osolmaz/aacr-bench-harbor

AACR-Bench Harbor

AACR-Bench Harbor packages AACR-Bench as independent Harbor tasks for code-review agents. Each task checks out one pull request at its pinned head commit, keeps the merge base of the recorded base and head as aacr-base, and grades a structured findings file with AACR-Bench's path, side, line, and semantic matching stages.

The release contains 190 of AACR-Bench's 196 positive-sample pull requests and 1,472 of its 1,505 reference comments. The other 6 pull requests are excluded because their exact head commits are gone from GitHub, the authors' forks, and Software Heritage.

This is an independent adaptation of AACR-Bench, with revised matching, configurable judges, and a separate AACR Judge Eval draft. The draft contains 236 comment pairs and model-proposed labels pending human approval. The implementation is independent of AACR-Bench and Harbor.

Install

Requirements:

• Python 3.12 or newer
• uv
• Git
• Harbor 0.21.0
• Docker for local task runs, or Harbor's hf-sandbox extra for Hugging Face Sandboxes

git clone https://huggingface.co/datasets/osolmaz/aacr-bench-harbor
cd aacr-bench-harbor
uv sync --locked --dev

The Hugging Face dataset repository is the single canonical repository. It contains the code and immutable source packs needed to build tasks without cloning upstream repositories. If a Git client leaves large files as pointers, hydrate the exact corpus revision:

uv run aacr-source-pack hydrate \
  --manifest source-packs/manifest.json \
  --revision 4e21a6fd5e821ef0764b3985658b8ff673081e03

Generate tasks

The default source is pinned to AACR-Bench commit b3072489eace26efca8bcf2b1ac6a24ba64f82c1 and SHA-256 d8683cb240249bc4e0aff6428802bdffa7b7573ace600552cab1cd0cb7e905c9.

Generate every released task from the local source packs:

uv run aacr-bench \
  --output-dir datasets/aacr-bench \
  --source-pack-manifest source-packs/manifest.json

--split selects full (the default), development, validation, or sealed-test. Generate one task by ID:

uv run aacr-bench \
  --output-dir datasets/aacr-bench \
  --source-pack-manifest source-packs/manifest.json \
  --task-ids appwrite-appwrite-pr-9999-710b8bda565c

Existing task directories are never replaced unless --overwrite is present. Task generation verifies the archive, commit, tree, diff, clean checkout, and absence of Git remotes before writing Harbor files. It performs no repository network access.

Maintainers can verify every released pack with:

uv run aacr-source-pack verify --manifest source-packs/manifest.json

Generated Harbor datasets are ignored by Git.

Run with Harbor

Verify generated tasks with the oracle first:

harbor run -p datasets/aacr-bench --agent oracle

Run GPT-5.6 Luna through Codex's native review harness. Put this repository on Python's import path so official Harbor can load the custom agent module:

PYTHONPATH="$PWD" harbor run \
  -p datasets/aacr-bench \
  --agent agents.codex_review:CodexReview \
  --model openai/gpt-5.6-luna

OPENAI_API_KEY must already be present in the invoking environment. The custom agent uses an isolated temporary CODEX_HOME, calls codex review --base aacr-base, validates the structured review, and writes /workspace/aacr-review.json outside the read-only review session.

For Hugging Face Sandbox execution, install the official harbor[hf-sandbox]==0.21.0 release and add:

-e agents.hf_sandbox:AACRHFSandboxEnvironment \
--ek flavor=cpu-basic \
--ek forward_hf_token=true \
--ek job_timeout=35m

The corrected DeepSeek runs keep the official Harbor package unchanged. A thin benchmark-owned environment adapter creates the configured task workdir from the image root and runs installed-agent commands with Bash, which Harbor's agent helper requires for pipefail. It uploads large task archives through public Sandbox file operations in 1 MiB chunks, with four concurrent writes and three bounded attempts per chunk. Artifact reads use three bounded transport attempts, and a transient status-poll transport error can recover for up to six 5-second polls without killing the review.

Run Pi Reviewer jobs through the AACR wrapper:

uv run aacr-harbor-run --config path/to/job-config.json

The wrapper leaves Harbor unchanged. It waits for Harbor to finish redacting trial files, then validates each final native Pi JSONL file and writes its mode-0600 receipt atomically. Judging runs in production by default and uses the selected judge's key from the invoking environment. Set AACR_JUDGE_MODE=mock to ask for diagnostic scoring explicitly, which needs no judge credential. A terminal trial without a valid native session makes the wrapper fail.

Select the Pi model through the Harbor agent entry and bind it to a reviewed manifest:

{
  "import_path": "agents.pi_reviewer:PiReviewer",
  "model_name": "huggingface/deepseek-ai/DeepSeek-V4.1-Flash:novita",
  "kwargs": {
    "model_manifest_path": "experiments/model-manifests/deepseek-v4.1-flash-novita.json",
    "model_manifest_sha256": "9003e7cd58eb29ce0f3cf5d1c4ee6a33e52aaec17c70b80e4b111925476f3169",
    "model_source_revision": "dba1be0a40aa45a94ad051997016db3960a90277",
    "expected_response_model": "deepseek/deepseek-v4.1-flash",
    "thinking": "high"
  }
}

The adapter rejects a model, provider route, manifest digest, or source revision mismatch. The complete reviewed configuration is in experiments/runner-manifests/deepseek-v4.1-flash-novita.json.

