| spec_version: 1 |
| name: cognitive-load-manager |
| type: space |
| runtime: fastapi |
| app: server.app:app |
| port: 7860 |
|
|
| description: > |
| Cognitive Load Manager (CLM) — a real-world productivity simulation where an AI agent |
| acts as a human task scheduler, managing energy, stress, and fatigue while completing |
| heterogeneous work items (emails, meetings, code reviews, reports, calls) under deadlines. |
| Features task dependencies, mid-episode interruptions, focus mode, and priority weighting. |
| version: "2.0.0" |
| author: "CLM Team" |
| tags: [openenv, scheduling, productivity, rl, agent-eval] |
|
|
| endpoints: |
| health: /health |
| reset: /reset |
| step: /step |
| state: /state |
| grade: /grader |
|
|
| action_space: |
| type: discrete |
| actions: |
| - name: work |
| description: "Work on task_id at normal pace (energy cost varies by task type)" |
| requires: task_id |
| - name: focus |
| description: "Deep-work mode: 2× progress, 2× energy cost; exits on break" |
| requires: task_id |
| - name: break |
| description: "Rest: +0.22 energy, -0.18 stress" |
| - name: switch |
| description: "Change active task (small context-switch cost)" |
| requires: task_id |
| - name: delay |
| description: "Wait one step; slight stress reduction" |
|
|
| observation_space: |
| tasks: |
| - id: string |
| - task_type: "email | meeting | code_review | report | call" |
| - priority: "critical | high | normal | low" |
| - progress: float [0.0, 1.0] |
| - deadline: int (step number) or null |
| - depends_on: task_id or null |
| - is_interrupted: bool |
| visible_state: |
| |
| - fatigue_level: "low | medium | high" |
| - stress_level: "calm | elevated | critical" |
| - stress_warning: bool |
| - focus_mode: bool |
| - upcoming_deadlines: list[task_id] |
| - blocked_tasks: list[task_id] |
| time_step: int |
|
|
| tasks: |
| - id: easy |
| difficulty: easy |
| description: > |
| 2 tasks (email + report), normal priority, no deadlines. |
| Agent must complete both without burning out. |
| Tests basic work/break balance. |
| max_steps: 50 |
| grader: "grader.clm_graders:EasyGrader" |
| baseline_score: 0.856 |
|
|
| - id: medium |
| difficulty: medium |
| description: > |
| 5 heterogeneous tasks (email/meeting/code_review/report/call) with mixed |
| priorities (critical→low) and real deadlines. Agent must triage intelligently. |
| Tests priority-aware scheduling and deadline management. |
| max_steps: 50 |
| grader: "grader.clm_graders:MediumGrader" |
| baseline_score: 0.523 |
|
|
| - id: hard |
| difficulty: hard |
| description: > |
| 8 tasks with explicit dependencies (task B cannot start until task A completes), |
| tight deadlines, and 2 mid-episode urgent email interruptions. |
| Tests dependency-aware scheduling under time pressure. |
| max_steps: 50 |
| grader: "grader.clm_graders:HardGrader" |
| baseline_score: 0.301 |
|
|
| - id: expert |
| difficulty: expert |
| description: > |
| 10 tasks in a deep dependency chain, 3 mid-episode interruptions, |
| mixed critical/high/normal priorities, and very tight deadlines. |
| Genuinely challenges frontier LLM agents. |
| max_steps: 60 |
| grader: "grader.clm_graders:ExpertGrader" |
| baseline_score: 0.221 |
|
|
| scoring: |
| reward_range: [-1.0, 1.0] |
| grader_range: [0.01, 0.99] |
| success_threshold: 0.50 |
| score_formula: deterministic_grader |
| components: |
| - weighted_completion: 0.60 |
| - deadline_adherence: 0.22 |
| - energy_efficiency: 0.10 |
| - dependency_bonus: 0.05 |
| - interruption_bonus: 0.03 |
|
|
| reward_shaping: |
| milestone_rewards: [0.25, 0.50, 0.75, 1.00] |
| burnout_penalty: -1.0 |
| context_switch_penalty: -0.07 |
| blocked_task_penalty: -0.15 |
| stress_penalty_threshold: 0.80 |
|
|
| constraints: |
| max_runtime_seconds: 1800 |
| max_memory_gb: 8 |
| max_vcpu: 2 |
|
|
| inference: |
| script: "inference.py" |
| env_vars: |
| - API_BASE_URL |
| - MODEL_NAME |
| - HF_TOKEN |
|
|
