Skip to content

fix(cloudflare): sync reasoning controls from model schemas #7920

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
superhighfives wants to merge 1 commit into from
Closed

fix(cloudflare): sync reasoning controls from model schemas #7920

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 10 additions & 4 deletions packages/core/src/sync/index.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -93,6 +93,8 @@ export interface SyncProvider<SourceModel> {
/** Replace existing leading comments with translateModel.header. */
authoritativeHeaders?: boolean;
sameModel?(current: ExistingModel, desired: SyncedModel): boolean;
/** Refresh provider-owned comments while preserving unrelated authored notes. */
updateHeader?(current: string, generated: string): string;
missingNotice?(paths: string[]): string[];
/**
* Remote ID to report when translateModel skips a source model. Return
Expand Down Expand Up @@ -368,12 +370,15 @@ export async function syncProvider<SourceModel>(
throw parsed.error;
}

const currentHeader = existing.get(relativePath)?.header ?? "";
const translatedHeader = translated.header === undefined
? undefined
: leadingComments(translated.header);
const header = provider.authoritativeHeaders
? translatedHeader ?? ""
: (existing.get(relativePath)?.header || translatedHeader) ?? "";
const header = translatedHeader !== undefined && provider.updateHeader !== undefined
? provider.updateHeader(currentHeader, translatedHeader)
: provider.authoritativeHeaders
? translatedHeader ?? ""
: currentHeader || translatedHeader || "";
desired.set(relativePath, {
model: parsed.data,
content: header + formatToml(parsed.data),
Expand Down Expand Up @@ -448,7 +453,8 @@ export async function syncProvider<SourceModel>(
continue;
}

const headerChanged = provider.authoritativeHeaders && current.header !== file.header;
const headerChanged = (provider.authoritativeHeaders || provider.updateHeader !== undefined)
&& current.header !== file.header;
if (
headerChanged
|| !(provider.sameModel?.(current.authored, file.model)
Expand Down
135 changes: 107 additions & 28 deletions packages/core/src/sync/providers/cloudflare-workers-ai.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,8 @@
import { z } from "zod";
import { readFileSync, readdirSync } from "node:fs";
import { readdirSync } from "node:fs";
import path from "node:path";

import { ReasoningOption } from "../../schema.js";
import type { ExistingModel, SyncedFullModel, SyncedModel, SyncProvider } from "../index.js";
import { MissingReasoningOptionsError } from "../missing-reasoning-options.js";
import { buildOpenRouterModel, OpenRouterModel } from "./openrouter.js";
Expand Down Expand Up @@ -34,7 +35,10 @@ const METADATA_PUBLISHERS: Record<string, string> = {
};

const WorkersAiReasoning = OpenRouterModel.shape.reasoning.unwrap().partial({ mandatory: true });
const WorkersAiModel = OpenRouterModel.extend({ reasoning: WorkersAiReasoning.optional() });
const WorkersAiModel = OpenRouterModel.extend({
reasoning: WorkersAiReasoning.optional(),
input_schema: z.record(z.unknown()).optional(),
});
type WorkersAiModel = z.infer<typeof WorkersAiModel>;

const CloudflareModel = z.object({
Expand All @@ -54,6 +58,7 @@ const CloudflareModel = z.object({
supported_sampling_parameters: z.array(z.string()).optional(),
// Validate reasoning separately so malformed controls do not discard the model.
reasoning: z.unknown().optional(),
input_schema: z.record(z.unknown()).optional(),
}).passthrough();

const CloudflareResponse = z.object({
Expand Down Expand Up @@ -85,7 +90,14 @@ export const cloudflareWorkersAi = {
name: "Cloudflare Workers AI",
modelsDir: "providers/cloudflare-workers-ai/models",
deleteMissing: false,
authoritativeHeaders: true,
updateHeader(current, generated) {
const notes = current.split("\n").filter((line) => {
const trimmed = line.trim();
return trimmed !== ""
&& !/^# (?:Reasoning controls declared by Cloudflare's model input schema\.|Toggle:|Effort:|Budget:)/.test(trimmed);
});
return [...new Set([...generated.trim().split("\n"), ...notes])].join("\n") + "\n";
},
async fetchModels() {
const accountID = process.env.CLOUDFLARE_WORKERS_AI_SYNC_ACCOUNT_ID;
const token = process.env.CLOUDFLARE_WORKERS_AI_SYNC_API_TOKEN;
Expand All @@ -104,7 +116,14 @@ export const cloudflareWorkersAi = {
if (response !== undefined) rows.push(...modelRows(response));
}

return { data: rows };
const models = parseCloudflareModels({ data: rows });
for (let i = 0; i < models.length; i += 4) {
await Promise.all(models.slice(i, i + 4).map(async (model) => {
model.input_schema = await fetchSchema(accountID, token, model.id.replace(/^workers-ai\//, ""));
}));
}

