Two-layer detection: grep finds candidates, agent reasons about context
Emit FIX_RACE, FIX_DEADLOCK, or CONTROL_ASYNC_SIDE_EFFECT
Calculate compliance score (X/10)
Inputs
MANDATORY READ: Load references/audit_worker_core_contract.md.
Tool policy: follow host AGENTS.md MCP preferences; load references/mcp_tool_preferences.md and references/mcp_integration_patterns.md only when host policy is absent or MCP behavior is unclear.
Use hex-graph first when dataflow or call-path analysis materially improves concurrency findings. Use hex-line first for local code reads when available. If MCP is unavailable, unsupported, or not indexed, continue with built-in Read/Grep/Glob/Bash and state the fallback in the report.
Workflow
Detection policy: use two-layer detection (candidate scan, then context verification); load references/two_layer_detection.md only when the verification method is ambiguous.
Parse context -- extract tech_stack, language, output_dir from contextStore
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
Collect confirmed findings with severity, location, effort, recommendation
Calculate score per references/audit_scoring.md
Write Report -- build in memory, write to {output_dir}/ln-628--global.md (atomic single Write)
Return Summary
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:
Language
Pattern
Grep
JS/TS
Read-modify-write across await
\w+\s*[+\-*/]?=\s*.*await (e.g., result += await something)
JS/TS
Check-then-initialize race
if\s*\(!?\w+\) followed by \w+\s*=\s*await in same block
Python
Read-modify-write across await
\w+\s*[+\-*/]?=\s*await inside async def
Python
Shared module-level state in async
Module-level \w+\s*= + modified inside async def
All
Shared 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:
Language
Pattern
Grep
Go
Map access without mutex
map\[.*\].*= in struct without sync.Mutex or sync.RWMutex
Go
Variable captured by goroutine
go func + variable from outer scope modified
Python
Global modified in threads
global\s+\w+ in function + threading.Thread in same file
Java
HashMap shared between threads
HashMap + Thread|Executor|Runnable in same class without synchronized|ConcurrentHashMap
Rust
Rc in multi-thread context
Rc<RefCell + thread::spawn|tokio::spawn in same file
Node.js
Worker Threads shared state
workerData|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:
Language
Check
Use
Grep
Python
os.path.exists()
open()
os\.path\.exists\( near open\( on same variable
Python
os.access()
os.open()
os\.access\( near os\.open\(|open\(
Node.js
fs.existsSync()
fs.readFileSync()
existsSync\( near readFileSync\(|readFile\(
Node.js
fs.accessSync()
fs.openSync()
accessSync\( near openSync\(
Go
os.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:
Language
Pattern
Grep
Python
Nested locks
with\s+\w+_lock: (multiline: two different locks nested)
Python
Lock in loop
for.*: with \.acquire\(\) inside loop body
Python
Lock + external call
\.acquire\(\) followed by await|requests\.|urllib before release
Go
Missing defer unlock
\.Lock\(\) without defer.*\.Unlock\(\) on next line
Go
Nested locks
Two \.Lock\(\) calls in same function without intervening \.Unlock\(\)
Java
Nested synchronized
synchronized\s*\( (multiline: nested blocks with different monitors)
JS
Async mutex nesting
await\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)
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:
Pattern
Risk
Grep
Shared memory without sync
Data corruption
SharedArrayBuffer|SharedMemory|shm_open|mmap without Atomics|Mutex|Lock nearby
IPC without coordination
Message ordering
process\.send|parentPort\.postMessage in concurrent loops
Concurrent file append
Interleaved writes
Multiple 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:
Pattern
Risk
Grep
Clipboard dual access
OSC 52 + native clipboard in same flow
osc52|\\x1b\\]52 AND clipboard|SetClipboardData|pbcopy|xclip in same file
Subprocess + shared file
Parent and child write same file
spawn|exec|Popen + writeFile|open.*"w" on same path
Layer 2 -- This check relies on reasoning more than any other:
Build Resource Inventory:
Resource
Exclusive?
Accessor 1
Accessor 2
Sync present?
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?
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.
Write JSON summary per references/audit_summary_contract.md. In managed mode the caller passes both runId and summaryArtifactPath; in standalone mode the worker generates its own run-scoped artifact path per shared contract.
Write report to {output_dir}/ln-628--global.md with category: "Concurrency" and checks: async_races, thread_safety, toctou, deadlock_potential, blocking_io, resource_contention, cross_process_races.
Return summary per references/audit_summary_contract.md.
When summaryArtifactPath is absent, write the standalone runtime summary under .hex-skills/runtime-artifacts/runs/{run_id}/evaluation-worker/{worker}--{identifier}.json and optionally echo the same summary in structured output.
