CLI reference

Every repo2rlenv command, subcommand and flag with its default, read from the argparse definitions.

repo2rlenv is one executable with thirteen commands. This page lists every flag and its default as defined in the v0.9.3 argparse code. Run repo2rlenv COMMAND --help (or repo2rlenv COMMAND SUBCOMMAND --help) for the same information in your terminal.

Global flags

repo2rlenv [--version] [-v] [--no-ui] COMMAND ...

Flag	Default	Description
`--version`	n/a	Print `repo2rlenv <version>` and exit.
`-v`, `--verbose`	off	Debug logging. On an error, also print the traceback to stderr.
`--no-ui`	off	Disable Rich live displays and print plain log lines.
`-h`, `--help`	n/a	Show help for the command it follows.

Global flags go before the command: repo2rlenv --no-ui generate …. Placed after the command, they fail with unrecognized arguments.

Conventions

Credentials. At startup the CLI loads a .env file without overriding variables you already set. See .env files for where it looks.
Exit codes. 0 means success. 1 means the command finished but produced nothing usable (for example generate emitted no tasks or validate found a failing task). 2 means invalid arguments or a setup error.
Errors. Printed to stderr as ErrorType: message. Commands that accept --json print {"error": "ErrorType", "message": "…"} to stdout instead.
Live displays. They turn off automatically when stdout isn't a terminal, when NO_COLOR or CI is set, or when TERM=dumb.

generate

Run a pipeline against a source and write Harbor tasks to a local directory.

repo2rlenv generate [--config FILE] [--repo REPO] [--pipeline NAME] [--recipe NAME]
                    [--pipeline-opt KEY=VALUE ...] [--llm PROVIDER/MODEL] [--out DIR] ...

generate routes on --recipe:

Native (no --recipe, or --recipe native) runs on your machine. pr_diff needs neither Docker nor an LLM. Every other native pipeline needs --llm and Docker, because it first bootstraps a container image for the repository.
Research recipes (--recipe swe_smith, …) and non-repository sources need a --config file with an execution: block and a running remote worker. There is no local fallback. See Remote execution.

Flag	Default	Description
`--config FILE`	none	Generation input as YAML (`.yaml`, `.yml`), TOML or JSON. Flags on the command line override its values.
`--repo REPO`	none	GitHub `owner/name`, a GitHub or GitLab URL, or a local path.
`--ref REF`	`HEAD`	Branch, tag or commit. Applied together with `--repo`.
`--access`	`auto`	`public`, `private` or `auto`. Applied together with `--repo`.
`--pipeline NAME`	none	Pipeline name, such as `pr_diff`. `repo2rlenv pipelines list` shows them all.
`--recipe NAME`	`native`	Research recipe to run for the pipeline, such as `swe_smith`.
`--pipeline-opt KEY=VALUE`	none	Pipeline option. Repeatable. `true`/`false` become booleans, JSON values (numbers, lists) are parsed, and anything else stays a string. Overrides options from `--config`.
`--llm PROVIDER/MODEL`	none	Model for synthesis and bootstrap, such as `anthropic/claude-sonnet-4-6`.
`--llm-fallback PROVIDER/MODEL`	none	Model to switch to when the primary returns 5xx, rate-limit or network errors.
`--llm-endpoint URL`	none	Base URL of a self-hosted OpenAI-compatible server for the primary model, such as `http://localhost:8000/v1`. The provider's default key is never sent there.
`--llm-key-env VAR`	provider default	Environment variable that holds the primary model's API key.
`--out DIR`	none	Output directory.
`--org ORG`	`default`	Task namespace: task names become `ORG/<slug>`. Applied together with `--out`.
`--dataset-name NAME`	name of `--out`	Dataset name recorded in the output spec. Applied together with `--out`.
`--visibility`	`public`	`public` or `private`. Applied together with `--out`.
`--max-spend-usd N`	`5.0`	Native only. Caps the bootstrap agent's LLM spend; `0` removes the cap. Research recipes reject it, because their spend comes from a campaign budget.
`--language LANG`	auto-detect	Bootstrap language: `python`, `node`, `go`, `rust`, `java` or `c_cpp`.
`--base-image IMAGE`	per language	Bootstrap base image, such as `ubuntu:24.04` or `python:3.11-slim`.
`--force-bootstrap`	off	Ignore the bootstrap cache and rebuild the image.
`--bootstrap-opt KEY=VALUE`	none	Set any bootstrap field, such as `cache_dir`, `max_iterations`, `max_seconds` or `user_dockerfile`. Repeatable.
`--force-language`	off	Skip the check that the repository's primary language is one the pipeline supports.
`--resume`	off	Recipes only. Continue a run without redispatching completed work.
`--json`	off	Recipes only. Stream progress as JSON Lines. Native generation rejects it.

