@in-the-loop-labs/pair-review 3.5.0 → 3.5.2

package/src/ai/claude-provider.js

@@ -17,19 +17,32 @@ const { StreamParser, parseClaudeLine } = require('./stream-parser');
 const BIN_DIR = path.join(__dirname, '..', '..', 'bin');
 
 /**
- * Claude model definitions with tier mappings
+ * Claude model definitions with tier mappings.
+ *
+ * Effort is set via the CLAUDE_CODE_EFFORT_LEVEL env var (highest-precedence way
+ * to control reasoning effort; takes precedence over the --effort CLI flag and is
+ * not deprecated). Extended thinking is forced on globally via `--thinking enabled`
+ * in the constructor's base args; individual models can override this via extra_args
+ * (e.g., Haiku uses adaptive thinking for efficiency).
+ *
+ * Effort support by model (newest CLIs): Opus 4.8 / 4.7 support low|medium|high|
+ * xhigh|max; Opus 4.6 & Sonnet 4.6 support low|medium|high|max (no xhigh); Haiku
+ * has no effort levels.
  */
 const CLAUDE_MODELS = [
+  // ── Thorough tier ───────────────────────────────────────────────────────
   {
-    id: 'opus-4.7-xhigh',
+    id: 'opus',
+    aliases: ['opus-4.7-xhigh'],
     cli_model: 'claude-opus-4-7',
     env: { CLAUDE_CODE_EFFORT_LEVEL: 'xhigh' },
     name: 'Opus 4.7 XHigh',
     tier: 'thorough',
-    tagline: 'Latest Gen',
-    description: 'Opus 4.7 (latest) with extra-high effort',
-    badge: 'Latest',
-    badgeClass: 'badge-power'
+    tagline: 'Maximum Depth',
+    description: 'Opus 4.7 with extra-high effort — deepest analysis',
+    badge: 'Most Thorough',
+    badgeClass: 'badge-power',
+    default: true
   },
   {
     id: 'opus-4.7-high',
@@ -37,33 +50,46 @@ const CLAUDE_MODELS = [
     env: { CLAUDE_CODE_EFFORT_LEVEL: 'high' },
     name: 'Opus 4.7 High',
     tier: 'thorough',
-    tagline: 'Latest Gen',
-    description: 'Opus 4.7 (latest) with high effort',
+    tagline: 'High Effort',
+    description: 'Opus 4.7 with high effort — thorough, quicker than XHigh',
+    badge: 'Thorough',
+    badgeClass: 'badge-power'
+  },
+  {
+    id: 'opus-4.8-xhigh',
+    cli_model: 'claude-opus-4-8',
+    env: { CLAUDE_CODE_EFFORT_LEVEL: 'xhigh' },
+    name: 'Opus 4.8 XHigh',
+    tier: 'thorough',
+    tagline: 'Newest',
+    description: 'Opus 4.8 (newest) with extra-high effort',
     badge: 'Latest',
     badgeClass: 'badge-power'
   },
   {
-    id: 'opus',
-    aliases: ['opus-4.6-high'],
-    cli_model: 'claude-opus-4-6',
+    id: 'opus-4.8-high',
+    cli_model: 'claude-opus-4-8',
     env: { CLAUDE_CODE_EFFORT_LEVEL: 'high' },
-    name: 'Opus 4.6 High',
+    name: 'Opus 4.8 High',
     tier: 'thorough',
-    tagline: 'Maximum Depth',
-    description: 'Opus 4.6 with high effort — deepest analysis',
-    badge: 'Most Thorough',
-    badgeClass: 'badge-power',
-    default: true
+    tagline: 'Newest',
+    description: 'Opus 4.8 (newest) with high effort',
+    badge: 'Latest',
+    badgeClass: 'badge-power'
   },
   {
-    id: 'haiku',
-    name: 'Haiku 4.6',
-    tier: 'fast',
-    tagline: 'Lightning Fast',
-    description: 'Quick analysis for simple changes',
-    badge: 'Fastest',
-    badgeClass: 'badge-speed'
+    id: 'opus-4.6-high',
+    aliases: ['opus-4.6-low', 'opus-4.6-medium', 'opus-4.5'],
+    cli_model: 'claude-opus-4-6',
+    env: { CLAUDE_CODE_EFFORT_LEVEL: 'high' },
+    name: 'Opus 4.6 High',
+    tier: 'thorough',
+    tagline: 'Previous Gen',
+    description: 'Opus 4.6 with high effort',
+    badge: 'Previous Gen',
+    badgeClass: 'badge-power'
   },
+  // ── Balanced tier ───────────────────────────────────────────────────────
   {
     id: 'sonnet-4.6',
     cli_model: 'claude-sonnet-4-6',
@@ -74,28 +100,6 @@ const CLAUDE_MODELS = [
     badge: 'Standard',
     badgeClass: 'badge-recommended'
   },
-  {
-    id: 'opus-4.6-low',
-    cli_model: 'claude-opus-4-6',
-    env: { CLAUDE_CODE_EFFORT_LEVEL: 'low' },
-    name: 'Opus 4.6 Low',
-    tier: 'balanced',
-    tagline: 'Fast Opus',
-    description: 'Opus 4.6 with low effort — quick and capable',
-    badge: 'Balanced',
-    badgeClass: 'badge-recommended'
-  },
-  {
-    id: 'opus-4.6-medium',
-    cli_model: 'claude-opus-4-6',
-    env: { CLAUDE_CODE_EFFORT_LEVEL: 'medium' },
-    name: 'Opus 4.6 Medium',
-    tier: 'balanced',
-    tagline: 'Balanced Opus',
-    description: 'Opus 4.6 with medium effort — balanced depth',
-    badge: 'Thorough',
-    badgeClass: 'badge-power'
-  },
   {
     id: 'opus-4.6-1m',
     cli_model: 'claude-opus-4-6[1m]',
@@ -106,15 +110,17 @@ const CLAUDE_MODELS = [
     badge: 'More Context',
     badgeClass: 'badge-power'
   },
+  // ── Fast tier ───────────────────────────────────────────────────────────
   {
-    id: 'opus-4.5',
-    cli_model: 'claude-opus-4-5-20251101',
-    name: 'Opus 4.5',
-    tier: 'thorough',
-    tagline: 'Deep Thinker',
-    description: 'Extended thinking for complex analysis',
-    badge: 'Previous Gen',
-    badgeClass: 'badge-power'
+    id: 'haiku',
+    cli_model: 'claude-haiku-4-5-20251001',
+    name: 'Haiku 4.5',
+    tier: 'fast',
+    tagline: 'Lightning Fast',
+    description: 'Quick analysis for simple changes',
+    badge: 'Fastest',
+    badgeClass: 'badge-speed',
+    extra_args: ['--thinking', 'adaptive']
   }
 ];
 
