nexrall-code 0.5.100 → 0.5.102

package/dist/index.js

@@ -153214,6 +153214,59 @@ function normaliseModelId(model) {
     return "claude-sonnet-5";
   return LEGACY_MODEL_ALIASES[raw.toLowerCase()] ?? raw;
 }
+var MODEL_EFFORT_STYLE = {
+  "gpt-5.4": "openai_gated",
+  "gpt-5.4-mini": "openai_gated",
+  "gpt-4.1": "none",
+  "gpt-4o-mini": "none",
+  "deepseek-v4-pro": "deepseek",
+  "deepseek-v4-flash": "deepseek",
+  "qwen3.7-max": "qwen",
+  "glm-5.3": "glm"
+  // Everything else (all claude-* ids) falls through to 'anthropic' below.
+};
+var EFFORT_STYLE_CONFIG = {
+  // Anthropic's real output_config.effort enum is low|medium|high|xhigh|max
+  // (docs.claude.com extended output). 'extra'/'ultra' are legacy UI labels
+  // this codebase invented, translated server-side by routes/code.js's
+  // EFFORT_MAP (extra->xhigh, ultra->max) — kept as the wire values here
+  // (rather than switching to 'xhigh'/'max' directly) only for backward
+  // compatibility with saved sessions/scripts already passing them.
+  anthropic: { levels: ["low", "medium", "high", "extra", "ultra"], names: ["Low", "Medium", "High", "Extra High", "Max"] },
+  // DeepSeek V4 has THREE real buckets — live-verified against
+  // api.deepseek.com 2026-08-30 (see streamFactory.test.js's DEEPSEEK case):
+  // low->low, medium/high/xhigh->high, max->max.
+  deepseek: { levels: ["low", "high", "max"], names: ["Low", "Standard", "Max"] },
+  // Qwen3.7-Max's knob is a boolean (enable_thinking), not a graded scale.
+  qwen: { levels: ["low", "high"], names: ["Thinking Off", "Thinking On"] },
+  // GLM-5.3 cannot disable reasoning at all (a hard 400 if you try) and has
+  // 3 real buckets (low/high/max) via a FLAT top-level `reasoning_effort`
+  // string — live-verified against api.z.ai 2026-08-30 (see
+  // streamFactory.test.js's GLM_THINKING_LEVEL case).
+  glm: { levels: ["low", "high", "max"], names: ["Low", "High", "Max"] },
+  // Effectively a no-op today whenever the turn has tools (almost always true
+  // for nexrall-code) — see backend modelRegistry.js's OPENAI_GATED doc
+  // comment. Still forwards a real value for the rare tool-less turn.
+  openai_gated: { levels: ["low", "medium", "high"], names: ["Low", "Medium", "High"] },
+  // No reasoning-effort concept on this model at all (gpt-4.1, gpt-4o-mini).
+  none: { levels: ["medium"], names: ["N/A"] }
+};
+function effortStyleFor(modelId) {
+  return MODEL_EFFORT_STYLE[normaliseModelId(modelId)] ?? "anthropic";
+}
+function effortConfigFor(modelId) {
+  return EFFORT_STYLE_CONFIG[effortStyleFor(modelId)];
+}
+function clampEffortForModel(effort, modelId) {
+  const config = effortConfigFor(modelId);
+  if (config.levels.includes(effort))
+    return effort;
+  return defaultEffortForModel(modelId);
+}
+function defaultEffortForModel(modelId) {
+  const levels = effortConfigFor(modelId).levels;
+  return levels[Math.min(1, levels.length - 1)];
+}
 function resolveModelLabel(alias) {
   const id = normaliseModelId(alias);
   return MODEL_LABELS[id] ?? id;
@@ -153664,6 +153717,41 @@ 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`
@@ -153755,7 +153843,7 @@ function printHelp() {
     ["/help", "Show this help"],
     ["/model [name]", "Switch model (e.g. claude-opus-5, gpt-5.4)"],
     ["/mode [ask|edit|plan|auto]", "Set agent mode"],
-    ["/effort [low|medium|high|extra]", "Set thinking effort level"],
+    ["/effort [level]", "Set thinking effort level (options depend on the current model \u2014 run /effort with no argument to see them)"],
     ["/yolo", "Auto-approve all permissions"],
     ["/balance", "Show wallet balance"],
     ["/add <filepath>", "Add a file to conversation context"],
@@ -153876,8 +153964,7 @@ Set ${TRUST_ENV_VAR}=1 to confirm non-interactively, or run \`nex\` interactivel
   let sessionId = crypto4.randomBytes(8).toString("hex");
   let sessionTitle = "";
   let agentMode = "auto";
-  const VALID_EFFORT_LEVELS = ["low", "medium", "high", "extra"];
-  let effortLevel = options.effort && VALID_EFFORT_LEVELS.includes(options.effort) ? options.effort : "medium";
+  let effortLevel = options.effort ? clampEffortForModel(options.effort, modelAlias) : defaultEffortForModel(modelAlias);
   const permRules = initPermissions(workDir);
   const ruleCount = permRules.allow.length + permRules.ask.length + permRules.deny.length;
   if (ruleCount && !headless)
@@ -154356,7 +154443,13 @@ ${dirList}`;
             console.log(source_default.dim(`  Options: ${SELECTABLE_MODELS.map(modelWithCostHint).join(" \xB7 ")}`));
           } else if (SELECTABLE_MODELS.includes(normaliseModelId(requested))) {
             modelAlias = normaliseModelId(requested);
+            const clamped = clampEffortForModel(effortLevel, modelAlias);
+            const effortChanged = clamped !== effortLevel;
