@aexol/spectral 0.2.22 → 0.3.1

package/dist/server/pi-bridge.js

@@ -48,8 +48,8 @@
  *  a single pi session also work normally (the same AgentSession
  *  instance is reused across `prompt()` calls).
  */
-import { AuthStorage, createAgentSession, DefaultResourceLoader, ModelRegistry, SessionManager, } from "@mariozechner/pi-coding-agent";
 import { createJiti } from "@mariozechner/jiti";
+import { AuthStorage, createAgentSession, DefaultResourceLoader, ModelRegistry, SessionManager, } from "@mariozechner/pi-coding-agent";
 import { randomUUID } from "node:crypto";
 import { existsSync } from "node:fs";
 import { dirname, resolve } from "node:path";
@@ -124,6 +124,12 @@ const MODEL_PRICING = [
     { prefix: "claude-3-sonnet", input: 3, output: 15, cacheWrite: 3.75, cacheRead: 0.30 },
     { prefix: "claude-3-haiku", input: 0.25, output: 1.25, cacheWrite: 1.25, cacheRead: 0.025 },
     // OpenAI models
+    { prefix: "gpt-5.5", input: 1.75, output: 14, cacheWrite: 14, cacheRead: 0.35 },
+    { prefix: "gpt-5.4", input: 1.75, output: 14, cacheWrite: 14, cacheRead: 0.35 },
+    { prefix: "gpt-5.3", input: 1.75, output: 14, cacheWrite: 14, cacheRead: 0.35 },
+    { prefix: "gpt-5.2", input: 1.75, output: 14, cacheWrite: 14, cacheRead: 0.35 },
+    { prefix: "gpt-5.1", input: 2, output: 8, cacheWrite: 8, cacheRead: 0.50 },
+    { prefix: "gpt-5", input: 1.25, output: 10, cacheWrite: 10, cacheRead: 0.25 },
     { prefix: "gpt-4.1", input: 2, output: 8, cacheWrite: 8, cacheRead: 0.50 },
     { prefix: "gpt-4o", input: 2.50, output: 10, cacheWrite: 10, cacheRead: 1.25 },
     { prefix: "gpt-4-turbo", input: 10, output: 30, cacheWrite: 0, cacheRead: 0 },
@@ -158,6 +164,21 @@ function lookupPricing(modelId) {
     }
     return null;
 }
+/**
+ * Model prefixes known to support reasoning/thinking.
+ * Mirrors pi-ai's supportsXhigh() + additional models.
+ */
+const REASONING_SUPPORT_PREFIXES = [
+    "gpt-5.2", "gpt-5.3", "gpt-5.4", "gpt-5.5",
+    "claude-opus-4",
+    "o3", "o4",
+    "deepseek-r1",
+    "gemini-2.5",
+];
+/** Check if a modelId prefix indicates reasoning/thinking support. */
+function supportsReasoning(modelId) {
+    return REASONING_SUPPORT_PREFIXES.some((p) => modelId.startsWith(p));
+}
 /**
  * Parse the newline-delimited JSON of wire events persisted alongside an
  * assistant message and extract all tool_call / tool_result events.
@@ -249,6 +270,12 @@ export class PiBridge {
      * `session.setModel()`. Phase 3 (Available Models whitelist).
      */
     modelRegistry;
+    /**
+     * Raw allowed models list from the backend, preserved in sortOrder.
+     * Used by `getFirstAvailableModelId()` to return the backend-curated
+     * top pick when no explicit model selection is made.
+     */
+    allowedModels;
     /**
      * Last `modelId` we successfully applied via `session.setModel()`, or
      * `undefined` if we never applied one (pi falls back to its own settings
@@ -426,6 +453,7 @@ export class PiBridge {
             throw new Error(`Failed to fetch allowed models from backend; check SPECTRAL_BACKEND_URL ` +
                 `and machine JWT. Underlying error: ${e.message}`);
         }
+        this.allowedModels = allowedModels;
         this.registerSyntheticProviders(allowedModels);
         console.info(`✓ Inference routed via backend proxy (${allowedModels.length} model(s) available)`);
         const result = await createAgentSession({
@@ -490,7 +518,7 @@ export class PiBridge {
                         // at our synthetic proxy provider so auth resolves to the machine JWT.
                         provider: SPECTRAL_PROXY_ANTHROPIC,
                         baseUrl,
-                        reasoning: false,
+                        reasoning: supportsReasoning(m.modelId),
                         input: ["text", "image"],
                         // Real pricing so pi can compute accurate token costs.
                         cost: pricing
@@ -520,7 +548,7 @@ export class PiBridge {
                         // breaking auth lookup against our synthetic proxy provider.
                         provider: SPECTRAL_PROXY_OPENAI,
                         baseUrl,
-                        reasoning: false,
+                        reasoning: supportsReasoning(m.modelId),
                         input: ["text", "image"],
                         // Real pricing so pi can compute accurate token costs.
                         cost: pricing
@@ -551,7 +579,7 @@ export class PiBridge {
                         api: "openai-completions",
                         provider: SPECTRAL_PROXY_USER_MODEL,
                         baseUrl,
-                        reasoning: false,
+                        reasoning: supportsReasoning(m.modelId),
                         input: ["text", "image"],
                         cost: pricing
                             ? { input: pricing.input, output: pricing.output, cacheRead: pricing.cacheRead, cacheWrite: pricing.cacheWrite }
@@ -579,6 +607,18 @@ export class PiBridge {
      *
      * Phase 3 (Available Models whitelist).
      */
+    /**
+     * Return the modelId of the first available model from the backend
+     * whitelist (preserving the backend's sortOrder, which is the same
+     * ordering the frontend uses in its model picker). Returns `undefined`
+     * when no models are available (e.g. backend unreachable at startup).
+     *
+     * Used by `SessionStreamManager.prompt()` as a defense-in-depth
+     * default when neither the envelope nor SQLite supply a modelId.
+     */
+    getFirstAvailableModelId() {
+        return this.allowedModels?.[0]?.modelId;
+    }
     async setModel(modelId) {
         if (!modelId)
             return true; // nothing to apply — pi keeps its current model

