@mewbleh/purrx 1.0.8 → 1.0.10

package/bin/purrx.js

@@ -173,6 +173,7 @@ async function cmdExec(args) {
       registry,
       approval,
       onChange: () => saveSession(session),
+      sessionId: session.id,
     });
   } finally {
     registry.shutdown();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mewbleh/purrx",
-  "version": "1.0.8",
+  "version": "1.0.10",
   "description": "purrx, a lightweight AI coding agent for your terminal",
   "type": "module",
   "main": "src/index.js",

package/src/api/client.js

@@ -1,5 +1,11 @@
 import { API_BASE, CHATGPT_BASE } from "../config.js";
 
+// Models that use the Responses API reasoning fields. gpt-5 family, o-series,
+// and any *-codex variant are reasoning models.
+function isReasoningModel(model) {
+  return /^(gpt-5|o\d|codex)/i.test(model) || /-codex\b/i.test(model);
+}
+
 // Builds the request URL and headers for the resolved auth mode.
 /**
  * @param {import("../types.js").AuthInfo} authInfo
@@ -11,8 +17,9 @@ function buildRequest(authInfo) {
     const headers = {
       "Content-Type": "application/json",
       Authorization: `Bearer ${authInfo.accessToken}`,
-      "OpenAI-Beta": "responses=experimental",
+      // Codex backend identifies the client via originator + session headers.
       originator: "codex_cli_rs",
+      "User-Agent": "purrx",
     };
     if (authInfo.accountId) {
       headers["chatgpt-account-id"] = authInfo.accountId;
@@ -40,10 +47,14 @@ function buildRequest(authInfo) {
  * @param {import("../types.js").HistoryItem[]} req.input
  * @param {import("../types.js").ToolDefinition[]} [req.tools]
  * @param {string} [req.instructions]
+ * @param {string} [req.sessionId] stable id used for prompt caching
  * @param {{ onText?: (delta: string) => void, onEvent?: (event: any) => void }} [handlers]
  * @returns {Promise<any>}
  */
-export async function streamResponse({ authInfo, model, input, tools, instructions }, handlers = {}) {
+export async function streamResponse(
+  { authInfo, model, input, tools, instructions, sessionId },
+  handlers = {}
+) {
   const { url, headers } = buildRequest(authInfo);
 
   /** @type {Record<string, any>} */
@@ -51,12 +62,26 @@ export async function streamResponse({ authInfo, model, input, tools, instructio
     model,
     input,
     stream: true,
+    // The Codex backend requires store:false; responses are not persisted.
     store: false,
+    // Stable cache key improves latency/cost across turns (matches Codex).
+    prompt_cache_key: sessionId || "purrx",
   };
   if (instructions) payload.instructions = instructions;
+
+  // Reasoning models (gpt-5*, o-series, *-codex) require a reasoning block and
+  // must request the encrypted reasoning content back since we don't persist
+  // responses (store:false). This mirrors the official Codex client and is the
+  // shape the ChatGPT/Codex backend expects.
+  if (isReasoningModel(model)) {
+    payload.reasoning = { effort: "medium", summary: "auto" };
+    payload.include = ["reasoning.encrypted_content"];
+  }
+
   if (tools && tools.length) {
     payload.tools = tools;
     payload.tool_choice = "auto";
+    payload.parallel_tool_calls = true;
   }
 
   const resp = await fetch(url, {

package/src/api/models.js

@@ -1,9 +1,9 @@
-import { API_BASE, CHATGPT_BASE, DEFAULT_MODEL } from "../config.js";
+import { API_BASE, DEFAULT_MODEL } from "../config.js";
 
 // Fetches the list of model ids the current account can actually use.
 // - API-key mode: GET https://api.openai.com/v1/models
 // - ChatGPT mode: the codex backend does not expose a public models list the
-//   same way, so we probe a known-good set and fall back to the default.
+//   same way, so we report a known-good set.
 export async function listModels(authInfo) {
   if (authInfo.mode === "apikey") {
     return listApiKeyModels(authInfo.apiKey);
@@ -24,18 +24,16 @@ async function listApiKeyModels(apiKey) {
       // Keep models usable for chat/agent work (gpt*, o*, codex).
       .filter((/** @type {string} */ id) => /^(gpt|o\d|codex|chatgpt)/i.test(id))
       .sort();
-    return ids.length ? ids : fallbackModels();
+    return ids.length ? dedupeWithDefault(ids) : fallbackModels();
   } catch {
     return fallbackModels();
   }
 }
 
