本文へ移動
cccskills
無料GitHub で公開

ln-628-concurrency-auditor

Checks async races, thread safety, TOCTOU, deadlocks, blocking I/O, resource contention, cross-process races. Two-layer detection: grep + agent reasoning.

インストール方法を見る

含まれるファイル(1)

  • SKILL.md13.7 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Paths: File paths (shared/, references/, ../ln-*) are relative to skills repo root. If not found at CWD, locate this SKILL.md directory and go up one level for repo root.

Concurrency Auditor (L3 Worker)

Specialized worker auditing concurrency, async patterns, and cross-process resource access.

Purpose & Scope

  • Worker in ln-620 coordinator pipeline
  • Audit concurrency (Category 11: High Priority)
  • 7 checks: async races, thread safety, TOCTOU, deadlocks, blocking I/O, resource contention, cross-process races
  • Two-layer detection: grep finds candidates, agent reasons about context
  • Calculate compliance score (X/10)

Inputs (from Coordinator)

MANDATORY READ: Load shared/references/audit_worker_core_contract.md.

Receives contextStore with: tech_stack, best_practices, codebase_root, output_dir.

Workflow

MANDATORY READ: Load shared/references/two_layer_detection.md for detection methodology.

  1. Parse context — extract tech_stack, language, output_dir from contextStore
  2. Per check (1-7):
    • Layer 1: Grep/Glob scan to find candidates
    • Layer 2: Read 20-50 lines around each candidate. Apply check-specific critical questions. Classify: confirmed / false positive / needs-context
  3. Collect confirmed findings with severity, location, effort, recommendation
  4. Calculate score per shared/references/audit_scoring.md
  5. Write Report — build in memory, write to {output_dir}/628-concurrency.md (atomic single Write)
  6. Return Summary to coordinator

Audit Rules

Unified severity escalation: For ALL checks — if finding affects payment/auth/financial code → escalate to CRITICAL regardless of other factors.

1. Async/Event-Loop Races (CWE-362)

What: Shared state corrupted across await/yield boundaries in single-threaded async code.

Layer 1 — Grep patterns:

LanguagePatternGrep
JS/TSRead-modify-write across await\w+\s*[+\-*/]?=\s*.*await (e.g., result += await something)
JS/TSCheck-then-initialize raceif\s*\(!?\w+\) followed by \w+\s*=\s*await in same block
PythonRead-modify-write across await\w+\s*[+\-*/]?=\s*await inside async def
PythonShared module-level state in asyncModule-level \w+\s*= + modified inside async def
AllShared cache without lock`.set(

Layer 2 — Critical questions:

  • Is the variable shared (module/global scope) or local?
  • Can two async tasks interleave at this await point?
  • Is there a lock/mutex/semaphore guarding the access?

Severity: CRITICAL (payment/auth) | HIGH (user-facing) | MEDIUM (background)

Safe pattern exclusions: Local variables, const declarations, single-use await (no interleaving possible).

Effort: M

2. Thread/Goroutine Safety (CWE-366)

What: Shared mutable state accessed from multiple threads/goroutines without synchronization.

Layer 1 — Grep patterns:

LanguagePatternGrep
GoMap access without mutexmap\[.*\].*= in struct without sync.Mutex or sync.RWMutex
GoVariable captured by goroutinego func + variable from outer scope modified
PythonGlobal modified in threadsglobal\s+\w+ in function + threading.Thread in same file
JavaHashMap shared between threadsHashMap + Thread|Executor|Runnable in same class without synchronized|ConcurrentHashMap
RustRc in multi-thread contextRc<RefCell + thread::spawn|tokio::spawn in same file
Node.jsWorker Threads shared stateworkerData|SharedArrayBuffer|parentPort + mutable access without Atomics

Layer 2 — Critical questions:

  • Is this struct/object actually shared between threads? (single-threaded code → FP)
  • Is mutex/lock in embedded struct or imported module? (grep may miss it)
  • Is go func capturing by value (safe) or by reference (unsafe)?

Severity: CRITICAL (payment/auth) | HIGH (data corruption possible) | MEDIUM (internal)

Safe pattern exclusions: Go map in init() or main() before goroutines start. Rust Arc<Mutex<T>> (already safe). Java Collections.synchronizedMap().

Effort: M

3. TOCTOU — Time-of-Check Time-of-Use (CWE-367)

What: Resource state checked, then used, but state can change between check and use.

