@askalf/dario 5.0.1 → 5.1.1

package/README.md

@@ -6,7 +6,7 @@
 
 <p>
   <a href="https://www.npmjs.com/package/@askalf/dario"><img src="https://img.shields.io/npm/v/@askalf/dario?color=6f42c1&label=npm&logo=npm" alt="npm version"></a>
-  <a href="https://github.com/askalf/dario/releases"><img src="https://img.shields.io/badge/release-v5.0-6f42c1?logo=github" alt="v5.0"></a>
+  <a href="https://github.com/askalf/dario/releases"><img src="https://img.shields.io/github/v/release/askalf/dario?color=6f42c1&label=release&logo=github" alt="Latest release"></a>
   <a href="https://github.com/askalf/dario/actions/workflows/ci.yml"><img src="https://github.com/askalf/dario/actions/workflows/ci.yml/badge.svg" alt="CI"></a>
   <a href="https://github.com/askalf/dario/actions/workflows/codeql.yml"><img src="https://github.com/askalf/dario/actions/workflows/codeql.yml/badge.svg" alt="CodeQL"></a>
   <a href="https://scorecard.dev/viewer/?uri=github.com/askalf/dario"><img src="https://img.shields.io/ossf-scorecard/github.com/askalf/dario?label=OpenSSF%20Scorecard&color=6f42c1" alt="OpenSSF Scorecard"></a>

package/dist/cli.js

@@ -348,8 +348,16 @@ async function proxy() {
     // proxy presents to Anthropic is Bun's BoringSSL ClientHello, not
     // Node's OpenSSL one. v3.23 (direction #3).
     const strictTls = args.includes('--strict-tls');
+    // --no-claude-auth: don't load/refresh the Claude OAuth pool (OpenAI-only
+    // proxies). Stops dario rotating a shared refresh token out from under an
+    // interactive Claude Code on the same machine.
+    const noClaudeAuth = args.includes('--no-claude-auth');
     const modelArg = args.find(a => a.startsWith('--model='));
     const model = modelArg ? modelArg.split('=')[1] : undefined;
+    // --fast-model=MODEL: route Haiku-tier (CC sub-agent) requests to this
+    // model instead of the forced --model, so sub-agents stay cheap.
+    const fastModelArg = args.find(a => a.startsWith('--fast-model='));
+    const fastModel = fastModelArg ? fastModelArg.split('=')[1] : undefined;
     // --pace-min=MS / --pace-jitter=MS (v3.24, direction #6 — behavioral
     // smoothing). Inter-request gap floor + optional uniform-random jitter.
     // v4: ~/.dario/config.json's `pacing.{minMs,jitterMs}` is the fallback
@@ -569,7 +577,7 @@ async function proxy() {
         console.error(`[dario] Override (not recommended): pass --unsafe-no-auth if you have out-of-band network controls and accept the risk.`);
         process.exit(1);
     }
-    await startProxy({ port, host, verbose, verboseBodies, model, passthrough, preserveTools, hybridTools, mergeTools, noAutoDetect, strictTls, pacingMinMs, pacingJitterMs, thinkTimeBaseMs, thinkTimePerTokenMs, thinkTimeJitterMs, thinkTimeMaxMs, sessionStartMinMs, sessionStartJitterMs, stealth, drainOnClose, sessionIdleRotateMs, sessionRotateJitterMs, sessionMaxAgeMs, sessionPerClient, preserveOrchestrationTags, noLiveCapture, strictTemplate, maxConcurrent, maxQueued, queueTimeoutMs, effort, maxTokens, logFile, passthroughBetas, skipFields, systemPrompt, overageGuardEnabled, overageGuardBehavior, overageGuardCooldownMs, overageGuardNotifyOs, honorClientThinking, preserveOutputFormat });
+    await startProxy({ port, host, verbose, verboseBodies, model, fastModel, noClaudeAuth, passthrough, preserveTools, hybridTools, mergeTools, noAutoDetect, strictTls, pacingMinMs, pacingJitterMs, thinkTimeBaseMs, thinkTimePerTokenMs, thinkTimeJitterMs, thinkTimeMaxMs, sessionStartMinMs, sessionStartJitterMs, stealth, drainOnClose, sessionIdleRotateMs, sessionRotateJitterMs, sessionMaxAgeMs, sessionPerClient, preserveOrchestrationTags, noLiveCapture, strictTemplate, maxConcurrent, maxQueued, queueTimeoutMs, effort, maxTokens, logFile, passthroughBetas, skipFields, systemPrompt, overageGuardEnabled, overageGuardBehavior, overageGuardCooldownMs, overageGuardNotifyOs, honorClientThinking, preserveOutputFormat });
 }
 /**
  * Parse `--system-prompt=<verbatim|partial|aggressive|filepath>` (or the
@@ -1198,6 +1206,16 @@ async function help() {
                              Provider prefix: openai:gpt-4o, groq:llama-3.3-70b,
                              claude:opus, local:qwen-coder (forces backend)
                              Default: passthrough (client decides)
+    --fast-model=MODEL       Route Haiku-tier sub-agent requests to MODEL
+                             instead of --model, so Claude Code's cheap
+                             sub-agents aren't upgraded to the forced model.
+                             Same MODEL forms as --model. No effect unless set.
+    --no-claude-auth         Don't load or refresh the Claude OAuth token —
+                             for OpenAI-only proxies (e.g. --model=openai:...).
+                             Prevents dario rotating a shared refresh token out
+                             from under an interactive Claude Code on the same
+                             machine. Claude-bound requests then return a clean
+                             unauthenticated error.
     --passthrough, --thin    Thin proxy — OAuth swap only, no injection
     --preserve-tools         Forward client tool schemas unchanged
                              Loses subscription routing; use for custom agents

