Skip to content

CHANGELOG.md

Latest commit

 

History

History
341 lines (276 loc) · 16.7 KB

CHANGELOG.md

File metadata and controls

341 lines (276 loc) · 16.7 KB

Changelog

All notable changes to Drako are documented here.

Format: Keep a Changelog

[3.0.0] - 2026-04-10

Changed

  • Unified versioning: SDK, backend, public repo, and MCP manifests now share version 3.0.0. All version references synchronized across pyproject.toml, __init__.py, server.json, and OpenAPI spec.
  • License classifier corrected to match BUSL-1.1 license field

Added

  • drako scan --simple — indie-developer-friendly output mode. Filters findings to a curated whitelist of 12 high-signal rules, renders plain text with severity markers (capped at 10 findings), and emits framework-aware remediation snippets for CrewAI, LangChain, LangGraph, AutoGen, and direct API integrations. Mutually exclusive with --format sarif and --details.
  • drako nocode — governance scanner for no-code agent workflows. Parses n8n and Flowise exports into a normalized graph, runs BFS taint propagation from user-input nodes, and evaluates 10 rules (NC-001 → NC-010) covering unsanitized input, code execution, credential exposure, missing human-in-the-loop, and more. Ships drako nocode scan and drako nocode bom with text/JSON/markdown output, reusing the existing scoring engine for coherence with drako scan.
  • drako[comply] — optional extra for EU AI Act documentation generation. Runs a 30-question interactive assessment, merges answers with the latest scan cache and BOM, and generates DOCX and PDF drafts for Articles 9 (Risk Management), 11 (Technical Documentation), 12 (Record-Keeping), and 14 (Human Oversight). Every generated document carries a hardcoded legal disclaimer stating that the output is a draft requiring qualified legal review. Install with pip install drako[comply].
  • Desktop Agent Scanning improvements and stabilization
  • TypeScript scanner with Tree-sitter (17 dedicated rules)
  • 97 total governance rules (up from 25 in v1.0.1)
  • Determinism scoring, benchmark system, baseline scanning
  • Proxy mode, policy templates, Helm chart
  • govern() universal wrapper for runtime compliance enforcement

[2.5.0] - 2026-04-01

Added

  • Desktop Agent Scanning — drako desktop scan/bom/govern commands
    • Auto-discovery of MCP servers across 8 clients: Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Codex CLI, Gemini CLI, Kiro
    • Cross-platform config path detection (macOS, Linux, Windows)
    • 8 MCP-specific security rules (MCP-001 → MCP-008): unrestricted filesystem, shell capability, untrusted source, unrestricted network, plaintext credentials, unencrypted transport, elevated privileges, compound capabilities
    • drako desktop govern — scan + activate runtime proxy to intercept MCP traffic
    • Desktop Governance Score separate from project scoring
    • Output formats: text, JSON, SARIF, markdown
  • RFC 3161 timestamping on audit trail — externally verifiable by EU AI Act auditors
    • TSA redundancy: FreeTSA primary, DFN backup
    • Offline verification via embedded certificate chain
    • tsa_token column on audit_logs table
    • GET /api/v1/audit/verify/{entry_id} endpoint
  • Threat Intel Bootstrap — automated daily sync from AlienVault OTX and abuse.ch
    • Quality gate: rejects pulses with <10 subscribers or zero indicators
    • Maps external threats to DRAKO-ABSS format with external: true metadata
    • GitHub Action cron: daily 6 AM UTC
  • Scanner Limitations Documentation — docs/scanner-limitations.md
    • Transparent disclosure of 5 known gaps: interprocedural taint, TS type resolution, semantic injection, in-process enforcement, cross-agent taint
    • Roadmap issue labels on GitHub for each gap

Fixed

  • MCP server Dockerfile: corrected CMD to use drako serve instead of broken module path
  • drako/mcp/__main__.py added to enable python -m drako.mcp
  • server.json version sync with SDK version

[2.4.2] - 2026-03-30

Fixed

  • Windows encoding crash — drako push and other commands no longer crash on cp1252 with box-drawing characters
    • Systemic fix via configure_output() at CLI startup
    • errors='replace' fallback for non-UTF-8 terminals
  • drako simulate now reads API key from .drako.yaml (was ignoring api_key_env)
  • drako validate auto-detects .drako.yaml in current directory when argument omitted
  • drako scan --threshold flag added (was only --threshold-det)
  • drako scan --benchmark panel now renders (was silent)
  • drako scan --share texts include getdrako.com URL (was only pip install)
  • drako init correctly validates API key against backend (was "Could not reach backend" on valid keys)
  • Backend upgrade URLs — all references to useagentmesh.com replaced with getdrako.com in 403 responses and feature-gating messages

