Skip to content

Use Claude Code with your ChatGPT, Kimi, Cursor or Grok subscription via a local Anthropic-compatible proxy

Rust

add model and effort override env vars

Omar Abdulla4d552f1d4b17
diff --git a/src/anthropic/schema.ts b/src/anthropic/schema.ts
index bda93bf..916a628 100644
--- a/src/anthropic/schema.ts
+++ b/src/anthropic/schema.ts
@@ -61,7 +61,7 @@ export interface AnthropicRequest {
stream?: boolean
thinking?: { type: string; [k: string]: unknown }
output_config?: {
- effort?: "low" | "medium" | "high"
+ effort?: "low" | "medium" | "high" | "xhigh"
format?: { type: "json_schema"; schema: unknown; name?: string; strict?: boolean }
}
context_management?: unknown
diff --git a/src/providers/codex/translate/model-allowlist.ts b/src/providers/codex/translate/model-allowlist.ts
index 5ff3744..580d57e 100644
--- a/src/providers/codex/translate/model-allowlist.ts
+++ b/src/providers/codex/translate/model-allowlist.ts
@@ -16,6 +16,20 @@ export const MODEL_ALIASES = new Map<string, string>([
])
export function resolveModel(model: string): string {
+ // The CLAUDE_CODEX_PROXY_OPEN_AI_MODEL_OVERRIDE environment variable allows
+ // for the model that's used to be overridden so that regardless of whatever
+ // model is being requested by the harness, the model which is provided in
+ // that env var is always returned.
+ //
+ // This is useful in cases where you just want the claude-code harness to use
+ // a specific model all across the way.
+ if (
+ process.env.CLAUDE_CODEX_PROXY_OPEN_AI_MODEL_OVERRIDE !== undefined &&
+ process.env.CLAUDE_CODEX_PROXY_OPEN_AI_MODEL_OVERRIDE !== ""
+ ) {
+ return process.env.CLAUDE_CODEX_PROXY_OPEN_AI_MODEL_OVERRIDE
+ }
+
return MODEL_ALIASES.get(model) ?? model
}
diff --git a/src/providers/codex/translate/request.ts b/src/providers/codex/translate/request.ts
index ac0f115..6899243 100644
--- a/src/providers/codex/translate/request.ts
+++ b/src/providers/codex/translate/request.ts
@@ -7,6 +7,8 @@ import type {
AnthropicTool,
} from "../../../anthropic/schema.ts"
+export type Effort = "none" | "low" | "medium" | "high" | "xhigh"
+
// Keep this aligned to the upstream Codex ResponsesApiRequest field set.
// Do not add plausible-looking top-level fields without source support or a confirmed live test.
export interface ResponsesRequest {
@@ -20,7 +22,7 @@ export interface ResponsesRequest {
| "required"
| { type: "function"; name: string }
parallel_tool_calls?: boolean
- reasoning?: { effort?: "low" | "medium" | "high"; summary?: unknown }
+ reasoning?: { effort?: Effort; summary?: unknown }
store: false
stream: true
include?: string[]
@@ -71,6 +73,24 @@ export interface TranslateOptions {
sessionId?: string
}
+function resolveEffort(effort?: Effort): Effort | undefined {
+ // The CLAUDE_CODEX_PROXY_OPEN_AI_EFFORT_OVERRIDE environment variable allows
+ // for the effort that's used to be overridden so that regardless of whatever
+ // effort is being requested by the harness, the effort which is provided in
+ // that env var is always returned.
+ //
+ // This is useful in cases where you just want the claude-code harness to use
+ // a specific effort all across the way.
+ if (
+ process.env.CLAUDE_CODEX_PROXY_OPEN_AI_EFFORT_OVERRIDE !== undefined &&
+ process.env.CLAUDE_CODEX_PROXY_OPEN_AI_EFFORT_OVERRIDE !== ""
+ ) {
+ return process.env.CLAUDE_CODEX_PROXY_OPEN_AI_EFFORT_OVERRIDE as Effort
+ }
+
+ return effort
+}
+
export function translateRequest(req: AnthropicRequest, opts: TranslateOptions = {}): ResponsesRequest {
const instructions = buildInstructions(req.system)
const input = buildInput(req.messages)
@@ -99,7 +119,7 @@ export function translateRequest(req: AnthropicRequest, opts: TranslateOptions =
if (instructions) out.instructions = instructions
if (tools && tools.length) out.tools = tools
if (opts.sessionId) out.prompt_cache_key = opts.sessionId
- const effort = req.output_config?.effort
+ const effort = resolveEffort(req.output_config?.effort)
if (effort) {
out.reasoning = { effort }
out.include = ["reasoning.encrypted_content"]
diff --git a/src/providers/kimi/translate/request.ts b/src/providers/kimi/translate/request.ts
index b74d694..5535436 100644
--- a/src/providers/kimi/translate/request.ts
+++ b/src/providers/kimi/translate/request.ts
@@ -86,7 +86,7 @@ export function translateRequest(
stream: true,
stream_options: { include_usage: true },
max_tokens: clampMaxTokens(req.max_tokens),
- reasoning_effort: req.output_config?.effort ?? "medium",
+ reasoning_effort: mapReasoningEffort(req.output_config?.effort),
thinking: { type: "enabled" },
}
if (tools && tools.length) out.tools = tools
@@ -101,6 +101,16 @@ function clampMaxTokens(requested: number | undefined): number {
return Math.min(requested, DEFAULT_MAX_TOKENS)
}
+// Kimi's reasoning_effort is capped at "high"; collapse the proxy's "xhigh"
+// extension to "high" since Kimi has no stronger tier, and default to "medium"
+// when no effort is requested.
+function mapReasoningEffort(
+ effort: NonNullable<AnthropicRequest["output_config"]>["effort"],
+): "low" | "medium" | "high" {
+ if (effort === "xhigh") return "high"
+ return effort ?? "medium"
+}
+
function mapToolChoice(choice: AnthropicRequest["tool_choice"]): KimiToolChoice {
if (!choice) return "auto"
switch (choice.type) {