From a224617abbfb9c4eeb20b13e6bfaa66f20d6a8db Mon Sep 17 00:00:00 2001 From: chs0813 <85471619@qq.com> Date: Sat, 4 Jul 2026 22:48:44 +0800 Subject: [PATCH] fix(hooks): do not echo raw input from plugin-hook-bootstrap.js Rebased onto origin/main (49128b576). Fixtures moved from scripts/hooks/ to /tmp/ecc-pr2380-fixtures/ per reviewer feedback. (Original commit b0e49036 was based on cc6724ee; main has since refactored spawnShell to use a shellArgs variable and added PowerShell .sh fallback paths. This rebase adapts the const result = spawnSync(...) + __rawInput tagging pattern to all three spawnSync call sites in spawnShell.) --- scripts/hooks/plugin-hook-bootstrap.js | 58 +++- .../plugin-hook-bootstrap-no-echo.test.js | 256 ++++++++++++++++++ tests/hooks/plugin-hook-bootstrap.test.js | 37 ++- 3 files changed, 326 insertions(+), 25 deletions(-) create mode 100644 tests/hooks/plugin-hook-bootstrap-no-echo.test.js diff --git a/scripts/hooks/plugin-hook-bootstrap.js b/scripts/hooks/plugin-hook-bootstrap.js index 00fce645a..c23ff0167 100644 --- a/scripts/hooks/plugin-hook-bootstrap.js +++ b/scripts/hooks/plugin-hook-bootstrap.js @@ -22,15 +22,40 @@ function writeStderr(stderr) { } } -function passthrough(raw, result) { +function passthrough(result) { const stdout = typeof result?.stdout === 'string' ? result.stdout : ''; if (stdout) { + // Most ECC hook scripts follow a `run(rawInput) -> rawInput` passthrough + // pattern: they do their work, then return the original input so the hook + // chain's tool result is preserved. The harness then writes the verbatim + // raw input (tool_input + tool_response, often 1-275 KB) into the session + // transcript as a hook_success attachment -- ~89% of every ECC session's + // transcript is this bloat. Detect the passthrough and emit empty stdout + // instead; the harness falls back to the tool_use's original result, the + // same path #2240 established for bash-hook-dispatcher. + // + // IMPORTANT: a strict `stdout === raw` check misses the common case where + // child processes' synchronous `process.stdout.write()` writes hit the + // ~64 KB Node.js pipe buffer and get truncated -- stdout is then exactly + // 65536 bytes and a strict prefix of raw. So we also detect that + // truncation sentinel. + const raw = typeof result?.__rawInput === 'string' ? result.__rawInput : ''; + const STDOUT_PIPE_CAP = 64 * 1024; + const looksLikePassthrough = + (stdout.length === STDOUT_PIPE_CAP && raw.startsWith(stdout)) || + (raw.length > 0 && stdout === raw); + if (looksLikePassthrough) { + writeStderr( + '[Hook] bootstrap: hook returned raw input as stdout; emitting empty to avoid transcript bloat\n' + ); + return; + } process.stdout.write(stdout); return; } if (!Number.isInteger(result?.status) || result.status === 0) { - process.stdout.write(raw); + writeStderr('[Hook] bootstrap: hook produced no output; emitting empty stdout\n'); } } @@ -146,7 +171,7 @@ function spawnNode(rootDir, relPath, raw, args) { CLAUDE_PLUGIN_ROOT: rootDir, ECC_PLUGIN_ROOT: rootDir, }; - return spawnSync(process.execPath, [resolveTarget(rootDir, relPath), ...args], { + const result = spawnSync(process.execPath, [resolveTarget(rootDir, relPath), ...args], { input: raw, encoding: 'utf8', env: hookEnv, @@ -154,6 +179,11 @@ function spawnNode(rootDir, relPath, raw, args) { timeout: 30000, windowsHide: true, }); + // Tag result with the raw input so passthrough() can detect the + // "hook returned raw input as stdout" pattern and suppress it + // (the dominant source of session-transcript bloat). + result.__rawInput = raw; + return result; } // spawnShell is not used by any hook in the shipped hooks.json configuration @@ -190,7 +220,7 @@ function spawnShell(rootDir, relPath, raw, args) { stderr: '[Hook] .sh script requested but no bash binary found on Windows; skipping\n', }; } - return spawnSync(bash, [scriptPath, ...args], { + const bashResult = spawnSync(bash, [scriptPath, ...args], { input: raw, encoding: 'utf8', env: hookEnv, @@ -198,6 +228,8 @@ function spawnShell(rootDir, relPath, raw, args) { timeout: 30000, windowsHide: true, }); + bashResult.