@link-assistant/hive-mind 2.1.9 → 2.2.0

package/CHANGELOG.md

@@ -1,5 +1,24 @@
 # @link-assistant/hive-mind
 
+## 2.2.0
+
+### Minor Changes
+
+- d3b7c97: Default `--tool codex` to `gpt-5.6-sol` and make the `--think` levels map predictably to Codex reasoning efforts (`off`→`none`, `low`/`medium`/`high`/`xhigh`/`ultra`/`max` as an identity mapping). GPT-5.6 Sol's multi-agent `ultra` mode is always paired with a `rollout_token_budget` cap (default `500000`, overridable via `--rollout-token-budget`), and budget-derived effort stays capped at `xhigh`. Align `--tool claude` by adding the matching `ultra` level (equivalent to `ultracode`). By default both tools run the model as-is with no thinking level enforced.
+
+## 2.1.11
+
+### Patch Changes
+
+- 208e123: Fix false-positive npm releases (issue #2028). `setup-npm.mjs` now pins npm to the 11.x line and validates the result instead of installing `npm@latest` (which pulled in npm 12.0.0, whose sigstore regression crashes provenance publishes — npm/cli#9722). `publish-to-npm.mjs` no longer trusts the publish command's exit status alone: it observes the real exit code, scans output for failure patterns, and verifies the version is actually live on npm before reporting success, so a failed publish can no longer be reported as a successful release. Added `sanitize-npm-userconfig.mjs` to remove the deprecated `always-auth` npm warning from release logs.
+
+## 2.1.10
+
+### Patch Changes
+
+- 9964e69: Retry Claude stream-json sessions with `--resume` when the stream ends without a terminal result event after tool output.
+- afbc353: Avoid fetching `use-m` at isolation-runner import time when tests only use pure helper exports.
+
 ## 2.1.9
 
 ### Patch Changes

package/README.hi.md

@@ -39,7 +39,7 @@ Hive Mind एक **सामान्यवादी AI** (मिनी-AGI) ह
 | सदस्यता                                                          | `--tool` के साथ     | डिफ़ॉल्ट मॉडल | किसके लिए बेहतर है                                |
 | ---------------------------------------------------------------- | ------------------- | ------------- | ------------------------------------------------- |
 | **Anthropic Claude MAX** (~$200/माह, अक्सर 50% छूट = $400 मूल्य) | `claude` (डिफ़ॉल्ट) | Sonnet/Haiku  | उच्चतम रचनात्मकता, मजबूत सामान्य कोड रीजनिंग      |
-| **OpenAI ChatGPT Pro** ($200/माह, Codex शामिल)                   | `codex`             | `gpt-5.5`     | भरोसेमंद deterministic refactors और तेज iteration |
+| **OpenAI ChatGPT Pro** ($200/माह, Codex शामिल)                   | `codex`             | `gpt-5.6-sol` | भरोसेमंद deterministic refactors और तेज iteration |
 
 दोनों टूल एक ही hive में साथ उपयोग किए जा सकते हैं। Worker अलग-अलग टूल समानांतर चला सकते हैं, और `/codex` या `/solve --tool codex` कार्यों को ChatGPT Pro पर भेजता है जबकि डिफ़ॉल्ट Claude MAX पर जाता है। किसी एक को चुनना आवश्यक नहीं है: किसी भी एक सदस्यता से संचालन संभव है, और दोनों का उपयोग per-tool/model concurrency mode (#1474) खोलता है।
 

package/README.md

@@ -39,7 +39,7 @@ Hive Mind is a **generalist AI** (mini-AGI) capable of working on a wide range o
 | Subscription                                                       | Pairs with `--tool` | Default model | Best for                                                |
 | ------------------------------------------------------------------ | ------------------- | ------------- | ------------------------------------------------------- |
 | **Anthropic Claude MAX** (~$200/month, often 50% off = $400 value) | `claude` (default)  | Sonnet/Haiku  | Highest creativity, strongest general code reasoning    |
-| **OpenAI ChatGPT Pro** ($200/month, includes Codex)                | `codex`             | `gpt-5.5`     | Strong deterministic refactors and fast iteration loops |
+| **OpenAI ChatGPT Pro** ($200/month, includes Codex)                | `codex`             | `gpt-5.6-sol` | Strong deterministic refactors and fast iteration loops |
 
 Both tools can be combined in the same hive. Workers can run different tools in parallel, and `/codex` or `/solve --tool codex` routes tasks to ChatGPT Pro while the default routes to Claude MAX. There is no requirement to pick one: either single subscription is enough to operate, and using both unlocks per-tool/model concurrency mode (#1474).
 