-// The ChatGPT/Codex backend exposes models tied to the plan. There is no
-// stable public list endpoint, so we report the commonly available Codex
-// models. DEFAULT_MODEL is always first.
+// The ChatGPT/Codex backend exposes models tied to the plan. There is no stable
+// public list endpoint, so we report the commonly available Codex models.
 function listChatGptModels() {
-  const known = ["gpt-5-codex", "gpt-5", "gpt-5-mini", "o4-mini", "o3"];
-  return dedupeWithDefault(known);
+  return dedupeWithDefault(["gpt-5-codex", "gpt-5", "gpt-5-mini", "o4-mini", "o3"]);
 }
 
 function fallbackModels() {
@@ -47,8 +45,13 @@ function dedupeWithDefault(list) {
   return [...set];
 }
 
-// Resolves the model to use: explicit override > config/session > a model the
-// account actually has > DEFAULT_MODEL.
+// Resolves the model to use: explicit override > a model the account actually
+// has > DEFAULT_MODEL.
+/**
+ * @param {import("../types.js").AuthInfo} authInfo
+ * @param {string} [preferred]
+ * @returns {Promise<string>}
+ */
 export async function resolveModel(authInfo, preferred) {
   if (preferred) return preferred;
   const available = await listModels(authInfo).catch(() => []);

package/src/config.js

@@ -13,8 +13,8 @@ export const OAUTH_SCOPE =
 export const API_BASE = "https://api.openai.com/v1";
 export const CHATGPT_BASE = "https://chatgpt.com/backend-api/codex";
 
-// Default model. Used as a fallback when discovery is unavailable.
-// gpt-5-codex is the default Codex coding model.
+// Default model. gpt-5-codex is the default Codex coding model and works on
+// both API-key and ChatGPT-account auth. Override with PURRX_MODEL.
 export const DEFAULT_MODEL = process.env.PURRX_MODEL || "gpt-5-codex";
 
 // Approximate context window (in tokens) used to decide when to auto-compact

package/src/core/agent.js

@@ -45,6 +45,7 @@ Guidelines:
  * @param {import("../tools/registry.js").ToolRegistry} opts.registry
  * @param {{requestApproval: (name: string, label: string) => Promise<string>}} [opts.approval]
  * @param {() => void} [opts.onChange]
+ * @param {string} [opts.sessionId] stable id used for prompt caching
  * @returns {Promise<import("../types.js").HistoryItem[]>}
  */
 export async function runTurn({
@@ -56,6 +57,7 @@ export async function runTurn({
   registry,
   approval,
   onChange,
+  sessionId,
 }) {
   history.push({
     type: "message",
@@ -84,22 +86,28 @@ export async function runTurn({
 
     spinner.start();
 
-    const finalResponse = await streamResponse(
-      {
-        authInfo,
-        model,
-        input: history,
-        tools,
-        instructions: systemInstructions(cwd, registry),
-      },
-      {
-        onText: (delta) => {
-          buffered += delta;
+    let finalResponse;
+    try {
+      finalResponse = await streamResponse(
+        {
+          authInfo,
+          model,
+          input: history,
+          tools,
+          instructions: systemInstructions(cwd, registry),
+          sessionId,
         },
-      }
-    );
-
-    spinner.stop();
+        {
+          onText: (delta) => {
+            buffered += delta;
+          },
+        }
+      );
+    } finally {
+      // Always stop the spinner, even if the request throws, so it never
+      // keeps spinning under an error message.
+      spinner.stop();
+    }
 
     if (!finalResponse) {
       throw new Error("No response received from the model.");

package/src/ui/tui.js

@@ -101,6 +101,7 @@ export async function startTui(authInfo, options = {}) {
         registry,
         approval: state.approval,
         onChange: persist,
+        sessionId: session.id,
       });
     } catch (err) {
       console.log(box(chalk.red(err.message), { title: "error", borderColor: "red" }));