openairev 0.3.9 → 0.3.11

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openairev",
-  "version": "0.3.9",
+  "version": "0.3.11",
   "description": "Cross-model AI code reviewer — independent review for AI-assisted coding workflows",
   "type": "module",
   "bin": {

package/src/agents/codex.js

@@ -24,7 +24,7 @@ export class CodexAdapter {
     }
 
     args.push(prompt);
-    args.push('--json');
+    args.push('--json', '--sandbox', 'read-only');
 
     if (useSchema) {
       const schemaPath = join(__dirname, '../config', schemaFile);

package/src/cli/init.js

@@ -257,13 +257,15 @@ ${marker}
 
 This project uses OpenAIRev for independent AI code review. When the user asks to "review", "get a review", or "check my changes":
 
-1. Call \`openairev_review\` (MCP tool) with \`executor: "claude_code"\` and \`diff_cmd\` scoped to the files you changed (e.g. \`"git diff HEAD -- src/auth.ts src/routes.ts"\`). It returns immediately.
-2. Run \`openairev wait\` via Bash. It streams live progress and blocks until the verdict is ready — one command, that's it.
+1. Call \`openairev_review\` (MCP tool) with \`executor: "claude_code"\` and \`diff_cmd\` scoped to the files you changed (e.g. \`"git diff HEAD -- src/auth.ts src/routes.ts"\`). It returns immediately with a progress file path.
+2. Run \`openairev wait --file <progress_file_path>\` via Bash using the path from step 1. It streams live progress and blocks until the verdict is ready.
 
 **Rules — you MUST follow these:**
+- ALWAYS pass \`--file\` to \`openairev wait\` with the exact path from the review response. Without it, wait may look in the wrong directory.
 - ALWAYS scope \`diff_cmd\` to your changed files. Never send the full repo diff — the reviewer will run out of budget.
-- NEVER use \`sleep\`, polling loops, or manually read progress.json. \`openairev wait\` handles everything.
+- NEVER use \`sleep\`, polling loops, or manually read progress.json. \`openairev wait --file\` handles everything.
 - NEVER call \`openairev_status\`. Use \`openairev wait\` instead.
+- NEVER re-call \`openairev_review\` if wait fails — the progress file may take a few seconds to appear. Just retry \`openairev wait --file\`.
 - If a review is stuck or you want to retry, call \`openairev_cancel\` to kill the running review, then start a new one.
 - You MAY launch the review + wait in a sub-agent and continue other work in parallel.
 - Treat review feedback as peer review — use your judgment, focus on \`critical_issues\` and \`repair_instructions\`.
@@ -313,13 +315,15 @@ ${marker}
 
 This project uses OpenAIRev for independent AI code review. When the user asks to "review", "get a review", or "check my changes":
 
-1. Call \`openairev_review\` (MCP tool) with \`executor: "codex"\` and \`diff_cmd\` scoped to the files you changed (e.g. \`"git diff HEAD -- src/auth.ts src/routes.ts"\`). It returns immediately.
-2. Run \`openairev wait\` via Bash. It streams live progress and blocks until the verdict is ready — one command, that's it.
+1. Call \`openairev_review\` (MCP tool) with \`executor: "codex"\` and \`diff_cmd\` scoped to the files you changed (e.g. \`"git diff HEAD -- src/auth.ts src/routes.ts"\`). It returns immediately with a progress file path.
+2. Run \`openairev wait --file <progress_file_path>\` via Bash using the path from step 1. It streams live progress and blocks until the verdict is ready.
 
 **Rules — you MUST follow these:**
+- ALWAYS pass \`--file\` to \`openairev wait\` with the exact path from the review response. Without it, wait may look in the wrong directory.
 - ALWAYS scope \`diff_cmd\` to your changed files. Never send the full repo diff — the reviewer will run out of budget.
-- NEVER use \`sleep\`, polling loops, or manually read progress.json. \`openairev wait\` handles everything.
+- NEVER use \`sleep\`, polling loops, or manually read progress.json. \`openairev wait --file\` handles everything.
 - NEVER call \`openairev_status\`. Use \`openairev wait\` instead.
+- NEVER re-call \`openairev_review\` if wait fails — the progress file may take a few seconds to appear. Just retry \`openairev wait --file\`.
 - If a review is stuck or you want to retry, call \`openairev_cancel\` to kill the running review, then start a new one.
 - Treat review feedback as peer review — use your judgment, focus on \`critical_issues\` and \`repair_instructions\`.
 ${marker}

package/src/cli/wait.js

@@ -61,6 +61,9 @@ function printResult(data) {
   }
 
   console.log('');
+  if (data.partial_notice) {
+    console.log(`\n⚠ ${data.partial_notice}\n`);
+  }
   if (data.executor_feedback) {
     console.log(data.executor_feedback);
   } else if (data.verdict) {

package/src/mcp/mcp-server.js

@@ -76,13 +76,16 @@ server.tool(
       session.status = 'completed';
       saveSession(session, cwd);
 
