git-coco 0.22.7 → 0.22.9

package/dist/index.d.ts

@@ -337,6 +337,7 @@ interface BaseParserOptions {
     llm: ReturnType<typeof getLlm>;
     git: SimpleGit;
     logger: Logger;
+    maxTokens?: number;
 }
 interface BaseParserInput {
     options: BaseParserOptions;

package/dist/index.esm.mjs

@@ -47,7 +47,7 @@ import { pathToFileURL } from 'url';
 /**
  * Current build version from package.json
  */
-const BUILD_VERSION = "0.22.7";
+const BUILD_VERSION = "0.22.9";
 
 const isInteractive = (config) => {
     return config?.mode === 'interactive' || !!config?.interactive;
@@ -7212,6 +7212,8 @@ const CONVENTIONAL_TEMPLATE = `Generate a commit message that strictly adheres t
    - Scope should be a noun in parentheses (e.g., (parser), (ui), (config))
    - Omit scope if the change is broad or affects multiple areas
 
+CRITICAL: You must respond with ONLY valid JSON. All string values must be properly quoted.
+
 Based on the following diff summary, generate a conventional commit message that follows these rules exactly:
 
 """"""
@@ -7955,6 +7957,55 @@ async function executeChainWithSchema(schema, llm, prompt, variables, options =
     }
 }
 
+/**
+ * Utility to repair common JSON formatting issues that LLMs make
+ * Specifically handles cases where string values are not properly quoted
+ */
+function repairJson(jsonString) {
+    // Remove any markdown code block wrapping
+    let cleaned = jsonString.replace(/```(?:json)?\s*([\s\S]*?)\s*```/g, '$1').trim();
+    // Remove inline code block wrapping
+    cleaned = cleaned.replace(/^`(.*)`$/, '$1').trim();
+    // If it doesn't look like JSON, return as-is
+    if (!cleaned.startsWith('{') || !cleaned.endsWith('}')) {
+        return jsonString;
+    }
+    try {
+        // First try parsing as-is
+        JSON.parse(cleaned);
+        return cleaned;
+    }
+    catch {
+        // Try to repair common issues
+        let repaired = cleaned;
+        // Fix unquoted string values in title and body fields
+        // Pattern: "title": unquoted_value, -> "title": "unquoted_value",
+        repaired = repaired.replace(/"(title|body)":\s*([^",\{\}\[\]]+?)(?=\s*[,\}])/g, (match, field, value) => {
+            // Clean up the value (remove leading/trailing whitespace)
+            const cleanValue = value.trim();
+            // If it's already quoted or looks like a number/boolean, leave it
+            if (cleanValue.startsWith('"') || /^(true|false|\d+)$/.test(cleanValue)) {
+                return match;
+            }
+            // Quote the value
+            return `"${field}": "${cleanValue}"`;
+        });
+        // Fix missing quotes around field names (though this should be rare)
+        repaired = repaired.replace(/([{,]\s*)([a-zA-Z_][a-zA-Z0-9_]*)\s*:/g, '$1"$2":');
+        // Remove trailing commas before closing braces
+        repaired = repaired.replace(/,(\s*[}\]])/g, '$1');
+        try {
+            // Test if the repair worked
+            JSON.parse(repaired);
+            return repaired;
+        }
+        catch {
+            // If repair failed, return original
+            return jsonString;
+        }
+    }
+}
+
 /**
  * Utility function to ensure commit messages are properly formatted as strings
  * rather than JSON objects, whether they come as parsed objects or stringified JSON
@@ -8004,7 +8055,23 @@ function formatCommitMessage(result, options = {}) {
                 }
             }
             catch {
-                // Not valid JSON, continue to fallback
+                // Try to repair the JSON and parse again
+                try {
+                    const repairedJson = repairJson(jsonString);
+                    const parsed = JSON.parse(repairedJson);
+                    if (parsed &&
+                        typeof parsed === 'object' &&
+                        typeof parsed.title === 'string' &&
+                        typeof parsed.body === 'string' &&
+                        parsed.title.length > 0 &&
+                        parsed.body.length > 0) {
+                        // Successfully repaired and parsed JSON
+                        return constructMessage(parsed.title, parsed.body);
+                    }
+                }
+                catch {
+                    // Repair failed, continue to fallback
+                }
             }
         }
         // If no JSON found and it's already formatted, return as-is
@@ -11140,10 +11207,7 @@ for (var i = 0; i < 256; i++) {
   simpleEscapeMap[i] = simpleEscapeSequence(i);
 }
 