@@ -507,15 +507,15 @@ Free Models via Kilo Gateway (with --tool agent):
 
 Current tool defaults in Hive Mind:
 
-| Tool       | Default model                                               | Default reasoning behavior                                                               |
-| ---------- | ----------------------------------------------------------- | ---------------------------------------------------------------------------------------- |
-| `claude`   | `sonnet`                                                    | No extra thinking is requested unless you pass `--think` or `--thinking-budget`          |
-| `codex`    | `gpt-5.5` preferred, with runtime fallback to local catalog | Codex runs with `reasoning_effort=none` unless you pass `--think` or `--thinking-budget` |
-| `opencode` | `grok-code-fast-1`                                          | No extra thinking prompt is added for the default model                                  |
-| `agent`    | `nemotron-3-super-free`                                     | No extra thinking prompt is added for the default model                                  |
-| `gemini`   | `flash`                                                     | No extra thinking prompt is added for the default model                                  |
-| `qwen`     | `qwen3-coder-plus`                                          | No extra thinking prompt is added for the default model                                  |
-| `gemini`   | `gemini-2.5-flash`                                          | No extra thinking prompt is added for the default model                                  |
+| Tool       | Default model                                                               | Default reasoning behavior                                                               |
+| ---------- | --------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------- |
+| `claude`   | `sonnet`                                                                    | No extra thinking is requested unless you pass `--think` or `--thinking-budget`          |
+| `codex`    | `gpt-5.6-sol` preferred, with runtime fallback to local catalog (`gpt-5.5`) | Codex runs with `reasoning_effort=none` unless you pass `--think` or `--thinking-budget` |
+| `opencode` | `grok-code-fast-1`                                                          | No extra thinking prompt is added for the default model                                  |
+| `agent`    | `nemotron-3-super-free`                                                     | No extra thinking prompt is added for the default model                                  |
+| `gemini`   | `flash`                                                                     | No extra thinking prompt is added for the default model                                  |
+| `qwen`     | `qwen3-coder-plus`                                                          | No extra thinking prompt is added for the default model                                  |
+| `gemini`   | `gemini-2.5-flash`                                                          | No extra thinking prompt is added for the default model                                  |
 
 See [docs/CONFIGURATION.md](./docs/CONFIGURATION.md) for the full per-tool defaults and reasoning mappings.
 

package/README.ru.md

@@ -39,7 +39,7 @@ Hive Mind — это **универсальный ИИ** (мини-AGI), спо
 | Подписка                                                          | Используется с `--tool` | Модель по умолчанию | Лучше всего подходит для                                   |
 | ----------------------------------------------------------------- | ----------------------- | ------------------- | ---------------------------------------------------------- |
 | **Anthropic Claude MAX** (~$200 в месяц, часто скидка 50% = $400) | `claude` (по умолчанию) | Sonnet/Haiku        | Максимальная креативность и сильное общее кодовое мышление |
-| **OpenAI ChatGPT Pro** ($200 в месяц, включает Codex)             | `codex`                 | `gpt-5.5`           | Надёжные детерминированные рефакторинги и быстрые итерации |
+| **OpenAI ChatGPT Pro** ($200 в месяц, включает Codex)             | `codex`                 | `gpt-5.6-sol`       | Надёжные детерминированные рефакторинги и быстрые итерации |
 
 Оба инструмента можно сочетать в одном hive. Воркеры могут параллельно запускать разные инструменты, а `/codex` или `/solve --tool codex` направляет задачи в ChatGPT Pro, тогда как маршрут по умолчанию идёт в Claude MAX. Выбирать только один вариант не требуется: любой одной подписки достаточно для работы, а использование обеих открывает режим параллелизма по инструментам/моделям (#1474).
 

package/README.zh.md

@@ -36,10 +36,10 @@ Hive Mind 是一款**通用 AI**（迷你 AGI），能够处理广泛的任务
 
 **费用**：Hive Mind 支持两种 $200/月订阅作为功能完整、近乎“无限”的选项：
 
-| 订阅                                                             | 搭配 `--tool`    | 默认模型     | 最适合场景                   |
-| ---------------------------------------------------------------- | ---------------- | ------------ | ---------------------------- |
-| **Anthropic Claude MAX**（约 $200/月，常有五折优惠 = 价值 $400） | `claude`（默认） | Sonnet/Haiku | 最高创造力、最强通用代码推理 |
-| **OpenAI ChatGPT Pro**（$200/月，包含 Codex）                    | `codex`          | `gpt-5.5`    | 稳定确定性的重构和快速迭代   |
+| 订阅                                                             | 搭配 `--tool`    | 默认模型      | 最适合场景                   |
+| ---------------------------------------------------------------- | ---------------- | ------------- | ---------------------------- |
+| **Anthropic Claude MAX**（约 $200/月，常有五折优惠 = 价值 $400） | `claude`（默认） | Sonnet/Haiku  | 最高创造力、最强通用代码推理 |
+| **OpenAI ChatGPT Pro**（$200/月，包含 Codex）                    | `codex`          | `gpt-5.6-sol` | 稳定确定性的重构和快速迭代   |
 
