AACR-Bench: Evaluating Automatic Code Review with Holistic Repository-Level Context
Paper • 2601.19494 • Published • 17
Datasets:
instance_id string | set string | key_version string | defect_id string | title string | statement string | kind string | severity string | sites list | evidence string | counts_if string | provenance_hunt bool | provenance_hunt_refs list | provenance_dataset_refs list | provenance_reviewer_comments list | provenance_added_in string | benchmark_limitation string | bot_derived_reference string | repo string | head_commit string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
comfyanonymous__ComfyUI@0bb6023 | smoke10 | 2 | D1 | Colab notebook: localtunnel and iframe launch cells lose the imports their threads still use (NameError, UI never exposed) | To silence ruff F811 (a cross-cell redefinition report), the PR deletes `import subprocess/time/socket/urllib.request` from the localtunnel launch cell and `import time/socket` from the colab-iframe launch cell. Both cells' `iframe_thread` still calls `time.sleep`, `socket.socket`, and in the localtunnel cell also `url... | bug | medium | [
{
"path": "notebooks/comfyui_colab.ipynb",
"from_line": 237,
"to_line": 260,
"role": "cause"
},
{
"path": "notebooks/comfyui_colab.ipynb",
"from_line": 284,
"to_line": 302,
"role": "other-site"
}
] | At head, the localtunnel cell (id jjjjjjjjjjjjj) imports only `threading` (line 240: "import threading\n"), yet uses `time.sleep(0.5)` (244), `socket.socket(socket.AF_INET, socket.SOCK_STREAM)` (245), `urllib.request.urlopen('https://ipv4.icanhazip.com')` (252) and `subprocess.Popen(["lt", "--port", ...])` (253). The i... | To count, a comment must say that imports removed from the localtunnel cell and/or the iframe cell (time, socket, subprocess, urllib.request) are still used by that cell's thread code, so running that cell raises NameError and the tunnel or iframe never starts. Flagging either cell is enough, since both share one root ... | true | [] | [
"3",
"4"
] | [
"R1",
"R2",
"R3",
"R4",
"R5",
"R6",
"R7",
"R8",
"R9",
"R10",
"R11",
"R12",
"R13"
] | null | null | null | comfyanonymous/ComfyUI | 0bb60235132cc72daea81427107a40ba1ccb5ddf |
infiniflow__ragflow@7051234 | smoke10 | 2 | D1 | get_doc_ids_by_doc_names returns None on no match, so DELETE /api/document with doc_ids only fails with TypeError | The new get_doc_ids_by_doc_names returns None, not an empty list, when no row matches. That includes the default `doc_names=[]`, because peewee renders an empty IN as `0 = 1`. document_rm then runs `doc_id not in doc_ids` against None and raises TypeError, which the except turns into server_error_response. So every req... | bug | high | [
{
"path": "api/db/services/document_service.py",
"from_line": 337,
"to_line": 340,
"role": "cause"
},
{
"path": "api/apps/api_app.py",
"from_line": 618,
"to_line": 621,
"role": "symptom"
},
{
"path": "api/apps/api_app.py",
"from_line": 628,
"to_line": 629,
"ro... | document_service.py:337-340 at head: `doc_ids = doc_ids.dicts()` / `if not doc_ids:` / `return` / `return [doc["id"] for doc in doc_ids]`. That returns None for zero rows; peewee evaluates truthiness through __len__, and the sibling get_doc_id_by_doc_name uses the same pattern. api_app.py:618-621: `doc_ids = DocumentSe... | To count, a comment must say that the new name lookup returns None, not an empty list, when no names match or doc_names is omitted or empty. It must also say that document_rm's membership test and append on doc_ids then fail with TypeError or a server error, so ids-only deletion breaks. A comment placed at either site ... | true | [] | [
"0"
] | [
"R1",
"R2",
"R3",
"R4",
"R6",
"R8",
"R10",
"R25",
"R26",
"R27",
"R28",
"R29"
] | null | null | null | infiniflow/ragflow | 705123406ec7cd5eb1f9be26f6c5c31357aeffc9 |
infiniflow__ragflow@7051234 | smoke10 | 2 | D2 | Delete-by-name now deletes every document with that name in the deployment, across knowledge bases and tenants | The pre-PR name lookup already ignored tenant and knowledge base, but deleted only the first document matching each name. The new get_doc_ids_by_doc_names selects every document with a matching name across the whole deployment, so one DELETE /v1/api/document request by name now irreversibly deletes every same-named doc... | data | high | [
{
"path": "api/db/services/document_service.py",
"from_line": 333,
"to_line": 340,
"role": "cause"
},
{
"path": "api/apps/api_app.py",
"from_line": 618,
"to_line": 618,
"role": "other-site"
},
{
"path": "api/apps/api_app.py",
"from_line": 641,
"to_line": 660,
... | document_service.py:335-336 `doc_ids = cls.model.select(*fields).where(cls.model.name.in_(doc_names))`, with every row returned at 340. Compare the merge-base helper at 322-329, `.where(cls.model.name == doc_name)` ... `return doc_id[0]["id"]`, which gives one id per name. db_models.py:624 `name = CharField(max_length=... | To count, a comment must identify that the new bulk name lookup is not scoped to the caller's tenant or knowledge base and/or returns ALL documents with a matching name. It must also say that a delete-by-name request can therefore remove documents beyond the one intended, either other tenants' documents or several same... | true | [] | [] | [
"R5",
"R7",
"R9",
"R17",
"R18",
"R19",
"R20",
"R21",
"R22",
"R23",
"R24"
] | null | null | null | infiniflow/ragflow | 705123406ec7cd5eb1f9be26f6c5c31357aeffc9 |
infiniflow__ragflow@7051234 | smoke10 | 2 | D3 | doc_names that match nothing are silently dropped instead of producing 'Document not found!' | Before, an unknown name put None into doc_ids, and the delete loop returned get_data_error_result('Document not found!') when it reached that entry. Now the IN query simply omits names with no match. A request that mixes valid names with a misspelled or already-deleted name therefore returns success (data=True), and th... | contract | low | [
{
"path": "api/db/services/document_service.py",
"from_line": 335,
"to_line": 340,
"role": "cause"
},
{
"path": "api/apps/api_app.py",
"from_line": 618,
"to_line": 626,
"role": "symptom"
},
{
"path": "api/apps/api_app.py",
"from_line": 641,
"to_line": 644,
"ro... | Merge base: `doc_ids = [DocumentService.get_doc_id_by_doc_name(doc_name) for doc_name in req.get("doc_names", [])]`. get_doc_id_by_doc_name returns None for no match, then `e, doc = DocumentService.get_by_id(doc_id); if not e: return get_data_error_result(message="Document not found!")` fires for that None. Head: docum... | To count, a comment must say that names with no matching document are now silently ignored (no 'not found' error), so a partly unmatched name list reports success. R4's 'requests with unknown names fail the same way' does NOT count. It describes the TypeError crash, which is D1. | true | [] | [] | [] | null | null | null | infiniflow/ragflow | 705123406ec7cd5eb1f9be26f6c5c31357aeffc9 |
langflow-ai__langflow@9513d15 | smoke10 | 2 | D1 | Non-reentrant vertex lock re-acquired inside get_result: every result lookup with a flow_id deadlocks | Vertex.get_result holds self._lock (a plain asyncio.Lock, which is not reentrant) while it awaits self._get_result. Both _get_result implementations now await _log_transaction_async, which does `async with self._lock` on the same lock, so the coroutine waits forever on a lock its own caller holds. Every graph with flow... | concurrency | high | [
{
"path": "src/backend/base/langflow/graph/vertex/base.py",
"from_line": 636,
"to_line": 637,
"role": "cause"
},
{
"path": "src/backend/base/langflow/graph/vertex/base.py",
"from_line": 621,
"to_line": 622,
"role": "other-site"
},
{
"path": "src/backend/base/langflow/grap... | base.py:62 `self._lock = asyncio.Lock()`. base.py:621-622 `async with self._lock:\n return await self._get_result(requester, target_handle_name)`. base.py:637 (inside the new async _log_transaction_async) `async with self._lock:`. base.py:683 `await self._log_transaction_async(str(flow_id), source=self, target=reque... | To count, a comment must say that _log_transaction_async acquires self._lock while its caller get_result already holds the same lock, and that asyncio.Lock is not reentrant, so the call deadlocks or hangs whenever flow_id is set. Pointing at either the lock acquisition (base.py:637) or any of the four awaited call site... | true | [] | [
"0"
] | [
"R2",
"R3",
"R4",
"R5",
"R6",
"R7",
"R8",
"R9",
"R10",
"R11",
"R23"
] | null | null | null | langflow-ai/langflow | 9513d152197e5336fbc00eb6b891f1ffb25ad62d |
langflow-ai__langflow@9513d15 | smoke10 | 2 | D2 | suppress(asyncio.CancelledError) around Task.cancel() is ineffective dead code | The new outer cancellation handler wraps `pending_task.cancel()` in `with suppress(asyncio.CancelledError)`. Task.cancel() only schedules cancellation and never raises CancelledError in the caller, so the wrapper and the new `from contextlib import suppress` import do nothing. They mislead readers about which cancellat... | maintainability | low | [
{
"path": "src/backend/base/langflow/graph/vertex/base.py",
"from_line": 658,
"to_line": 659,
"role": "cause"
},
{
"path": "src/backend/base/langflow/graph/vertex/base.py",
"from_line": 10,
"to_line": 10,
"role": "other-site"
}
] | base.py:658-659 `with suppress(asyncio.CancelledError):\n pending_task.cancel()`; base.py:10 `from contextlib import suppress` was added only for this. asyncio.Task.cancel() returns a bool and delivers CancelledError inside the target task on its next step, never synchronously to the caller. | To count, a comment must say that the suppress(CancelledError) around task.cancel() cannot catch anything, because cancel() does not raise in the caller, and so the wrapper is redundant or dead. Near-misses that do NOT count: a claim that the outer handler iterates an already-cleared set and therefore misses cancellati... | false | [] | [
"2"
] | [
"R19",
"R20",
"R21"
] | null | null | null | langflow-ai/langflow | 9513d152197e5336fbc00eb6b891f1ffb25ad62d |
immich-app__immich@7e2009f | smoke10 | 2 | D1 | Motion-video asset's fileModifiedAt now comes from the still's EXIF ModifyDate, and that value then feeds the new min() fallback | When the PR creates the hidden motion-video asset, it now sets fileModifiedAt to getDates(asset, tags).modifyDate. That is the still image's EXIF ModifyDate whenever one is present. Before, it was the still's stored file modification timestamp, and every other writer stores a file or client mtime in that field. The vid... | contract | low | [
{
"path": "server/src/services/metadata.service.ts",
"from_line": 453,
"to_line": 460,
"role": "cause"
},
{
"path": "server/src/services/metadata.service.ts",
"from_line": 608,
"to_line": 611,
"role": "other-site"
},
{
"path": "server/src/services/metadata.service.ts",
... | Head metadata.service.ts:453-460 has `const dates = this.getDates(asset, tags); ... fileCreatedAt: dates.dateTimeOriginal, fileModifiedAt: dates.modifyDate, localDateTime: dates.localDateTime`. At the merge base it was `const createdAt = asset.fileCreatedAt ?? asset.createdAt; ... fileModifiedAt: asset.fileModifiedAt`.... | To count, a comment must say that the motion-video creation now stores the still's EXIF ModifyDate (getDates().modifyDate) as the video's fileModifiedAt instead of the file modification timestamp, i.e. that the field's meaning or provenance changes. Also count a comment that explains how the dates written here, fed int... | true | [] | [] | [
"R6",
"R7",
"R12",
"R13",
"R14"
] | null | null | null | immich-app/immich | 7e2009f038cbf3f26403af6582af3cf8ba91672f |
cline__cline@7d18fb9 | smoke10 | 2 | D1 | ASCII-only files are detected as "ascii" and re-encoded as ASCII on save, so non-ASCII characters added by the edit become "?" | For any buffer with no byte >= 0x80 (most source files), jschardet returns encoding "ascii". "ascii" !== "utf8", so after VS Code has saved the document, saveChanges re-encodes it with iconv's ASCII codec and overwrites the file. Every non-ASCII character the model or user added (accents, em dashes, arrows, emoji, CJK ... | data | high | [
{
"path": "src/integrations/editor/DiffViewProvider.ts",
"from_line": 49,
"to_line": 57,
"role": "cause"
},
{
"path": "src/integrations/editor/DiffViewProvider.ts",
"from_line": 237,
"to_line": 240,
"role": "symptom"
}
] | DiffViewProvider.ts:50-56 `const detected = chardet.detect(fileBuffer); this.fileEncoding = typeof detected === "string" ? detected : detected && (detected as any).encoding ? (detected as any).encoding : "utf8"`. jschardet's UniversalDetector.close() returns {encoding:"ascii", confidence:1} in the pure-ASCII input stat... | To count, a comment must say that a pure-ASCII original (detected as "ascii") makes the save path encode with ASCII, so newly added non-ASCII characters are silently replaced with "?" or lost. Accept equivalent wording, for example "an ASCII-detected file edited to contain an emoji is passed to iconv.encode(..., 'ascii... | true | [] | [] | [
"R10",
"R27",
"R28",
"R32",
"R34"
] | null | null | null | cline/cline | 7d18fb9df620c61420c6939df22fb6f7af836bb2 |
cline__cline@7d18fb9 | smoke10 | 2 | D2 | package-lock.json not updated: jschardet (and the direct iconv-lite dependency) missing from the lockfile, so npm ci / CI and publishing break | package.json adds iconv-lite and jschardet as dependencies, but package-lock.json is unchanged. jschardet appears nowhere in it, and neither package is in the root dependencies block. On a CI cache miss, npm ci in test.yml and publish.yml refuses the mismatched manifest and lock. On a cache hit (the key hashes the unch... | build | high | [
{
"path": "package.json",
"from_line": 292,
"to_line": 295,
"role": "cause"
}
] | The diff touches only .changeset, package.json, DiffViewProvider.ts and extract-text.ts. At head, `git show HEAD:package-lock.json | grep -c jschardet` gives 0. The root packages[""].dependencies lists "isbinaryfile": "^5.0.2" but neither new package; iconv-lite exists only as a transitive node_modules/iconv-lite 0.6.3... | To count, a comment must say the lockfile was not updated for the new dependency or dependencies (jschardet at minimum), so clean or CI installs (npm ci) or the build fail. A comment that only questions the choice of library or version does not count. | true | [] | [] | [
"R1",
"R2",
"R3",
"R4",
"R5",
"R6",
"R7",
"R8"
] | null | null | null | cline/cline | 7d18fb9df620c61420c6939df22fb6f7af836bb2 |
cline__cline@7d18fb9 | smoke10 | 2 | D3 | jschardet's guess is trusted with no confidence/UTF-8 validity check: valid UTF-8 files with few non-ASCII chars are misdetected as a single-byte encoding, read as mojibake and corrupted on revert/save | The detected encoding is used unconditionally. jschardet's UTF-8 prober confidence for one to three multibyte characters is only 0.5 to 0.88, while the Latin1 and SBCS probers can score higher. A UTF-8 file containing, for example, a single ©, é or em dash is therefore labelled windows-1252 or windows-1250. Those files... | bug | high | [
{
"path": "src/integrations/editor/DiffViewProvider.ts",
"from_line": 49,
"to_line": 57,
"role": "cause"
},
{
"path": "src/integrations/editor/DiffViewProvider.ts",
"from_line": 269,
"to_line": 279,
"role": "symptom"
},
{
"path": "src/integrations/editor/DiffViewProvider.... | DiffViewProvider.ts:50-57 uses `chardet.detect(fileBuffer)` with no confidence or UTF-8 validity check, then `this.originalContent = iconv.decode(fileBuffer, this.fileEncoding)`. revertChanges :276-279 `edit.replace(updatedDocument.uri, fullRange, this.originalContent ?? "") ... await updatedDocument.save()` saves the ... | To count, a comment must say that jschardet's guess is used without a confidence or UTF-8 validity check, so real UTF-8 files (typically with few non-ASCII characters) can be detected as a legacy single-byte encoding, AND name a consequence: garbled read_file, diff or notebook text, or corruption when the file is rever... | true | [] | [] | [] | null | null | null | cline/cline | 7d18fb9df620c61420c6939df22fb6f7af836bb2 |
cline__cline@7d18fb9 | smoke10 | 2 | D4 | Encoding-preserving write overwrites the saved file with the normalized comparison string (trimEnd, forced EOL, EOLs converted) | The new write serializes normalizedPostSaveContent, not the text VS Code saved. That string is trimEnd()'d, has exactly one EOL appended, and has every line ending rewritten to the EOL of the model's newContent. So a write on a CRLF file converts it to LF, trailing blank lines and whitespace at EOF are removed, and a f... | data | medium | [
{
"path": "src/integrations/editor/DiffViewProvider.ts",
"from_line": 211,
"to_line": 213,
"role": "cause"
},
{
"path": "src/integrations/editor/DiffViewProvider.ts",
"from_line": 237,
"to_line": 240,
"role": "symptom"
}
] | DiffViewProvider.ts:237 `if (this.fileEncoding.toLowerCase() !== "utf8") {`, where fileEncoding is jschardet's name (UTF8Prober reports "UTF-8"; pure ASCII reports "ascii"). :211-213 `const newContentEOL = this.newContent.includes("\r\n") ? "\r\n" : "\n"; const normalizedPostSaveContent = postSaveContent.replace(/\r\n|... | To count, a comment must say the overwrite writes normalizedPostSaveContent (trimmed / EOL-normalized / forced final newline) instead of the text actually saved, so the file's bytes (trailing whitespace, final newline, CRLF/LF) differ from what was approved. A comment naming only the 'utf8' vs 'UTF-8'/'ascii' guard mis... | true | [] | [] | [
"R25",
"R30",
"R33"
] | null | null | null | cline/cline | 7d18fb9df620c61420c6939df22fb6f7af836bb2 |
