[Engine] Tool-call arguments ignore the schema type: booleans come back as the strings "True"/"False"

serve/frontend.pyを読み、issueで変換経路の候補として挙げられているcoerce_value(value, declared)を調べてください。json.loadsへのフォールバックより前に、宣言されたブール値がどのように処理されるかを確認し、関連するツール呼び出しテストを実行してください。issueにはテストファイルの指定がありません。完了条件は、モデルがTrueまたはFalseと記述した場合、ツール引数がJSONのブール値として出力され、宣言された文字列は文字列のままであることです。

[Engine] Tool-call arguments ignore the schema type: booleans come back as the strings "True"/"False"

Area: [Engine] (the server and API) · How: Docker · Reproduction status: Reproduced locally
(0.2.7 / 0.2.9 / 0.2.11; the same code path on current main is read, not run — see Evidence)

Problem

For a declared tool whose parameter is {"type": "boolean"}, the OpenAI-compatible response carries the
argument as a string: "formatted": "True" instead of "formatted": true. The model's chosen value is
right; only the JSON type is wrong.

Environment

Docker ghcr.io/niko1221/strata:0.2.7-cuda13 (engine 0.1.52) — still present in 0.2.9 (engine 0.1.55) and
0.2.11 (engine 0.1.57). Our image is our own build of the upstream release tag (buildah, upstream Dockerfile
unchanged, the release's signed cuda13 engine package); same code as upstream's image.
Linux · RTX PRO 5000 48 GB, one card (the 2026-11-10 capture below; the 0.2.7 and 0.2.9 observations were on
two cards: the same 48 GB plus an RTX 5060 Ti 16 GB) · 62.7 GB RAM · Qwen3.8-Flash-Next
FAMILY=nvfp4 MODEL=NVFP4-NVIDIA, 225,280 context, 2 slots. Seen with MODEL=NVFP4 (GPTQ),
NVFP4-QAD and FAMILY=unsloth MODEL=UD-Q4_K_XL too, so it is not pack-specific.

Reproduce

1. Start any Qwen3.8-Flash-Next pack and POST a request whose tools declare a boolean parameter. Keep
   "reasoning_effort": "none" — with thinking on the model usually spells it true, so the bug shows on a
   minority of requests; thinking off, this prompt reproduced 8 times in 8:

   curl -s http://<host>/v1/chat/completions -H 'content-type: application/json' -d '{
     "model": "qwen3.8-flash-next-nvfp4-nvfp4-nvidia",
     "reasoning_effort": "none",
     "messages": [{"role": "user", "content": "Get the prime factors of 450, formatted."}],
     "tools": [{"type": "function", "function": {
       "name": "get_prime_factors",
       "description": "Return the prime factors of a number.",
       "parameters": {"type": "object", "properties": {
         "number": {"type": "integer"},
         "formatted": {"type": "boolean", "description": "Return the factors formatted as a string"}},
       "required": ["number", "formatted"]}}}],
     "tool_choice": "auto"}'
   

2. Read choices[0].message.tool_calls[0].function.arguments.

What we got (captured live 2026-11-10, engine 0.1.57)

Same request, only reasoning_effort and temperature varied; arguments is
choices[0].message.tool_calls[0].function.arguments verbatim:

attempts	arguments returned	formatted type
8 (thinking off via "reasoning_effort": "none", temp 0.7 and 0.0)	{"number":450,"formatted":"True"}	string, 8/8
4 (thinking on, server default, temp 0.7)	1× {"number":450,"formatted":"True"}, 3× {"number":450,"formatted":true}	string 1, boolean 3

So the failure is the model's spelling, not the sampling settings: when it writes True the server hands the
text through untouched, when it writes true the value arrives correctly typed. Thinking off, this prompt
fails every time; thinking on it is intermittent, which is why a single request can look fine. Full body of the
failing case is pasted under Logs; the correct case for contrast:

"tool_calls": [{"id": "call_fcd6accc1a144cd7913b8250", "type": "function", "function": {
  "name": "get_prime_factors", "arguments": "{\"number\":450,\"formatted\":\"True\"}"}}]     -- actual
"tool_calls": [{"id": "call_…",  "type": "function", "function": {
  "name": "get_prime_factors", "arguments": "{\"number\":450,\"formatted\":true}"}}]         -- expected

Expected

Each <parameter=NAME>VALUE</parameter> value is converted by its declared JSON-schema type (boolean,
integer, number, array/object) before it is emitted in tool_calls[].function.arguments, so a boolean
parameter is a JSON true/false.

Actual

"formatted": "True" — a string where the schema says boolean.