--llm-fallback, --llm-endpoint and --llm-key-env need --llm or an llm: block in --config. The bootstrap flags (--language through --force-language) only affect native pipelines that bootstrap.

The pipeline option limit counts candidates listed (for PR pipelines, merged PRs), not tasks emitted. Filters usually drop some, so generate can finish with fewer tasks than the limit. It exits 1 when it emits none.

repo2rlenv generate \
  --repo pallets/click \
  --pipeline pr_diff \
  --pipeline-opt limit=10 \
  --out ./tasks

A research recipe takes its source, options and execution: block from the config file:

repo2rlenv --no-ui generate --config examples/owned-swe-smith.yaml

validate

Statically check a dataset or task directory. Nothing is built or run.

repo2rlenv validate [--deep] [--oracle] PATH

Flag	Default	Description
`PATH`	required	Dataset or single task directory. Every `task.toml` below it is checked.
`--deep`	off	Also check each task's assets and metadata: instruction, environment definition, `tests/test.sh`, graded verifier files and the reproducibility table.
`--oracle`	off	Implies `--deep`. Also require `solution/solve.sh` and, for native pipelines, a non-empty `solution/patch.diff`.

Without flags, validate confirms each task.toml parses and has a [task] name. It exits 1 if any task fails or if no task.toml is found. Deep validation lists every check.

validate currently rejects research-recipe and Tasksmith output (schema 1.3 bundles) with missing top-level ['version'], because it still requires the Harbor 1.0 version key. A fix is in progress. Until then, use it on native pipeline output.

repo2rlenv validate ./tasks --oracle

push

Upload a local dataset directory to a Hugging Face dataset repository.

repo2rlenv push [--private] [-m MESSAGE] [--image-registry PREFIX] [--inline-dockerfile]
                [--require-registry] [--skip-image-push] [--image-visibility VIS]
                [LOCAL_DIR] [DATASET]
repo2rlenv push --check-auth [--fast] [--json]

push stages the tasks under tasks/, writes a dataset card and manifest.json, uploads them, then uploads a registry.json pinned to that commit. For tasks whose Dockerfile starts from a locally built bootstrap image, it first pushes the image to a container registry, or inlines the build recipe. It rewrites environment/Dockerfile and task.toml in place to record the choice. See Publishing.

Flag	Default	Description
`LOCAL_DIR`	`.`	Dataset directory, as written by `generate`.
`DATASET`	none	Target as `owner/name` or `hf://owner/name`. Required unless `--check-auth`. `harbor://` and GitHub targets print the equivalent `harbor publish` or `git push` command instead.
`--private`	off	Create the dataset repository as private.
`-m`, `--message TEXT`	`Repo2RLEnv: add N tasks`	Commit message.
`--image-registry PREFIX`	auto-detect	Force a registry prefix, such as `ghcr.io/my-org`. It's probed for write access first.
`--inline-dockerfile`	off	Don't push images; write the bootstrap recipe into each `environment/Dockerfile`. Rebuilds from the recipe, not bit-exact.
`--require-registry`	off	Fail if no verified registry is available instead of falling back to inline mode.
`--skip-image-push`	off	Rewrite tasks against an image that already exists at the registry; don't run `docker push`.
`--image-visibility`	`inherit`	`public`, `private` or `inherit` (match the dataset's visibility).
`--check-auth`	off	Probe every registry you're logged in to, report, and exit. Uploads nothing.
`--fast`	off	With `--check-auth`: stop after the reachability and authentication levels.
`--json`	off	With `--check-auth`: print the probe results as JSON.

