nexrall-code 0.5.101 → 0.5.103

package/dist/index.js

@@ -103250,6 +103250,94 @@ var require_audit = __commonJS({
   }
 });
 
+// ../core/dist/agent/modelCatalogue.js
+var require_modelCatalogue = __commonJS({
+  "../core/dist/agent/modelCatalogue.js"(exports2) {
+    "use strict";
+    Object.defineProperty(exports2, "__esModule", { value: true });
+    exports2.getModelCatalogue = getModelCatalogue2;
+    exports2.getCachedModelCatalogue = getCachedModelCatalogue;
+    exports2.liveContextWindowFor = liveContextWindowFor;
+    exports2.liveMaxOutputTokensFor = liveMaxOutputTokensFor;
+    exports2.liveSupportsImageInput = liveSupportsImageInput2;
+    exports2.liveSupportsPdfInput = liveSupportsPdfInput2;
+    exports2.liveReasoningEffortStyle = liveReasoningEffortStyle2;
+    exports2.liveVendorFor = liveVendorFor2;
+    exports2.liveLabelFor = liveLabelFor2;
+    exports2.liveCostMultiplierFor = liveCostMultiplierFor2;
+    exports2.liveSelectableModelIds = liveSelectableModelIds2;
+    exports2._resetModelCatalogueCacheForTests = _resetModelCatalogueCacheForTests;
+    var client_1 = require_client();
+    var CACHE_TTL_MS = 5 * 60 * 1e3;
+    var _state2 = null;
+    var _inFlight = null;
+    async function getModelCatalogue2(opts = {}) {
+      const fresh = _state2 && Date.now() - _state2.fetchedAt < CACHE_TTL_MS;
+      if (fresh && !opts.force)
+        return _state2;
+      if (_inFlight)
+        return _inFlight;
+      _inFlight = (async () => {
+        try {
+          const { models, defaultModel } = await (0, client_1.getCodeModels)();
+          if (!models.length)
+            return _state2;
+          const next = { models, defaultModel, fetchedAt: Date.now() };
+          _state2 = next;
+          return next;
+        } catch {
+          return _state2;
+        } finally {
+          _inFlight = null;
+        }
+      })();
+      return _inFlight;
+    }
+    function getCachedModelCatalogue() {
+      return _state2;
+    }
+    function findModel(id) {
+      if (!id || !_state2)
+        return void 0;
+      return _state2.models.find((m2) => m2.id === id);
+    }
+    function liveContextWindowFor(id, fallback) {
+      return findModel(id)?.contextWindow ?? fallback;
+    }
+    function liveMaxOutputTokensFor(id, fallback) {
+      return findModel(id)?.maxOutputTokens ?? fallback;
+    }
+    function liveSupportsImageInput2(id, fallback) {
+      const m2 = findModel(id);
+      return m2 ? m2.supportsImageInput : fallback;
+    }
+    function liveSupportsPdfInput2(id, fallback) {
+      const m2 = findModel(id);
+      return m2 ? m2.supportsPdfInput : fallback;
+    }
+    function liveReasoningEffortStyle2(id) {
+      return findModel(id)?.reasoningEffortStyle;
+    }
+    function liveVendorFor2(id) {
+      return findModel(id)?.vendor;
+    }
+    function liveLabelFor2(id, fallback) {
+      return findModel(id)?.label ?? fallback;
+    }
+    function liveCostMultiplierFor2(id, fallback) {
+      const m2 = findModel(id);
+      return m2 && typeof m2.costMultiplier === "number" ? m2.costMultiplier : fallback;
+    }
+    function liveSelectableModelIds2() {
+      return _state2 ? _state2.models.filter((m2) => m2.available !== false).map((m2) => m2.id) : null;
