`env_repair / cli_gym`¶

CLI-Gym derives repair tasks by deliberately breaking a healthy development environment and verifying that a recovery restores its existing tests.

Pipeline, step by step¶

flowchart TD
  S["Pinned repo + offline recovery dependencies"] --> B["Healthy image, tests and protected file hashes"]
  B --> P1["P1 · Inversion goal from sampled tests + direction"]
  P1 --> P2["P2 · Destruction and recovery scripts"]
  P2 --> E["Remote damage, protected-file check, tests, recovery, tests"]
  E -->|"No selected failure or recovery incomplete"| P2
  E --> P3["P3 · Symptom-based repair instruction"]
  P3 --> H["Independent Harbor rebuild"]
  H --> R["nop 0; restoration oracle 1"]
  R -->|"Failure feedback within max_rounds"| P2
  R -->|"Pass"| O["Export environment repair task"]

P1, P2, … identify actual model calls. Unlabelled stages are code or remote execution.

Establish a healthy system. Collect real passing test identities, installed packages and file paths, plus hashes protecting source/tests. The author samples at most fifty test identities per candidate.

Invert and restore. The goal specifies an environment failure. Scripts must change persistent filesystem state without editing protected repository code/tests. A valid inversion affects at least one selected test; recovery restores the healthy suite.

Describe symptoms. The instruction author gets actual baseline results and a recovery goal, not a license to invent assertion failures. A fresh Harbor rebuild verifies the packaged destruction/restoration path.

Every prompt and its data¶

One goal call, then up to max_rounds inversion calls. Instruction writing occurs only for a successful disruption/recovery attempt.

Call	System prompt composition	User / input material	Output	Retry or branch
P1 · Goal	inversion_prompt.md with candidate_uts_list, directions and existing_tasks substitutions	Candidate, installed packages and file paths.	InversionGoal: title, category, selected_tests, description, expected_result, recovery_strategy	Existing titles discourage repeats within a run.
P2 · Inversion / repair	Inline system prompt in CLIGymPipeline.author_export	Goal, observed healthy environment and feedback.	Inversion: destruction_shell, recovery_shell, explanation	Up to max_rounds, default three.
P3 · Instruction	instruction_prompt.md with task_description and symptoms_UTs + adaptation	Actual baseline and goal.recovery_strategy.	RepairInstruction: instruction	Called after a valid contrast; repeated if a later Harbor failure returns to the loop.

Read the complete cli_gym prompt reference for every retained template, appended instruction, substitution, example and output schema. The shared prompt guide explains how to inspect the fully resolved request from a real run.

Follow one task¶

Illustration: a Python path configuration causes the healthy CLI package to stop importing. The learner sees the failure and offline wheel cache, and must restore the environment. The source implementation itself must remain unchanged.

What repeats, what is checked¶

P1 is fixed for the candidate. P2 receives destruction/recovery failures and later Harbor feedback. The final task runs as root because it is an environment repair problem; its isolation/shortcut risks still require the later quality review. Neither empty tests nor incomplete recovery is a successful generation.

An exported bundle is a generation result. Independent leakage review, shortcut probes and blind solver traces belong to the later quality campaign.

Implementation map¶

Run and supported profile¶

Run repo2rlenv generate --config examples/owned-cli-gym.yaml. This first profile supports public GitHub Python repositories with a tests/ directory. The configuration supplies source paths, test/build dependencies, offline recovery assets and optional disruption directions. The target repository is cloned in the remote bootstrap; the CLI-Gym research repository is never a runtime dependency.

The inversion author samples up to fifty real passing test identities, proposes a distinct environment failure, then revises its scripts using actual execution feedback. Source and test files are protected. Changes must persist in files; temporary shell exports cannot represent a separate environment state.

The baseline must lose at least one selected test and the recovery must restore all required healthy tests. Empty, malformed and incomplete results cannot earn success. A broken Python import or collection step is a valid environment failure, but is reported accurately. The final Harbor bundle undergoes another fresh baseline/reference pair before export.

The solver repairs the container as root, offline, without changing repository source or tests. The example makes dependency wheels available in /opt/wheelhouse. The exported image preinstalls tmux so Harbor's Terminus-2 agent can start without downloading terminal tooling. Baseline/reference success alone does not test this agent setup path; include a blind solver run in the quality pilot. The build-time destruction script stays outside the learner filesystem; its inverse remains a private reference. Full adversarial review of the root runtime is deferred to the later quality campaign.

Options include target, max_candidates, max_rounds, seed, directions and the common Python build/test profile. The target remains at least 20 tasks at the user's request; all 25 generated tasks are retained in the published release. See the release inventory and shared Modal/Daytona and progress interface.

Credit: CLI-Gym, MIT, commit 48bb920b728a25a55a5b442303e901919654599e. See RFC 0021 and packaged recipes/cli_gym/provenance.md for source mapping and profile restrictions.

env_repair / cli_gym¶