package/dist/live-fingerprint.d.ts

@@ -282,7 +282,7 @@ export declare function _resetInstalledVersionProbeForTest(): void;
  */
 export declare const SUPPORTED_CC_RANGE: {
     readonly min: "1.0.0";
-    readonly maxTested: "2.1.207";
+    readonly maxTested: "2.1.209";
 };
 /**
  * Compare two dotted-numeric version strings. Returns negative if `a<b`,

package/dist/live-fingerprint.js

@@ -806,7 +806,7 @@ export function _resetInstalledVersionProbeForTest() {
  */
 export const SUPPORTED_CC_RANGE = {
     min: '1.0.0',
-    maxTested: '2.1.207',
+    maxTested: '2.1.209',
 };
 /**
  * Compare two dotted-numeric version strings. Returns negative if `a<b`,

package/dist/oauth.d.ts

@@ -137,6 +137,15 @@ export declare function readLineFromStdin(prompt: string): Promise<string>;
  * Retries with exponential backoff on transient failures.
  * Uses a mutex to prevent concurrent refresh races.
  */
+/**
+ * A refresh failure that retrying cannot recover: the refresh token is
+ * invalid, revoked, or rotated out (e.g. a second dario on the same account
+ * consumed the single-use token). Anthropic signals this as HTTP 401/403, or
+ * as a 400 with an `invalid_grant` body ("Refresh token not found or
+ * invalid"). Terminal -> fail fast with a re-login prompt instead of burning
+ * doomed retries and masking as healthy.
+ */
+export declare function isTerminalRefreshFailure(status: number, body: string): boolean;
 export declare function refreshTokens(): Promise<OAuthTokens>;
 /**
  * Get a valid access token, refreshing if needed.

