@in-the-loop-labs/pair-review 3.0.5 → 3.1.0

package/src/ai/prompts/baseline/level3/fast.js

@@ -96,6 +96,14 @@ Do NOT modify files.
 {{sparseCheckoutGuidance}}
 </section>
 
+<section name="severity-classification" required="true" tier="fast">
+### Severity
+Assign severity to each suggestion (omit for praise):
+- **critical**: Production incidents, crashes, security vulnerabilities, data corruption/loss, race conditions, breaking changes, test failures
+- **medium**: Degraded functionality, missing error handling, missing validation, missing/poor test coverage
+- **minor**: Code quality, docs gaps, minor optimizations
+</section>
+
 <section name="output-schema" locked="true">
 ## Output Format
 
@@ -109,6 +117,7 @@ Output JSON with this structure:
     "line": 42,
     "old_or_new": "NEW",
     "type": "bug|improvement|praise|suggestion|design|performance|security|code-style",
+    "severity": "critical|medium|minor (omit for praise)",
     "title": "Brief title",
     "description": "Detailed explanation mentioning why codebase context was needed",
     "suggestion": "How to fix/improve based on codebase context (omit for praise items)",
@@ -117,6 +126,7 @@ Output JSON with this structure:
   "fileLevelSuggestions": [{
     "file": "path/to/file",
     "type": "bug|improvement|praise|suggestion|design|performance|security|code-style",
+    "severity": "critical|medium|minor (omit for praise)",
     "title": "Brief title describing file-level concern",
     "description": "Explanation of the file-level observation from codebase perspective",
     "suggestion": "How to address the file-level concern (omit for praise items)",
@@ -169,6 +179,7 @@ const sections = [
   { name: 'focus-areas', required: true, tier: ['fast'] },
   { name: 'available-commands', required: true, tier: ['fast'] },
   { name: 'sparse-checkout', optional: true, tier: ['fast', 'balanced', 'thorough'] },
+  { name: 'severity-classification', required: true, tier: ['fast'] },
   { name: 'output-schema', locked: true },
   { name: 'diff-instructions', required: true, tier: ['fast'] },
   { name: 'guidelines', required: true, tier: ['fast'] }
@@ -191,6 +202,7 @@ const defaultOrder = [
   'focus-areas',
   'available-commands',
   'sparse-checkout',
+  'severity-classification',
   'output-schema',
   'diff-instructions',
   'guidelines'

package/src/ai/prompts/baseline/level3/thorough.js

@@ -238,6 +238,15 @@ Note: You may optionally use parallel read-only Tasks to explore different areas
 {{sparseCheckoutGuidance}}
 </section>
 
+<section name="severity-classification" required="true">
+### Severity Classification
+Assign a severity to each suggestion (except praise):
+- **critical**: Production incidents, system failures, or security vulnerabilities — runtime crashes, data corruption or loss, race conditions, deadlocks, breaking changes, changes that will cause existing tests to fail
+- **medium**: Degraded functionality or reliability — missing error handling, N+1 queries, missing validation, missing or poor test coverage for new functionality
+- **minor**: Code quality concerns — documentation gaps, minor optimizations, style inconsistencies
+Omit severity for praise items.
+</section>
+
 <section name="output-schema" locked="true">
 ## Output Format
 
