source/docs/generation.md

Generation

An agent (a coding CLI and a model) rebuilds one benchmark item, a room, an object or a demonstration's scene, in Docker, from the item's public input, under the contract of its track and engine. Code: src/simready/generation/.

• contracts.py: the eight contracts, the images, the CLIs, the egress hosts, prompt rendering.
• prompts/: every word an agent is told ({{name}} placeholders, filled per item and budget; image_tools.json swaps Claude's "Read tool" for Codex's and OpenCode's own image tool; resume.md is the message of each later budget).
• stage.py: the agent's snapshot of data/benchmark/<track>/<id>/input, anonymised by anonymize.py.
• session.py: one isolated session over one or more budgets; cost.py: tokens and money; campaign.py: job lists.
• runtime/: the two Dockerfiles, the pinned Blender and OpenCode installers, the egress proxy, the MuJoCo room self-check and the Isaac Sim SDK example.

Contracts

A contract has a name and no version. Its identity is contract.sha256 in run.json: the hash of its record, every prompt file it renders and every file it stages into the container. Change any of them and the hash changes.

Contract	Given (/work/input, read-only)	Tools (/work/tools, read-only)	Output (/work/output)	Self-check
scenes-mujoco	video.mp4, scale_anchor.json, anchor_reference.png, anchor_marked.png, prompt.txt, generation_spec.json, self_check.py		scene.xml, assets/, objects.json, README.md, refinement_log.json, previews/	python /work/input/self_check.py /work/output (runtime/self_check_mujoco.py: passive limits, missing labels)
scenes-isaac	as MuJoCo, plus sdk_smoke.py		scene.usda, build_scene.py, objects.json, render_preview.py, preview.mp4, interaction.mp4, README.md, refinement_log.json	the evaluator's isaac_physics.py as self_check.py
objects-rigid-mujoco	object.json (name, category), photos/		scene.xml (body object, freejoint, ground), assets/, objects.json, README.md	none
objects-rigid-isaac	as MuJoCo		scene.usda (PhysX scene, box ground), objects.json, README.md	none
objects-deformable-mujoco	case.txt, object_facts.json, media_manifest.json, media/	check_package.py (the evaluator's deformable_admission.py)	scene.xml with one <flexcomp> (dof="full"), assets/ if needed, README.md	python /work/tools/check_package.py /work/output 2d|3d
objects-deformable-isaac	as MuJoCo	check_package.py, deformable_inspect.py (the evaluator's)	scene.usda (/World/Object/Material), meta.json	/isaac-sim/python.sh /work/tools/check_package.py
demos-mujoco	instruction.txt, ext1.mp4, ext2.mp4, timestamps.json, cameras.json (both exterior poses and intrinsics)	check_package.py, mjcf_facts.py, rules.py (the evaluator's)	scene.xml (moved, target, table), assets/, objects.json, README.md	python /work/tools/check_package.py
demos-isaac	as MuJoCo	check_package.py, usd_facts.py, rules.py	scene.usda (/World/moved, /World/target, /World/table), objects.json, README.md	/isaac-sim/python.sh /work/tools/check_package.py

Never mounted: the item's groundtruth/ and provenance/ (sizes, masses, scans, reference shapes, the wrist video and calibration, stereo, the robot log, end-state labels, a room's source tag), the frozen protocol (copied beside the run as evaluator_protocol.json), other runs. The evaluator files are staged byte for byte from src/simready/evaluation/ (paths in STRUCTURE.md); a session refuses to start while one is missing.

What the prompts say. Labels are free text (no class list). The room prompt says the anchor's real dimensions give every other size and distance (measure against it where both are visible, or scale a reconstruction to it). The demonstration prompt says its calibration is metric (a point seen by both cameras can be triangulated), and that if the object is put on the table itself, the table is the target: built once as target, no table, listed as {"id": "target", "label": "table"}. No prompt, spec or staged file names an evaluation, a score, a metric, a threshold, a held-out view, a replay, ground truth or the source dataset (tested). Room refinement: at least 3 render-critique-refine rounds, 6 reference views, 12 inspected frames, the last 90 s for validation.