The benchmark image uses the official release tag node:24.7.0-bullseye. Its recorded OCI index digest is sha256:1684133274a44010e6d6011d6eec100cf756309678c77700779ac44a5ac36715, and its linux/amd64 manifest digest is sha256:14671502e8e7b000cd644c1ad56934bab4a82077457af6461d084fd5d0e85c05. HF Sandbox cannot execute Docker digest references reliably, so runs must attest that the release tag still resolves to those digests before launch.

Do not place credentials in task files, generated images, command arguments, logs, or run manifests.

The hosted task set

The tracked tasks directory holds all 190 released task definitions, metric.py, and split-manifest.json. Regenerate it from the source data and the pack manifest with one command:

uv run aacr-hosted-tasks

Never edit tasks by hand. scripts/check.sh runs uv run aacr-hosted-tasks --check, which fails when the tracked task set differs from a fresh generation.

A local task keeps its review checkout in environment/repo, and that checkout is itself a Git repository. Git records a nested repository as a gitlink, so a clone of this dataset would contain an empty repo directory. A hosted task therefore carries environment/aacr-source.json instead. The file names the task's immutable source pack, its SHA-256 and size, the recorded base commit, the merge base, and the head commit. The pack revision and dataset come from experiments/source-corpus.json. A reviewed harness fetches the pack from that revision, verifies its SHA-256, restores repo, and checks that HEAD is the head commit and aacr-base is the merge base before the agent starts. Hosted tasks use /root as their workdir because the pinned image has no /app.

The files in workbench/recipes run the two DeepSeek models through Pi Reviewer 0.2.0 and GPT-5.6 Terra, GPT-5.6 Luna, and GPT-6 Luna through Codex Review 0.147.0. Every recipe requests xhigh, pins its runner and harness, writes the AACR findings file, records native usage, and saves raw output and native session files in aacr-artifacts.tar.gz. Codex connects directly to the reviewed Responses API URL and reads its native rollout, so no request or response proxy is used.

Harbor owns task expansion, attempts, concurrency, retry behavior, and result storage. The Harbor-HF benchmark presets set three attempts for each task: all-tasks-3-trials covers all 190 released tasks, and sealed-test-3-trials covers the 57 sealed-test tasks. Keep these runs diagnostic while the scorer is under review, then rescore the stored findings without rerunning the agents.

Preset source

The presets directory makes this dataset the pinned preset source for AACR-Bench in Harbor-HF. The Harbor-HF catalog holds presets for very popular benchmarks only, so the reviewed AACR presets live here:

• presets/agents holds one agent-preset-v1 document for each reviewed review harness.
• presets/benchmarks holds the benchmark-preset-v1 documents that name the AACR tasks.

uv run python scripts/generate-run-configs.py generates the Workbench recipes and both preset directories from one setup script and one set of pins. Never edit them by hand; scripts/check.sh runs the generator with --check.

A control Space reads this directory at one exact commit, named in HARBOR_HF_PRESET_SOURCES:

[{"repository": "<namespace>/<dataset>", "kind": "dataset", "revision": "<40-character commit>", "path": "presets"}]

The service requires the commit to resolve to itself and every file to match the git object of that commit. It validates every file with the same preset schemas the Harbor-HF catalog uses. A file outside agents and benchmarks, a file that does not parse into a preset, a loose pin, and a preset identity that another source also provides all stop the service before it can launch anything. Keep the directory free of other files for the same reason.

A preset carries no credential value. The agent presets name a reviewed credential binding that the control service resolves in place.

Findings

Agents submit one strict JSON document. See the findings format.

The verifier writes raw expected, generated, line-match, and semantic-match counts plus per-task metrics. The dataset-level metric.py sums counts across tasks before computing precision, recall, and F1, matching AACR-Bench's micro aggregation. Missing rewards and judge errors fail closed and suppress headline aggregate scores.

Judge eval

Use the judge eval workflow to select up to 300 real comment pairs from saved reviews, collect blind annotations, and compare judges through the canonical verifier. The default is 240 pairs. Approved cases belong in a separate judge-eval configuration in this Dataset. Raw trial evidence stays in the existing private archive. Selection and label approval do not call a model.

Reference agreement measures agreement with the supplied notes, not verified bug precision. The pinned source also includes AI reference annotations. The adapter preserves that provenance and includes concrete maintainability and readability improvements in instructions for newly generated tasks.

Judge selection and replay

The canonical verifier supports JUDGE_NAME=luna (OpenAI Responses), JUDGE_NAME=jev (TypeSafe System One), and JUDGE_NAME=deepseek (Hugging Face Chat Completions through Novita). Set AACR_JUDGE_MODE=production and an explicit JUDGE_MODEL. For Luna, use JUDGE_MODEL=gpt-5.6-luna with OPENAI_API_KEY in the verifier process environment. For Jev, use JUDGE_MODEL=jev-1.13.0 with TYPESAFE_API_KEY and an explicit JUDGE_THRESHOLD, for example 0.5. This example threshold is not calibrated. Jev accepts a match when its finite probability is greater than or equal to the threshold. Text judges reject a threshold. No judge falls back to another provider or model.