@@ -251,6 +260,7 @@ Output JSON with this structure:
     "line": 42,
     "old_or_new": "NEW",
     "type": "bug|improvement|praise|suggestion|design|performance|security|code-style",
+    "severity": "critical|medium|minor (omit for praise)",
     "title": "Brief title",
     "description": "Detailed explanation mentioning why codebase context was needed",
     "suggestion": "How to fix/improve based on codebase context (omit for praise items)",
@@ -260,6 +270,7 @@ Output JSON with this structure:
   "fileLevelSuggestions": [{
     "file": "path/to/file",
     "type": "bug|improvement|praise|suggestion|design|performance|security|code-style",
+    "severity": "critical|medium|minor (omit for praise)",
     "title": "Brief title describing file-level concern",
     "description": "Explanation of the file-level observation from codebase perspective",
     "suggestion": "How to address the file-level concern (omit for praise items)",
@@ -403,6 +414,7 @@ const sections = [
   { name: 'focus-areas', required: true, tier: ['thorough'] },
   { name: 'available-commands', required: true, tier: ['thorough'] },
   { name: 'sparse-checkout', optional: true, tier: ['fast', 'balanced', 'thorough'] },
+  { name: 'severity-classification', required: true },
   { name: 'output-schema', locked: true },
   { name: 'diff-instructions', required: true, tier: ['thorough'] },
   { name: 'confidence-guidance', required: true, tier: ['thorough'] },
@@ -429,6 +441,7 @@ const defaultOrder = [
   'focus-areas',
   'available-commands',
   'sparse-checkout',
+  'severity-classification',
   'output-schema',
   'diff-instructions',
   'confidence-guidance',

package/src/ai/prompts/baseline/orchestration/balanced.js

@@ -25,6 +25,7 @@ const { ORCHESTRATION_INPUT_SCHEMA_DOCS } = require('../../shared/output-schema'
  * - {{reviewIntro}} - Review introduction line
  * - {{prContext}} - PR context section (optional, may be empty)
  * - {{customInstructions}} - Custom instructions section (optional)
+ * - {{dedupInstructions}} - Dedup instructions section (optional)
  * - {{lineNumberGuidance}} - Line number guidance section
  * - {{level1Count}} - Number of Level 1 suggestions
  * - {{level2Count}} - Number of Level 2 suggestions
@@ -58,6 +59,10 @@ You are helping a human reviewer by intelligently curating and merging suggestio
 {{customInstructions}}
 </section>
 
+<section name="dedup-instructions" optional="true">
+{{dedupInstructions}}
+</section>
+
 <section name="input-suggestions" locked="true">
 ## Input: Multi-Level Analysis Results
 
@@ -82,6 +87,12 @@ ${ORCHESTRATION_INPUT_SCHEMA_DOCS}
 - **Preserve unique insights** that only one level discovered
 - **Prefer preserving line-level suggestions** over file-level suggestions when curating
 - **Do NOT mention which level found the issue** - focus on the insight itself
+- **Assess severity** based on the evidence and reasoning across input levels. When levels assign different severities, evaluate the supporting evidence rather than defaulting to the highest. When truly uncertain, preserve the highest severity. Omit severity for praise items.
+
+**Severity Definitions:**
+- **critical**: Production incidents, system failures, or security vulnerabilities — runtime crashes, data corruption or loss, race conditions, deadlocks, breaking changes, changes that will cause existing tests to fail
+- **medium**: Degraded functionality or reliability — missing error handling, N+1 queries, missing validation, missing or poor test coverage for new functionality
+- **minor**: Code quality concerns — documentation gaps, minor optimizations, style inconsistencies
 </section>
 
 <section name="priority-curation" required="true">
@@ -123,6 +134,7 @@ Output JSON with this structure:
     "line": 42,
     "old_or_new": "NEW",
     "type": "bug|improvement|praise|suggestion|design|performance|security|code-style",
+    "severity": "critical|medium|minor (omit for praise)",
     "title": "Brief title describing the curated insight",
     "description": "Clear explanation of the issue and why this guidance matters to the human reviewer",
     "suggestion": "Specific, actionable guidance for the reviewer. For praise items this can be omitted. For other types always include specific, actionable suggestions.",
@@ -132,6 +144,7 @@ Output JSON with this structure:
   "fileLevelSuggestions": [{
     "file": "path/to/file",
     "type": "bug|improvement|praise|suggestion|design|performance|security|code-style",
+    "severity": "critical|medium|minor (omit for praise)",
     "title": "Brief title describing file-level concern",
     "description": "Explanation of the file-level observation",
     "suggestion": "How to address the file-level concern (omit for praise items)",
@@ -199,6 +212,7 @@ const sections = [
   { name: 'critical-output', locked: true },
   { name: 'role-description', required: true },
   { name: 'custom-instructions', optional: true },
+  { name: 'dedup-instructions', optional: true },
   { name: 'input-suggestions', locked: true },
   { name: 'intelligent-merging', required: true },
   { name: 'priority-curation', required: true },
@@ -220,6 +234,7 @@ const defaultOrder = [
   'critical-output',
   'role-description',
   'custom-instructions',
+  'dedup-instructions',
   'input-suggestions',
   'intelligent-merging',
   'priority-curation',

