earakely-scale/LycheeEnv-REA

lychee

lychee

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

Tables

• lychee_tasks: one row per task of the bundle, with its prompt, envs, verifiers and steps.
• lychee_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/lychee.jsonl holds each run's record and native trajectory, and bundles/lychee/ the bundle itself.

Run it

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

Lychee on AgentEnv, from a REA capture

What this is. A replay environment of Lychee 6.10.4 (MIT), built from one recorded browser session against a self-hosted instance seeded with synthetic data. REA projected the capture into an environment manifest (15 tools), and rea environment-export wrote the bundle you can run from this dataset. Every tool call replays a retained record from the capture, cites the record and the Evidence it came from, and nothing executes Lychee. Tasks are graded deterministically, with no judge.

What's in it

• 15 tools (11 GET, 2 POST, 1 DELETE, 1 PATCH) over 134 recorded exchanges on http://127.0.0.1:3402; 66 recorded groups excluded (66 as UI pages or assets).
• Path parameters: 1 sibling-variance. REA's four format rules (numeric, uuid, hex, iso-date) applied unconditionally; the three opt-in rules (after-uid-literal, opaque-token, sibling-variance) were requested and are labelled per parameter. With the four format rules alone the manifest has 27 tools.
• Unknowns the manifest states explicitly: authorization model, response semantics beyond recorded exchanges, state transitions between tools.
• 12 retrieval tasks (evals/lychee-retrieval.toml): each asks one factual question about a seeded object, with a distinct answer that the app does not ship by default, answerable through a recorded exchange; reward 1 if the response contains the recorded value, else 0. Plus smoke, which deploys the environment and checks replay through the MCP endpoint.
• 24 recorded episodes 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_4c93356ec80d86dddda8881ba10b66ea5bb7bb3dc22787cff2d986fe12319669 (a capture_browser_scenario session REA recorded itself; capture file sha256 420c6f740ee3b72b6d7217abde543a280ad51b66dac844786b87729db1473dd9, 2,395,283 bytes), manifest view ev_9b22b236f85033d95ac596940e36b339f2c0283a934e6fb1c604743c0ead1a86, confidence inferred, authority controlled-replay. The REA build is a prerelease of the environment commands proposed upstream in morluto/rea#4954, #4955, #4975, #4999 and #5182; the bundle installs it from the tarball it ships, so the environment runs without waiting for an npm release.

What replay means, and does not

A tool call returns the recorded response whose path and query values best match the arguments, with replay.exact_match telling you whether it matched exactly or fell back to the closest record. Reads of seeded objects behave like the application. Writes return the recorded answer of the one write the session made, and the environment has no state: a created object does not appear in a later list. This is what makes retrieval tasks hard-gradable and write tasks weak; the manifest's unknowns say so.

Tasks and grading

task	tool	arguments	field	answer
qa-01-get_by_v2_id	get_by_v2_id	v2_id=Gallery::Init	title	Lychee v6
qa-02-get_albums	get_albums	no arguments	albums[0].title	Saltmarsh Depot
qa-03-get_timeline	get_timeline	no arguments	photos[0].title	Annex facade bands, option C
qa-04-get_users	get_users	no arguments	[0].username	ada.finch
qa-05-get_tag	get_tag	tag_id=6	name	lamp room
qa-06-get_album	get_album	album_id=7essfu5uxVTOtu3jvbCzG-we, page=1	resource.albums[0].title	Signal Box Nine
qa-07-get_search	get_search	terms=bGFtcA==, page=1	photos[0].title	Lamp oil swatch, winter budget
qa-08-get_flow	get_flow	page=1	albums[0].title	Copperfield Annex
qa-09-get_tags	get_tags	no arguments	tags[0].name	brass rail
qa-10-get_albums	get_albums	no arguments	pinned_albums[0].title	Lighthouse Rota
qa-11-get_tag	get_tag	tag_id=6	photos[0].title	Fresnel rings, shift three
qa-12-get_album	get_album	album_id=7essfu5uxVTOtu3jvbCzG-we, page=1	resource.photos[0].title	Ballast wagon stripes

Episodes

model	tasks	passed	pass rate	mean tool calls	mean input tokens
openai/gpt-oss-120b:groq	12	11	92%	0.9	6,514
zai-org/GLM-5.3-Flash:zai-org	12	12	100%	1.0	7,364

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

openai/gpt-oss-120b:groq missed qa-03-get_timeline. Miss by gpt-oss-120b:groq on qa-03-get_timeline: the model provider rejected a malformed tool call the model generated (HTTP 400), so the episode ended without an answer.

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/LycheeEnv-REA --bundle lychee --task smoke          # Docker, no model
agent-env hf run earakely-scale/LycheeEnv-REA --bundle lychee --eval lychee-retrieval --model openai/gpt-oss-120b:groq

The retrieval eval needs 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. Lychee ran in Docker on a throwaway VM, bound to loopback, seeded through its own API with synthetic records (admin: rea-admin (created with php artisan lychee:create_user; password generated at seed time, never recorded); albums: Saltmarsh Depot, Signal Box Nine (nested in Saltmarsh Depot), Lighthouse Rota (pinned), Lantern Gallery Repairs (nested in Lighthouse Rota), Blue Heron Kitchen, Copperfield Annex; album fields: description, copyright holder and licence per album (CC-BY-4.0, CC0, reserved, CC-BY-SA-4.0); photos: 14; photo fields: title, description with invented measurements and codes (SD-0412, HK-208, CA-31), tags, taken_at date; 3 starred; images: 14 generated PNGs (stdlib zlib +...); people and projects are invented.
2. REA's capture-browser-scenario launched Chromium and drove the app's own UI, with any credential given as a declared secret; it recorded every network transaction with credential headers and declared secrets redacted, and the environment-manifest view read that capture directly, with no HAR or scrubbing step.
3. rea inspect-analysis-view projected the environment-manifest view; rea environment-export wrote this bundle; gen_tasks.py derived the retrieval tasks from the manifest and the recorded bodies; 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	Lychee (this dataset)
Environment	A simulation: a container quay, ships, cranes, weather, with A2A ship agents	A replay of a real application's API, captured from one session
Observation	Tool results computed by the simulator as the week unfolds	Recorded responses, cited to Evidence
Action space	Planning tools (check_plan, confirm_berths, advance, message_agent)	15 tools mirroring the application's endpoints
State	Full: a virtual clock, triggers, frozen windows, agents that answer	None: every call is independent; writes replay one recorded answer
Reward	Deterministic, against a CP-SAT optimum or a re-planner anchor; continuous 0 to 1	Deterministic, recorded-value match; 0 or 1 per task
What it measures	Planning under uncertainty and negotiation	Tool discovery, parameter inference and faithful retrieval in a real API
Data provenance	Real 2024 port calls and weather; simulated workloads	Synthetic seed data on a self-hosted Lychee; 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 list exactly what a stateful version would still have to specify.

Licence and credits

Bundle code and tasks: Apache-2.0. Recorded content: responses generated by a self-hosted Lychee instance over synthetic data; Lychee is MIT software by LycheeOrg 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.