repo2rlenv push ./tasks my-org/click-pr-diff

pull

Download a dataset from the Hugging Face Hub, a Harbor registry or GitHub.

repo2rlenv pull [--task NAME] [--registry-url URL] [--force] DATASET [LOCAL_DIR]

Flag	Default	Description
`DATASET`	required	`owner/name`, `owner/name@rev`, `hf://owner/name`, or a bare `name` (owner resolved from your Hub login) for the Hub. `harbor://name[@tag]` for a Harbor registry. `gh://owner/repo[@ref]` or a `https://github.com/owner/repo` URL for GitHub.
`LOCAL_DIR`	`./datasets/<owner>__<name>`	Destination directory.
`--task NAME`	none	Hub only: download one task, such as `encode__httpx-3367`.
`--registry-url URL`	Harbor's public registry	Harbor only: a custom registry. Requires the `harbor` CLI on your `PATH`.
`--force`	off	Delete the local copy and download again.

A Hub pull flattens the published tasks/<id>/ layout back to LOCAL_DIR/<id>/, ready for harbor run -p. A GitHub pull is a shallow git clone.

repo2rlenv pull FineEnvs/repo2rlenv-pr-runtime ./pr-runtime --task encode__httpx-3367

bootstrap

Build and cache a working Docker image for a repository with an LLM agent, without generating tasks. A later generate against the same repository and commit reuses the cached image.

repo2rlenv bootstrap --repo REPO --llm PROVIDER/MODEL [--ref REF] [--max-spend-usd N] ...

Flag	Default	Description
`--repo REPO`	required	GitHub `owner/name` or URL.
`--ref REF`	`HEAD`	Branch, tag or commit.
`--access`	`auto`	`public`, `private` or `auto`.
`--llm PROVIDER/MODEL`	required	Model that drives the build agent.
`--llm-endpoint URL`	none	Self-hosted OpenAI-compatible server for `--llm`.
`--llm-key-env VAR`	provider default	Environment variable that holds the API key.
`--max-iterations N`	`25`	Maximum agent turns.
`--max-seconds N`	`1800`	Wall-clock limit.
`--cache-dir DIR`	`$R2E_CACHE_DIR`, else `./workspace/bootstrap`	Image cache root.
`--image-registry PREFIX`	none	Push the built image here, such as `ghcr.io/my-org/r2e`.
`--platform`	`linux/amd64`	`linux/amd64` or `linux/arm64`.
`--language LANG`	auto-detect	`python`, `node`, `go`, `rust`, `java` or `c_cpp`.
`--base-image IMAGE`	per language	Base image, such as `ubuntu:24.04`.
`--max-spend-usd N`	`5.0`	Abort when cumulative LLM cost passes this; `0` removes the cap.
`--force`	off	Ignore the cache and rebuild.

The agent's turn limit comes from --max-iterations here. Inside generate, it comes from the bootstrap spec (default 20; set it with --bootstrap-opt max_iterations=N). Bootstrap explains the cache and the agent loop.

repo2rlenv bootstrap --repo pallets/click --llm anthropic/claude-sonnet-4-6

pipelines

Discover native pipelines and research recipes. Makes no network calls and needs no credentials.

pipelines list

List every native pipeline and research recipe with its status (stable, experimental or planned) and source kind.

repo2rlenv pipelines list [--json]

Flag	Default	Description
`--json`	off	Print `{"native": […], "recipes": […]}`.

repo2rlenv pipelines list --json

pipelines describe

Show one recipe's input, scope, upstream source pin and RFC.