package/src/ai/prompts/baseline/orchestration/fast.js

@@ -28,6 +28,7 @@ const { ORCHESTRATION_INPUT_SCHEMA_DOCS } = require('../../shared/output-schema'
  * - {{reviewIntro}} - Review introduction line
  * - {{prContext}} - PR context section (optional, may be empty)
  * - {{customInstructions}} - Custom instructions section (optional)
+ * - {{dedupInstructions}} - Dedup instructions section (optional)
  * - {{lineNumberGuidance}} - Line number guidance section
  * - {{level1Count}} - Number of Level 1 suggestions
  * - {{level2Count}} - Number of Level 2 suggestions
@@ -61,6 +62,10 @@ Curate and merge 3-level suggestions. Remove duplicates. Keep high-value items o
 {{customInstructions}}
 </section>
 
+<section name="dedup-instructions" optional="true">
+{{dedupInstructions}}
+</section>
+
 <section name="input-suggestions" locked="true">
 ## Input: Multi-Level Analysis Results
 
@@ -79,6 +84,8 @@ ${ORCHESTRATION_INPUT_SCHEMA_DOCS}
 <section name="intelligent-merging" required="true" tier="fast">
 ## Rules
 Combine related suggestions. Merge overlaps. Preserve unique insights. Never mention levels.
+Assess severity using evidence across levels; preserve highest when uncertain. Omit for praise.
+Severity: critical (crashes, security, data loss, race conditions, breaking changes, test failures), medium (degraded functionality, missing error handling/validation/test coverage), minor (code quality, docs, optimizations).
 </section>
 
 <section name="priority-curation" required="true" tier="fast">
@@ -105,6 +112,7 @@ Use "Consider...", "Worth noting..." - guidance not mandates.
     "line": 42,
     "old_or_new": "NEW",
     "type": "bug|improvement|praise|suggestion|design|performance|security|code-style",
+    "severity": "critical|medium|minor (omit for praise)",
     "title": "Brief title",
     "description": "Why it matters",
     "suggestion": "What to do (omit for praise)",
@@ -113,6 +121,7 @@ Use "Consider...", "Worth noting..." - guidance not mandates.
   "fileLevelSuggestions": [{
     "file": "path/to/file",
     "type": "...",
+    "severity": "critical|medium|minor (omit for praise)",
     "title": "Brief title",
     "description": "File-level observation",
     "suggestion": "How to fix (omit for praise)",
@@ -152,6 +161,7 @@ const sections = [
   { name: 'critical-output', locked: true },
   { name: 'role-description', required: true, tier: ['fast'] },
   { name: 'custom-instructions', optional: true, tier: ['fast', 'balanced', 'thorough'] },
+  { name: 'dedup-instructions', optional: true },
   { name: 'input-suggestions', locked: true },
   { name: 'intelligent-merging', required: true, tier: ['fast'] },
   { name: 'priority-curation', required: true, tier: ['fast'] },
@@ -173,6 +183,7 @@ const defaultOrder = [
   'critical-output',
   'role-description',
   'custom-instructions',
+  'dedup-instructions',
   'input-suggestions',
   'intelligent-merging',
   'priority-curation',

