Version history

Current release highlights followed by historical notes. CLAUDE.md carries only the compressed summary; the detail lives here.

For per-release deep dives see the sibling pages (v0.8.2.post3.md, v0.8.3/).

v0.9.3: FrontierSmith optimization synthesisCopy Anchor Link

Released September 29, 2026.

Adds the experimental optimization_synth / frontiersmith recipe: original seed problems become standalone Harbor optimization tasks with deterministic continuous rewards. OpenAI authors and reviews formulations, independent solution samples, and private test infrastructure; Daytona runs construction checks and optional blind rollouts. Bounded repair and durable API receipts retain uncertain-call reservations without blindly redispatching requests.

The local collection contains 100 tasks across 20 problem families, selected from 153 candidate attempts (65.4% yield). All passed construction checks; blind rollouts completed with fully feasible solutions on 21 of 24 tested tasks, with three command timeouts. All tasks remain unverified, stage construction; sampled references are not proven optima. Accounted generation, review and compute cost was $68.42, with $0.67 reserved for unresolved API responses. These are estimates, not reconciled invoices. The task collection remains local and is not included in the package or published to the Hub.

The pipeline guide includes diagrams, prompt contracts, costs and validation limits. RFC 0032 records the differences from FrontierSmith: this is a credited, independent Python adaptation, not the withheld upstream orchestrator or a reproduction of its training results. No upstream research package is required.

Also includes runtime fixes for Mocha spec-reporter parsing (#162) and validation of test output written to stderr (#164), plus the setup-uv workflow update (#169). Existing generated datasets are unchanged; regenerate affected runtime tasks and rerun their baseline/oracle checks to adopt the parser fixes.

Upgrade with pip install --upgrade repo2rlenv==0.9.3, adding the extras your pipeline needs. Python 3.12–3.14 remain supported; use Linux, macOS or WSL for research generation. The shared fixed-success-reward quality loop does not yet provide full acceptance for continuous optimization tasks.

v0.9.2: CodeMidas and runtime correctnessCopy Anchor Link

Released September 26, 2026.

Adds the experimental codemidas recipe under repo_reconstruct: reconstruct existing functionality from pinned GitHub source or inline Stack v3 repository rows, using Daytona execution and GPT-6 Luna/Sol authoring and review. Outputs are standard Harbor tasks with deterministic verifiers. Includes source provenance, bounded repairs, independent rollout review, difficulty screening, and portable JSON release evidence. RFC 0031 and the pipeline guide document the method and full prompts.

The 100-task local collection passed 400 oracle and 200 failing-baseline controls. All tasks retain blocked labels because adversarial checks could not run; no full-method acceptance is claimed. Sol solved 94 tasks at least once across four attempts each. The collection remains unpublished. See the dataset audit and release notes for sources, difficulty, costs and limitations.

This release also includes fixes merged since 0.9.1:

• PR-diff tasks deliver the oracle and verifier through Harbor's private mounts, keeping reference artifacts out of learner images (#145). Diff normalization ignores Git extended headers and mode-only lines (#137); the judge supports OpenAI-compatible endpoints (#135).
• Runtime test parsing handles Go subtests, verbose Vitest, pytest-xdist output, and Rust should-panic/doctest results (#127, #143, #151, #139).
• Source links and forge blocking respect the repository's source host (#133, #153). Windows fixes cover ownership changes and CRLF patch application; TMax reads UTF-8 explicitly (#136, #141).
• Dependency and workflow updates include the patched AnyIO minimum, coding-agent runtime updates and tighter GitHub Actions permissions (#154, #147, #131, #159).

Upgrade with pip install --upgrade --upgrade-strategy eager repo2rlenv==0.9.2, adding your required extras. Python 3.12–3.14 remain supported. Examples point to the 0.9.2 runtime wheel. Existing generated datasets are not rewritten: to adopt changed verifier, parser or image behavior, regenerate affected tasks and rerun baseline/oracle checks before publishing a new dataset revision. Full native Windows research-pipeline execution remains unsupported; use Linux, macOS or WSL.

v0.9.1: Windows CLI startup and release checksCopy Anchor Link

Released September 15, 2026.