Changed

  • Command ordering in drako --help — journey-linear (scan → init → push) instead of alphabetical
  • Quickstart hint added to drako help output
  • Post-scan CTA no longer suggests pip install drako (user already has it)
  • First-scan output truncated to top 10 findings with ... and X more footer
  • Benchmark panel labeled "benchmark dataset" instead of "100 scanned projects"

[2.4.1] - 2026-03-29

Added

  • SSO via WorkOS (Enterprise tier) — Okta, Azure AD, SAML, Google Workspace
    • GET /api/v1/auth/sso/authorize?email=X
    • GET /api/v1/auth/sso/callback
    • POST /api/v1/settings/sso/enable (owner only, enterprise plan)
    • CSRF state tokens via Redis (10min TTL, one-time use) with HMAC fallback
    • Auto-provisioning of users matching tenant's sso_domain
    • Password login blocked server-side for SSO-enforced domains
    • Domain normalization (case-insensitive comparison)
  • 28 SSO tests (CSRF state, password blocking, domain validation, enable/disable flows)

[2.4.0] - 2026-03-28

Added

  • Production Dashboard — full observability UI at getdrako.com/dashboard
    • Stats grid with MetricCards (audit entries, agents, policy blocks, trust score)
    • Governance Score Trend (TimeSeriesChart)
    • Tool Health Grid (circuit breaker status per tool)
    • Recent Activity feed with 30s auto-refresh
    • Governance Roadmap component (conversion hook with tier-tagged action items)
  • Observability page — 4 tabs (Overview, Metrics, Violations, Alerts)
    • P50/P95/P99 latency charts
    • Bottleneck detection
    • Cost breakdown by model (DonutChart)
    • Violation heatmap (7 days × severity)
  • FinOps page — cost tracking, budget burn-down, model/agent breakdown
  • Settings page — config viewer, snapshot history, feature status
  • recharts integration for all visualizations
  • PlanGate component for Pro/Enterprise feature gating

[2.3.1] - 2026-03-26

Fixed

  • 10 integration test issues discovered in E2E testing
  • govern() module/function shadowing — renamed govern.py → governance_wrapper.py to eliminate namespace collision
  • Scoring recalibration — reduced per-finding deductions (CRITICAL -10, HIGH -5, MEDIUM -2, LOW -1) to prevent score floor at 0
  • DET-001 now detects missing temperature on CrewAI Agent() constructors
  • LangGraph tools detected in BOM via ToolNode([...]) and tools=[...] patterns
  • VCR-001 extracts vendors from ChatOpenAI(), ChatAnthropic(), ChatGoogleGenerativeAI() constructors
  • Global governance rules reclassified — ODD-001, MAG-001, ID-001, HOOK-001, CV-001, FIN-001, RES-001, CI-001, MULTI-004 now finding_type="recommendation" (do not affect score)
  • validate command UnicodeEncodeError on Windows (systemic fix via CAN_UNICODE pattern)
  • drako bom --format text exit code (now 0 even with 0 agents found)
  • Telemetry endpoint no longer requires auth (anonymous events)
  • Config push API schema aligned (flat payload accepted)

[2.3.0] - 2026-03-25

Added

  • TypeScript Scanner — Tree-sitter based TS/JS scanning alongside Python
    • Framework detection: LangChain.js, Vercel AI SDK, Mastra, AutoGen.js
    • 17 TypeScript rules covering Security (10), Governance (3), Determinism (2), Compliance (1), Operational (1)
    • TS BOM extraction: agents, tools, models, prompts from .ts/.tsx/.js/.jsx/.mts/.mjs files
    • Mixed Python+TypeScript projects supported with merged scoring
    • Optional dependency: pip install drako[typescript]
    • 15 TypeScript rule fixtures with vulnerable/safe pairs
    • 34 TypeScript tests
  • Clean project output — projects without AI agent components show friendly message instead of generic findings
  • Graceful degradation — TS files detected without tree-sitter installed show install hint, Python scan continues

Changed

  • Policy count: 97 total (80 Python + 17 TypeScript)
  • Rule count badge updated across README, docs, and scoring section

