earakely-scale/MetabaseEnv-REA

metabase

metabase

An AgentEnv bundle with its recorded runs, published with agentenv-hf.

Tables

• metabase_tasks: one row per task of the bundle, with its prompt, envs, verifiers and steps.
• metabase_episodes: one row per recorded run, with its status, model, reward and scores, the prompt and response, and the agent's transcript as chat messages when its trajectory format is one agentenv-hf reads.

raw/metabase.jsonl holds each run's record and native trajectory, and bundles/metabase/ the bundle itself.

Run it

pip install agentenv-framework huggingface_hub
hf download earakely-scale/MetabaseEnv-REA --repo-type dataset --local-dir agentenv-dataset
agent-env run agentenv-dataset/bundles/metabase

Metabase on AgentEnv, from a REA capture

What this is. An environment of Metabase 0.64.2.8 (AGPL-3.0), built from one recorded browser session against a self-hosted instance seeded with synthetic data. REA projected the capture into an environment manifest (42 tools) and an environment state, the entities the capture's reads showed, and rea environment-export wrote the bundle you can run from this dataset. Reads the state can render are answered from it, write tools whose effect the capture confirmed change it, and data/reset restores it. Nothing executes Metabase. Tasks are graded deterministically, with no judge.

v0.2.0 rebuilds this environment from a new capture of an instance seeded the same way, recorded so each write's effect is shown: two scratch objects per write target, every write read back at once. A REA build that serves a state, confirmed writes and list queries projects it; the tag v0.1.0 keeps the earlier replay-only version.

What's in it

• 42 tools (37 GET, 3 POST, 1 DELETE, 1 PUT) over 137 recorded exchanges on http://127.0.0.1:3201; no recorded group excluded. 41 are offered; the other one is a write tool the environment cannot apply (post_card), left out rather than faked.
• Path parameters: 25 numeric. REA's four format rules (numeric, uuid, hex, iso-date) applied unconditionally; the four opt-in rules (after-uid-literal, opaque-token, sibling-variance, prefix-variance) were requested and are labelled per parameter. With the four format rules alone the manifest has 42 tools.
• State: 46 entities of 16 types. 33 read tools answer from it and 6 replay their recorded responses. 8 list tools also answer filters, search, sorting or paging from it. Served write tools: delete_card_by_id (delete), put_card_by_id (update).
• Unknowns the manifest states explicitly: authorization model, response semantics beyond recorded exchanges, state transitions between tools.
• 12 retrieval tasks (evals/metabase-retrieval.toml): each asks one factual question about a seeded object, with a distinct answer that the app does not ship by default; reward 1 if the response contains the recorded value, else 0. Plus smoke, which deploys the environment and checks it through the MCP endpoint.
• 2 write tasks (evals/metabase-writes.toml): each asks for a change in plain language and is graded from the environment's state through data/get, not from the agent's reply; reward 1 only if every check passes. Their grade is in scores.state; reward holds the retrieval verifier's score, so it is empty for them, as for smoke.
• 24 retrieval episodes and 4 write episodes, plus the smoke run, by openai/gpt-oss-120b:groq, zai-org/GLM-5.3-Flash:zai-org through the Hugging Face router, with their chat transcripts.
• Provenance: capture ev_f49b41d700f68477693f35ae3f0fafaa5d0ce3fbb225f49861e539ce8b582c3f (scoped HAR sha256 c41e443c811d16b5f70f8a9d7e3249fff5f40d9e42346b16e877824be4601f05, 780,237 bytes), manifest view ev_cbf0a5453a22abc796f14cd76b12852eb6eaf46b2274f502d56281e4831bbda6, confidence inferred, authority historical-reference. The REA build is rea-agents 6.7.0 from main 3e7e102a with open pull requests merged: morluto/rea#6385 at 342933ea and morluto/rea#6897 at 08c2532a. The bundle installs it from the tarball it ships, so the environment runs without waiting for an npm release.