repo2rlenv pipelines describe --recipe RECIPE [--json] PIPELINE

Flag	Default	Description
`PIPELINE`	required	Pipeline the recipe belongs to, such as `repo_mutate`.
`--recipe RECIPE`	required	Recipe id, such as `swe_smith`.
`--json`	off	Print the description as JSON.

repo2rlenv pipelines describe repo_mutate --recipe swe_smith

tasksmith

Build Harbor tasks from merged PRs with a remote coding agent (Pi or OpenCode). Needs the tasksmith extra plus modal or daytona, a campaign and a runtime wheel. See Tasksmith.

tasksmith run

Run or resume a fixed PR panel.

repo2rlenv tasksmith run PANEL --campaign DIR --output DIR --runtime-wheel WHEEL [options]

Flag	Default	Description
`PANEL`	required	JSON panel of PRs, such as `examples/tasksmith-pr.json`.
`--campaign DIR`	required	Existing campaign directory.
`--output DIR`	required	Run directory. Rerunning the same command reuses matching completed work.
`--runtime-wheel WHEEL`	required	Wheel built from the installed version (`uv build --wheel`).
`--options FILE`	none	Options JSON, including quality models and spending limits, such as `examples/tasksmith-options.json`.
`--provider`	`modal`	`modal` or `daytona`. Used when `--options` is absent; must agree with it otherwise.
`--author`	`pi`	`pi` or `opencode`. Used when `--options` is absent; must agree with it otherwise.
`--stop-after N`	none	Process only the first N inputs. The report's denominator stays the full panel.
`--sources FILE`	none	Frozen PR source records, in panel order.
`--prepared-task DIR`	none	Recover one task matching `--sources` with fresh quality checks; the input is preserved.
`--generation-run DIR`	none	Reuse verified generation artifacts from an identical panel and rerun only the current quality policy.
`--reuse-evidence`	off	With `--generation-run`: also reuse matching baseline, reference and solver evidence. Semantic probes still rerun.
`--bootstrap-report FILE`	none	Reuse the snapshot and dependency hints from a completed CPU bootstrap report.
`--env-file FILE`	none	Load credentials from this file, without overriding variables already set.
`--json`	off	Print the report as JSON.

Exits 0 only when every input produced a usable task.

repo2rlenv tasksmith run examples/tasksmith-pr.json \
  --options examples/tasksmith-options.json \
  --campaign workspace/tasksmith \
  --output workspace/tasksmith/run \
  --runtime-wheel dist/repo2rlenv-0.9.3-py3-none-any.whl

tasksmith batch

Run a bounded parallel campaign over many PRs, retaining every generated task. The plan format is in Run Tasksmith on many PRs.

repo2rlenv tasksmith batch PLAN --campaign DIR --output DIR --runtime-wheel WHEEL [--env-file FILE] [--json]

Flag	Default	Description
`PLAN`	required	Batch plan JSON (target, parallelism, spending cap, candidates).
`--campaign DIR`	required	Existing campaign directory.
`--output DIR`	required	Batch directory.
`--runtime-wheel WHEEL`	required	Runtime wheel.
`--env-file FILE`	none	Credentials file.
`--json`	off	Print the batch report as JSON.

Exits 0 when the plan's verified target was reached.

repo2rlenv tasksmith batch plan.json --campaign workspace/campaign \
  --output workspace/tasksmith-batch --runtime-wheel dist/repo2rlenv-0.9.3-py3-none-any.whl

tasksmith bootstrap

Build and smoke-test pinned repositories on remote CPU and GPU workers before a campaign.

repo2rlenv tasksmith bootstrap MATRIX --campaign DIR --output DIR --runtime-wheel WHEEL
                               [--resource cpu|gpu|both] [--env-file FILE] [--json]

Flag	Default	Description
`MATRIX`	required	JSON list of distinct repository bootstrap specs.
`--resource`	`both`	`cpu`, `gpu` or `both`. The GPU pass stops at the first repository that isn't ready.
`--campaign DIR`	required	Existing campaign directory.
`--output DIR`	required	Receipts, logs and compute estimates.
`--runtime-wheel WHEEL`	required	Runtime wheel.
`--env-file FILE`	none	Credentials file.
`--json`	off	Print the reports as JSON.