__rawInput = raw; + return bashResult; } const shellArgs = isPs @@ -206,7 +238,7 @@ function spawnShell(rootDir, relPath, raw, args) { ? ['-NoProfile', '-NonInteractive', '-ExecutionPolicy', 'Bypass', '-File', scriptPath, ...args] : [scriptPath, ...args]; - return spawnSync(shell, shellArgs, { + const result = spawnSync(shell, shellArgs, { input: raw, encoding: 'utf8', env: hookEnv, @@ -214,6 +246,8 @@ function spawnShell(rootDir, relPath, raw, args) { timeout: 30000, windowsHide: true, }); + result.__rawInput = raw; + return result; } function main() { @@ -224,7 +258,9 @@ function main() { ); if (!mode || !relPath || !rootDir) { - process.stdout.write(raw); + writeStderr( + '[Hook] bootstrap: missing required args (mode/relPath/rootDir); emitting empty stdout\n' + ); process.exit(0); } @@ -235,17 +271,15 @@ function main() { } else if (mode === 'shell') { result = spawnShell(rootDir, relPath, raw, args); } else { - writeStderr(`[Hook] unknown bootstrap mode: ${mode}\n`); - process.stdout.write(raw); + writeStderr(`[Hook] unknown bootstrap mode: ${mode}; emitting empty stdout\n`); process.exit(0); } } catch (error) { - writeStderr(`[Hook] bootstrap resolution failed: ${error.message}\n`); - process.stdout.write(raw); + writeStderr(`[Hook] bootstrap resolution failed: ${error.message}; emitting empty stdout\n`); process.exit(0); } - passthrough(raw, result); + passthrough(result); writeStderr(result.stderr); if (result.error || result.signal || result.status === null) { @@ -275,4 +309,4 @@ if (require.main === module || require.main === undefined) { module.exports = { main, normalizePluginRootForPlatform, -}; +}; \ No newline at end of file diff --git a/tests/hooks/plugin-hook-bootstrap-no-echo.test.js b/tests/hooks/plugin-hook-bootstrap-no-echo.test.js new file mode 100644 index 000000000..67662340b --- /dev/null +++ b/tests/hooks/plugin-hook-bootstrap-no-echo.test.js @@ -0,0 +1,256 @@ +/** + * Regression tests for plugin-hook-bootstrap.js raw-echo bloat. + * + * Before the fix, every fallthrough path in plugin-hook-bootstrap.js + * (the actual entry point used by ECC plugin hooks, NOT run-with-flags.js) + * echoed the full raw hook input JSON to stdout. For a typical + * PostToolUse:Edit payload this is 10-130 KB of tool_input + tool_response + * per tool call. The harness then wrote that stdout into the session + * transcript as a hook_success attachment, ballooning 51 transcripts + * to a combined 1.06 GB (89% of which was raw-echo bloat). + * + * The fix removes the 4 echo-raw sites in plugin-hook-bootstrap.js: + * - line 137: missing mode/relPath/rootDir + * - line 149: unknown mode + * - line 154: catch on spawn failure + * - line 31: passthrough() default when hook outputs nothing + * + * For each, we emit empty stdout and a stderr explanation. The harness + * then falls back to the tool_use's original result, mirroring the + * pattern already shipped in #2240 (bash-hook-dispatcher.js) and #2227 + * (run-with-flags.js truncation path). + * + * Related: + * - #2222 / #2227 — fixed the *truncated* path of run-with-flags.js + * - #2239 / #2240 — fixed the same bug in bash-hook-dispatcher.js + * - #1575 — "token limit so fast" (symptom caused in part by this) + * + * Fixtures live under `/tmp/ecc-pr2380-fixtures/` (per reviewer feedback + * on #2380 — keep temp fixture files out of the live scripts/hooks/ tree). + */ + +'use strict'; + +const assert = require('assert'); +const fs = require('fs'); +const path = require('path'); +const { spawnSync } = require('child_process'); + +const repoRoot = path.join(__dirname, '..', '..'); +const bootstrap = path.join(repoRoot, 'scripts', 'hooks', 'plugin-hook-bootstrap.js'); +const FIXTURE_DIR = '/tmp/ecc-pr2380-fixtures'; + +function ensureFixtureDir() { + fs.mkdirSync(FIXTURE_DIR, { recursive: true }); +} + +function test(name, fn) { + try { + fn(); + console.log(` ✓ ${name}`); + return true; + } catch (error) { + console.log(` ✗ ${name}`); + console.log(` Error: ${error.message}`); + return false; + } +} + +function runBootstrap(args, input, env) { + return spawnSync('node', [bootstrap, ...args], { + input, + encoding: 'utf8', + cwd: repoRoot, + env: { ...process.env, ...