tokens), /user, /accounts, /memberships will return 401 — only /user/tokens/verify and /accounts/{id}/ai/* are accessible. That's fine; the fan-out only needs the AI endpoint.

Subcommands

list — show all text-gen models on CF

curl -s "https://api.cloudflare.com/client/v4/accounts/$CF_ACCOUNT_ID/ai/models/search?per_page=200" \
  -H "Authorization: Bearer $CF_API_TOKEN" \
  | python3 -c "
import sys, json
d = json.load(sys.stdin)
for m in d.get('result', []):
    if 'text-generation' in (m.get('task') or {}).get('name','').lower():
        print(m['name'])
"

search <keyword> — filter

KW="${KW:-llama}"
curl -s "https://api.cloudflare.com/client/v4/accounts/$CF_ACCOUNT_ID/ai/models/search?per_page=200" \
  -H "Authorization: Bearer $CF_API_TOKEN" \
  | python3 -c "
import sys, json, os
kw = os.environ['KW'].lower()
for m in json.load(sys.stdin).get('result', []):
    if kw in m['name'].lower(): print(m['name'])
"

Useful keywords: llama, gpt-oss, nemotron, qwen, glm, mistral, gemma, phi, deepseek.

run <model> <prompt> — single-model call

Use consult-cloudflare for this. Or directly (note: model name needs the @cf/ prefix):

MODEL="${MODEL:-@cf/meta/llama-3.3-70b-instruct-fp8-fast}"
PROMPT="${PROMPT:-Give me a one-sentence answer.}"
curl -s --max-time 90 \
  "https://api.cloudflare.com/client/v4/accounts/$CF_ACCOUNT_ID/ai/run/$MODEL" \
  -H "Authorization: Bearer $CF_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d "$(python3 -c "
import json, os
print(json.dumps({
  'messages': [{'role': 'user', 'content': os.environ['PROMPT']}],
  'max_tokens': 1500,
}))")" \
  | python3 -c "
import sys, json
d = json.load(sys.stdin)
print(d.get('result', {}).get('response', d))
"

fan-out — fan one prompt to N CF models (the headline command)

# 1. Pick the panel — diversify families (note: CF uses @cf/ prefix on names)
PANEL=(
  "@cf/meta/llama-3.3-70b-instruct-fp8-fast"
  "@cf/openai/gpt-oss-120b"
  "@cf/nvidia/nemotron-3-120b"
  "@cf/qwen/qwen3-30b-a3b-fp8"
  "@cf/z-ai/glm-4.7-flash"
  "@cf/mistralai/mistral-small-3.1-24b-instruct"
)

# 2. Write the prompt to a file
cat > /tmp/cf_prompt.txt <<'EOF'
Your prompt here. Self-contained — no tool access. Ask for quantitative + cited answers only.
EOF

# 3. Fan out (sequential)
mkdir -p /tmp/cf_fanout
for MODEL in "${PANEL[@]}"; do
  SLUG=$(echo "$MODEL" | tr '/@.-' '_____')
  PROMPT="$(cat /tmp/cf_prompt.txt)" MODEL="$MODEL" python3 -c "
import json, os, subprocess
model = os.environ['MODEL']
payload = {
  'messages': [{'role': 'user', 'content': os.environ['PROMPT']}],
  'max_tokens': 1500,
  'temperature': 0.3,
}
url = f'https://api.cloudflare.com/client/v4/accounts/{os.environ[\"CF_ACCOUNT_ID\"]}/ai/run/{model}'
r = subprocess.run(['curl','-s','--max-time','90', url,
  '-H', f'Authorization: Bearer {os.environ[\"CF_API_TOKEN\"]}',
  '-H', 'Content-Type: application/json',
  '-d', json.dumps(payload)], capture_output=True, text=True, timeout=95)
try:
    d = json.loads(r.stdout)
    if d.get('success') is False:
        out = f'ERROR: {d.get(\"errors\", d)}'
    else:
        out = d.get('result', {}).get('response') or json.dumps(d)[:1000]
except Exception as exc:
    out = f'ERROR: {exc}; raw={r.stdout[:400]}'
print(out)
" > "/tmp/cf_fanout/$SLUG.md" 2>&1
  echo "  -> /tmp/cf_fanout/$SLUG.md ($(wc -c < /tmp/cf_fanout/$SLUG.md) bytes)"
  sleep 0.5
done