-      writeProgress({
-        status: 'completed',
+      const progressData = {
+        status: review.error ? 'error' : 'completed',
         reviewer: reviewerName,
         progress: review.progress || [],
         verdict: review.verdict,
         executor_feedback: review.executor_feedback,
-      });
+      };
+      if (review.error) progressData.error = review.error;
+      if (review.partial_notice) progressData.partial_notice = review.partial_notice;
+      writeProgress(progressData);
       activeReview = null;
       activeAbort = null;
       return review;
@@ -96,7 +99,7 @@ server.tool(
     return {
       content: [{
         type: 'text',
-        text: `Review started. Reviewer: ${reviewerName}\nProgress file: ${PROGRESS_FILE}\n\nRun \`openairev wait\` from ${cwd} or read ${PROGRESS_FILE} directly to stream progress and get the verdict.`,
+        text: `Review started. Reviewer: ${reviewerName}\nProgress file: ${PROGRESS_FILE}\n\nRun \`openairev wait --file ${PROGRESS_FILE}\` via Bash to stream progress and get the verdict. The file may take a few seconds to appear — wait will handle this automatically. Do NOT re-call openairev_review or use sleep/polling.`,
       }],
     };
   }

package/src/review/input-stager.js

@@ -3,6 +3,108 @@ import { join } from 'path';
 
 const INLINE_THRESHOLD = 8_000; // characters — inline if under this
 
+// Rough estimate: ~4 characters per token for code/diffs
+const CHARS_PER_TOKEN = 4;
+
+// Reviewer context budgets (in tokens). Leave room for prompt, schema, and output.
+// These are conservative — better to compact than to blow the budget.
+const REVIEWER_BUDGETS = {
+  codex: 100_000,       // Codex has ~200k context, reserve half for output + tools
+  claude_code: 150_000, // Claude has ~200k context, needs less reserve
+};
+
+const DEFAULT_BUDGET = 100_000;
+
+/**
+ * Estimate token count from character length.
+ */
+export function estimateTokens(text) {
+  return Math.ceil(text.length / CHARS_PER_TOKEN);
+}
+
+/**
+ * Get the token budget for a reviewer.
+ */
+export function getReviewerBudget(reviewerName) {
+  return REVIEWER_BUDGETS[reviewerName] || DEFAULT_BUDGET;
+}
+
+/**
+ * Compact a diff to fit within a token budget.
+ * Strategy: parse into per-file hunks, sort by size ascending,
+ * drop the largest files first until it fits.
+ * Returns { content, compacted, stats }.
+ */
+export function compactDiff(diff, { maxTokens }) {
+  const tokens = estimateTokens(diff);
+  if (tokens <= maxTokens) {
+    return { content: diff, compacted: false, partial: false, stats: { originalTokens: tokens, finalTokens: tokens, filesDropped: 0 } };
+  }
+
+  const files = parseDiffFiles(diff);
+  // Sort by size descending — drop largest first
+  files.sort((a, b) => b.content.length - a.content.length);
+
+  // Reserve chars for the omission notice (worst case)
+  const noticeOverhead = 500;
+  const maxChars = (maxTokens * CHARS_PER_TOKEN) - noticeOverhead;
+  let totalChars = diff.length;
+  const dropped = [];
+
+  while (totalChars > maxChars && files.length > 1) {
+    const largest = files.shift();
+    totalChars -= largest.content.length;
+    dropped.push(largest.name);
+  }
+
+  // If single file still too large, truncate it
+  if (files.length === 1 && totalChars > maxChars) {
+    const file = files[0];
+    const truncateAt = maxChars - 200;
+    file.content = file.content.slice(0, Math.max(0, truncateAt)) + `\n\n... [TRUNCATED — file too large for reviewer context] ...\n`;
+  }
+
+  let compactedDiff = files.map(f => f.content).join('\n');
+  if (dropped.length > 0) {
+    const notice = `\n\n--- FILES OMITTED (too large for reviewer context) ---\n${dropped.map(f => `  ${f}`).join('\n')}\n--- This is a PARTIAL review. Only the files below were included. ---\n`;
+    compactedDiff = notice + compactedDiff;
+  }
+
+  const finalTokens = estimateTokens(compactedDiff);
+
+  return {
+    content: compactedDiff,
+    compacted: true,
+    partial: dropped.length > 0,
+    stats: {
+      originalTokens: tokens,
+      finalTokens,
+      filesDropped: dropped.length,
+      droppedFiles: dropped,
+    },
+  };
+}
+
+/**
+ * Parse a unified diff into per-file sections.
+ */
+function parseDiffFiles(diff) {
+  const files = [];
+  // Split on diff headers (diff --git, or --- a/ lines preceded by blank)
+  const parts = diff.split(/(?=^diff --git )/m);
+
+  for (const part of parts) {
+    if (!part.trim()) continue;
+    const nameMatch = part.match(/^diff --git a\/(.+?) b\//m)
+      || part.match(/^---\s+a\/(.+)/m)
+      || part.match(/^\+\+\+\s+b\/(.+)/m);
+    const name = nameMatch ? nameMatch[1] : 'unknown';
+    files.push({ name, content: part });
+  }
+
+  return files;
+}
+
 /**
  * Stage review input. For small content, returns it inline.
  * For large content, writes to .openairev/tmp/ and returns a file reference.