 两个工具可以在同一个 hive 中组合使用。Worker 可以并行运行不同工具，`/codex` 或 `/solve --tool codex` 会将任务路由到 ChatGPT Pro，而默认路由到 Claude MAX。无需二选一：任意单一订阅都足以运行，同时使用两者可解锁按工具/模型划分的并发模式（#1474）。
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@link-assistant/hive-mind",
-  "version": "2.1.9",
+  "version": "2.2.0",
   "description": "AI-powered issue solver and hive mind for collaborative problem solving",
   "main": "src/hive.mjs",
   "type": "module",

package/src/agent-commander.lib.mjs

@@ -62,8 +62,13 @@ const buildClaudeToolOptions = (argv = {}) => {
 
 const buildCodexToolOptions = (argv = {}) => {
   const options = {};
-  const { reasoningEffort } = resolveCodexReasoningEffort(argv);
-  appendExtraArgs(options, ['-c', `model_reasoning_effort=${reasoningEffort}`, '-c', 'model_reasoning_summary=auto']);
+  const { reasoningEffort, rolloutTokenBudget } = resolveCodexReasoningEffort(argv);
+  const reasoningArgs = ['-c', `model_reasoning_effort=${reasoningEffort}`, '-c', 'model_reasoning_summary=auto'];
+  // Issue #2027: pair GPT-5.6 Sol's multi-agent `ultra` effort with a rollout token budget cap.
+  if (rolloutTokenBudget) {
+    reasoningArgs.push('-c', `rollout_token_budget=${rolloutTokenBudget}`);
+  }
+  appendExtraArgs(options, reasoningArgs);
 
   appendExtraArgs(options, buildCodexDisable1mContextConfigArgs(!!argv.disable1mContext));
   try {

package/src/claude.lib.mjs

@@ -1163,6 +1163,11 @@ export const executeClaudeCommand = async params => {
           queuedFeedback,
         };
       }
+      if (shouldFailClaudeStreamWithoutResult({ commandFailed, streamingInput, resultEventReceived })) {
+        commandFailed = true;
+        lastMessage = buildMissingClaudeResultMessage({ lastToolResultError, lastMessage });
+        await log(`\n\n❌ Command failed: ${lastMessage}`, { level: 'error' });
+      }
       const retryableLastError = classifyRetryableError(lastMessage);
       // Issue #1834: Corrupted extended-thinking blocks → try to resume the session first, then fall
       // back to a fresh restart (PR #1835 feedback). When both caps are reached, tryThinkingBlockRecovery
@@ -1291,11 +1296,6 @@ export const executeClaudeCommand = async params => {
           .join('\n');
         await log(`\n\n❌ Command failed: No messages processed and errors detected in stderr\nStderr errors:\n${errorsPreview}`, { level: 'error' });
       }
-      if (shouldFailClaudeStreamWithoutResult({ commandFailed, streamingInput, resultEventReceived })) {
-        commandFailed = true;
-        lastMessage = buildMissingClaudeResultMessage({ lastToolResultError, lastMessage });
-        await log(`\n\n❌ Command failed: ${lastMessage}`, { level: 'error' });
-      }
       if (commandFailed) {
         // Take resource snapshot after failure
         const resourcesAfter = await getResourceSnapshot();

package/src/codex.lib.mjs

@@ -804,7 +804,7 @@ export const executeCodexCommand = async params => {
 
     let execCommand;
     const mappedModel = mapModelToId(argv.model);
-    const { reasoningEffort, source: reasoningEffortSource } = resolveCodexReasoningEffort(argv);
+    const { reasoningEffort, source: reasoningEffortSource, rolloutTokenBudget } = resolveCodexReasoningEffort(argv);
     const isResumeMode = !!argv.resume;
     const codexEnv = getCodexExecEnv(argv.verbose);
 
@@ -839,7 +839,10 @@ export const executeCodexCommand = async params => {
     for (const arg of codexPlaywrightMcpDisableConfigArgs) {
       codexArgs += ` ${shellQuote(arg)}`;
     }
-    codexArgs += ` --json --skip-git-repo-check -o ${shellQuote(lastMessageFile)} -c ${shellQuote(`model_reasoning_effort=${reasoningEffort}`)} -c ${shellQuote('model_reasoning_summary=auto')} --dangerously-bypass-approvals-and-sandbox`;
+    codexArgs += ` --json --skip-git-repo-check -o ${shellQuote(lastMessageFile)} -c ${shellQuote(`model_reasoning_effort=${reasoningEffort}`)} -c ${shellQuote('model_reasoning_summary=auto')}`;
+    // Issue #2027: pair GPT-5.6 Sol's multi-agent `ultra` effort with a rollout token budget cap so it stays predictable and does not run away on cost.
+    if (rolloutTokenBudget) codexArgs += ` -c ${shellQuote(`rollout_token_budget=${rolloutTokenBudget}`)}`;
+    codexArgs += ' --dangerously-bypass-approvals-and-sandbox';
 
