dsh-plugin-subscriptions 0.5.0 → 0.5.2

package/lib/providers/claude.js

@@ -6,16 +6,17 @@
 import { execFileSync } from 'node:child_process';
 import { EMPTY_RESPONSE_CODE, errorChain, LlmAdapter, LlmError, ReasoningEffortId, resolveRetryPolicy } from '@deepseek-ai/dsh-llm';
 import { resolveImages } from '../translate/resolved.js';
-import { streamAnthropic, toAnthropicMessages, toAnthropicSystem, toAnthropicTools, } from '../translate/anthropic.js';
-import { httpLlmError, idleWatchdog, mapFetchFailure, ModelCatalogCache, oauthEndpointError, OAuthEndpointError, TokenManager, } from './common.js';
+import { markMessageCache, streamAnthropic, toAnthropicMessages, toAnthropicSystem, toAnthropicTools, } from '../translate/anthropic.js';
+import { httpLlmError, idleWatchdog, mapFetchFailure, ModelCatalogCache, discoverOrRetryAuth, isMissingOrInvalidCredential, oauthEndpointError, OAuthEndpointError, TokenManager, } from './common.js';
+import { proxiedFetch } from '../http.js';
 export const CLAUDE_CLIENT_ID = '9d1c250a-e61b-44d9-88ed-5944d1962f5e';
-export const CLAUDE_AUTHORIZE_URL = 'https://claude.com/cai/oauth/authorize';
+export const CLAUDE_AUTHORIZE_URL = 'https://claude.ai/oauth/authorize';
 export const CLAUDE_TOKEN_URL = 'https://claude.ai/v1/oauth/token';
 export const CLAUDE_API_URL = 'https://api.anthropic.com/v1/messages?beta=true';
 export const CLAUDE_PROFILE_URL = 'https://api.anthropic.com/api/oauth/profile';
 export const CLAUDE_MODELS_URL = 'https://api.anthropic.com/v1/models?beta=true';
-const CLAUDE_SCOPE = 'org:create_api_key user:profile user:inference user:sessions:claude_code user:mcp_servers user:file_upload';
-const CLAUDE_CALLBACK_PATH = '/callback';
+export const CLAUDE_SCOPE = 'org:create_api_key user:profile user:inference user:sessions:claude_code user:mcp_servers user:file_upload';
+export const CLAUDE_CALLBACK_PATH = '/callback';
 const CLAUDE_CONTEXT_WINDOW = 200_000;
 const CLAUDE_DEFAULT_MAX_TOKENS = 32_000;
 /** Refresh when the access token has less than this much life left. */
@@ -61,13 +62,13 @@ const CLAUDE_BETA_FLAGS = CLAUDE_BETA_FALLBACK;
 export const claudeFlow = {
     callbackPath: CLAUDE_CALLBACK_PATH,
     // The redirect URI embeds the port, so it must be an ephemeral one.
-    listen: { host: '127.0.0.1', ports: [0] },
+    listen: { host: 'localhost', ports: [0] },
     buildAuthorizeUrl({ redirectUri, state, pkce }) {
         const params = new URLSearchParams({
             code: 'true',
             client_id: CLAUDE_CLIENT_ID,
             response_type: 'code',
-            redirect_uri: 'https://platform.claude.com/oauth/code/callback',
+            redirect_uri: redirectUri,
             scope: CLAUDE_SCOPE,
             code_challenge: pkce.challenge,
             code_challenge_method: 'S256',
@@ -79,7 +80,7 @@ export const claudeFlow = {
 /** Best-effort account profile; login must not fail when this does. */
 async function fetchClaudeProfile(accessToken) {
     try {
-        const response = await fetch(CLAUDE_PROFILE_URL, {
+        const response = await proxiedFetch(CLAUDE_PROFILE_URL, {
             headers: { authorization: `Bearer ${accessToken}` },
         });
         if (!response.ok)
@@ -129,7 +130,7 @@ async function claudeSession(tokens, fallbackRefreshToken, withProfile) {
  * @returns the session to store.
