@almightygpt/core 0.3.0 → 0.5.1

package/src/review/run-diff-review.ts

@@ -37,12 +37,16 @@ import {
 } from "../runs/folder.js";
 import type { RunMetadata } from "../runs/types.js";
 import { collectGitDiff } from "./diff.js";
+import { filterDiffByIgnoreLists } from "./diff-filter.js";
 import { assembleMemory } from "./memory.js";
 import {
   buildReviewerSystemFraming,
   buildReviewerUserMessage,
 } from "./prompts.js";
-import { writeHumanReviewFile } from "./write.js";
+import {
+  preflightReviewFileCollision,
+  writeHumanReviewFile,
+} from "./write.js";
 import { BudgetTracker, BudgetExceededError } from "./budget.js";
 
 export interface DiffReviewOptions {
@@ -63,6 +67,8 @@ export interface DiffReviewResult {
   provider: string;
   modelUsed: string;
   tokensIn: number;
+  /** Portion of tokensIn served from prompt cache (0 if no cache hit). */
+  cachedTokensIn: number;
   tokensOut: number;
   costUsd: number;
   latencyMs: number;
@@ -109,6 +115,16 @@ export async function runDiffReview(
     );
   }
 
+  // Preflight the review-file collision BEFORE any paid adapter call.
+  // (Codex review v0.5: previously the check ran post-adapter, burning API
+  // money on duplicate-topic runs that were going to fail anyway.)
+  await preflightReviewFileCollision(
+    opts.repoRoot,
+    config.reviewsDir,
+    opts.topic,
+    opts.force ?? false,
+  );
+
   // Create run folder up front so partial failures still leave an artifact.
   const runFolder = await createRunFolder({
     repoRoot: opts.repoRoot,
@@ -155,14 +171,29 @@ export async function runDiffReview(
       );
     }
 
+    // Codex review v0.5 P1 #2: filter ignored files BEFORE secret redaction
+    // and BEFORE the provider call. Honors .almightyignore + .gitignore +
+    // config.context.exclude. Redaction still runs on the survivors as
+    // defense-in-depth.
+    const filtered = await filterDiffByIgnoreLists({
+      repoRoot: opts.repoRoot,
+      diffText: diffResult.diff,
+      files: diffResult.files,
+      configExclude: config.context.exclude,
+    });
+
     const redaction = config.security.redactSecrets
-      ? redactSecrets(diffResult.diff)
-      : { text: diffResult.diff, redactions: [], totalCount: 0 };
+      ? redactSecrets(filtered.filteredDiff)
+      : { text: filtered.filteredDiff, redactions: [], totalCount: 0 };
 
     const manifest = buildContextManifest({
       inputSource: opts.range ? "diff-range" : "diff",
-      filesIncluded: diffResult.files.map((p) => ({ path: p, bytes: 0 })),
-      filesSkipped: [],
+      filesIncluded: filtered.filesIncluded.map((p) => ({ path: p, bytes: 0 })),
+      filesSkipped: filtered.filesSkipped.map((s) => ({
+        path: s.path,
+        bytes: 0,
+        skippedReason: s.reason,
+      })),
       diffText: redaction.text,
       redaction: {
         enabled: config.security.redactSecrets,
@@ -182,15 +213,18 @@ export async function runDiffReview(
     const userMessage = buildReviewerUserMessage({
       topic: opts.topic,
       diff: redaction.text,
-      files: diffResult.files,
+      files: filtered.filesIncluded,
     });
 
     const budget = new BudgetTracker(config.budget);
     const estimatedTokensIn = Math.ceil(
       (systemPrompt.length + userMessage.length) / 4,
     );
+    // Codex review v0.5 P2 #5: previously hardcoded "gpt-4o" here, which
+    // could block a cheap Gemini run on an OpenAI cost estimate. Use the
+    // actual adapter's default model so the preflight matches reality.
     budget.preflightCheck({
-      model: "gpt-4o", // estimator only; adapter resolves real model
+      model: adapter.defaultModel,
       estimatedTokensIn,
       maxOutputTokens: 4096,
     });
@@ -247,6 +281,9 @@ export async function runDiffReview(
           provider: adapter.provider,
           model: adapterOut.modelUsed,
           tokensIn: adapterOut.tokensIn,
+          ...(adapterOut.cachedTokensIn !== undefined && adapterOut.cachedTokensIn > 0
+            ? { cachedTokensIn: adapterOut.cachedTokensIn }
+            : {}),
           tokensOut: adapterOut.tokensOut,
           costUsd: adapterOut.costUsd,
           latencyMs: adapterOut.latencyMs,
@@ -268,11 +305,12 @@ export async function runDiffReview(
       provider: adapter.provider,
       modelUsed: adapterOut.modelUsed,
       tokensIn: adapterOut.tokensIn,
+      cachedTokensIn: adapterOut.cachedTokensIn ?? 0,
       tokensOut: adapterOut.tokensOut,
       costUsd: adapterOut.costUsd,
       latencyMs: adapterOut.latencyMs,
       diffEmpty: diffResult.empty,
-      filesReviewed: diffResult.files,
+      filesReviewed: filtered.filesIncluded,
       memorySources: memory.sources,
       memoryMissing: memory.missing,
       runId: runFolder.id,

