@delegance/claude-autopilot 2.1.0 → 2.3.0

package/CHANGELOG.md

@@ -1,5 +1,24 @@
 # Changelog
 
+## [2.3.0] — 2026-04-22
+
+### Added
+- **Parallel chunk review** — file-level chunks are now reviewed concurrently (default parallelism: 3, configurable via `chunking.parallelism`); serial fallback preserved when `cost.budgetUSD` is set so budget enforcement remains accurate
+- **`.autopilot-ignore`** — project-level suppression file; format: `<rule-id> <glob>` or bare `<glob>` (matches any finding on that path); comments and blank lines ignored; suppressed count printed dim after run
+- **`--delta` mode** — only reports findings new since the previous run; pre-existing findings are hidden and the count is printed dim; findings always persisted to `.autopilot-cache/findings.json` after each run (gitignored)
+- `src/core/ignore/index.ts` — `loadIgnoreRules()`, `applyIgnoreRules()`
+- `src/core/persist/findings-cache.ts` — `loadCachedFindings()`, `saveCachedFindings()`, `filterNewFindings()`
+- 15 new tests — **248 total**
+
+## [2.2.0] — 2026-04-22
+
+### Added
+- **`reviewStrategy: diff`** — new chunking strategy that sends `git diff` unified hunks instead of full file contents; typically ~70% fewer tokens and more focused findings (LLM sees exactly what changed)
+- **`--diff` flag** on `run` and `ci` subcommands — shorthand to activate diff strategy without editing config
+- **`src/core/git/diff-hunks.ts`** — `getFileDiffs()`, `parseUnifiedDiff()`, `formatDiffContent()`; per-file diff sections in fenced code blocks; files that exceed `maxChars` are omitted with a count notice
+- `BuildChunksInput.base` / `ReviewPhaseInput.base` / `RunInput.base` — threads git base ref through pipeline to diff engine
+- 9 new tests for `parseUnifiedDiff` and `formatDiffContent` — **233 total**
+
 ## [2.1.0] — 2026-04-22
 
 ### Added

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@delegance/claude-autopilot",
-  "version": "2.1.0",
+  "version": "2.3.0",
   "type": "module",
   "description": "Claude Code automation pipeline: spec → plan → implement → validate → PR",
   "keywords": [

package/src/cli/ci.ts

@@ -6,6 +6,7 @@ export interface CiCommandOptions {
   base?: string;
   postComments?: boolean;
   sarifOutput?: string;
+  diff?: boolean;
 }
 
 /**
@@ -34,5 +35,6 @@ export async function runCi(options: CiCommandOptions = {}): Promise<number> {
     postComments: options.postComments ?? true,
     format: 'sarif',
     outputPath: sarifOutput,
+    diff: options.diff,
   });
 }

package/src/cli/index.ts

@@ -66,6 +66,8 @@ Options (run):
   --config <path>      Path to config file (default: ./autopilot.config.yaml)
   --files <a,b,c>      Explicit comma-separated file list (skips git detection)
   --dry-run            Show what would run without executing
+  --diff               Send git diff hunks instead of full files (~70% fewer tokens)
+  --delta              Only report findings new since last run (suppress pre-existing)
   --post-comments      Post/update a summary comment on the open PR
   --format <text|sarif>  Output format (default: text)
   --output <path>        Output file path (required with --format sarif)
@@ -120,6 +122,8 @@ switch (subcommand) {
     const config = flag('config');
     const filesArg = flag('files');
     const dryRun = boolFlag('dry-run');
+    const diff = boolFlag('diff');
+    const delta = boolFlag('delta');
     const postComments = boolFlag('post-comments');
     const formatArg = flag('format');
     const outputPath = flag('output');
@@ -138,6 +142,8 @@ switch (subcommand) {
       configPath: config,
       files: filesArg ? filesArg.split(',').map(f => f.trim()) : undefined,
       dryRun,
+      diff,
+      delta,
       postComments,
       format: formatArg as 'text' | 'sarif' | undefined,
       outputPath,
@@ -151,11 +157,13 @@ switch (subcommand) {
     const config = flag('config');
     const outputPath = flag('output');
     const noPostComments = boolFlag('no-post-comments');
+    const diff = boolFlag('diff');
     const code = await runCi({
       configPath: config,
       base,
       sarifOutput: outputPath,
       postComments: noPostComments ? false : undefined,
+      diff,
     });
     process.exit(code);
     break;

