@agentprojectcontext/apx 1.77.3 → 1.79.0

package/src/core/deck/manifest.js

@@ -1,6 +1,6 @@
 // APX Deck manifest — the data model the companion clients (deck, desktop
 // capsule) read on boot. Pure data + decoration; no HTTP or filesystem.
-// host/daemon/api/deck.js wraps this for the /deck/manifest endpoint.
+// host/daemon/api/deck.js wraps this for the /api/deck/manifest endpoint.
 
 export const CORE_WIDGETS = [
   {
@@ -75,19 +75,19 @@ export const SAFE_ACTIONS = [
     id: "apx.copy_context",
     title: "Copiar contexto APX",
     risk: "safe",
-    endpoint: "/projects/:pid/agents",
+    endpoint: "/api/projects/:pid/agents",
   },
   {
     id: "apx.voice_turn",
     title: "Hablar con APX",
     risk: "safe",
-    endpoint: "/voice/turn",
+    endpoint: "/api/voice/turn",
   },
   {
     id: "apx.super_agent",
     title: "Pedir acción a APX",
     risk: "confirm",
-    endpoint: "/projects/:pid/super-agent/chat",
+    endpoint: "/api/projects/:pid/super-agent/chat",
   },
 ];
 
@@ -131,7 +131,7 @@ export function decorateExternalWidgets(pluginStatus = {}, overrides = {}) {
 }
 
 /**
- * Build the full /deck/manifest response body.
+ * Build the full /api/deck/manifest response body.
  *
  * Inputs are *resolved* runtime values, not the live managers — caller is
  * responsible for catching errors in projects.list()/plugins.status() and
@@ -167,13 +167,13 @@ export function buildDeckManifest({
       projects: projectList,
       plugins: pluginStatus,
       endpoints: {
-        health: "/health",
-        projects: "/projects",
-        plugins: "/plugins",
-        voice_turn: "/voice/turn",
-        transcribe_chunk: "/transcribe/chunk",
-        super_agent_chat: "/projects/:pid/super-agent/chat",
-        super_agent_stream: "/projects/:pid/super-agent/chat/stream",
+        health: "/api/health",
+        projects: "/api/projects",
+        plugins: "/api/plugins",
+        voice_turn: "/api/voice/turn",
+        transcribe_chunk: "/api/transcribe/chunk",
+        super_agent_chat: "/api/projects/:pid/super-agent/chat",
+        super_agent_stream: "/api/projects/:pid/super-agent/chat/stream",
       },
     },
     safety: {

package/src/core/desktop/autostart.js

@@ -1,6 +1,6 @@
 // Per-user "launch APX Desktop at login" wiring. Used by BOTH the CLI
 // (`apx desktop install` / `uninstall`) and the daemon's
-// /desktop/autostart endpoint so the web admin can toggle the same setting
+// /api/desktop/autostart endpoint so the web admin can toggle the same setting
 // without shelling out.
 //
 // Per-user, never sudo, fully reversible. Three platforms supported:
@@ -16,6 +16,7 @@
 import fs from "node:fs";
 import os from "node:os";
 import path from "node:path";
+import { DESKTOP_AUTOSTART_LOG_PATH } from "#core/config/paths.js";
 import { execFileSync } from "node:child_process";
 import { fileURLToPath } from "node:url";
 import { augmentedPath } from "#core/util/path-env.js";
@@ -28,7 +29,7 @@ export const MAC_PLIST_PATH       = path.join(os.homedir(), "Library", "LaunchAg
 export const LINUX_DESKTOP_PATH   = path.join(os.homedir(), ".config", "autostart", "apx-desktop.desktop");
 export const WIN_RUN_KEY          = "HKCU\\Software\\Microsoft\\Windows\\CurrentVersion\\Run";
 export const WIN_RUN_NAME         = "APXDesktop";
-const AUTOSTART_LOG_PATH          = path.join(os.homedir(), ".apx", "desktop-autostart.log");
+const AUTOSTART_LOG_PATH          = DESKTOP_AUTOSTART_LOG_PATH;
 