repo2rlenv tasksmith bootstrap matrix.json --resource cpu --campaign workspace/campaign \
  --output workspace/bootstrap-matrix --runtime-wheel dist/repo2rlenv-0.9.3-py3-none-any.whl

tasksmith install-runtime

Install the pinned Pi and OpenCode libraries into your user cache. Takes no flags.

repo2rlenv tasksmith install-runtime

tasksmith show

Read a saved run or batch report. Makes no paid calls.

repo2rlenv tasksmith show [--json] OUTPUT

Flag	Default	Description
`OUTPUT`	required	Run or batch directory containing `report.json`.
`--json`	off	Print the report as JSON.

repo2rlenv tasksmith show workspace/tasksmith/run --json

campaign

Create and inspect the budget ledger (budget.sqlite3) that research recipes, Tasksmith, workers and the quality loop reserve spend against. See Remote execution.

campaign init

Create a campaign directory with a fixed spending limit. Running it again with the same amount is harmless; a different amount is refused.

repo2rlenv campaign init --budget-usd N [--json] PATH

Flag	Default	Description
`PATH`	required	Campaign directory.
`--budget-usd N`	required	Spending limit in USD.
`--json`	off	Print the budget as JSON.

repo2rlenv campaign init workspace/my-campaign --budget-usd 25

campaign status

Show the limit, accounted spend, open reservations and remaining allowance. Warns when operations are waiting for reconciliation.

repo2rlenv campaign status [--json] PATH

Flag	Default	Description
`PATH`	required	Campaign directory created with `campaign init`.
`--json`	off	Also list every operation with its status (`reserved`, `uncertain` or `settled`).

repo2rlenv campaign status workspace/my-campaign --json

campaign settle

Close a reservation with its actual cost and an evidence file. The ledger records the file's path and SHA-256.

repo2rlenv campaign settle --operation ID --cost-usd N --evidence FILE [--json] PATH

Flag	Default	Description
`PATH`	required	Campaign directory.
`--operation ID`	required	Operation id from `campaign status --json`. Workers use `worker:<provider>:<name>`.
`--cost-usd N`	required	Actual cost in USD.
`--evidence FILE`	required	Usage or billing receipt, or a clearly labeled conservative estimate.
`--json`	off	Print the budget as JSON.

repo2rlenv campaign settle workspace/my-campaign --operation worker:modal:my-worker \
  --cost-usd 0.30 --evidence workspace/my-campaign/worker-usage.json

workers

Create, check and terminate remote Modal or Daytona workers. Each worker has a JSON receipt at CAMPAIGN/workers/NAME.json that later commands take as input. Needs the modal or daytona extra.

workers start

Reserve spend in the campaign, create the worker and write its receipt.

repo2rlenv workers start --campaign DIR --provider modal|daytona --name NAME --reserve-usd N
                         [--cpus N] [--memory-mb N] [--timeout-sec N] [--json]

Flag	Default	Description
`--campaign DIR`	required	Existing campaign directory.
`--provider`	required	`modal` or `daytona`.
`--name NAME`	required	Lowercase letters, digits and hyphens, starting with a letter. Also the receipt's file name, so it can't be reused in the same campaign.
`--reserve-usd N`	required	Amount to reserve for the worker's lifetime.
`--cpus N`	`2`	CPUs (1–16).
`--memory-mb N`	`4096`	Memory in MB (1024–65536).
`--timeout-sec N`	`3600`	Lifetime in seconds (60–14400). On Daytona this becomes an idle auto-stop.
`--json`	off	Print the receipt as JSON.

repo2rlenv workers start --campaign workspace/my-campaign --provider modal \
  --name my-worker --reserve-usd 3 --timeout-sec 3600

workers probe

Check a running worker: Docker daemon readiness, byte-exact file transfer, an image build and an isolated container run with no network access.