package/src/cli/run.ts

@@ -38,6 +38,8 @@ import { detectProtectedPaths } from '../core/detect/protected-paths.ts';
 import { detectGitContext } from '../core/detect/git-context.ts';
 import { detectProject } from './detector.ts';
 import { detectPrNumber, formatComment, postPrComment } from './pr-comment.ts';
+import { loadIgnoreRules, applyIgnoreRules } from '../core/ignore/index.ts';
+import { loadCachedFindings, saveCachedFindings, filterNewFindings } from '../core/persist/findings-cache.ts';
 
 function readToolVersion(): string {
   const pkgPath = path.join(path.dirname(fileURLToPath(import.meta.url)), '../../package.json');
@@ -64,6 +66,8 @@ export interface RunCommandOptions {
   base?: string;        // git base ref (default HEAD~1)
   files?: string[];     // explicit file list (skips git detection)
   dryRun?: boolean;     // skip review, print what would run
+  diff?: boolean;       // use diff strategy (send git hunks instead of full files)
+  delta?: boolean;      // only report findings not present in last run's baseline
   format?: 'text' | 'sarif';
   outputPath?: string;
   postComments?: boolean; // post/update summary comment on the open PR
@@ -167,6 +171,11 @@ export async function runCommand(options: RunCommandOptions = {}): Promise<numbe
     ? await loadRulesFromConfig(config.staticRules)
     : [];
 
+  // Apply --diff flag: override reviewStrategy to 'diff'
+  if (options.diff && config.reviewStrategy !== 'diff') {
+    config = { ...config, reviewStrategy: 'diff' };
+  }
+
   // Execute pipeline
   const input: RunInput = {
     touchedFiles,
@@ -175,11 +184,42 @@ export async function runCommand(options: RunCommandOptions = {}): Promise<numbe
     staticRules,
     cwd,
     gitSummary: gitCtx.summary ?? undefined,
+    base: options.base,
   };
 
   console.log('');
   const result = await runAutopilot(input);
 
+  // Apply .autopilot-ignore suppression rules
+  const ignoreRules = loadIgnoreRules(cwd);
+  if (ignoreRules.length > 0) {
+    const before = result.allFindings.length;
+    result.allFindings = applyIgnoreRules(result.allFindings, ignoreRules);
+    for (const phase of result.phases) {
+      phase.findings = applyIgnoreRules(phase.findings, ignoreRules);
+    }
+    const suppressed = before - result.allFindings.length;
+    if (suppressed > 0) {
+      console.log(fmt('dim', `  [run] ${suppressed} finding${suppressed !== 1 ? 's' : ''} suppressed by .autopilot-ignore`));
+    }
+  }
+
+  // Delta mode: filter to only new findings vs last run's baseline, then persist
+  if (options.delta) {
+    const cached = loadCachedFindings(cwd);
+    const before = result.allFindings.length;
+    result.allFindings = filterNewFindings(result.allFindings, cached);
+    for (const phase of result.phases) {
+      phase.findings = filterNewFindings(phase.findings, cached);
+    }
+    const existing = before - result.allFindings.length;
+    if (existing > 0) {
+      console.log(fmt('dim', `  [run] ${existing} pre-existing finding${existing !== 1 ? 's' : ''} hidden (--delta mode)`));
+    }
+  }
+  // Always persist the unfiltered findings as the new baseline
+  saveCachedFindings(cwd, result.allFindings);
+
   // emitAnnotations is a no-op unless GITHUB_ACTIONS=true
   emitAnnotations(result.allFindings);
 