cline__cline@7d18fb9 | smoke10 | 2 | D12 | The `!== "utf8"` guard never matches jschardet's names, so every existing file is re-encoded and rewritten behind VS Code | saveChanges re-encodes and overwrites the file when this.fileEncoding.toLowerCase() !== "utf8", but jschardet returns "UTF-8", "ascii" and so on, never "utf8". So every existing file — not only legacy-encoded ones — goes through iconv and fs.writeFile after VS Code has already saved it. This is the enabler that routes ... | bug | medium | [
{
"path": "src/integrations/editor/DiffViewProvider.ts",
"from_line": 237,
"to_line": 237,
"role": "cause"
}
] | DiffViewProvider.ts:237 `if (this.fileEncoding.toLowerCase() !== "utf8") {`, where fileEncoding is jschardet's name (UTF8Prober reports "UTF-8"; pure ASCII reports "ascii"). :211-213 `const newContentEOL = this.newContent.includes("\r\n") ? "\r\n" : "\n"; const normalizedPostSaveContent = postSaveContent.replace(/\r\n|... | To count, a comment must name the case/naming mismatch: jschardet's 'UTF-8'/'ascii' never equal 'utf8', so the re-encode-and-overwrite path runs for ordinary UTF-8/ASCII files too. A comment that names only the lossy '?' consequence of encoding to 'ascii' counts for D1, not here. | true | [] | [] | [
"R25",
"R30",
"R33"
] | null | null | null | cline/cline | 7d18fb9df620c61420c6939df22fb6f7af836bb2 |
cline__cline@7d18fb9 | smoke10 | 2 | D5 | Save strips the byte-order mark from UTF-8-BOM and UTF-16/32-BOM files | jschardet reports BOM-prefixed files as UTF-8, UTF-16LE/BE and so on. iconv.decode strips the BOM by default, and iconv.encode without addBOM writes the file back with none, although VS Code's own save had preserved it. After every accepted edit, UTF-8-BOM files silently lose their BOM, which BOM-dependent consumers ne... | data | medium | [
{
"path": "src/integrations/editor/DiffViewProvider.ts",
"from_line": 57,
"to_line": 57,
"role": "cause"
},
{
"path": "src/integrations/editor/DiffViewProvider.ts",
"from_line": 237,
"to_line": 240,
"role": "symptom"
}
] | DiffViewProvider.ts:57 `this.originalContent = iconv.decode(fileBuffer, this.fileEncoding)` and :238 `iconv.encode(normalizedPostSaveContent, this.fileEncoding)` pass no addBOM option. Verified with iconv-lite 0.6.3: decode(EF BB BF 68 69 0a, 'UTF-8') gives 'hi\n', and encode('hi\n','UTF-8') gives <68 69 0a> (no BOM). ... | To count, a comment must say that the new re-encode/overwrite drops a file's BOM, because iconv strips it on decode and does not add it on encode, so BOM-bearing files change on save. Comments about the charset alone do not count. | true | [] | [] | [
"R31"
] | null | null | null | cline/cline | 7d18fb9df620c61420c6939df22fb6f7af836bb2 |
cline__cline@7d18fb9 | smoke10 | 2 | D6 | Binary check demoted to a fallback: any binary jschardet assigns an encoding is decoded and returned to the model as text | Before the PR, the default branch always ran isBinaryFile first and refused binaries. Now isBinaryFile runs only when jschardet returns no encoding. jschardet has no notion of binary content: any 7-bit buffer, including NULs and control bytes, is reported as "ascii", and single-byte probers can put a legacy charset gue... | bug | medium | [
{
"path": "src/integrations/misc/extract-text.ts",
"from_line": 30,
"to_line": 43,
"role": "cause"
}
] | extract-text.ts:30-43 `const detected = chardet.detect(fileBuffer) ... else if (detected && (detected as any).encoding) { encoding = (detected as any).encoding } else { const isBinary = await isBinaryFile(fileBuffer)... }` then `return iconv.decode(fileBuffer, encoding)`. Base: `const isBinary = await isBinaryFile(file... | To count, a comment must say that isBinaryFile now runs only when charset detection returns nothing, so a binary file that gets a charset guess (for example "ascii" for 7-bit data) bypasses binary rejection and is decoded and sent to the model. Near-misses that do not count: requests for binary-rejection tests, and com... | true | [] | [] | [
"R54",
"R57",
"R58",
"R61",
"R62",
"R63",
"R64",
"R65",
"R66"
] | null | null | null | cline/cline | 7d18fb9df620c61420c6939df22fb6f7af836bb2 |
cline__cline@7d18fb9 | smoke10 | 2 | D7 | extractTextFromFile reads the entire file into memory before the 300KB size check | The old default branch rejected binaries with a 512-byte sniff and rejected large files with fs.stat (getFileSizeInKB) before reading. The new code does fs.readFile(filePath) on the whole file first and checks byteLength afterwards. A read_file or folder mention on a multi-hundred-MB log, dataset or media file now buff... | performance | medium | [
{
"path": "src/integrations/misc/extract-text.ts",
"from_line": 26,
"to_line": 29,
"role": "cause"
},
{
"path": "src/integrations/misc/extract-text.ts",
"from_line": 7,
"to_line": 7,
"role": "other-site"
}
] | extract-text.ts:26-29 `const fileBuffer = await fs.readFile(filePath); if (fileBuffer.byteLength > 300 * 1024) { throw new Error(`File is too large to read into context.`) }`. Base: `const fileSizeInKB = await getFileSizeInKB(filePath); if (fileSizeInKB > 300) { throw ... } return await fs.readFile(filePath, "utf8")`, ... | To count, a comment must say that the whole file is now read into memory before the 300KB limit is enforced, a regression from the stat-based pre-check, with a memory or performance consequence for large files. A comment that only says "remove the unused getFileSizeInKB import" does not count. Dataset ref 4 counts beca... | true | [] | [
"4",
"5"
] | [
"R36",
"R37",
"R40",
"R41",
"R42",
"R43",
"R44",
"R45"
] | null | null | null | cline/cline | 7d18fb9df620c61420c6939df22fb6f7af836bb2 |
cline__cline@7d18fb9 | smoke10 | 2 | D8 | Reject/revert path does not preserve the detected encoding: rejecting an edit transcodes a legacy-encoded file | On the reject/abort/failed-diff path (revertChanges), the PR restores the decoded original text through VS Code's own save, so a legacy-encoded (Shift-JIS, ISO-8859-1) file is converted to UTF-8 with its content intact instead of keeping its original encoding. That contradicts the PR's promise to preserve the encoding ... | incomplete-fix | low | [
{
"path": "src/integrations/editor/DiffViewProvider.ts",
"from_line": 269,
"to_line": 279,
"role": "cause"
},
{
"path": "src/integrations/editor/DiffViewProvider.ts",
"from_line": 237,
"to_line": 240,
"role": "other-site"
}
] | DiffViewProvider.ts:271-279 `edit.replace(updatedDocument.uri, fullRange, this.originalContent ?? ""); await vscode.workspace.applyEdit(edit); await updatedDocument.save()`. This path has no iconv.encode/fs.writeFile with this.fileEncoding, unlike saveChanges:237-240. revertChanges is called from Cline.ts:1752 (diff er... | To count, a comment must say that revertChanges (reject, abort or failed diff) restores the decoded text through VS Code's document save without the detected encoding, so the file's bytes or encoding change when an edit is rejected. Near-misses that do not count: comments only about VS Code decoding the editable side o... | true | [] | [] | [
"R11",
"R13",
"R14",
"R16",
"R19",
"R21",
"R23",
"R24",
"R26"
] | null | null | null | cline/cline | 7d18fb9df620c61420c6939df22fb6f7af836bb2 |
cline__cline@7d18fb9 | smoke10 | 2 | D9 | Encodings jschardet can return but iconv-lite cannot decode now make read_file and edits throw | The detected name goes straight to iconv.decode with no encodingExists check or fallback. jschardet can return ISO-2022-JP, ISO-2022-KR, ISO-2022-CN, HZ-GB-2312, EUC-TW, and X-ISO-10646-UCS-4-3412/2143 (the last two from a BOM). iconv-lite 0.6.3 supports none of these. Such files, which were previously read as UTF-8, n... | compat | low | [
{
"path": "src/integrations/editor/DiffViewProvider.ts",
"from_line": 57,
"to_line": 57,
"role": "cause"
},
{
"path": "src/integrations/misc/extract-text.ts",
"from_line": 43,
"to_line": 43,
"role": "cause"
},
{
"path": "src/integrations/misc/extract-text.ts",
"from_l... | DiffViewProvider.ts:57 `iconv.decode(fileBuffer, this.fileEncoding)`, extract-text.ts:43 `return iconv.decode(fileBuffer, encoding)` and :63: none has a guard. Verified with iconv-lite 0.6.3: encodingExists is false for 'ISO-2022-JP', 'ISO-2022-KR', 'ISO-2022-CN', 'HZ-GB-2312', 'EUC-TW', 'X-ISO-10646-UCS-4-3412' and 'X... | To count, a comment must say that jschardet can report encodings iconv-lite does not support, so iconv.decode throws with no encodingExists check or fallback. Near-miss that does not count: a generic "wrap in try/catch" (dataset refs 1 and 2), since callers already catch and the comment does not identify the unsupporte... | true | [] | [] | [] | null | null | null | cline/cline | 7d18fb9df620c61420c6939df22fb6f7af836bb2 |
cline__cline@7d18fb9 | smoke10 | 2 | D10 | @-mention path still pre-screens with isBinaryFile, so non-UTF-8 text files remain misclassified as binary there | The PR says it fixes binary misclassification of non-UTF-8 files, but only extractTextFromFile changed. getFileOrFolderContent still calls isBinaryFile(absPath) before extractTextFromFile. isbinaryfile counts high bytes that are not valid UTF-8 as suspicious, so a Shift-JIS/GBK file is still reported as "(Binary file, ... | incomplete-fix | low | [
{
"path": "src/core/mentions/index.ts",
"from_line": 154,
"to_line": 159,
"role": "cause"
},
{
"path": "src/core/mentions/index.ts",
"from_line": 175,
"to_line": 180,
"role": "cause"
},
{
"path": "src/integrations/misc/extract-text.ts",
"from_line": 30,
"to_line":... | mentions/index.ts:155-159 `const isBinary = await isBinaryFile(absPath).catch(() => false); if (isBinary) { return "(Binary file, unable to display content)" } const content = await extractTextFromFile(absPath)`. :176-180 has the same pre-check in the folder loop and returns undefined. This file is untouched by the PR,... | To count, a comment must say that file and folder @-mentions still gate on isBinaryFile before extractTextFromFile, so legacy-encoded text is still rejected as binary there and the claimed fix does not reach mentions. Comments about binaries getting through extractTextFromFile do not count (that is D6). | true | [] | [] | [
"R49",
"R50",
"R51",
"R53",
"R56",
"R59"
] | null | null | null | cline/cline | 7d18fb9df620c61420c6939df22fb6f7af836bb2 |
cline__cline@7d18fb9 | smoke10 | 2 | D11 | Encoding-detection/fallback policy copy-pasted in three places (with a dead string branch and as-any casts) | The same detect-then-fallback logic is duplicated in DiffViewProvider.open, extractTextFromFile and extractTextFromIPYNB. Each copy carries an unreachable `typeof detected === "string"` branch and `(detected as any).encoding` casts, although jschardet.detect is typed to return {encoding, confidence}. Every fix the dete... | maintainability | low | [
{
"path": "src/integrations/editor/DiffViewProvider.ts",
"from_line": 50,
"to_line": 56,
"role": "other-site"
},
{
"path": "src/integrations/misc/extract-text.ts",
"from_line": 30,
"to_line": 42,
"role": "other-site"
},
{
"path": "src/integrations/misc/extract-text.ts",
... | DiffViewProvider.ts:51-56 `typeof detected === "string" ? detected : detected && (detected as any).encoding ? (detected as any).encoding : "utf8"`, repeated as an if/else chain at extract-text.ts:32-41 and as a ternary at :61-62. | To count, a comment must point out that the charset detection and fallback logic is duplicated across these readers/DiffViewProvider and should be one shared helper so the policy cannot drift. Comments only about the nested ternary's readability, or only about `as any`, do not count. | true | [] | [
"6"
] | [
"R67",
"R68",
"R69",
"R70"
] | null | null | null | cline/cline | 7d18fb9df620c61420c6939df22fb6f7af836bb2 |
mrdoob__three.js@96ab072 | smoke10 | 2 | D1 | Front and back caps vanish for every non-bevelled extrusion (bevelEnabled:false, extrudePath, bevelSegments:0, TextGeometry default) | The PR triangulates contractedContourVertices/expandedHoleVertices, but these are filled only inside the bevel-segment loop (`if ( t == 0 ) ...push`). When bevel is disabled, bevelSegments is forced to 0, so the loop never runs. triangulateShape([], []) then returns no faces, flen is 0, and buildLidFaces emits no botto... | bug | high | [
{
"path": "src/geometries/ExtrudeGeometry.js",
"from_line": 387,
"to_line": 435,
"role": "cause"
},
{
"path": "src/geometries/ExtrudeGeometry.js",
"from_line": 104,
"to_line": 134,
"role": "other-site"
},
{
"path": "src/geometries/ExtrudeGeometry.js",
"from_line": 553... | L109 `bevelEnabled = false; // bevels not supported for path extrusion` and L127-129 `if ( ! bevelEnabled ) { bevelSegments = 0; ...`. L392 `for ( let b = 0; b < bevelSegments; b ++ ) {` is the only filler of the arrays: L407 `if ( t == 0 ) contractedContourVertices.push( vert );` and L427 `if ( t == 0 ) expandedHoleVe... | To count, a comment must say that with bevel disabled (bevelEnabled:false, extrudePath, or bevelSegments 0) the arrays passed to triangulateShape are empty or never populated because the bevel loop does not run, so the lids/caps are missing. These do not count: comments that only ask for tests of the non-bevel path (R1... | true | [] | [] | [
"R13",
"R14",
"R16",
"R17",
"R18",
"R19",
"R22",
"R23",
"R24"
] | null | null | null | mrdoob/three.js | 96ab0720fc65368e32ee3f2389cb192f96d8adff |
mrdoob__three.js@96ab072 | smoke10 | 2 | D2 | A hole that welds down to zero points crashes the constructor inside earcut (TypeError) | cleanPoints compares with the last kept point and, at the wrap-around, compares points[0] with it. Once a ring is down to one point, that point is compared with itself (distance 0) and spliced, so a tiny or all-identical hole becomes []. With bevel enabled (the default), triangulateShape still registers a zero-length h... | bug | medium | [
{
"path": "src/geometries/ExtrudeGeometry.js",
"from_line": 170,
"to_line": 194,
"role": "cause"
},
{
"path": "src/geometries/ExtrudeGeometry.js",
"from_line": 433,
"to_line": 433,
"role": "symptom"
}
] | L172-181: `const currentIndex = i % points.length; ... if ( distSq <= THRESHOLD_SQ ) { points.splice( currentIndex, 1 ); if ( currentIndex == 0 ) break;`. With length 1, currentIndex is 0 and currentPos === prevPos, so the last point is removed. L194 `holes.forEach( cleanPoints );`. src/extras/ShapeUtils.js triangulate... | To count, a comment must say that cleanPoints can reduce a hole (or ring) to zero points, and that the empty hole then breaks triangulation (throw/crash in triangulateShape or earcut). These do not count: saying only that small holes or contours are erased or lose detail (that is D3); dataset ref 1's claim that cleanPo... | true | [] | [] | [] | null | null | null | mrdoob/three.js | 96ab0720fc65368e32ee3f2389cb192f96d8adff |
mrdoob__three.js@96ab072 | smoke10 | 2 | D3 | Absolute 0.0002 world-unit weld threshold decimates or deletes valid small-scale shapes and holes | The threshold is a hard-coded world-space distance that ignores the shape's scale, and each point is measured against the last kept point. Any ring sampled more finely than 0.0002 loses vertices, and any contour or hole whose edges are all shorter than 0.0002 is spliced away entirely. Legitimate small-unit geometry tha... | bug | medium | [
{
"path": "src/geometries/ExtrudeGeometry.js",
"from_line": 165,
"to_line": 191,
"role": "cause"
},
{
"path": "src/geometries/ExtrudeGeometry.js",
"from_line": 193,
"to_line": 194,
"role": "other-site"
}
] | L167-168 `const THRESHOLD = 0.0002; const THRESHOLD_SQ = THRESHOLD * THRESHOLD;`. L178-183 splice on `distSq <= THRESHOLD_SQ`. L187 `prevPos = currentPos;` runs only for kept points. L193 `cleanPoints( vertices );`. Measured by running the sources, base -> head: a 0.0001 square with bevel off went from lid 12 / side 24... | To count, a comment must say that the fixed or absolute 0.0002 threshold (not scale-aware) removes vertices from legitimately small or densely sampled shapes, or erases small contours or holes entirely. These do not count: pure renaming or `===` style comments; R9's seam re-check edge case (a different, contrived issue... | true | [] | [] | [
"R1",
"R2",
"R3",
"R4",
"R5",
"R6",
"R7",
"R8"
] | null | null | null | mrdoob/three.js | 96ab0720fc65368e32ee3f2389cb192f96d8adff |
ollama__ollama@b6002f6 | smoke10 | 2 | D1 | Fit-mode graph reservation reads positions through a no-alloc placeholder pointer, so llama4 and mllama cannot load | The runner always runs LoadOperationFit first, with AllocMemory=false. In that mode FromIntSlice never writes the positions, and ggml gives the tensor a placeholder data pointer equal to the buffer alignment (about 0x20). The new Ints() branch returns unsafe.Slice over that non-nil pointer. reserveWorstCaseGraph then d... | bug | high | [
{
"path": "runner/ollamarunner/runner.go",
"from_line": 1066,
"to_line": 1080,
"role": "cause"
},
{
"path": "ml/backend/ggml/ggml.go",
"from_line": 1029,
"to_line": 1041,
"role": "cause"
},
{
"path": "kvcache/encoder.go",
"from_line": 92,
"to_line": 95,
"role"... | llm/server.go:709 `for operation := LoadOperationFit; operation < LoadOperationCommit; operation++` means Fit always runs first. runner.go:1204 `AllocMemory: req.Operation != llm.LoadOperationFit`, and ggml.go:162 `C.ggml_backend_buft_set_alloc(bt, C.bool(params.AllocMemory))` covers the CPU buffer type, which backs In... | To count, a comment must say that during the fit / no-alloc (AllocMemory=false) reservation, the positions tensor is never populated (or has only a placeholder data pointer), and that reading it back through Ints() in the encoder cache (mllama, encoder.go:94) or in llama4's temperature-scale loop (model_text.go:222) cr... | true | [] | [] | [
"R2",
"R3",
"R4",
"R7",
"R15",
"R16",
"R17",
"R18",
"R19"