Layer 1 — Grep patterns:

LanguageCheckUseGrep
Pythonos.path.exists()open()os\.path\.exists\( near open\( on same variable
Pythonos.access()os.open()os\.access\( near os\.open\(|open\(
Node.jsfs.existsSync()fs.readFileSync()existsSync\( near readFileSync\(|readFile\(
Node.jsfs.accessSync()fs.openSync()accessSync\( near openSync\(
Goos.Stat()os.Open()os\.Stat\( near os\.Open\(|os\.Create\(
Java.exists()new FileInputStream\.exists\(\) near new File|FileInputStream|FileOutputStream

Layer 2 — Critical questions:

  • Is the check used for control flow (vulnerable) or just logging (safe)?
  • Is there a lock/retry around the check-then-use sequence?
  • Is the file in a temp directory controlled by the application (lower risk)?
  • Could an attacker substitute the file (symlink attack)?

Severity: CRITICAL (security-sensitive: permissions, auth tokens, configs) | HIGH (user-facing file ops) | MEDIUM (internal/background)

Safe pattern exclusions: Check inside try/catch with retry. Check for logging/metrics only. Check + use wrapped in file lock.

Effort: S-M (replace check-then-use with direct use + error handling)

4. Deadlock Potential (CWE-833)

What: Lock acquisition in inconsistent order, or lock held during blocking operation.

Layer 1 — Grep patterns:

LanguagePatternGrep
PythonNested lockswith\s+\w+_lock: (multiline: two different locks nested)
PythonLock in loopfor.*: with \.acquire\(\) inside loop body
PythonLock + external call\.acquire\(\) followed by await|requests\.|urllib before release
GoMissing defer unlock\.Lock\(\) without defer.*\.Unlock\(\) on next line
GoNested locksTwo \.Lock\(\) calls in same function without intervening \.Unlock\(\)
JavaNested synchronizedsynchronized\s*\( (multiline: nested blocks with different monitors)
JSAsync mutex nestingawait\s+\w+\.acquire\(\) (two different mutexes in same function)

Layer 2 — Critical questions:

  • Are these the same lock (reentrant = OK) or different locks (deadlock risk)?
  • Is the lock ordering consistent across all call sites?
  • Does the external call inside lock have a timeout?

Severity: CRITICAL (payment/auth) | HIGH (app freeze risk)

Safe pattern exclusions: Reentrant locks (same lock acquired twice). Locks with explicit timeout (asyncio.wait_for, tryLock).

Effort: L (lock ordering redesign)

5. Blocking I/O in Async Context (CWE-400)

What: Synchronous blocking calls inside async functions or event loop handlers.

Layer 1 — Grep patterns:

LanguageBlocking CallGrepReplacement
Pythontime.sleep in async deftime\.sleep inside async defawait asyncio.sleep
Pythonrequests.* in async defrequests\.(get|post|put|delete) inside async defhttpx or aiohttp
Pythonopen() in async defopen\( inside async defaiofiles.open
Node.jsfs.readFileSync in asyncfs\.readFileSync|fs\.writeFileSync|fs\.mkdirSyncfs.promises.*
Node.jsexecSync in asyncexecSync|spawnSync in async handlerexec with promises
Node.jsSync crypto in asynccrypto\.pbkdf2Sync|crypto\.scryptSynccrypto.pbkdf2 (callback)

Layer 2 — Critical questions:

  • Is this in a hot path (API handler) or cold path (startup script)?
  • Is the blocking duration significant (>100ms)?
  • Is there a legitimate reason (e.g., sync read of small config at startup)?

Severity: HIGH (blocks event loop/async context) | MEDIUM (minor blocking <100ms)

Safe pattern exclusions: Blocking call in if __name__ == "__main__" (startup). readFileSync in config loading at init time. Sync crypto for small inputs.

Effort: S-M (replace with async alternative)

6. Resource Contention (CWE-362)

What: Multiple concurrent accessors compete for same resource without coordination.

Layer 1 — Grep patterns:

PatternRiskGrep
Shared memory without syncData corruptionSharedArrayBuffer|SharedMemory|shm_open|mmap without Atomics|Mutex|Lock nearby
IPC without coordinationMessage orderingprocess\.send|parentPort\.postMessage in concurrent loops
Concurrent file appendInterleaved writesMultiple appendFile|fs\.write to same path from parallel tasks

Layer 2 — Critical questions:

  • Are multiple writers actually concurrent? (Sequential = safe)
  • Is there OS-level atomicity guarantee? (e.g., O_APPEND for small writes)
  • Is ordering important for correctness?