(env || {}) }, + timeout: 30000, + maxBuffer: 16 * 1024 * 1024, + stdio: ['pipe', 'pipe', 'pipe'] + }); +} + +function realisticPostToolUseEditPayload() { + return JSON.stringify({ + session_id: 'test-session', + transcript_path: '/tmp/test.jsonl', + cwd: '/tmp', + permission_mode: 'auto', + hook_event_name: 'PostToolUse', + tool_name: 'Edit', + tool_input: { + file_path: '/tmp/example.ts', + old_string: 'a'.repeat(200), + new_string: 'b'.repeat(200) + }, + tool_response: { filePath: '/tmp/example.ts', diff: 'c'.repeat(100 * 1024) }, + tool_use_id: 'call_test_1' + }); +} + +console.log('\nplugin-hook-bootstrap raw-echo (no bloat) tests:'); + +ensureFixtureDir(); + +let passed = 0; +let failed = 0; + +// --- Bug site #1: line 137 (missing args) --- +if ( + test('fallthrough 1: missing mode emits empty stdout (no raw echo)', () => { + const payload = realisticPostToolUseEditPayload(); + const result = runBootstrap([], payload, { + CLAUDE_PLUGIN_ROOT: repoRoot + }); + assert.strictEqual(result.status, 0); + assert.strictEqual(result.stdout, '', 'missing-args path must NOT echo raw input (was ' + result.stdout.length + ' bytes)'); + }) +) + passed++; +else failed++; + +// --- Bug site #2: line 149 (unknown mode) --- +if ( + test('fallthrough 2: unknown mode emits empty stdout (no raw echo)', () => { + const payload = realisticPostToolUseEditPayload(); + const result = runBootstrap(['bogus-mode', path.join(FIXTURE_DIR, 'noop-hook-fixture.js')], payload, { + CLAUDE_PLUGIN_ROOT: repoRoot + }); + assert.strictEqual(result.status, 0); + assert.strictEqual(result.stdout, '', 'unknown-mode path must NOT echo raw input (was ' + result.stdout.length + ' bytes)'); + assert.match(result.stderr, /unknown bootstrap mode/); + }) +) + passed++; +else failed++; + +// --- Bug site #3: line 31 (passthrough default) — THE CORE BUG --- +// This is what fires on EVERY successful hook call where the hook script +// itself didn't write to stdout. The default `passthrough` behavior is +// to echo raw input — which is the bulk of the bloat. +if ( + test('fallthrough 3: silent hook does NOT echo raw input (the core bug)', () => { + const payload = realisticPostToolUseEditPayload(); + // A no-op node hook that reads stdin and exits silently. Lives in + // /tmp/ecc-pr2380-fixtures/ — NOT in the live scripts/hooks/ tree. + const noopHookPath = path.join(FIXTURE_DIR, 'noop-hook-fixture.js'); + fs.writeFileSync(noopHookPath, "process.stdin.resume(); process.stdin.on('end', () => process.exit(0));"); + try { + const result = runBootstrap(['node', noopHookPath], payload, { + CLAUDE_PLUGIN_ROOT: repoRoot + }); + assert.strictEqual(result.status, 0); + assert.strictEqual(result.stdout, '', 'silent hook must NOT echo raw input (was ' + result.stdout.length + ' bytes)'); + } finally { + fs.unlinkSync(noopHookPath); + } + }) +) + passed++; +else failed++; + +// --- Bug site #4: tool_response leak guard (the user-visible symptom) --- +if ( + test('fallthrough 4: tool_response contents never leak into stdout', () => { + const marker = 'PAYLOAD_MARKER_DO_NOT_LEAK_X9Z42'; + const payload = JSON.stringify({ + session_id: 'test', + hook_event_name: 'PostToolUse', + tool_name: 'Edit', + tool_input: { file_path: '/tmp/x', old_string: 'A', new_string: 'B' }, + tool_response: { filePath: '/tmp/x', leaked: marker, diff: 'x'.repeat(50 * 1024) } + }); + const result = runBootstrap([], payload, { + CLAUDE_PLUGIN_ROOT: repoRoot + }); + assert.strictEqual(result.status, 0); + assert.ok(!result.stdout.includes(marker), 'tool_response contents must not appear in stdout'); + }) +) + passed++; +else failed++; + +// --- GREEN-side: behavior we want preserved --- +if ( + test('GREEN: hook that outputs JSON is passed through unchanged', () => { + // When the hook legitimately produces output (e.g., PreToolUse + // additionalContext), we must preserve that output verbatim. + const fixturePath = path.join(FIXTURE_DIR, 'echo-fixture.js'); + const expectedOutput = '{"hookSpecificOutput":{"permissionDecision":"allow"}}\n'; + fs.writeFileSync( + fixturePath, + "process.stdin.resume(); process.stdin.on('end', () => { process.stdout.write('" + expectedOutput.replace(/\n/g, '\\n') + "'); process.exit(0); });" + ); + try { + const payload = realisticPostToolUseEditPayload(); + const result = runBootstrap(['node', fixturePath], payload, { + CLAUDE_PLUGIN_ROOT: repoRoot + }); + assert.strictEqual(result.status, 0); + assert.ok(result.stdout.length > 0, 'hook that produced output should have non-empty stdout'); + // Must not contain the raw input — only the hook's own output + assert.ok(!result.stdout.includes('tool_response'), 'when hook outputs its own stdout, raw input must not also be echoed'); + } finally { + fs.unlinkSync(fixturePath); + } + }) +) + passed++; +else failed++; + +// --- THE CORE ECC PATTERN: most ECC hooks do `process.stdout.write(run(data))` +// where run(data) returns the raw input unchanged. Bootstrap must detect +// this and emit empty stdout instead of writing raw back. --- +if ( + test('CORE ECC PATTERN: hook returning raw input as stdout is suppressed', () => { + // Simulate the post-edit-accumulator pattern: read stdin, return it + // unchanged via process.stdout.write. This is THE dominant source of + // transcript bloat — 12+ ECC hook scripts use this exact pattern. + const fixturePath = path.join(FIXTURE_DIR, 'passthrough-fixture.js'); + fs.writeFileSync( + fixturePath, + "let d=''; process.stdin.setEncoding('utf8'); process.stdin.on('data', c => d += c); process.stdin.on('end', () => { process.stdout.write(d); process.exit(0); });" + ); + try { + const payload = realisticPostToolUseEditPayload(); + const result = runBootstrap(['node', fixturePath], payload, { + CLAUDE_PLUGIN_ROOT: repoRoot + }); + assert.strictEqual(result.status, 0); + assert.strictEqual(result.stdout, '', 'hook that returned raw input as stdout must be suppressed (was ' + result.stdout.length + ' bytes)'); + assert.match(result.stderr, /returned raw input as stdout/, 'stderr should explain the suppression'); + } finally { + fs.unlinkSync(fixturePath); + } + }) +) + passed++; +else failed++; + +// --- Regression guard: hook with its OWN non-raw output (not equal to raw) +// must still pass through unchanged. --- +if ( + test('hook with its own non-raw output passes through unchanged', () => { + const fixturePath = path.join(FIXTURE_DIR, 'own-output-fixture.js'); + const ownOutput = '{"hookSpecificOutput":{"additionalContext":"hello"}}\n'; + fs.writeFileSync( + fixturePath, + "process.stdin.resume(); process.stdin.on('end', () => { process.stdout.write('" + ownOutput.replace(/\n/g, '\\n').replace(/"/g, '\\"') + "'); process.exit(0); });" + ); + try { + const payload = realisticPostToolUseEditPayload(); + const result = runBootstrap(['node', fixturePath], payload, { + CLAUDE_PLUGIN_ROOT: repoRoot + }); + assert.strictEqual(result.status, 0); + // Should contain the hook's own output, not the raw input + assert.ok(result.stdout.includes('additionalContext'), 'hook own output must be preserved'); + assert.ok(!result.stdout.includes('tool_response'), 'raw input must NOT be echoed when hook has its own output'); + } finally { + fs.unlinkSync(fixturePath); + } + }) +) + passed++; +else failed++; + +console.log('\n ' + passed + ' passed, ' + failed + ' failed\n'); +process.exit(failed > 0 ? 