] | null | null | null | ollama/ollama | b6002f6b6c69370f8c18a9e6c1cafa8463273a6b |
ollama__ollama@b6002f6 | smoke10 | 2 | D2 | The new Bytes/Floats/Ints fallback treats any non-nil data pointer as readable, already-computed host memory | For tensors whose sync is nil, the added branch reads ggml_get_data() directly and assumes it is host memory that already holds the values. That fails in two cases. A tensor in a GPU buffer (for example a weight from Layer()) has a device pointer, and GoBytes or unsafe.Slice on it crashes the process. A pending op whos... | contract | low | [
{
"path": "ml/backend/ggml/ggml.go",
"from_line": 1001,
"to_line": 1041,
"role": "cause"
},
{
"path": "ml/backend.go",
"from_line": 551,
"to_line": 573,
"role": "symptom"
}
] | ggml.go:1007-1009 `} else if bts := C.ggml_get_data(t.t); bts != nil { // tensor is in cpu memory; data = C.GoBytes(bts, C.int(C.ggml_nbytes(t.t)))`, with the same pattern at :1021-1024 (Floats) and :1035-1038 (Ints). ggml_get_data returns tensor->data whatever buffer holds it. ggml.c:3252 ggml_cpy_impl `struct ggml_te... | To count, a comment must say that a non-nil ggml data pointer does not mean the data is host-accessible and/or already computed, with a consequence: a crash when reading a GPU/device-buffer tensor, or stale or uninitialized values for a pending or uncomputed tensor (for example ml.Dump skipping Compute because Bytes() ... | true | [] | [] | [
"R5",
"R6"
] | null | null | null | ollama/ollama | b6002f6b6c69370f8c18a9e6c1cafa8463273a6b |
ollama__ollama@b6002f6 | smoke10 | 2 | D3 | Floats()/Ints() return a view into C memory on the new path but a Go-owned copy on the computed path, and Causal keeps the view after its context is freed | On the new branch, Floats() and Ints() return unsafe.Slice views into the tensor's ggml buffer. On the computed (sync) path they return freshly allocated Go copies, and Bytes() copies on both paths. Callers cannot tell whether the result outlives ctx.Close() or whether writing to it changes the tensor. Causal stores th... | maintainability | low | [
{
"path": "ml/backend/ggml/ggml.go",
"from_line": 1015,
"to_line": 1041,
"role": "cause"
},
{
"path": "kvcache/causal.go",
"from_line": 200,
"to_line": 204,
"role": "symptom"
}
] | ggml.go:1024 `data = unsafe.Slice((*float32)(bts), C.ggml_nelements(t.t))` and :1038 `data = unsafe.Slice((*int32)(bts), C.ggml_nelements(t.t))` versus :1031-1034 `data = make([]int32, ...); ... ggml_backend_tensor_get(...)`. Context.Close (ggml.go) calls ggml_backend_buffer_free on every allocated buffer. causal.go:20... | To count, a comment must say that the new Floats/Ints path returns a slice aliasing C/ggml memory that is freed when the context closes (or that writes alias the tensor), inconsistent with the copying sync path, ideally noting that Causal retains it in curPositions. Near-misses that do NOT count: pure performance remar... | true | [] | [] | [
"R8"
] | null | null | null | ollama/ollama | b6002f6b6c69370f8c18a9e6c1cafa8463273a6b |
dbeaver__dbeaver@bc3ac7b | smoke10 | 2 | D1 | InputStream obtained for a pasted remote file is never closed | dropNodes gets a newly opened stream from node.getAdapter(InputStream.class), which calls IFile.getContents(), and passes it to copyInputStream. That helper calls Files.copy(InputStream, Path, ...), which does not close its source. Nothing else closes the stream either: there is no try-with-resources. Every pasted remo... | resource | medium | [
{
"path": "plugins/org.jkiss.dbeaver.model/src/org/jkiss/dbeaver/model/navigator/fs/DBNPathBase.java",
"from_line": 290,
"to_line": 302,
"role": "cause"
},
{
"path": "plugins/org.jkiss.dbeaver.model/src/org/jkiss/dbeaver/model/navigator/fs/DBNPathBase.java",
"from_line": 363,
"to_lin... | DBNPathBase.java:293-296 at head reads `InputStream adapter = node.getAdapter(InputStream.class); if (adapter != null) { String fileName = node.getNodeDisplayName(); copyInputStream(adapter, folder, fileName); }`, and its `finally` only calls `monitor.worked(1)`. DBNPathBase.java:363-369 `copyInputStream(...) { ... Fil... | To count, a comment must say that the InputStream from getAdapter(InputStream.class) / IFile.getContents() in the new dropNodes branch or in copyInputStream is never closed, for example because Files.copy does not close it or because there is no try-with-resources. Mentioning the failure path is optional. Near-misses t... | true | [] | [
"0"
] | [
"R8",
"R9",
"R10",
"R11",
"R12",
"R13",
"R14",
"R15",
"R16",
"R19"
] | null | null | null | dbeaver/dbeaver | bc3ac7bb47092a934b344791d8aa3a1c41e6c8f7 |
dbeaver__dbeaver@bc3ac7b | smoke10 | 2 | D2 | New InputStream adapter duplicates the existing DBNStreamData.openInputStream contract | DBNResource already implements DBNStreamData. Its supportsStreamData() is true exactly for remote IFiles with no location, and openInputStream() opens the same IFile.getContents() stream and reports failure through checked IOException. The PR adds a second way to get that stream: getAdapter(InputStream.class). It has a... | maintainability | low | [
{
"path": "plugins/org.jkiss.dbeaver.model.rcp/src/org/jkiss/dbeaver/model/navigator/DBNResource.java",
"from_line": 423,
"to_line": 430,
"role": "cause"
},
{
"path": "plugins/org.jkiss.dbeaver.model.rcp/src/org/jkiss/dbeaver/model/navigator/DBNResource.java",
"from_line": 440,
"to_l... | New code at DBNResource.java:423-429 is `else if (adapter == InputStream.class && resource instanceof IFile file) { try { return adapter.cast(file.getContents()); } catch (CoreException e) { log.error(...); throw new RuntimeException(e); } }`. The existing code in the same class at :441-443 is `supportsStreamData() { r... | To count, a comment must point out that the new InputStream adapter duplicates DBNResource's existing DBNStreamData.openInputStream() / supportsStreamData() and that the paste path should use that contract, or must say that getAdapter opening a stream or throwing an unchecked exception is a bad adapter contract. A comm... | true | [] | [] | [
"R1",
"R2",
"R3",
"R4"
] | null | null | null | dbeaver/dbeaver | bc3ac7bb47092a934b344791d8aa3a1c41e6c8f7 |
dbeaver__dbeaver@bc3ac7b | smoke10 | 2 | D3 | DBNResource uses DBNNode's package-private `log` across OSGi bundles, so the error path throws IllegalAccessError | DBNResource declares no `log` field of its own, so the new `log.error(...)` binds to DBNNode's package-private `static final Log log`. DBNNode is in the org.jkiss.dbeaver.model bundle, while DBNResource is in the separate org.jkiss.dbeaver.model.rcp bundle, which reuses the same package name. The two classes are define... | bug | low | [
{
"path": "plugins/org.jkiss.dbeaver.model.rcp/src/org/jkiss/dbeaver/model/navigator/DBNResource.java",
"from_line": 426,
"to_line": 428,
"role": "cause"
}
] | DBNResource.java:427 `log.error("Error get file content");` is the only `log` reference in the file, and there is no Log field declaration. DBNNode.java:40 `static final Log log = Log.getLog(DBNNode.class);` is package-private. The org.jkiss.dbeaver.model.rcp MANIFEST.MF declares a standalone bundle (`Bundle-SymbolicNa... | To count, a comment must say that `log` in DBNResource is not declared there and resolves to DBNNode's package-private field across bundles / split package, and that this causes IllegalAccessError at runtime, or at least that it is an invalid cross-bundle access. A comment that only says the log message should include ... | true | [] | [] | [] | null | null | null | dbeaver/dbeaver | bc3ac7bb47092a934b344791d8aa3a1c41e6c8f7 |
dbeaver__dbeaver@bc3ac7b | smoke10 | 2 | D4 | Target file name comes from the node's display name instead of the resource's file name | The new branch names the copied file after node.getNodeDisplayName(), which is a UI label, not the file name. The existing Path branch uses resource.getFileName(). For DBNResource subclasses that override the display name, the copy gets the wrong name: DBNBookmark returns the bookmark title, so a pasted remote bookmark... | bug | low | [
{
"path": "plugins/org.jkiss.dbeaver.model/src/org/jkiss/dbeaver/model/navigator/fs/DBNPathBase.java",
"from_line": 295,
"to_line": 296,
"role": "cause"
},
{
"path": "plugins/org.jkiss.dbeaver.model/src/org/jkiss/dbeaver/model/navigator/fs/DBNPathBase.java",
"from_line": 363,
"to_lin... | DBNPathBase.java:295 `String fileName = node.getNodeDisplayName();` and :364 `Path targetFile = folder.resolve(str);`, while the existing branch at :320 uses `folder.resolve(resource.getFileName().toString())`. DBNBookmark.java:62-64 (extends DBNResource) is `public String getNodeDisplayName() { return storage.getTitle... | To count, a comment must say that using getNodeDisplayName() as the target file name is wrong: it is a display label rather than the resource or file name (for example, a bookmark title without its extension), or it is unsanitised and can contain path separators. Comments that only rename the `str` parameter do not cou... | true | [] | [] | [] | null | null | null | dbeaver/dbeaver | bc3ac7bb47092a934b344791d8aa3a1c41e6c8f7 |
linera-io__linera-protocol@3535ddb | smoke10 | 2 | D1 | Client pagination loop is unbounded: a validator returning full pages keeps it running forever and growing remote_log | The new loop in synchronize_received_certificates_from_validator exits only when a page has fewer than CHAIN_INFO_MAX_RECEIVED_LOG_ENTRIES entries. Nothing bounds iterations, checks for progress, or compares against count_received_log. The client verifies only the response signature, not that the entries belong to the ... | security | medium | [
{
"path": "linera-core/src/client/mod.rs",
"from_line": 846,
"to_line": 857,
"role": "cause"
},
{
"path": "linera-core/src/client/mod.rs",
"from_line": 3711,
"to_line": 3713,
"role": "symptom"
},
{
"path": "linera-core/src/client/mod.rs",
"from_line": 4019,
"to_li... | client/mod.rs:848-857 `loop { let query = ChainInfoQuery::new(chain_id).with_received_log_excluding_first_n(offset); let info = remote_node.handle_chain_info_query(query).await?; let received_entries = info.requested_received_log.len(); offset += received_entries as u64; remote_log.extend(info.requested_received_log); ... | To count, a comment must say the new pagination loop can fail to terminate, or grows without bound, because its only exit is receiving a page shorter than the limit, and must name a trigger: a malicious or faulty validator that keeps returning full pages or ignores the offset, or entries being appended as fast as the c... | true | [] | [
"4"
] | [
"R12",
"R15",
"R16",
"R17",
"R18"
] | null | null | null | linera-io/linera-protocol | 3535ddb83de7dfae5db1732a7f1535031d31380a |
linera-io__linera-protocol@3535ddb | smoke10 | 2 | D2 | Hotfixed validators silently truncate the received log for existing (pre-PR) SDK clients | The PR ships as a validator hotfix. After it, every existing SDK client, which makes exactly one request_received_log_excluding_first_n query per sync and treats the vector it gets back as the complete tail, receives only the first 1000 unseen entries. Nothing in the response marks it as truncated, and legacy clients i... | compat | low | [
{
"path": "linera-core/src/chain_worker/state.rs",
"from_line": 1416,
"to_line": 1421,
"role": "cause"
},
{
"path": "linera-core/src/client/mod.rs",
"from_line": 846,
"to_line": 858,
"role": "other-site"
}
] | state.rs:1418-1420 `let end = (start.saturating_add(CHAIN_INFO_MAX_RECEIVED_LOG_ENTRIES)).min(chain.received_log.count()); info.requested_received_log = chain.received_log.read(start..end).await?;` changes what the unchanged wire field returns: the doc on data_types.rs:215-217 still reads as "the response to request_re... | To count, a comment must say that the validator-side cap changes behaviour for older or already-deployed clients that make a single request: they silently get only the first 1000 entries per sync, so syncs are partial and catch-up is slow. Ideally it also suggests gating the cap behind a pagination-aware or versioned r... | false | [] | [] | [
"R1",
"R5"
] | null | null | null | linera-io/linera-protocol | 3535ddb83de7dfae5db1732a7f1535031d31380a |
lvgl__lvgl@4a57db3 | smoke10 | 2 | D1 | New public lv_draw_label_dsc_t.text_size is dead and misleading: only lv_label sets it and its only consumer's value is ignored | The PR adds a public field to lv_draw_label_dsc_t, documented as 'The size of the text', and makes lv_draw_label_iterate_characters use dsc->text_size.x as the line width when EXPAND is set. Only lv_label's draw_main fills it. The roller, the table, the one-line textarea placeholder and any user code that sets LV_TEXT_... | maintainability | low | [
{
"path": "src/draw/lv_draw_label.h",
"from_line": 38,
"to_line": 39,
"role": "cause"
},
{
"path": "src/draw/lv_draw_label.c",
"from_line": 215,
"to_line": 222,
"role": "cause"
},
{
"path": "src/widgets/label/lv_label.c",
"from_line": 826,
"to_line": 826,
"rol... | Head lv_draw_label.h:38-39 adds '/**The size of the text*/ lv_point_t text_size;' to the public descriptor. lv_draw_label.c:221 reads 'w = dsc->text_size.x;' in the EXPAND branch. w is used only as max_width in lv_text_get_next_line calls at lv_draw_label.c:259, 265 and 509. lv_text.c:336 is 'if((flag & LV_TEXT_FLAG_EX... | To count, a comment must say that the new descriptor field (or the w it feeds) is unnecessary, dead or misleading. Either it says the value is ignored because lv_text_get_next_line disregards max_width under EXPAND, or it says only lv_label populates it, so other EXPAND users (roller, table, textarea placeholder, user ... | true | [] | [] | [
"R1",
"R6"
] | null | null | null | lvgl/lvgl | 4a57db33476f1caf3e85726a6b1e189c68c1f36b |
lvgl__lvgl@4a57db3 | smoke10 | 2 | D2 | draw_main now uses a size cached at refr_text time; paths that change the effective font without STYLE_CHANGED/SIZE_CHANGED leave the alignment override and the circular second copy using a stale width | Before the PR, draw_main measured the text at draw time with the font and letter and line spacing actually used for drawing. It now reads label->text_size, which only lv_label_refr_text writes. refr_text runs only on LV_EVENT_STYLE_CHANGED or LV_EVENT_SIZE_CHANGED, or from the label setters. Some paths change the inher... | bug | low | [
{
"path": "src/widgets/label/lv_label.c",
"from_line": 893,
"to_line": 899,
"role": "symptom"
},
{
"path": "src/widgets/label/lv_label.c",
"from_line": 854,
"to_line": 857,
"role": "symptom"
},
{
"path": "src/widgets/label/lv_label.c",
"from_line": 967,
"to_line":... | Head lv_label.c:894 'lv_point_t size = label->text_size;' then 898 'label_draw_dsc.ofs_x = label->offset.x + size.x + lv_font_get_glyph_width(label_draw_dsc.font, ...)', which mixes the cached width with the draw-time font. lv_label.c:856 does the same for the align override. The only writer is lv_label.c:968 'label->t... | To count, a comment must say that the SCROLL or SCROLL_CIRCULAR draw logic (alignment override or circular copy offset) now relies on label->text_size, which is refreshed only in lv_label_refr_text. It must also say that the value can go stale when the font, letter_space or line_space in effect at draw time changes wit... | true | [] | [
"1"
] | [] | null | null | null | lvgl/lvgl | 4a57db33476f1caf3e85726a6b1e189c68c1f36b |
lvgl__lvgl@4a57db3 | smoke10 | 2 | D3 | Nema GFX's private copy of the character iterator still runs lv_text_get_size for every EXPAND label draw | The PR removes the per-draw lv_text_get_size call from the shared lv_draw_label_iterate_characters. The Nema GFX backend keeps its own duplicated _draw_label_iterate_characters, and that copy was not updated. It still measures the whole text on every EXPAND label draw, so the claimed draw-time saving does not reach tha... | performance | low | [
{
"path": "src/draw/nema_gfx/lv_draw_nema_gfx_label.c",
"from_line": 433,
"to_line": 442,
"role": "cause"
},
{
"path": "src/draw/lv_draw_label.c",
"from_line": 215,
"to_line": 222,
"role": "other-site"
}
] | Head lv_draw_nema_gfx_label.c:140 calls '_draw_label_iterate_characters(t, dsc, coords);' and lines 438-441 keep 'lv_point_t p; lv_text_get_size(&p, dsc->text, dsc->font, dsc->letter_space, dsc->line_space, LV_COORD_MAX, dsc->flag); w = p.x;' in the EXPAND branch. The shared iterator at lv_draw_label.c:221 now reads 'w... | To count, a comment must point out that the Nema GFX label path (lv_draw_nema_gfx_label.c, its private iterator) still calls lv_text_get_size per draw and did not get the optimization or the matching change. A question form ('should Nema also...') counts. Near-misses that do NOT count: generic 'other backends may need ... | true | [] | [] | [
"R2",
"R3",
"R4",
"R5",
"R7"
] | null | null | null | lvgl/lvgl | 4a57db33476f1caf3e85726a6b1e189c68c1f36b |
bitcoin__bitcoin@724546e | dev10 | 2 | D1 | std::string_view `services` dangles: it is bound to the temporary std::string from strprintf, then read on the next line | `services` is declared as `const std::string_view` and initialized from a conditional expression. Because strprintf returns std::string, the expression is a temporary std::string (the "" branch converts to std::string too), and that temporary is destroyed at the end of the declaration. The next statement passes the dan... | bug | high | [
{
"path": "src/bitcoin-cli.cpp",
"from_line": 585,
"to_line": 585,
"role": "cause"
},
{
"path": "src/bitcoin-cli.cpp",
"from_line": 586,
"to_line": 586,
"role": "symptom"
}