Fixes the Windows CLI crash introduced in 0.9.0: even --version and --help previously failed while importing the POSIX-only fcntl module. A shared standard-library lock now uses flock on POSIX and byte-range locking on Windows, preserving exclusive controller ownership and the blocking history checkout lock. Thanks to KNambiarDJsc for the report and initial fix (#128, #129).

CI and publication now require fresh Windows wheel checks on Python 3.12–3.14: CLI startup, recipe discovery, native UTF-8 task emission/static validation and real process-lock contention. Full native Windows Tasksmith, research-recipe and quality-controller execution remains unsupported; use Linux, macOS or WSL. The remaining artifact-permission and process-cleanup work is tracked in #130.

This release also includes the Python, documentation/build-tool and Pi/OpenCode runtime dependency updates merged after 0.9.0. Upgrade with pip install --upgrade --upgrade-strategy eager repo2rlenv==0.9.1, including any extras your pipeline needs. The Windows CLI/locking fix does not require changes to existing generated datasets. Versioned examples use the 0.9.1 runtime wheel.

v0.9.0: Tasksmith and owned generation recipesCopy Anchor Link

Released September 15, 2026.

Tasksmith and 14 research-inspired generation recipes now ship alongside the six native pipelines. Tasksmith uses LangGraph with Pi or OpenCode to investigate a merged PR, bootstrap its repository, design a task and private verifier, and run bounded review and repair. CPU execution supports Daytona and Modal; the implemented GPU route uses Modal L4 GPUs. Tasksmith and the recipes remain experimental; their current scope is documented in the pipeline guide.

The owned recipes cover repository mutation, PR and commit mining, function reconstruction, terminal synthesis, task evolution, environment repair and reasoning instances. CLI discovery exposes each method's options and source provenance. Campaigns record budgets, resumable work and execution evidence; the shared quality loop reviews Harbor bundles and learner traces and writes explicit evaluation labels. Generation alone does not establish task quality.

The README now provides a shorter quickstart and complete route table. Detailed guides document stage diagrams, exact prompts, dataset validation scope and measured yield and costs. Versioned examples point to the 0.9.0 runtime wheel.

Python 3.12–3.14 remain supported. To upgrade an existing environment, use pip install --upgrade --upgrade-strategy eager repo2rlenv==0.9.0, adding any extras your pipeline needs.

Dependency maintenance updates the locked HTTP clients and LiteLLM to patched versions, with security minimums for direct dependencies and transitive lock constraints. Existing pip environments should upgrade transitive dependencies as well; uv constraints are not included in wheel metadata. Dependabot now tracks the Python and coding-agent locks alongside GitHub Actions.

Native generation preserves parametrized pytest IDs containing spaces in both test discovery and the copied runtime verifier. GitLab tasks use the source host for clone URLs and commit references. Hub publishing accepts both flat datasets and tasks/<id> layouts, and Windows generation uses portable paths and explicit UTF-8 output.

validate --deep checks task assets and metadata without running a sandbox. --oracle also checks reference-solution assets; named recipes can provide a solve script without a patch. These checks do not establish oracle success or task quality. Use Harbor execution and the quality workflow for that evidence.

Generated datasets retain the code they were emitted with. Updating the package does not replace existing tests/verifier.py files or repair saved F2P/P2P lists. For an affected dataset, recover complete test IDs from the original test results, update the verifier, and rerun baseline and oracle checks before publishing a new revision. Metadata-only reference corrections do not change rewards; clone-URL corrections need a new image build. Preserve earlier dataset revisions and keep validation claims tied to the revision actually tested.

generate and bootstrap now accept --llm-endpoint and --llm-key-env. Keyless self-hosted providers can use LiteLLM's native credential handling.