What the state means, and does not

Reads the state can render, item, list and document views of the entities the capture's responses showed, answer from it. List reads filtered, searched, sorted or paged answer from it too wherever the capture's recorded calls explain what each query parameter does, including values no call recorded; a parameter the recordings do not explain, or a flag's unrecorded value, replays. Other reads replay the recorded response whose path and query values best match the arguments, with replay.exact_match saying whether it matched exactly; a replayed read does not change after a write. A served write was confirmed by the capture itself: replaying the session against the extracted state, a later recorded read reproduced its effect and none contradicted it. A create generates ids and clock values deterministically, so the same calls after a reset give the same state; an update changes the members the call sends; a delete removes the entity. Each takes only the body members the recorded call showed and refuses others with a message that names the accepted ones. Its effect shows in the reads the state renders, lists included wherever a list lists its whole collection or a learned query answers the call; a list call nothing explains keeps its recorded members, dropping a deleted entity but not adding a new one. The data plane at /agentenv takes data/reset (restore the extracted state), data/add and data/get (return it), which is how write tasks are graded.

Not modelled: side effects a write has beyond its own entity (counters, copies in other collections, related entities), preconditions the application enforces before a write, list query parameters the recorded calls do not explain, paging members such as page counts and next links when a call asks for another page size or page, and anything the capture never showed. The manifest's unknowns list what a capture cannot establish.

Tasks and grading

Retrieval

task	tool	arguments	field	answer
qa-01-get_timeline_by_id	get_timeline_by_id	id=1, include=events	events[0].name	Quarry audit signed off by Rui Castellanos
qa-02-get_user_by_id	get_user_by_id	id=2	email	ada.finch@example.test
qa-03-get_permissions_group_by_id	get_permissions_group_by_id	id=5	name	Night Shift Analysts
qa-04-get_timeline_event_by_id	get_timeline_event_by_id	id=1	description	412 blocks north quarry, 97 in the tidal yard.
qa-05-get_dashboard_by_id	get_dashboard_by_id	id=2	description	Tide-table menu metrics for the 36-cover sittings.
qa-06-get_segment_by_id	get_segment_by_id	id=1	name	Granite Ledger large orders
qa-07-get_card_by_id	get_card_by_id	id=2	description	Mean product rating by category, read out at the 22:00 kitchen handover.
qa-08-get_native_query_snippet_by_id	get_native_query_snippet_by_id	id=1	name	Saltmarsh Depot gizmo filter
qa-09-get_collection_items_by_id	get_collection_items_by_id	id=7	data[0].name	Blue Heron Kitchen average rating per category
qa-10-get_dashboard	get_dashboard	f=all	[0].name	Blue Heron Kitchen Service Board
qa-11-get_segment	get_segment	no arguments	[0].description	Orders whose ledger total is above 120 tide-tokens.
qa-12-get_timeline	get_timeline	include=events	[0].name	Granite Ledger milestones

Writes

Graded from the environment's state through data/get (scores.state in the episodes table; reward 1 only if every check passes).

task	prompt	checks
write-update	In this Metabase instance, change the description of the saved question "Lighthouse Rota sign-ups by source" to "Sign-ups per acquisition source, recounted for the spring rota."	the question has the new description; no saved question was added or removed; another question's description is unchanged
write-delete-lookup	In this Metabase instance, one model in the Granite Ledger collection holds the order book. Find it and delete it.	the Granite Ledger order book model is gone; the other four saved questions are still there; the Granite Ledger question Copperfield Annex orders by shelf family is still there

Episodes

model	retrieval episodes	passed	pass rate	mean tool calls	mean input tokens
openai/gpt-oss-120b:groq	12	12	100%	1.0	23,066
zai-org/GLM-5.3-Flash:zai-org	12	12	100%	1.0	26,870

Every episode is in the episodes config with its transcript; scores are the deterministic verifier scores.

model	write episodes	passed
openai/gpt-oss-120b:groq	2	1
zai-org/GLM-5.3-Flash:zai-org	2	2

The card create is not served: after it the read-backs show background statistics Metabase computes for the new card and its list form, which the build counts as a side effect, so the write tasks cover the card update and delete only. In write-update, openai/gpt-oss-120b:groq never finished: after 11 tool calls Groq rejected its next turn ("Tool choice is none, but model called a tool", HTTP 400), so the episode ended as a provider error rather than a grade.