package/src/review/run-worker-reviewer.ts

@@ -45,6 +45,7 @@ import {
 } from "../runs/folder.js";
 import type { AgentMetrics, RunMetadata } from "../runs/types.js";
 import { collectGitDiff } from "./diff.js";
+import { filterDiffByIgnoreLists } from "./diff-filter.js";
 import { assembleMemory } from "./memory.js";
 import {
   buildReviewerSystemFraming,
@@ -52,7 +53,10 @@ import {
   buildWorkerSystemFraming,
   buildWorkerUserMessage,
 } from "./prompts.js";
-import { writeHumanReviewFile } from "./write.js";
+import {
+  preflightReviewFileCollision,
+  writeHumanReviewFile,
+} from "./write.js";
 import { BudgetTracker, BudgetExceededError } from "./budget.js";
 import type { ReviewEventHandler } from "./events.js";
 
@@ -153,6 +157,16 @@ export async function runWorkerReviewerReview(
       ? `Worker and Reviewer are both ${workerCfg.provider}. Cross-vendor pairing catches more issues.`
       : undefined;
 
+  // Preflight the review-file collision BEFORE Worker or Reviewer fires.
+  // (Codex review v0.5: previously the check ran post-adapter, burning two
+  // adapter calls of API money on duplicate-topic runs.)
+  await preflightReviewFileCollision(
+    opts.repoRoot,
+    config.reviewsDir,
+    opts.topic,
+    opts.force ?? false,
+  );
+
   const runFolder = await createRunFolder({
     repoRoot: opts.repoRoot,
     runsDir: config.runsDir,
@@ -207,9 +221,19 @@ export async function runWorkerReviewerReview(
       );
     }
 
+    // Codex review v0.5 P1 #2: filter ignored files BEFORE redaction and
+    // BEFORE either adapter call. Honors .almightyignore + .gitignore +
+    // config.context.exclude.
+    const filtered = await filterDiffByIgnoreLists({
+      repoRoot: opts.repoRoot,
+      diffText: diffResult.diff,
+      files: diffResult.files,
+      configExclude: config.context.exclude,
+    });
+
     const redaction = config.security.redactSecrets
-      ? redactSecrets(diffResult.diff)
-      : { text: diffResult.diff, redactions: [], totalCount: 0 };
+      ? redactSecrets(filtered.filteredDiff)
+      : { text: filtered.filteredDiff, redactions: [], totalCount: 0 };
     emit({
       type: "redaction_complete",
       totalCount: redaction.totalCount,
@@ -218,8 +242,12 @@ export async function runWorkerReviewerReview(
 
     const manifest = buildContextManifest({
       inputSource: opts.range ? "diff-range" : "diff",
-      filesIncluded: diffResult.files.map((p) => ({ path: p, bytes: 0 })),
-      filesSkipped: [],
+      filesIncluded: filtered.filesIncluded.map((p) => ({ path: p, bytes: 0 })),
+      filesSkipped: filtered.filesSkipped.map((s) => ({
+        path: s.path,
+        bytes: 0,
+        skippedReason: s.reason,
+      })),
       diffText: redaction.text,
       redaction: {
         enabled: config.security.redactSecrets,
@@ -245,11 +273,14 @@ export async function runWorkerReviewerReview(
     const workerUser = buildWorkerUserMessage({
       topic: opts.topic,
       diff: redaction.text,
-      files: diffResult.files,
+      files: filtered.filesIncluded,
     });
 
+    // Codex review v0.5 P2 #5: use the actual adapter's default model
+    // for the preflight estimate so a cheap Gemini Worker isn't blocked
+    // by an OpenAI cost estimate.
     budget.preflightCheck({
-      model: "gpt-4o",
+      model: workerAdapter.defaultModel,
       estimatedTokensIn: Math.ceil(
         (workerSystem.length + workerUser.length) / 4,
       ),
@@ -281,6 +312,9 @@ export async function runWorkerReviewerReview(
       model: workerOut.modelUsed,
       outputPath: `${runFolder.relPath}/outputs/worker.md`,
       tokensIn: workerOut.tokensIn,
+      ...(workerOut.cachedTokensIn !== undefined && workerOut.cachedTokensIn > 0
+        ? { cachedTokensIn: workerOut.cachedTokensIn }
+        : {}),
       tokensOut: workerOut.tokensOut,
       costUsd: workerOut.costUsd,
       latencyMs: workerOut.latencyMs,
@@ -296,14 +330,14 @@ export async function runWorkerReviewerReview(
     const reviewerUser = buildReviewerOfWorkerUserMessage({
       topic: opts.topic,
       diff: redaction.text,
-      files: diffResult.files,
+      files: filtered.filesIncluded,
       workerOutput: workerOut.content,
       workerAgent: workerName,
       workerProvider: workerAdapter.provider,
     });
 