+    }
+    function _resetModelCatalogueCacheForTests() {
+      _state2 = null;
+      _inFlight = null;
+    }
+  }
+});
+
 // ../core/dist/agent/loop.js
 var require_loop = __commonJS({
   "../core/dist/agent/loop.js"(exports2) {
@@ -103355,6 +103443,7 @@ var require_loop = __commonJS({
     var audit_1 = require_audit();
     var memory_1 = require_memory();
     var safeSlice_1 = require_safeSlice();
+    var modelCatalogue_1 = require_modelCatalogue();
     var fs9 = __importStar(__require("fs"));
     var path8 = __importStar(__require("path"));
     var child_process_1 = __require("child_process");
@@ -104184,6 +104273,16 @@ ${partial}` : "",
       "gpt-5.4-mini": 272e3,
       "gpt-4.1": 1047576,
       "gpt-4o-mini": 128e3,
+      // GPT-5.6 family: MEASURED 2026-09-03 via a 400 (same technique as
+      // gpt-5.4's 922_000 above) — and it is the EXACT SAME 922,000-token
+      // ceiling on all three sizes. This CONTRADICTS the publicly documented
+      // figure (openai.com/index/gpt-5-6 + OpenRouter's model card both
+      // advertise 1,050,000) — see backend services/providers/modelRegistry.js's
+      // own comment on these rows for the measured 400 body. Guessing 1.05M here
+      // would fire auto-compaction ~12% past the real wall.
+      "gpt-5.6-sol": 922e3,
+      "gpt-5.6-terra": 922e3,
+      "gpt-5.6-luna": 922e3,
       // DeepSeek, by real model id (documented — see backend/services/providers/
       // modelRegistry.js's own TODO(unverified-by-400): DeepSeek accepts an
       // oversized max_completion_tokens without rejecting it, so there was no 400
@@ -104202,7 +104301,8 @@ ${partial}` : "",
       "glm-5.3": 1048576
     };
     function contextWindowFor2(model) {
-      return MODEL_CONTEXT_TOKENS[model ?? "turbo"] ?? 128e3;
+      const fallback = MODEL_CONTEXT_TOKENS[model ?? "turbo"] ?? 128e3;
+      return (0, modelCatalogue_1.liveContextWindowFor)(model, fallback);
     }
     function envFraction(name, fallback) {
       const v = Number(process.env[name]);
@@ -104523,7 +104623,7 @@ Continue the work from here.` }] });
       const settings = (0, rules_1.loadSettings)(opts.workDir);
       if (!resolveAutoCompact(void 0, settings.raw))
         return false;
-      const contextWindow = MODEL_CONTEXT_TOKENS[opts.model ?? "turbo"] ?? 1e6;
+      const contextWindow = (0, modelCatalogue_1.liveContextWindowFor)(opts.model, MODEL_CONTEXT_TOKENS[opts.model ?? "turbo"] ?? 1e6);
       let bodyBytes = estimateBodyBytes2(messages);
       let tokenGuess = estimateTokensRough2(messages);
       const overPruneThreshold = () => tokenGuess > contextWindow * AUTO_PRUNE_THRESHOLD || bodyBytes > MAX_BODY_BYTES;
@@ -104638,7 +104738,7 @@ ${options.nexrallMd}` : "") : options.nexrallMd;
       const autoContinue = resolveAutoContinue(options.autoContinue, settings.raw);
       const autoCompact = resolveAutoCompact(options.autoCompact, settings.raw);
       const verifyNudgeOn = resolveVerificationNudge(settings.raw);
-      const contextWindow = MODEL_CONTEXT_TOKENS[model] ?? 2e5;
+      const contextWindow = (0, modelCatalogue_1.liveContextWindowFor)(model, MODEL_CONTEXT_TOKENS[model] ?? 2e5);
       let lastPromptTokens = 0;
       let compacting = false;
