tools/test_glimmer_fmt.py

"""Byte-equality check of tools/glimmer_fmt.py against chat_template.jinja.

    python tools/test_glimmer_fmt.py --model-dir /workspace/gguf/ref/meta

Renders every case through transformers' jinja sandbox and through the port and
asserts the two strings are identical.  If this fails, the calibration markup is
wrong and the build must not run.
"""

from __future__ import annotations

import argparse
import json
import sys

import glimmer_fmt as G

DATE = "2026-08-10"
CUTOFF = "2026-01-04"

WEATHER = {
    "type": "function",
    "function": {
        "name": "weather.get_forecast",
        "description": "Get the weather forecast for a location.",
        "parameters": {
            "type": "object",
            "properties": {
                "city": {"type": "string", "description": "City name"},
                "days": {"type": "integer", "description": "How many days ahead"},
            },
            "required": ["city"],
        },
    },
}
SHELL = {
    "type": "function",
    "function": {
        "name": "shell.run",
        "description": "Run a shell command and return stdout/stderr.",
        "parameters": {
            "type": "object",
            "properties": {
                "cmd": {"type": "string"},
                "timeout_s": {"type": "integer"},
                "env": {"type": "object"},
                "cwd": {"type": "string"},
                "capture_stderr": {"type": "boolean"},
            },
            "required": ["cmd"],
        },
    },
}


def cases() -> list[dict]:
    out = []

    # 1. plain turn, no system, no tools
    out.append(dict(name="plain", messages=[
        {"role": "user", "content": "Hi."},
        {"role": "assistant", "content": "Hello."},
    ]))

    # 2. explicit system + every reasoning strength
    for rs in G.REASONING_STRENGTHS:
        out.append(dict(name=f"system-{rs}", reasoning_strength=rs, messages=[
            {"role": "system", "content": "You are a build engineer."},
            {"role": "user", "content": "Explain rpath."},
            {"role": "assistant", "reasoning_content": "rpath is a linker search path.",
             "content": "It is a search path baked into the binary."},
        ]))

    # 3. one-argument tool call plus its result
    out.append(dict(name="tool-1arg", tools=[WEATHER], messages=[
        {"role": "user", "content": "Weather in Kyoto?"},
        {"role": "assistant", "tool_calls": [
            {"id": "c1", "type": "function", "function": {
                "name": "weather.get_forecast", "arguments": {"city": "Kyoto"}}}]},
        {"role": "tool", "name": "weather.get_forecast", "content": "18C, clear"},
        {"role": "assistant", "content": "18 degrees and clear."},
    ]))

    # 4. five arguments, nested JSON, booleans, null
    out.append(dict(name="tool-5arg-nested", tools=[SHELL], messages=[
        {"role": "user", "content": "Build it."},
        {"role": "assistant", "tool_calls": [
            {"id": "c2", "type": "function", "function": {
                "name": "shell.run",
                "arguments": {
                    "cmd": "cmake --build build -j",
                    "timeout_s": 900,
                    "env": {"CC": "clang", "FLAGS": ["-O2", "-g"], "NESTED": {"a": [1, 2]}},
                    "cwd": None,
                    "capture_stderr": True,
                }}}]},
        {"role": "tool", "name": "shell.run", "content": "ninja: build stopped."},
    ]))

    # 5. two tool calls in one assistant turn -> first gets EOM
    out.append(dict(name="tool-parallel", tools=[WEATHER, SHELL], messages=[
        {"role": "user", "content": "Both please."},
        {"role": "assistant", "tool_calls": [
            {"id": "a", "type": "function", "function": {
                "name": "weather.get_forecast", "arguments": {"city": "Oslo", "days": 3}}},
            {"id": "b", "type": "function", "function": {
                "name": "shell.run", "arguments": {"cmd": "uptime"}}},
        ]},
    ]))

    # 6. tool name resolved through tool_call_id
    out.append(dict(name="tool-by-id", tools=[WEATHER], messages=[
        {"role": "user", "content": "?"},
        {"role": "assistant", "tool_calls": [
            {"id": "xyz", "type": "function", "function": {
                "name": "weather.get_forecast", "arguments": {"city": "Lima"}}}]},
        {"role": "tool", "tool_call_id": "xyz", "content": "rain"},
    ]))

    # 7. consecutive same-role user messages -> EOM boundary logic
    out.append(dict(name="consecutive", messages=[
        {"role": "user", "content": "one"},
        {"role": "user", "content": "two"},
        {"role": "assistant", "content": "ok"},
    ]))

    # 8. image part
    out.append(dict(name="image", messages=[
        {"role": "user", "content": [
            {"type": "text", "text": "What is this? "},
            {"type": "image"},
        ]},
        {"role": "assistant", "content": "A plot."},
    ]))

    # 9. generation prompt
    out.append(dict(name="genprompt", add_generation_prompt=True, messages=[
        {"role": "user", "content": "go"},
    ]))

    # 10. non-user recipient
    out.append(dict(name="recipient", messages=[
        {"role": "user", "content": "think"},
        {"role": "assistant", "recipient": "self", "content": "planning", "end_turn": False},
        {"role": "assistant", "content": "done"},
    ]))
    return out


def main() -> int:
    ap = argparse.ArgumentParser()
    ap.add_argument("--model-dir", required=True,
                    help="directory holding chat_template.jinja + tokenizer_config.json")
    args = ap.parse_args()

    from transformers import AutoTokenizer
    tok = AutoTokenizer.from_pretrained(args.model_dir)

    failures = 0
    for c in cases():
        kwargs = dict(
            tools=c.get("tools"),
            reasoning_strength=c.get("reasoning_strength", G.DEFAULT_REASONING),
            knowledge_cutoff=CUTOFF,
            current_date=DATE,
            add_generation_prompt=c.get("add_generation_prompt", False),
        )
        want = tok.apply_chat_template(c["messages"], tokenize=False, **kwargs)
        got = G.render(c["messages"], **kwargs)
        if want != got:
            failures += 1
            print(f"FAIL {c['name']}")
            for i, (a, b) in enumerate(zip(want, got)):
                if a != b:
                    print(f"  first difference at char {i}")
                    print(f"    template: {want[max(0,i-60):i+60]!r}")
                    print(f"    port    : {got[max(0,i-60):i+60]!r}")
                    break
            else:
                print(f"  length {len(want)} vs {len(got)}")
                print(f"    template tail: {want[len(got)-60:][:160]!r}")
                print(f"    port tail    : {got[len(want)-60:][:160]!r}")
        else:
            print(f"ok   {c['name']}  ({len(got)} chars)")

    print(f"\n{len(cases()) - failures}/{len(cases())} cases byte-identical")
    return 1 if failures else 0


if __name__ == "__main__":
    sys.exit(main())