BFCL v4 simple_python (400 entries, thinking off) counts these as type_error:simple. Entries affected, by
pack, on the same config: 15/400 NVFP4-NVIDIA (90.25% as scored → 94.0% counted correct), 25/400
NVFP4-QAD (88.75% → 95.0%), 20/400 NVFP4 GPTQ, 24/400 Unsloth UD-Q4_K_XL. Example, BFCL simple_python_17:

{"get_prime_factors": {"number": 450, "formatted": "True"}}   →  Expected type boolean, got str

Why it matters

Clients with strict tool validation (MCP servers, typed agent frameworks) reject the call. Worse, a client
that is lenient treats the non-empty string "False" as truthy, so the behaviour silently inverts —
the opposite of what the model asked for.

Evidence

• Reproduced locally: the table above, run against our deployment on 2026-11-10, and seen on 0.2.7, 0.2.9 and
  0.2.11 before that (BFCL suite plus direct requests).

• Suspected path (read on main 2026-11-05 and again 2026-11-10, not run against a 0.2.13 build):
  serve/frontend.py coerce_value(value, declared) — it does receive the declared type, but for anything
  but string it only tries json.loads(value); JSON spells booleans true/false, so the model's
  Python-style True fails to parse, and the fallback lenient path only fires when the text starts with
  [, { or a backtick, so the raw text is returned unchanged. As of main on 2026-11-10 that function is
  unchanged (line 1118), and the 0.2.12 and 0.2.13 release notes mention neither the symptom nor a fix.

  if declared == "string":
      return value
  try:
      v = json.loads(value)
  except ValueError:
      if not value.lstrip().startswith(("[", "{", "`")) or ...:
          return value          # <-- "True" leaves here, still a string
  

• Precedent: SGLang's Qwen3-coder detector lowercases the text first for boolean/bool
  (param_value.lower() == "true"); vLLM's and SGLang's qwen3_xml parsers coerce by schema type.

Suggested fix (not tested as a patch — offered as the shape of it)

# in coerce_value, before json.loads(value)
if declared in ("boolean", "bool") and value.strip().lower() in ("true", "false"):
    return value.strip().lower() == "true"

Safe to do here because the declared == "string" branch already returns before it, so a string parameter
whose value is literally True is untouched.

Searched first (2026-11-10)

boolean argument, True False string argument, schema type argument, coerce, type_error,
tool call parameter True. Closest hits, all distinct: #4037 (Chat/Messages coerce boolean token
limits — the request side, not tool arguments), #4630 (array/object parameter with a raw newline),
#4751 / #4071 (calls emitted as text or not read as a call of a declared tool), #5281 (streamed
MCP call landing in content), #4622 (undeclared tool-call names). Happy to move this into one of those
threads if a maintainer considers it a duplicate.

Workaround we use

Nothing server-side: our benchmark harness counts those entries separately so the number is not silently
wrong, and clients have to be told not to trust truthiness on tool arguments.

Logs

The failing response, verbatim (nothing but token counts and timings in it — no keys, no private text):

{
 "id": "chatcmpl-575bfad3a8c14fb68dad4de7",
 "object": "chat.completion",
 "created": 1794343744,
 "model": "qwen3.8-flash-next-nvfp4-nvfp4-nvidia",
 "choices": [
  {
   "index": 0,
   "message": {
    "role": "assistant",
    "content": null,
    "reasoning_content": "We need answer user: \"Get the prime factors of 450, formatted.\" Need call get_prime_factors with number 450 formatted true. Need final likely factorization. Let's compute? 450=2*3^2*5^2. Formatted maybe string. Need call tool.\n",
    "tool_calls": [
     {
      "id": "call_fcd6accc1a144cd7913b8250",
      "type": "function",
      "function": {
       "name": "get_prime_factors",
       "arguments": "{\"number\":450,\"formatted\":\"True\"}"
      }
     }
    ]
   },
   "finish_reason": "tool_calls"
  }
 ],
 "usage": {
  "prompt_tokens": 335,
  "completion_tokens": 110,
  "total_tokens": 445,
  "prompt_tokens_details": {
   "cached_tokens": 330
  }
 },
 "timings": {
  "cache_n": 330,
  "prompt_n": 5,
  "prompt_ms": 109.0,
  "prompt_per_token_ms": 21.8,
  "prompt_per_second": 45.9,
  "predicted_n": 110,
  "predicted_ms": 1210.3,
  "predicted_per_token_ms": 11.003,
  "predicted_per_second": 90.9,
  "draft_n": 85,
  "draft_n_accepted": 70
 }
}