package/src/ai/prompts/baseline/orchestration/thorough.js

@@ -32,6 +32,7 @@ const { ORCHESTRATION_INPUT_SCHEMA_DOCS } = require('../../shared/output-schema'
  * - {{reviewIntro}} - Review introduction line
  * - {{prContext}} - PR context section (optional, may be empty)
  * - {{customInstructions}} - Custom instructions section (optional)
+ * - {{dedupInstructions}} - Dedup instructions section (optional)
  * - {{lineNumberGuidance}} - Line number guidance section
  * - {{level1Count}} - Number of Level 1 suggestions
  * - {{level2Count}} - Number of Level 2 suggestions
@@ -85,6 +86,10 @@ Quality matters more than speed for this orchestration level. It's better to sur
 {{customInstructions}}
 </section>
 
+<section name="dedup-instructions" optional="true">
+{{dedupInstructions}}
+</section>
+
 <section name="input-suggestions" locked="true">
 ## Input: Multi-Level Analysis Results
 
@@ -134,6 +139,12 @@ When levels disagree (e.g., Level 1 flags an issue that Level 3 says follows cod
 - Use the clearest framing, regardless of which level provided it
 - Do NOT mention which level found the issue - focus on the insight itself
 - When merging would lose important nuance, keep suggestions distinct
+- **Assess severity** based on the evidence and reasoning across input levels. When levels assign different severities, evaluate the supporting evidence rather than defaulting to the highest. When truly uncertain, preserve the highest severity. Omit severity for praise items.
+
+**Severity Definitions:**
+- **critical**: Production incidents, system failures, or security vulnerabilities — runtime crashes, data corruption or loss, race conditions, deadlocks, breaking changes, changes that will cause existing tests to fail
+- **medium**: Degraded functionality or reliability — missing error handling, N+1 queries, missing validation, missing or poor test coverage for new functionality
+- **minor**: Code quality concerns — documentation gaps, minor optimizations, style inconsistencies
 </section>
 
 <section name="priority-curation" required="true" tier="thorough">
@@ -266,6 +277,7 @@ Output JSON with this structure:
     "line": 42,
     "old_or_new": "NEW",
     "type": "bug|improvement|praise|suggestion|design|performance|security|code-style",
+    "severity": "critical|medium|minor (omit for praise)",
     "title": "Brief title describing the curated insight",
     "description": "Clear explanation of the issue and why this guidance matters to the human reviewer",
     "suggestion": "Specific, actionable guidance for the reviewer (omit for praise items)",
@@ -275,6 +287,7 @@ Output JSON with this structure:
   "fileLevelSuggestions": [{
     "file": "path/to/file",
     "type": "bug|improvement|praise|suggestion|design|performance|security|code-style",
+    "severity": "critical|medium|minor (omit for praise)",
     "title": "Brief title describing file-level concern",
     "description": "Explanation of the file-level observation",
     "suggestion": "How to address the file-level concern (omit for praise items)",
@@ -380,6 +393,7 @@ const sections = [
   { name: 'role-description', required: true, tier: ['thorough'] },
   { name: 'reasoning-encouragement', required: true, tier: ['thorough'] },
   { name: 'custom-instructions', optional: true, tier: ['balanced', 'thorough'] },
+  { name: 'dedup-instructions', optional: true },
   { name: 'input-suggestions', locked: true },
   { name: 'intelligent-merging', required: true, tier: ['thorough'] },
   { name: 'priority-curation', required: true, tier: ['thorough'] },
@@ -405,6 +419,7 @@ const defaultOrder = [
   'role-description',
   'reasoning-encouragement',
   'custom-instructions',
+  'dedup-instructions',
   'input-suggestions',
   'intelligent-merging',
   'priority-curation',

package/src/ai/prompts/render-for-skill.js

@@ -39,6 +39,9 @@ const SKILL_DEFAULTS = {
   changedFiles:
     '[Changed files list provided by the orchestrating agent]',
 
