fix: safely convert line number string arguments to integers in dataflow agent tools

fix: disable HTTP2 globally in httpx to prevent ConnectionTerminated errors on long scans

fix: retrieve fresh Supabase clients to prevent connection termination on long scans

fix: implement missing _extract_calls_regex fallback method

feat: upgrade scanner with route-middleware modeling, adaptive context triage, and memory scaling

fix: auto-enable DAST whenever runtime_url is provided

fix: auto-enable DAST for DAST-only runtime URL scans

fix: resolve runtime url validation crash and add short file base64 decoding

fix: return 100/100 score for clean scans with 0 findings

fix: add synchronous scanning and optimize for Hugging Face Spaces 16GB memory

feat: upgrade Layer 7 triage validator to use Llama 3.1 70B and enhanced prompts

bugfix: import ScanDomain in layer7_validator.py

tuning: update deprecated nvidia nim models

feat: sync proximity deduplication, max_iterations=10, timeout extensions, and secrets entropy fixes to backend

fix(orchestrator): mark AST-confirmed sanitizer findings as false positive instead of real

fix(triage): route all SAST/LLM taint findings to Joern CPG validator

fix(triage): route injection findings to Joern for path validation and constrain prompt-injection rules

Fix VibeSec pipeline logic bugs, routing, feedback store, and RLS test cases

changed model LLM

feat(scanner): implement dynamic subscription guard, unvalidated DB structural check, and optimize Semgrep chunk execution

perf(sast): reduce batch chunk size to 10 for Render free-tier RAM ceiling

refactor(scanner): optimize imports of Confidence in orchestrator

fix(chains): deduplicate LLM prompt injection chains to prevent duplicates

fix(report): add rate-limit recovery and batch exception handling to Layer 9

fix: resolve logic, security, and memory leak issues in scanner layers

perf(cf): use fast llama model and fix TaintPath instantiation typo

feat(scanner): improve precision in auth, rls, and llm rules, add public repo test script

perf: hybrid AST+LLM taint tracer and tier 0 pre-filter

fix(feedback): resolve rule_fp_rates SQLite table initialization crash

feat(scanner): implement 5 targeted improvements to scanning engine and rules

perf(fallback): fail fast on 429 errors to invoke NVIDIA NIM immediately

fix(security): remove hardcoded Gemini API key fallback from client

feat: implement primary Gemini with NVIDIA NIM fallback and robust regex JSON parsing

Migrate backend to Gemini 3.1 Flash Lite and optimize JS/TS rulesets

refactor(backend): use asyncio and thread pools for background tasks, optimize scanner AST utils and checks

Refactor scanner layers to use AST architecture, fix uvicorn graceful shutdown

fix(sast): resolve false positives in Layer 0 indexer

fix: optimize Layer 7 memory footprints and prevent OOM/timeouts on Render

fix(pipeline): resolve container port binding, checkov execution path, and secrets scanner test path filtration

feat(scanner): sync dynamic secrets validation filters and entropy triaging

fix(reporter): bypass finding enrichment on Render or low-memory containers to avoid timeouts

fix(scanner): exclude .bevel, .continue, .vscode, .idea, .github and minified files from indexing

Fix local scan hangs: parallelize Layer 9 LLM calls, offload report generation, and add SAST tracing logs

Fix backend scan hangs: parallelize Layer 9 report enrichment and offload to loop executor

fix: bypass TruffleHog subprocess on large repositories or low-memory environments to prevent OOM

fix: use filesystem mode for TruffleHog secrets scan on shallow clones to prevent OOM

feat(scanner): transition CORS, debug mode, and prompt exfiltration to AST-based checks

perf: implement generic comment-filtering logic across scanner layers to prevent false positives in commented code and documentation

security: add security headers middleware, upgrade package dependencies, and resolve scanner false positives

perf: increase SAST scan chunk size to 30