     // Issue #1706: Append --disable-1m-context and --sub-session-size as Codex -c overrides.
     let parsedSubSessionSize;

package/src/codex.options.lib.mjs

@@ -4,13 +4,34 @@ import { codexModels } from './models/index.mjs';
 
 export const mapModelToId = model => codexModels[model] || model;
 
+// Issue #2027: Map the shared hive-mind --think levels to Codex `model_reasoning_effort`
+// values. GPT-5.6 Sol (the default Codex model) keeps the full ladder inherited from the
+// GPT-5.5/GPT-5.4 generation — low/medium/high/xhigh — and adds `max` *above* xhigh for the
+// deepest single-agent reasoning, plus a multi-agent `ultra` mode. Because every hive level
+// has a same-named Codex reasoning effort, the mapping is a predictable identity: `xhigh`
+// stays `xhigh` (natively supported per `codex debug models`), `ultra` selects GPT-5.6's
+// multi-agent ultra mode (the counterpart of Claude's "ultracode"), and `max` selects the
+// deepest single-agent effort. `off` disables reasoning (`none`). See docs/case-studies/issue-2027.
 const THINK_LEVEL_TO_CODEX_REASONING = {
   off: 'none',
   low: 'low',
   medium: 'medium',
   high: 'high',
   xhigh: 'xhigh',
-  max: 'xhigh',
+  ultra: 'ultra',
+  max: 'max',
+};
+
+// Issue #2027: GPT-5.6 Sol's multi-agent `ultra` mode spawns subagents and consumes far more
+// tokens per turn than single-agent reasoning. OpenAI's guidance is explicit: never use `ultra`
+// reasoning effort without a `rollout_token_budget` cap, or it can run away on cost. We pair
+// every `ultra` selection with this budget (their recommended default) so `--think ultra` stays
+// predictable. Override with `--rollout-token-budget`.
+export const CODEX_ULTRA_ROLLOUT_TOKEN_BUDGET = 500000;
+
+const resolveUltraRolloutTokenBudget = argv => {
+  const override = argv?.rolloutTokenBudget;
+  return Number.isFinite(override) && override > 0 ? override : CODEX_ULTRA_ROLLOUT_TOKEN_BUDGET;
 };
 
 export const resolveCodexReasoningEffort = argv => {
@@ -26,6 +47,10 @@ export const resolveCodexReasoningEffort = argv => {
     }
 
     const ratio = Math.min(1, thinkingBudget / maxBudget);
+    // Issue #2027: the budget-derived effort caps at `xhigh` — the deepest tier every Codex
+    // model (including the gpt-5.5 runtime fallback) supports. `max` is GPT-5.6-only and `ultra`
+    // needs a paired rollout token budget, so both require an explicit `--think max`/`--think ultra`
+    // to stay predictable rather than being reached implicitly through a token budget.
     const reasoningEffort = ratio <= 0.2 ? 'minimal' : ratio <= 0.4 ? 'low' : ratio <= 0.6 ? 'medium' : ratio <= 0.8 ? 'high' : 'xhigh';
 
     return {
@@ -35,10 +60,15 @@ export const resolveCodexReasoningEffort = argv => {
   }
 
   if (argv?.think && THINK_LEVEL_TO_CODEX_REASONING[argv.think]) {
-    return {
-      reasoningEffort: THINK_LEVEL_TO_CODEX_REASONING[argv.think],
+    const reasoningEffort = THINK_LEVEL_TO_CODEX_REASONING[argv.think];
+    const result = {
+      reasoningEffort,
       source: `--think ${argv.think}`,
     };
+    if (reasoningEffort === 'ultra') {
+      result.rolloutTokenBudget = resolveUltraRolloutTokenBudget(argv);
+    }
+    return result;
   }
 
   return {
@@ -50,4 +80,5 @@ export const resolveCodexReasoningEffort = argv => {
 export default {
   mapModelToId,
   resolveCodexReasoningEffort,
+  CODEX_ULTRA_ROLLOUT_TOKEN_BUDGET,
 };

package/src/config.lib.mjs

@@ -397,6 +397,7 @@ export const getThinkingLevelToTokens = (maxBudget = DEFAULT_MAX_THINKING_BUDGET
   medium: Math.floor(maxBudget / 2), // ~16000 for default 31999
   high: Math.floor((maxBudget * 3) / 4), // ~24000 for default 31999
   xhigh: maxBudget, // same as max when represented as MAX_THINKING_TOKENS
+  ultra: maxBudget, // Issue #2027: ultra is max-class in token terms for Claude
   max: maxBudget, // 31999 by default
 });
 