Staging hides the source. Videos are remuxed with no container or stream metadata (no re-encode, rotation kept, audio dropped); JPEG and PNG lose EXIF, comments and text chunks, losslessly. A room's anchor gets the id scale_anchor and neutral measurement wording. Case ids and translation notes are dropped from object_facts.json and media_manifest.json, whose hashes are recomputed for the staged media. Episode ids are never staged. Tests scan every staged file, name and media tag for a deny-list of dataset, device and lab names (anonymize.SOURCE_WORDS).

Budget. The agent is told the first of the session's cumulative budgets (wall-clock seconds from container start, default 900). The runner kills the container at the deadline; the output as it stands is the result. Each later budget resumes the same CLI conversation for the rest of its time (prompts/resume.md) and is its own run (continued_from).

Tools and network

Every prompt: any tools, including web search, and the installed software, including Blender (blender); download nothing (no files, models, assets, datasets or packages); do not look up this item, its media or the dataset it comes from; do not read authentication files.

• Claude Code (host ~/.local/bin/claude, mounted read-only): every tool but WebFetch; WebSearch runs on Anthropic's side. --claude-config-dir picks the login folder (default ~/.claude, e.g. ~/.claude-huan).
• Codex CLI 0.167.0 (Linux x64, installed in both images, pinned by archive and binary sha256 in runtime/install_codex.sh): web_search="live", browser features off. Model gpt-6.1-sol. The runner copies only the host login (~/.codex/auth.json) into the session state, not the host executable.
• OpenCode 1.19.5 (in both images, pinned by sha256 with the ripgrep its glob and grep tools need): opencode run --pure --auto --format json -m opencode-go/<model>; webfetch denied by its permission config, so not offered; websearch (Exa or Parallel) on. Its config, cache, login and session database live in the session's state folder (XDG_*); updates, the models.dev fetch, plugins, project config and Claude Code compatibility are off. On a V2 host, the runner exports only the model provider's active account with opencode auth export <provider>. It converts V2 API-key credentials into the V1 schema in the private temporary state folder. The host needs no auth.json; its database is neither mounted nor copied. Rotated credentials return through opencode auth import only if the exported account and credentials still match the initial login. A V1 host can still use its existing ~/.local/share/opencode/auth.json. CLI credential output is captured and excluded from errors and run logs.
• Isolation. Read-only input and tools, writable output and scratch, the login copied into a private state folder and saved back when it rotates (a backup first; a login changed meanwhile by someone else is kept). Every 20 s a running session exchanges rotated tokens with the saved login, so sessions sharing a login do not invalidate each other. MuJoCo containers are read-only, non-root, 4 CPUs, 12 GB; Isaac containers one GPU, uid 1234, 12 CPUs, 48 GB.
• Network, enforced. Containers sit on simready-egress, an internal Docker network. Its only exit is simready-egress-proxy (runtime/egress_proxy.py), an HTTPS CONNECT proxy that tunnels only to api.anthropic.com, platform.claude.com, chatgpt.com, auth.openai.com, opencode.ai, mcp.exa.ai and search.parallel.ai (and their subdomains). Plain HTTP, pypi.org, github.com and registry.npmjs.org are refused (tested in Docker).
• Cost. run.json cost: tokens and USD at list price (Claude's own figure; Codex per request from its session counts, which survive a budget stop; OpenCode's own figure from its session database, billed as a subscription).