+  // Dedup instructions — empty by default (section collapses)
+  dedupInstructions: '',
+
   // Orchestration data markers
   level1Count: '[N]',
   level2Count: '[N]',

package/src/ai/prompts/shared/output-schema.js

@@ -22,6 +22,7 @@ const ORCHESTRATION_INPUT_SCHEMA_DOCS = `Each level provides suggestions as a JS
 - title: brief title
 - description: full explanation
 - suggestion: remediation advice
+- severity: "critical", "medium", or "minor" (omit for praise items)
 - confidence: 0.0-1.0 score
 - reasoning: (optional) array of strings with step-by-step reasoning
 - is_file_level: true if this is a file-level suggestion (no line numbers)`;
@@ -37,6 +38,7 @@ const LEVEL1_OUTPUT_SCHEMA = {
     line_end: 42,
     old_or_new: 'NEW',
     type: 'bug|improvement|praise|suggestion|design|performance|security|code-style',
+    severity: 'critical|medium|minor (omit for praise)',
     title: 'Brief title',
     description: 'Detailed explanation',
     suggestion: 'How to fix/improve (omit this field for praise items - no action needed)',
@@ -57,6 +59,7 @@ const LEVEL2_OUTPUT_SCHEMA = {
     line_end: 42,
     old_or_new: 'NEW',
     type: 'bug|improvement|praise|suggestion|design|performance|security|code-style',
+    severity: 'critical|medium|minor (omit for praise)',
     title: 'Brief title',
     description: 'Detailed explanation mentioning why full file context was needed',
     suggestion: 'How to fix/improve based on file context (omit for praise items)',
@@ -66,6 +69,7 @@ const LEVEL2_OUTPUT_SCHEMA = {
   fileLevelSuggestions: [{
     file: 'path/to/file',
     type: 'bug|improvement|praise|suggestion|design|performance|security|code-style',
+    severity: 'critical|medium|minor (omit for praise)',
     title: 'Brief title describing file-level concern',
     description: 'Explanation of the file-level observation (architecture, organization, naming, etc.)',
     suggestion: 'How to address the file-level concern (omit for praise items)',
@@ -86,6 +90,7 @@ const LEVEL3_OUTPUT_SCHEMA = {
     line_end: 42,
     old_or_new: 'NEW',
     type: 'bug|improvement|praise|suggestion|design|performance|security|code-style',
+    severity: 'critical|medium|minor (omit for praise)',
     title: 'Brief title',
     description: 'Detailed explanation mentioning why codebase context was needed',
     suggestion: 'How to fix/improve based on codebase context (omit for praise items)',
@@ -95,6 +100,7 @@ const LEVEL3_OUTPUT_SCHEMA = {
   fileLevelSuggestions: [{
     file: 'path/to/file',
     type: 'bug|improvement|praise|suggestion|design|performance|security|code-style',
+    severity: 'critical|medium|minor (omit for praise)',
     title: 'Brief title describing file-level concern',
     description: 'Explanation of the file-level observation from codebase perspective',
     suggestion: 'How to address the file-level concern (omit for praise items)',
@@ -115,6 +121,7 @@ const ORCHESTRATION_OUTPUT_SCHEMA = {
     line_end: 42,
     old_or_new: 'NEW',
     type: 'bug|improvement|praise|suggestion|design|performance|security|code-style',
+    severity: 'critical|medium|minor (omit for praise)',
     title: 'Brief title describing the curated insight',
     description: 'Clear explanation of the issue and why this guidance matters to the human reviewer',
     suggestion: 'Specific, actionable guidance for the reviewer (omit for praise items)',
@@ -124,6 +131,7 @@ const ORCHESTRATION_OUTPUT_SCHEMA = {
   fileLevelSuggestions: [{
     file: 'path/to/file',
     type: 'bug|improvement|praise|suggestion|design|performance|security|code-style',
+    severity: 'critical|medium|minor (omit for praise)',
     title: 'Brief title describing file-level concern',
     description: 'Explanation of the file-level observation',
     suggestion: 'How to address the file-level concern (omit for praise items)',