@@ -482,10 +483,13 @@ export const OPUS_46_EFFORT_LEVELS = ['low', 'medium', 'high', 'max'];
 export const OPUS_47_EFFORT_LEVELS = ['low', 'medium', 'high', 'xhigh', 'max'];
 
 /**
- * Convert thinking level to effort level (Issue #1238, Issue #1620)
+ * Convert thinking level to effort level (Issue #1238, Issue #1620, Issue #2027)
  * Models with max support keep max as max. Opus 4.7 keeps xhigh as xhigh.
  * Models with effort but without max support use high for max/xhigh.
- * @param {string|undefined} thinkLevel - The thinking level (off/low/medium/high/xhigh/max)
+ * `ultra` (Issue #2027) requests Claude "ultracode"-class reasoning; Claude Code has no
+ * distinct `ultra`/`ultracode` effort value, so it clamps to the highest supported effort
+ * (max, else xhigh, else high).
+ * @param {string|undefined} thinkLevel - The thinking level (off/low/medium/high/xhigh/ultra/max)
  * @param {Object} [options] - Options
  * @param {boolean} [options.isOpus47] - Backward-compatible shorthand for supportsXHigh
  * @param {boolean} [options.supportsXHigh] - Whether the model supports xhigh effort
@@ -509,6 +513,8 @@ export const thinkLevelToEffortLevel = (thinkLevel, options = {}) => {
       return 'high';
     case 'xhigh':
       return supportsXHigh ? 'xhigh' : supportsMax ? 'max' : 'high';
+    case 'ultra':
+      return supportsMax ? 'max' : supportsXHigh ? 'xhigh' : 'high';
     case 'max':
       return supportsMax ? 'max' : 'high';
     default:

package/src/isolation-runner.lib.mjs

@@ -20,11 +20,26 @@ import os from 'node:os';
 import path from 'node:path';
 import { isExecutingSessionStatus, isTerminalSessionStatus } from './session-status.lib.mjs';
 
-if (typeof use === 'undefined') {
-  await ensureUseM();
-}
+let commandStreamDollarPromise = null;
+
+async function getCommandStreamDollar() {
+  if (!commandStreamDollarPromise) {
+    commandStreamDollarPromise = (async () => {
+      if (typeof globalThis.use === 'undefined') {
+        await ensureUseM();
+      }
+      const { $ } = await globalThis.use('command-stream');
+      return $;
+    })();
+  }
 
-const { $ } = await use('command-stream');
+  try {
+    return await commandStreamDollarPromise;
+  } catch (error) {
+    commandStreamDollarPromise = null;
+    throw error;
+  }
+}
 
 // Re-export the shared status predicates so existing callers that reach them via
 // the isolation-runner module (e.g. session-monitor's `runner.isExecutingSessionStatus`)
@@ -516,6 +531,7 @@ export function readSessionExitFromLog(logPath, options = {}) {
  */
 async function findStartCommandBinary() {
   try {
+    const $ = await getCommandStreamDollar();
     const result = await $({ mirror: false })`which $`;
     const path = result.stdout?.toString().trim() || '';
     return path || null;
@@ -677,6 +693,7 @@ export async function querySessionStatus(sessionId, verbose = false) {
   }
 
   try {
+    const $ = await getCommandStreamDollar();
     const result = await $({ mirror: false })`${binPath} --status ${sessionId} --output-format json`;
 
     const stdout = result.stdout?.toString().trim() || '';
@@ -755,6 +772,7 @@ export async function listIsolationSessions(verbose = false) {
     return [];
   }
   try {
+    const $ = await getCommandStreamDollar();
     const result = await $({ mirror: false })`${binPath} --list --output-format json`;
     const stdout = result.stdout?.toString().trim() || '';
     const sessions = parseSessionListOutput(stdout);
@@ -791,6 +809,7 @@ export async function stopIsolatedSession(sessionId, verbose = false) {
   }
 