  */
 export async function exchangeClaudeCode(code, verifier, redirectUri, state) {
-    const response = await fetch(CLAUDE_TOKEN_URL, {
+    const response = await proxiedFetch(CLAUDE_TOKEN_URL, {
         method: 'POST',
         headers: { 'content-type': 'application/json' },
         body: JSON.stringify({
@@ -151,7 +152,7 @@ export async function exchangeClaudeCode(code, verifier, redirectUri, state) {
  * @returns the fresh session to store.
  */
 export async function refreshClaude(session) {
-    const response = await fetch(CLAUDE_TOKEN_URL, {
+    const response = await proxiedFetch(CLAUDE_TOKEN_URL, {
         method: 'POST',
         headers: { 'content-type': 'application/json' },
         body: JSON.stringify({
@@ -237,7 +238,7 @@ function claudeLimitsWindows(value) {
  * @param signal - caller cancellation from the RPC transport.
  * @returns the mapped usage snapshot.
  */
-export async function fetchClaudeUsage(session, fetchFn = fetch, signal) {
+export async function fetchClaudeUsage(session, fetchFn = proxiedFetch, signal) {
     const response = await fetchFn(CLAUDE_USAGE_URL, {
         headers: {
             'authorization': `Bearer ${session.accessToken}`,
@@ -288,7 +289,7 @@ function claudeReasoning(capabilities) {
     return efforts.length > 0 ? { efforts } : undefined;
 }
 /** Fetch the live model catalog from the subscription endpoint. */
-export async function fetchClaudeModels(session, fetchFn = fetch) {
+export async function fetchClaudeModels(session, fetchFn = proxiedFetch) {
     const response = await fetchFn(CLAUDE_MODELS_URL, {
         headers: {
             'authorization': `Bearer ${session.accessToken}`,
@@ -331,6 +332,38 @@ const CLAUDE_RETRY_MAX_DELAY_MS = 60_000;
 const CLAUDE_RETRY_JITTER_RATIO = 0.2;
 /** The Claude 4.5 family accepts image input. */
 const CLAUDE_MODALITIES = ['text', 'image'];
+/**
+ * Assemble the Anthropic request body.
+ *
+ * Extracted from the adapter so the wire shape — cache breakpoints above all —
+ * is testable without a network round trip. The message array is marked before
+ * it is placed so the breakpoints land on the blocks the body ships: one on the
+ * last `system` block (covering `tools` + `system`, which render ahead of it)
+ * and up to three across the history, Anthropic's four-slot maximum.
+ * @param options - the generate request.
+ * @param messages - conversation messages with images already resolved.
+ * @param maxTokens - the resolved output cap.
+ * @param thinking - the thinking parameter, when the model takes one.
+ * @param effort - the reasoning effort, when the model advertises efforts.
+ * @returns the JSON body to POST.
+ */
+export function claudeRequestBody(options, messages, maxTokens, thinking, effort) {
+    const anthropicMessages = toAnthropicMessages(messages);
+    markMessageCache(anthropicMessages);
+    return {
+        model: options.model,
+        max_tokens: maxTokens,
+        system: toAnthropicSystem(options.system, messages),
+        messages: anthropicMessages,
+        ...options.tools !== undefined && options.tools.length > 0
+            ? { tools: toAnthropicTools(options.tools) }
+            : {},
+        ...thinking === undefined ? {} : { thinking },
+        ...effort === undefined ? {} : { output_config: { effort } },
+        stream: true,
+        ...options.sessionId !== undefined ? { metadata: { user_id: String(options.sessionId) } } : {},
+    };
+}
 /** Claude wire adapter: one instance serves the `claude` provider route. */
 export class ClaudeAdapter extends LlmAdapter {
     options;
@@ -379,7 +412,7 @@ export class ClaudeAdapter extends LlmAdapter {
         if (!this.options.discovery)
             return this.staticModels(provider);
         try {
-            const models = await this.catalog.get(() => this.fetchCatalog());
+            const models = await discoverOrRetryAuth(force => this.options.tokens.session(force), this.catalog, () => this.catalog.get(() => this.fetchCatalog()));
             return models.map(model => ({
                 provider,
                 id: model.id,