package/src/ai/provider.js

@@ -424,9 +424,43 @@ function resolveDefaultModel(models) {
 }
 
 /**
- * Apply configuration overrides for all providers
- * Call this after all providers have registered and config is loaded
- * Clears any existing overrides before applying new ones.
+ * Create an aliased provider class that reuses an existing provider's implementation
+ * but with a different ID, name, and config overrides.
+ *
+ * @param {string} aliasId - New provider ID (e.g., 'pi-reskin')
+ * @param {typeof AIProvider} BaseClass - The base provider class to alias
+ * @param {Object} aliasConfig - Config for the alias (name, models, etc.)
+ * @returns {typeof AIProvider} A subclass with overridden static metadata
+ */
+function createAliasedProviderClass(aliasId, BaseClass, aliasConfig) {
+  const processedModels = Array.isArray(aliasConfig.models) && aliasConfig.models.length > 0
+    ? aliasConfig.models.map(inferModelDefaults)
+    : null;
+
+  class AliasedProvider extends BaseClass {}
+
+  // Override static metadata so the alias has its own identity
+  AliasedProvider.getProviderName = () => aliasConfig.name || aliasId;
+  AliasedProvider.getProviderId = () => aliasId;
+  if (processedModels) {
+    AliasedProvider.getModels = () => processedModels;
+    AliasedProvider.getDefaultModel = () => resolveDefaultModel(processedModels);
+  }
+  if (aliasConfig.installInstructions) {
+    AliasedProvider.getInstallInstructions = () => aliasConfig.installInstructions;
+  }
+  if (aliasConfig.defaultTimeout != null) {
+    AliasedProvider.defaultTimeout = aliasConfig.defaultTimeout;
+  }
+
+  return AliasedProvider;
+}
+
+/**
+ * Apply configuration overrides for all providers.
+ * Called once at startup after all providers have self-registered.
+ * Does not support re-application — the provider registry is intentionally
+ * not cleaned between calls (aliased/executable classes persist).
  * @param {Object} config - Configuration object from loadConfig()
  */
 function applyConfigOverrides(config) {
@@ -446,6 +480,47 @@ function applyConfigOverrides(config) {
   for (const [providerId, providerConfig] of Object.entries(providersConfig)) {
     logger.debug(`Applying config overrides for provider: ${providerId}`);
 
+    // Executable providers: dynamically create and register a provider class
+    if (providerConfig.type === 'executable') {
+      if (!providerConfig.command) {
+        logger.warn(`Executable provider "${providerId}" missing required "command" field`);
+        continue;
+      }
+      // Lazy-require to avoid circular dependency
+      const { createExecutableProviderClass } = require('./executable-provider');
+      const ExecClass = createExecutableProviderClass(providerId, providerConfig);
+      registerProvider(providerId, ExecClass);
+      providerConfigOverrides.set(providerId, { ...providerConfig, models: ExecClass.getModels() });
+      logger.debug(`Registered executable provider: ${providerId}`);
+      continue;
+    }
+
+    // Type matching a registered provider ID creates an alias of that provider
+    if (providerConfig.type && providerConfig.type !== providerId && providerRegistry.has(providerConfig.type)) {
+      const BaseClass = providerRegistry.get(providerConfig.type);
+      const AliasClass = createAliasedProviderClass(providerId, BaseClass, providerConfig);
+      registerProvider(providerId, AliasClass);
+
+      // Aliases reuse the base provider's implementation class, not its config.
+      // Only universal override fields are forwarded; provider-specific fields
+      // (e.g. codex args) must be explicitly set in the alias config.
+      providerConfigOverrides.set(providerId, {
+        command: providerConfig.command,
+        installInstructions: providerConfig.installInstructions,
+        extra_args: providerConfig.extra_args,
+        env: providerConfig.env,
+        models: AliasClass.getModels() !== BaseClass.getModels() ? AliasClass.getModels() : null
+      });
+      logger.debug(`Registered aliased provider: ${providerId} (base: ${providerConfig.type})`);
+      continue;
+    }
+
+    // Unknown type: warn and skip (self-referential type falls through to standard override path)
+    if (providerConfig.type && providerConfig.type !== providerId) {
+      logger.warn(`Provider "${providerId}" has unknown type "${providerConfig.type}" — no matching registered provider`);
+      continue;
+    }
+
     // Process models if specified - infer defaults for each
     let processedModels = null;
     if (Array.isArray(providerConfig.models) && providerConfig.models.length > 0) {
@@ -543,12 +618,21 @@ function getAllProvidersInfo() {
     // Use overridden install instructions if available
     const installInstructions = overrides?.installInstructions || ProviderClass.getInstallInstructions();
 