repo2rlenv workers probe --out DIR [--json] RECEIPT

Flag	Default	Description
`RECEIPT`	required	Receipt of a running worker.
`--out DIR`	required	New directory for `probe.json` and command logs. It must not exist.
`--json`	off	Print the result as JSON.

repo2rlenv workers probe workspace/my-campaign/workers/my-worker.json \
  --out workspace/my-campaign/probes/first

workers stop

Terminate the worker and mark its reservation for reconciliation. Stopping confirms cleanup; it doesn't record a cost. Settle it with campaign settle.

repo2rlenv workers stop [--json] RECEIPT

Flag	Default	Description
`RECEIPT`	required	Worker receipt. Stopping an already terminated worker is a no-op.
`--json`	off	Print the receipt as JSON.

repo2rlenv workers stop workspace/my-campaign/workers/my-worker.json

release

Publish an explicit, immutable selection of tasks as a Hub dataset, checked against each task's bundle hash. Only research-recipe and Tasksmith bundles record a bundle_hash (and the recipe the plan names); native output doesn't, so publish it with push. See Releasing Harbor task collections.

release stage

Check each selected task against its expected bundle hash, then build the release directory: tasks, tasks.tar.gz, manifests, index and dataset card. Needs the harbor extra, which parses every task.

repo2rlenv release stage --out DIR [--json] PLAN

Flag	Default	Description
`PLAN`	required	Release plan JSON.
`--out DIR`	required	New staging directory. It must not exist or sit inside a selected task.
`--json`	off	Print the result as JSON.

repo2rlenv release stage release-plan.json --out workspace/releases/swe-smith

release verify

Check that every staged file still matches its recorded hash and mode. Runs nothing.

repo2rlenv release verify [--json] DIRECTORY

Flag	Default	Description
`DIRECTORY`	required	Staging directory.
`--json`	off	Print the result as JSON.

repo2rlenv release verify workspace/releases/swe-smith

release publish

Upload a verified release, pin registry.json to the upload commit and optionally add the dataset to a collection. Needs HF_TOKEN or a Hub login.

repo2rlenv release publish --receipt FILE [--collection SLUG] [--batch-size N | --recover-empty] [--json] DIRECTORY

Flag	Default	Description
`DIRECTORY`	required	Verified staging directory.
`--receipt FILE`	required	Publication receipt. A completed receipt returns without uploading again; an incomplete one stops for reconciliation.
`--collection SLUG`	none	Collection to add the dataset to.
`--batch-size N`	none	Upload to a new, empty repository in commits of N files (1–500) instead of one large commit.
`--recover-empty`	off	Recover an upload that timed out before creating a commit. Confirms the repository is still empty, then uploads in bounded commits.
`--json`	off	Print the receipt as JSON.

repo2rlenv release publish workspace/releases/swe-smith \
  --receipt workspace/releases/swe-smith-publication.json

quality

Review a Harbor task against evidence, and optionally run controls, probes and a blind solver remotely and repair the task. See Review and repair.

quality run

repo2rlenv quality run TASK --campaign DIR --out DIR [--repair] [--run-rollout] [options]

A run without --repair or --run-rollout makes model calls but creates no worker. Those two flags need a remote worker: --runtime-wheel for CPU tasks, or native Modal for tasks that request GPUs. They also accept only single-container tasks with network_mode = "no-network", such as research-recipe and Tasksmith bundles. Native pipeline output stops with Remote validation currently requires a single-container no-network Dockerfile task.

