Vibelifebench

Benchmark version: 1.0.0

Long-horizon — 20+ stages per task, spanning simulated weeks.
Multi-service — task-local environments across 22 services.
Verifiable — atomic checks reading real backend state, not prose.

---

Overview

Vibelifebench evaluates agents on the messy, consequential work of managing someone's life over weeks: a lawsuit, a mortgage escrow shortfall, a cross-city apartment hunt, a licensing exam, an office fit-out.

Each task unfolds as a timeline. The user sends messages, the world changes underneath the agent (a hearing gets rescheduled, a price moves, a policy updates), and the agent has to keep goals, constraints, commitments, and open items coherent across every stage — while knowing which actions it may take on its own and which require asking first.

What makes this hard is not any single step. It is that stage 20 depends on what the agent understood at stage 3, and nothing re-states the context along the way.

Two formats. harbor/ holds 100 self-contained tasks in Harbor format: each ships its own docker compose stack of mock services and runs with the Harbor CLI. eval_set/ holds 20 of those tasks in native format. These are task bundles only; they run on the Terrarium harness from the GitHub repository. See External Runtime Requirement.

Tasks

100 tasks across 10 domains, 10 per domain, in Harbor task format under harbor/ (see Harbor-format subset). 20 of them, two per domain, additionally ship in native format under eval_set/, with bilingual (zh/en) task cards, for the Terrarium harness.

The native-format subset

20 tasks, 489 stages, 1247 atomic checks.

Domain	Task ID	Title	Stages	Envs	Checks	Weight	Difficulty
Career / 职业与劳动权益	career_equity_buyback_recovery	Equity Buyback Reconciliation and Re-employment	25	7	42	100	hard
Career / 职业与劳动权益	career_espp_refund_recovery	ESPP Redemption Reconciliation and Re-employment	24	7	42	100	hard
Exam prep / 考试准备	civil_service_written_to_interview_audit	Civil Service Written Exam to Interview Qualification Audit	25	9	40	62	hard
Exam prep / 考试准备	pharmacist_western_registration_shift_prep	Licensed Pharmacist Registration, Course Purchase, and Shift-Based Preparation	30	5	53	89	hard
Finance / 个人金融	arm_escrow_shortfall_reset_guard_30d	ARM Escrow Shortfall Reset Guard — 30-Day Plan	24	6	123	386	hard
Finance / 个人金融	hsa_medical_bill_liquidity_guard_30d	HSA Medical Bill Liquidity Guard — 30-Day Plan	24	6	123	386	hard
Fitness / 运动与体能	broadcast_exam_posture_breathing_32d	Broadcast Arts Exam Posture, Breathing, and Taper Maintenance	26	5	50	82.5	hard
Fitness / 运动与体能	dragon_boat_newcomer_upper_body_endurance_037	Dragon Boat Newcomer Upper-Body Endurance Preparation	28	6	44	68.5	hard
Litigation / 诉讼管理	food_safety_dispute_33d	Food Safety E-commerce Dispute Litigation — 33 Days	22	5	51	100	hard
Litigation / 诉讼管理	private_lending_33d	Private Lending Recovery Litigation — 33 Days	22	5	71	100	medium
Renovation / 装修与改造	garage_adu_rental_conversion_25d	Legal Garage-to-Rental ADU Conversion	25	8	45	77.5	medium
Renovation / 装修与改造	office_fitout_15d	Commercial Office Fit-Out Project Management	21	7	105	289.657	hard
Rental / 住房租赁	cross_city_remote_viewing_rental	Cross-City Remote Viewing Rental and Address Proof	24	8	66	97.5	hard
Rental / 住房租赁	wheelchair_student_accessible_rental	Accessible Campus Housing for a Wheelchair-Using Student	24	8	58	88.75	hard
Shopping / 购物与履约	baby_stroller_safety_standard_30d	Baby Stroller Safety and Accessory Coordination	24	7	69	134	medium
Shopping / 购物与履约	central_ac_install_30d	Central Air-Conditioning Installation and After-Sales Reconciliation	24	7	70	136.5	hard
Team building / 团队活动	factory_visit_safety_day	Supply Chain Factory Visit Team Day	25	7	46	100	hard
Team building / 团队活动	pottery_invoice_compliance_day	Indoor Pottery Team-Building Planning	25	7	64	100	medium
Travel / 差旅与出行	east_china_bereavement_docs_reissue	Low-Disruption Bereavement Travel and Document Reissue Assistance	22	10	44	100	hard
Travel / 差旅与出行	galapagos_no_us_transit	Galapagos Travel Without U.S. Transit	25	7	41	100	hard

Difficulty labels are audited for v1.0.0. The native-format subset contains 16 tasks labelled hard, 4 labelled medium, and 0 labelled easy. Scores are normalized independently per task under the flat_pool contract.