 // ── runner resolution ────────────────────────────────────────────────────
 

package/src/core/desktop/process.js

@@ -1,5 +1,5 @@
 // Desktop (Electron floating window) process control — shared by the CLI
-// (`apx desktop start/stop/restart`) and the daemon's /desktop/{start,stop}
+// (`apx desktop start/stop/restart`) and the daemon's /api/desktop/{start,stop}
 // HTTP endpoints, so both spawn/kill the window the exact same way.
 //
 // The window is a detached Electron process (it must survive the spawner so a
@@ -8,8 +8,8 @@
 
 "use strict";
 import fs from "node:fs";
-import os from "node:os";
 import path from "node:path";
+import { DESKTOP_LOG_PATH, DESKTOP_PID_PATH } from "#core/config/paths.js";
 import { spawn, execFileSync } from "node:child_process";
 import { fileURLToPath } from "node:url";
 
@@ -19,8 +19,8 @@ const __dirname = path.dirname(__filename);
 // src/core/desktop/ → repo root is three levels up.
 const ROOT = path.resolve(__dirname, "..", "..", "..");
 export const DESKTOP_MAIN = path.resolve(__dirname, "..", "..", "interfaces", "desktop", "main.js");
-export const DESKTOP_PID = path.join(os.homedir(), ".apx", "desktop.pid");
-const DESKTOP_LOG = path.join(os.homedir(), ".apx", "desktop.log");
+export const DESKTOP_PID = DESKTOP_PID_PATH;
+const DESKTOP_LOG = DESKTOP_LOG_PATH;
 
 // ── PID file ────────────────────────────────────────────────────────────────
 export function readPid() {

package/src/core/engines/gemini.js

@@ -6,7 +6,82 @@ import { randomUUID } from "node:crypto";
 const API_BASE = "https://generativelanguage.googleapis.com/v1beta/models";
 
 function getKey(config) {
-  return config.api_key || process.env.GEMINI_API_KEY || process.env.GOOGLE_API_KEY || "";
+  return getKeys(config)[0] || "";
+}
+
+/**
+ * Every key we may use, in order.
+ *
+ * WHY THIS IS A LIST. Gemini's free tier meters per KEY per MODEL per DAY, and
+ * the good models are metered tightly — 20 requests a day on gemini-3.5-flash.
+ * One key runs out mid-morning and the whole chain collapses to whatever is
+ * last, which in practice was a free OpenRouter router that answers with its
+ * raw chain of thought. Several keys against the same tier multiply the day's
+ * capacity by the number of keys, with no change in behaviour until one is
+ * exhausted.
+ *
+ * `api_key` stays the primary so nothing about the existing single-key config
+ * changes; `api_keys` is additive.
+ */
+export function getKeys(config = {}) {
+  const raw = [
+    config.api_key,
+    ...(Array.isArray(config.api_keys) ? config.api_keys : []),
+    process.env.GEMINI_API_KEY,
+    process.env.GOOGLE_API_KEY,
+  ];
+  const seen = new Set();
+  const out = [];
+  for (const k of raw) {
+    const key = String(k || "").trim();
+    if (!key || seen.has(key)) continue;
+    seen.add(key);
+    out.push(key);
+  }
+  return out;
+}
+
+/**
+ * Keys known to be out of quota for a given model, until the daily reset.
+ *
+ * In memory on purpose: quotas reset on Google's clock, not ours, and a
+ * persisted "this key is dead" file would outlive the reset and permanently
+ * shrink the pool. Losing this on restart costs one wasted request per key.
+ */
+const exhausted = new Map(); // `${model}\u0000${key}` -> epoch ms when it expires
+
+function cooldownKey(model, key) {
+  return `${model}\u0000${key}`;
+}
+
+function isExhausted(model, key, now = Date.now()) {
+  const until = exhausted.get(cooldownKey(model, key));
+  if (!until) return false;
+  if (until <= now) {
+    exhausted.delete(cooldownKey(model, key));
+    return false;
+  }
+  return true;
+}
+
+/** Park a key until the next UTC midnight, which is when Google's RPD resets. */
+function markExhausted(model, key, now = Date.now()) {
+  const reset = new Date(now);
+  reset.setUTCHours(24, 0, 0, 0);
+  exhausted.set(cooldownKey(model, key), reset.getTime());
+}
+
+/** Did this response mean "this key has no quota left", as opposed to a real error? */
+function isQuotaError(status, json) {
+  if (status === 429) return true;
+  const msg = String(json?.error?.message || "");
+  const reason = String(json?.error?.status || "");
+  return reason === "RESOURCE_EXHAUSTED" || /quota|rate limit/i.test(msg);
+}
+
+/** Test-only: forget which keys are parked. */
+export function _resetKeyCooldowns() {
+  exhausted.clear();
 }
 