Severity: HIGH (data corruption) | MEDIUM (ordering issues)

Safe pattern exclusions: Single writer pattern. OS-guaranteed atomic operations (small pipe writes, O_APPEND). Message queues with ordering guarantees.

Effort: M

7. Cross-Process & Invisible Side Effects (CWE-362, CWE-421)

What: Multiple processes or process+OS accessing same exclusive resource, including operations with non-obvious side effects on shared OS resources.

Layer 1 — Grep entry points:

PatternRiskGrep
Clipboard dual accessOSC 52 + native clipboard in same flowosc52|\\x1b\\]52 AND clipboard|SetClipboardData|pbcopy|xclip in same file
Subprocess + shared fileParent and child write same filespawn|exec|Popen + writeFile|open.*"w" on same path
OS exclusive resourceWin32 clipboard, serial port, named pipeOpenClipboard|serial\.Serial|CreateNamedPipe|mkfifo
Terminal escape sequencesstdout triggers terminal OS access\\x1b\\]|\\033\\]|writeOsc|xterm
External clipboard toolsClipboard via spawned processpbcopy|xclip|xsel|clip\.exe

Layer 2 — This check relies on reasoning more than any other:

  1. Build Resource Inventory:

    ResourceExclusive?Accessor 1Accessor 2Sync present?
  2. Trace Timeline:

    t=0ms  operation_A() -> resource_X accessed
    t=?ms  side_effect   -> resource_X accessed by external process
    t=?ms  operation_B() -> resource_X accessed again -> CONFLICT?
    
  3. Critical Questions:

    • Can another process (terminal, OS, child) access this resource simultaneously?
    • Does this operation have invisible side effects on shared OS resources?
    • What happens if the external process is slower/faster than expected?
    • What happens if user triggers this action twice rapidly?

Severity: CRITICAL (two accessors to exclusive OS resource without sync) | HIGH (subprocess + shared file without lock) | HIGH (invisible side effect detected via reasoning)

Safe pattern exclusions: Single accessor. Retry/backoff pattern present. Operations sequenced with explicit delay/await.

Effort: M-L (may require removing redundant access path)

Scoring Algorithm

MANDATORY READ: Load shared/references/audit_worker_core_contract.md and shared/references/audit_scoring.md.

Output Format

MANDATORY READ: Load shared/references/audit_worker_core_contract.md and shared/templates/audit_worker_report_template.md.

Write report to {output_dir}/628-concurrency.md with category: "Concurrency" and checks: async_races, thread_safety, toctou, deadlock_potential, blocking_io, resource_contention, cross_process_races.

Return summary to coordinator:

Report written: docs/project/.audit/ln-620/{YYYY-MM-DD}/628-concurrency.md
Score: X.X/10 | Issues: N (C:N H:N M:N L:N)

Critical Rules

MANDATORY READ: Load shared/references/audit_worker_core_contract.md.

  • Do not auto-fix: Report only — concurrency fixes require careful human review
  • Two-layer detection: Always apply Layer 2 reasoning after Layer 1 grep. Never report raw grep matches without context analysis
  • Language-aware detection: Use language-specific patterns per check
  • Unified CRITICAL escalation: Any finding in payment/auth/financial code = CRITICAL
  • Effort realism: S = <1h, M = 1-4h, L = >4h
  • Exclusions: Skip test files, skip single-threaded CLI tools, skip generated code

Definition of Done

MANDATORY READ: Load shared/references/audit_worker_core_contract.md.

  • contextStore parsed (language, concurrency model, output_dir)
  • All 7 checks completed with two-layer detection:
    • async races, thread safety, TOCTOU, deadlock potential, blocking I/O, resource contention, cross-process races
  • Layer 2 reasoning applied to each candidate (confirmed / FP / needs-context)
  • Findings collected with severity, location, effort, recommendation
  • Score calculated per shared/references/audit_scoring.md
  • Report written to {output_dir}/628-concurrency.md (atomic single Write call)
  • Summary returned to coordinator

Reference Files

  • Two-layer detection methodology: shared/references/two_layer_detection.md
  • Audit output schema: shared/references/audit_output_schema.md

Version: 4.0.0 Last Updated: 2026-03-04

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Design, implement, and audit inclusive digital products using WCAG 2.2 Level AA standards. Use this skill to generate semantic ARIA for Web and accessibility traits for Web and Native platforms (iOS/Android).