+            effortLevel = clamped;
             console.log(source_default.green(`  Model \u2192 ${source_default.bold(resolveModelLabel(modelAlias))}`));
+            if (effortChanged) {
+              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(", ")}`));
           }
@@ -154383,13 +154476,15 @@ ${dirList}`;
         }
         case "/effort": {
           const effortArg = arg.toLowerCase();
+          const config = effortConfigFor(modelAlias);
+          const optionsLine = config.levels.map((v, i2) => `${v} (${config.names[i2]})`).join(" \xB7 ");
           if (!effortArg) {
-            console.log(source_default.dim(`  Current: ${source_default.cyan(effortLevel)}  Options: low \xB7 medium \xB7 high \xB7 extra`));
-          } else if (["low", "medium", "high", "extra"].includes(effortArg)) {
+            console.log(source_default.dim(`  Current: ${source_default.cyan(effortLevel)}  Options: ${optionsLine}`));
+          } else if (config.levels.includes(effortArg)) {
             effortLevel = effortArg;
             console.log(source_default.green(`  Effort \u2192 ${source_default.bold(effortArg)}`));
           } else {
-            console.log(source_default.red(`  Unknown: ${effortArg}`));
+            console.log(source_default.red(`  Unknown: ${effortArg}. Options: ${optionsLine}`));
           }
           rl.prompt();
           return;
@@ -155151,7 +155246,7 @@ program2.command("sessions").description("List or delete saved chat sessions").o
   console.log(source_default.dim("  nex sessions --delete <id>  to remove one"));
   console.log();
 });
-program2.option("-p, --pro", "Use Claude Opus 5 (most capable)").option("-t, --turbo", "Use Claude Sonnet 5 (fast, default)").option("-u, --ultra", "Use Claude Fable 5 (most powerful)").option("-y, --yolo", "Auto-approve all tool calls (no permission prompts)").option("-d, --dir <path>", "Set working directory (default: current directory)").option("-m, --model <name>", "Explicit model, e.g. claude-opus-5 | gpt-5.4 | gpt-4.1").option("-r, --resume [id]", "Resume last session, or a specific session id").option("-e, --effort <level>", "Thinking effort: low | medium | high | extra (default: medium)").option("--no-banner", "Skip the ASCII banner (useful in scripts/pipes)").option("--output-format <fmt>", "One-shot output format: text | json | stream-json (implies auto-approve)").option("--audit [path]", "Log every tool call (and refusal) to a JSONL trail (default: .nexrall/audit/<session>.jsonl)").argument("[prompt...]", "One-shot prompt \u2014 if omitted, starts interactive mode").action(async (promptParts, options) => {
+program2.option("-p, --pro", "Use Claude Opus 5 (most capable)").option("-t, --turbo", "Use Claude Sonnet 5 (fast, default)").option("-u, --ultra", "Use Claude Fable 5 (most powerful)").option("-y, --yolo", "Auto-approve all tool calls (no permission prompts)").option("-d, --dir <path>", "Set working directory (default: current directory)").option("-m, --model <name>", "Explicit model, e.g. claude-opus-5 | gpt-5.4 | gpt-4.1").option("-r, --resume [id]", "Resume last session, or a specific session id").option("-e, --effort <level>", "Thinking effort \u2014 real levels depend on the model (e.g. DeepSeek: low|high|max, Qwen: low|high); run `nex --model <id>` then `/effort` with no argument to see a model's scale").option("--no-banner", "Skip the ASCII banner (useful in scripts/pipes)").option("--output-format <fmt>", "One-shot output format: text | json | stream-json (implies auto-approve)").option("--audit [path]", "Log every tool call (and refusal) to a JSONL trail (default: .nexrall/audit/<session>.jsonl)").argument("[prompt...]", "One-shot prompt \u2014 if omitted, starts interactive mode").action(async (promptParts, options) => {
   if (!(0, import_code_core7.isAuthenticated)()) {
     console.error("Not logged in. Run: nex auth");
     process.exit(1);
@@ -155186,8 +155281,9 @@ program2.option("-p, --pro", "Use Claude Opus 5 (most capable)").option("-t, --t
   let effort;
   if (options.effort) {
     const lvl = String(options.effort).toLowerCase();
-    if (!["low", "medium", "high", "extra"].includes(lvl)) {
-      console.error(`Invalid --effort "${options.effort}". Use: low | medium | high | extra`);
+    const config = effortConfigFor(normaliseModelId(model));
+    if (!config.levels.includes(lvl)) {
+      console.error(`Invalid --effort "${options.effort}" for model "${model}". Use: ${config.levels.join(" | ")}`);
       process.exit(1);
     }
     effort = lvl;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nexrall-code",
-  "version": "0.5.100",
+  "version": "0.5.102",
   "description": "Nexrall Code — AI coding assistant for your terminal (headless agent for scripts, CI and automation)",
   "keywords": [
     "ai",