package/dist/oauth.js

@@ -39,6 +39,12 @@ const REFRESH_COOLDOWN_MS = 60 * 1000;
 let consecutiveRefreshFailures = 0;
 let lastRefreshError;
 const REFRESH_BROKEN_THRESHOLD = 3;
+// Set the moment a refresh fails TERMINALLY (invalid_grant / 401 / 403 — a
+// dead, revoked, or rotated-out refresh token). Unlike the count-based
+// threshold, one terminal failure is conclusive, so surface `broken`
+// immediately instead of masking as healthy for THRESHOLD x cooldown. Cleared
+// whenever new credentials are saved (a successful refresh or `dario login`).
+let refreshTokenDead = false;
 // In-memory credential cache — avoids disk reads on every request
 let credentialsCache = null;
 let credentialsCacheTime = 0;
@@ -401,6 +407,9 @@ export function pickFreshestCredentials(candidates) {
     return best;
 }
 async function saveCredentials(creds) {
+    // New credentials (a successful refresh or a fresh `dario login`) mean the
+    // token is live again — clear the terminal-dead flag.
+    refreshTokenDead = false;
     const path = getDarioCredentialsPath();
     await mkdir(dirname(path), { recursive: true });
     // Write atomically: write to temp file, then rename
@@ -688,6 +697,17 @@ export async function readLineFromStdin(prompt) {
  * Retries with exponential backoff on transient failures.
  * Uses a mutex to prevent concurrent refresh races.
  */
+/**
+ * A refresh failure that retrying cannot recover: the refresh token is
+ * invalid, revoked, or rotated out (e.g. a second dario on the same account
+ * consumed the single-use token). Anthropic signals this as HTTP 401/403, or
+ * as a 400 with an `invalid_grant` body ("Refresh token not found or
+ * invalid"). Terminal -> fail fast with a re-login prompt instead of burning
+ * doomed retries and masking as healthy.
+ */
+export function isTerminalRefreshFailure(status, body) {
+    return status === 401 || status === 403 || /invalid_grant/i.test(body);
+}
 export async function refreshTokens() {
     // Prevent concurrent refreshes — if one is already in progress, wait for it
     if (refreshInProgress)
@@ -723,8 +743,13 @@ async function doRefreshTokens() {
         if (!res.ok) {
             const errBody = await res.text().catch(() => '');
             console.error(`[dario] Refresh attempt ${attempt + 1}/3 failed: HTTP ${res.status} — ${redactSecrets(errBody.slice(0, 200))}`);
-            if (res.status === 401 || res.status === 403) {
-                throw new Error(`Refresh token rejected (${res.status}). Run \`dario login\` to re-authenticate.`);
+            // Terminal failures (invalid_grant / 401 / 403) can't be retried away —
+            // the refresh token is dead. Flag it so /status, /health, and doctor
+            // report `broken` immediately (not after THRESHOLD x cooldown), and fail
+            // fast with an actionable message instead of a vague "3 attempts" throw.
+            if (isTerminalRefreshFailure(res.status, errBody)) {
+                refreshTokenDead = true;
+                throw new Error(`Refresh token invalid or revoked (HTTP ${res.status}) — run \`dario login\` to re-authenticate.`);
             }
             continue;
         }
@@ -795,7 +820,7 @@ export async function getStatus() {
     }
     const { expiresAt } = creds.claudeAiOauth;
     const now = Date.now();
-    const broken = consecutiveRefreshFailures >= REFRESH_BROKEN_THRESHOLD;
+    const broken = refreshTokenDead || consecutiveRefreshFailures >= REFRESH_BROKEN_THRESHOLD;
     if (expiresAt < now) {
         // Expired but has refresh token — can be refreshed (unless refresh itself is dead)
         const canRefresh = !!creds.claudeAiOauth.refreshToken && !broken;

package/dist/proxy.d.ts

@@ -37,6 +37,21 @@ export declare function buildBillingTag(cliVersion: string, cch: string | null):
  * that needs to Just Work.
  */
 export declare function resolveClaudeAlias(model: string): string;
+/**
+ * Pick the per-request model override under a forced `--model`.
+ *
+ * A forced `--model` normally rewrites the model on *every* request. But
+ * Claude Code dispatches its throwaway sub-agent (Explore/Task) turns on the
+ * cheap Haiku tier, so forcing a frontier model silently upgrades those
+ * sub-agents too — quietly multiplying cost. When `--fast-model` is set, a
+ * Haiku-tier inbound request routes there instead, keeping sub-agents cheap
+ * while the main conversation stays on `--model`.
+ *
+ * No-op unless `fastModelOverride` is set: with it null, this returns
+ * `modelOverride` for every request (the pre-existing behavior), so the
+ * change is inert until the operator opts in.
+ */
+export declare function selectModelOverride(incomingModel: string, modelOverride: string | null, fastModelOverride: string | null): string | null;
 export declare function parseProviderPrefix(model: string): {
     provider: 'openai' | 'claude';
     model: string;
@@ -209,12 +224,21 @@ export declare const OPENAI_MODELS_LIST: {
         owned_by: string;
     }>;
 };
+/**
+ * Whether dario must have a Claude login to start. False (an empty pool is
+ * expected, not a fatal "run dario login") for the modes that serve requests
+ * without the Claude OAuth pool: admin-bootstrap, upstream-api-key, and
+ * --no-claude-auth. Pure so the startup gate is unit-testable.
+ */
+export declare function requiresClaudeLogin(poolSize: number, adminEnabled: boolean, hasUpstreamApiKey: boolean, noClaudeAuth: boolean): boolean;
 interface ProxyOptions {
     port?: number;
     host?: string;
     verbose?: boolean;
     verboseBodies?: boolean;
     model?: string;
+    fastModel?: string;
+    noClaudeAuth?: boolean;
     passthrough?: boolean;
     preserveTools?: boolean;
     hybridTools?: boolean;