return { data: models };
},
parseModels(raw) {
const models = parseCloudflareModels(raw);
Expand All @@ -114,35 +133,46 @@ export const cloudflareWorkersAi = {
translateModel(model, context) {
const id = model.id;
const existing = context.existing(id);
if (existing === undefined && hasReasoning(model) && workersAiReasoningOptions(model) === undefined) {
const apiControls = workersAiReasoningOptions(model);
const controls = apiControls === undefined
&& model.supported_parameters.some((value) => value === "reasoning" || value === "include_reasoning")
? schemaReasoningOptions(model.input_schema)
: undefined;
if (
existing === undefined
&& model.input_schema !== undefined
&& hasReasoning(model)
&& controls === undefined
&& apiControls === undefined
) {
throw new MissingReasoningOptionsError(id, "Workers AI Search does not specify concrete reasoning controls; manual authoring is needed");
}
const translated = buildWorkersAiModel(model, existing);
// Only read paths already present in the sync runner's catalogue map.
const header = existing === undefined ? "" : modelHeader(path.resolve(ROOT_DIR, this.modelsDir, `${id}.toml`));
if (header === undefined) return undefined;
// Remove only the exact generated line; preserve every other leading comment.
const preservedHeader = header.replace(TOGGLE_HEADER, "");
// Respect curated model-specific wire instructions, including Qwen's split lines.
const hasToggleWire = /\b(?:thinking\.type|chat_template_kwargs\.thinking|enable_thinking)\b/.test(preservedHeader);
const translated = {
...buildWorkersAiModel(model, existing),
...(controls && { reasoning_options: controls.options }),
};
if (controls !== undefined) return { id, model: translated, header: controls.header };

const toggleHeader = ENABLE_THINKING_MODELS.has(id) ? TOGGLE_HEADER : "";
if (translated.reasoning_options?.some((option) => option.type === "toggle") && !hasToggleWire && !toggleHeader) {
if (existing === undefined) {
throw new MissingReasoningOptionsError(id, "Workers AI toggle wire path needs manual verification");
}
if (
translated.reasoning_options?.some((option) => option.type === "toggle")
&& model.input_schema !== undefined
&& !toggleHeader
) {
if (existing === undefined) throw new MissingReasoningOptionsError(id, "Workers AI toggle wire path needs manual verification");
console.warn(`Keeping catalogue reasoning for ${id}: Workers AI toggle wire path is unknown`);
// Keep the reasoning boundary small: other usable properties still update.
return {
id,
model: buildWorkersAiModel({ ...model, reasoning: undefined }, existing),
header,
};
}
return {
id,
model: translated,
header: (translated.reasoning_options?.some((option) => option.type === "toggle") && !hasToggleWire ? toggleHeader : "")
+ preservedHeader,
header: translated.reasoning_options?.some((option) => option.type === "toggle")
? toggleHeader || undefined
: undefined,
};
},
} satisfies SyncProvider<WorkersAiModel>;
Expand All @@ -153,15 +183,64 @@ function hasReasoning(model: WorkersAiModel) {
|| model.reasoning?.supports_max_tokens === true;
}

function modelHeader(file: string) {
try {
const lines = readFileSync(file, "utf8").split("\n");
const firstKey = lines.findIndex((line) => line.trim() !== "" && !line.trim().startsWith("#"));
return lines.slice(0, firstKey === -1 ? lines.length : firstKey).join("\n") + "\n";
} catch (error) {
console.warn(`Skipping Workers AI model with an unreadable catalogue header: ${file}`, error);
return undefined;
const SchemaNode = z.object({
type: z.string().optional(),
enum: z.array(z.unknown()).optional(),
const: z.unknown().optional(),
properties: z.record(z.unknown()).optional(),
anyOf: z.array(z.unknown()).optional(),
oneOf: z.array(z.unknown()).optional(),
});

// Follow request properties and alternatives, not arbitrary nested message/tool schemas.
function schemaFields(input: unknown, path: string[]): z.infer<typeof SchemaNode>[] {
const parsed = SchemaNode.safeParse(input);
if (!parsed.success) return [];
const node = parsed.data;
const fields = path.length === 0 ? [node] : schemaFields(node.properties?.[path[0]!], path.slice(1));
return fields.concat((node.anyOf ?? []).concat(node.oneOf ?? []).flatMap((branch) => schemaFields(branch, path)));
}

function schemaReasoningOptions(input: unknown) {
const options: z.infer<typeof ReasoningOption>[] = [];
const paths = [["reasoning_effort"], ["reasoning", "effort"]];
const values = [...new Set(paths.flatMap((path) => schemaFields(input, path))
.flatMap((node) => node.enum ?? []).filter((value) => typeof value === "string"))];
const effort = ReasoningOption.safeParse({ type: "effort", values });
const toggle = schemaFields(input, ["chat_template_kwargs", "enable_thinking"])
.some((node) => node.type === "boolean" && node.const === undefined &&
(node.enum === undefined || (node.enum.includes(true) && node.enum.includes(false))));
const comments = ["# Reasoning controls declared by Cloudflare's model input schema."];
if (toggle && !values.includes("none")) {
options.push({ type: "toggle" });
comments.push("# Toggle: chat_template_kwargs.enable_thinking = true|false");
}
if (effort.success && values.some((value) => value !== "none")) {
options.push(effort.data);
for (const path of paths) {
if (schemaFields(input, path).some((node) => node.enum?.length)) {
comments.push(`# Effort: ${path.join(".")} = ${values.join("|")}`);
}
}
}
if (options.length === 0) return undefined;
return { options, header: comments.join("\n") + "\n" };
}

async function fetchSchema(accountID: string, token: string, model: string) {
const url = new URL(`${API_BASE}/${accountID}/ai/models/schema`);
url.searchParams.set("model", model);
const response = await fetch(url, {
headers: { Authorization: `Bearer ${token}` },
signal: AbortSignal.timeout(30_000),
});
if (!response.ok) {
throw new Error(`Cloudflare Workers AI schema request failed for ${model}: ${response.status}`);
}
return z.object({
success: z.literal(true),
result: z.object({ input: z.record(z.unknown()) }),
}).parse(await response.json()).result.input;
}

export function buildWorkersAiModel(
Expand Down
Loading
Loading