] | src/bitcoin-cli.cpp:585 at head: `const std::string_view services{DetailsRequested() ? strprintf(" - services %s", FormatServices(networkinfo["localservicesnames"])) : ""};` and :586: `std::string result{strprintf("%s client %s%s - server %i%s%s\n\n", CLIENT_NAME, FormatFullVersion(), ChainToString(), networkinfo["prot... | To count, a comment must say that the string_view `services` at line 585 refers to a temporary std::string (the result of strprintf or the ternary) that is destroyed before line 586 uses it: a dangling view, use-after-free or UB. It counts with or without the fix `std::string services`. Near-misses that do not count: a... | true | [] | [
"3"
] | [
"R1",
"R2",
"R3",
"R4",
"R5"
] | null | null | null | bitcoin/bitcoin | 724546e28a5778e1c3f5100406ed7237f76aab55 |
FreeCAD__FreeCAD@699b45c | dev10 | 2 | D1 | SoDatumLabel angle-wrap casts leave the right-hand side double, so the float += double narrowing (C4244) the PR targets remains | The PR rewrites `range += 2.0 * M_PI` and `endangle += 2. * M_PI` as `2.0 * static_cast<float>(M_PI)` and `2. * static_cast<float>(M_PI)`. Because `2.0` and `2.` are double literals, double * float still produces a double. Each compound assignment is still float += double with an implicit narrowing, so the MSVC C4244 w... | incomplete-fix | low | [
{
"path": "src/Gui/SoDatumLabel.cpp",
"from_line": 82,
"to_line": 85,
"role": "cause"
},
{
"path": "src/Gui/SoDatumLabel.cpp",
"from_line": 551,
"to_line": 553,
"role": "other-site"
},
{
"path": "src/Gui/SoDatumLabel.cpp",
"from_line": 694,
"to_line": 696,
"ro... | The head has `float range = endAngle - startAngle; if (range < 0.) { range += 2.0 * static_cast<float>(M_PI); }` at SoDatumLabel.cpp:82-84. At :552 and :695 it has `float endangle = atan2f(...); if (endangle < startangle) { endangle += 2. * static_cast<float>(M_PI); }`. Under the usual arithmetic conversions, a double ... | To count, a comment must say that `2.0`/`2.` is a double literal, so the expression stays double. It must also say either that the double-to-float conversion/warning is not fixed, or that `2.0F`/`2.0f` is needed. Pointing at any of the three sites counts. Near-misses that do not count: comments only about float vs doub... | true | [] | [
"10",
"11"
] | [] | null | null | null | FreeCAD/FreeCAD | 699b45c086f53858897aa8c688bbedaf570c644f |
FreeCAD__FreeCAD@699b45c | dev10 | 2 | D2 | StackWalker prints DWORD (unsigned long) arguments with the signed %ld specifier instead of %lu | The PR replaces `%d` with `%ld` in every StackWalker diagnostic format that takes DWORD arguments: module size and result, the file-version words v1..v4, entry.lineNumber, GetLastError `gle`, symOptions, and the OS-version fields. DWORD is `unsigned long`, so the matching specifier is `%lu`. As written, specifier and a... | incomplete-fix | low | [
{
"path": "src/Base/StackWalker.cpp",
"from_line": 1473,
"to_line": 1473,
"role": "cause"
},
{
"path": "src/Base/StackWalker.cpp",
"from_line": 1489,
"to_line": 1490,
"role": "other-site"
},
{
"path": "src/Base/StackWalker.cpp",
"from_line": 1540,
"to_line": 1540,... | StackWalker.cpp:1455-1462 has `void StackWalker::OnLoadModule(..., DWORD size, DWORD result, ...)`, and :1473 formats them with `"... size: %ld (result: %ld) ..."`. At :1483-1486, `DWORD v4 = (DWORD)(fileVersion & 0xFFFF); ...` is printed with `"%ld.%ld.%ld.%ld"` at :1490. `entry.lineNumber` (DWORD) uses `%ld` at :1540... | To count, a comment must say that the new `%ld` is signed while the DWORD argument is unsigned (unsigned long), and that `%lu` or another unsigned conversion is the right specifier. Any one of the listed lines counts. Near-misses that do not count: claims that `%ld` is the wrong size, since long and DWORD are both 32-b... | true | [] | [
"1",
"2",
"3",
"4",
"5",
"6"
] | [
"R1",
"R2",
"R3",
"R4",
"R5",
"R6",
"R7",
"R8",
"R9"
] | null | null | null | FreeCAD/FreeCAD | 699b45c086f53858897aa8c688bbedaf570c644f |
CherryHQ__cherry-studio@83a2594 | dev10 | 2 | D1 | useSystemAgents / useRemoteSystemAgents call hooks conditionally (rules-of-hooks violation): CI lint fails, and hook order differs between the local and remote sources | useSystemAgents returns `useLocalSystemAgents()` or `useRemoteSystemAgents()` from different `if` branches. The two hooks call different hook sequences: local calls useState, then useRuntime (a selector hook), then useEffect; remote calls useState, then useEffect. useRemoteSystemAgents also calls `useEffect` after an e... | build | high | [
{
"path": "src/renderer/src/pages/agents/index.ts",
"from_line": 19,
"to_line": 30,
"role": "cause"
},
{
"path": "src/renderer/src/pages/agents/index.ts",
"from_line": 32,
"to_line": 40,
"role": "cause"
},
{
"path": "src/renderer/src/pages/agents/index.ts",
"from_line... | index.ts:20-29 `const { defaultaides } = store.getState().settings; if (defaultaides === null || defaultaides === undefined) { ...; return useLocalSystemAgents(); } if (!defaultaides.startsWith('http')) { return useLocalSystemAgents(); } else { return useRemoteSystemAgents(); }`. index.ts:36-40 `if (defaultaides === nu... | To count, a comment must say one of these: the local and remote hooks are selected/called conditionally; the two branches have different hook sequences, so hook order changes and React can throw a hook-order error; useEffect is called after an early return; or this violates the Rules of Hooks / fails react-hooks lint. ... | true | [] | [] | [
"R14",
"R15",
"R16",
"R17",
"R18"
] | null | null | null | CherryHQ/cherry-studio | 83a25941859179b635105c9b4fd79d5542e3701e |
CherryHQ__cherry-studio@83a2594 | dev10 | 2 | D2 | New settings UI strings exist only in zh-CN; en-US (fallback), ja, ru, zh-TW and the machine-translated locales show raw i18n keys | The new menu entry and panel call t('settings.data.default_aides.title'), '...repo_url' and '...repo_url_placeholder'. Only zh-cn.json defines `settings.data.default_aides`. en-us, ja-jp, ru-ru and zh-tw instead add a flat, unused `settings.defaultaides` string at the wrong path under a different name. en-US is the fal... | bug | medium | [
{
"path": "src/renderer/src/i18n/locales/en-us.json",
"from_line": 1391,
"to_line": 1391,
"role": "cause"
},
{
"path": "src/renderer/src/i18n/locales/ja-jp.json",
"from_line": 1391,
"to_line": 1391,
"role": "other-site"
},
{
"path": "src/renderer/src/i18n/locales/ru-ru.js... | DataSettings.tsx:116 `title: 'settings.data.default_aides.title'`, rendered at :180 `title={t(item.title || '')}`. DefaultaidesSettings.tsx:26/29/36 use `t('settings.data.default_aides.title' | '.repo_url' | '.repo_url_placeholder')`. Parsing the head JSON gives: en-us/ja-jp have `settings.defaultaides` = 'Built-in'/'内... | To count, a comment must say that the keys used by the new UI (`settings.data.default_aides.*`) are missing from the non-zh-CN locales, or that en/ja/ru/zh-TW define `defaultaides` at a different name/path, so users of those languages (and the en-US fallback) see raw keys or untranslated labels. Pointing at any one of ... | true | [] | [
"0",
"1",
"2",
"4"
] | [
"R1",
"R2",
"R3",
"R4",
"R5",
"R6",
"R7",
"R8",
"R29",
"R30",
"R32",
"R33"
] | null | null | null | CherryHQ/cherry-studio | 83a25941859179b635105c9b4fd79d5542e3701e |
CherryHQ__cherry-studio@83a2594 | dev10 | 2 | D3 | A failed remote load leaves the system assistant list empty; no fallback to the bundled agents despite the PR's 'fallback mechanism' claim | With an http URL configured, useRemoteSystemAgents starts from `[]`. Any failure (offline, non-2xx, invalid JSON) is caught and only console.error'd. The bundled resources/data/agents.json is never read. The Agents page and the Add Assistant popup then show no built-in assistants at all, and the user gets no indication... | incomplete-fix | medium | [
{
"path": "src/renderer/src/pages/agents/index.ts",
"from_line": 50,
"to_line": 52,
"role": "cause"
},
{
"path": "src/renderer/src/pages/agents/index.ts",
"from_line": 33,
"to_line": 33,
"role": "other-site"
},
{
"path": "src/renderer/src/pages/agents/AgentsPage.tsx",
... | index.ts:33 `const [agents, setAgents] = useState<Agent[]>([]);`; :45-47 `if (!response.ok) { throw new Error(...) }`; :50-52 `} catch (error) { console.error("Failed to load agents:", error); }`; :58 `return agents;`. There is no call into the local loader or `_agents` on failure. Compare useLocalSystemAgents at :61-7... | To count, a comment must say that when the remote fetch/parse fails, the built-in/bundled assistants are not used as a fallback and the system agent list stays empty. Does NOT count: a generic 'handle errors / show an error toast' comment that does not mention losing the bundled list; payload validation (D4); abort/cle... | true | [] | [] | [
"R24",
"R25",
"R26",
"R27"
] | null | null | null | CherryHQ/cherry-studio | 83a25941859179b635105c9b4fd79d5542e3701e |
CherryHQ__cherry-studio@83a2594 | dev10 | 2 | D4 | Remote JSON is cast to Agent[] without runtime validation; a wrongly shaped payload crashes the Agents page and the Add Assistant popup | `await response.json() as Agent[]` publishes whatever the user-configured URL returns. If it is a JSON object (for example a wrapped `{agents:[...]}` or an API response), AgentsPage's effect calls `groupByCategories(systemAgents)`, and `data.forEach` throws. AddAssistantPopup spreads `...systemAgents` inside useMemo du... | bug | medium | [
{
"path": "src/renderer/src/pages/agents/index.ts",
"from_line": 48,
"to_line": 49,
"role": "cause"
},
{
"path": "src/renderer/src/pages/agents/index.ts",
"from_line": 77,
"to_line": 80,
"role": "symptom"
},
{
"path": "src/renderer/src/pages/agents/AgentsPage.tsx",
"f... | index.ts:48-49 `const agentsData = await response.json() as Agent[]; setAgents(agentsData);`. AgentsPage.tsx:33 `const systemAgentsGroupList = groupByCategories(systemAgents)` leads to index.ts:79-80 `data.forEach((item) => { item.group?.forEach(`. AddAssistantPopup.tsx:38 `const allAgents = [...userAgents, ...systemAg... | To count, a comment must say that the fetched remote JSON is trusted/cast as Agent[] without checking its shape, and that a non-array or malformed payload breaks the consumers (a crash in groupByCategories/forEach/spread/name access) or should be treated as a load error. Does NOT count: the no-fallback-on-failure point... | true | [] | [] | [
"R21",
"R22",
"R23"
] | null | null | null | CherryHQ/cherry-studio | 83a25941859179b635105c9b4fd79d5542e3701e |
CherryHQ__cherry-studio@83a2594 | dev10 | 2 | D5 | The new setting's initial value is the raw i18n key 'settings.defaultaides', shown verbatim in the URL input | initialState sets `defaultaides: 'settings.defaultaides'`: a translation key stored as the setting's value and never passed through t(). On a fresh install the Assistant URL input binds directly to it and displays the literal 'settings.defaultaides'. The 'Built-in' strings added to the locales are never rendered. The v... | bug | low | [
{
"path": "src/renderer/src/store/settings.ts",
"from_line": 209,
"to_line": 210,
"role": "cause"
},
{
"path": "src/renderer/src/pages/settings/DataSettings/DefaultaidesSettings.tsx",
"from_line": 16,
"to_line": 37,
"role": "symptom"
}
] | settings.ts:210 `defaultaides: 'settings.defaultaides', // 默认智能体`. DefaultaidesSettings.tsx:16 `const Aides = useSelector((state: RootState) => state.settings.defaultaides)` and :33 `value={Aides || ''}`, with no t(). index.ts:25 `if (!defaultaides.startsWith('http')) { return useLocalSystemAgents(); }`. | To count, a comment must say that the default/initial value is an i18n key (or an internal-looking literal) instead of an empty/null/real value, and that it is shown to users in the input or only works because it is not http. Does NOT count: comments about the missing `settings.data.default_aides` translations (D2); a ... | true | [] | [
"16"
] | [
"R35",
"R36"
] | null | null | null | CherryHQ/cherry-studio | 83a25941859179b635105c9b4fd79d5542e3701e |
CherryHQ__cherry-studio@83a2594 | dev10 | 2 | D6 | useSystemAgents reads the source setting via store.getState() during render: consumers never subscribe to changes | The source selection reads `store.getState().settings.defaultaides` instead of subscribing via useSelector/useAppSelector. A change to the setting therefore never re-renders the hook's consumers, and a mounted consumer keeps the old source until something unrelated re-renders it. This is latent today, because the two c... | maintainability | low | [
{
"path": "src/renderer/src/pages/agents/index.ts",
"from_line": 20,
"to_line": 20,
"role": "cause"
},
{
"path": "src/renderer/src/pages/agents/index.ts",
"from_line": 34,
"to_line": 35,
"role": "other-site"
}
] | index.ts:20 `const { defaultaides } = store.getState().settings` and :34 the same again inside useRemoteSystemAgents. The writer is DefaultaidesSettings.tsx:18-19 `dispatch(setefaultAides(e.target.value))`. Consumers are AgentsPage.tsx:29 and AddAssistantPopup.tsx:32. App.tsx routes have no keep-alive, so AgentsPage re... | To count, a comment must say that store.getState() is a non-reactive snapshot, so consumers do not update/re-fetch when the URL setting changes, and should use useSelector/useAppSelector. Does NOT count: a comment that only flags the conditional hook calls or hook order (that is D1). A comment that makes both points co... | true | [] | [
"6"
] | [
"R9",
"R10",
"R11",
"R12",
"R13"
] | null | null | {"bots": ["Sourcery"], "dataset_refs": [6], "references": ["src/renderer/src/pages/agents/index.ts:19-19:2956a397d0b4"], "hunt": true, "note": "Not scored. The cited AACR reference is a reworded review-bot comment from the original PR; the defect itself was checked against the code."} | CherryHQ/cherry-studio | 83a25941859179b635105c9b4fd79d5542e3701e |
CherryHQ__cherry-studio@83a2594 | dev10 | 2 | D7 | Remote agents are fetched on every mount after a hard-coded 500 ms sleep and are never cached | Unlike the local path, which caches in the module-level `_agents`, useRemoteSystemAgents starts from `[]`, sleeps 500 ms for no reason, and re-fetches the URL every time AgentsPage mounts or the Add Assistant popup opens. Each open costs at least 500 ms plus a network round-trip, during which the list holds only user a... | performance | low | [
{
"path": "src/renderer/src/pages/agents/index.ts",
"from_line": 41,
"to_line": 56,
"role": "cause"
}
] | index.ts:33 `useState<Agent[]>([])`; :43 `await new Promise(resolve => setTimeout(resolve, 500));`; :44 `await fetch(resourcesPath)`. Nothing writes `_agents`. Compare :67 `if (!resourcesPath || _agents.length > 0) return` in the local hook. | To count, a comment must say that the artificial 500 ms delay slows every load, or that the remote list is refetched on every mount with no caching (unlike the local `_agents` cache). Does NOT count: a comment that only asks for an AbortController/cleanup or warns about setState after unmount (in React 19 that is a no-... | true | [] | [] | [
"R20"
] | null | null | null | CherryHQ/cherry-studio | 83a25941859179b635105c9b4fd79d5542e3701e |
CherryHQ__cherry-studio@83a2594 | dev10 | 2 | D8 | No redux-persist migration for the new settings field: upgraded installs get undefined and log console.error on every render | The store persists `settings` with redux-persist's default level-1 merge, so a rehydrated settings object replaces initialState.settings. The repo adds each new settings field through a numbered migration in migrate.ts plus a version bump. This PR adds neither, so every existing install has `defaultaides === undefined`... | compat | low | [
{
"path": "src/renderer/src/store/index.ts",
"from_line": 41,
"to_line": 50,
"role": "cause"
},
{
"path": "src/renderer/src/store/migrate.ts",
"from_line": 1232,
"to_line": 1242,
"role": "other-site"
},
{
"path": "src/renderer/src/pages/agents/index.ts",
"from_line": ... | store/index.ts:41-47 `persistReducer({ key: 'cherry-studio', storage, version: 96, blacklist: ['runtime', 'messages'], migrate }, rootReducer)` has no stateReconciler, so it uses the default autoMergeLevel1. migrate.ts has no entry setting `state.settings.defaultaides`; repo convention is shown by e.g. '94': `state.set... | To count, a comment must say that there is no migration/persist version bump for `defaultaides`, so existing users' persisted settings lack it (undefined), with any consequence such as the error log or an inconsistent default. Does NOT count: the raw-i18n-key default value (D5); the duplicated null check (ref 7). | true | [] | [] | [] | null | null | null | CherryHQ/cherry-studio | 83a25941859179b635105c9b4fd79d5542e3701e |
CherryHQ__cherry-studio@83a2594 | dev10 | 2 | D9 | The exported Redux action is misspelled `setefaultAides` | The new reducer and exported action creator are named `setefaultAides`, missing the 'D' of Default. It becomes part of the settings slice's public API. Anyone searching for or importing `setDefaultAides` will not find it, and renaming it after merge means churn across consumers. | maintainability | low | [
{
"path": "src/renderer/src/store/settings.ts",
"from_line": 459,
"to_line": 461,
"role": "cause"
},
{
"path": "src/renderer/src/store/settings.ts",
"from_line": 576,
"to_line": 576,
"role": "other-site"
},
{
"path": "src/renderer/src/pages/settings/DataSettings/Defaultai... | settings.ts:459 `setefaultAides: (state, action: PayloadAction<string>) => {`; :576 `setefaultAides,` in the export list; DefaultaidesSettings.tsx:4 `import { setefaultAides } from '@renderer/store/settings'`. | To count, a comment must point out the `setefaultAides` typo (missing D) in the action/reducer name. Does NOT count: casing preferences for `defaultaides` / `DefaultaidesSettings` (taste, excluded). | false | [] | [
"12",
"17",
"18"
] | [
"R31",
"R37"