For DeepSeek V4.1 Flash, set JUDGE_MODEL=deepseek/deepseek-v4.1-flash, JUDGE_REQUEST_MODEL=deepseek-ai/DeepSeek-V4.1-Flash:novita, and HF_TOKEN in the verifier process environment. JUDGE_MODEL pins the expected response model; JUDGE_REQUEST_MODEL pins the distinct Hub request ID and provider route. Both are recorded. DeepSeek uses the same question as Luna, with reasoning_effort=low and max_tokens=16384. Only the completed assistant answer is scored; reasoning text is not an answer. The native token counts are saved with normalized input and output counts. DeepSeek is also a benchmark arm; trying it as a judge does not make it independent of the contestants.

Newly generated task configurations default to Luna and production judging. Configure both JUDGE_NAME and JUDGE_MODEL when selecting Jev. Set AACR_JUDGE_MODE=mock to ask for diagnostic scoring explicitly; mock results remain invalid for benchmark reporting, including exact-text matches. These source changes do not update the task revisions pinned by existing runs.

All judges require the response body's model identifier to equal JUDGE_MODEL. Luna and DeepSeek send aacr-semantic-v2; Jev sends aacr-semantic-jev-v2. The verifier records the selected judge, endpoint, model, prompt version and digest, probability, and threshold in grading_details.json along with one decision per judged pair. The metrics and matching rules are the same for all judges. Independent maximum bipartite matching produces the line and semantic counts and is recorded as aacr-maximum-matching-v1. An exact-text pair is counted as a match without a judge call, and a mock run never calls the judge. Both decisions carry their own source value.

The verifier runs inside the Harbor process through aacr_bench_harbor.harbor_verifier:AacrVerifier, so the judge credential stays on the host and is never placed in the task container. Harbor reads the verifier class from the run configuration, not from task.toml, so a run selects it with verifier.import_path = "aacr_bench_harbor.harbor_verifier:AacrVerifier". The recipe in this repository sets that value, and the generated tasks carry no in-container verifier. The verifier reads references.json from the task's tests directory and the submission from the collected trial output. Harbor remains responsible for trial scheduling and result collection.

A failed pair is recorded in judge_errors with its message in judge_errors_detail and skipped. It is never a negative answer and never aborts the trial. A trial whose judged pairs all failed is marked invalid through benchmark_valid. The judge-eval harness saves rejected responses as judge_failure-*.json after credential checks, and it can judge a saved split on the command line or replay saved responses offline with AACR_JUDGE_TRANSPORT=replay and a matching AACR_JUDGE_RECORDS_PATH.

JUDGE_BASE_URL can override the provider's base URL. HTTPS is required except for local fake endpoints on loopback HTTP. URLs must not contain credentials or query parameters. Never put credential values in task files, command arguments, records, or logs. Tests use local fake endpoints only.

Model study status

The fixed five-task canary completed for Luna, Terra, and two earlier DeepSeek/Pi configurations. Luna and Terra completed every trial. The earlier DeepSeek runs used a destructive 20-minute deadline and are retained only as diagnostic evidence. The corrected DeepSeek protocol pins the provenance-signed @osolmaz/pi-reviewer 0.1.4 npm package, gives one Pi session 10 minutes to investigate and 2 minutes to submit, retries one stalled final-submission request halfway through that grace period, and uses official Harbor 0.21.0 without source patches. The registered 10-task Harbor HF canary produced 9 valid submissions, exactly meeting its operational gate. The current execution runs each of the 57 released tasks once. The canary has no margin and does not establish a stable 90% completion rate.

The frozen split names 98 development, 39 validation, and 59 sealed-test pull requests. Four deleted Keycloak commits and one deleted ClickHouse commit were rebuilt exactly from Software Heritage. Keycloak PR #35645, Node.js PR #56185, ComfyUI PRs #7952 and #9560, uv PR #11088, and ClickHouse PR #85266 were absent from checked public archives. That leaves 94 development, 39 validation, and 57 sealed-test tasks with 755, 255, and 462 references.

The model-study results above were produced with the first source-pack format. Those packs set aacr-base to the recorded base commit, so a two-dot git diff aacr-base HEAD also showed target-branch changes made after the pull request branched, reversed. Pi Reviewer computes the merge base itself and reviewed the pull-request range. The finding converter used the two-dot diff to place findings, which can differ only for files that both the pull request and the target branch changed. Current packs set aacr-base to the merge base, so every tool sees only the pull-request change. The historical study design contains 684 trials. Each new campaign needs its own measured cumulative cost ceiling. See the source corpus, canary results, and implementation plan.

Development

scripts/check.sh
uv run slophammer-py check . --execute
npx -y @simpledoc/simpledoc check
git diff --check

The implementation plan records the source baseline, hidden-verifier boundary, model-study protocol, and remaining publication work.

License

Apache-2.0