+    // Build capabilities: executable providers define their own, others get defaults
+    const capabilities = ProviderClass.capabilities || {
+      review_levels: true,
+      custom_instructions: true
+    };
+
     providers.push({
       id,
       name: ProviderClass.getProviderName(),
       models,
       defaultModel,
-      installInstructions
+      installInstructions,
+      capabilities,
+      isExecutable: ProviderClass.isExecutable || false,
+      ...(ProviderClass.defaultTimeout != null ? { defaultTimeout: ProviderClass.defaultTimeout } : {})
     });
   }
   return providers;
@@ -656,6 +740,7 @@ module.exports = {
   testProviderAvailability,
   // Config override support
   applyConfigOverrides,
+  createAliasedProviderClass,
   getProviderConfigOverrides,
   inferModelDefaults,
   resolveDefaultModel,

package/src/chat/prompt-builder.js

@@ -18,9 +18,10 @@ const logger = require('../utils/logger');
  * @param {Object} options.review - Review metadata {id, pr_number, repository, review_type, local_path, name}
  * @param {Object} [options.prData] - PR data with base_sha/head_sha (for PR reviews)
  * @param {string} [options.chatInstructions] - Custom instructions from repo settings to append to system prompt
+ * @param {string} [options.commentFormatTemplate] - Resolved comment format template for consistent comment formatting
  * @returns {string} System prompt for the chat agent
  */
-function buildChatPrompt({ review, prData, chatInstructions }) {
+function buildChatPrompt({ review, prData, chatInstructions, commentFormatTemplate }) {
   const sections = [];
 
   // Role
@@ -53,6 +54,21 @@ function buildChatPrompt({ review, prData, chatInstructions }) {
   }
   sections.push(domainLines.join('\n'));
 
+  // Comment format template — injected when the user has a configured format
+  if (commentFormatTemplate) {
+    sections.push(
+      '## Comment format\n\n' +
+      'When creating or editing review comments, use this template:\n' +
+      '<template>\n' +
+      commentFormatTemplate + '\n' +
+      '</template>\n\n' +
+      'Template placeholders: {emoji}, {category}, {title}, {severity}, {SEVERITY}, {description}, {suggestion}.\n' +
+      '{severity} renders as Title Case (e.g. "Critical"), {SEVERITY} renders as UPPERCASE (e.g. "CRITICAL").\n' +
+      'Conditional sections: {?field}...{/field} — include content only when the field has a value.\n' +
+      'Always follow this format for consistency with the reviewer\'s preferences.'
+    );
+  }
+
   // API Access — cheat-sheet is injected into initial context; full docs available via GET /api.md
   sections.push(
     '## API Access\n\n' +

package/src/chat/session-manager.js

@@ -44,6 +44,12 @@ class ChatSessionManager {
    * @returns {Promise<{id: number, status: string}>}
    */
   async createSession({ provider, model, reviewId, contextCommentId, systemPrompt, cwd, initialContext }) {
+    // Resolve provider definition once — used for model fallback and bridge construction
+    const providerDef = getChatProvider(provider);
+
+    // Resolve model: explicit request value > provider config default
+    const resolvedModel = model || providerDef?.model || null;
+
     // Insert session record into DB
     const stmt = this._db.prepare(`
       INSERT INTO chat_sessions (review_id, context_comment_id, provider, model, status)
@@ -53,20 +59,20 @@ class ChatSessionManager {
       reviewId,
       contextCommentId || null,
       provider,
-      model || null
+      resolvedModel
     );
     const sessionId = Number(result.lastInsertRowid);
 