] | null | null | {"bots": ["Sourcery"], "dataset_refs": [17], "references": ["src/renderer/src/store/settings.ts:459-459:12e8c1434ab6"], "hunt": false, "note": "Not scored. The cited AACR reference is a reworded review-bot comment from the original PR; the defect itself was checked against the code."} | CherryHQ/cherry-studio | 83a25941859179b635105c9b4fd79d5542e3701e |
bluewave-labs__Checkmate@29e03a6 | dev10 | 2 | D1 | Manual and scheduled runs use different English codes (en vs gb); one default path writes a stray en.json and leaves gb.json stale, the other asks POEditor for 'gb' | The workflow_dispatch input defaults to "tr,en", but the env fallback that applies on the nightly schedule is 'tr,gb'. The script's own default is 'tr,en' and the README says the default is tr and gb. The script writes temp/<code>.json verbatim and the workflow copies it into src/locales, while the app is keyed on "gb"... | bug | medium | [
{
"path": ".github/workflows/poeditor-sync.yml",
"from_line": 7,
"to_line": 10,
"role": "cause"
},
{
"path": ".github/workflows/poeditor-sync.yml",
"from_line": 60,
"to_line": 60,
"role": "other-site"
},
{
"path": ".github/scripts/download-translations.js",
"from_line... | poeditor-sync.yml:10 `default: "tr,en"` vs :60 `LANGUAGES: ${{ github.event.inputs.languages || 'tr,gb' }}`. The fallback applies only when the input is empty, which is always the case on schedule. download-translations.js:12 `(process.env.LANGUAGES || 'tr,en').split(',')`; :64 `path.join(process.cwd(), 'temp', `${lang... | To count, a comment must point out the en-vs-gb language-code mismatch between any two of: the manual input default (tr,en), the scheduled fallback (tr,gb), the script default (tr,en) and the README (tr,gb). Stronger hits add the consequence: the app uses gb.json, so the manual default writes en.json and leaves gb stal... | true | [] | [
"3",
"6",
"7"
] | [
"R5",
"R6",
"R7",
"R8",
"R12"
] | null | null | {"bots": ["CodeRabbit"], "dataset_refs": [6, 7], "references": [".github/workflows/README.md:25-25:ef424660ad45", ".github/workflows/poeditor-sync.yml:60-61:dc5b3f20fd61"], "hunt": true, "note": "Not scored. The cited AACR reference is a reworded review-bot comment from the original PR; the defect itself was checked ag... | bluewave-labs/Checkmate | 29e03a6294edfe43b49eba20fe49099cef601d34 |
bluewave-labs__Checkmate@29e03a6 | dev10 | 2 | D2 | key_value_json export is written unfiltered, so untranslated terms land in tr.json as "" and render blank instead of falling back to gb | The default key_value_json path saves POEditor's export exactly as received. The export request passes no filters or fallback_language, and POEditor includes untranslated terms with empty-string values. Today tr.json holds only the translated subset (105 lines vs 178 in gb.json), so missing Turkish keys fall back to gb... | bug | medium | [
{
"path": ".github/scripts/download-translations.js",
"from_line": 30,
"to_line": 35,
"role": "cause"
},
{
"path": ".github/scripts/download-translations.js",
"from_line": 49,
"to_line": 65,
"role": "symptom"
}
] | The export body is only `{api_token, id, language, type}` (script:29-35): no `filters: 'translated'` and no `fallback_language`. Script:50 `let formattedTranslations = translations;`. Only the `Array.isArray` branch (:53-60) filters, via `if (item.term && item.definition)`, and :65 writes the object unchanged. src/Util... | To count, a comment must say that untranslated terms come back from POEditor as empty strings, are written into the locale file, and show blank text instead of falling back to gb/English. Acceptable fixes it might propose: filter empty values, or use filters=translated or fallback_language. A near-miss that does not co... | true | [] | [] | [] | null | null | null | bluewave-labs/Checkmate | 29e03a6294edfe43b49eba20fe49099cef601d34 |
bluewave-labs__Checkmate@29e03a6 | dev10 | 2 | D3 | One-way download sync overwrites locale files wholesale, deleting keys developers add in the repo before they reach POEditor | The workflow only downloads. It replaces src/locales/<lang>.json with POEditor's export (cp temp/* src/locales/) and pushes the result straight to the branch. The script does no merge with the existing file and no upload. Developers in this repo add strings by editing gb.json and tr.json directly in PRs, so every sync ... | data | low | [
{
"path": ".github/workflows/poeditor-sync.yml",
"from_line": 80,
"to_line": 92,
"role": "cause"
},
{
"path": ".github/scripts/download-translations.js",
"from_line": 63,
"to_line": 65,
"role": "other-site"
}
] | poeditor-sync.yml:83 `cp -r temp/* src/locales/`, :90-92 `git add src/locales/*.json` / `git commit` / `git push`. script:65 `fs.writeJson(outputPath, formattedTranslations, ...)` writes only what POEditor returned. History at the merge base shows direct locale edits: d40507e (+30 gb.json), 93462d7 (+34 gb.json), a72d0... | To count, a comment must say the sync replaces the repo's locale files with the POEditor export, with no merge and no upload of repo-side terms, so keys added directly in the repo are lost or reverted on the next run. Near-misses that do not count: the 'empty {} export' comments (R13-R16), which are about an empty expo... | true | [] | [] | [] | null | null | null | bluewave-labs/Checkmate | 29e03a6294edfe43b49eba20fe49099cef601d34 |
bluewave-labs__Checkmate@29e03a6 | dev10 | 2 | D4 | Workflow pushes with GITHUB_TOKEN but requests no contents: write permission | The Commit step runs `git push` with the checkout's GITHUB_TOKEN, but the workflow and job declare no `permissions:` block. On a repo or org whose default workflow token is read-only (GitHub's default for newer orgs and repos), the job downloads and commits and then fails on the push with a 403. Translations are never ... | build | low | [
{
"path": ".github/workflows/poeditor-sync.yml",
"from_line": 20,
"to_line": 28,
"role": "cause"
},
{
"path": ".github/workflows/poeditor-sync.yml",
"from_line": 86,
"to_line": 92,
"role": "symptom"
}
] | poeditor-sync.yml has no `permissions:` key anywhere. Line 28 is `token: ${{ secrets.GITHUB_TOKEN }}` and line 92 is `git push`. There are no other workflows in the repo at the merge base, so nothing shows the repo's default token setting. | To count, a comment must say the workflow needs `permissions: contents: write`, or equivalent, for the push, because the default GITHUB_TOKEN may be read-only. Near-misses that do not count: comments about which branch is pushed to (R19-R22), and token-exposure or supply-chain comments (R11). | true | [] | [] | [
"R9",
"R17",
"R18",
"R23"
] | null | null | null | bluewave-labs/Checkmate | 29e03a6294edfe43b49eba20fe49099cef601d34 |
bluewave-labs__Checkmate@29e03a6 | dev10 | 2 | D5 | LANGUAGES is split on ',' without trimming or dropping empties, so 'tr, en' or a trailing comma requests ' en' or '' | The workflow_dispatch languages input is free text, and the script only splits it on ','. An operator who types 'tr, en' or leaves a trailing comma sends ' en' or '' as the POEditor language. That either fails the whole run (exit 1 after the first bad export) or writes an oddly named temp/' en'.json, which is copied in... | bug | low | [
{
"path": ".github/scripts/download-translations.js",
"from_line": 12,
"to_line": 12,
"role": "cause"
},
{
"path": ".github/workflows/poeditor-sync.yml",
"from_line": 7,
"to_line": 10,
"role": "other-site"
}
] | script:12 `const LANGUAGES = (process.env.LANGUAGES || 'tr,en').split(',');` with no `.map(s => s.trim()).filter(Boolean)`. The values are used verbatim at :38 `language: language` and :64 `` `${language}.json` ``. The input at workflow:7-10 is free text described as 'comma separated'. | To count, a comment must say the language list is not trimmed or filtered, so spaces or empty entries produce invalid or oddly named language codes. Near-misses that do not count: the en/gb default mismatch (D1) and generic 'validate env vars' comments (dataset ref 1). | false | [] | [
"2"
] | [] | null | null | null | bluewave-labs/Checkmate | 29e03a6294edfe43b49eba20fe49099cef601d34 |
bluewave-labs__Checkmate@29e03a6 | dev10 | 2 | D6 | json (array) export conversion keys on item.term only, ignoring context, so nested keys collapse and collide | The workflow offers format=json as a supported input. For the array that export returns, the script builds a flat map keyed only by item.term and ignores item.context. POEditor stores the repo's nested keys (notifications.slack.label, notifications.discord.label, ...) as a term plus a parent-path context. Terms like la... | bug | low | [
{
"path": ".github/scripts/download-translations.js",
"from_line": 53,
"to_line": 60,
"role": "cause"
},
{
"path": ".github/workflows/poeditor-sync.yml",
"from_line": 11,
"to_line": 14,
"role": "other-site"
}
] | script:56-58 `translations.forEach(item => { if (item.term && item.definition) { formattedTranslations[item.term] = item.definition; } })`, which never reads context. workflow:12 `description: "Export format (key_value_json or json)"`. src/locales/gb.json nests `"notifications": { "slack": { "label" ... }, "discord": {... | To count, a comment must say the array/json conversion drops context or nesting, so nested keys are flattened or terms collide and the locale no longer matches the nested lookups. A near-miss that does not count is R2, which only asks whether definitions can be objects (plurals). Comments about dropping empty definitio... | true | [] | [] | [] | null | null | null | bluewave-labs/Checkmate | 29e03a6294edfe43b49eba20fe49099cef601d34 |
mudler__LocalAI@e466450 | dev10 | 2 | D1 | Non-streaming /v1/completions now always reports 0 prompt/completion/total tokens | In CompletionEndpoint's non-streaming loop, the PR deleted `totalTokenUsage.Prompt += tokenUsage.Prompt` and `totalTokenUsage.Completion += tokenUsage.Completion` and put the two timing accumulations in their place. totalTokenUsage's token fields therefore stay at their zero value, and the response's usage.prompt_token... | bug | high | [
{
"path": "core/http/endpoints/openai/completion.go",
"from_line": 183,
"to_line": 184,
"role": "cause"
},
{
"path": "core/http/endpoints/openai/completion.go",
"from_line": 188,
"to_line": 192,
"role": "symptom"
},
{
"path": "core/http/endpoints/openai/completion.go",
... | Merge base completion.go (non-streaming loop) had `totalTokenUsage.Prompt += tokenUsage.Prompt` / `totalTokenUsage.Completion += tokenUsage.Completion`. At head, lines 183-184 are only `totalTokenUsage.TimingTokenGeneration += tokenUsage.TimingTokenGeneration` / `totalTokenUsage.TimingPromptProcessing += tokenUsage.Tim... | To count, a comment must say that the non-streaming completion path no longer accumulates Prompt/Completion token counts, because the timing lines replaced them, so the reported prompt_tokens/completion_tokens/total_tokens are 0 (or wrong). Pointing at completion.go ~183-184 or the usage block at 188-192 with that mech... | true | [] | [
"12"
] | [
"R22",
"R23",
"R24",
"R25",
"R26"
] | null | null | null | mudler/LocalAI | e466450c007539a2b1e4d2951c6756ff004a2b52 |
mudler__LocalAI@e466450 | dev10 | 2 | D2 | Machine tag defaults to the server hostname and is sent to every client on most endpoints | WithMachineTag falls back to os.Hostname() whenever LOCALAI_MACHINE_TAG is unset, and the CLI always applies it. About 30 handlers then emit it unconditionally as the LocalAI-Machine-Tag response header. As a result, every LocalAI deployment, not only P2P users who opt in, discloses its internal hostname to any API cli... | security | low | [
{
"path": "core/config/application_config.go",
"from_line": 100,
"to_line": 108,
"role": "cause"
},
{
"path": "core/cli/run.go",
"from_line": 111,
"to_line": 111,
"role": "other-site"
},
{
"path": "core/http/endpoints/openai/chat.go",
"from_line": 174,
"to_line": ... | application_config.go:100-107: `func WithMachineTag(tag string) AppOption { return func(o *ApplicationConfig) { if tag == "" { hostname, _ := os.Hostname(); tag = hostname }; o.MachineTag = tag } }`. run.go:111 always calls `config.WithMachineTag(r.MachineTag)`. Handlers such as chat.go:174 `c.Set("LocalAI-Machine-Tag"... | To count, a comment must say that the hostname (os.Hostname default) is exposed to clients by default or unconditionally in a response header, as information disclosure, and/or that the machine-tag header should be opt-in or omitted when no tag is configured. Near-misses that do NOT count: 'the os.Hostname error is ign... | true | [] | [
"16"
] | [] | null | null | null | mudler/LocalAI | e466450c007539a2b1e4d2951c6756ff004a2b52 |
mudler__LocalAI@e466450 | dev10 | 2 | D3 | Machine-tag header set by hand in each handler instead of a middleware; coverage already inconsistent | Instead of one middleware, the LocalAI-Machine-Tag response header is set by a hand-copied `c.Set(...)` line at the top of about 30 individual handlers. Coverage is already inconsistent. The ElevenLabs text-to-speech and sound-generation inference endpoints, /v1/tokenize, /stores/*, the gallery, backend-monitor, /syste... | maintainability | low | [
{
"path": "core/http/endpoints/openai/assistant.go",
"from_line": 79,
"to_line": 79,
"role": "cause"
},
{
"path": "core/http/endpoints/openai/files.go",
"from_line": 86,
"to_line": 86,
"role": "other-site"
},
{
"path": "core/http/endpoints/openai/list.go",
"from_line"... | `git grep LocalAI-Machine-Tag` at head matches assistant.go (9 times: 79, 141, 251, 284, 314, 361, 419, 459, 514), files.go (5 times), chat.go (twice: 174 and 330), and rerank.go, localai/tts.go, vad.go, completion.go, edit.go, embeddings.go, image.go, list.go and transcription.go (once each). It has no match in core/h... | To count, a comment must say one of two things. Either: the header is duplicated per handler and belongs in one shared middleware or route scope, with the drift risk. Or: specific routes (ElevenLabs TTS/sound-generation, tokenize, stores, etc.) are missing the machine-tag header, so coverage is inconsistent. Either hal... | true | [] | [
"15"
] | [
"R4",
"R5",
"R6",
"R7",
"R27"
] | null | null | null | mudler/LocalAI | e466450c007539a2b1e4d2951c6756ff004a2b52 |
mudler__LocalAI@e466450 | dev10 | 2 | D4 | New --machine-tag CLI flag ships with placeholder help text 'TODO: write a help string' | The new MachineTag option on RunCMD has the literal help text "TODO: write a help string" and no `group:` tag. `local-ai run --help` therefore shows a placeholder to operators, and the flag's purpose is undocumented. Unlike its neighbours, it also lacks the un-prefixed env alias, a MACHINE_TAG alongside LOCALAI_MACHINE... | maintainability | low | [
{
"path": "core/cli/run.go",
"from_line": 74,
"to_line": 74,
"role": "cause"
}
] | run.go:74: `MachineTag string `env:"LOCALAI_MACHINE_TAG" help:"TODO: write a help string"``. Neighbouring flags, e.g. run.go:73 `LoadToMemory []string `env:"LOCALAI_LOAD_TO_MEMORY,LOAD_TO_MEMORY" help:"A list of models to load into memory at startup" group:"models"``, all carry a real help string, a group and an alias. | To count, a comment must point out that the new MachineTag flag's help string is a TODO placeholder (user-visible in --help) or that it is otherwise undocumented or unfinished. A comment only about the missing group or env alias, without the placeholder, is a weak match. Near-misses: general 'document the timing units'... | true | [] | [
"6"
] | [] | null | null | null | mudler/LocalAI | e466450c007539a2b1e4d2951c6756ff004a2b52 |
alibaba__spring-ai-alibaba@4bc7305 | dev10 | 2 | D1 | move_to_and_click scrolls the page, then clicks the unadjusted (x,y), so it hits the wrong element | The action now calls window.scrollTo(x - viewportWidth/2, y - viewportHeight/2) and then still clicks at the original (x, y). Playwright's mouse.click uses viewport coordinates, and callers get (x, y) from get_element_position, which is Locator.boundingBox() and so also viewport-relative. Whenever the scroll moves the ... | bug | high | [
{
"path": "spring-ai-alibaba-jmanus/src/main/java/com/alibaba/cloud/ai/example/manus/tool/browser/actions/MoveToAndClickAction.java",
"from_line": 48,
"to_line": 52,
"role": "cause"
},
{
"path": "spring-ai-alibaba-jmanus/src/main/java/com/alibaba/cloud/ai/example/manus/tool/browser/actions/M... | MoveToAndClickAction.java:49-52 `int scrollX = Math.max(0, x - page.viewportSize().width / 2); int scrollY = Math.max(0, y - page.viewportSize().height / 2); page.evaluate("(args) => window.scrollTo({left: args[0], top: args[1], behavior: 'instant'})", new Object[] { scrollX, scrollY });` then :69-70 `page.mouse().move... | To count, a comment must say that the new scroll moves the page before the click, but the click still uses the same (x,y), which are viewport coordinates from boundingBox/get_element_position. The click therefore lands on the wrong element (or the coordinates are not adjusted by the scroll offset). It is enough to note... | true | [] | [
"6"
] | [
"R13",
"R14",
"R15",
"R16",
"R17"
] | null | null | null | alibaba/spring-ai-alibaba | 4bc73055c7b33f53433d9725eab5010b85a59632 |
alibaba__spring-ai-alibaba@4bc7305 | dev10 | 2 | D2 | input_text falls back to a raw JS el.value assignment and reports success for fields that rejected input | When both Playwright fill attempts fail, the action runs `el.value = value; el.dispatchEvent(new Event('input'))` and returns "成功输入" (success). Fill fails on disabled, readonly and hidden fields and on non-text types such as checkbox or radio. The JS fallback skips the editability and type checks, so it forces a value ... | bug | medium | [
{
"path": "spring-ai-alibaba-jmanus/src/main/java/com/alibaba/cloud/ai/example/manus/tool/browser/actions/InputTextAction.java",