Existing configurations with a custom llm.endpoint must explicitly name llm.api_key_env to send a hosted provider's default key there. For example, an authenticated OpenAI-compatible gateway that previously used OPENAI_API_KEY implicitly now needs --llm-key-env OPENAI_API_KEY or llm.api_key_env: OPENAI_API_KEY. The CLI warns when that default key is set but withheld. An explicitly named, unset key variable now raises an error instead of falling back to a different key. Default hosted endpoints are unchanged.

---

v0.1.0: first releaseCopy Anchor Link

pr_diff (originally pr_mining_lite) + HF Hub publish + diff-similarity reward.

v0.2: bootstrapCopy Anchor Link

Merged into main (not separately released): bootstrap phase, Rich UI module, cost tracking, content-addressed cache keyed on bootstrap options.

v0.3.0: sandbox verificationCopy Anchor Link

pr_runtime (sandbox-verified PR mining) + auto-trigger bootstrap from generate + structural quality filters + targeted test invocation + CI/CD (ruff + matrix tests + release workflow).

v0.4.0: polyglot + Harbor complianceCopy Anchor Link

Polyglot log parsers (Go / Cargo / Jest) + Harbor end-to-end compliance fixes: task.name format, solve.sh shim, /logs/verifier/reward.txt, PATH prelude for non-Python toolchains, defensive git install.

v0.5.0: commit-level miningCopy Anchor Link

pr_stream (continuous PR mining with watermark state) + commit_runtime (commit-level mining, SWE-GEN style). Both Harbor-verified.

pr_stream was removed in v0.8.3 as scope-creep: pr_runtime handles the same niche on its own.

v0.6.0: first LLM-synthesized pipelinesCopy Anchor Link

mutation_bugs (procedural AST bug injection, inspired by SWE-smith) + code_instruct (repo-anchored OSS-Instruct with executable verifiers, inspired by Magicoder). Both Harbor-verified on pallets/click (mean reward 1.000).

v0.7.0: function-level synthesis + CVEsCopy Anchor Link

equivalence_tests (R2E-style function-level synthesis: extract a real function, LLM writes an equivalence test against a reference_<name> oracle, gold patch fills the candidate) + cve_patches (OSV-driven CVE → fix-commit pipeline, reuses the pr_runtime validation harness). Both Harbor-verified.

v0.8.0: refactor miningCopy Anchor Link

refactor_synthesis (Python-native rename-refactor mining, dropping the v1.0-planned JVM RefactoringMiner dependency; commit-message regex + diff verification + multi-criteria structural+behavioral verifier). Harbor-verified on pallets/click (mean reward 1.000).

v0.8.3: pipeline auditCopy Anchor Link

Removed mutation_bugs + refactor_synthesis. Both were binary-reward, Python-only, and the lowest-signal pipelines in the set: synthetic AST bugs are unrealistic, and renames are a near-no-op RL target. pr_stream removed in the same pass.

Shared helpers (make_unified_diff, build_binary_eval_script) moved from mutation_bugs.py into pipelines/_eval_script.py.

v0.8.4: input sources + GitLabCopy Anchor Link

Input-source abstraction (GitHub / GitLab / local, capability-gated) + GitLab MR mining for pr_diff / pr_runtime (#62) + commit_runtime promoted to stable.

commit_runtime gained LLM-synthesized leak-free instructions (synthesize_with_llm, default on) + a max_pass_to_pass cap. Audit went 33% → 100% clean; Opus solves the sampled tasks (4/4 genuine). Reference dataset: …commit-runtime-test (100 envs).

Anti-contamination pass (PR #69)Copy Anchor Link

Sandbox-verified tasks were gameable: an agent could fetch the published fix (and the hidden tests) for the repo it was asked to fix. Now baked into the emitter for every task via pipelines/_env_guard.py:

• git-history scrub: strip the repo to base_commit: remove origin, prune future refs/commits, gc.
• egress guard: environment/docker-compose.yaml blackholes PyPI + GitHub hosts so pip download / git fetch / web-fetch fail. Model API and agent install stay up.

