Skip to content
Python

OpenAI-compatible path names an output ceiling instead of trusting server defaults

Rohit C Prasad44e0e8566f3b
diff --git a/coworker/providers/openai_provider.py b/coworker/providers/openai_provider.py
index fd733f93..48e6b6a0 100644
--- a/coworker/providers/openai_provider.py
+++ b/coworker/providers/openai_provider.py
@@ -82,6 +82,12 @@ def _strip_foreign_sidecars(messages: list[dict[str, Any]]) -> list[dict[str, An
_MAX_TOKENS_ERROR = "'max_tokens' is not supported"
+# Ceiling, not a spend target — same rationale as the Anthropic provider's default: a
+# coworker writing a report ships the whole file inside one tool call's arguments, and
+# compat servers left to their OWN defaults cap completions absurdly low (observed
+# 2026-08-15: Together defaulted Kimi K3 to ~2k tokens — every ~5KB write truncated).
+DEFAULT_MAX_TOKENS = 32000
+
def _param_fix_retry(kwargs: dict[str, Any], exc: Exception) -> dict[str, Any]:
"""Kwargs for the one retry an unsupported-parameter error earns, or re-raise.
@@ -103,6 +109,14 @@ def _param_fix_retry(kwargs: dict[str, Any], exc: Exception) -> dict[str, Any]:
fixed = dict(kwargs)
fixed.pop("stream_options")
return fixed
+ if ("max_tokens" in msg or "max_new_tokens" in msg) and "max_tokens" in kwargs:
+ # Our 32k default exceeded this model's completion limit (each server words the
+ # 400 differently, so no number parsing) — drop the param and retry on the
+ # server's own default rather than surfacing the 400. Worst case is exactly
+ # yesterday's behavior; best case the server allows far more once asked.
+ fixed = dict(kwargs)
+ fixed.pop("max_tokens")
+ return fixed
raise exc
@@ -179,11 +193,13 @@ class OpenAIProvider(ProviderClient):
}
if tools:
kwargs["tools"] = tools
+ kwargs.setdefault("max_tokens", DEFAULT_MAX_TOKENS)
_pin_reasoning_effort(kwargs)
client = self._ensure_client()
- # Up to two param-fix retries: effort and max_tokens can BOTH need fixing.
- for _ in range(2):
+ # Up to three param-fix retries: effort, the max_tokens rename, and the
+ # max_tokens over-limit drop can ALL need fixing on one call.
+ for _ in range(3):
try:
response = client.chat.completions.create(**kwargs)
break
@@ -227,6 +243,7 @@ class OpenAIProvider(ProviderClient):
}
if tools:
kwargs["tools"] = tools
+ kwargs.setdefault("max_tokens", DEFAULT_MAX_TOKENS)
_pin_reasoning_effort(kwargs)
client = self._ensure_client()
@@ -236,8 +253,9 @@ class OpenAIProvider(ProviderClient):
finish_reason = None
usage: Optional[TokenUsage] = None
- # Up to two param-fix retries: effort and max_tokens can BOTH need fixing.
- for _ in range(2):
+ # Up to three param-fix retries: effort, the max_tokens rename, and the
+ # max_tokens over-limit drop can ALL need fixing on one call.
+ for _ in range(3):
try:
chunks = client.chat.completions.create(**kwargs)
break
diff --git a/tests/test_providers.py b/tests/test_providers.py
index 943c6e4c..cd068b67 100644
--- a/tests/test_providers.py
+++ b/tests/test_providers.py
@@ -456,3 +456,47 @@ def test_complete_picks_up_reasoning_content():
provider = OpenAIProvider(client=_FakeClient(SimpleNamespace(choices=[choice])))
turn = provider.complete(model="deepseek-v4-pro", messages=[{"role": "user", "content": "x"}])
assert turn.text == "Answer" and turn.reasoning == "deep thought"
+
+
+def test_default_max_tokens_injected_and_caller_setting_wins():
+ """Compat servers left to their OWN defaults cap completions absurdly low
+ (owner-hit 2026-08-15: Together defaulted Kimi K3 to ~2k tokens, so every report
+ write truncated mid-arguments). The request always names a ceiling now."""
+ from coworker.providers.openai_provider import DEFAULT_MAX_TOKENS
+
+ client = _FakeClient(_response(content="ok"))
+ provider = OpenAIProvider(client=client)
+ provider.complete(model="kimi-k3", messages=[])
+ assert client.chat.completions.calls[0]["max_tokens"] == DEFAULT_MAX_TOKENS
+
+ client2 = _FakeClient(_response(content="ok"))
+ provider2 = OpenAIProvider(client=client2)
+ provider2.complete(model="kimi-k3", messages=[], max_tokens=512)
+ assert client2.chat.completions.calls[0]["max_tokens"] == 512
+
+
+def test_over_limit_max_tokens_is_dropped_and_retried():
+ """A model whose completion limit sits below our default must not surface the 400:
+ drop the param, retry on the server's own default (yesterday's behavior, at worst)."""
+
+ class _LimitRejecting:
+ def __init__(self, response):
+ self._response = response
+ self.calls: list[dict] = []
+
+ def create(self, **kwargs):
+ self.calls.append(kwargs)
+ if "max_tokens" in kwargs:
+ raise RuntimeError(
+ "Error code: 400 - max_tokens must be at most 8193 for this model"
+ )
+ return self._response
+
+ client = _FakeClient(_response(content="ok"))
+ client.chat.completions = _LimitRejecting(_response(content="ok"))
+ provider = OpenAIProvider(client=client)
+
+ turn = provider.complete(model="tiny-model", messages=[])
+ calls = client.chat.completions.calls
+ assert turn.text == "ok" and len(calls) == 2
+ assert "max_tokens" not in calls[1]