日本語の概要は準備中です。原文の説明を表示しています。

tensology/decisionsai172026年10月13日 更新

Whole site or product — a full web accessibility (a11y) audit against WCAG 2.2, following the WCAG-EM methodology. Defines scope, samples representative pages and flows, runs the automated tier (`accessibility-scan`) and the hands-on manual tier (`accessibility-inspect`), and produces one conformance report. Grades each finding by severity and evidence basis, and states per-criterion conformance as pass, fail, or undetermined (needs a human). Use it for 'audit my site for accessibility', 'is this product accessible', 'a11y audit', 'WCAG or Section 508 conformance report', or any multi-page assessment. Assesses; does not fix (use `accessibility-fix`) or diff (use `accessibility-diff`). For a single page use `accessibility-scan`; for hands-on keyboard and screen-reader checks use `accessibility-inspect`.

日本語の概要は準備中です。原文の説明を表示しています。

tensology/decisionsai172026年10月13日 更新

Regression check — diff a live page's web accessibility (a11y) violations against a baseline. By default it compares your uncommitted changes (stash-based); pass `--branch [<name>]` to compare against a branch. Reports the new WCAG violations introduced, the ones fixed, and the count of pre-existing ones. Use it for 'did my change break accessibility', 'what a11y issues did this PR add', or as a CI gate. For a full scan of one page use `accessibility-scan`; for a whole site use `accessibility-audit`.

日本語の概要は準備中です。原文の説明を表示しています。

tensology/decisionsai172026年10月13日 更新

Remediation only — repair web accessibility (a11y) violations against WCAG 2.2 with a baseline, edit, and verify loop. Takes a target (URL, files, directory) or a findings worklist from `accessibility-scan`/`accessibility-inspect`/`accessibility-audit`, applies mechanical fixes as given, leaves TODOs for visual or contextual judgment, and verifies by re-running the baseline check. It only fixes. To find issues use `accessibility-scan` (one page, automated), `accessibility-inspect` (one page, manual), or `accessibility-audit` (whole site, WCAG-EM); to check for regressions use `accessibility-diff`. Use it for 'fix the a11y issues in X', 'make this accessible', 'add missing alt text and labels', 'apply these accessibility fixes', 'remediate these violations'.

日本語の概要は準備中です。原文の説明を表示しています。

tensology/decisionsai172026年10月13日 更新

One page, hands-on manual tier — drive a live page through the web accessibility (a11y) checks a rule engine can't decide: keyboard operation and focus order, screen-reader names, roles and states from the accessibility tree, reflow and zoom, reduced motion, form errors, and target size. Grades each finding by evidence basis (verified / confirm-with-a-human / human-required) and severity, and closes every criterion in a ledger: verified, flagged, not exercised, or N/A. Locates and assesses; does not fix (use `accessibility-fix`). Use it for keyboard testing, focus-order checks, screen-reader or a11y-tree review, reflow and zoom at 200%, or 'is this operable, not just lint-clean'. The automated tier is `accessibility-scan`; `accessibility-audit` runs both across a sampled site.

日本語の概要は準備中です。原文の説明を表示しています。

tensology/decisionsai172026年10月13日 更新

One page, automated tier — run the web accessibility (a11y) rule engine against a live page and locate every violation it can detect mechanically. Pass a URL, a config target name (e.g. `accesslint:accessibility-scan dev`), or nothing to use the default target from `accesslint.config.json`. Ensures a debuggable Chrome, runs the @accesslint/core engine over CDP, and returns a worklist of live-DOM WCAG 2.2 violations, each grounded to its DOM selector and source `file:line`. Locates; doesn't edit. Use it for 'is this page accessible', 'check a11y on this URL', 'find contrast and alt-text issues', or to verify a UI change. For hands-on keyboard and screen-reader checks use `accessibility-inspect`; for a whole site or product use `accessibility-audit`; to diff against uncommitted changes or a branch use `accessibility-diff`.

日本語の概要は準備中です。原文の説明を表示しています。

tensology/decisionsai172026年10月13日 更新

tensology のスキルをすべて見る

このスキルの問題を報告する