Skip to content

ROADMAP.md

Latest commit

 

History

History
263 lines (230 loc) · 12.9 KB

ROADMAP.md

File metadata and controls

263 lines (230 loc) · 12.9 KB

OpenRange Roadmap

Note

Direction, not a release schedule. Each item links to its tracking issue — that's where status, scope, and discussion live.

The bet behind OpenRange: agents trained against fresh, runnable, admission-checked worlds will generalize better than agents that overfit to static benches. Making that real means (a) good packs across many domains, (b) a clean way to plug a training loop into them, and (c) a core that scales from "single process" to "production training fleet."

Status tags on each item:

  • 🚧 in progress — actively being worked on
  • 🟢 help wanted — well-scoped, contributor-friendly
  • 🟡 design needed — design doc PR before code
  • 🔮 research — exploratory, no near-term plan
  • ✅ shipped — landed; kept here for context

Browse all roadmap issues: label:roadmap.

Two near-term priorities:

  1. From eval to training. Today OpenRange only runs eval loops — closing them into training is the highest-leverage next step. That means a training-integration standard for our dynamic, evolving worlds (#243), with open-trajectory-gym (#198) as the reference consumer.
  2. Cyber depth → enterprise scale. Deepen the cyber pack and push toward enterprise scale via graph-driven lazy realization (#212 / #235).

Physical-space simulation (#219) is explicitly far-future.

Packs

Packs are the reusable starting points for a family of worlds — ontology, realizer, defaults. See docs/start_here.md for the contract.

Cyber (priority)

Cyber is the first-class pack. Both offense and defense are first-class task families on the same pack — same world, different role, different success check. Long-term goal: dual-mode scenarios where one realized world hosts an offense agent and a defense agent in parallel.

  • 🚧 webapp — procedural builder + AST-spliced vuln injection + NPCs + per-task feasibility/success checks via TaskFamily. Already shipping. LLM enrichment is wired only on the pentest family's curriculum (available_mutations); task-instruction enrichment is unwired.
  • 🚧 webapp.build rigorous grader — staged rollout. Stage 1 has landed: agent submits a handler source string into result.json; check_success runs it through a sandboxed subprocess against a held-out behavioral contract. Admission validates the contract is well-posed (clean reference passes; bug-injecting mutation breaks). Only the api service kind is wired; other kinds emit no build task. Stages 2–4 (#237) expand the mutation/contract library, add curriculum mutations on build, and plumb live realizer reload + an agent test-runner tool so the loop is multi-turn instead of single-shot.
  • 🟢 Kind / Kubernetes runtime backing (#189) — unblocks real cross-service exploit chains.
  • 🟢 Expand the vulnerability catalog (#190) — command injection, path traversal, deserialization, weak creds, etc.
  • 🟢 Defense-cyber task family (#191) — same pack, detection / mitigation / patching tasks instead of flag retrieval. Enables dual-mode adversarial scenarios.
  • 🟡 LLM-driven naming realism (#192) — svc_web → customer-portal, acct_0 → alice@corp.example.
  • 🔮 MCTS-based world generator (#193) — search over graph mutations instead of rejection sampling.
  • 🟡 Enterprise-scale worlds (#212) — 100–1000× today's scale: org chart, AD-shaped identity, multi-site networks, population NPCs. Umbrella; map the shape before building.
  • 🟡 Graph-driven lazy realization (#235) — the mechanism that makes enterprise scale tractable: use the world graph's reachability to spin up only what the agent can reach (hot / warm / cold tiers) instead of realizing the whole estate at once.
  • 🔮 Phishing / mail channel (#87) — non-HTTP mail entrypoint + an email-handling NPC. Longer-term; the multi-channel direction overlaps the physical-space vision below.

Other domains

A non-cyber pack proves the Pack ABC isn't accidentally cyber-shaped. Trading is the first; more domains welcome.

  • ✅ Trading pack — shipped (#194). Order-book backtest + P&L success check; the first non-HTTP backing and the load-bearing proof that the pack contract is domain-agnostic.
  • 🚧 SWE pack — MVP shipped, generation next. (#248) — code repos + test-suite grading; generalizes the webapp.build grader from one handler to a real repo + test runner. The imported source shipped (#247, packs/swe/DESIGN.md): a world freezes to a SWE-bench-shaped triple (repo.base_files + held-out test_suite + HIDDEN solution), admission proves it well-posed (the gold overlay greens the suite; the bare base fails its grading tests), and a sandboxed filesystem realizer grades the agent's edited tree — fed either by a bundled fixture or a cloned SWE-bench row. Two task shapes ride the one world: swe.fix (all-or-nothing repair) and swe.build (units shape partial credit, integration tests gate success — the long-horizon sibling). Next: the injected (AST-splice) and authored (LLM) world-sources + curriculum, plus seccomp / container hardening for adversarial public traffic (#202).
  • 🔮 Social / negotiation pack (#195) — multi-turn dialogue with NPC counterparties.
  • 🔮 Robotics pack (#196) — MuJoCo / Isaac integration via non-HTTP backing.
  • 🟢 Pack author guide (#197) — step-by-step doc lowering the bar for first-time pack authors.