@@ -388,11 +421,8 @@ export class ClaudeAdapter extends LlmAdapter {
             }));
         }
         catch (error) {
-            if (error instanceof LlmError
-                && (error.code === 'MISSING_CREDENTIAL' || error.code === 'INVALID_CREDENTIAL'))
+            if (isMissingOrInvalidCredential(error))
                 return [];
-            if (error instanceof LlmError && error.code === 'AUTH')
-                this.catalog.invalidate();
             this.options.onWarn?.(`claude model discovery failed; using the built-in catalog (${errorChain(error)})`);
             return this.staticModels(provider);
         }
@@ -462,22 +492,10 @@ export class ClaudeAdapter extends LlmAdapter {
         const disc = await this.discovered(options.model);
         const thinking = this.thinkingParam(disc?.thinkingType, maxTokens);
         const effort = options.reasoningEffort !== undefined && disc?.reasoning !== undefined
-            ? { output_config: { effort: String(options.reasoningEffort) } }
-            : {};
-        const body = {
-            model: options.model,
-            max_tokens: maxTokens,
-            system: toAnthropicSystem(options.system, messages),
-            messages: toAnthropicMessages(messages),
-            ...options.tools !== undefined && options.tools.length > 0
-                ? { tools: toAnthropicTools(options.tools) }
-                : {},
-            ...thinking === undefined ? {} : { thinking },
-            ...effort,
-            stream: true,
-            ...options.sessionId !== undefined ? { metadata: { user_id: String(options.sessionId) } } : {},
-        };
-        return fetch(CLAUDE_API_URL, {
+            ? String(options.reasoningEffort)
+            : undefined;
+        const body = claudeRequestBody(options, messages, maxTokens, thinking, effort);
+        return proxiedFetch(CLAUDE_API_URL, {
             method: 'POST',
             headers: {
                 'authorization': `Bearer ${session.accessToken}`,

package/lib/providers/codex.js

@@ -3,12 +3,13 @@
  * Codex CLI client id, and streaming against the ChatGPT backend Responses
  * endpoint.
  */
-import { randomUUID } from 'node:crypto';
+import { createHash, randomUUID } from 'node:crypto';
 import { attributionHeaders, EMPTY_RESPONSE_CODE, errorChain, LlmAdapter, LlmError, ReasoningEffortId } from '@deepseek-ai/dsh-llm';
 import { decodeJwtPayload } from '../auth/jwt.js';
 import { resolveImages } from '../translate/resolved.js';
 import { streamResponses, toResponsesInput, toResponsesTools } from '../translate/responses.js';
-import { httpLlmError, idleWatchdog, mapFetchFailure, ModelCatalogCache, oauthEndpointError, OAuthEndpointError, TokenManager, } from './common.js';
+import { httpLlmError, idleWatchdog, mapFetchFailure, ModelCatalogCache, discoverOrRetryAuth, isMissingOrInvalidCredential, oauthEndpointError, OAuthEndpointError, TokenManager, } from './common.js';
+import { proxiedFetch } from '../http.js';
 export const CODEX_CLIENT_ID = 'app_EMoamEEZ73f0CkXaXp7hrann';
 export const CODEX_AUTHORIZE_URL = 'https://auth.openai.com/oauth/authorize';
 export const CODEX_TOKEN_URL = 'https://auth.openai.com/oauth/token';
@@ -152,7 +153,7 @@ function codexSession(tokens, fallback) {
  * @returns the session to store.
  */
 export async function exchangeCodexCode(code, verifier, redirectUri) {
-    const response = await fetch(CODEX_TOKEN_URL, {
+    const response = await proxiedFetch(CODEX_TOKEN_URL, {
         method: 'POST',
         headers: { 'content-type': 'application/x-www-form-urlencoded' },
         body: new URLSearchParams({
@@ -173,7 +174,7 @@ export async function exchangeCodexCode(code, verifier, redirectUri) {
  * @returns the fresh session to store.
  */
 export async function refreshCodex(session) {
-    const response = await fetch(CODEX_TOKEN_URL, {
+    const response = await proxiedFetch(CODEX_TOKEN_URL, {
         method: 'POST',
         headers: { 'content-type': 'application/json' },
         body: JSON.stringify({
@@ -253,7 +254,7 @@ function codexUsageWindow(value, fallbackKind) {
  * @param signal - caller cancellation from the RPC transport.