Flag	Default	Description
`TASK`	required	Harbor task directory.
`--campaign DIR`	required	Existing campaign directory.
`--out DIR`	required	Run directory for `result.json`, revisions, trials and receipts.
`--env-file FILE`	none	Load credentials from this file. They're never copied into task artifacts.
`--baseline PATH`	none	Existing trial evidence: an owned trial receipt, a Harbor trial `result.json`, or a single-trial directory.
`--oracle PATH`	none	Same, for the reference solution.
`--rollout PATH`	none	Same, for a solver rollout.
`--probes FILE`	none	Task-bound JSON manifest of known counterexamples and valid alternatives.
`--review-model PROVIDER/MODEL`	`anthropic/claude-sonnet-4-6`	Reviewer. Only `openai/` and `anthropic/` routes are accepted, for this and the other model flags.
`--repair-model PROVIDER/MODEL`	the reviewer	Repair author.
`--escalation-model PROVIDER/MODEL`	none	One stronger attempt when a review stays unresolved.
`--solver-model PROVIDER/MODEL`	`anthropic/claude-sonnet-4-6`	Blind solver.
`--repair`	off	Author repairs and validate each new revision remotely.
`--run-rollout`	off	Run missing controls, probes and a blind solver remotely, without editing.
`--max-repairs N`	`3`	Repair rounds (0–5).
`--max-read-rounds N`	`2`	Extra file-reading rounds (0–4).
`--max-probes N`	`2`	Semantic probes (0–4).
`--context-chars N`	`100000`	Document context per call (16000–250000).
`--model-tokens N`	`6000`	Output tokens per review or repair call (1024–16000).
`--max-turns N`	`24`	Solver turns (1–100).
`--solver-tokens N`	`4096`	Solver output tokens per call (256–8192).
`--trial-timeout-sec N`	`900`	Timeout per trial (30–3600).
`--model-reservation-usd N`	`1.00`	Reservation per model call.
`--solver-reservation-usd N`	`4.00`	Reservation per solver rollout.
`--max-spend-usd N`	`15.00`	Spending limit for this run, inside the campaign limit.
`--success-reward N`	`1.0`	Reward that counts as a solver success.
`--provider`	`modal`	`modal` or `daytona` for CPU workers. GPU tasks always use native Modal.
`--runtime-wheel WHEEL`	none	Required for CPU `--repair` and `--run-rollout`.
`--worker-receipt RECEIPT`	none	Reuse a running worker from the same campaign. You stay responsible for stopping it.
`--worker-reservation-usd N`	`3.00`	Reservation for a worker the run creates.
`--resume`	off	Reuse completed, identical model requests and trial evidence.
`--json`	off	Print the result as JSON.

Exits 0 for usable and reviewed, 1 for other dispositions. Automation that needs validated tasks should check status == "usable", not the exit code.

repo2rlenv quality run ./tasks/example \
  --campaign workspace/my-campaign \
  --out workspace/reviews/example \
  --run-rollout --provider modal \
  --runtime-wheel dist/repo2rlenv-0.9.3-py3-none-any.whl

quality show

Read a completed quality result. Makes no paid calls.

repo2rlenv quality show [--json] PATH

Flag	Default	Description
`PATH`	required	`result.json`, or a run directory containing it.
`--json`	off	Print the result as JSON.

repo2rlenv quality show workspace/reviews/example

tasks

Inspect and set the advisory evaluation label stored in each task's [metadata.repo2env.evaluation]. See Evaluation labels.

tasks show

Show one task's label.

repo2rlenv tasks show [--json] PATH

Flag	Default	Description
`PATH`	required	Task directory.
`--json`	off	Print the label as JSON.

repo2rlenv tasks show ./tasks/example

tasks list

List the labels of every task under a directory.

repo2rlenv tasks list [--status STATUS] [--json] PATH

Flag	Default	Description
`PATH`	required	Dataset directory.
`--status`	all	Keep only `unverified`, `verified`, `needs_repair` or `blocked`.
`--json`	off	Print the labels as JSON.

repo2rlenv tasks list ./tasks --status needs_repair

tasks label

Write a labeled copy of a task to a new directory. The original is left untouched.

repo2rlenv tasks label --out DIR [--quality-result FILE | --status STATUS] [--stage STAGE]
                       [--reason-code CODE ...] [--detail TEXT] [--provenance P] [--json] TASK