 // Convert OpenAI-style tool schemas (`{ type: "function", function: { name,
@@ -27,59 +102,193 @@ function toGeminiTools(toolSchemas) {
   ];
 }
 
-// Map our message history into Gemini's `contents` array. Tool results land
-// as `role: "function"` parts with a `functionResponse`. Function calls
-// emitted by the model in earlier turns become `functionCall` parts under
-// `role: "model"`.
-function toGeminiContents(messages) {
+function signatureOf(part) {
+  return part?.thoughtSignature || part?.thought_signature || null;
+}
+
+function callSignatureOf(tc) {
+  return tc?._thoughtSignature || tc?.thought_signature || null;
+}
+
+function isFunctionCallPart(p) {
+  return !!(p?.functionCall || p?.function_call);
+}
+
+// A raw part is replayable if it still holds content. Empty objects would be
+// rejected by the API, and we never invent parts we didn't receive.
+function isUsablePart(p) {
+  return !!p && typeof p === "object" && Object.keys(p).length > 0;
+}
+
+function callArgsOf(tc) {
+  const raw =
+    typeof tc.function?.arguments === "string"
+      ? safeParseJson(tc.function.arguments)
+      : tc.function?.arguments || tc.arguments || {};
+  return raw && typeof raw === "object" ? raw : {};
+}
+
+// Which models stamp a thoughtSignature on their own turns — and therefore
+// demand it back on every functionCall part we replay. Declarative on purpose:
+// covering a new family is one glob in this list, never a new branch in the
+// code below.
+//
+// Patterns are shell-style globs matched against the bare model id (no
+// provider prefix): `*` is any run of characters, `?` a single one.
+// Verified live, one tool call per model — all return the signature:
+//   gemini-2.5-flash, gemini-3-flash-preview, gemini-3.1-flash-lite,
+//   gemini-3.6-flash. (gemini-2.0-* is retired by Google and never signed.)
+//
+// The 4.x/5.x entries are a forward default: the mechanism is a family trait,
+// not a per-version quirk, so a new release is covered the day it ships.
+// Anything still unlisted — a tuned endpoint, a family we didn't predict — is
+// picked up at runtime by the evidence check in signaturesRequired().
+export const THOUGHT_SIGNATURE_MODELS = [
+  "gemini-2.5*",
+  "gemini-3*",
+  "gemini-4*",
+  "gemini-5*",
+];
+
+function globToRegExp(pattern) {
+  const escaped = String(pattern).replace(/[.+^${}()|[\]\\]/g, "\\$&");
+  return new RegExp(`^${escaped.replace(/\*/g, ".*").replace(/\?/g, ".")}$`, "i");
+}
+
+// Per-install override, no code change required:
+//   engines.gemini.thought_signature_models: ["gemini-3*", "my-tuned-model*"]
+// A configured list REPLACES the built-in default — it is the whole answer to
+// "which models use this mechanism", so an install can also opt out entirely
+// with an explicit empty list.
+export function modelUsesThoughtSignatures(model, config = {}) {
+  const configured = config?.thought_signature_models;
+  const patterns = Array.isArray(configured) ? configured : THOUGHT_SIGNATURE_MODELS;
+  const id = String(model || "");
+  return patterns.some((p) => globToRegExp(p).test(id));
+}
+
+function historyHasSignature(messages) {
+  for (const m of messages) {
+    if (Array.isArray(m?._geminiRawParts) && m._geminiRawParts.some(signatureOf)) return true;
+    if (Array.isArray(m?.tool_calls) && m.tool_calls.some(callSignatureOf)) return true;
+  }
+  return false;
+}
+
+function signaturesRequired(model, messages, config) {
+  if (modelUsesThoughtSignatures(model, config)) return true;
+  // Evidence beats the list: a model nobody declared, but which signed
+  // something earlier in THIS conversation, gets the same treatment.
+  return historyHasSignature(messages);
+}
+
+// Map our message history into Gemini's `contents` array. Tool results land as
+// `functionResponse` parts on a `role: "user"` turn. Function calls emitted by
+// the model in earlier turns become `functionCall` parts under `role: "model"`.
+//
+// Thinking-model history fidelity: when APX stores a Gemini-originated
+// assistant turn it attaches `_geminiRawParts` — the verbatim `parts` array
+// from the API response. We replay those raw parts here instead of
+// reconstructing from `tool_calls`, which would lose the thought parts (and
+// their thoughtSignature) that the signing families require in every
+// subsequent turn. Falling back to reconstruction handles messages that came
+// from a non-Gemini engine or were serialised before this field was added.
+function toGeminiContents(messages, { model = "", config = {} } = {}) {
+  const requireSignatures = signaturesRequired(model, messages, config);
   const out = [];
+  // tool_call ids whose functionCall part we could not replay with a valid
+  // signature and had to narrate as text instead. Their tool results must be
+  // degraded to text too: Gemini rejects a functionResponse that answers a
