bug: max_tokens not clamped to max_output_size for OpenAI-compatible providers — 400 on models with output limit < context window (v0.20.1)

Bug: max_tokens not clamped to max_output_size for OpenAI-compatible providers — causes 400 on models with output limits lower than context window

What version of Kimi Code is running?

0.20.1

Which open platform/subscription were you using?

Self-configured third-party provider (Ollama Cloud — OpenAI-compatible API)

Which model were you using?

ollama-cloud/deepseek-v4-pro

What platform is your computer?

Windows 11 x64

What issue are you seeing?

The LLM request fails immediately with a 400 error on the very first turn:

400 "max_tokens (131072) exceeds model's maximum output tokens (65536) for model deepseek-v4-pro"

The session is completely blocked — no retry can recover because the max_tokens value is computed incorrectly every time.

What steps can reproduce the bug?

1. Configure a model via an OpenAI-compatible provider where the model's actual max output tokens (enforced server-side by the API) is less than the remaining context window:

[providers.ollama-cloud]
type = "openai"
api_key = "..."
base_url = "https://ollama.com/v1"

[models."ollama-cloud/deepseek-v4-pro"]
provider = "ollama-cloud"
model = "deepseek-v4-pro"
max_context_size = 1048576
max_output_size = 1048576
capabilities = [ "thinking", "tool_use" ]

2. Set this model as default: default_model = "ollama-cloud/deepseek-v4-pro"
3. Start a new session and send any prompt.
4. The request fails with 400 because Kimi Code sends max_tokens=131072 while the API enforces a hard limit of 65536 for this model.

Root Cause Analysis

This is the same underlying bug tracked in #306 and #834, persisting through v0.20.1 despite the partial fix in #1131.

Factor 1: max_output_size is ignored for openai providers (#306)

In provider-manager.ts, maxOutputSize is passed to anthropic providers as defaultMaxTokens, but not to openai providers. So the configured max_output_size has zero effect on the max_tokens sent to the API.

Factor 2: The #1131 fix caps to remaining context window, not to max_output_size

The fix in v0.20.1 (f1c8175) caps completion tokens to the remaining context window. However, when max_output_size is not threaded through (Factor 1), the cap falls back to max_context_tokens (1,048,576 in this case). The remaining context window (131,072) is still larger than the API's actual output limit (65,536).

The result: max_tokens = min(remaining_context, max_context_size) = 131072, which exceeds the API's hard limit of 65536.

Expected Behavior

Kimi Code should clamp max_tokens to min(max_output_size, remaining_context_window) for all provider types, including openai. If max_output_size is configured, it should be the hard cap.

Additionally, Kimi Code should gracefully handle the case where the configured max_output_size itself exceeds the API's actual limit — ideally by catching the 400 and retrying with a smaller value, or by clamping max_tokens more conservatively.

Impact

• All OpenAI-compatible models where the actual API output limit is less than the configured max_context_size are affected.
• This is increasingly common with third-party providers (Ollama Cloud, OpenRouter, etc.) that aggregate models with varying output limits.
• The session is completely blocked — the error is deterministic and no retry helps.

Workaround

Set the environment variable:

export KIMI_MODEL_MAX_COMPLETION_TOKENS=65536

This acts as a hardCap and overrides the broken max_output_size path.

Alternatively, there is no config-only workaround because max_output_size is ignored for openai providers.

Related Issues

• #306 — Original bug report: max_output_size not consumed for openai providers. Open since v0.7.0.
• #834 — Same bug in the compaction path. Open since v0.16.0.
• #476 — Related: max_context_size being sent as max_output_size during compaction.
• #1131 (v0.20.1) — Partial fix: caps to remaining context window but does not address the max_output_size ignoring bug for openai providers.

Session Evidence

Session log (kimi-code.log):

2026-06-26T16:12:51.345Z INFO  llm config  provider=openai model=deepseek-v4-pro modelAlias=ollama-cloud/deepseek-v4-pro thinkingEffort=high systemPromptChars=36103 toolCount=460
2026-06-26T16:12:51.345Z INFO  llm request  turnStep=0.1
2026-06-26T16:12:52.446Z WARN  llm request failed  attempt=1/3 model=deepseek-v4-pro errorName=APIStatusError errorMessage="400 \"max_tokens (131072) exceeds model's maximum output tokens (65536)\"" statusCode=400
2026-06-26T16:12:52.446Z ERROR turn failed  turnId=0
  APIStatusError: 400 "max_tokens (131072) exceeds model's maximum output tokens (65536) for model deepseek-v4-pro"
    at normalizeAPIStatusError (.../intermediates/main.cjs:1185:9)
    at convertOpenAIError (.../intermediates/main.cjs:52368:10)
    at OpenAILegacyChatProvider.generate (.../intermediates/main.cjs:53249:10)

Start in provider-manager.ts and inspect how maxOutputSize is passed to OpenAI-compatible providers, then trace the v0.20.1 change from f1c8175 that caps against the remaining context window. Reproduce the Ollama Cloud configuration from the issue and verify that max_tokens is clamped to the smaller of max_output_size and the remaining context window for OpenAI providers.