   try {
+    const $ = await getCommandStreamDollar();
     const result = await $({ mirror: false })`${binPath} --stop ${sessionId}`;
     const stdout = result.stdout?.toString() || '';
     const stderr = result.stderr?.toString() || '';
@@ -827,6 +846,7 @@ export async function stopIsolatedSession(sessionId, verbose = false) {
  */
 export async function checkScreenSessionRunning(sessionName, verbose = false) {
   try {
+    const $ = await getCommandStreamDollar();
     const result = await $({ mirror: false })`screen -ls`;
     const output = result.stdout?.toString() || '';
     const exists = output.includes(sessionName);
@@ -857,6 +877,7 @@ export async function checkScreenSessionRunning(sessionName, verbose = false) {
  */
 export async function checkDockerContainerRunning(containerName, verbose = false) {
   try {
+    const $ = await getCommandStreamDollar();
     const result = await $({ mirror: false })`docker inspect -f ${'{{.State.Running}}'} ${containerName}`;
     const running = (result.stdout?.toString() || '').trim() === 'true';
     if (verbose) {
@@ -890,6 +911,7 @@ export function parseDockerContainerWritableLayerSizeOutput(output) {
 export async function getDockerContainerWritableLayerSize(containerName, verbose = false) {
   if (!containerName) return null;
   try {
+    const $ = await getCommandStreamDollar();
     const result = await $({ mirror: false })`docker inspect --size -f ${'{{.SizeRw}}'} ${containerName}`;
     const bytes = parseDockerContainerWritableLayerSizeOutput(result.stdout?.toString() || '');
     if (verbose) {
@@ -922,6 +944,7 @@ export async function releaseDockerContainerStartGate(containerName, verbose = f
 
   for (let attempt = 1; attempt <= 5; attempt++) {
     try {
+      const $ = await getCommandStreamDollar();
       await $({ mirror: false })`docker exec ${containerName} sh -c ${releaseCommand}`;
       if (verbose) {
         console.log(`[VERBOSE] isolation-runner: released docker start gate for '${containerName}'`);
@@ -960,6 +983,7 @@ export async function removeDockerContainer(containerName, verbose = false) {
   }
 
   try {
+    const $ = await getCommandStreamDollar();
     const result = await $({ mirror: false })`docker rm -f ${containerName}`;
     const stdout = result.stdout?.toString() || '';
     const stderr = result.stderr?.toString() || '';
@@ -992,6 +1016,7 @@ export async function removeDockerContainer(containerName, verbose = false) {
  */
 export async function checkTmuxSessionRunning(sessionName, verbose = false) {
   try {
+    const $ = await getCommandStreamDollar();
     await $({ mirror: false })`tmux has-session -t ${sessionName}`;
     if (verbose) console.log(`[VERBOSE] isolation-runner: tmux has-session '${sessionName}': running`);
     return true;
@@ -1035,6 +1060,7 @@ export async function checkBackendSessionAlive(sessionId, backend, verbose = fal
  */
 export async function checkDockerImagePresent(image, verbose = false) {
   try {
+    const $ = await getCommandStreamDollar();
     await $({ mirror: false })`docker image inspect ${image}`;
     if (verbose) console.log(`[VERBOSE] isolation-runner: docker image inspect '${image}': present`);
     return true;
@@ -1062,6 +1088,7 @@ export async function checkDockerImagePresent(image, verbose = false) {
  */
 export async function checkDockerStorageDriver(verbose = false) {
   try {
+    const $ = await getCommandStreamDollar();
     const result = await $({ mirror: false })`docker info --format ${'{{.Driver}}'}`;
     const driver = (result.stdout?.toString() || '').trim().toLowerCase() || null;
     if (verbose) console.log(`[VERBOSE] isolation-runner: docker storage driver: ${driver || '(unknown)'}`);
@@ -1091,6 +1118,7 @@ export async function checkDockerDiskSpace(verbose = false) {
   try {
     let dataRoot = '/var/lib/docker';
     try {
+      const $ = await getCommandStreamDollar();
       const info = await $({ mirror: false })`docker info --format ${'{{.DockerRootDir}}'}`;
       const root = (info.stdout?.toString() || '').trim();
       if (root) dataRoot = root;
@@ -1099,6 +1127,7 @@ export async function checkDockerDiskSpace(verbose = false) {
       // fails on it (e.g. the path does not exist) we return null below.
     }
 
+    const $ = await getCommandStreamDollar();
     const df = await $({ mirror: false })`df -Pk ${dataRoot}`;
     // `df -P` guarantees one logical line per filesystem (no wrapping). The last
     // line is the data row: Filesystem 1024-blocks Used Available Capacity Mount

package/src/locales/en.lino

@@ -578,7 +578,7 @@ en
         branch
           option "• `--base-branch <branch>` or `-b` - Target branch for PR (default: repo default branch)"
       think
-        option "• `--think <level>` - Thinking level (off/low/medium/high/xhigh/max) | `--thinking-budget <num>` - Token budget (0-63999)"
+        option "• `--think <level>` - Thinking level (off/low/medium/high/xhigh/ultra/max) | `--thinking-budget <num>` - Token budget (0-63999)"
       verbose
         option "• `--verbose` or `-v` - Verbose output | `--attach-logs` - Attach logs to PR"
       show