package/src/core/chunking/index.ts

@@ -3,6 +3,7 @@ import * as path from 'node:path';
 import type { ReviewEngine, ReviewInput } from '../../adapters/review-engine/types.ts';
 import type { AutopilotConfig } from '../config/types.ts';
 import { rankByRisk } from './risk-ranker.ts';
+import { getFileDiffs, formatDiffContent } from '../git/diff-hunks.ts';
 
 export interface ReviewChunk {
   content: string;
@@ -12,11 +13,12 @@ export interface ReviewChunk {
 
 export interface BuildChunksInput {
   touchedFiles: string[];
-  strategy: 'auto' | 'single-pass' | 'file-level';
+  strategy: 'auto' | 'single-pass' | 'file-level' | 'diff';
   chunking?: AutopilotConfig['chunking'];
   engine: ReviewEngine;
   cwd?: string;
   protectedPaths?: string[];
+  base?: string;  // git base ref — required for 'diff' strategy
 }
 
 const DEFAULT_SMALL_TIER_TOKENS = 8000;
@@ -26,6 +28,11 @@ export async function buildReviewChunks(input: BuildChunksInput): Promise<Review
   const smallMax = input.chunking?.smallTierMaxTokens ?? DEFAULT_SMALL_TIER_TOKENS;
   const fileMax = input.chunking?.perFileMaxTokens ?? DEFAULT_FILE_TIER_TOKENS;
 
+  // Diff strategy: send unified diff hunks instead of full file contents
+  if (input.strategy === 'diff') {
+    return buildDiffChunks(input);
+  }
+
   const ranked = rankByRisk(input.touchedFiles, { protectedPaths: input.protectedPaths });
   const fileContents = await readFiles(ranked, input.cwd);
 
@@ -51,6 +58,19 @@ export async function buildReviewChunks(input: BuildChunksInput): Promise<Review
   return chunks;
 }
 
+function buildDiffChunks(input: BuildChunksInput): ReviewChunk[] {
+  const cwd = input.cwd ?? process.cwd();
+  const base = input.base ?? 'HEAD~1';
+  const ranked = rankByRisk(input.touchedFiles, { protectedPaths: input.protectedPaths });
+  const diffs = getFileDiffs(cwd, base, ranked);
+
+  if (diffs.length === 0) return [];
+
+  // Single chunk — diff content is already compact; truncation handled in formatDiffContent
+  const content = formatDiffContent(diffs);
+  return [{ content, kind: 'file-batch', files: diffs.map(d => d.file) }];
+}
+
 async function readFiles(touchedFiles: string[], cwd?: string): Promise<Map<string, string>> {
   const result = new Map<string, string>();
   for (const f of touchedFiles) {

package/src/core/config/types.ts

@@ -27,7 +27,7 @@ export interface AutopilotConfig {
     maxCodexRetries?: number;
     maxBugbotRounds?: number;
   };
-  reviewStrategy?: 'auto' | 'single-pass' | 'file-level';
+  reviewStrategy?: 'auto' | 'single-pass' | 'file-level' | 'diff';
   chunking?: {
     smallTierMaxTokens?: number;
     partialReviewTokens?: number;