+  // call it can no longer see in the history.
+  const degraded = new Set();
+
   for (const m of messages) {
     if (m.role === "tool") {
-      out.push({
-        role: "function",
-        parts: [
-          {
-            functionResponse: {
-              name: m.name || m.tool_name || "tool",
-              response: { content: m.content },
-            },
-          },
-        ],
-      });
+      const name = m.name || m.tool_name || "tool";
+      const id = m.tool_call_id || m.id;
+      if (id && degraded.has(id)) {
+        out.push({
+          role: "user",
+          parts: [{ text: `[tool result: ${name}] ${asText(m.content)}` }],
+        });
+        continue;
+      }
+      // Tool results ride under role "user", NOT "function": the newer models
+      // (3.6/3.7) reject `role: "function"` outright —
+      //   400 Role 'function' is not supported. Please use a valid role: …
+      // — while every version accepts a user turn carrying functionResponse
+      // parts. Parallel results merge into a single turn, mirroring the one
+      // model turn that emitted the calls.
+      const part = { functionResponse: { name, response: { content: m.content } } };
+      const prev = out[out.length - 1];
+      if (prev && prev.role === "user" && prev.parts.every((p) => p.functionResponse)) {
+        prev.parts.push(part);
+      } else {
+        out.push({ role: "user", parts: [part] });
+      }
       continue;
     }
+
     if (m.role === "assistant" && Array.isArray(m.tool_calls) && m.tool_calls.length > 0) {
-      out.push({
-        role: "model",
-        parts: m.tool_calls.map((tc) => {
-          const part = {
-            functionCall: {
-              name: tc.function?.name || tc.name,
-              args:
-                typeof tc.function?.arguments === "string"
-                  ? safeParseJson(tc.function.arguments)
-                  : tc.function?.arguments || tc.arguments || {},
-            },
-          };
-          // Gemini 3.x thinking models require us to echo back the
-          // thoughtSignature that came attached to the original functionCall
-          // part, or the API rejects the next turn with 400. We captured it
-          // in the response parser; replay it verbatim when present.
-          const sig = tc._thoughtSignature || tc.thought_signature;
-          if (sig) part.thoughtSignature = sig;
-          return part;
-        }),
-      });
+      // Preferred path: replay the raw parts as Gemini returned them. This
+      // preserves thought parts and their thoughtSignatures verbatim — the
+      // signature belongs to the PART, and rebuilding loses whichever part
+      // was carrying it.
+      const raw = Array.isArray(m._geminiRawParts) ? m._geminiRawParts.filter(isUsablePart) : null;
+      if (
+        raw &&
+        raw.length > 0 &&
+        raw.some(isFunctionCallPart) &&
+        (!requireSignatures || raw.some(signatureOf))
+      ) {
+        out.push({ role: "model", parts: raw });
+        continue;
+      }
+
+      // Fallback: rebuild from tool_calls. Used for calls that never came from
+      // a Gemini response — pseudo-tool calls parsed out of plain text, or a
+      // turn inherited from another engine after a retry-chain model switch.
+      //
+      // The signature is per TURN, not per call: on parallel calls Gemini
+      // stamps only the first part and accepts the siblings bare. So one
+      // signature anywhere in the turn is enough to send it as real calls; a
+      // turn with none at all is what earns the 400.
+      const turnHasSignature = m.tool_calls.some(callSignatureOf);
+      const parts = [];
+      const text = typeof m.content === "string" ? m.content.trim() : "";
+      if (text) parts.push({ text });
+      for (const tc of m.tool_calls) {
+        const name = tc.function?.name || tc.name;
+        const sig = callSignatureOf(tc);
+        if (requireSignatures && !turnHasSignature) {
+          // Nothing to replay: sending this as a functionCall is a guaranteed
+          // 400. Narrate it instead so the model keeps the context without the
+          // API rejecting the turn.
+          if (tc.id) degraded.add(tc.id);
+          parts.push({ text: `[tool call: ${name}] ${JSON.stringify(callArgsOf(tc))}` });
+          continue;
+        }
+        const part = { functionCall: { name, args: callArgsOf(tc) } };
+        if (sig) part.thoughtSignature = sig;
+        parts.push(part);
+      }
+      if (parts.length === 0) parts.push({ text: "" });
+      out.push({ role: "model", parts });
       continue;
     }
+
     out.push({
       role: m.role === "assistant" ? "model" : "user",
-      parts: [{ text: typeof m.content === "string" ? m.content : JSON.stringify(m.content) }],
+      parts: [{ text: asText(m.content) }],
     });
   }
   return out;
 }
 