package/dist/proxy.js

@@ -235,6 +235,25 @@ const MODEL_ALIASES = {
 export function resolveClaudeAlias(model) {
     return resolveAliasAgainst(model, getCachedBases()) ?? MODEL_ALIASES[model] ?? model;
 }
+/**
+ * Pick the per-request model override under a forced `--model`.
+ *
+ * A forced `--model` normally rewrites the model on *every* request. But
+ * Claude Code dispatches its throwaway sub-agent (Explore/Task) turns on the
+ * cheap Haiku tier, so forcing a frontier model silently upgrades those
+ * sub-agents too — quietly multiplying cost. When `--fast-model` is set, a
+ * Haiku-tier inbound request routes there instead, keeping sub-agents cheap
+ * while the main conversation stays on `--model`.
+ *
+ * No-op unless `fastModelOverride` is set: with it null, this returns
+ * `modelOverride` for every request (the pre-existing behavior), so the
+ * change is inert until the operator opts in.
+ */
+export function selectModelOverride(incomingModel, modelOverride, fastModelOverride) {
+    if (fastModelOverride && /haiku/i.test(incomingModel))
+        return fastModelOverride;
+    return modelOverride;
+}
 // Provider prefix in the `model` field — `<provider>:<model>`. Forces
 // routing regardless of model-name regex. Only recognized prefixes are
 // parsed, so ollama-style `llama3:8b` (without a recognized prefix)
@@ -683,6 +702,15 @@ export function createOpenAIStreamTranslator() {
 // autodetected catalog (model-catalog.ts); `[1m]` variants are GENERATED by
 // the one shared long-context rule, never hand-listed per model.
 export const OPENAI_MODELS_LIST = buildOpenAIModelsList(withLongContextVariants(BAKED_BASE_MODELS));
+/**
+ * Whether dario must have a Claude login to start. False (an empty pool is
+ * expected, not a fatal "run dario login") for the modes that serve requests
+ * without the Claude OAuth pool: admin-bootstrap, upstream-api-key, and
+ * --no-claude-auth. Pure so the startup gate is unit-testable.
+ */
+export function requiresClaudeLogin(poolSize, adminEnabled, hasUpstreamApiKey, noClaudeAuth) {
+    return poolSize === 0 && !adminEnabled && !hasUpstreamApiKey && !noClaudeAuth;
+}
 /**
  * Append a JSON-ND line to the proxy log file. No-op when stream is
  * null (logFile not configured). Errors are swallowed — log writes
@@ -1060,17 +1088,29 @@ export async function startProxy(opts = {}) {
         console.error(`[dario] overage-guard resumed (${info.reason}). Normal request handling restored.`);
     });
     let status;
-    for (const acc of accountsList) {
-        pool.add(acc.alias, {
-            accessToken: acc.accessToken,
-            refreshToken: acc.refreshToken,
-            expiresAt: acc.expiresAt,
-            deviceId: acc.deviceId,
-            accountUuid: acc.accountUuid,
-        });
+    // --no-claude-auth: don't populate the Claude OAuth pool. An empty pool makes
+    // the background refresh a no-op, so dario never touches (and never rotates)
+    // the shared Claude refresh token — the fix for a local OpenAI-only proxy
+    // logging out an interactive Claude Code on the same machine (dario#737 class,
+    // locally). OpenAI-compat requests use their own backend key, unaffected.
+    if (opts.noClaudeAuth) {
+        console.error('[dario] --no-claude-auth: Claude OAuth pool NOT loaded — serving OpenAI-compatible backends only; the Claude refresh token is never touched. Claude-bound requests return an unauthenticated error.');
+    }
+    else {
+        for (const acc of accountsList) {
+            pool.add(acc.alias, {
+                accessToken: acc.accessToken,
+                refreshToken: acc.refreshToken,
+                expiresAt: acc.expiresAt,
+                deviceId: acc.deviceId,
+                accountUuid: acc.accountUuid,
+            });
+        }
     }
     // Background refresh — keep every account's token fresh without blocking requests
     const refreshInterval = setInterval(async () => {
+        if (opts.noClaudeAuth)
+            return; // never touch the Claude token in OpenAI-only mode
         for (const acc of pool.all()) {
             if (acc.expiresAt < Date.now() + 45 * 60 * 1000) {
                 try {
@@ -1098,7 +1138,7 @@ export async function startProxy(opts = {}) {
     // dead-but-refreshable token (a container restarted right after a normal
     // expiry, gap #1) is refreshed, then back-filled so the recovered login