-// Max tokens for GPT-3 is 4096
-// const MAX_TOKENS_PER_SUMMARY = 4096
-const MAX_TOKENS_PER_SUMMARY = 12288;
-async function fileChangeParser({ changes, commit, options: { tokenizer, git, llm: model, logger }, }) {
+async function fileChangeParser({ changes, commit, options: { tokenizer, git, llm: model, logger, maxTokens }, }) {
     const textSplitter = new RecursiveCharacterTextSplitter({ chunkSize: 10000, chunkOverlap: 250 });
     const summarizationChain = loadSummarizationChain(model, {
         type: 'map_reduce',
@@ -11161,7 +11225,7 @@ async function fileChangeParser({ changes, commit, options: { tokenizer, git, ll
     logger.startTimer();
     const summary = await summarizeDiffs(diffs, {
         tokenizer,
-        maxTokens: MAX_TOKENS_PER_SUMMARY,
+        maxTokens: maxTokens || 4096,
         textSplitter,
         chain: summarizationChain,
         logger,
@@ -11464,7 +11528,7 @@ const handler$3 = async (argv, logger) => {
         return await fileChangeParser({
             changes,
             commit: '--staged',
-            options: { tokenizer, git, llm, logger },
+            options: { tokenizer, git, llm, logger, maxTokens: config.service.tokenLimit },
         });
     }
     const commitMsg = await generateAndReviewLoop({
@@ -11498,12 +11562,23 @@ const handler$3 = async (argv, logger) => {
             const schema = USE_CONVENTIONAL_COMMITS
                 ? ConventionalCommitMessageResponseSchema
                 : CommitMessageResponseSchema;
-            const formatInstructions = `You must always return a valid JSON object. Do not return any additional text. The JSON object you return should match the following schema:
+            const formatInstructions = `CRITICAL: You must return ONLY a valid JSON object with no additional text, explanations, or markdown formatting.
+
+REQUIRED JSON FORMAT:
 ${schema.description}
+
+EXAMPLE (follow this exact structure):
 {
-  "title": "The commit title",
-  "body": "The commit body"
-}`;
+  "title": "feat(auth): add user authentication system",
+  "body": "Implement JWT-based authentication with login and logout functionality. Includes password hashing and session management."
+}
+
+IMPORTANT RULES:
+- ALL string values MUST be enclosed in double quotes
+- NO trailing commas
+- NO comments or additional text outside the JSON
+- The "title" and "body" values must be properly quoted strings
+- Return ONLY the JSON object, nothing else`;
             // Use conventional commit prompt if enabled
             const promptTemplate = USE_CONVENTIONAL_COMMITS ? CONVENTIONAL_COMMIT_PROMPT : COMMIT_PROMPT;
             const prompt = getPrompt({

package/dist/index.js

@@ -69,7 +69,7 @@ var readline__namespace = /*#__PURE__*/_interopNamespaceDefault(readline);
 /**
  * Current build version from package.json
  */
-const BUILD_VERSION = "0.22.7";
+const BUILD_VERSION = "0.22.9";
 
 const isInteractive = (config) => {
     return config?.mode === 'interactive' || !!config?.interactive;
@@ -7234,6 +7234,8 @@ const CONVENTIONAL_TEMPLATE = `Generate a commit message that strictly adheres t
    - Scope should be a noun in parentheses (e.g., (parser), (ui), (config))
    - Omit scope if the change is broad or affects multiple areas
 
+CRITICAL: You must respond with ONLY valid JSON. All string values must be properly quoted.
+
 Based on the following diff summary, generate a conventional commit message that follows these rules exactly:
 
 """"""
@@ -7977,6 +7979,55 @@ async function executeChainWithSchema(schema, llm, prompt, variables, options =
     }
 }
 