  * @returns the mapped usage snapshot.
  */
-export async function fetchCodexUsage(session, fetchFn = fetch, signal) {
+export async function fetchCodexUsage(session, fetchFn = proxiedFetch, signal) {
     const response = await fetchFn(CODEX_USAGE_URL, {
         headers: {
             'authorization': `Bearer ${session.accessToken}`,
@@ -309,7 +310,7 @@ function supportsFastTier(entry) {
  * @param fetchFn - fetch implementation (injectable for tests).
  * @returns discovered models: hidden entries dropped, sorted by priority.
  */
-export async function fetchCodexModels(session, fetchFn = fetch) {
+export async function fetchCodexModels(session, fetchFn = proxiedFetch) {
     const url = `${CODEX_MODELS_URL}?client_version=${CODEX_CLIENT_VERSION}`;
     const response = await fetchFn(url, {
         headers: {
@@ -373,6 +374,53 @@ export async function fetchCodexModels(session, fetchFn = fetch) {
     }
     return discovered;
 }
+const CODEX_CALL_ID_MAX_LENGTH = 64;
+const CODEX_CALL_ID_PREFIX = 'call_';
+/**
+ * Bound tool-call ids at the Codex wire boundary without changing the shared
+ * Responses translation used by Grok. Short ids stay verbatim. Oversized ids
+ * become deterministic hashes, and every id already present in this request
+ * is reserved first so a generated id cannot collide with a legitimate short
+ * one (or another oversized id).
+ */
+function normalizeCodexCallIds(input) {
+    const mapping = new Map();
+    const used = new Set();
+    const callId = (item) => (item.type === 'function_call' || item.type === 'function_call_output') && typeof item.call_id === 'string'
+        ? item.call_id
+        : undefined;
+    for (const item of input) {
+        const id = callId(item);
+        if (id !== undefined && id.length <= CODEX_CALL_ID_MAX_LENGTH) {
+            mapping.set(id, id);
+            used.add(id);
+        }
+    }
+    for (const item of input) {
+        const id = callId(item);
+        if (id === undefined || mapping.has(id))
+            continue;
+        let attempt = 0;
+        let normalized;
+        do {
+            const hash = createHash('sha256');
+            if (attempt > 0)
+                hash.update(String(attempt)).update('\0');
+            const digest = hash.update(id).digest('hex');
+            normalized = `${CODEX_CALL_ID_PREFIX}${digest.slice(0, CODEX_CALL_ID_MAX_LENGTH - CODEX_CALL_ID_PREFIX.length)}`;
+            attempt += 1;
+        } while (used.has(normalized));
+        mapping.set(id, normalized);
+        used.add(normalized);
+    }
+    return input.map((item) => {
+        const id = callId(item);
+        if (id === undefined)
+            return item;
+        const normalized = mapping.get(id) ?? id;
+        return normalized === id ? item : { ...item, call_id: normalized };
+    });
+}
 /**
  * The Responses request body for one generation. A fast-tier request (the
  * composer Speed toggle, the codex CLI's fast mode) carries
@@ -383,7 +431,7 @@ export function codexRequestBody(options, resolved, fast) {
     return {
         model: options.model,
         instructions: resolved.instructions ?? DEFAULT_CODEX_INSTRUCTIONS,
-        input: resolved.input,
+        input: normalizeCodexCallIds(resolved.input),
         ...options.tools !== undefined && options.tools.length > 0
             ? { tools: toResponsesTools(options.tools) }
             : {},
@@ -434,7 +482,7 @@ export class CodexAdapter extends LlmAdapter {
             // The fetcher runs only on a cache miss, and resolves the session
             // through the refresh-aware path so an expired access token renews here
             // instead of failing discovery into the static fallback.
-            const discovered = await this.catalog.get(() => this.fetchCatalog());
+            const discovered = await discoverOrRetryAuth(force => this.options.tokens.session(force), this.catalog, () => this.catalog.get(() => this.fetchCatalog()));
             return discovered.map(model => ({
                 provider,
                 id: model.id,
@@ -446,11 +494,8 @@ export class CodexAdapter extends LlmAdapter {
         catch (error) {
             // A permanent refresh failure deletes the stored session: the provider
             // is logged out, so hide it instead of showing a stale static catalog.