1 : 0); \ No newline at end of file diff --git a/tests/hooks/plugin-hook-bootstrap.test.js b/tests/hooks/plugin-hook-bootstrap.test.js index 694e44004..45011e1ce 100644 --- a/tests/hooks/plugin-hook-bootstrap.test.js +++ b/tests/hooks/plugin-hook-bootstrap.test.js @@ -61,12 +61,14 @@ function runTests() { let passed = 0; let failed = 0; - if (test('passes stdin through when required bootstrap inputs are missing', () => { + if (test('emits empty stdout and stderr warning when required bootstrap inputs are missing', () => { const result = run([], { input: '{"ok":true}' }); assert.strictEqual(result.status, 0); - assert.strictEqual(result.stdout, '{"ok":true}'); - assert.strictEqual(result.stderr, ''); + // Empty stdout (not the raw input) so the harness falls back to the + // tool_use's original result -- prevents session-transcript bloat. + assert.strictEqual(result.stdout, ''); + assert.ok(result.stderr.includes('missing required args')); })) passed++; else failed++; if (test('normalizes Windows Git Bash POSIX drive roots', () => { @@ -143,7 +145,7 @@ process.stdout.write(JSON.stringify({ } })) passed++; else failed++; - if (test('node mode passes original stdin when child exits cleanly without stdout', () => { + if (test('node mode emits empty stdout when child exits cleanly without stdout', () => { const root = createTempDir(); try { writeFile(root, path.join('scripts', 'silent.js'), 'process.exit(0);\n'); @@ -154,7 +156,10 @@ process.stdout.write(JSON.stringify({ }); assert.strictEqual(result.status, 0); - assert.strictEqual(result.stdout, 'raw-input'); + // Empty stdout (not the raw input) -- the dominant source of + // session-transcript bloat pre-fix. + assert.strictEqual(result.stdout, ''); + assert.ok(result.stderr.includes('emitting empty stdout')); } finally { cleanup(root); } @@ -225,7 +230,7 @@ process.exit(7); } })) passed++; else failed++; - if (test('shell mode fails open when no shell runtime is available', () => { + if (test('shell mode fails open with empty stdout when no shell runtime is available', () => { const root = createTempDir(); try { writeFile(root, path.join('scripts', 'hook.sh'), 'printf unreachable\n'); @@ -237,14 +242,16 @@ process.exit(7); }); assert.strictEqual(result.status, 0); - assert.strictEqual(result.stdout, 'raw-input'); + // Empty stdout (not the raw input) so the harness falls back to the + // tool_use's original result. + assert.strictEqual(result.stdout, ''); assert.ok(result.stderr.includes('shell runtime unavailable')); } finally { cleanup(root); } })) passed++; else failed++; - if (test('rejects target paths that escape the plugin root', () => { + if (test('rejects target paths that escape the plugin root with empty stdout', () => { const root = createTempDir(); try { const result = run(['node', path.join('..', 'outside.js')], { @@ -253,14 +260,16 @@ process.exit(7); }); assert.strictEqual(result.status, 0); - assert.strictEqual(result.stdout, 'raw-input'); + // Empty stdout (not the raw input) -- the resolver throws, fallthrough + // path emits empty + stderr explanation. + assert.strictEqual(result.stdout, ''); assert.ok(result.stderr.includes('Path traversal rejected')); } finally { cleanup(root); } })) passed++; else failed++; - if (test('unknown mode fails open with stderr warning', () => { + if (test('unknown mode fails open with empty stdout and stderr warning', () => { const root = createTempDir(); try { const result = run(['python', 'hook.py'], { @@ -269,7 +278,9 @@ process.exit(7); }); assert.strictEqual(result.status, 0); - assert.strictEqual(result.stdout, 'raw-input'); + // Empty stdout (not the raw input) -- unknown mode fallthrough path + // emits empty + stderr explanation. + assert.strictEqual(result.stdout, ''); assert.ok(result.stderr.includes('unknown bootstrap mode: python')); } finally { cleanup(root); @@ -375,7 +386,7 @@ process.exit(7); }); assert.strictEqual(result.status, 0); - assert.strictEqual(result.stdout, 'raw-input'); + assert.strictEqual(result.stdout, ''); assert.ok( result.stderr.includes('no bash binary found') || result.stderr.includes('shell runtime unavailable'), @@ -391,4 +402,4 @@ process.exit(7); process.exit(failed > 0 ? 1 : 0); } -runTests(); +runTests(); \ No newline at end of file