Also in this pass:

• cve_patches ships a graded F2P/P2P verifier (was binary whole-suite, which scored the gold patch 0.0 on unrelated suite failures).
• A leak-stripped instruction for cve_patches.
• Agentic PoC-test synthesis (_poc_agent.py): an LLM with shell access in the vulnerable sandbox writes a regression test for no-test CVEs.
• bootstrap re-bootstraps when a cache hit points at an evicted image.

Reference dataset …-cve-patches (19 verified envs).

The principle: the environment enforces, the prompt never asks.

v0.8.6: code_instruct self-improvementCopy Anchor Link

The v0.6 prompt explicitly forbade using repo APIs, producing generic Codeforces-lite tasks (baseline audit: mean repo-anchoring 1.4/5; zero of 20 tasks imported the target package).

Reoriented around genuine repo anchoring with four post-synthesis gates in _oss_instruct.py (check_repo_anchoring, check_symbol_collision, check_test_strength, task_fingerprints), plus max_attempts_per_seed retries (default 1 → 3). Post-fix audit: RA = 4.95, TR = 4.95, zero scores ≤ 2.

Multi-agent validation (claude-code + Sonnet 4.6, codex + GPT-5.3-Codex, openhands-sdk + Qwen3.6-35B via HF Router) surfaced a second bug: all three models correctly implemented the requested logic but wrote to natural filenames (ranged_float.py, fetcher.py), failing pytest from task_module import ... collection with ModuleNotFoundError. Fix: append a task_module.py delivery-contract paragraph to the emitted instruction.md. Solve rate 40% → 80% at fixed dataset size.

Reference dataset …-code-instruct (100 envs across 5 Python repos), the first published dataset for this pipeline. code_instruct stays experimental pending graded reward + polyglot support.

v0.8.7: equivalence_tests self-improvementCopy Anchor Link

The v0.7 pipeline had a full-source-in-instruction leak and a 97% Stage-B failure rate on click. Root cause: extracted functions referenced repo-internal types (Argument, FC) that don't exist in the standalone task_module.py, so the import crashed before pytest reached the LLM's assertions.

Landed: leak-free instruction (signature_only_source), annotation-strip at bake time (strip_annotations), scope-aware purity + self-containment filter in the extractor (_references_only_safe_names), is_module_importable pre-flight, recursion-safe rename (rename_function_ast), feedback-driven retry (max_attempts_per_function default 1 → 3, documented but not previously wired), test-strength gate (check_equivalence_test_strength), task dedup (_equivalence_fingerprint), a sharper prompt, .debug_skips dumps, and all_tests_passed moved to the shared _eval_script.py.

Reference dataset deferred. The 5-repo click/flask/requests/attrs/starlette survey yielded only ~8 pure candidates combined. Framework-heavy repos are structurally weak fits for equivalence testing. v0.8.7 is pipeline-hardening only; the dataset ships once we survey utility-heavy libs (packaging, itsdangerous, markupsafe, dateutil, …). equivalence_tests stays experimental.

v0.8.8: docs siteCopy Anchor Link

Version bump + docs site published, Google Search Console verification, README pointed at the live site, scripts/ untracked.

In progress: pr_to_env (RFC 0007)Copy Anchor Link

Designed, not yet on main. The import-shape sibling of pr_runtime: consumes an explicit list of curated PR URLs rather than mining a repo's history. One URL → one Harbor task, or fail closed with a per-URL reason. Same task shape, same graded F2P/P2P verifier, same anti-contamination guards; the reused machinery is imported verbatim from pr_runtime.py.

Ships with a 12-gate quality layer (M3) landing across milestones M1–M4; the gate list lives in RFC 0007.

PlannedCopy Anchor Link

• env_setup (RFC 0008): Repo2Run / SetupBench-style. The agent makes a bare repo's tests run green.
• test_synthesis (RFC 0009): SWE-Flow-style TDD.
• issue_runtime (RFC 0010).
• Graded rewards for the binary synthesis pipelines.
• LLM-judged QA gate.