     // becomes the pool-of-one it should be — rather than crash-looping on exit(1).
-    if (pool.size === 0 && !adminEnabled && !upstreamApiKey) {
+    if (requiresClaudeLogin(pool.size, adminEnabled, !!upstreamApiKey, opts.noClaudeAuth ?? false)) {
         const single = await resolveSingleAccountStartupStatus();
         if (!single.authenticated) {
             console.error('[dario] Not authenticated. Run `dario login` first.');
@@ -1150,6 +1190,12 @@ export async function startProxy(opts = {}) {
     const cliModelRaw = modelPrefix ? modelPrefix.model : opts.model;
     const cliProviderOverride = modelPrefix ? modelPrefix.provider : null;
     const modelOverride = cliModelRaw ? resolveClaudeAlias(cliModelRaw) : null;
+    // --fast-model: the model that Haiku-tier (Claude Code sub-agent) requests
+    // route to instead of the forced `--model`, so sub-agents stay cheap. Parsed
+    // like --model (alias + optional provider prefix). Null = disabled.
+    const fastModelPrefix = opts.fastModel ? parseProviderPrefix(opts.fastModel) : null;
+    const fastModelRaw = fastModelPrefix ? fastModelPrefix.model : opts.fastModel;
+    const fastModelOverride = fastModelRaw ? resolveClaudeAlias(fastModelRaw) : null;
     const identity = loadClaudeIdentity();
     if (identity.deviceId) {
         console.log('  Device identity: detected');
@@ -2034,7 +2080,13 @@ export async function startProxy(opts = {}) {
                     const parsed = parsedBody ?? JSON.parse(body.toString());
                     // Strip orchestration tags from messages (Aider, Cursor, etc.)
                     sanitizeMessages(parsed, opts.preserveOrchestrationTags);
-                    const result = isOpenAI ? openaiToAnthropic(parsed, modelOverride) : (modelOverride ? { ...parsed, model: modelOverride } : parsed);
+                    // Tier-aware routing: under a forced --model, a Haiku-tier sub-agent
+                    // request routes to --fast-model (when set) instead of the forced
+                    // model, so CC's cheap sub-agents aren't silently upgraded. Inert
+                    // when --fast-model is unset (effectiveOverride === modelOverride).
+                    const incomingModel = typeof parsed.model === 'string' ? parsed.model : '';
+                    const effectiveOverride = selectModelOverride(incomingModel, modelOverride, fastModelOverride);
+                    const result = isOpenAI ? openaiToAnthropic(parsed, effectiveOverride) : (effectiveOverride ? { ...parsed, model: effectiveOverride } : parsed);
                     const r = result;
                     requestModel = (r.model || '').toLowerCase();
                     // Suspended-model guard. Empty by default (Fable 5 returned globally
@@ -3364,7 +3416,10 @@ export async function startProxy(opts = {}) {
         const modeLine = passthrough
             ? 'Mode: passthrough (OAuth swap only, no injection)'
             : `OAuth: ${status.status} (expires in ${status.expiresIn})`;
-        const modelLine = modelOverride ? `Model: ${modelOverride} (all requests)` : 'Model: passthrough (client decides)';
+        const fastNote = fastModelOverride ? `, ${fastModelOverride} (Haiku-tier sub-agents)` : '';
+        const modelLine = modelOverride
+            ? `Model: ${modelOverride} (all requests)${fastNote}`
+            : `Model: passthrough (client decides)${fastNote}`;
         // Pool line surfaces the account state on every startup. The pool is the
         // one credential model now (v5.0): a plain `dario login` shows as a pool of
         // one. An empty pool is only reachable in admin-bootstrap / api-key mode.
@@ -3388,6 +3443,8 @@ export async function startProxy(opts = {}) {
         console.log('');
         console.log(`  ${modeLine}`);
         console.log(`  ${modelLine}`);
+        if (opts.noClaudeAuth)
+            console.log('  Claude auth: disabled (--no-claude-auth) — OpenAI-compatible backends only; Claude token untouched');
         console.log(`  ${poolLine}`);
         if (!isLoopbackHost(host)) {
             console.log('');

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@askalf/dario",
-  "version": "5.0.1",
+  "version": "5.1.1",
   "description": "Use your Claude Pro/Max subscription in any tool — Cursor, Cline, Aider, the Agent SDK, your scripts — at subscription pricing, not per-token API bills. One local Anthropic + OpenAI-compatible endpoint.",
   "type": "module",
   "bin": {