+function asText(v) {
+  return typeof v === "string" ? v : JSON.stringify(v);
+}
+
 function safeParseJson(s) {
   try { return JSON.parse(s); } catch { return {}; }
 }
@@ -108,12 +317,12 @@ export default {
     config = {},
     signal,
   }) {
-    const key = getKey(config);
-    if (!key) throw new Error("gemini: no api_key (set GEMINI_API_KEY or engines.gemini.api_key)");
+    const keys = getKeys(config);
+    if (!keys.length) throw new Error("gemini: no api_key (set GEMINI_API_KEY or engines.gemini.api_key)");
     if (!model) throw new Error("gemini: model required");
 
     const body = {
-      contents: toGeminiContents(messages),
+      contents: toGeminiContents(messages, { model, config }),
       generationConfig: { temperature, maxOutputTokens: maxTokens },
     };
     if (system) body.systemInstruction = { parts: [{ text: system }] };
@@ -130,26 +339,74 @@ export default {
       }
     }
 
-    const url = `${API_BASE}/${encodeURIComponent(model)}:generateContent?key=${key}`;
-    const res = await fetch(url, {
-      method: "POST",
-      headers: { "content-type": "application/json" },
-      body: JSON.stringify(body),
-      signal,
-    });
-    const json = await res.json();
-    if (!res.ok) {
+    // Try each key in turn, skipping any already known to be out of quota for
+    // THIS model today. A key exhausted on gemini-3.5-flash usually still has
+    // its full allowance on the -lite tiers, so the cooldown is per model.
+    const fresh = keys.filter((k) => !isExhausted(model, k));
+    // All parked? Try them anyway rather than failing without asking — the
+    // cooldown is a guess about Google's clock, and being wrong should cost a
+    // request, not the turn.
+    const order = fresh.length ? fresh : keys;
+
+    let json;
+    let lastError = null;
+    for (let i = 0; i < order.length; i++) {
+      const key = order[i];
+      const url = `${API_BASE}/${encodeURIComponent(model)}:generateContent?key=${key}`;
+      const res = await fetch(url, {
+        method: "POST",
+        headers: { "content-type": "application/json" },
+        body: JSON.stringify(body),
+        signal,
+      });
+      const payload = await res.json();
+
+      if (res.ok) { json = payload; break; }
+
+      if (isQuotaError(res.status, payload)) {
+        markExhausted(model, key);
+        lastError = new Error(
+          `gemini ${res.status}: ${payload?.error?.message || "quota exhausted"}`
+        );
+        continue; // next key
+      }
+
+      // A real error — a bad request, a missing model, a revoked key. Rotating
+      // would just repeat it against every key and turn one clear failure into
+      // N slow ones.
       throw new Error(
-        `gemini ${res.status}: ${json?.error?.message || JSON.stringify(json)}`
+        `gemini ${res.status}: ${payload?.error?.message || JSON.stringify(payload)}`
       );
     }
 
+    if (!json) {
+      throw lastError || new Error("gemini: every key is out of quota for this model");
+    }
+
     const parts = json.candidates?.[0]?.content?.parts || [];