"from_line": 76,
"to_line": 88,
"role": "cause"
},
{
"path": "spring-ai-alibaba-jmanus/src/main/java/com/alibaba/cloud/ai/example/manus/tool/browser/actions/InputT... | InputTextAction.java:62-75 `try { elementLocator.fill(""); ... elementLocator.pressSequentially(text, options); } catch (Exception e) { try { elementLocator.fill(""); elementLocator.fill(text); } catch (Exception e2) { ... elementLocator.evaluate("(el, value) => { el.value = value; el.dispatchEvent(new Event('input', {... | To count, a comment must say that the JS `el.value` fallback bypasses Playwright's editability/actionability (or type) checks and still returns success. The result is a false success for disabled, readonly or hidden inputs, or input that is not accepted, or form handlers that never see the change. Near-misses that do n... | true | [] | [] | [
"R10",
"R11",
"R12"
] | null | null | null | alibaba/spring-ai-alibaba | 4bc73055c7b33f53433d9725eab5010b85a59632 |
alibaba__spring-ai-alibaba@4bc7305 | dev10 | 2 | D3 | navigate prepends https:// to every URL without a lowercase http(s) prefix, breaking other schemes | The new auto-prefix rewrites any URL that does not start with lowercase "http://" or "https://". about:blank becomes https://about:blank, file:///tmp/x.html becomes https://file:///tmp/x.html, and data:, ftp:// and chrome:// URLs are broken the same way. An uppercase HTTP://example.com becomes https://HTTP://example.co... | bug | low | [
{
"path": "spring-ai-alibaba-jmanus/src/main/java/com/alibaba/cloud/ai/example/manus/tool/browser/actions/NavigateAction.java",
"from_line": 41,
"to_line": 46,
"role": "cause"
}
] | NavigateAction.java:42-44 `if (!url.startsWith("http://") && !url.startsWith("https://")) { url = "https://" + url; }` then :46 `page.navigate(url, ...)`. BrowserUseTool.java:223-224 forwards the URL the agent supplies directly to NavigateAction. The merge base passed the URL to page.navigate unchanged. | To count, a comment must say the prefix check only recognises (lowercase) http/https, so URLs that already have another scheme get mangled. Examples are about:, file:, data:, ftp:, or uppercase HTTP. The fix is to prefix only when no scheme is present. A comment that only mentions protocol-relative '//host' also counts... | true | [] | [
"9"
] | [
"R18",
"R19"
] | null | null | null | alibaba/spring-ai-alibaba | 4bc73055c7b33f53433d9725eab5010b85a59632 |
alibaba__spring-ai-alibaba@4bc7305 | dev10 | 2 | D4 | Agent-save fix uses a " " sentinel instead of fixing the inverted `||` guard, so the deprecated-prompt branch fires on every read and write | The NPE the PR works around comes from mergePrompts' guard `getSystemPrompt() != null || !getSystemPrompt().trim().isEmpty()`. That guard is true for every non-null value and dereferences null. Rather than fix it, the PR makes the field never null (a " " default, and " " written back as the 'cleared' value). The cleare... | incomplete-fix | low | [
{
"path": "spring-ai-alibaba-jmanus/src/main/java/com/alibaba/cloud/ai/example/manus/dynamic/agent/service/AgentServiceImpl.java",
"from_line": 190,
"to_line": 190,
"role": "cause"
},
{
"path": "spring-ai-alibaba-jmanus/src/main/java/com/alibaba/cloud/ai/example/manus/dynamic/agent/service/A... | AgentServiceImpl.java:190 `if (entity.getSystemPrompt() != null || !entity.getSystemPrompt().trim().isEmpty()) {` ... :197-199 `log.warn("Agent[{}]的SystemPrompt不为空， 但属性已经废弃...本次将agent 的内容合并... 当前制定的值: {}", agentName, nextPrompt);` :200 `entity.setSystemPrompt(" ");`. DynamicAgentEntity.java:45-47 `@Column(length = 4000... | To count, a comment must identify the mergePrompts guard at :190 (`!= null || !...trim().isEmpty()`) as the real defect the " " sentinel papers over, so it should be a null-safe && / isBlank check. It may name either consequence: the misleading deprecation warning firing for every agent on every read or save, or the gu... | false | [] | [] | [
"R1",
"R2",
"R3",
"R4",
"R5",
"R6",
"R7"
] | null | null | null | alibaba/spring-ai-alibaba | 4bc73055c7b33f53433d9725eab5010b85a59632 |
alibaba__spring-ai-alibaba@4bc7305 | dev10 | 2 | D5 | Debug marker cleanup is not guaranteed, and its failure can turn a successful click into a reported failure | In debug mode (manus.browser.debug=true) the red marker is removed by a page.evaluate placed in the same try block right after mouse.click, not in a finally block. If the click's navigation has already destroyed the page's execution context, that evaluate throws. The catch then returns "Failed to click at position" for... | bug | low | [
{
"path": "spring-ai-alibaba-jmanus/src/main/java/com/alibaba/cloud/ai/example/manus/tool/browser/actions/MoveToAndClickAction.java",
"from_line": 54,
"to_line": 82,
"role": "cause"
}
] | MoveToAndClickAction.java:57-65 inject: `let dot = document.getElementById(id); if (!dot) { dot = document.createElement('div'); ... document.body.appendChild(dot); }`. Then :69-70 `page.mouse().move(x, y); page.mouse().click(x, y);`, then :73-77 `if (isDebug) { page.evaluate("(id) => { const dot = document.getElementB... | To count, a comment must say either that the marker removal is not in a finally or guarded cleanup, so the dot leaks when the click path throws, or that the post-click removal evaluate can itself throw (for example after navigation) and make a successful click report failure. Near-misses that do not count: 'extract the... | true | [] | [
"8"
] | [] | null | null | null | alibaba/spring-ai-alibaba | 4bc73055c7b33f53433d9725eab5010b85a59632 |
openai__codex@8c952af | dev10 | 2 | D1 | codex-core unit tests no longer compile: `store` is removed from the ResponsesApiRequest test literals but the field still exists | The PR keeps `pub(crate) store: bool` on `ResponsesApiRequest` (client_common.rs:128) but deletes `store: true,` from both struct literals in the client_common.rs test module. Neither literal uses `..Default::default()`, and the struct derives only Debug and Serialize. The lib's test target therefore fails with E0063 (... | build | high | [
{
"path": "codex-rs/core/src/client_common.rs",
"from_line": 190,
"to_line": 204,
"role": "cause"
},
{
"path": "codex-rs/core/src/client_common.rs",
"from_line": 219,
"to_line": 231,
"role": "cause"
},
{
"path": "codex-rs/core/src/client_common.rs",
"from_line": 128,
... | Head client_common.rs:117-136 `#[derive(Debug, Serialize)] pub(crate) struct ResponsesApiRequest<'a> { ... pub(crate) reasoning: Option<Reasoning>, pub(crate) store: bool, pub(crate) stream: bool, ... }`. The diff removes `- store: true,` from both test literals. At head, `let req = ResponsesApiRequest { mod... | To count, a finding must say that the ResponsesApiRequest literals in the client_common.rs tests no longer initialize the still-required `store` field, so the core test target fails to compile (missing field). Either literal, or both, is enough. Near-misses that do not count: comments about the misleading `store` doc c... | true | [] | [] | [
"R1",
"R2",
"R3",
"R4",
"R5"
] | null | null | null | openai/codex | 8c952af20fe791cc0c5d0a59089d3329968b4373 |
openai__codex@8c952af | dev10 | 2 | D2 | mcp-server integration tests no longer compile: they still set the removed `Profile.disable_response_storage` field | The PR removes `disable_response_storage` from `codex_protocol::mcp_protocol::Profile`, but the unchanged `codex-rs/mcp-server/tests/suite/config.rs` still builds a `Profile { ..., disable_response_storage: Some(false), ... }`. That file is compiled into the mcp-server `all` integration-test binary through `tests/all.r... | build | high | [
{
"path": "codex-rs/protocol/src/mcp_protocol.rs",
"from_line": 368,
"to_line": 378,
"role": "cause"
},
{
"path": "codex-rs/mcp-server/tests/suite/config.rs",
"from_line": 107,
"to_line": 116,
"role": "symptom"
}
] | Head mcp_protocol.rs:368 `pub struct Profile { pub model, pub model_provider, pub approval_policy, pub model_reasoning_effort, pub model_reasoning_summary, pub model_verbosity, pub chatgpt_base_url }`, from which the diff removes `- pub disable_response_storage: Option<bool>,`. At head, mcp-server/tests/suite/config... | To count, a finding must say that the mcp-server test `Profile` literal (tests/suite/config.rs) still sets `disable_response_storage`, which no longer exists on `Profile`, so that test target fails to compile. Near-misses that do not count: saying only that the TOML fixture line `disable_response_storage = false` (conf... | true | [] | [] | [
"R13",
"R14",
"R15",
"R16"
] | null | null | null | openai/codex | 8c952af20fe791cc0c5d0a59089d3329968b4373 |
openai__codex@8c952af | dev10 | 2 | D3 | Reasoning items are written to rollouts without their required `id`, so resumed sessions silently drop all reasoning | `ResponseItem::Reasoning { #[serde(skip_serializing)] id: String, ... }` has no `#[serde(default)]`, unlike the other skipped ids, which are `Option`s and default to None. The rollout writer serializes ResponseItems with the same serde impl, so every persisted Reasoning line lacks `id`. On resume, `serde_json::from_val... | data | medium | [
{
"path": "codex-rs/protocol/src/models.rs",
"from_line": 50,
"to_line": 57,
"role": "cause"
},
{
"path": "codex-rs/core/src/rollout/recorder.rs",
"from_line": 291,
"to_line": 334,
"role": "other-site"
},
{
"path": "codex-rs/core/src/rollout/recorder.rs",
"from_line":... | models.rs:50-56 at head: `Reasoning { #[serde(skip_serializing)] id: String, summary: Vec<...>, #[serde(default, ...)] content: ..., encrypted_content: Option<String> }`. Writer: recorder.rs:293-297 `RolloutCmd::AddItems(items) => for item in items { if is_persisted_response_item(&item) { writer.write_line(&item).await... | To count, a finding must connect the non-optional `Reasoning.id: String` + `skip_serializing` to local persistence: rollout items are written without `id` and cannot be deserialized back (missing required field), so resume drops reasoning items. Accepted fixes include `#[serde(default)]`/Option, or a separate wire repr... | true | [] | [
"5"
] | [
"R17",
"R18",
"R19",
"R20"
] | null | null | null | openai/codex | 8c952af20fe791cc0c5d0a59089d3329968b4373 |
openai__codex@8c952af | dev10 | 2 | D4 | Now-unused `request_body` bindings in two auth tests fail `cargo clippy --tests -D warnings` | In `prefers_chatgpt_token_when_config_prefers_chatgpt` and `prefers_apikey_when_config_prefers_apikey_even_with_chatgpt_tokens`, the PR replaces the only use of `request_body` (the store assertion) with a comment but keeps `let request_body = request.body_json::<serde_json::Value>().unwrap();`. That triggers rustc's `u... | build | high | [
{
"path": "codex-rs/core/tests/suite/client.rs",
"from_line": 578,
"to_line": 582,
"role": "cause"
},
{
"path": "codex-rs/core/tests/suite/client.rs",
"from_line": 659,
"to_line": 663,
"role": "cause"
}
] | Head core/tests/suite/client.rs:579-581 `let request = &server.received_requests().await.unwrap()[0]; let request_body = request.body_json::<serde_json::Value>().unwrap(); // store flag removed; responses are never stored.`, followed by `}`. Lines 660-662 are the same. The diff removes `assert!(!request_body["store"].a... | To count, a finding must say that `request_body` is left unused at client.rs ~580 and/or ~661, which produces an unused-variable warning or clippy -D warnings failure (a dead binding). Near-misses that do not count: comments that only ask to restore or assert `store == false` for coverage without noting that the bindin... | true | [] | [] | [
"R10",
"R11"
] | null | CI gate (lint): only fatal because CI runs `cargo clippy --tests -D warnings`; in production that job goes red and the review reads it. Kept and scored for the benchmark's integrity. | null | openai/codex | 8c952af20fe791cc0c5d0a59089d3329968b4373 |
mpv-player__mpv@0729afe | dev10 | 2 | D1 | The startup 'run all update handlers' pass is only queued, so it runs against live playback state after playback has started | mp_initialize's 'Run all update handlers' call now only queues {co=NULL, flags=UPDATE_OPTS_MASK}. The handlers therefore run at the first mp_idle, which for a CLI run with no clients (no user scripts, Linux) is inside play_current_file's open loop, after stop_play was reset to 0. There the UPDATE_DVB_PROG branch sets s... | bug | high | [
{
"path": "player/main.c",
"from_line": 383,
"to_line": 387,
"role": "cause"
},
{
"path": "player/command.c",
"from_line": 7705,
"to_line": 7725,
"role": "cause"
},
{
"path": "player/command.c",
"from_line": 7827,
"to_line": 7830,
"role": "symptom"
},
{
... | main.c:387 `mp_option_change_callback(mpctx, NULL, UPDATE_OPTS_MASK, false);` now just reaches `MP_TARRAY_APPEND(mpctx, mpctx->option_callbacks, ...)` (command.c:7723). UPDATE_OPTS_MASK includes UPDATE_DVB_PROG (m_option.h:440,452). mp_create sets `.stop_play = PT_NEXT_ENTRY` (main.c:280), so the old synchronous run wa... | To count, a comment must say that the startup UPDATE_OPTS_MASK 'run all update handlers' call (main.c:383-387) is now only queued and runs later, together with ANY one consequence: UPDATE_DVB_PROG aborting and reloading the first file, UPDATE_VO recreating the force-window VO, builtin scripts/hooks missing or not waite... | true | [] | [] | [
"R4",
"R5",
"R6",
"R8"
] | null | null | null | mpv-player/mpv | 0729afe8d7067f0e4d69000309ec6f3e24ac3926 |
mpv-player__mpv@0729afe | dev10 | 2 | D3 | Queued option changes never wake the core: synchronous property/option sets (mpv_set_property, IPC set_property, script mp.set_property) do not take effect while the player sleeps | mp_option_change_callback now only appends to the queue and never calls mp_wakeup_core or interrupts the dispatch queue. Synchronous client paths (mpv_set_property, mpv_set_option, and through them JSON IPC set_property and Lua/JS mp.set_property*) run under mp_dispatch_lock/unlock. Unlocking does not end the core's ti... | bug | medium | [
{
"path": "player/command.c",
"from_line": 7705,
"to_line": 7725,
"role": "cause"
},
{
"path": "player/client.c",
"from_line": 1059,
"to_line": 1064,
"role": "other-site"
},
{
"path": "player/client.c",
"from_line": 1309,
"to_line": 1335,
"role": "other-site"
... | The new callback body ends with `MP_TARRAY_APPEND(mpctx, mpctx->option_callbacks, mpctx->num_option_callbacks, callback);` and contains no wakeup. The path is mpv_set_property -> `run_locked(ctx, setproperty_fn, &req)` -> `mp_dispatch_lock(...); fn(fn_data); mp_dispatch_unlock(...)`. mp_dispatch_unlock only decrements ... | To count, a comment must say that queuing the option callback does not wake the core (no mp_wakeup_core or dispatch interrupt). As a result, a change made through a synchronous client/IPC/script set while the player is paused or idle isn't applied until an unrelated event wakes the loop, possibly never. Near-misses tha... | true | [] | [] | [] | null | null | null | mpv-player/mpv | 0729afe8d7067f0e4d69000309ec6f3e24ac3926 |
mpv-player__mpv@0729afe | dev10 | 2 | D4 | Setting aid/vid/sid no longer switches the track synchronously: a read right after the set, and the command's OSD message, report the previous track | For aid/vid/sid/secondary-sid, SET goes through mp_property_generic_option, which only writes the option and queues the callback; update_track_switch now runs later in mp_option_run_callbacks. GET and PRINT for these properties read mpctx->current_track while playback is initialized. So `set sid 2` / `cycle-values sid ... | contract | low | [
{
"path": "player/command.c",
"from_line": 7928,
"to_line": 7933,
"role": "cause"
},
{
"path": "player/command.c",
"from_line": 1948,
"to_line": 1971,
"role": "symptom"
},
{
"path": "player/command.c",
"from_line": 5215,
"to_line": 5219,
"role": "symptom"
},... | mp_property_switch_track: `struct track *track = mpctx->current_track[order][type]; ... case M_PROPERTY_GET: if (mpctx->playback_initialized) { *(int *)arg = track ? track->user_tid : -2; }`, and PRINT formats the same `track`. M_PROPERTY_SET falls through to `return mp_property_generic_option(...)`, which calls m_conf... | To count, a comment must say that because update_track_switch is deferred, setting aid/vid/sid (or another option whose effect is read back through derived state) followed immediately by a read or by the command's OSD message reports the old track, which breaks read-after-write for property sets. Near-misses that do NO... | true | [] | [] | [] | null | null | null | mpv-player/mpv | 0729afe8d7067f0e4d69000309ec6f3e24ac3926 |
mpv-player__mpv@0729afe | dev10 | 2 | D5 | Coalescing a round-trip hidpi-window-scale toggle leaves one unpaired relative resize, so the window is rescaled although the option ends where it started | update_hidpi_window_scale does not re-derive the window size from state. It applies a relative resize: it multiplies the current unfullscreened window size by the cached HiDPI scale when the option is true, and by the inverse when it is false. The new queue keeps only the newest entry for an option, and the handler rea... | bug | low | [
{
"path": "player/command.c",
"from_line": 7706,
"to_line": 7724,
"role": "cause"
},
{
"path": "player/command.c",
"from_line": 7898,
"to_line": 7899,
"role": "other-site"
},
{
"path": "player/command.c",
"from_line": 2781,
"to_line": 2799,
"role": "symptom"