@@ -196,7 +202,12 @@ class ClaudeProvider extends AIProvider {
     // user's configured environment. To disable skills, add --disable-slash-commands
     // to extra_args in provider/model config.
     const hooksArgs = ['--settings', '{"disableAllHooks":true}'];
-    const baseArgs = ['-p', '--verbose', ...cliModelArgs, '--output-format', 'stream-json', ...hooksArgs, ...permissionArgs];
+    // Force extended thinking on for every analysis call. The Claude CLI's
+    // `--thinking` flag accepts enabled|adaptive|disabled; we always want
+    // reasoning engaged for code review. User config extra_args appended later
+    // win over this (commander uses the last occurrence) if an override is set.
+    const thinkingArgs = ['--thinking', 'enabled'];
+    const baseArgs = ['-p', '--verbose', ...cliModelArgs, '--output-format', 'stream-json', ...thinkingArgs, ...hooksArgs, ...permissionArgs];
     if (maxBudget) {
       const budgetNum = parseFloat(maxBudget);
       if (isNaN(budgetNum) || budgetNum <= 0) {
@@ -242,7 +253,8 @@ class ClaudeProvider extends AIProvider {
     // - string: use this exact value for --model
     // - null: explicitly suppress --model (for tools that want the model set via env instead)
     const builtIn = CLAUDE_MODELS.find(m => m.id === modelId || (m.aliases && m.aliases.includes(modelId)));
-    const configModel = configOverrides.models?.find(m => m.id === modelId);
+    const modelKeys = new Set([modelId, builtIn?.id, ...(builtIn?.aliases || [])].filter(Boolean));
+    const configModel = configOverrides.models?.find(m => modelKeys.has(m.id));
     const resolvedCliModel = configModel?.cli_model !== undefined
       ? configModel.cli_model
       : (builtIn?.cli_model !== undefined ? builtIn.cli_model : modelId);