[2.2.2] - 2026-03-24

Fixed

  • Re-release of 2.2.1 (PyPI upload conflict resolved)

[2.2.1] - 2026-03-23

Fixed

  • Framework detection for LangGraph/AutoGen in subdirectories (parent directory walk-up)
  • AutoGen new package names support (autogen_agentchat, autogen_core, ag2)
  • Jupyter notebook .ipynb scanning with IPython magic line stripping
  • CrewAI @CrewBase pattern: agent extraction from YAML config files
  • Unconditional exit code 1 on critical findings removed (only with --threshold-det)
  • MULTI-002 false positive on constructor calls (Crew, Task, Agent)

Added

  • Finding type classification: vulnerability vs recommendation
  • Terminal report split into FINDINGS and RECOMMENDATIONS sections
  • Recommendations excluded from governance score
  • finding_type field in JSON and SARIF output formats

[2.2.0] - 2026-03-22

Added

  • Determinism Score — second score (0-100) measuring agent reliability (unset temperatures, missing timeouts, no retries)
  • Execution path reachability analysis on findings — unreachable tools are dimmed
  • 25 DRAKO-ABSS security advisories (OWASP Top 10 for LLMs, MITRE ATLAS, framework CVEs)
  • Vendor Concentration Risk rules (VCR-001, VCR-002, VCR-003)
  • Framework-specific security rules (FW-AUTOGEN-001/002, FW-CREWAI-001/002/003, FW-LANGGRAPH-001/002, FW-PYDANTIC-001, FW-SK-001/002)
  • Detection rules (DET-001 → DET-007) for determinism scoring
  • Benchmark comparison system (--benchmark flag)
  • Baseline system for incremental scanning (--baseline flag, drako baseline show/reset)
  • Structured impact explanations on all rules (--details flag with impact, attack_scenario, references)
  • GitHub Action for PR governance comments
  • Score card SVG generation (--share with scorecard + badge SVGs)
  • PQL telemetry (anonymous, opt-in)
  • Framework security monitor (6 frameworks, daily)
  • HITL test harness (test_mode(), MockHITLResolver)
  • Policy simulation (drako simulate)
  • Config validation (drako validate)
  • Behavioral versioning
  • Policy templates (fintech, healthcare, eu-ai-act, startup, enterprise)
  • Out-of-process enforcement proxy (drako proxy start/stop/status)
  • Property-based testing (28 Hypothesis tests)
  • Helm chart for Kubernetes deployment
  • Type safety (mypy --strict on new modules)

Changed

  • Renamed from AgentMesh to Drako
  • 80 rules total (up from 55)
  • 647 tests (up from 277)
  • Domain: getdrako.com
  • PyPI: drako (was useagentmesh)

[2.1.0] - 2026-03-18

Added

  • Autopilot Mode — drako init now generates smart defaults from scan results (audit-first). Governance levels: --autopilot (default), --balanced, --strict, --manual
  • drako upgrade command — upgrade governance level in-place (--balanced, --strict)
  • Proxy Mode — out-of-process LLM API governance proxy. Intercepts every LLM call at the network layer with full governance pipeline (ODD, DLP, Magnitude, HITL). drako proxy start/stop/status
  • Policy Templates — 6 industry-specific governance presets: base, fintech, healthcare, eu-ai-act, startup, enterprise. Template inheritance via extends: in config. drako templates list/show
  • Property-Based Testing — 28 Hypothesis tests covering scoring invariants, grade monotonicity, deep merge, ODD/DLP/Magnitude enforcement
  • Helm Chart — Kubernetes deployment for the governance proxy (deploy/helm/drako-proxy/)
  • governance_level config field — autopilot | balanced | strict | custom
  • extends config field — template inheritance (e.g. extends: fintech)
  • test_mode() context manager and MockHITLResolver for testing governed agents in CI
  • Docker Compose file for proxy deployment

Changed

  • drako init default mode is now autopilot (was manual)
  • Config load() now resolves template inheritance automatically

[2.0.1] - 2026-03-17

Fixed

  • Sync __version__ in __init__.py with pyproject.toml — CLI now reports correct version

[2.0.0] - 2026-03-17

Changed

  • Version bump to 2.0.0 — reflects production maturity and comprehensive governance coverage
  • CI: add FORCE_JAVASCRIPT_ACTIONS_TO_NODE24 to publish workflow for Node.js 24 migration