... | The coalescing loop in mp_option_change_callback (command.c:7706-7724), `if (mpctx->option_callbacks[i].co && opt_ptr == mpctx->option_callbacks[i].co->data) drop = true;`, overwrites the older same-option entry and keeps a single one. The drain calls `if (opt_ptr == &opts->vo->hidpi_window_scale) update_hidpi_window_s... | To count, a comment must name hidpi-window-scale / update_hidpi_window_scale (or the relative window resize it performs) as a handler that coalescing breaks, because it applies a delta instead of re-deriving state from the final value. It must also name the consequence: changing the option and changing it back before t... | false | [] | [] | [
"judge candidate: N1"
] | null | null | null | mpv-player/mpv | 0729afe8d7067f0e4d69000309ec6f3e24ac3926 |
ppy__osu@748ab3d | dev10 | 2 | D1 | The smoothstep bell curve is not equivalent to sin²: the 'optimisation' changes the rhythm bonus and therefore SR/pp | Before the 0.5 clamp, SmoothstepBellCurve(f) equals 12f²-16f³ for f<0.25 (mirrored for f>0.75). The old term was sin²(πf). The two agree only at f=0 and wherever both are clamped to 0.5 (f in [0.25, 0.75]). For f in (0, 0.25) ∪ (0.75, 1) the new value is always higher, by up to +0.010, so currRatio = 1 + 12·term rises ... | bug | low | [
{
"path": "osu.Game.Rulesets.Osu/Difficulty/Evaluators/RhythmEvaluator.cs",
"from_line": 71,
"to_line": 76,
"role": "cause"
},
{
"path": "osu.Game/Rulesets/Difficulty/Utils/DifficultyCalculationUtils.cs",
"from_line": 76,
"to_line": 81,
"role": "other-site"
},
{
"path": "... | Merge base: `double currRatio = 1.0 + rhythm_ratio_multiplier * Math.Min(0.5, Math.Pow(Math.Sin(Math.PI / deltaDifferenceRatio), 2));`. Head, RhythmEvaluator.cs:71-76: `reverseDeltaDifferenceRatio = Math.Max(prevDelta, currDelta) / Math.Min(prevDelta, currDelta); deltaDifferenceFraction = reverseDeltaDifferenceRatio - ... | To count, a comment must say that the new smoothstep bell curve is not a drop-in equivalent of the old sin²(π·ratio) term. It must say that it gives different (higher) values for some fractional delta ratios, so rhythm/speed difficulty, star rating or pp change, meaning the PR is not a pure performance change. Naming a... | true | [] | [] | [
"R2"
] | null | null | null | ppy/osu | 748ab3d4863f6ea4bf6a3fffccb50e5cd616937b |
symfony__symfony@2b9721f | dev10 | 2 | D1 | vibrate documented as int[] only; the Notification API VibratePattern also accepts a single integer | The new shape declares `vibrate?: int[]` in both the constructor @param and the getContent() @return. The Web Notification `vibrate` option is a VibratePattern (`unsigned long or sequence<unsigned long>`), so a valid value such as `['vibrate' => 200]` gets flagged by PHPStan/Psalm even though MercureTransport forwards ... | contract | low | [
{
"path": "src/Symfony/Component/Notifier/Bridge/Mercure/MercureOptions.php",
"from_line": 39,
"to_line": 39,
"role": "cause"
},
{
"path": "src/Symfony/Component/Notifier/Bridge/Mercure/MercureOptions.php",
"from_line": 96,
"to_line": 96,
"role": "other-site"
}
] | MercureOptions.php:39 ` * vibrate?: int[],` (in `@param array{...}|null $content`) and :96 ` * vibrate?: int[],` (in `@return array{...}|null` of getContent()). MercureTransport.php doSend: `json_encode([... 'content' => $options->getContent(), ])` passes the array through with no normalisation, so the docbloc... | To count, a comment must say that `vibrate` is typed as an int array only, while the Notification/Vibration API (VibratePattern) also accepts a single integer, so the type should be `int|int[]` / `int|list<int>`. Flagging either line 39 or line 96 counts. Near-misses that do NOT count: saying `int[]` should be `list<in... | true | [] | [
"1"
] | [] | null | null | null | symfony/symfony | 2b9721f04744fb213808f754ac98cd563384b3e5 |
End of preview. Expand in Data Studio
Real defects in code review
The data behind every benchmark number QuickReview publishes, and every other benchmark run its authors hold except those listed under "What is not here, and why". It comes in three parts, all in this repository:
- The tables (the configs below,
data/*.parquet). A verified defect key for 185 pull requests of AACR-Bench v1.0. Every review from every shipped AACR-Bench run, and every judge verdict that scored it. AACR-Bench's own metric. The reviews and verdicts on Martian's offline code-review benchmark and on Qodo's PR-Review-Bench. The pairwise Elo judgments and ratings. The answer key's build data (which reviewer raised each candidate, each PR's size). A run index. - The run archive (
archive/). The run directories as the runs wrote them, without their raw model session logs: reviews and findings, judge and scorer outputs with their reasoning, run configurations, per-PR summaries, driver logs, and a numbers-only extract of every JSONL session log (timestamps, token usage, cost, lens outcomes); the other session logs, the process logs and the prompt files are left out with no extract. There is one.tar.zstvolume per run family, 928 run index units (run and scoring directories, driver files, support trees) in 50 families. Its README lists every volume, every rewrite and every kind of file left out. - The slim re-score bundle (
slim/). The part of the archive that the scoring and recompute scripts read, with no transcript text.
Beside them, second_judges/deepseek-v4.1-flash-1017/ holds a second judge's verdicts on the
reviewers of the repository README's full tables, on all three benchmarks,
next to the published judge's ("Second judge").
DOI: 10.57967/hf/11598. The code, the benchmark definition and both papers are in the
QuickReview repository, https://github.com/quickchatai/quickreview, under
bench/aacr/.
The two papers:
- Measuring whether a code review finds the real defects: a verified defect
benchmark on AACR-Bench, applied to eight reviewers at their best-shot
configurations (the benchmark paper, arXiv
<arXiv pending>). - QuickReview: a multi-pass automated code reviewer, its design, its cost,
and how it performs on a verified defect benchmark (the technical report,
arXiv
<arXiv pending>).
bench/aacr/claims.json in the repository lists every benchmark number of
both papers, the README, bench/aacr/docs/ and this card, with the data and
the script behind it, and one command recomputes them all from this data with
no model call ("How to verify"). Every one of them passes: a number the release could not check was cut from the pages or reworded to the number the data gives. Published numbers that
differed from the data were corrected to it before this release ("Versions").
Third-party figures the pages quote (AACR-Bench's own metadata, Qodo's and
Martian's published scores and leaderboards) are attributed to their source
and listed in the manifest as quotes, not as claims.
Read this first
- No human has checked the key or the judge yet. LLM agents verified the 520 defects and the 3,093 exclusions against the code, and an LLM judge graded every review. A human audit of a sample is planned and will be published with its agreement numbers. Until then, treat the key as an LLM-built answer key, with the error rate that implies.
- The key was partly built from QuickReview's own earlier output. The
candidates checked for each PR included the comments that reviewers posted
in our runs, QuickReview's among them. They were anonymised, and every
candidate was checked against the code. A reviewer whose earlier comments
seeded the key can still be favoured by it.
defects.provenance_*records where each defect came from: 369 of the 520 were raised by the blind hunt or by an AACR reference, and the papers report recall on those separately. - The authors build one of the reviewers (QuickReview), and it was tuned on
these benchmarks. The last tuning round was mined from misses on part of
each benchmark, including 145 of the 185 AACR-Bench PRs. The tuning sets are
named per PR (
prs.set,prs.in_s40,martian_goldens.audit_split), and the papers and the README report the held-out results beside the tuned ones. - Most results rest on one judge from one model family. In the tables, a
second judge from another family scored only the earlier gpt-6-sol study.
The final comparison, the frozen build, the open-model runs, Martian and
Qodo each have one GPT judge there.
second_judges/adds a DeepSeek judge on the frozen build, the final comparison's third-party reviewers, Martian and Qodo; it is unaudited, and on AACR-Bench it could not read the code. See "Judges" and "Second judge" below. - AACR-Bench's reference labels are agent-produced, not human labels
(
aacr_references.is_ai_comment,source_model). At least 80 of the references marked human are rewritten review-bot comments (aacr_references.bot_derived). - Each reviewer ran once and each review was judged once, except where a
config says otherwise: the stability runs, the earlier study's second judge
(
claude-opus-5-5), the judge re-run.second_judges/, which no config reads, judges the reviews of the repository README's full tables once more. - Builds and judges differ between rows. The open-model runs used a
mid-October build and the
gpt-6.1-sol-highjudge; the frozen build and the papers' comparisons usedgpt-6-sol-high. Compare rows within one judge.
What is in here, and which claim each part backs
The claim ids in the second column are claim groups of the repository's
bench/aacr/claims.json, which describes each one (groups). The runs
config lists a group on each run directory its claims' recompute keys read
(runs.claims): a key that names the run, its arm or its scoring directory,
and for a script that reads the run archive the keys its reads ties to each
unit (the keys whose numbers moved when that unit's files were changed and the
script rerun). A group none of whose keys reads a run is on no run:
do:sets, second-judge.
Volume names drop the quickreview-runs- prefix and the .tar.zst suffix.
The scripts are in the QuickReview repository under bench/aacr/.
| Published claim | Claim id | Tables | Archive volumes | Recomputed by |
|---|---|---|---|---|
README "AACR-Bench, verified defects" chart and full table; benchmark paper tab:final, tab:grades, tab:lang, tab:costfinal, tab:selection; QuickReview paper tab:final |
readme:aacr, bench:final-tables, qr:tab:final |
prs, defects, arms, reviews, defect_verdicts, comment_verdicts |
aacr-frozen-build, aacr-final-comparison-g61, aacr-claude-code-opus55, answer-key |
final/score.sh on the slim bundle; validation.json |
The frozen build on 185 PRs (benchmark paper tab:finalbuild; QuickReview paper tab:frozen) |
bench:tab:finalbuild, qr:tab:frozen |
as above, arm quickreview-frozen |
aacr-frozen-build |
validation.json; recompute/aacr_bootstrap.py |
Bootstrap intervals and paired differences (README "+8.3 points [+1.9, +14.6]"; tab:ci in both papers) |
readme:aacr-ci |
prs, defects, defect_verdicts, reviews |
as above, analysis |
recompute/aacr_bootstrap.py |
| README "Same model, two reviewers" chart | readme:same-model |
arms, defect_verdicts |
aacr-opus55-basemodel, aacr-gpt6sol-study, aacr-claude-code-opus55, aacr-frozen-build, aacr-final-comparison-g61 |
validation.json; final/score.sh |
| README "Other models" table | readme:other-models |
arms (quickreview-glm-53-flash, quickreview-deepseek-v41-flash, quickreview-luna, quickreview-opus55), reviews, defect_verdicts |
openmodels-glm53-full, openmodels-dsv41f-full, openmodels-dsv41f-low, openmodels-dsv41f-mix, openmodels-driver, aacr-luna-basemodel, aacr-opus55-basemodel |
recompute/open_models.py; validation.json |
| README "Tuning and held-out PRs" | readme:heldout-aacr, readme:heldout-martian, readme:heldout-qodo |
prs, defect_verdicts, martian_goldens, martian_verdicts, qodo_issues, qodo_reviews, qodo_verdicts; the repository's bench/aacr/sets/ (Qodo's held-out PRs: sets/qodo-heldout31.json) |
aacr-frozen-build, martian-final, qodo-final, analysis |
recompute/heldout.py |
The earlier gpt-6-sol study, two judges (benchmark paper tab:sol-summary, tab:second-sol and the rest of that section) |
bench:sol-study |
arms (*-sol, claude-code-opus55), defect_verdicts, comment_verdicts |
aacr-gpt6sol-study, aacr-claude-code-opus55, aacr-v2-default-rerun |
validation.json; final/score.sh |
Base models (QuickReview paper tab:bm185, tab:bm40, tab:bmsplit) |
qr:basemodel |
arms, defect_verdicts |
aacr-final-comparison-g61, aacr-gpt6sol-study, aacr-v2-default-rerun, aacr-opus55-basemodel, aacr-luna-basemodel |
validation.json |
Stability (tab:stab-sol, tab:stab) |
bench:tab:stab-sol, qr:tab:stab |
arms (quickreview-stab-r1 to -r4) |
aacr-stability-repeats, aacr-v2-default-rerun |
validation.json |
Best-shot configuration search (tab:bestshot, tab:transfer) |
bench:bestshot |
arms (*-s40) |
aacr-bestshot-search, aacr-final-comparison-g61 |
final/score.sh |
AACR-Bench's own metric and the reference-label study (tab:aacr, tab:labels-sol) |
bench:final-tables, bench:sol-study |
aacr_matches, aacr_references, aacr_reference_labels |
aacr-final-comparison-g61, answer-key |
validation.json |
Harvested Copilot 2025 reviews (tab:copilot2025) |
bench:tab:copilot2025 |
arms (copilot-2025-upstream, links only, and the arms below the table), reviews, defect_verdicts, comment_verdicts, aacr_matches |
aacr-upstream-copilot, aacr-gpt6sol-study, aacr-v2-default-rerun, aacr-final-comparison-g61, aacr-claude-code-opus55 |
final/score.sh (its tab:copilot2025 call) on the slim bundle; verify_claims.py from the tables |
README Martian chart and table; benchmark paper tab:martianfinal and the golden-comment audit |
readme:martian, bench:tab:martianfinal |
martian_prs, martian_goldens, martian_reviews, martian_verdicts, martian_scores |
martian-final, analysis |
Martian's scorer on the slim bundle (verify_slim.py); validation.json |
Martian development rounds, posting policies, judge noise (tab:martianraw, fig:martianpolicy) |
bench:tab:martianraw, bench:tab:martianfinal (the judge-noise numbers) |
martian_scores, martian_verdicts |
martian-tuning-rounds, martian-posting-variants, martian-judge-calibration |
Martian's scorer on the slim bundle |
README Qodo chart and table; benchmark paper tab:qodo |
readme:qodo, bench:tab:qodo |
qodo_prs, qodo_issues, qodo_reviews, qodo_verdicts (QuickReview's two builds: work_dir and tool final1-qodo-pr100 and quickreview-frozen, ext-r6-qodo-pr100 and quickreview-2026-10-02) |
qodo-final, qodo-r6-and-vendors |
our scorer, ext/qodo_judge.py, on the slim bundle (verify_slim.py --qodo-gt; Qodo publishes none); validation.json |
| README "Judged side by side" (pairwise Elo) and its chart | readme:elo |
elo_judgments, elo_ratings |
elo-pairwise; the reviews: ccvq-hard20-claude-max-sonnet, ccvq-claude-code-high, ccvq-quickreview-on-claude, martian-final, openmodels-glm53-full, openmodels-dsv41f-mix |
recompute/elo_ratings.py --check-ratings |
docs/DIGITALOCEAN.md: speed and cost table, the smokes, the review time limit |
do:speed, do:glm-smoke, do:gemma-smoke, do:timeouts |
none (archive only) | openmodels-smokes, openmodels-gpt61-baselines, openmodels-glm53-full, openmodels-driver, aacr-frozen-build |
recompute/do_speed.py |
| QuickReview paper: cost mechanism (pass counts, lens shares, validator dispositions) | qr:mechanism |
reviews |
aacr-gpt6sol-study (the transcripts' numbers-only extracts) |
recompute/passes.py, recompute/lens_credit.py, recompute/tokens.py |
Benchmark paper tab:martiantiers (QuickReview on Martian by its own severity) |
bench:tab:martiantiers, bench:tab:martianfinal |
martian_goldens |
analysis (analysis/relabel/martian_tiers.labels.jsonl, its prompt, LABELS.json; the blind tune20 classification), martian-posting-variants, martian-tuning-rounds |
relabel/martian_tiers.py --compute |
| QuickReview paper: the validator's kills and the lost-credit attribution | qr:mechanism, qr:tab:final |
defects, reviews, defect_verdicts |
analysis (analysis/relabel/lost_credit.labels.jsonl, its prompt, LABELS.json), aacr-gpt6sol-study |
relabel/lost_credit.py --compute |
| QuickReview paper: tuning appendix (prompt iterations, levers, lenses) | qr:app:tuning |
arms, defect_verdicts |
aacr-dev-iterations, aacr-dev-levers-presets |
verify_claims.py (its table checks, recompute/claim_metrics.py); recompute/missed_defects.py |
| QuickReview paper: the max preset example | qr:basemodel |
arms (quickreview-max-g61) |
aacr-g61-smokes |
compare.py |
The second judge beside the published judges (this card's "Second judge"; benchmark paper tab:dsjudge and its judge-reliability section; QuickReview paper's limits) |
second-judge |
none: the folder second_judges/deepseek-v4.1-flash-1017/, beside the published judge's rows in defect_verdicts, comment_verdicts, martian_verdicts, martian_reviews, qodo_verdicts |
none | recompute/score_verdicts.py, recompute/martian_scores.py, recompute/qodo_scores.py (the folder README's commands); recompute/second_judges.py; validation.json |
tab:copilot2025's rows are these arms, on the 19 PRs that have a
copilot-2025-upstream review (final/copilot19.txt):
| Row | Arm |
|---|---|
| QuickReview default | quickreview-default-sol (the tab:final row of that name is quickreview-default-g61) |
| QuickReview default, current build | quickreview-default-sol-rerun |
| QuickReview default, gpt-6.1-sol | quickreview-default-g61 |
| QuickReview low | quickreview-low-sol |
| Claude Code (Opus 5.5) | claude-code-opus55 |
| open-code-review | ocr-medium-sol |
| pi-review | pireview-medium-sol |
| OpenCode | opencode-bridge-sol |
| Codex CLI | codex-medium-sol |
| Copilot 2025 (upstream) | copilot-2025-upstream |
The claim ids come from bench/aacr/claims.json alone: a run or an arm
carries exactly the claim groups whose recompute keys read it (runs.claims,
arms.paper_tables, arms.in_paper), and validation.json lists the run
families and archive volumes of each group whose keys name a run (a group
whose keys read only the key, a set or a published number lists none). A
claim's q key names its exact rows.