Image	Dockerfile	Contents
simready-generation:latest	runtime/Dockerfile	Python 3.12, MuJoCo 4.0.0, NumPy, SciPy, Pillow, imageio-ffmpeg, ffmpeg, Blender 5.2.3, Codex 0.167.0, OpenCode 1.19.5 + ripgrep 15.1.0, runtime/requirements-tools.txt (trimesh, Open3D, OpenCV, scikit-image, PyMeshLab, manifold3d, usd-core, shapely, pandas, matplotlib, scikit-learn, networkx, pygltflib, rtree)
simready-isaac-generation:latest	runtime/Dockerfile.isaac	nvcr.io/nvidia/isaac-sim:6.2.0 plus the same Blender, Codex and OpenCode; the host ffmpeg is mounted

docker build -t simready-generation:latest -f src/simready/generation/runtime/Dockerfile src/simready/generation/runtime
docker build -t simready-isaac-generation:latest -f src/simready/generation/runtime/Dockerfile.isaac src/simready/generation/runtime

Both images target Linux AMD64; add --platform linux/amd64 when building on Apple Silicon. Docker Desktop must share the repository and run directories. Codex and OpenCode use image binaries; Claude Code still needs a compatible Linux host executable. Isaac Sim also requires a supported NVIDIA GPU host with NVIDIA Container Toolkit.

An Isaac session needs its CLI's shader-cache volume warmed once (data/runtime/isaac/smoke_<cli>/smoke.json):

CLI=claude; mkdir -p data/runtime/isaac/smoke_$CLI
docker run --rm --user 0:0 --entrypoint bash --mount type=volume,src=simready-isaac-$CLI-cache,dst=/c simready-isaac-generation:latest -lc "chown 1234:1234 /c"
docker run --rm --gpus device=GPU-... --user 1234:1234 --entrypoint bash -e ACCEPT_EULA=Y -e NVIDIA_DRIVER_CAPABILITIES=all \
  --mount type=volume,src=simready-isaac-$CLI-cache,dst=/isaac-sim/.cache \
  --mount type=bind,src=$PWD/src/simready/generation/runtime/isaac_smoke.py,dst=/work/smoke.py,readonly \
  --mount type=bind,src=$PWD/data/runtime/isaac/smoke_$CLI,dst=/work/output simready-isaac-generation:latest -lc "/isaac-sim/python.sh /work/smoke.py"

Commands

Campaigns can enforce subscription reserves with an optional top-level quota object:

"quota": {
  "codex_weekly_remaining_percent": 30,
  "poll_seconds": 30,
  "opencode_pause_on_limit": true
}

The host's Codex app-server account/rateLimits/read supplies the weekly window (10080 minutes). At or below the configured reserve, the runner stops dispatching Codex jobs and stops active Codex containers at their next monitoring check (at most 20 seconds after detection). With a positive reserve, missing or invalid quota data also pause Codex. Polling and provider reporting delays mean this is not an exact account-wide billing cap; other sessions also use the account. No reset credits are consumed. Set the reserve to 0 to use the full weekly allowance. Quota policy edits reload into the existing guard. A reserve-threshold pause otherwise stays paused for this invocation; another CLI may continue. With a positive reserve, a quota-query failure pauses spending while quota remains unknown. With a zero reserve, unavailable quota reporting leaves exhaustion enforcement to the provider. Queries retry at the configured poll interval; a fresh valid reading above the reserve releases only the query-failure pause and its partial job. OpenCode HTTP 402/429 errors pause its login without retrying the same rejected job. These responses can indicate either exhausted allowance or a temporary provider rate limit; they do not establish a remaining balance. Partial quota-stopped runs are retained as quota_paused, skipped by evaluation, and never counted as generated. An explicit restart retries these jobs from the beginning while retaining the partial attempts as provenance. Completed runs are kept. Existing saved adaptive concurrency cannot exceed the configured per_cli_max. Only one campaign run process can hold a campaign's .run.lock at once.

For unattended local runs, redirect stdout/stderr to a persistent log and detach stdin; a closed terminal pipe must not be the process's output destination. Keep Docker, network connectivity and the workstation available.

simready generate --track scenes --engine mujoco --id 41159420 --cli codex --model gpt-6.1-sol --budget 900 1800
simready generate --track objects --engine mujoco --id input_2d_001 --cli claude --claude-config-dir ~/.claude-huan --model claude-opus-5-5
simready generate --track demos --engine isaac --id ID --cli opencode --model opencode-go/kimi-k3.1-pro --gpu GPU-...
simready generate ... --campaign smoke-test --runs-root data/runtime/smoke    # a labelled run outside data/runs
simready campaign run configs/campaigns/example.json --generation-gpus GPU-a --evaluation-gpus GPU-b --render-gpu GPU-b
simready campaign status configs/campaigns/example.json