-            if (error instanceof LlmError
-                && (error.code === 'MISSING_CREDENTIAL' || error.code === 'INVALID_CREDENTIAL'))
+            if (isMissingOrInvalidCredential(error))
                 return [];
-            if (error instanceof OAuthEndpointError && error.status === 401)
-                this.catalog.invalidate();
             this.options.onWarn?.(`codex model discovery failed; using the built-in catalog (${errorChain(error)})`);
             return this.staticModels(provider);
         }
@@ -529,7 +574,7 @@ export class CodexAdapter extends LlmAdapter {
         const fast = this.options.speedFor !== undefined
             && await this.options.speedFor(options.sessionId, options.model);
         const body = codexRequestBody(options, toResponsesInput(messages, options.system), fast);
-        return fetch(CODEX_API_URL, {
+        return proxiedFetch(CODEX_API_URL, {
             method: 'POST',
             headers: {
                 'authorization': `Bearer ${session.accessToken}`,

package/lib/providers/common.d.ts

@@ -19,6 +19,12 @@ export interface ModelEntry {
     maxTokens?: number;
     /** Accepted request modalities; when set, wins over the provider default. */
     inputModalities?: ('text' | 'image')[];
+    /**
+     * Force this model's upstream protocol. Only the copilot adapter consumes
+     * the semantics; the union is inlined here to avoid a circular import of
+     * the copilot module's `CopilotWire`.
+     */
+    wire?: 'chat-completions' | 'responses';
 }
 /**
  * Validate a configured model catalog (mirrors llm-deepseek's resolveModels).
@@ -175,10 +181,20 @@ export interface DiscoveredModel {
         }[];
         defaultEffort?: ReasoningEffortId;
     };
+    /** Accepted request modalities the endpoint advertised (e.g. Copilot's vision support flag). */
+    inputModalities?: ('text' | 'image')[];
     /** Claude-specific: which extended-thinking wire shape this model accepts. */
     thinkingType?: 'enabled' | 'adaptive';
     /** Codex-specific: the catalog advertises a fast (priority) service tier. */
     fastTier?: boolean;
+    /** Copilot-specific: which upstream protocol the model's endpoints speak. */
+    copilotWire?: 'chat-completions' | 'responses';
+    /**
+     * Copilot-specific: the catalog also lists `/responses` for this model
+     * (dual-protocol entries, e.g. gpt-5.4), so a chat-wire request may reroute
+     * there when it combines function tools with a reasoning effort.
+     */
+    copilotResponses?: boolean;
 }
 /** How long a discovered catalog is trusted before re-fetching. */
 export declare const DISCOVERY_TTL_MS: number;
@@ -207,7 +223,7 @@ export interface CatalogPersistence {
  * while a stale entry refreshes in the background, and only awaits the fetch
  * when nothing is known yet. An optional {@link CatalogPersistence} seeds the
  * last-known state across restarts and receives every successful fetch. A 401
- * during a fetch must call {@link invalidate}.
+ * that still fails after a forced token refresh must call {@link invalidate}.
  */
 export declare class ModelCatalogCache {
     private readonly persistence?;
@@ -224,6 +240,12 @@ export declare class ModelCatalogCache {
      * @returns the cached models, or `undefined` when absent or stale.
      */
     cached(): readonly DiscoveredModel[] | undefined;
+    /**
+     * The last successfully fetched catalog, ignoring TTL. Used to carry
+     * capability metadata forward when a later fetch cannot re-enrich.
+     * @returns the last-known models, or `undefined` when nothing has been stored.
+     */
+    lastKnown(): readonly DiscoveredModel[] | undefined;
     /** Load the persisted snapshot once; a fetch or invalidate that landed first wins. */
     private ensureSeeded;
     /** Run (or join) the single in-flight fetch, updating memory and disk on success. */
@@ -248,4 +270,13 @@ export declare class ModelCatalogCache {
     /** Drop the cached catalog (e.g. after a 401 proved the credential changed). */
     invalidate(): void;
 }
+/** Whether discovery failed because the stored login is gone. */
+export declare function isMissingOrInvalidCredential(error: unknown): boolean;
+/**
+ * Run a catalog fetch, retrying once after a forced token refresh when the
+ * first attempt is a 401/AUTH. Only {@link ModelCatalogCache.invalidate}s
+ * when the retry is also an auth failure, so a refresh race cannot erase
+ * last-known capability metadata.