Envs counts the task's service-environment bindings. Weight is the task's declared total scoring weight; scores are normalized per task, so weights are not comparable across tasks.

Harbor-format subset

The 100 tasks under harbor/, 10 per domain and including the 20 native-format tasks above, are self-contained Harbor tasks. Each ships its own docker compose stack of mock services with seeded data, a timeline of 21–64 steps with per-step instructions and world mutations, an oracle reference trajectory, and a rubric-based final verifier that scores backend state. Same domains and scoring philosophy as the native format, with a different runner.

You need Docker, the Harbor CLI (uv tool install harbor), and an agent such as the claude-code or codex CLI with its credentials in the environment (or pass --env-file .env).

# Fetch only the Harbor-format tasks
hf download EvolventAI/Vibelifebench --repo-type dataset --include 'harbor/*' --local-dir vibelifebench
cd vibelifebench

# Smoke test: one task
harbor run --path harbor/career --include-task-name '*espp*' --agent claude-code --model <model>

# One domain, four trials in parallel
harbor run --path harbor/travel --agent claude-code --model <model> -n 4

harbor/ holds about 49k small files, and the Hub rate-limits per-file downloads (5,000 requests per 5 minutes when logged in, 3,000 anonymously), so fetching all of it takes about an hour. To start sooner, fetch a single domain (4–5.5k files), e.g. --include 'harbor/career/*', or clone the GitHub repository, which carries the same harbor/ tree.

Most tasks request 8 CPUs and 16 GB of memory for their stack, so size -n to your host. harbor/README.md covers the layout, scoring, and the full task inventory. The scripts/run_harbor.sh wrapper it mentions is a thin loop over harbor run and lives in the GitHub repository.

What a task looks like

# event.yaml — the timeline the runtime replays
stages:
  0:
    - id: S00_user_initial_request
      time: 2026-07-24T09:10:00+08:00
      type: user_message
      from: 林乔
      body: |
        我 8 月要去厄瓜多尔参加一个加拉帕戈斯生态数据工作坊...
        能 hold 的先 hold，最终付款和不可退项目要先跟我确认。
  1:
    - id: S01_mutation_workshop_calendar_publish
      time: 2026-07-25T09:40:00+08:00
      type: mutation          # the world changes without being announced
      target: calendar_mock

Stages are checkpoints, not calendar days. Event types include user_message, mutation, notification, world, and policy_update. In Harbor format the same timeline is laid out as steps/event-*/, one directory per step holding its instruction.md and the world mutations applied by workdir/setup.sh.

Capability Coverage

• Long-horizon state maintenance — keep goals, constraints, commitments, and open items consistent across many stages.
• Tool use — query and act across email, calendar, banking, booking, maps, knowledge-base, and notification services.
• Dynamic world updates — apply staged changes in event order without using future facts early.
• Authorization and risk control — distinguish reads and drafts from payments, orders, cancellations, and other high-impact actions.
• Evidence and traceability — keep business state, tool results, and workspace deliverables in agreement.
• Cross-service coordination — reconcile times, amounts, statuses, identities, policies, and dependencies across services.

Service Coverage

Number of tasks using each service in the 100-task Harbor-format subset and in the 20-task native-format subset.

Service	Harbor	Native	Service	Harbor	Native
email	99	20	calendar	96	20
notion	82	18	notification_hub	58	12
ecommerce	39	5	banking	37	6
credit_card	36	5	weather	34	5
maps	32	8	legal_search	29	9
listing_platform	27	4	review_platform	27	7
delivery_logistics	26	2	health_tracker	18	2
job_board	14	2	brokerage	13	4
hotel_booking	13	3	flight_booking	10	2
content_platform	9	1	visa_and_advisory	8	1
rail_booking	5	1	car_rental	3	0