package/src/review/review-runner.js

@@ -1,7 +1,7 @@
 import { mkdirSync, writeFileSync } from 'fs';
 import { join } from 'path';
 import { createAdapter } from '../agents/registry.js';
-import { stageInput, buildInputReference } from './input-stager.js';
+import { stageInput, buildInputReference, compactDiff, getReviewerBudget, estimateTokens } from './input-stager.js';
 import { loadPromptFile } from './prompt-loader.js';
 
 export async function runReview(content, {
@@ -22,7 +22,12 @@ export async function runReview(content, {
   }
 
   const reviewerPrompt = loadPromptFile(promptFile, cwd);
-  const staged = stageInput(content, { cwd });
+
+  // Compact diff if it exceeds the reviewer's context budget
+  const budget = getReviewerBudget(reviewerName);
+  const { content: compactedContent, compacted, stats } = compactDiff(content, { maxTokens: budget });
+
+  const staged = stageInput(compactedContent, { cwd });
   const inputRef = buildInputReference(staged);
 
   let prompt = reviewerPrompt;
@@ -51,7 +56,7 @@ export async function runReview(content, {
 
   logReviewerOutput(rawOutput, reviewerName, cwd);
 
-  return {
+  const reviewResult = {
     reviewer: reviewerName,
     verdict,
     executor_feedback: executorFeedback,
@@ -59,6 +64,31 @@ export async function runReview(content, {
     progress: result?.progress || [],
     session_id: adapter.sessionName || adapter.sessionId,
   };
+
+  if (compacted) {
+    reviewResult.context_stats = stats;
+  }
+
+  // Mark partial reviews so callers know files were omitted
+  if (compacted && stats.filesDropped > 0 && verdict) {
+    reviewResult.partial = true;
+    reviewResult.partial_notice = `PARTIAL REVIEW: ${stats.filesDropped} files were omitted to fit the reviewer's context budget (${stats.droppedFiles.join(', ')}). Re-run with a smaller diff to cover them.`;
+  }
+
+  // Diagnose null verdicts — check for explicit errors before blaming context budget
+  if (!verdict) {
+    const explicitError = result?.error;
+    if (explicitError) {
+      reviewResult.error = `Reviewer (${reviewerName}) failed: ${explicitError}`;
+    } else {
+      reviewResult.error = `Reviewer (${reviewerName}) produced no verdict. ` +
+        `Diff was ~${stats.originalTokens} tokens` +
+        (compacted ? `, compacted to ~${stats.finalTokens} tokens (${stats.filesDropped} files dropped)` : '') +
+        `. Possible causes: context budget exceeded, auth failure, or schema mismatch. Check .openairev/logs/ for details.`;
+    }
+  }
+
+  return reviewResult;
 }
 
 function logReviewerOutput(rawOutput, reviewerName, cwd) {