package/src/locales/hi.lino

@@ -578,7 +578,7 @@ hi
         branch
           option "• `--base-branch <branch>` या `-b` - PR के लिए target branch (default: repo default branch)"
       think
-        option "• `--think <level>` - thinking level (off/low/medium/high/xhigh/max) | `--thinking-budget <num>` - token budget (0-63999)"
+        option "• `--think <level>` - thinking level (off/low/medium/high/xhigh/ultra/max) | `--thinking-budget <num>` - token budget (0-63999)"
       verbose
         option "• `--verbose` या `-v` - verbose output | `--attach-logs` - logs को PR से attach करें"
       show

package/src/locales/ru.lino

@@ -578,7 +578,7 @@ ru
         branch
           option "• `--base-branch <branch>` или `-b` - Целевая ветка для PR (по умолчанию ветка репозитория)"
       think
-        option "• `--think <level>` - уровень размышления (off/low/medium/high/xhigh/max) | `--thinking-budget <num>` - бюджет токенов (0-63999)"
+        option "• `--think <level>` - уровень размышления (off/low/medium/high/xhigh/ultra/max) | `--thinking-budget <num>` - бюджет токенов (0-63999)"
       verbose
         option "• `--verbose` или `-v` - подробный вывод | `--attach-logs` - прикрепить логи к PR"
       show

package/src/locales/zh.lino

@@ -578,7 +578,7 @@ zh
         branch
           option "• `--base-branch <branch>` 或 `-b` - PR 目标分支（默认：仓库默认分支）"
       think
-        option "• `--think <level>` - 思考级别（off/low/medium/high/xhigh/max）| `--thinking-budget <num>` - token 预算（0-63999）"
+        option "• `--think <level>` - 思考级别（off/low/medium/high/xhigh/ultra/max）| `--thinking-budget <num>` - token 预算（0-63999）"
       verbose
         option "• `--verbose` 或 `-v` - 详细输出 | `--attach-logs` - 将日志附加到 PR"
       show

package/src/models/index.mjs

@@ -192,7 +192,7 @@ export const defaultModels = {
   claude: 'sonnet',
   agent: 'nemotron-3-super-free', // Issue #1563: changed from qwen3.6-plus-free (free promotion ended) per agent PR #243
   opencode: 'grok-code-fast-1',
-  codex: 'gpt-5.5',
+  codex: 'gpt-5.6-sol', // Issue #2027: GPT-5.6 Sol is the released Codex flagship; runtime falls back to gpt-5.5 when Sol is not in the local catalog
   qwen: 'qwen3-coder-plus',
   gemini: 'flash',
 };
@@ -366,7 +366,10 @@ export const getDefaultModelForTool = tool => {
 };
 
 let cachedInstalledCodexModelsPromise = null;
-const CODEX_DEFAULT_FALLBACK_CHAIN = ['openai.gpt-5.5', 'gpt-5.6-sol', 'gpt-5.6-terra', 'gpt-5.6-luna', 'openai.gpt-5.6-sol', 'openai.gpt-5.6-terra', 'openai.gpt-5.6-luna', 'gpt-5.4', 'openai.gpt-5.4', 'gpt-5.5-mini', 'gpt-5.4-mini', 'gpt-5.3-codex', 'gpt-5.3-codex-spark', 'gpt-5.2', 'gpt-5.2-codex', 'gpt-5.5-nano', 'gpt-5.4-nano'];
+// Issue #2027: With gpt-5.6-sol as the preferred default, the fallback chain is only
+// consulted when Sol is absent from the local catalog. Prefer the previous stable
+// default (gpt-5.5) first, then the remaining GPT-5.6 preview tiers, then older models.
+const CODEX_DEFAULT_FALLBACK_CHAIN = ['gpt-5.5', 'openai.gpt-5.5', 'gpt-5.6-terra', 'gpt-5.6-luna', 'openai.gpt-5.6-sol', 'openai.gpt-5.6-terra', 'openai.gpt-5.6-luna', 'gpt-5.4', 'openai.gpt-5.4', 'gpt-5.5-mini', 'gpt-5.4-mini', 'gpt-5.3-codex', 'gpt-5.3-codex-spark', 'gpt-5.2', 'gpt-5.2-codex', 'gpt-5.5-nano', 'gpt-5.4-nano'];
 
 export const getInstalledCodexModels = async () => {
   if (!cachedInstalledCodexModelsPromise) {
@@ -490,7 +493,7 @@ export const getValidModelsForTool = tool => {
 export const primaryModelNames = {
   claude: ['opus', 'sonnet', 'haiku', 'opusplan', 'fable'],
   opencode: ['grok', 'gpt4o'],
-  codex: ['gpt-5.5', 'gpt-5.6-sol', 'gpt-5.6-terra', 'gpt-5.6-luna', 'gpt-5.4', 'gpt-5.4-mini', 'gpt-5.3-codex-spark'],
+  codex: ['gpt-5.6-sol', 'gpt-5.5', 'gpt-5.6-terra', 'gpt-5.6-luna', 'gpt-5.4', 'gpt-5.4-mini', 'gpt-5.3-codex-spark'],
   agent: ['nemotron-3-super-free', 'minimax-m2.5-free', 'big-pickle', 'gpt-5-nano', 'glm-5-free', 'deepseek-r1-free'],
   qwen: ['qwen3-coder-plus', 'qwen3-coder', 'qwen3-coder-flash'],
   gemini: ['flash', 'pro', 'flash-lite', 'auto'],

