v0.1.6

What's Changed

• fix(sandbox): collect each artifact under its name, not its filename by @Luolc in #121
• chore(release): 0.1.6 by @Luolc in #122

Full Changelog: v0.1.5...v0.1.6

v0.1.5

What's Changed

• fix(swebench_pro): guarantee a writable HOME in the eval script by @Luolc in #119
• chore(release): 0.1.5 by @Luolc in #120

Full Changelog: v0.1.4...v0.1.5

v0.1.4

What's Changed

• feat: suffix artifact names with their format by @Luolc in #113
• refactor: name a shared artifact for its file (patch.raw.diff) by @Luolc in #114
• feat(eval): make a grading explainable after the fact by @Luolc in #115
• refactor: name the agent stderr artifact .log, not .txt by @Luolc in #116
• feat: keep the execution's own outcome on both paths by @Luolc in #117
• chore(release): 0.1.4 by @Luolc in #118

Full Changelog: v0.1.3...v0.1.4

v0.1.3

What's Changed

• docs(review): audit the rollout outcome state space by @Luolc in #106
• docs(review): add the retry policy for a fair eval by @Luolc in #107
• feat: give run_rollout observers and run_unit_test eval_env by @Luolc in #108
• fix(swebench_pro): pin core.autocrlf=false in the eval script by @Luolc in #109
• docs(swebench_pro): record the autocrlf pin as a divergence from the reference by @Luolc in #110
• fix(swebench_pro): also pin core.eol=lf in the eval script by @Luolc in #111
• chore(release): 0.1.3 by @Luolc in #112

Full Changelog: v0.1.2...v0.1.3

v0.1.2

What's Changed

• refactor(rollout): inject the harness and proxy into run_rollout by @Luolc in #98
• feat(persist): multi-rollout run-record layout (ADR-0004, task 16) by @Luolc in #99
• feat(harness): let the caller inject env into Harness.run by @Luolc in #100
• feat(harness): HarnessOutcomeObserver collects byproducts + completion by @Luolc in #101
• refactor(sandbox): contribute derived artifacts inline, no sandbox round trip by @Luolc in #102
• feat(harness): namespace a harness's native artifacts by harness name by @Luolc in #103
• chore(release): 0.1.2 by @Luolc in #105

Full Changelog: v0.1.1...v0.1.2

v0.1.1

What's Changed

• docs(adr): ADR-0003 — remote-sandbox support (up-first lifecycle + sb.fs) by @Luolc in #58
• docs(adr): revise ADR-0003 after design review (merge Sandbox/Backend; defer transfer seam) by @Luolc in #59
• docs(adr): keep Mount/Mounts; "asset" is a wording convention by @Luolc in #60
• docs(adr): simplify Resource extensibility (subclass both axes, import-only) by @Luolc in #61
• docs(adr): scope out shipping a remote backend (internal users own theirs) by @Luolc in #62
• docs(plan): task-14 — merged Sandbox + up-first lifecycle + transfer seam by @Luolc in #63
• refactor(sandbox): merge Sandbox + backend into one lifecycle object (task 14) by @Luolc in #64
• refactor(patch): move patch.py out of sandbox/ to a top-level utility by @Luolc in #65
• feat(claude_code): add --bare flag; drop CLAUDE.md-disable env by @Luolc in #66
• docs: purge stale references to deleted/moved modules by @Luolc in #67
• refactor: rename solve.py -> rollout.py (its intended post-cutover name) by @Luolc in #68
• refactor(w1): relocate trace.py into pipelines/related_files as exchange.py by @Luolc in #69
• refactor(w1): unify capture vocabulary on the Capture enum by @Luolc in #70
• docs: single source of truth for task status by @Luolc in #71
• docs: reconcile spec.md body with ADR-0003 by @Luolc in #72
• docs: add doc-map + single-source-of-truth guards by @Luolc in #73
• chore: delete vendored .claude/commands + add stale-reference CI guard by @Luolc in #74
• docs(plan): task-12 persistence design (Store seam + post-run persist + manifest) by @Luolc in #75
• feat(sandbox): task-12 Store seam + post-run persist/promote/index by @Luolc in #76
• feat(cli): task-12 persistence wiring — --persist + promote by @Luolc in #77
• refactor(cli): rename _persist -> persist_wiring (it's imported, not private) by @Luolc in #78
• refactor: move verify out of cli/ into datasets/swebench_pro/ (dataset-QA) by @Luolc in #79
• refactor(cli): dataset-agnostic eval/rollout via the TaskInstance interface by @Luolc in #80
• feat(verify): persist golden-verification to the T1 store, not git by @Luolc in #81
• refactor: TaskInstance.unit_test -> unit_test_spec, return only UnitTestSpec by @Luolc in #82
• refactor(swebench_pro): the instance owns its harness (run_script/parser/golden) by @Luolc in #83
• refactor: unit_spec -> unit_test_spec; add "don't shorten names" rule by @Luolc in #84
• refactor: inject the Sandbox into run_unit_test (dependency injection) by @Luolc in #85
• refactor: use instance.golden_test_checkout_cmd directly in _build_eval_script by @Luolc in #86
• docs: drop a redundant call-site comment in _build_eval_script by @Luolc in #87
• refactor(rollout): inject the sandbox into run_rollout by @Luolc in #88
• refactor(swebench_pro): break the record↔unit_test import cycle by @Luolc in #89
• docs: refresh stale symbol references across code + docs by @Luolc in #90
• refactor(swebench_pro): rename execution.py → auxiliary.py; inline image_ref by @Luolc in #91
• refactor(sandbox): drop the unused stream_to exec arg by @Luolc in #92
• refactor(sandbox): move workspace into SandboxConfig, off the factory seam by @Luolc in #93
• docs(adr): ADR-0004 multi-rollout run-record layout (pass@K) by @Luolc in #94
• refactor: adopt etils.epath for filesystem paths (loose in, strict out) by @Luolc in #95
• refactor: use epath's copy/rmtree instead of shutil by @Luolc in #96
• chore(release): 0.1.1 + document the release process by @Luolc in #97

Full Changelog: v0.1.0...v0.1.1

v0.1.0

First public release of swe-lab — tooling to build, run, and evaluate SWE-agent evaluation data (SWE-Bench Pro).

• A sandboxed-task engine with pluggable harness / dataset / eval-method axes; rollout and eval are two compositions of one engine.
• Two backends behind one Sandbox: A-host (Docker) and A-ghjob (the CI job is the container).
• CLI: swe-lab {eval,rollout,verify}.

Alpha — APIs may change.