项目文件夹

文件
Ali Khokhar 6455c63e1d Make reasoning policy provider-neutral and client-aware (#1148)
## Problem

FCC reduced reasoning to global and route booleans, mixing client
intent, configuration, provider wire capabilities, output visibility,
and history replay. That discarded named client efforts, encouraged
model-name checks, and made provider behavior inconsistent.

## Changes

| Before | After |
| --- | --- |
| Admin exposed global and route thinking toggles. | Admin exposes
**Off**, **From client**, **Low**, **Medium**, **High**, **X-High**, and
**Max**; Fable, Opus, Sonnet, and Haiku also expose **Inherit**. |
| Request intent was repeatedly reduced to a boolean across routing and
providers. | The application boundary resolves one immutable
`ReasoningPolicy` with independent control, named effort, and exact
positive token budget. |
| Provider adapters could infer reasoning behavior from upstream model
names or versions. | Provider profiles translate only documented
provider-wide wire capabilities; architecture and contributor rules
prohibit model-specific reasoning branches. |
| Gateway reasoning controls were ad hoc. |
[OpenRouter](https://openrouter.ai/docs/guides/best-practices/reasoning-tokens)
and [Vercel AI
Gateway](https://vercel.com/docs/ai-gateway/models-and-providers) use
documented reasoning objects, including exact budgets where
representable. |
| Named effort forwarding was inconsistent or absent. |
[Gemini](https://ai.google.dev/gemini-api/docs/openai),
[Ollama](https://docs.ollama.com/api/openai-compatibility), [LM
Studio](https://lmstudio.ai/changelog/lmstudio-v0.4.8),
[Fireworks](https://docs.fireworks.ai/guides/querying-text-models/reasoning),
[Cohere](https://docs.cohere.com/docs/compatibility-api),
[Wafer](https://docs.wafer.ai/serverless/api-reference),
[Groq](https://console.groq.com/docs/reasoning),
[Cerebras](https://inference-docs.cerebras.ai/capabilities/reasoning),
[SambaNova](https://docs.sambanova.ai/docs/api-reference/chat-completions/create-chat-based-completion),
and
[Mistral](https://docs.mistral.ai/studio-api/conversations/reasoning)
receive their documented named vocabularies with explicit provider-owned
downgrades. |
| Boolean thinking controls were mixed into shared conversion. |
[DeepSeek](https://api-docs.deepseek.com/guides/thinking_mode/),
[Kimi](https://platform.kimi.ai/docs/guide/use-kimi-k2-thinking-model),
[Z.ai](https://docs.z.ai/guides/capabilities/thinking-mode), [Cloudflare
Workers
AI](https://developers.cloudflare.com/changelog/post/2026-04-20-kimi-k2-6-workers-ai/),
and [NVIDIA
NIM](https://docs.nvidia.com/nim/large-language-models/1.15.0/thinking-budget-control.html)
use provider-owned thinking-object or chat-template controls. |
| Effort names and output limits could become fabricated reasoning
budgets. | Exact budgets remain exact and are forwarded only through
documented fields for OpenRouter, Fireworks, LM Studio, NIM, and
[llama.cpp](https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md);
named efforts and output limits are never converted into token budgets.
|
| New-turn reasoning and prior-turn replay shared one switch. | Every
profile independently declares native reasoning replay, `<think>` tag
replay, provider-specific replay, or no replay; **Off** suppresses new
reasoning output without corrupting required history. |
| Providers without a stable generic compute control received guessed
controls. |
[MiniMax](https://platform.minimax.io/docs/api-reference/text-openai-api)
requests split output only, while [GitHub
Models](https://docs.github.com/en/rest/models/inference), [Hugging Face
Inference
Providers](https://huggingface.co/docs/inference-providers/en/tasks/chat-completion),
Codestral, and OpenCode keep provider defaults and use only their
explicit replay profile. |
| OpenAI Responses effort became a lossy Anthropic thinking boolean. |
Responses preserves `reasoning.effort` through `output_config`, then
resolves it through the same application policy as Messages without
inventing a budget. |
| Legacy booleans remained the persisted contract. | FCC-owned dotenv
files migrate to typed `REASONING_*` values, explicit env files receive
an actionable warning, documentation describes the ownership boundary,
and the package advances to 4.8.0. |
| Reasoning behavior was covered by scattered boolean assertions. | New
policy, routing, encoder, provider, Admin, migration, Responses, and
smoke contracts pass all five local CI checks: 2,368 tests passed, 40
skipped; 92 smoke tests collect and both live config migration checks
pass. |

<!-- greptile_comment -->

<details open><summary><h3>Greptile Summary</h3></summary>

This PR makes reasoning policy client-aware and independent of provider
model names. The main changes are:

- Adds one immutable reasoning policy resolved at the application
boundary.
- Adds typed root and route reasoning settings with Admin UI support.
- Moves wire controls and history replay behavior into provider
profiles.
- Migrates owned dotenv files from legacy thinking booleans.
- Expands provider, routing, migration, API, and smoke coverage.
</details>

<h3>Confidence Score: 5/5</h3>

This looks safe to merge.

No blocking issues found in the changed code.