A run is data/runs/<track>/<engine>/<id>/<UTC time>/: run.json (track, engine, item, cli, model, budgets, stage, continued_from, contract with name, sha256 and the frozen protocol's sha256, image id, agent_cli with version and login folder, status, timings, cli_session_id, throttling, credential_refresh_persistence, cost), submitted_prompt.txt, input/, tools/, output/, generation.jsonl (the CLI's log), evaluator_protocol.json, and token_usage.json(l) for Codex and OpenCode.

Campaign config

{
  "name": "example",
  "budgets": [900, 1800],
  "groups": [{"track": "scenes", "engine": "mujoco", "items": ["41159420"]},
             {"track": "objects", "engine": "isaac", "items": ["input_2d_001"], "budgets": [900]}],
  "models": [{"cli": "claude", "claude_config_dir": "~/.claude-huan", "model": "claude-opus-5-5", "effort": "medium"},
             {"cli": "codex", "model": "gpt-6.1-sol"}, {"cli": "opencode", "model": "opencode-go/kimi-k3.1-pro"}],
  "limits": {"sessions": 24, "per_cli": {"claude:~/.claude-huan": 3, "codex": 4}, "per_cli_max": {"codex": 8},
             "isaac_evaluations": 32, "mujoco_evaluations": 8, "pause_new_jobs_above_load": 170},
  "evaluate": true
}

Every item of every group runs with every model (budgets per group overrides the campaign's). per_cli limits the concurrent sessions of one login, keyed by the login (claude:<config dir>) or by the CLI; a login not listed runs one session at a time. A login whose session fails early or hits a rate limit is held (takes no new sessions); a CLI crash after 120 s is retried once. State is data/campaigns/<name>/state.json; rerunning the command resumes, restarts interrupted sessions from nothing and retries failures. Each result is evaluated with simready evaluate (report in <run>/evaluation/); a campaign refuses to start until every group's protocol is frozen. A campaign stops and removes only its own leftover containers (label simready.campaign=<name>). A group may list models (model names) to run only those. Adding groups or items to a config and rerunning adds their jobs to the campaign. Scheduling adapts to the host: Isaac sessions and evaluations go to the GPU with the most free memory, counting jobs started there in the last 3 minutes as gpu_claim_mib each (default 16000) and starting a job only while at least gpu_keep_free_mib (default 16000) is left after those claims (other users' work is seen through nvidia-smi). Both are limits keys and apply to sessions and evaluations alike; with the defaults a 32 GB GPU takes one Isaac job at a time, so lower both on small GPUs (docs/handoff-jiarui.md: 10000 and 10000 for 32 GB RTX 5090s). A login starts at its per_cli limit, gains one session after 15 minutes at its limit with no rate limit, up to per_cli_max (default twice the start), and on a rate limit halves its limit, pauses 10 minutes and requeues a failed session. Edits to the config (groups, models, limits) apply to a running campaign within seconds; a changed per_cli resets that login's limit to the new value. Isaac work is capped by the NVIDIA driver, not GPU memory: about 50 Isaac processes stalled the driver for every user of the host. isaac_sessions and isaac_evaluations are starting caps; each grows by one after 5 minutes at the cap while nvidia-smi answers within 3 s, up to isaac_sessions_max and isaac_evaluations_max, and both drop to 70% (held for 10 minutes) when it is slower or silent; a changed starting cap applies at once. isaac_evaluations and mujoco_evaluations are worker counts. GPU containers start at most one per gpu_start_gap_seconds (default 15) per pool, and free memory comes from one nvidia-smi reading shared by all schedulers every 15 s: dozens of Isaac containers starting at once (each runs NVIDIA's container hook) stalled the driver for every user of the host. Evaluations share GPUs, so real-time factors are not comparable: run.json has evaluation_gpu_shared: true.