# 4. Aggregate
for f in /tmp/cf_fanout/*.md; do
  echo "=== $(basename "$f" .md) ==="
  head -20 "$f"
  echo
done

Curated panels

# Wide diversity (default for triangulation)
PANEL_WIDE=(
  "@cf/meta/llama-3.3-70b-instruct-fp8-fast"
  "@cf/openai/gpt-oss-120b"
  "@cf/nvidia/nemotron-3-120b"
  "@cf/qwen/qwen3-30b-a3b-fp8"
  "@cf/z-ai/glm-4.7-flash"
)

# Smaller / cheaper (rapid checks)
PANEL_FAST=(
  "@cf/meta/llama-3.1-8b-instruct"
  "@cf/qwen/qwen3-8b-instruct"
  "@cf/google/gemma-3-4b-it"
)

# Reasoning-heavy
PANEL_REASON=(
  "@cf/openai/gpt-oss-120b"
  "@cf/nvidia/nemotron-3-120b"
  "@cf/mistralai/mistral-large-2"
)

CF-specific pitfalls

• @cf/ prefix is required in the run URL path — meta/llama-3.3-70b-instruct-fp8-fast won't resolve.
• Some models reject temperature — if you get a 400 about "unsupported field", drop it from the payload.
• Rate limits are per-account, not per-model — fan-out hits the same bucket. 5 models with sleep 0.5 is fine; 20 will throttle.
• Response shape varies: most return {"result": {"response": "..."}} but some (especially older models) return {"result": [{"response": "..."}]} or just {"result": "..."}. The fan-out template above handles the common case; if a model gives empty output, inspect the raw response in /tmp/cf_fanout/<slug>.md.
• fp8-fast variants have ~30% lower latency at ~equivalent quality on Llama 3.3 — prefer them for fan-outs.

When NOT to use

• Single-model question — use consult-cloudflare instead, simpler payload.
• Network-restricted environments (CF API needs egress to api.cloudflare.com) — fall back to NIM if NVIDIA egress is open.
• Production decision-making — multi-model votes are inputs to a human decision, not the decision itself.

🔴 MANDATORY: Leakage-Context Block

Every prompt sent to CF models for asset-class / edge / pick-performance questions MUST include a leakage-context block. Multi-model consensus is worthless when all models see the same corrupted inputs. (Incident P1: 5/5 NIM models converged on a false COMMODITY alpha in 2026-05-25 because the prompt omitted leakage evidence.)

Template — paste at the TOP of any analysis prompt

LEAKAGE CONTEXT — Read before analyzing:

The following known data-quality incidents affect the dataset you are about
to analyze. Your first task is to determine whether ANY of these incidents
invalidate the numbers presented. If so, flag it explicitly and adjust your
analysis accordingly.

Known incidents (see reports/hypothesis_registry.json for full state):
 - H-101 (COMMODITY): COT publication-lag look-ahead. Pre-dedup commodity
   numbers show PF~4.3/WR~58%; post-policy ground truth is PF~0.31/WR~11%.
 - M-095 (COMMODITY): CT=F cotton over-emission. 230 closed picks from 16
   unique signal dates (14x re-emission ratio).
 - M-069 (NET): slippage reclassification — gross-PF vs net-PF can diverge
   by 2-3x.
 - QUALITY_BUCKETS (ALL): quality_bucket in performance_report_*.json is
   POST-HOC and circular. Do NOT treat as a pre-trade filter.

Be skeptical:
 - If one symbol/source dominates, flag concentration risk.
 - If PF>3.0 on a single asset class, check for look-ahead or survivorship
   bias before recommending allocation.

When to include

• ALWAYS for asset-class performance questions citing dashboard numbers
• ALWAYS for COMMODITY questions (H-101 / M-095)
• ALWAYS when referencing quality_bucket data
• SKIP only for purely theoretical / no-dataset prompts

Post-fan-out cross-check

After collecting model replies, flag any model that endorses a known-bad signal without addressing the cited leakage evidence.

Related

• consult-cloudflare — single-model CF consult, has its own list/search/run subcommands
• consult-nvidia-models — same fan-out shape but NVIDIA NIM provider
• swarm-run / swarm-second-opinion — provider-agnostic fan-out / 3-engine consensus
• tools/swarm/api_consult.py — backend behind the existing consult skills