Sync Kilo models with upstream gateway

[Ardakilic]

Sync Kilo models with upstream gateway

Updates providers/kilo/models/ to match https://api.kilo.ai/api/gateway/models.

Summary

Metric	Value
Models added	6
Models removed	5
Upstream count	345
Local count	345 (was 344)

Models Added (6)

ID	Name	Input	Output	Ctx → Out	Key details
google/gemini-3.5-flash	Google: Gemini 3.5 Flash	$1.50	$9.00	1M → 65K	multimodal (audio+image+pdf+text+video), reasoning=$9, cache_read=$0.15, cache_write=$0.08333
qwen/qwen3.7-max	Qwen: Qwen3.7 Max	$1.625	$4.875	1M → 65K	text-only, cache_read=$0.1625, cache_write=$2.03125
stealth/claude-opus-4.6	Stealth: Claude Opus 4.6 (20% off)	$4.00	$20.00	1M → 128K	cache_read=$0.40, cache_write=$5.00, third-party Anthropic resale
stealth/claude-opus-4.7	Stealth: Claude Opus 4.7 (20% off)	$4.00	$20.00	1M → 128K	cache_read=$0.40, cache_write=$5.00, third-party Anthropic resale
stealth/claude-sonnet-4.6	Stealth: Claude Sonnet 4.6 (20% off)	$2.40	$12.00	1M → 64K	cache_read=$0.24, cache_write=$3.00, third-party Anthropic resale
x-ai/grok-build-0.1	xAI: Grok Build 0.1	$1.00	$2.00	256K → 256K	cache_read=$0.20, coding-optimized model

Nuances in Added Models

• New provider directory: stealth/ is a new provider under Kilo containing 3 third-party Anthropic Claude resale models at 20% discount ($4/$20 vs $5/$25 for Opus models, $2.4/$12 vs $3/$15 for Sonnet).
• google/gemini-3.5-flash: "file" input modality mapped to "pdf" per convention; internal_reasoning priced at $9 (same as output).
• x-ai/grok-build-0.1: max_completion_tokens = null falls back to context_length (256K). Supports structured_outputs upstream but omitted per Kilo convention.
• qwen/qwen3.7-max: Text-only model with no attachment support. Supports reasoning despite zero internal_reasoning upstream pricing.

Models Removed (5)

Model ID	Notes
alibaba/tongyi-deepresearch-30b-a3b	Deprecated upstream
arcee-ai/trinity-large-preview	Deprecated upstream
deepseek/deepseek-v4-flash:free	Free tier discontinued upstream
stepfun/step-3.5-flash:free	Free tier discontinued upstream
x-ai/grok-code-fast-1:optimized:free	Free tier discontinued upstream

Sync Methodology

All field mapping conventions follow established patterns from previous Kilo gateway sync PRs:

1. Pricing: Multiply upstream per-token prices ×1,000,000 → per-1M-tokens TOML values
2. Modalities: "file" → "pdf", arrays sorted alphabetically, duplicates deduplicated
3. max_completion_tokens = null: Falls back to context_length
4. Cache pricing: Only included when upstream value is non-zero
5. Free models: All cost fields set to 0, cache fields omitted
6. temperature: false when absent from upstream supported_parameters
7. reasoning: true when include_reasoning or reasoning in upstream supported_parameters
8. tool_call: true when tools in upstream supported_parameters
9. structured_output: Never emitted in Kilo TOML files
10. knowledge: Never emitted — not present in upstream data
11. release_date: Derived from upstream created Unix timestamp
12. attachment: true iff input modalities include non-text modality

Sync Check Script

The following Python script (check.py) was used to detect differences between the local TOML files and the upstream Kilo API:

#!/usr/bin/env python3
"""
Usage:
    python3 check.py

Compares providers/kilo/models/**/*.toml against the live Kilo gateway API
and prints which models need to be added or removed.

Run from the repo root.
"""

import json
import subprocess
from pathlib import Path

GATEWAY_URL = "https://api.kilo.ai/api/gateway/models"
MODELS_DIR  = Path("providers/kilo/models")


def fetch_upstream() -> set[str]:
    result = subprocess.run(
        ["curl", "-sS", GATEWAY_URL],
        capture_output=True, text=True, check=True,
    )
    data = json.loads(result.stdout)
    return {m["id"] for m in data["data"]}


def fetch_local() -> set[str]:
    return {
        str(p.relative_to(MODELS_DIR)).removesuffix(".toml")
        for p in MODELS_DIR.rglob("*.toml")
    }


def main() -> None:
    print("Fetching upstream model list …")
    upstream = fetch_upstream()
    local    = fetch_local()

    to_add    = sorted(upstream - local)
    to_remove = sorted(local - upstream)

    if to_add:
        print(f"\n### Models to ADD ({len(to_add)}) — present upstream, missing locally ###")
        for mid in to_add:
            print(f"  + {mid}")
    else:
        print("\nNo models to add.")

    if to_remove:
        print(f"\n### Models to REMOVE ({len(to_remove)}) — local only, gone from gateway ###")
        for mid in to_remove:
            print(f"  - {mid}")
    else:
        print("No models to remove.")

    print(f"\nSummary: upstream={len(upstream)}  local={len(local)}  "
          f"to_add={len(to_add)}  to_remove={len(to_remove)}")

    if not to_add and not to_remove:
        print("✓ Local tree is in sync with the upstream gateway.")


if __name__ == "__main__":
    main()

Run it with python3 check.py from the repo root.