package/src/ai/codex-provider.js

@@ -34,6 +34,29 @@ const BIN_DIR = path.join(__dirname, '..', '..', 'bin');
  * Deprecated (April 2026): gpt-5.1-codex-mini, gpt-5.1-codex-max, gpt-5.1-codex
  */
 const CODEX_MODELS = [
+  {
+    id: 'gpt-5.5-high',
+    cli_model: 'gpt-5.5',
+    extra_args: ['-c', 'model_reasoning_effort="high"'],
+    name: 'GPT-5.5 High',
+    tier: 'thorough',
+    tagline: 'Latest Deep',
+    description: 'Latest-generation GPT model with high reasoning effort for demanding PR reviews, strong code understanding, and careful cross-file analysis.',
+    badge: 'Recommended',
+    badgeClass: 'badge-recommended',
+    default: true
+  },
+  {
+    id: 'gpt-5.5-xhigh',
+    cli_model: 'gpt-5.5',
+    extra_args: ['-c', 'model_reasoning_effort="xhigh"'],
+    name: 'GPT-5.5 XHigh',
+    tier: 'thorough',
+    tagline: 'Frontier Depth',
+    description: 'GPT-5.5 with extra-high reasoning effort for the hardest reviews: architecture, concurrency, security-sensitive changes, and large codebase context.',
+    badge: 'Max Reasoning',
+    badgeClass: 'badge-power'
+  },
   {
     id: 'gpt-5.4-high',
     // Alias keeps results/councils saved under the previous bare `gpt-5.4`
@@ -45,9 +68,8 @@ const CODEX_MODELS = [
     tier: 'thorough',
     tagline: 'Deep Review',
     description: 'GPT-5.4 with high reasoning effort for complex multi-file reviews, architectural consistency, and subtle behavioral regressions.',
-    badge: 'Recommended',
-    badgeClass: 'badge-recommended',
-    default: true
+    badge: 'Previous Gen',
+    badgeClass: 'badge-power'
   },
   {
     id: 'gpt-5.4-xhigh',
@@ -60,28 +82,6 @@ const CODEX_MODELS = [
     badge: 'Extra High',
     badgeClass: 'badge-power'
   },
-  {
-    id: 'gpt-5.5-high',
-    cli_model: 'gpt-5.5',
-    extra_args: ['-c', 'model_reasoning_effort="high"'],
-    name: 'GPT-5.5 High',
-    tier: 'thorough',
-    tagline: 'Latest Deep',
-    description: 'Latest-generation GPT model with high reasoning effort for demanding PR reviews, strong code understanding, and careful cross-file analysis.',
-    badge: 'High Effort',
-    badgeClass: 'badge-power'
-  },
-  {
-    id: 'gpt-5.5-xhigh',
-    cli_model: 'gpt-5.5',
-    extra_args: ['-c', 'model_reasoning_effort="xhigh"'],
-    name: 'GPT-5.5 XHigh',
-    tier: 'thorough',
-    tagline: 'Frontier Depth',
-    description: 'GPT-5.5 with extra-high reasoning effort for the hardest reviews: architecture, concurrency, security-sensitive changes, and large codebase context.',
-    badge: 'Max Reasoning',
-    badgeClass: 'badge-power'
-  },
   {
     id: 'gpt-5.3-codex',
     name: 'GPT-5.3 Codex',
@@ -121,7 +121,7 @@ class CodexProvider extends AIProvider {
    * @param {Object} configOverrides.env - Additional environment variables
    * @param {Object[]} configOverrides.models - Custom model definitions
    */
-  constructor(model = 'gpt-5.4-high', configOverrides = {}) {
+  constructor(model = 'gpt-5.5-high', configOverrides = {}) {
     super(model);
 