     budget.preflightCheck({
-      model: "gpt-4o",
+      model: reviewerAdapter.defaultModel,
       estimatedTokensIn: Math.ceil(
         (reviewerSystem.length + reviewerUser.length) / 4,
       ),
@@ -335,6 +369,9 @@ export async function runWorkerReviewerReview(
       model: reviewerOut.modelUsed,
       outputPath: `${runFolder.relPath}/outputs/reviewer.md`,
       tokensIn: reviewerOut.tokensIn,
+      ...(reviewerOut.cachedTokensIn !== undefined && reviewerOut.cachedTokensIn > 0
+        ? { cachedTokensIn: reviewerOut.cachedTokensIn }
+        : {}),
       tokensOut: reviewerOut.tokensOut,
       costUsd: reviewerOut.costUsd,
       latencyMs: reviewerOut.latencyMs,
@@ -390,6 +427,9 @@ export async function runWorkerReviewerReview(
         provider: workerAdapter.provider,
         model: workerOut.modelUsed,
         tokensIn: workerOut.tokensIn,
+        ...(workerOut.cachedTokensIn !== undefined && workerOut.cachedTokensIn > 0
+          ? { cachedTokensIn: workerOut.cachedTokensIn }
+          : {}),
         tokensOut: workerOut.tokensOut,
         costUsd: workerOut.costUsd,
         latencyMs: workerOut.latencyMs,
@@ -400,6 +440,9 @@ export async function runWorkerReviewerReview(
         provider: reviewerAdapter.provider,
         model: reviewerOut.modelUsed,
         tokensIn: reviewerOut.tokensIn,
+        ...(reviewerOut.cachedTokensIn !== undefined && reviewerOut.cachedTokensIn > 0
+          ? { cachedTokensIn: reviewerOut.cachedTokensIn }
+          : {}),
         tokensOut: reviewerOut.tokensOut,
         costUsd: reviewerOut.costUsd,
         latencyMs: reviewerOut.latencyMs,
@@ -442,7 +485,7 @@ export async function runWorkerReviewerReview(
       },
       metrics,
       totals,
-      filesReviewed: diffResult.files,
+      filesReviewed: filtered.filesIncluded,
       redactionsTotal: redaction.totalCount,
       memoryMissing: [
         ...new Set([...workerMemory.missing, ...reviewerMemory.missing]),

package/src/review/write.ts

@@ -86,6 +86,48 @@ export async function writeHumanReviewFile(
   return { path: relPath, bytes: content.length };
 }
 
+/**
+ * Compute the review file path for a topic without writing anything. Used
+ * by the review pipelines to **preflight** collisions *before* any paid
+ * adapter call. Same sanitization rules as writeHumanReviewFile.
+ */
+export function reviewFilePathFor(
+  reviewsDir: string,
+  topic: string,
+): string {
+  return join(reviewsDir, `${sanitizeTopic(topic)}.md`);
+}
+
+/**
+ * Preflight check: throws ReviewFileExistsError if the topic's review file
+ * already exists and `force` isn't set. Surfaces the error BEFORE any
+ * adapter is invoked so the user doesn't spend API money discovering an
+ * easily-avoidable collision.
+ *
+ * Returns the (relative) path so the caller doesn't have to recompute it.
+ */
+export async function preflightReviewFileCollision(
+  repoRoot: string,
+  reviewsDir: string,
+  topic: string,
+  force = false,
+): Promise<string> {
+  const relPath = reviewFilePathFor(reviewsDir, topic);
+  const absPath = join(repoRoot, relPath);
+  if (existsSync(absPath) && !force) {
+    // Also run the git-status check to give a richer dirty-file message
+    // when applicable; falls through to the existence error below.
+    await assertSafeToWrite(repoRoot, relPath, false);
+    throw new ReviewFileExistsError(
+      `Refusing to overwrite ${relPath}. Pass --force to overwrite (the ` +
+        `previous version remains in git history). Alternatively pick a ` +
+        `different --topic to keep both reviews.`,
+      relPath,
+    );
+  }
+  return relPath;
+}
+
 function sanitizeTopic(topic: string): string {
   // Allow letters, digits, dot, dash, underscore. Replace everything else
   // with a single dash. Trim leading/trailing dashes. Lowercase.

package/src/runs/types.ts

@@ -24,6 +24,12 @@ export interface AgentMetrics {
   provider: string;
   model: string;
   tokensIn: number;
+  /**
+   * Portion of tokensIn served from the provider's prompt cache (Anthropic
+   * cache_read, OpenAI prompt_tokens_details.cached_tokens). Always less
+   * than or equal to tokensIn. Zero when caching didn't trigger.
+   */
+  cachedTokensIn?: number;
   tokensOut: number;
   costUsd: number;
   latencyMs: number;