Spaces:
Runtime error
Runtime error
Commit History
-
Make mutation policy weights explicit
Lars Talian228ed67 -
Remove hardcoded fallbacks, add snapshot-driven service lifecycle
Aaron Brown- Remove _LEGACY_STOP_DAEMONS, _start_services_legacy, _declared_service_daemons - _stop_services and _capture_service_pids derive daemons from snapshot.services - _start_snapshot_services skips when no service specs (no legacy fallback) - _resolve_target reads topology roles/zones, raises if missing - _select_snapshot raises RuntimeError when no snapshot provided - _get_pending_alerts returns [] instead of synthetic fallback - Renderer auto-populates snapshot.services via generate_service_specs - start.sh supports ServiceSpec-driven startup alongside legacy path - Tests updated: explicit snapshot in all reset() calls, 114 env tests pass - 70 new service spec tests (models, generation, lifecycle, renderer)7fedc25 -
Add episode CLI, synthetic data pipeline, NPC generalization, service manifest
Aaron Brown- Add `openrange episode` command (golden-path replay + interactive REPL) - Add synthetic data generation pipeline with teacher agents and dataset helpers - Generalize NPC actions/traffic scripts to derive targets from SnapshotSpec - Add service_manifest.py for snapshot-declared service lifecycle - Add replay_agent, dataset module, NPC reward coupling tests - Graceful container resolution when Docker is available but no containers running - Skip tests that depend on deleted snapshot fixtures - Remove scripted_agent (replaced by replay_agent)f016eb7 -
Add NPC actions, synthetic data pipeline, CLI enhancements
Aaron BrownAgent-generated supporting changes from architecture review: - NPC action system (actions.py) for structured NPC behaviors - Synthetic data generation pipeline (training/synthetic.py) - CLI improvements and NPC traffic script updates - NPC agent loop enhancements for topology-driven operationfb68239 -
Render admitted snapshots into artifact bundles
Lars Talian4e3f8b1 -
Add managed snapshot runtime behind OpenEnv server
Lars Talian906af9d -
fix: align package with openenv contract
Lars Talianfda4cbc -
Full OpenRange implementation: environment, builder, validator, agents, tests
Aaron Brown +1: Claude Opus 4.6- src/open_range/: complete package with server, builder, validator, client, agents, training, and protocols modules - manifests/: Pydantic schema + 3 enriched YAML manifests (tier 1-3) with realism profiles (company context, tech stack, credential policy, monitoring coverage, trust relationships, operational context) - vulns/catalog.yaml: 23 vulnerability classes across 4 realism layers (application, infrastructure, identity, process) - Builder LLM pipeline: structured JSON output, 18-rule system prompt using all manifest realism layers, template renderer, snapshot store - Validator: 10-check hybrid pipeline (8 mechanical + 2 LLM advisory) - Environment: container-as-constraint design (no artificial command allowlists), Red routes to attacker container, Blue routes to siem container - Rewards: Red (flag, efficiency, stealth, evidence, anti-hallucination) and Blue (detection, patch, availability, false-positive penalty) - Agents: protocol, scripted, LLM, and human agent i…8c486a8 -
Add Red & Blue agent design: RangeAgent protocol, BYO adapters, episode loop
Aaron Brown +1: Claude Opus 4.6New doc: docs/red-blue-agents.md - RangeAgent protocol (structural subtyping, same pattern as other protocols) - BYO agent adapters: LiteLLM (any model), scripted (demo), open-weight (GRPO), human - Episode orchestration loop (agents can't reset, RFC 001 compliant) - Tool sets for Red (attacker container) and Blue (SIEM container) - Observation asymmetry (Red=partial, Blue=full logs with NPC noise) - System prompts for both roles - Metrics: flag capture, stealth, detection TP, FP rate, patch success, availability - Training integration: SFT data generation, asymmetric GRPO, trajectory logging - YAML config with env var override, resolve via existing resolve_component() Issue #36 created for implementation. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>63891ca -
Align docs with hackathon statements: multi-agent, self-improvement, simulated experts
Aaron Brown +1: Claude Opus 4.6- README: zero-sum multi-agent framing (Statement 1), long-horizon sparse reward framing (Statement 2), curriculum feedback loop section (Statement 4), simulated expert NPC framing (Snorkel), tier-scaled reward mention (Mercor) - architecture.md: ComplexityBonus in reward tree, tier-scaled reward ceiling table, curriculum section upgraded from post-hackathon to core - IMPLEMENTATION_PLAN.md: curriculum feedback moved to Phase 3.8 - CLAUDE.md: r_complexity reward signal added to both Red and Blue tables - Issue #11: updated with ComplexityBonus and tier multiplier table - Issue #34: created for curriculum feedback loop Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>819cfef -
Validator: hybrid pipeline with task feasibility + difficulty checks
Aaron Brown +1: Claude Opus 4.6Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>7bb17ad -
Agent protocols: pluggable Builder, NPC, Validator
Aaron Brown +1: Claude Opus 4.6Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>fe0a566 -
LLM NPC personas and social engineering surface
Aaron Brown +1: Claude Opus 4.6Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>8c10e98 -
Align docs with snapshot architecture
Aaron Browna9ac5cf
-
Add docs and README
Aaron Browncebc7ff