     // Command precedence: ENV > config > default
@@ -149,9 +149,9 @@ class CodexProvider extends AIProvider {
     // 2. "read-only" prevents ALL shell commands including git-diff-lines
     // 3. The AI is instructed to only analyze code, not modify it
     //
-    // --full-auto: Non-interactive mode that auto-approves within sandbox bounds.
-    // Combined with workspace-write sandbox, this limits damage to the worktree only.
-    // Note: The -a flag is for interactive mode only; exec subcommand uses --full-auto.
+    // Newer Codex CLI versions deprecate --full-auto; `codex exec` is already
+    // non-interactive, and `--sandbox workspace-write` selects the required
+    // sandbox policy.
     //
     // Shell environment config:
     // - allow_login_shell=false: Prevents zsh from using -l flag, which would
@@ -164,7 +164,7 @@ class CodexProvider extends AIProvider {
     // (--dangerously-bypass-approvals-and-sandbox is the Codex CLI equivalent of Claude's --dangerously-skip-permissions)
     const sandboxArgs = configOverrides.yolo
       ? ['--dangerously-bypass-approvals-and-sandbox']
-      : ['--sandbox', 'workspace-write', '--full-auto'];
+      : ['--sandbox', 'workspace-write'];
     // Shell env args prevent login shell from reconstructing PATH (orthogonal to
     // sandbox permissions). Overridable via configOverrides.args following the
     // same two-tier pattern as chat-providers.js: args replaces, extra_args appends.
@@ -352,7 +352,7 @@ class CodexProvider extends AIProvider {
 
         if (code !== 0) {
           logger.error(`${levelPrefix} Codex CLI exited with code ${code}`);
-          settle(reject, new Error(`${levelPrefix} Codex CLI exited with code ${code}: ${stderr}`));
+          settle(reject, this.createExitError(code, stderr, levelPrefix));
           return;
         }
 
@@ -433,6 +433,37 @@ class CodexProvider extends AIProvider {
     });
   }
 
+  /**
+   * Build an actionable error for Codex CLI process failures.
+   *
+   * @param {number} code - Process exit code
+   * @param {string} stderr - Captured stderr
+   * @param {string} levelPrefix - Logging prefix
+   * @returns {Error}
+   */
+  createExitError(code, stderr, levelPrefix) {
+    const stderrText = stderr.trim();
+
+    if (this.isAuthError(stderrText)) {
+      return new Error(
+        `${levelPrefix} Codex CLI authentication failed. Check Codex CLI authentication and try again. ` +
+        `Original stderr: ${stderrText}`
+      );
+    }
+
+    return new Error(`${levelPrefix} Codex CLI exited with code ${code}: ${stderr}`);
+  }
+
+  /**
+   * Detect authentication failures reported by the Codex CLI.
+   *
+   * @param {string} stderr - Captured stderr
+   * @returns {boolean}
+   */
+  isAuthError(stderr) {
+    return /(?:401\s+Unauthorized|HTTP error:\s*401|Unauthorized)/i.test(stderr);
+  }
+
   /**
    * Parse Codex CLI JSONL response
    * Codex outputs JSONL with multiple event types:
@@ -664,7 +695,7 @@ class CodexProvider extends AIProvider {
 