Training integration

Today the examples are eval loops. Closing them into actual training is the next big piece. Because OpenRange worlds are dynamic and admission-checked — they evolve — the integration seam carries more than a static gym does: the aim is a small, trainer-agnostic standard, EpisodeResult / EpisodeReport → (trajectory, reward) plus the curriculum/evolve dimension, that any trainer can consume.

  • 🟡 Training-integration standard for dynamic worlds (#243) — define the seam + adapter protocol so many trainers plug in. Umbrella for the items below.
  • 🟡 In-house reference: open-trajectory-gym (#198) — pair OpenRange with open-trajectory-gym, vecna's own SFT + Online RL + GEPA trainer; the first / reference consumer of the seam.
  • 🟢 Per-domain reward adapters (#199) — structured EpisodeResult → scalar / vector reward signal (the reward half of the seam).
  • 🟢 Reach adapters: SkyRL / TRL (#244 / #245) — trainer-side adapters that bridge the seam out to external trainers. verl / TorchForge are later candidates.
  • 🟢 Curriculum-driven training demo (#200) — notebook showing success-rate curves as evolve(...) hardens worlds.

Core

Primitives every pack and harness depends on.

  • 🚧 Per-NPC LLM backend selection (#236) — extracting openrange-pack-sdk removed the model: config key from the cyber webapp NPC factories (it used to construct a concrete StrandsAgentBackend and so couldn't survive the strict pack-cannot-import-openrange boundary). Every NPC in an episode now shares the single RunConfig.npc_agent_backend. To restore per-NPC model selection cleanly, add a backend_id: config key on NPC factories
    • a RunConfig.npc_backend_factory: Callable[[str], AgentBackend] | None the runtime resolves at NPC start. Mirror of the entry-point pattern; no openrange import in pack code.
  • 🟢 API-key LLM backends (#188) — a LiteLLMBackend + a zero-dep OpenAICompatibleBackend so anyone can build worlds with their own provider, not just the local Codex CLI. PR #220 in flight.
  • 🟢 Restore 100% test coverage (#201) — coverage dropped to 80% during the typed-property-graph + pack / admission refactor; bring it back.
  • 🟡 Mature the pack SDK (#239) — stabilize, version, document, and decide on publishing openrange-pack-sdk + graphschema. The whole architecture hinges on this contract; every new pack is another consumer, so churn gets more expensive over time.
  • 🟡 TaskFamily check sandboxing (#202) — the pack/admission refactor replaced exec'd verifier source with regular Python methods on TaskFamily; revisit isolation/timeouts at training scale. The cyber webapp.build grader exec's untrusted agent source as a subprocess with wall-clock + best-effort RLIMIT_CPU only — no filesystem, network, or syscall isolation. Production use needs firejail / bwrap / seccomp / container.
  • 🟡 Lineage as a pool, not a chain (#203) — multi-parent reference graph for curriculum / multi-snapshot training.
  • 🟡 Multi-host EpisodeService (#204) — distributed training fleets.
  • 🔮 Container snapshotting for stateful runtimes (#205) — CRIU / overlayfs / docker commit for fork/restore.

Dashboard

The inspection surface for builders, snapshots, and live episodes. See docs/dashboard.md.

  • 🟢 Snapshot inspector improvements (#206) — better UX for large world graphs, oracle-path highlighting.
  • 🟢 Live training view (#207) — curriculum evolution + per-rollout request log when training lands.
  • 🟡 Pluggable per-pack dashboard widgets (#208) — cyber's natural view is service topology; trading's is order book.

Docs

  • 🟢 Tutorial: build → eval → curriculum (#209) — uv sync to working eval to curriculum walk in one notebook.
  • 🟢 Architecture decisions log (#210) — short notes on the why of big shape choices.

Far-future / research

Big directional bets with no near-term plan. Here to capture the vision, not to schedule it.

  • 🔮 Physical-space simulation substrate (#219) — OpenRange as a grounded multi-agent environment (rooms, bodies, objects, conversations) with the cyber pack as one scenario rendered on top. Below enterprise-scale in priority; the cyber-relevant slices surface nearer-term via #87.

How to help

  1. Pick something tagged 🟢 → comment on the issue saying you want it.
  2. For new packs (trading, social, robotics, anything else), open the issue with a sketch of the ontology + TaskFamily shape; we'll work through the contract with you.
  3. 🟡 items want a design doc PR before code — saves cycles.
  4. Friday community call — link in README.md.

See CONTRIBUTING.md for setup and workflow.