<details><summary><h3><a href="https://www.greptile.com/trex"><img
alt="T-Rex"
src="https://greptile-static-assets.s3.amazonaws.com/trex/trex_green.svg"
height="20" align="absmiddle"></a> T-Rex Logs</h3></summary>

**What T-Rex did**
- Ran the contract-validation test suite with the specified test
modules, and the tests reported 78 passed in 1.53s with exit code 0.
- Reviewed the complete captured output artifact
reasoning-contract-02-after.log to verify the final test outcomes and
successful contract validation.

<a
href="https://app.greptile.com/trex/runs/14792858/artifacts"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://greptile-static-assets.s3.amazonaws.com/badges/ViewAllArtifactsDark.svg?v=4"><source
media="(prefers-color-scheme: light)"
srcset="https://greptile-static-assets.s3.amazonaws.com/badges/ViewAllArtifacts.svg?v=4"><img
alt="View all artifacts"
src="https://greptile-static-assets.s3.amazonaws.com/badges/ViewAllArtifacts.svg?v=4"></picture></a>

<sub><a href="https://www.greptile.com/trex"><img alt="T-Rex"
src="https://greptile-static-assets.s3.amazonaws.com/trex/trex_green.svg"
height="14" align="absmiddle"></a> Ran code and verified through
T-Rex</sub>
</details>

<details open><summary><h3>Important Files Changed</h3></summary>

| Filename | Overview |
|----------|----------|
| src/free_claude_code/config/env_migrations.py | Migrates legacy
reasoning booleans in owned dotenv files and warns for explicit
environment files. |
| src/free_claude_code/application/reasoning.py | Resolves client
controls and configured preferences into one provider-neutral reasoning
policy. |
| src/free_claude_code/application/routing.py | Carries route-level
reasoning preferences into request-scoped policy resolution. |
| src/free_claude_code/providers/openai_chat/reasoning.py | Provides
shared provider encoders for reasoning controls and replay behavior. |

</details>

<sub>Reviews (2): Last reviewed commit: ["chore: release reasoning
controls as
4.8..."](https://github.com/alishahryar1/free-claude-code/commit/9d4be767f7dbdca5709474012f43dcdc6f4347e3)
| [Re-trigger
Greptile](https://app.greptile.com/api/retrigger?id=44984039)</sub>

<!-- /greptile_comment -->
2026-07-16 19:57:12 -07:00

100 行
3.1 KiB
Python

"""Resolve client reasoning input and FCC configuration exactly once."""
from collections.abc import Mapping
from typing import Any
from free_claude_code.config.reasoning import ReasoningPreference
from free_claude_code.core.anthropic.models import MessagesRequest, ThinkingConfig
from free_claude_code.core.reasoning import (
ReasoningControl,
ReasoningEffort,
ReasoningPolicy,
)
def resolve_reasoning_policy(
request: MessagesRequest,
preference: ReasoningPreference,
) -> ReasoningPolicy:
"""Apply one resolved configuration preference to the client request."""
if preference is ReasoningPreference.INHERIT:
raise ValueError("Reasoning preference must be resolved before application.")
if preference is ReasoningPreference.OFF:
return ReasoningPolicy.off()
if preference is not ReasoningPreference.CLIENT:
return ReasoningPolicy.on(effort=ReasoningEffort(preference.value))
return client_reasoning_policy(request)
def client_reasoning_policy(request: MessagesRequest) -> ReasoningPolicy:
"""Return the lossless reasoning intent expressed by one client request."""
budget_tokens = _positive_budget(request.thinking)
thinking_control = _thinking_control(
request.thinking,
budget_tokens=budget_tokens,
)
effort, effort_disables = _output_effort(request.output_config)
if effort_disables:
return ReasoningPolicy.off()
if thinking_control is ReasoningControl.OFF:
return ReasoningPolicy(
control=ReasoningControl.OFF,
effort=effort,
)
if thinking_control is ReasoningControl.ON or budget_tokens is not None:
return ReasoningPolicy.on(
effort=effort,
budget_tokens=budget_tokens,
)
return ReasoningPolicy(
control=ReasoningControl.DEFAULT,
effort=effort,
)
def _thinking_control(
thinking: ThinkingConfig | None,
*,
budget_tokens: int | None,
) -> ReasoningControl:
if thinking is None:
return ReasoningControl.DEFAULT
if thinking.type == "disabled" or (
"enabled" in thinking.model_fields_set and thinking.enabled is False
):
return ReasoningControl.OFF
if (
thinking.type in {"adaptive", "enabled"}
or ("enabled" in thinking.model_fields_set and thinking.enabled is True)
or budget_tokens is not None
):
return ReasoningControl.ON
return ReasoningControl.DEFAULT
def _output_effort(value: Any) -> tuple[ReasoningEffort | None, bool]:
if not isinstance(value, Mapping):
return None, False
raw = value.get("effort")
if not isinstance(raw, str):
return None, False
normalized = raw.strip().lower()
if normalized == "none":
return None, True
try:
return ReasoningEffort(normalized), False
except ValueError:
return None, False
def _positive_budget(thinking: ThinkingConfig | None) -> int | None:
if thinking is None:
return None
value = thinking.budget_tokens
if isinstance(value, int) and not isinstance(value, bool) and value > 0:
return value
return None