     // Base args for extraction (read-only sandbox, no shell access needed)
     // Note: '-' (stdin marker) must come LAST, after any extra_args
-    const baseArgs = ['exec', '-m', cliModel, '--json', '--sandbox', 'read-only', '--full-auto'];
+    const baseArgs = ['exec', '-m', cliModel, '--json', '--sandbox', 'read-only'];
 
     // Append stdin marker '-' at the end after all other args
     return [...baseArgs, ...extraArgs, '-'];
@@ -790,7 +821,7 @@ class CodexProvider extends AIProvider {
   }
 
   static getDefaultModel() {
-    return 'gpt-5.4-high';
+    return 'gpt-5.5-high';
   }
 
   static getInstallInstructions() {

package/src/chat/api-reference.js

@@ -235,7 +235,7 @@ curl -s -X POST http://localhost:{{PORT}}/api/pr/OWNER/REPO/PR_NUMBER/analyses \
   -H 'Content-Type: application/json' \\
   -d '{
     "provider": "claude",
-    "model": "claude-sonnet-4-5-20250929",
+    "model": "claude-opus-4-7",
     "tier": "balanced",
     "customInstructions": "Focus on security issues."
   }'

package/src/chat/chat-providers.js

@@ -12,6 +12,7 @@ const logger = require('../utils/logger');
 
 // Default dependencies (overridable for testing)
 const defaults = { spawn };
+const CODEX_SANDBOX_MODES = new Set(['workspace-write', 'read-only']);
 
 /**
  * Built-in chat provider definitions.
@@ -68,6 +69,7 @@ const CHAT_PROVIDERS = {
     name: 'Codex (JSON-RPC)',
     type: 'codex',
     command: 'codex',
+    sandbox: 'workspace-write',
     // Shell environment config prevents zsh -l from reconstructing PATH,
     // ensuring git-diff-lines and other bin/ scripts remain findable.
     args: [
@@ -126,6 +128,9 @@ function getChatProvider(id) {
     }
     if (overrides.load_skills !== undefined) provider.load_skills = overrides.load_skills;
     if (overrides.app_extensions !== undefined) provider.app_extensions = overrides.app_extensions;
+    if (provider.type === 'codex' && overrides.sandbox !== undefined) {
+      provider.sandbox = normalizeCodexSandbox(overrides.sandbox, id);
+    }
     if (provider.command.includes(' ')) {
       provider.useShell = true;
     }
@@ -152,6 +157,9 @@ function getChatProvider(id) {
   }
   if (overrides.load_skills !== undefined) merged.load_skills = overrides.load_skills;
   if (overrides.app_extensions !== undefined) merged.app_extensions = overrides.app_extensions;
+  if (base.type === 'codex' && overrides.sandbox !== undefined) {
+    merged.sandbox = normalizeCodexSandbox(overrides.sandbox, id);
+  }
   // For multi-word commands (e.g. "devx claude"), use shell mode
   if (merged.command && merged.command.includes(' ')) {
     merged.useShell = true;
@@ -159,6 +167,24 @@ function getChatProvider(id) {
   return merged;
 }
 
+/**
+ * Validate the small user-facing Codex sandbox config surface.
+ * @param {string} sandbox
+ * @param {string} providerId
+ * @returns {string}
+ */
+function normalizeCodexSandbox(sandbox, providerId = 'codex') {
+  if (CODEX_SANDBOX_MODES.has(sandbox)) {
+    return sandbox;
+  }
+
+  logger.warn(
+    `[ChatProviders] Invalid sandbox "${sandbox}" for ${providerId}; ` +
+    'falling back to workspace-write. Supported values: workspace-write, read-only.'
+  );
+  return 'workspace-write';
+}
+
 /**
  * Get all chat provider definitions (built-in + dynamic from config).
  * @returns {Array<Object>}