| name | openenv-hackathon-env-designer |
| description | Design OpenEnv-compliant training environments for the OpenEnv Hackathon (India) aligned to themes like multi-agent interactions, long-horizon planning, world modeling, and self-improvement. Use when the user mentions OpenEnv hackathon themes, building an OpenEnv environment, reward design/rubrics, HF Spaces hosting, TRL/Unsloth training scripts, or needs a concrete environment spec + scaffold checklist. |
OpenEnv Hackathon Environment Designer
Quick start (default workflow)
When the user gives theme text (multi-agent / long-horizon / world-modeling / self-improvement / wildcard), produce one concrete environment with:
- Capability target: what the model should learn (1 sentence)
- Core loop: observation → action → transition (3–7 bullets)
- Partial observability: what’s hidden, how it’s revealed
- Actions: allowed tool/API calls + natural language outputs
- Rewards: rubric-style, dense enough to learn, hard to game
- Eval: baseline vs trained comparison + plots
- Packaging: OpenEnv manifest + HF Spaces plan
- Training: minimal TRL/Unsloth script/notebook plan that actually runs
If the request is ambiguous, choose the most “trainable” variant (dense signals, short episodes at first, curriculum knobs).
Additional templates:
Output format (what to write back)
Use this template (headings required; keep it to 1–2 pages):
## Environment: <name>
### Why this matters (capability gap)
- ...
### World + episode loop
- **Episode length**: ...
- **Hidden state**: ...
- **Observation**: ...
- **Actions**: ...
- **Transitions**: ...
### Reward (rubric)
- **R_total** = ...
- **Rubric items**:
- <item>: <signal> (range, frequency), anti-gaming note
### Difficulty & curriculum knobs
- ...
### Evaluation (show improvement)
- **Baselines**: ...
- **Metrics**: ...
- **Plots**: ...
### OpenEnv + HF Spaces packaging plan
- Repo tree:
- ...
- Demo UX:
- ...
### Minimal training script plan (TRL/Unsloth)
- ...
Design rules (judging-aligned defaults)
- Default to “trainability”: prefer rewards that produce learning curves within hours, not days.
- Dense feedback, sparse objective: combine shaping rewards with a terminal objective reward.
- Anti-gaming: for every reward component, name a plausible exploit and block it (caps, normalization, counter-metrics, consistency checks).
- Partial observability is real: hide key variables; provide noisy/limited probes and logs.
- Tool use must be necessary: the environment should require interaction (APIs, simulated tools, dynamic state), not just “answer from prior knowledge”.
- Evidence beats polish: make sure the environment can run + train + plot, even if UI is minimal.
Theme-to-environment mapping heuristics
Theme 1: Multi-agent interactions
Pick one:
- Negotiation (buyer/seller/arbiter) with private values
- Coalitions (3–6 agents) with shifting incentives
- Mixed coop/comp (shared constraint + individual payoff)
Must-have:
- Private info per agent
- Communication channel(s) with bounded budget
- Mechanism constraints (deadlines, compute budget, limited offers)
Theme 2: (Super) long-horizon planning & instruction following
Default pattern:
- Break a 200–300 step goal into checkpoints and memory constraints (e.g., short context windows, periodic “amnesia”, external scratchpad with cost).
- Introduce recoverability: allow re-plans but with penalties.
Must-have:
- Delayed rewards + intermediate rubric scores
- State that persists across sessions/episodes (project-style)
- Evaluation that tests “did it finish?” and “did it recover?”
Theme 3: World modeling (professional / personal)
Default pattern:
- A dynamic system with latent variables that drift (queues, deadlines, trust, inventory, hidden constraints).
- Tools produce partial, sometimes inconsistent evidence (logs, emails, tickets).
Must-have:
- Causal interventions (actions change world)
- Observation updates based on outcomes (belief update)
- Consistency scoring (no contradictions across time)
Theme 4: Self-improvement
Default pattern:
- The agent proposes new tasks/cases; the environment validates and scores them; difficulty auto-scales.
Must-have:
- Curriculum knobs + automatic difficulty adjustment
- “Task quality” rubric (novelty, solvability, non-triviality)
- Overfitting checks (holdout task families)
Theme 5: Wild card
Rule:
- Keep the environment runnable + trainable; novelty is only valuable if you can show improvement.
What to implement first (tracer bullet)
When asked to “build it”, do this in order:
- Single-episode minimal env with deterministic transitions and a simple rubric.
- One exploit test: show a dumb policy that tries to game the reward; patch it.
- Training loop that runs 50–200 episodes and logs reward.
- Curriculum: add 1–2 knobs and show curves improve.
- HF Spaces demo with a small interactive episode runner + plots.