-    logger.info(`[ChatSession] Creating session ${sessionId} (provider=${provider}, review=${reviewId})`);
+    logger.info(`[ChatSession] Creating session ${sessionId} (provider=${provider}, model=${resolvedModel}, review=${reviewId})`);
 
     // Create and start the bridge
     // Chat sessions get bash for git commands; review analysis uses the safe default
     const bridge = this._createBridge(provider, {
       provider,
-      model,
+      model: resolvedModel,
       cwd,
       systemPrompt,
-    });
+    }, providerDef);
 
     const listeners = {
       delta: new Set(),
@@ -530,39 +536,38 @@ class ChatSessionManager {
    * ACP providers get an AcpBridge; everything else gets a PiBridge with tools/skills.
    * @param {string} provider
    * @param {Object} options - Bridge constructor options
+   * @param {Object} [providerDef] - Pre-resolved provider definition (avoids redundant getChatProvider calls)
    * @returns {PiBridge|AcpBridge}
    */
-  _createBridge(provider, options) {
+  _createBridge(provider, options, providerDef) {
+    const def = providerDef || getChatProvider(provider);
     if (isAcpProvider(provider)) {
-      const providerDef = getChatProvider(provider);
       return new AcpBridge({
         ...options,
-        model: options.model || providerDef?.model,
-        acpCommand: providerDef?.command,
-        acpArgs: providerDef?.args,
-        env: providerDef?.env,
-        useShell: providerDef?.useShell,
+        model: options.model || def?.model,
+        acpCommand: def?.command,
+        acpArgs: def?.args,
+        env: def?.env,
+        useShell: def?.useShell,
       });
     }
     if (isClaudeCodeProvider(provider)) {
-      const providerDef = getChatProvider(provider);
       return new ClaudeCodeBridge({
         ...options,
-        model: options.model || providerDef?.model,
-        claudeCommand: providerDef?.command,
-        env: providerDef?.env,
-        useShell: providerDef?.useShell,
+        model: options.model || def?.model,
+        claudeCommand: def?.command,
+        env: def?.env,
+        useShell: def?.useShell,
       });
     }
     if (isCodexProvider(provider)) {
-      const providerDef = getChatProvider(provider);
       return new CodexBridge({
         ...options,
-        model: options.model || providerDef?.model,
-        codexCommand: providerDef?.command,
-        codexArgs: providerDef?.args,
-        env: providerDef?.env,
-        useShell: providerDef?.useShell,
+        model: options.model || def?.model,
+        codexCommand: def?.command,
+        codexArgs: def?.args,
+        env: def?.env,
+        useShell: def?.useShell,
       });
     }
     // Pi provider — resolve config overrides (command, model, env) from provider def.
@@ -570,14 +575,13 @@ class ChatSessionManager {
     // which would forward it as `--provider pi` to the Pi CLI.  The CLI's --provider flag
     // expects a model provider ("google", "anthropic", etc.) and should only come from
     // explicit user configuration (providerDef.provider).
-    const providerDef = getChatProvider(provider);
     return new PiBridge({
       ...options,
-      provider: providerDef?.provider || null,
-      model: options.model || providerDef?.model,
-      piCommand: providerDef?.command,
-      env: providerDef?.env,
-      useShell: providerDef?.useShell,
+      provider: def?.provider || null,
+      model: options.model || def?.model,
+      piCommand: def?.command,
+      env: def?.env,
+      useShell: def?.useShell,
       tools: CHAT_TOOLS,
       extensions: [taskExtensionDir],
     });