Run it yourself

uv tool install agentenv-framework --with 'agentenv-hf @ git+https://github.com/earakely-scale/agentenv-hf-plugin@v0.3.0'
agent-env hf run earakely-scale/MetabaseEnv-REA --bundle metabase --task smoke          # Docker, no model
agent-env hf run earakely-scale/MetabaseEnv-REA --bundle metabase --eval metabase-retrieval --model openai/gpt-oss-120b:groq
agent-env hf run earakely-scale/MetabaseEnv-REA --bundle metabase --eval metabase-writes --model openai/gpt-oss-120b:groq

The evals need a model endpoint: set [model] base_url and api_key in .agentenv/config.toml (the Hugging Face router works with any hf_ token). The bundle's agents/replay-qa is a small OpenAI-compatible tool-calling agent; any A2A agent can replace it.

How it was made

1. Metabase ran in Docker on a throwaway VM, bound to loopback, seeded through its own API with synthetic records (users Ada Finch, Rui Castellanos and Mina Okafor; group Night Shift Analysts; collections Granite Ledger, Blue Heron Kitchen and Saltmarsh Depot; 5 saved questions over the bundled Sample Database; 2 dashboards; a timeline with 2 events; a native-query snippet and a segment); people and places are invented.
2. Playwright drove a login, a scripted read sequence through the app's JSON API and the main pages, then wrote two scratch saved questions (created, renamed and deleted) and re-described a seeded model, reading each write back at once by item and by the card list; the HAR was scoped to the API, login, session and settings exchanges dropped, secret-looking string fields scrubbed, headers and cookies removed.
3. rea inspect-web-network-capture retained the exchanges as Evidence; rea inspect-analysis-view projected the environment-manifest view; rea environment-export wrote this bundle and the state it serves; gen_tasks.py derived the retrieval tasks from the manifest and the recorded bodies; the write tasks were written by hand against the served write tools and the seeded names; agent-env run recorded the episodes; agent-env hf publish wrote this dataset. Scripts: https://github.com/earakely-scale/rea-environments.

Compared with PortSimEnv on AgentEnv

	PortSimEnv-AgentEnv	Metabase (this dataset)
Environment	A simulation: a container quay, ships, cranes, weather, with A2A ship agents	An application's API rebuilt from one captured session
Observation	Tool results computed by the simulator as the week unfolds	Answers rendered from the extracted state, or recorded responses, cited to Evidence
Action space	Planning tools (check_plan, confirm_berths, advance, message_agent)	41 tools mirroring the application's endpoints
State	Full: a virtual clock, triggers, frozen windows, agents that answer	A state extracted from the capture (46 entities); served writes change it; data/reset restores it
Reward	Deterministic, against a CP-SAT optimum or a re-planner anchor; continuous 0 to 1	Deterministic: recorded-value match for retrieval, state checks for writes; 0 or 1 per task
What it measures	Planning under uncertainty and negotiation	Tool discovery, parameter inference and faithful retrieval in a real API, and changing its state
Data provenance	Real 2024 port calls and weather; simulated workloads	Synthetic seed data on a self-hosted Metabase; nothing real
How it was built	Hand-written plugin, months of work	Generated from a capture in one scripted pass
Where it is strong	Long-horizon decisions with hard rewards	Breadth: any app you can capture becomes an environment

The two are complementary: PortSimEnv shows what a hand-built simulation can grade; this dataset shows how far a capture alone gets you, and the manifest's unknowns and the state's limits above say what it still leaves out.

Licence and credits

Bundle code and tasks: Apache-2.0. Recorded content: responses generated by a self-hosted Metabase instance over synthetic data; Metabase is AGPL-3.0 software by Metabase, Inc. and is not redistributed here beyond the recorded JSON responses. REA is MIT software by morluto; AgentEnv Framework is Apache-2.0 by Scale AI.