+/**
+ * Utility to repair common JSON formatting issues that LLMs make
+ * Specifically handles cases where string values are not properly quoted
+ */
+function repairJson(jsonString) {
+    // Remove any markdown code block wrapping
+    let cleaned = jsonString.replace(/```(?:json)?\s*([\s\S]*?)\s*```/g, '$1').trim();
+    // Remove inline code block wrapping
+    cleaned = cleaned.replace(/^`(.*)`$/, '$1').trim();
+    // If it doesn't look like JSON, return as-is
+    if (!cleaned.startsWith('{') || !cleaned.endsWith('}')) {
+        return jsonString;
+    }
+    try {
+        // First try parsing as-is
+        JSON.parse(cleaned);
+        return cleaned;
+    }
+    catch {
+        // Try to repair common issues
+        let repaired = cleaned;
+        // Fix unquoted string values in title and body fields
+        // Pattern: "title": unquoted_value, -> "title": "unquoted_value",
+        repaired = repaired.replace(/"(title|body)":\s*([^",\{\}\[\]]+?)(?=\s*[,\}])/g, (match, field, value) => {
+            // Clean up the value (remove leading/trailing whitespace)
+            const cleanValue = value.trim();
+            // If it's already quoted or looks like a number/boolean, leave it
+            if (cleanValue.startsWith('"') || /^(true|false|\d+)$/.test(cleanValue)) {
+                return match;
+            }
+            // Quote the value
+            return `"${field}": "${cleanValue}"`;
+        });
+        // Fix missing quotes around field names (though this should be rare)
+        repaired = repaired.replace(/([{,]\s*)([a-zA-Z_][a-zA-Z0-9_]*)\s*:/g, '$1"$2":');
+        // Remove trailing commas before closing braces
+        repaired = repaired.replace(/,(\s*[}\]])/g, '$1');
+        try {
+            // Test if the repair worked
+            JSON.parse(repaired);
+            return repaired;
+        }
+        catch {
+            // If repair failed, return original
+            return jsonString;
+        }
+    }
+}
+
 /**
  * Utility function to ensure commit messages are properly formatted as strings
  * rather than JSON objects, whether they come as parsed objects or stringified JSON
@@ -8026,7 +8077,23 @@ function formatCommitMessage(result, options = {}) {
                 }
             }
             catch {
-                // Not valid JSON, continue to fallback
+                // Try to repair the JSON and parse again
+                try {
+                    const repairedJson = repairJson(jsonString);
+                    const parsed = JSON.parse(repairedJson);
+                    if (parsed &&
+                        typeof parsed === 'object' &&
+                        typeof parsed.title === 'string' &&
+                        typeof parsed.body === 'string' &&
+                        parsed.title.length > 0 &&
+                        parsed.body.length > 0) {
+                        // Successfully repaired and parsed JSON
+                        return constructMessage(parsed.title, parsed.body);
+                    }
+                }
+                catch {
+                    // Repair failed, continue to fallback
+                }
             }
         }
         // If no JSON found and it's already formatted, return as-is
@@ -11162,10 +11229,7 @@ for (var i = 0; i < 256; i++) {
   simpleEscapeMap[i] = simpleEscapeSequence(i);
 }
 
-// Max tokens for GPT-3 is 4096
-// const MAX_TOKENS_PER_SUMMARY = 4096
-const MAX_TOKENS_PER_SUMMARY = 12288;
-async function fileChangeParser({ changes, commit, options: { tokenizer, git, llm: model, logger }, }) {
+async function fileChangeParser({ changes, commit, options: { tokenizer, git, llm: model, logger, maxTokens }, }) {
     const textSplitter = new RecursiveCharacterTextSplitter({ chunkSize: 10000, chunkOverlap: 250 });
     const summarizationChain = loadSummarizationChain(model, {
         type: 'map_reduce',
@@ -11183,7 +11247,7 @@ async function fileChangeParser({ changes, commit, options: { tokenizer, git, ll
     logger.startTimer();
     const summary = await summarizeDiffs(diffs, {
         tokenizer,
-        maxTokens: MAX_TOKENS_PER_SUMMARY,
+        maxTokens: maxTokens || 4096,
         textSplitter,
         chain: summarizationChain,
         logger,
@@ -11486,7 +11550,7 @@ const handler$3 = async (argv, logger) => {
         return await fileChangeParser({
             changes,
             commit: '--staged',
-            options: { tokenizer, git, llm, logger },
+            options: { tokenizer, git, llm, logger, maxTokens: config.service.tokenLimit },
         });
     }
     const commitMsg = await generateAndReviewLoop({
@@ -11520,12 +11584,23 @@ const handler$3 = async (argv, logger) => {
             const schema = USE_CONVENTIONAL_COMMITS
                 ? ConventionalCommitMessageResponseSchema
                 : CommitMessageResponseSchema;
-            const formatInstructions = `You must always return a valid JSON object. Do not return any additional text. The JSON object you return should match the following schema:
+            const formatInstructions = `CRITICAL: You must return ONLY a valid JSON object with no additional text, explanations, or markdown formatting.
+
+REQUIRED JSON FORMAT:
 ${schema.description}
+
+EXAMPLE (follow this exact structure):
 {
-  "title": "The commit title",
-  "body": "The commit body"
-}`;
+  "title": "feat(auth): add user authentication system",
+  "body": "Implement JWT-based authentication with login and logout functionality. Includes password hashing and session management."
+}
+
+IMPORTANT RULES:
+- ALL string values MUST be enclosed in double quotes
+- NO trailing commas
+- NO comments or additional text outside the JSON
+- The "title" and "body" values must be properly quoted strings
+- Return ONLY the JSON object, nothing else`;
             // Use conventional commit prompt if enabled
             const promptTemplate = USE_CONVENTIONAL_COMMITS ? CONVENTIONAL_COMMIT_PROMPT : COMMIT_PROMPT;
             const prompt = getPrompt({

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "git-coco",
-  "version": "0.22.7",
+  "version": "0.22.9",
   "description": "zero-effort git commits with coco.",
   "author": "gfargo <ghfargo@gmail.com>",
   "license": "MIT",