#!/usr/bin/env bash
# hardstop-test.sh — prueft die drei Governor-Pfade mit direkten curls:
#   klein  -> passthrough (Modell antwortet normal)
#   mittel -> soft-Einschub (Modell bekommt die Stopp-Anweisung, antwortet)
#   gross  -> HART-STOPP (Governor antwortet selbst, KEIN Modell-Call)
set -u
GOV="http://127.0.0.1:8100/v1/chat/completions"

python3 - <<'PY' > /tmp/gov_small.json
import json
print(json.dumps({"model":"Qwen3-Coder-Next","messages":[{"role":"user","content":"Reply with exactly: SMALL"}],"max_tokens":8,"stream":False}))
PY
python3 - <<'PY' > /tmp/gov_mid.json
import json
print(json.dumps({"model":"Qwen3-Coder-Next","messages":[{"role":"user","content":"BEGIN "+"lorem ipsum "*2500+" END"}],"max_tokens":40,"stream":False}))
PY
python3 - <<'PY' > /tmp/gov_big.json
import json
print(json.dumps({"model":"Qwen3-Coder-Next","messages":[{"role":"user","content":"x"*40000}],"max_tokens":16,"stream":False}))
PY

echo "=== TEST A: klein (passthrough) ==="
curl -s -m 60 "$GOV" -H "Content-Type: application/json" --data @/tmp/gov_small.json | jq -r '.choices[0].message.content'

echo "=== TEST B: mittel ~30k Zeichen (soft-Einschub, geht ans Modell) ==="
curl -s -m 120 "$GOV" -H "Content-Type: application/json" --data @/tmp/gov_mid.json | jq '{id, content: (.choices[0].message.content|.[0:80])}'

echo "=== TEST C: gross ~40k Zeichen (HART-STOPP, kein Modell-Call) ==="
curl -s -m 30 "$GOV" -H "Content-Type: application/json" --data @/tmp/gov_big.json | jq '{id, finish: .choices[0].finish_reason, content: (.choices[0].message.content|.[0:70]), usage}'