       const hardCap = autoContinue ? Math.max(maxIterations, MAX_ITERATIONS_CEILING) : maxIterations;
@@ -107326,6 +107426,7 @@ var require_dist3 = __commonJS({
     __exportStar(require_client(), exports2);
     __exportStar(require_executor(), exports2);
     __exportStar(require_loop(), exports2);
+    __exportStar(require_modelCatalogue(), exports2);
     __exportStar(require_testIntegrity(), exports2);
     __exportStar(require_editCompleteness(), exports2);
     __exportStar(require_securityLint(), exports2);
@@ -153152,6 +153253,9 @@ var MODEL_LABELS = {
   "claude-opus-5": "Claude Opus 5",
   "claude-fable-5": "Claude Fable 5",
   "claude-haiku-4-5-20251001": "Claude Haiku 4.5",
+  "gpt-5.6-sol": "GPT-5.6 Sol",
+  "gpt-5.6-terra": "GPT-5.6 Terra",
+  "gpt-5.6-luna": "GPT-5.6 Luna",
   "gpt-5.4": "GPT-5.4",
   "gpt-5.4-mini": "GPT-5.4 Mini",
   "gpt-4.1": "GPT-4.1",
@@ -153172,6 +153276,9 @@ var SELECTABLE_MODELS = [
   "claude-sonnet-5",
   "claude-opus-5",
   "claude-fable-5",
+  "gpt-5.6-sol",
+  "gpt-5.6-terra",
+  "gpt-5.6-luna",
   "gpt-5.4",
   "gpt-5.4-mini",
   "gpt-4.1",
@@ -153184,6 +153291,13 @@ var MODEL_COST_MULTIPLIER = {
   "claude-sonnet-5": 1.5,
   "claude-opus-5": 3,
   "claude-fable-5": 5,
+  // GPT-5.6 family (2026-07 launch, Terra/Luna repriced 2026-07-30) — same
+  // relative-to-gpt-4.1/qwen3.7-max (=1x) signal as every other row here,
+  // off backend migration 132's per-1M-token seed prices (Sol $5/$30, Terra
+  // $2/$12, Luna $0.20/$1.20 vs gpt-4.1's $2/$8).
+  "gpt-5.6-sol": 2,
+  "gpt-5.6-terra": 1.2,
+  "gpt-5.6-luna": 0.1,
   "gpt-5.4": 2,
   "gpt-5.4-mini": 0.5,
   "gpt-4.1": 1,
@@ -153197,17 +153311,29 @@ var MODEL_COST_MULTIPLIER = {
   "glm-5.3": 0.6
 };
 function modelWithCostHint(id) {
-  const mult = MODEL_COST_MULTIPLIER[id];
+  const mult = (0, import_code_core4.liveCostMultiplierFor)(id, MODEL_COST_MULTIPLIER[id]);
   return mult === void 0 ? id : `${id} (${mult}x)`;
 }
 var NO_VISION_MODELS = /* @__PURE__ */ new Set(["deepseek-v4-pro", "deepseek-v4-flash", "qwen3.7-max", "glm-5.3"]);
 var NO_PDF_MODELS = /* @__PURE__ */ new Set([
+  "gpt-5.6-sol",
+  "gpt-5.6-terra",
+  "gpt-5.6-luna",
   "gpt-5.4",
   "gpt-5.4-mini",
   "gpt-4.1",
   "gpt-4o-mini",
   ...NO_VISION_MODELS
 ]);
+function needsVisionSidecar(model) {
+  return !(0, import_code_core4.liveSupportsImageInput)(model, !NO_VISION_MODELS.has(model));
+}
+function needsPdfSidecar(model) {
+  return !(0, import_code_core4.liveSupportsPdfInput)(model, !NO_PDF_MODELS.has(model));
+}
+function selectableModelIds() {
+  return (0, import_code_core4.liveSelectableModelIds)() ?? SELECTABLE_MODELS;
+}
 function normaliseModelId(model) {
   const raw = (model ?? "").trim();
   if (!raw)
@@ -153215,6 +153341,14 @@ function normaliseModelId(model) {
   return LEGACY_MODEL_ALIASES[raw.toLowerCase()] ?? raw;
 }
 var MODEL_EFFORT_STYLE = {
+  // GPT-5.6 Sol/Terra/Luna hit the SAME "reasoning_effort + tools" 400 as
+  // gpt-5.4/mini below (confirmed by multiple third-party integrations
+  // 2026-07: github.com/BerriAI/litellm#33221, github.com/danny-avila/
+  // LibreChat#14231) — OPENAI_GATED already handles that (skip when tools
+  // are present), so no new style is needed for this family.