package/src/core/git/diff-hunks.ts

@@ -0,0 +1,86 @@
+import { runSafe } from '../shell.ts';
+
+export interface FileDiff {
+  file: string;
+  hunks: string;  // unified diff content for this file (header + hunks)
+  additions: number;
+  deletions: number;
+}
+
+/**
+ * Returns per-file unified diffs for the given files between base and HEAD.
+ * Falls back to working-tree diff (unstaged) when base diff is empty for a file.
+ */
+export function getFileDiffs(cwd: string, base: string, files: string[]): FileDiff[] {
+  if (files.length === 0) return [];
+
+  // Get full diff in one shot — more efficient than per-file calls
+  const raw = runSafe('git', ['diff', base, 'HEAD', '--unified=3', '--', ...files], { cwd })
+    ?? runSafe('git', ['diff', 'HEAD', '--unified=3', '--', ...files], { cwd })
+    ?? '';
+
+  return parseUnifiedDiff(raw, files);
+}
+
+/**
+ * Parses unified diff output into per-file FileDiff entries.
+ * Only returns files that actually have diff content.
+ */
+export function parseUnifiedDiff(raw: string, requestedFiles: string[]): FileDiff[] {
+  if (!raw.trim()) return [];
+
+  const results: FileDiff[] = [];
+  const sections = raw.split(/^(?=diff --git )/m).filter(Boolean);
+
+  const requested = new Set(requestedFiles.map(f => f.replace(/\\/g, '/')));
+
+  for (const section of sections) {
+    // Extract b/ filename from diff header: diff --git a/src/foo.ts b/src/foo.ts
+    const headerMatch = section.match(/^diff --git a\/.+ b\/(.+)$/m);
+    if (!headerMatch) continue;
+    const file = headerMatch[1]!.trim();
+    if (!requested.has(file)) continue;
+
+    // Strip the git binary/index header lines, keep hunk content
+    const hunkStart = section.indexOf('@@');
+    const hunks = hunkStart >= 0 ? section.slice(hunkStart) : '';
+    if (!hunks.trim()) continue;
+
+    let additions = 0;
+    let deletions = 0;
+    for (const line of hunks.split('\n')) {
+      if (line.startsWith('+') && !line.startsWith('+++')) additions++;
+      if (line.startsWith('-') && !line.startsWith('---')) deletions++;
+    }
+
+    results.push({ file, hunks: hunks.trimEnd(), additions, deletions });
+  }
+
+  return results;
+}
+
+/**
+ * Formats FileDiff entries into a review-ready string.
+ * Total size is bounded by maxChars (default 120K chars ≈ 30K tokens).
+ */
+export function formatDiffContent(diffs: FileDiff[], maxChars = 120_000): string {
+  const parts: string[] = [];
+  let total = 0;
+  let skipped = 0;
+
+  for (const d of diffs) {
+    const section = `## ${d.file} (+${d.additions}/-${d.deletions})\n\`\`\`diff\n${d.hunks}\n\`\`\``;
+    if (total + section.length > maxChars) {
+      skipped++;
+      continue;
+    }
+    parts.push(section);
+    total += section.length;
+  }
+
+  if (skipped > 0) {
+    parts.push(`[${skipped} file${skipped !== 1 ? 's' : ''} omitted — diff exceeded size limit]`);
+  }
+
+  return parts.join('\n\n');
+}

package/src/core/ignore/index.ts

@@ -0,0 +1,42 @@
+import * as fs from 'node:fs';
+import * as path from 'node:path';
+import { minimatch } from 'minimatch';
+import type { Finding } from '../findings/types.ts';
+
+export interface IgnoreRule {
+  ruleId: string | '*';  // finding id prefix or '*' for any
+  pathGlob: string | null; // null = match all paths
+}
+
+export function loadIgnoreRules(cwd: string): IgnoreRule[] {
+  const filePath = path.join(cwd, '.autopilot-ignore');
+  if (!fs.existsSync(filePath)) return [];
+
+  const rules: IgnoreRule[] = [];
+  for (const raw of fs.readFileSync(filePath, 'utf8').split('\n')) {
+    const line = raw.trim();
+    if (!line || line.startsWith('#')) continue;
+
+    const parts = line.split(/\s+/);
+    if (parts.length === 1) {
+      // bare glob — suppress any finding whose file matches
+      rules.push({ ruleId: '*', pathGlob: parts[0]! });
+    } else {
+      // <rule-id-or-*> <path-glob>
+      rules.push({ ruleId: parts[0]!, pathGlob: parts[1]! });
+    }
+  }
+  return rules;
+}
+
+function matchesRule(finding: Finding, rule: IgnoreRule): boolean {
+  const ruleMatches = rule.ruleId === '*' || finding.id.startsWith(rule.ruleId);
+  if (!ruleMatches) return false;
+  if (rule.pathGlob === null) return true;
+  return minimatch(finding.file.replace(/\\/g, '/'), rule.pathGlob, { matchBase: true });
+}
+
+export function applyIgnoreRules(findings: Finding[], rules: IgnoreRule[]): Finding[] {
+  if (rules.length === 0) return findings;
+  return findings.filter(f => !rules.some(r => matchesRule(f, r)));
+}