package/dist/server/session-stream.js

@@ -232,7 +232,9 @@ export class SessionStreamManager {
         //   b) Else, look up the per-session persisted modelId in SQLite
         //      (cross-restart recovery — server restart wipes pi's in-memory
         //      session model state, but our durable store has the last value).
-        //   c) Else, leave model selection to pi (pre-Phase-3 behaviour).
+        //   c) Else, ask the bridge for the first available model from the
+        //      backend whitelist (same sortOrder the frontend uses for its
+        //      default picker display).
         //
         // We apply BEFORE persisting the user message: if the bridge can't
         // resolve the model (unknown id, registry unavailable), it has already
@@ -242,7 +244,9 @@ export class SessionStreamManager {
         // Persistence: only the envelope-supplied value is written back. A
         // recovery-only application (case b) doesn't update the row — the
         // value is already there.
-        const effectiveModelId = modelId ?? this.store.getSessionModel(sessionId) ?? undefined;
+        const effectiveModelId = modelId ??
+            this.store.getSessionModel(sessionId) ??
+            stream.bridge.getFirstAvailableModelId?.();
         if (effectiveModelId && stream.bridge.setModel) {
             const ok = await stream.bridge.setModel(effectiveModelId);
             if (!ok) {
@@ -252,9 +256,9 @@ export class SessionStreamManager {
                 return;
             }
         }
-        if (modelId) {
+        if (effectiveModelId) {
             try {
-                this.store.setSessionModel(sessionId, modelId);
+                this.store.setSessionModel(sessionId, effectiveModelId);
             }
             catch (err) {
                 // Persisting the sticky model is best-effort: the live turn will

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aexol/spectral",
-  "version": "0.2.22",
+  "version": "0.3.1",
   "description": "Always-on coding agent for Aexol — branded pi wrapper with relay-based browser access.",
   "type": "module",
   "private": false,