+  "gpt-5.6-sol": "openai_gated",
+  "gpt-5.6-terra": "openai_gated",
+  "gpt-5.6-luna": "openai_gated",
   "gpt-5.4": "openai_gated",
   "gpt-5.4-mini": "openai_gated",
   "gpt-4.1": "none",
@@ -153251,8 +153385,21 @@ var EFFORT_STYLE_CONFIG = {
   // No reasoning-effort concept on this model at all (gpt-4.1, gpt-4o-mini).
   none: { levels: ["medium"], names: ["N/A"] }
 };
+var _BACKEND_STYLE_TO_LOCAL = {
+  openai_gated: "openai_gated",
+  deepseek: "deepseek",
+  qwen_thinking_toggle: "qwen",
+  glm_thinking_level: "glm",
+  none: "none"
+};
 function effortStyleFor(modelId) {
-  return MODEL_EFFORT_STYLE[normaliseModelId(modelId)] ?? "anthropic";
+  const id = normaliseModelId(modelId);
+  if ((0, import_code_core4.liveVendorFor)(id) === "Anthropic")
+    return "anthropic";
+  const liveStyle = (0, import_code_core4.liveReasoningEffortStyle)(id);
+  if (liveStyle && _BACKEND_STYLE_TO_LOCAL[liveStyle])
+    return _BACKEND_STYLE_TO_LOCAL[liveStyle];
+  return MODEL_EFFORT_STYLE[id] ?? "anthropic";
 }
 function effortConfigFor(modelId) {
   return EFFORT_STYLE_CONFIG[effortStyleFor(modelId)];
@@ -153269,7 +153416,7 @@ function defaultEffortForModel(modelId) {
 }
 function resolveModelLabel(alias) {
   const id = normaliseModelId(alias);
-  return MODEL_LABELS[id] ?? id;
+  return (0, import_code_core4.liveLabelFor)(id, MODEL_LABELS[id] ?? id);
 }
 function tryExec(cmd, cwd2) {
   try {
@@ -153717,6 +153864,41 @@ async function runTurnHeadless(messages, modelAlias, workDir, _abortSignal, env4
       resultText += text;
       emit({ type: "text", text });
     },
+    // ── Reasoning-phase liveness ──────────────────────────────────────
+    // These used to be UNWIRED here while runTurn() (interactive) wired all
+    // three, and the asymmetry was not cosmetic: on a reasoning model at high
+    // effort the thinking phase emits no text and no tool calls, so a headless
+    // consumer saw ZERO events for the entire phase and could not distinguish
+    // "actively reasoning" from "process wedged".
+    //
+    // Measured on a Terminal-Bench 4.0 trial (2026-08-30, deepseek-v4-pro
+    // --effort max): nex-output.jsonl sat at 4 lines for 65 MINUTES while
+    // tcpdump inside the container's netns showed ~25 packets/s still flowing
+    // and the turn ultimately reported 310,872 output tokens. Nothing was
+    // actually wrong — but every signal available to the harness (log line
+    // count, file mtime) said "hung", and the socket/CPU forensics needed to
+    // prove otherwise are not something a CI wrapper can do.
+    //
+    // `thinking_progress` is the load-bearing one: it carries the backend's
+    // cumulative output-token count (routes/code.js's sendProgress, already
+    // throttled to <=5/s server-side, so this cannot flood the log) and fires
+    // DURING the phase. That makes it a real heartbeat.