Fixed

  • Publish workflow: suppress Node.js 20 deprecation warnings on GitHub Actions

[0.2.0] - 2026-03-16

Added

  • 26 new governance rules (60 total, up from 34)
  • New rule categories: Multi-Agent (MULTI-001→004), Hooks (HOOK-001→003), Context Versioning (CV-001→002), FinOps (FIN-001→003), Resilience (RES-001→002), A2A (A2A-001→003), Best Practices (BP-001→005)
  • drako scan --share viral sharing with pre-written posts for LinkedIn, X, Bluesky
  • drako history, diff, rollback commands for policy version management
  • govern() universal one-liner — auto-detects CrewAI, LangGraph, AutoGen
  • Per-tool interception: every tool call evaluated through the enforcement pipeline
  • Prompt Injection Detection (5 pattern categories, bidirectional)
  • Circuit Breaker per-tool (CLOSED/OPEN/HALF_OPEN state machine)
  • Human-in-the-Loop checkpoints (configurable triggers, webhook notifications)
  • Intent Fingerprinting (SHA-256 + Ed25519 two-gate verification)
  • Programmable Hooks (pre/post action, Python scripts + YAML conditions)
  • Context Versioning (immutable snapshots on every push)
  • Agentic FinOps (cost tracking, model routing, semantic cache, budgets)
  • Deterministic Fallback (failover to code/agent/human/retry queue)
  • Secure A2A Gateway (DID auth, channel policies, prompt worm prevention)
  • Multi-Agent Topology (directed graph, conflict detection, cascade analysis)
  • Chaos Engineering (7 fault types, governance grading A-F)
  • Collective Intelligence (anonymous IOC sharing, quality scoring)
  • Observability platform (session traces, latency metrics, drift detection)
  • Alerting engine (configurable rules, Slack/email/PagerDuty)
  • OTEL & SIEM export (OpenTelemetry, STIX 2.1, CEF)
  • MCP local server for Claude Desktop integration
  • AutoGen middleware support
  • drako serve command for local MCP server
  • drako verify command for configuration validation

Changed

  • govern() now wraps individual tools (was session-level only)
  • Config-as-code: full governance configurable from .drako.yaml
  • drako push translates YAML to enforcement engine
  • Scoring updated for 13 rule categories (was 4)
  • Policy engine evaluates all 60 rules in <2ms (P50)

[0.1.9] - 2026-03-10

Fixed

  • CLI UX improvements across all commands
  • API quota handling — graceful fallback on rate limit

[0.1.8] - 2026-03-06

Added

  • Compact output mode (drako scan --compact)
  • Autofix dry-run (drako fix --dry-run)
  • --project flag to specify project root
  • Governance badge URL generation for CI/CD
  • BUSL-1.1 license version update

Changed

  • Scan output refined for readability

[0.1.5] - 2026-03-04

Added

  • Operational Design Domain (ODD) scan rules: ODD-001 to ODD-004
  • Magnitude limit scan rules: MAG-001 to MAG-003
  • Agent Identity scan rules: ID-001 to ID-003
  • 10 new rules total (34 total, up from 24)

[0.1.4] - 2026-03-03

Added

  • API key YAML fallback — api_key_env field in .drako.yaml
  • Cross-platform environment variable hint in scan output

Fixed

  • drako scan and drako init command edge cases

[0.1.2] - 2026-02-28

Fixed

  • Windows MAX_PATH crash in directory scanner
  • Broken benchmarks link (sdk/ prefix removed)
  • Corrected paths and BSL-1.1 license note for contributors

Added

  • Professional upgrade CTA in scan output

[0.1.1] - 2026-02-27

Added

  • Initial release of Drako SDK
  • 24 governance rules across 4 categories (Security, Governance, Compliance, Best Practices)
  • Agent BOM (Bill of Materials) — AST-based discovery of agents, tools, models, prompts
  • Framework support: CrewAI, LangGraph, AutoGen, LangChain, LlamaIndex, PydanticAI
  • CLI commands: drako scan, drako init
  • Output formats: Terminal (Rich), JSON, SARIF 2.1.0
  • Governance scoring: A-F grades (0-100 scale)
  • GitHub Code Scanning integration via SARIF
  • Pre-commit hook support
  • Performance benchmarks (<2ms full scan P50)
  • PyPI package: drako