-    const text = parts.map((p) => p.text || "").join("");
+
+    // Thinking models (gemini-2.5-flash, gemini-2.5-pro, etc.) emit thought
+    // parts (p.thought === true) that must NOT be shown to the user — they are
+    // internal reasoning. Only non-thought text parts form the visible reply.
+    const text = parts.filter((p) => !p.thought).map((p) => p.text || "").join("");
+
     // Extract function calls and translate them into the OpenAI-shaped
     // tool_calls the run-agent loop expects.
+    //
+    // KEY FIX: On thinking models the thoughtSignature lives on the THOUGHT
+    // text part (p.thought === true), NOT on the functionCall part. We track
+    // the most recently seen signature as we walk the parts array, so the
+    // functionCall that follows a thought part picks it up correctly.
+    // Signatures that appear directly on a functionCall part are also captured
+    // (some non-thinking Gemini variants may do this).
     const toolCalls = [];
+    let pendingThoughtSig = null;
     for (const p of parts) {
+      // Collect thoughtSignature from wherever Gemini puts it: thought parts
+      // or (less commonly) directly on the functionCall part.
+      const partSig = p.thoughtSignature || p.thought_signature;
+      if (partSig) pendingThoughtSig = partSig;
+
       const fc = p.functionCall || p.function_call;
       if (fc?.name) {
         const tc = {
@@ -160,13 +417,13 @@ export default {
             arguments: typeof fc.args === "string" ? fc.args : JSON.stringify(fc.args || {}),
           },
         };
-        // Thinking models (Gemini 3.x) attach a thoughtSignature to the part
-        // alongside the functionCall. We must replay it on the next request
-        // or the API 400s. Carry it on the tool_call so the next call to
-        // toGeminiContents() can put it back. Underscore prefix marks it as
-        // adapter-private metadata other engines should ignore.
-        const sig = p.thoughtSignature || p.thought_signature;
-        if (sig) tc._thoughtSignature = sig;
+        // Attach the signature we collected above. On the next request
+        // toGeminiContents() puts it back on the part so Gemini 2.5+ doesn't
+        // reject the turn with 400 "missing thought_signature".
+        if (pendingThoughtSig) {
+          tc._thoughtSignature = pendingThoughtSig;
+          pendingThoughtSig = null; // consumed; reset for potential next call
+        }
         toolCalls.push(tc);
       }
     }
@@ -174,6 +431,11 @@ export default {
     return {
       text,
       tool_calls: toolCalls.length > 0 ? toolCalls : undefined,
+      // Raw parts stored so toGeminiContents() can replay the model turn
+      // faithfully (thought parts + functionCalls) on the next request.
+      // This is the most robust way to satisfy Gemini's thought-signature
+      // requirement without any reconstruction logic.
+      _geminiRawParts: parts.length > 0 ? parts : undefined,
       finish_reason: json.candidates?.[0]?.finishReason || null,
       usage: {
         input_tokens: json.usageMetadata?.promptTokenCount || 0,