Flag	Default	Description
`TASK`	required	Task directory.
`--out DIR`	required	New task directory.
`--quality-result FILE`	none	Derive the label from a saved quality `result.json`. This is the only way to set `verified`.
`--status`	`unverified`	`unverified`, `needs_repair` or `blocked`. Can't be combined with `--quality-result`.
`--stage`	`generation`	`generation`, `bootstrap`, `construction`, `review`, `controls`, `probes`, `rollout`, `repair`, `complete` or `unknown`.
`--reason-code CODE`	`validation_not_run`	Repeatable snake_case diagnosis. `needs_repair` and `blocked` require at least one, plus `--detail`.
`--detail TEXT`	none	Human-readable reason and next diagnostic step.
`--provenance`	`unknown`	`unknown`, `assisted` or `unattended`.
`--json`	off	Print the new label as JSON.

repo2rlenv tasks label ./tasks/example --out ./labeled/example \
  --quality-result workspace/reviews/example/result.json

codemidas

Commands specific to the CodeMidas recipe. Both print JSON.

codemidas source

Validate a pinned Stack v3 repository manifest and print its provenance. Executes nothing.

repo2rlenv codemidas source [--materialization inline|hydrated] MANIFEST

Flag	Default	Description
`MANIFEST`	required	Stack v3 row manifest JSON.
`--materialization`	`inline`	`inline` uses the row's files; `hydrated` restores the GitHub checkout at the same commit.

repo2rlenv codemidas source workspace/stack-row.json --materialization inline

codemidas audit

Run adversarial and solver audits on a generated CodeMidas task, plus a separate curriculum screen.

repo2rlenv codemidas audit TASK --controls DIR --out DIR --campaign DIR
                           --worker-receipt RECEIPT --runtime-wheel WHEEL [options]

Flag	Default	Description
`TASK`	required	Generated task directory.
`--controls DIR`	required	The candidate's generation directory, whose control receipts are reused.
`--out DIR`	required	Audit directory.
`--campaign DIR`	required	Campaign directory.
`--worker-receipt RECEIPT`	required	Running worker receipt.
`--runtime-wheel WHEEL`	required	Runtime wheel.
`--screen-attempts N`	`4`	Curriculum screening attempts.
`--attempt-concurrency N`	`2`	Independent solver attempts in parallel (1–4).
`--max-cost N`	`6`	Trial allowance in USD. Independent review has its own maximum of $1.25.
`--resume`	off	Reuse unchanged attempts.

repo2rlenv codemidas audit workspace/codemidas/tasks/TASK \
  --controls workspace/codemidas/runs/RUN/tasks/CANDIDATE \
  --campaign workspace/codemidas \
  --worker-receipt workspace/codemidas/workers/worker.json \
  --runtime-wheel dist/repo2rlenv-0.9.3-py3-none-any.whl \
  --out workspace/codemidas/audits/TASK

CLI reference

Global flagsCopy Anchor Link

ConventionsCopy Anchor Link

generateCopy Anchor Link

validateCopy Anchor Link

pushCopy Anchor Link

pullCopy Anchor Link

bootstrapCopy Anchor Link

pipelinesCopy Anchor Link

pipelines listCopy Anchor Link

pipelines describeCopy Anchor Link

tasksmithCopy Anchor Link

tasksmith runCopy Anchor Link

tasksmith batchCopy Anchor Link

tasksmith bootstrapCopy Anchor Link

tasksmith install-runtimeCopy Anchor Link

tasksmith showCopy Anchor Link

campaignCopy Anchor Link

campaign initCopy Anchor Link

campaign statusCopy Anchor Link

campaign settleCopy Anchor Link

workersCopy Anchor Link

workers startCopy Anchor Link

workers probeCopy Anchor Link

workers stopCopy Anchor Link

releaseCopy Anchor Link

release stageCopy Anchor Link

release verifyCopy Anchor Link

release publishCopy Anchor Link

qualityCopy Anchor Link

quality runCopy Anchor Link

quality showCopy Anchor Link

tasksCopy Anchor Link

tasks showCopy Anchor Link

tasks listCopy Anchor Link

tasks labelCopy Anchor Link

codemidasCopy Anchor Link

codemidas sourceCopy Anchor Link

codemidas auditCopy Anchor Link

On this page