package/src/solve.config.lib.mjs

@@ -299,13 +299,18 @@ export const SOLVE_OPTION_DEFINITIONS = {
   },
   think: {
     type: 'string',
-    description: 'Thinking level hint. For Claude, translated to --thinking-budget for Claude Code >= 2.1.12 (off=0, low=~8000, medium=~16000, high=~24000, xhigh/max=31999) and to CLAUDE_CODE_EFFORT_LEVEL when supported. Fable 5/Mythos 5/Sonnet 5/Opus 4.8/4.7 support xhigh and max; Opus 4.6/Sonnet 4.6/Mythos Preview support max; Opus 4.5 uses high for xhigh/max. For Codex, mapped to reasoning effort (off=none, low=low, medium=medium, high=high, xhigh/max=xhigh).',
-    choices: ['off', 'low', 'medium', 'high', 'xhigh', 'max'],
+    description: 'Thinking level hint. For Claude, translated to --thinking-budget for Claude Code >= 2.1.12 (off=0, low=~8000, medium=~16000, high=~24000, xhigh/ultra/max=31999) and to CLAUDE_CODE_EFFORT_LEVEL when supported. Fable 5/Mythos 5/Sonnet 5/Opus 4.8/4.7 support xhigh and max; Opus 4.6/Sonnet 4.6/Mythos Preview support max; Opus 4.5 uses high for xhigh/max. `ultra` maps to the highest supported Claude effort (Claude "ultracode"-class reasoning). For Codex (GPT-5.6 Sol), mapped 1:1 to reasoning effort (off=none, low=low, medium=medium, high=high, xhigh=xhigh, ultra=ultra, max=max); GPT-5.6 keeps xhigh and adds max above it, and ultra runs the multi-agent mode paired with a rollout token budget cap. Default: no thinking level is enforced (models run as-is).',
+    choices: ['off', 'low', 'medium', 'high', 'xhigh', 'ultra', 'max'],
     default: undefined,
   },
   'thinking-budget': {
     type: 'number',
-    description: 'Thinking token budget. For Claude Code, controls MAX_THINKING_TOKENS (0-31999 by default). For Codex, enables finer reasoning-effort mapping including minimal/low/medium/high/xhigh.',
+    description: 'Thinking token budget. For Claude Code, controls MAX_THINKING_TOKENS (0-31999 by default). For Codex, enables finer reasoning-effort mapping (minimal/low/medium/high) capped at xhigh; the deepest single-agent `max` and the multi-agent `ultra` effort must be requested explicitly via --think max / --think ultra.',
+    default: undefined,
+  },
+  'rollout-token-budget': {
+    type: 'number',
+    description: "Codex rollout token budget (turn-level cap) paired with the multi-agent `--think ultra` effort so GPT-5.6 Sol's ultra mode stays predictable and does not run away on cost. Default: 500000. Only applied when --think ultra selects the ultra reasoning effort.",
     default: undefined,
   },
   'thinking-budget-claude-minimum-version': {

package/src/thinking-prompt.lib.mjs

@@ -8,6 +8,7 @@ export const THINK_PROMPT_MESSAGES = Object.freeze({
   medium: 'Think hard.',
   high: 'Think harder.',
   xhigh: 'Ultrathink.',
+  ultra: 'Ultrathink.',
   max: 'Ultrathink.',
 });
 

package/src/tool-retry.lib.mjs

@@ -59,6 +59,13 @@ export const classifyRetryableError = value => {
     return { message, isRetryable: true, isCapacity: false, label: 'Stream disconnected before completion' };
   }
 
+  // Issue #2023: Claude Code can close a stream-json process after tool output
+  // (for example "Exit code 144") without emitting the terminal result event.
+  // The session is still resumable; failing immediately loses in-progress work.
+  if (lower.includes('claude stream ended without a terminal result event')) {
+    return { message, isRetryable: true, isCapacity: false, label: 'Claude stream ended without terminal result' };
+  }
+
   // Issue #1937: Stream idle timeout. When the Anthropic streaming response stalls
   // (no bytes for the SDK's idle window) after the model has already emitted part of
   // its answer, the Claude CLI aborts the turn and surfaces a synthetic assistant /