package/src/core/engines/openai-compatible.js

@@ -1,6 +1,19 @@
 // Shared OpenAI-compatible chat adapter (OpenAI, Groq, OpenRouter, …).
 import { pingUrl } from "./_health.js";
 
+// Adapters stash provider-private metadata on tool_calls under an underscore
+// prefix (e.g. gemini's `_thoughtSignature`). Strict OpenAI-shaped APIs reject
+// unknown properties, and a retry-chain switch can hand us a turn produced by
+// another engine — so scrub those keys on the way out.
+function stripPrivateFields(tc) {
+  if (!tc || typeof tc !== "object") return tc;
+  const out = {};
+  for (const [k, v] of Object.entries(tc)) {
+    if (!k.startsWith("_")) out[k] = v;
+  }
+  return out;
+}
+
 export function createOpenAiCompatibleEngine({
   id,
   defaultBaseUrl,
@@ -84,7 +97,7 @@ export function createOpenAiCompatibleEngine({
           role: m.role,
           content: typeof m.content === "string" ? m.content : JSON.stringify(m.content),
         };
-        if (m.tool_calls)    entry.tool_calls   = m.tool_calls;
+        if (m.tool_calls)    entry.tool_calls   = m.tool_calls.map(stripPrivateFields);
         if (m.tool_call_id)  entry.tool_call_id = m.tool_call_id;
         // Some adapters expect `name` on tool messages; we map from tool_name
         // (what run-agent.js writes) to be safe.
@@ -127,7 +140,13 @@ export function createOpenAiCompatibleEngine({
       }
 
       const choice = json.choices?.[0];
-      const text = choice?.message?.content || "";
+      // Reasoning models on OpenRouter/Groq return their chain of thought in a
+      // SEPARATE field (`reasoning`, or `reasoning_content` on some routers).
+      // It is not an answer and must never reach a channel — fold it into the
+      // <think> form APX already knows how to strip per channel.
+      const reasoning = choice?.message?.reasoning || choice?.message?.reasoning_content || "";
+      const answer = choice?.message?.content || "";
+      const text = reasoning ? `<think>${reasoning}</think>${answer}` : answer;
       const toolCalls = choice?.message?.tool_calls;
 
       return {

package/src/core/engines/presets.js

@@ -4,11 +4,11 @@
 // Who consumes this:
 //   • CLI  — `apx setup` builds its provider/model menus from here
 //     (src/interfaces/cli/commands/setup.js).
-//   • Web  — the admin panel fetches it via `GET /engines/presets` and hydrates
+//   • Web  — the admin panel fetches it via `GET /api/engines/presets` and hydrates
 //     its provider forms (src/interfaces/web/.../providers/typeStyles.ts).
 //
 // This is the OFFLINE / no-key fallback list. When the user has an api_key
-// configured, the daemon's `POST /engines/models` returns the provider's LIVE
+// configured, the daemon's `POST /api/engines/models` returns the provider's LIVE
 // catalog instead (see ./catalog.js). The model field is ALWAYS free-text, so
 // any id can be typed even if it is not listed here.
 //

package/src/core/http-tools/browser.js

@@ -9,7 +9,7 @@
 // browser_* tool is actually called. HTTP-only fetching lives in fetch.js
 // (no Chromium needed).
 //
-// Endpoints (mounted at /tools/browser by api.js):
+// Endpoints (mounted at /api/tools/browser by api.js):
 //   POST /navigate          { url, launch_options?, allow_dangerous? }
 //   POST /screenshot        { selector?, full_page?, width?, height?, encoded? }
 //   POST /click             { selector }