Skip to content

RFC 0021: cli_gym recipe for env_repair

Status: experimental implementation in PR #109; release evidence below Author: @adithya-s-k Created: 2026-09-11

Summary

Start with a healthy development workspace and its nonempty test identities. Apply a reproducible environment-breaking change, construct a restoration instruction, and retain a working setup restoration oracle. Grade against trusted original tests.

Motivation

Bring the released method into Repo2RLEnv as owned, maintainable code, with standalone Harbor output and independent quality evidence. The upstream project remains the attribution and comparison baseline, not a runtime dependency. Shared operations use RFC 0011; method-specific generation decisions stay in this recipe.

Design

Input

Native input: Healthy SWE-smith repository images and upstream destruction settings.

Select pipeline.name: env_repair and pipeline.recipe: cli_gym in a typed configuration. Source data, resolved revisions, resource limits, model roles, random seeds and recipe options are recorded before spending. Strict options reject unknown keys. Existing native pipeline defaults remain compatible.

Algorithm

Start with a healthy development workspace and its nonempty test identities. Apply a reproducible environment-breaking change, construct a restoration instruction, and retain a working setup restoration oracle. Grade against trusted original tests.

Each substantive stage emits a typed progress event and an artifact-bound receipt. Classification separates source eligibility, infrastructure, oracle, verifier and solver failures. Repairs are bounded and invalidate affected downstream evidence.

Output

A complete Harbor bundle: instruction, task configuration, environment, reference entry point and trusted verifier. Metadata records recipe/version, input lineage, upstream source pin, adaptations, reward scale, image/asset digests and the complete task hash. Exported is distinct from accepted.

Verification

The linked pipeline guide specifies the implemented native generation checks. Generation exports and independent quality acceptance are separate: reference success does not establish verifier coverage or shortcut resistance. Shared review, repair and labeling contracts are defined in RFC 0027.

Anti-contamination

Learner-visible snapshots exclude the reference, future Git objects, credentials and private tests. Execute grading so learner code cannot inspect the private oracle. Prefetch pinned assets; enforce and probe the actual learner network policy. The prompt is not an access-control mechanism. Preserve legitimate source context rather than indiscriminately deleting it.

LLM use

Where the algorithm requires synthesis or review, use recorded role-specific models through the common metered client or agent adapter. Reference execution is deterministic. Cost includes failures, retries, bootstrap, cloud runtime and independent audits; unknown costs are not zero.

Yield and suitability

See the pipeline guide for supported inputs and measured economics for sample sizes, yield definitions and cost coverage. Results on a selected source profile do not imply universal input conversion.

Dependencies

Repository-owned recipe code, existing source/auth/LLM/bootstrap helpers, remote execution adapters and Harbor. Essential SDKs and ordinary libraries are permitted. No install/import/clone of the upstream research implementation at runtime. No dependency on ignored local reference folders or private pilot artifacts.

Alternatives considered

A wrapper around upstream commands would preserve an uncontrolled runtime dependency. One generic generator for every method would lose method-specific behavior. Use owned stages with common execution and quality contracts instead; explicitly version deviations from the upstream baseline.

References

Implementation

The owned implementation is in recipes/cli_gym/. It constructs the healthy repository image remotely, samples real test identities, authors an inversion goal, and iterates destruction/recovery scripts against execution. A separate author describes observed symptoms. The standalone export keeps the inversion out of the copied source and grades against required original test identities. The guide and example record supported scope. The twenty-task generation campaign precedes detailed root-runtime audits.

Current release evidence

25 published Harbor tasks. The pipeline walkthrough documents the implemented profile, actual model calls, bounded repairs and limitations. The release inventory records source diversity, scoped economics, artifact revisions and evaluation labels. Generation checks, independent review and blind solver success are separate claims.