package/src/core/persist/findings-cache.ts

@@ -0,0 +1,43 @@
+import * as fs from 'node:fs';
+import * as path from 'node:path';
+import type { Finding } from '../findings/types.ts';
+
+const CACHE_DIR = '.autopilot-cache';
+const CACHE_FILE = 'findings.json';
+
+function cacheFilePath(cwd: string): string {
+  return path.join(cwd, CACHE_DIR, CACHE_FILE);
+}
+
+function findingKey(f: Finding): string {
+  return `${f.id}::${f.file}::${f.line ?? ''}`;
+}
+
+export function loadCachedFindings(cwd: string): Finding[] {
+  const p = cacheFilePath(cwd);
+  if (!fs.existsSync(p)) return [];
+  try {
+    return JSON.parse(fs.readFileSync(p, 'utf8')) as Finding[];
+  } catch {
+    return [];
+  }
+}
+
+export function saveCachedFindings(cwd: string, findings: Finding[]): void {
+  const dir = path.join(cwd, CACHE_DIR);
+  fs.mkdirSync(dir, { recursive: true });
+  // atomic write
+  const tmp = cacheFilePath(cwd) + '.tmp';
+  fs.writeFileSync(tmp, JSON.stringify(findings, null, 2), 'utf8');
+  fs.renameSync(tmp, cacheFilePath(cwd));
+}
+
+/**
+ * Returns only findings not present in the cached baseline.
+ * Two findings are considered the same when id + file + line all match.
+ */
+export function filterNewFindings(current: Finding[], cached: Finding[]): Finding[] {
+  if (cached.length === 0) return current;
+  const seen = new Set(cached.map(findingKey));
+  return current.filter(f => !seen.has(findingKey(f)));
+}

package/src/core/pipeline/review-phase.ts

@@ -1,7 +1,7 @@
 import type { ReviewEngine } from '../../adapters/review-engine/types.ts';
 import type { Finding } from '../findings/types.ts';
 import type { AutopilotConfig } from '../config/types.ts';
-import { buildReviewChunks } from '../chunking/index.ts';
+import { buildReviewChunks, type ReviewChunk } from '../chunking/index.ts';
 
 export interface ReviewPhaseResult {
   phase: 'review';
@@ -19,6 +19,49 @@ export interface ReviewPhaseInput {
   cwd?: string;
   gitSummary?: string;
   budgetRemainingUSD?: number;
+  base?: string;
+}
+
+interface ChunkResult {
+  findings: Finding[];
+  inputTokens: number;
+  outputTokens: number;
+  costUSD: number;
+}
+
+async function reviewChunk(chunk: ReviewChunk, input: ReviewPhaseInput): Promise<ChunkResult> {
+  const output = await input.engine.review({
+    content: chunk.content,
+    kind: chunk.kind,
+    context: { stack: input.config.stack, cwd: input.cwd, gitSummary: input.gitSummary },
+  });
+  return {
+    findings: output.findings,
+    inputTokens: output.usage?.input ?? 0,
+    outputTokens: output.usage?.output ?? 0,
+    costUSD: output.usage?.costUSD ?? 0,
+  };
+}
+
+/** Run up to `limit` promises concurrently, preserving result order. */
+async function pMap<T, R>(
+  items: T[],
+  fn: (item: T, index: number) => Promise<R>,
+  limit: number,
+): Promise<R[]> {
+  const results: R[] = new Array(items.length);
+  let next = 0;
+
+  async function worker(): Promise<void> {
+    while (next < items.length) {
+      const i = next++;
+      results[i] = await fn(items[i]!, i);
+    }
+  }
+
+  const workers = Array.from({ length: Math.min(limit, items.length) }, () => worker());
+  await Promise.all(workers);
+  return results;
 }
 
 export async function runReviewPhase(input: ReviewPhaseInput): Promise<ReviewPhaseResult> {
@@ -35,43 +78,54 @@ export async function runReviewPhase(input: ReviewPhaseInput): Promise<ReviewPha
     engine: input.engine,
     cwd: input.cwd,
     protectedPaths: input.config.protectedPaths,
+    base: input.base,
   });
 