+ */
+export declare function discoverOrRetryAuth<T>(session: (forceRefresh?: boolean) => Promise<unknown>, catalog: ModelCatalogCache, run: () => Promise<T>): Promise<T>;
 export {};

package/lib/providers/common.js

@@ -32,6 +32,9 @@ export function validateModels(models, label) {
                 || model.inputModalities.some(modality => modality !== 'text' && modality !== 'image'))) {
             throw new Error(`${label}: catalog model "${model.id}" inputModalities must be a non-empty list of "text"/"image"`);
         }
+        if (model.wire !== undefined && model.wire !== 'chat-completions' && model.wire !== 'responses') {
+            throw new Error(`${label}: catalog model "${model.id}" wire must be "chat-completions" or "responses"`);
+        }
         if (seen.has(model.id))
             throw new Error(`${label}: duplicate catalog model "${model.id}"`);
         seen.add(model.id);
@@ -41,6 +44,7 @@ export function validateModels(models, label) {
             ...model.contextWindow === undefined ? {} : { contextWindow: model.contextWindow },
             ...model.maxTokens === undefined ? {} : { maxTokens: model.maxTokens },
             ...model.inputModalities === undefined ? {} : { inputModalities: [...model.inputModalities] },
+            ...model.wire === undefined ? {} : { wire: model.wire },
         };
     });
 }
@@ -271,7 +275,7 @@ export const DISCOVERY_TTL_MS = 5 * 60_000;
  * while a stale entry refreshes in the background, and only awaits the fetch
  * when nothing is known yet. An optional {@link CatalogPersistence} seeds the
  * last-known state across restarts and receives every successful fetch. A 401
- * during a fetch must call {@link invalidate}.
+ * that still fails after a forced token refresh must call {@link invalidate}.
  */
 export class ModelCatalogCache {
     persistence;
@@ -295,6 +299,14 @@ export class ModelCatalogCache {
             return undefined;
         return this.entry.models;
     }
+    /**
+     * The last successfully fetched catalog, ignoring TTL. Used to carry
+     * capability metadata forward when a later fetch cannot re-enrich.
+     * @returns the last-known models, or `undefined` when nothing has been stored.
+     */
+    lastKnown() {
+        return this.entry?.models;
+    }
     /** Load the persisted snapshot once; a fetch or invalidate that landed first wins. */
     ensureSeeded() {
         if (this.persistence === undefined)
@@ -363,3 +375,38 @@ export class ModelCatalogCache {
         void this.persistence?.clear().catch(() => undefined);
     }
 }
+/** Whether discovery failed because the stored login is gone. */
+export function isMissingOrInvalidCredential(error) {
+    return error instanceof LlmError
+        && (error.code === 'MISSING_CREDENTIAL' || error.code === 'INVALID_CREDENTIAL');
+}
+/** Whether discovery failed because the access token was rejected. */
+function isDiscoveryAuthFailure(error) {
+    return (error instanceof OAuthEndpointError && error.status === 401)
+        || (error instanceof LlmError && error.code === 'AUTH');
+}
+/**
+ * Run a catalog fetch, retrying once after a forced token refresh when the
+ * first attempt is a 401/AUTH. Only {@link ModelCatalogCache.invalidate}s
+ * when the retry is also an auth failure, so a refresh race cannot erase
+ * last-known capability metadata.
+ */
+export async function discoverOrRetryAuth(session, catalog, run) {
+    try {
+        return await run();
+    }
+    catch (error) {
+        if (isMissingOrInvalidCredential(error) || !isDiscoveryAuthFailure(error))
+            throw error;
+        try {
+            await session(true);
+            return await run();
+        }
+        catch (retryError) {
+            if (!isMissingOrInvalidCredential(retryError) && isDiscoveryAuthFailure(retryError)) {
+                catalog.invalidate();
+            }
+            throw retryError;
+        }
+    }
+}