Repository Structure

Vibelifebench/
├── README.md
├── harbor/                     # 100 tasks in Harbor format (see harbor/README.md)
│   ├── README.md
│   └── <domain>/
│       └── <task>/
│           ├── task.toml       # multi-step task definition, final-step reward
│           ├── environment/    # docker compose stack: mock servers, seeds,
│           │                   # world controller, evidence collector, workspace
│           ├── steps/event-*/  # per step: instruction.md, workdir/setup.sh, solution/
│           └── tests/          # rubrics/ + run_verifier.py, the final scorer
└── eval_set/                   # 20 tasks in native format
    └── <domain>/
        └── <task>/
            ├── task.py          # entrypoint, service binding, event dispatch, aggregation
            ├── task.md          # bilingual public task card
            ├── task.toml        # metadata, dependencies, scenario window, scoring summary
            ├── event.yaml       # stage timeline: user messages, notifications, updates
            ├── run.toml         # runner configuration
            ├── workspace/       # initial agent workspace and durable deliverables
            ├── envs/
            │   └── <service>/<env_name>/    # task-local seed, env card, staged data
            ├── mutations/       # standalone staged updates (when present)
            └── rubrics/         # formal scoring modules

The repository root contains only README.md, harbor/, and eval_set/. Every environment payload is task-local; no shared top-level envs/ directory is distributed.

Environments

Each Harbor-format task builds its mock services from environment/servers/ and seeds them from the SQL files under environment/seeds/.

In the native-format subset, all 137 bindings use the layout eval_set/<domain>/<task>/envs/<service>/<env_name>/. Every environment contains a non-empty init.sql and a bilingual README.md; some also include init.json, JSONL records, or SQL updates referenced by event.yaml. Every SQL seed was freshly loaded against its service schema and checked with SQLite integrity_check and foreign-key validation.

Environments are seeded from SQL at load time, so no database files are committed.

Evaluation

Checks are written to read backend state and workspace artifacts rather than reward narration, so describing an action does not earn the credit for performing it.

Harbor format. Tasks score with multi_step_reward_strategy = "final": the last step recomputes every stage's rubrics from the immutable evidence tree, so the reward reflects the whole timeline rather than just the ending. The scorer is tests/run_verifier.py over tests/rubrics/; per-trial diagnostics land in reward.json and checks.json.

Native format. All 20 tasks use flat_pool scoring: atomic checks draw from a single weighted pool, and the task score is the earned fraction of total weight. Rubric modules live in rubrics/ and are loaded and aggregated by task.py:

Module	Scope
stage_<N>.py	Per-stage execution
cross_stage.py	Consistency across stages
final.py	Final deliverables
_helpers.py	Shared predicates and backend-state assertions

Every task provides stage_<N>.py, cross_stage.py, final.py, and _helpers.py; a few carry additional task-specific helper modules. task.toml records each task's post-cleanup atomic-check count and declared total weight.

External Runtime Requirement

Harbor-format tasks need only Docker, the Harbor CLI, and an agent; see Harbor-format subset.

Native-format task content is complete as a bundle, but execution depends on an external runtime that this dataset does not include. The mock-service images, Terrarium capability bindings, and runner scripts live in the GitHub repository, which installs Terrarium as a pinned dependency; its README has a step-by-step quickstart. A compatible environment must:

1. provide the Terrarium/OpenClaw task APIs referenced by task.py;
2. provide implementations and schemas for the declared services;
3. load environments from task-local envs/ per [dependencies.envs] in task.toml;
4. apply event.yaml events in stage and time order;
5. supply workspace persistence, tool traces, and the runtime context the rubrics need.

Set the model in each task's run.toml before running:

[[agents]]
name = "openclaw"
model_name = "<your-provider>/<your-model>"

Synthetic Data and Privacy

All people, organizations, accounts, communications, orders, transactions, places, policy summaries, health records, and other business entities are offline synthetic data. The environments require no internet access and contain no real personal data.

License

No repository-level license file is bundled. Use and redistribution of the task bundles are governed by the terms supplied by the publisher for this repository.

The mock services bundled inside Harbor-format tasks (harbor/<domain>/<task>/environment/servers/) carry their own MIT license files, except car_rental_mock, flight_booking_mock, and rail_booking_mock, which ship without one.

Citation

@misc{vibelifebench_2026,
  title        = {Vibelifebench: A 20-Task Long-Horizon Agent Evaluation Set},
  year         = {2026},
  howpublished = {Task-only release},
  note         = {Long-horizon agent tasks with task-local synthetic environments}
}

---

Vibelifebench · Evolvent AI