+    onThinkingProgress: (tokens) => {
+      emit({ type: "thinking_progress", tokens });
+    },
+    // Fires once per turn at message_complete with the full reasoning text.
+    // Typed `thinking` deliberately: the bench's ATIF converter (atif.py)
+    // already routes exactly this event type into the trajectory's reasoning
+    // buffer, so wiring it here also fills in reasoning that was previously
+    // dropped on the floor for every headless run.
+    onThinking: (text) => {
+      emit({ type: "thinking", text });
+    },
+    // onThinkingDelta is intentionally NOT wired. Its chunks concatenate to the
+    // same string `onThinking` emits in full above, so emitting both would
+    // duplicate the entire reasoning trace — on the 310k-token turn measured
+    // above that is megabytes of redundant NDJSON — while adding no liveness
+    // signal `thinking_progress` does not already provide.
     // System notice (mid-run auto-prune/auto-compact) — emit as its own event
     // type instead of falling through to onText, so a stream-json consumer
     // doesn't see compaction housekeeping text mixed into the model's `text`
@@ -153862,6 +154044,8 @@ async function startChatSession(options) {
   const headless = options.outputFormat === "json" || options.outputFormat === "stream-json";
   const { workDir } = options;
   let modelAlias = options.model;
+  void (0, import_code_core4.getModelCatalogue)().catch(() => {
+  });
   const canPrompt = !headless && !options.prompt && !options.stdinText && process.stdin.isTTY === true && process.stdout.isTTY === true;
   if (!trustGrantedByEnv()) {
     if (!canPrompt) {
@@ -154403,10 +154587,11 @@ ${dirList}`;
           return;
         case "/model": {
           const requested = arg.trim();
+          const models = selectableModelIds();
           if (!requested) {
             console.log(source_default.dim(`  Current: ${source_default.cyan(resolveModelLabel(modelAlias))}`));
-            console.log(source_default.dim(`  Options: ${SELECTABLE_MODELS.map(modelWithCostHint).join(" \xB7 ")}`));
-          } else if (SELECTABLE_MODELS.includes(normaliseModelId(requested))) {
+            console.log(source_default.dim(`  Options: ${models.map(modelWithCostHint).join(" \xB7 ")}`));
+          } else if (models.includes(normaliseModelId(requested))) {
             modelAlias = normaliseModelId(requested);
             const clamped = clampEffortForModel(effortLevel, modelAlias);
             const effortChanged = clamped !== effortLevel;
@@ -154416,7 +154601,7 @@ ${dirList}`;
               console.log(source_default.dim(`  Effort reset \u2192 ${effortLevel} (previous level not available on this model)`));
             }
           } else {
-            console.log(source_default.red(`  Unknown: ${requested}. Options: ${SELECTABLE_MODELS.join(", ")}`));
+            console.log(source_default.red(`  Unknown: ${requested}. Options: ${models.join(", ")}`));
           }
           rl.prompt();
           return;
@@ -154545,7 +154730,7 @@ ${content}
           const label = caption ? `[Attached: ${rel}]
 ${caption}` : `[Attached: ${rel}]`;
           const content = [{ type: "text", text: label }];
-          const needsSidecar = isPdf ? NO_PDF_MODELS.has(model) : NO_VISION_MODELS.has(model);
+          const needsSidecar = isPdf ? needsPdfSidecar(model) : needsVisionSidecar(model);
           if (needsSidecar) {
             console.log(source_default.dim(`  ${resolveModelLabel(modelAlias)} can't read ${isPdf ? "PDFs" : "images"} directly \u2014 describing ${rel} with Claude Sonnet 5 first\u2026`));
             try {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nexrall-code",
-  "version": "0.5.101",
+  "version": "0.5.103",
   "description": "Nexrall Code — AI coding assistant for your terminal (headless agent for scripts, CI and automation)",
   "keywords": [
     "ai",