-  const allFindings: Finding[] = [];
-  let totalInputTokens = 0;
-  let totalOutputTokens = 0;
-  let totalCostUSD = 0;
-  let budgetExceeded = false;
+  const parallelism = input.config.chunking?.parallelism ?? 3;
+  const budgetUSD = input.budgetRemainingUSD;
 
-  for (const chunk of chunks) {
-    if (input.budgetRemainingUSD !== undefined && totalCostUSD >= input.budgetRemainingUSD) {
-      budgetExceeded = true;
-      break;
+  // For budget tracking we still need to enforce it — run serially if budget set,
+  // parallel otherwise (budget check between serial chunks is the safe path).
+  let chunkResults: ChunkResult[];
+  if (budgetUSD !== undefined) {
+    chunkResults = [];
+    let spent = 0;
+    let budgetExceeded = false;
+    for (const chunk of chunks) {
+      if (spent >= budgetUSD) { budgetExceeded = true; break; }
+      const r = await reviewChunk(chunk, input);
+      spent += r.costUSD;
+      chunkResults.push(r);
     }
-    const output = await input.engine.review({
-      content: chunk.content,
-      kind: chunk.kind,
-      context: { stack: input.config.stack, cwd: input.cwd, gitSummary: input.gitSummary },
-    });
-    allFindings.push(...output.findings);
-    if (output.usage) {
-      totalInputTokens += output.usage.input;
-      totalOutputTokens += output.usage.output;
-      if (output.usage.costUSD !== undefined) totalCostUSD += output.usage.costUSD;
+    if (budgetExceeded) {
+      chunkResults.push({
+        findings: [{
+          id: 'budget-exceeded',
+          source: 'pipeline',
+          severity: 'warning',
+          category: 'budget',
+          file: '<pipeline>',
+          message: `Review budget of $${budgetUSD} USD exceeded — remaining chunks skipped`,
+          protectedPath: false,
+          createdAt: new Date().toISOString(),
+        }],
+        inputTokens: 0, outputTokens: 0, costUSD: 0,
+      });
     }
+  } else {
+    chunkResults = await pMap(chunks, chunk => reviewChunk(chunk, input), parallelism);
   }
 
-  if (budgetExceeded) {
-    allFindings.push({
-      id: 'budget-exceeded',
-      source: 'pipeline',
-      severity: 'warning',
-      category: 'budget',
-      file: '<pipeline>',
-      message: `Review budget of $${input.budgetRemainingUSD} USD exceeded — remaining chunks skipped`,
-      protectedPath: false,
-      createdAt: new Date().toISOString(),
-    });
+  let totalInputTokens = 0;
+  let totalOutputTokens = 0;
+  let totalCostUSD = 0;
+  const allFindings: Finding[] = [];
+
+  for (const r of chunkResults) {
+    allFindings.push(...r.findings);
+    totalInputTokens += r.inputTokens;
+    totalOutputTokens += r.outputTokens;
+    totalCostUSD += r.costUSD;
   }
 
   const hasCritical = allFindings.some(f => f.severity === 'critical');

package/src/core/pipeline/run.ts

@@ -18,6 +18,7 @@ export interface RunInput {
   staticRules?: StaticRule[];
   cwd?: string;
   gitSummary?: string;
+  base?: string;
 }
 
 export interface RunResult {
@@ -62,6 +63,7 @@ export async function runAutopilot(input: RunInput): Promise<RunResult> {
       cwd: input.cwd,
       gitSummary: input.gitSummary,
       budgetRemainingUSD: budgetUSD,
+      base: input.base,
     });
     phases.push(reviewResult);
     if (reviewResult.costUSD !== undefined) {