The other run families back no published number and ship anyway: the fast
option of the frozen build, the v1.0.7 re-baseline, the review-efficiency
arms, the refactor smoke, configuration ablations, development and plumbing
runs, freeze validations, a duplicate-merge check, a partial base-model probe,
the Opus 5.5 Martian smoke, one-PR probes, and the SWE-PRBench runs (archive
only). Their AACR reviews and verdicts are in the tables under arms of their
own (arms.in_paper is false); runs.purpose says what each one was for.
Configs
One split, test, in every config. Which PR set a row belongs to is the set
column (AACR: smoke10, dev10, holdout10, fresh10, rest145); the
papers' 40-PR selection set S40 is prs.in_s40. Every review, verdict and
match row carries run_family, the family of the run directory it was read
from (the archive volume of the same name).
from datasets import load_dataset
defects = load_dataset("quickchatai/quickreview-bench", "defects", split="test")
verdicts = load_dataset("quickchatai/quickreview-bench", "defect_verdicts", split="test")
runs = load_dataset("quickchatai/quickreview-bench", "runs", split="test")
Run it with uv run --no-project --with datasets python when the datasets
package is not installed.
| Config | Rows |
|---|---|
defects |
520 |
exclusions |
3,093 |
prs |
185 |
aacr_references |
1,418 |
aacr_reference_labels |
253 |
arms |
187 |
reviews |
5,593 |
defect_verdicts |
18,928 |
comment_verdicts |
25,810 |
aacr_matches |
28,894 |
martian_prs |
50 |
martian_goldens |
173 |
martian_reviews |
4,010 |
martian_verdicts |
13,093 |
martian_scores |
529 |
qodo_prs |
100 |
qodo_issues |
580 |
qodo_reviews |
2,215 |
qodo_verdicts |
12,888 |
elo_judgments |
654 |
elo_ratings |
140 |
runs |
928 |
aacr_candidates |
3,748 |
pr_diffstats |
185 |
The run index
runs— one row per shipped run index unit, 928 rows. Most are run directories and their scoring directories; the open-model driver's files (scripts, logs, gate reports) and a few top-level trees (the answer key, the analysis, the re-run logs, the Elo runs, the what-if study) are units of their own, andpathshows which. Each row gives its path in the archive and its volume (archive_volumes), its family and the family's title, the benchmark and PR set, the number of cases, start and end times, the reviewer configuration (arm_config), the model it billed (model_billed), its judges, what it was for (purpose), the claim groups whose recompute keys read it (claims, the ids above) and the tables that hold its rows (hf_tables).qcr_commitis the QuickReview commit the run's image was built from (for a third-party tool's run, the harness around that tool). The repository's history was rewritten before publication, so the ids the runs recorded are not in the public history:qcr_commitnames the public main commit the recorded one became, withqcr_commit_on_public_maintrue andqcr_commit_matchsaying how it was matched:same, the rewrite's commit maps carry the recorded commit to it;ancestor, the rewrite dropped the recorded commit because it changed only files that were never published, so it maps to the nearest ancestor the rewrite kept, whose published files are the same. A build with no public equivalent (from a side branch that was never merged, or from another clone whose commits no rewrite map holds) keeps its recorded id, withqcr_commit_on_public_mainfalse andqcr_commit_matchnull. The dirty flag is separate: a dirty build (built from a checkout with uncommitted or untracked files; the recorded id ends in-dirty) maps to the commit it was built on, keeps the-dirtysuffix and is flaggedqcr_commit_dirtytrue (false for every other build): it is that commit plus whatever was not committed, which was not saved, so not necessarily the commit itself. A row that recorded no build hasqcr_commit,qcr_commit_on_public_main,qcr_commit_matchandqcr_commit_dirtyall null. The archive'srun.jsonandcase-summary.jsonfiles carry the same remap, and itsRUNS-INDEX.jsonholds the same rows, plus the families that were withheld and a description of each claim group the rows carry.
The answer key (AACR-Bench, 185 PRs)
prs— the 185 runnable PRs: repo, PR URL, title, base and head commit, merge base, language, set, S40 membership, counts, and the repo's licence (SPDX).merge_baseis the merge base of the PR's base and head on GitHub. The reviewers and the judge saw the diffmerge_base..head_commit, notbase_commit..head_commit: on most PRs the two differ.defects— 520 defects (90 high, 145 medium, 285 low): title, mechanism and consequence (statement), sites, kind, severity, evidence (file:line and short code quotes), what a finding must say to count (counts_if), provenance, and two never-scored annotations (benchmark_limitation,bot_derived_reference). The provenance is the key files': on the T40 sets it records the older runs those keys were built from, and the answer-key volume'sorig40-code/<id>/truth-ext.jsonadds the extension reviewers' matches; each paper number says which reading it uses.key_versionis the key version the judge used (fresh10 is on version 1, the other sets on version 2).exclusions— 3,093 candidates checked and rejected, each with its reason.aacr_references— AACR-Bench's own 1,418 reference comments on these PRs (Apache-2.0), with their category and context labels,ref_key(the keyaacr_matchesuses), and whether a review bot wrote the original comment.aacr_reference_labels— our check of the 253 references on 30 PRs: real defect, real maintainability point, trivial, dubious or duplicate.aacr_candidates— how the key was built: one row per candidate the agents checked, with the earlier run or the reviewer that raised it (raised_by), its file, lines and text.roundiscandidatesfor the first key of the rest145 PRs (each candidate mapped to the run that raised it) andnew-candidatesfor the extended key (each mapped to the comparison reviewer). These back the papers' provenance table, the defects only QuickReview's earlier runs raised, and the leave-one-reviewer-out key.pr_diffstats— each PR's size as the papers bucket it: files changed, insertions, deletions and their sum (changed_lines), from the PR'sgit diff --stat. It backs recall by PR size. AACR-Bench's ownprs.change_line_countcounts differently.
The reviews and verdicts (AACR)
arms— one row per reviewer configuration: display name, the run directories it is scored from, the study and the paper tables it feeds, its run family, the model it billed, its primary judge (judge) and every judge with verdicts for it (judges).in_papermarks the configurations a published number reads (a claim's key inclaims.json), andpaper_tablesnames the claim groups that read them. The final comparison is eight configurations:quickreview-default-g61,quickreview-low-g61,pireview-panel-high-g61,ocr-high-g61,codex-high-g61,opencode-native-pr-g61,pragent-g61andclaude-code-opus55(the other*-g61arms,quickreview-max-g61,best-pragent-g61andpi099-smoke-g61, are not in it). The frozen build isquickreview-frozen; the earlier gpt-6-sol study isquickreview-default-sol,quickreview-low-sol,ocr-medium-sol,codex-medium-sol,opencode-bridge-sol,pireview-medium-solandclaude-code-opus55(pi099-smoke-solis a smoke, not in it);*-s40arms are the best-shot selection runs on S40; the open-model rows arequickreview-glm-53-flashandquickreview-deepseek-v41-flash. Every other shipped AACR run is an arm named after its run.reviews— one row per arm and PR: the review body, its comments exactly as the judge numbered them (C1,C2, …), list-price cost and wall-clock minutes, and the number of comments AACR-Bench's harness scored. A comment'stextis what the judge read: QuickReview's tag ([blocker],[question],[nit]) moved totag, and its footer (priority, confidence and lens, in italics) removed when it starts with the priority (_P1 · …_). A footer of another form (_nit · confidence … · lens …_, a bare_confidence …_, or one not in italics) was not removed: the judge read it, and it stays intext.raw_textis the comment as the run's result file holds it, after the release's name rewrite ("Names rewritten"): QuickReview's comment as it posted it, a third-party harness's finding text, and Claude Code's summary and failure scenario, which it never posted.priority(P0toP4, ornit) andconfidenceare read from QuickReview's footer in every form; both are null for a comment with no footer, as for every other reviewer. A link-only arm'sraw_textis null. The archive has the review as it was written:results/all/<id>.json(content),cases/<id>/comments.json,review.mdandfindings-record.json.defect_verdicts— per arm, PR, defect and judge: found, partial or missed (credit1, 0.5, 0), the comments that found it, and the judge's reason.comment_verdicts— per arm, PR, comment and judge: correct-valuable, correct-minor, partly-correct or incorrect, the defect it maps to, and whether the judge flagged a possible new defect.aacr_matches— AACR-Bench's own metric: per arm, PR and reference, the share of the harness's three judge rounds that matched it. Only runs the AACR harness scored have rows.
A defect with no row for an arm and PR counts as missed. Every arm of the final comparison has a judged review for all 185 PRs. Some reviews list a comment twice: once on the deleted line it was written on, and once re-anchored to the new side with the same text. The judge read and graded both copies, and both count: 12, 15 and 24 comments in the frozen build and the two open-model arms (51 in all), and a few in arms no paper reports.
Martian offline benchmark (50 PRs)
martian_prs,martian_goldens— the 173 golden comments (MIT), with our adversarial audit's bucket for each (86 real, 77 not real, 10 disputed) and its evidence.martian_reviews,martian_verdicts,martian_scores— the Martian scorings of our runs (the frozen build, the three third-party harnesses we ran, the development rounds, the posting-policy variants, the open-model and Claude runs), and the published vendor reviews re-judged by our judge (rejudged-*), each golden matched or not.work_dirnames the scoring directory.martian_scoresalso carries Martian's own published tool-level judgements (judge=published: …).- A QuickReview row's
commentsare what Martian's step 1 reads from GitHub: the first 30 inline comments (one API page) and the review body.n_posted_inlinecounts every inline comment posted, andcomments_beyond_first_pageholds the ones past the page, which Martian's judge never saw. The text is as posted, after the name rewrite (which also touches a few sandbox paths the comments quote). fast1-martian50scored the fast option beside the re-judged vendors, and its QuickReview reviews were ingested under the tool nameclaude-code. They are relabelledquickreviewhere. That dir'sclaude-coderows inmartian_scoresare Martian's published scores for Claude Code.- The posting-policy runs (
mv-*,mt-test30-frozen,qr-all50-*) are filters of stored reviews. 5 of the 56 also shorten each comment, to its title or its title and first line. Each run'sVARIANT.jsonin the archive states its policy.
Qodo PR-Review-Bench (100 PRs)
qodo_prs,qodo_issues— the 580 injected issues (MIT; 309 functional, 271 rule violations).qodo_prs.base_commitandhead_commitare the real GitHub commits: the PR's merge-base and its head, whose trees are the trees our runs reviewed. The runs reviewed them as replay commits, the same trees under a synthetic author, date and message so the reviewer could not read the PR's history; those ids never existed on GitHub. The run files carry them, and the case directories are named by the short replay head, soreplay_base_commitandreplay_head_commitkeep them for the join.repo_licenseis per PR: the redis fork's licence changed between its PRs (read fromLICENSE.txtat each PR's head).qodo_reviews,qodo_verdicts— the Qodo scorings of our runs (the frozen build, an earlier build of 2026-10-02, the same three harnesses, the fast option, the open-model runs) and Qodo's nine published vendor runs, each issue matched or not, with the location rule applied at ±5 lines and at "same file". Every verdict row has its review. QuickReview's two published builds are the rows withwork_dirandtoolfinal1-qodo-pr100andquickreview-frozen(the frozen build), andext-r6-qodo-pr100andquickreview-2026-10-02, 100 reviews each.qodo_reviews.commentsholds what our judge was sent:ask()inbench/aacr/ext/qodo_judge.pybuilds the message from it and cuts a long comment's text and a long issue's code (its limits are there). The two builds were sent their deleted-line comments differently: the frozen build's 12 were re-anchored to the new side and judged, and the 2026-10-02 build's 18 were left out (2 of them on held-out PRs). So the rise in rule violations found, 14.8% to 37.6%, compares builds under slightly different inputs. The held-out PRs aresets/qodo-heldout31.jsonin the repository.
Pairwise Elo (Martian, 20 PRs)
elo_judgments— one row per pairwise judgment: the plan, the PR, the two configurations (a,b) as the judge saw them, the judge and its effort, the winner overall and in four categories (recall, precision, finding quality, writing) with the judge's reasoning, and the sha256 of the prompt.used_in_ratingsmarks, per plan, the judgments that plan's ratings use: the 596 of the README's plan (martian-hard20-all-gpt) and the smoke plan's own. The rest are attempts at a superseded prompt, live Opus probes and an exploratory DeepSeek judge, kept for completeness.elo_ratings— per plan, judge table (gpt,opus,combined) and category: each configuration's Elo, its 95% bootstrap interval, its wins, ties and losses, and its games. The pairing plans themselves are in theelo-pairwisevolume.
Every comment, by reviewer
What you can read before you download: the comment rows of reviews
(AACR-Bench), martian_reviews (the comments past Martian's first page
included) and qodo_reviews, counted per reviewer from the tables when this
card was built. Every shipped run counts, so a PR that several runs reviewed
counts once per run. A count is of comments whose text the table holds; a
"location or link only" comment carries its file and line, or its link, and
its text is third-party ("What is not here"). A Martian vendor row carries no
comment rows: Martian's own file holds them. An AACR arm counts under the tool
its id names (claude-code-opus55, best-pragent-g61, ccvq-cc303-* for
Claude Code); the development, ablation, smoke and open-model arms are
QuickReview's.
| Reviewer | AACR-Bench | Martian | Qodo |
|---|---|---|---|
| QuickReview | 16,687 | 9,934 | 4,511 |
| open-code-review | 1,852 | 347 | 346 |
| OpenCode | 1,292 | 229 | 368 |
| Claude Code | 474 | 1,261 | – |
| pi-review | 1,439 | – | – |
| PR-Agent | 616 | 211 | 579 |
| Codex CLI | 963 | – | – |
| qodo-exhaustive (Qodo's published run) | – | – | 508 location or link only |
| copilot (Qodo's published run) | – | – | 467 location or link only |
| qodo-precise (Qodo's published run) | – | – | 340 location or link only |
| augment (Qodo's published run) | – | – | 283 location or link only |
| greptile (Qodo's published run) | – | – | 268 location or link only |
| cursor (Qodo's published run) | – | – | 212 location or link only |
| coderabbit (Qodo's published run) | – | – | 209 location or link only |
| codex (Qodo's published run) | – | – | 181 location or link only |
| sentry (Qodo's published run) | – | – | 107 location or link only |
| Copilot 2025 (upstream) | 52 location or link only | – | – |
| Claude Code (Martian's published review) | – | no comment rows (305 reviews) | – |
| cubic-v2 (Martian's published review) | – | no comment rows (1,405 reviews) | – |
| qodo-extended-v2 (Martian's published review) | – | no comment rows (305 reviews) | – |
| All | 23,323 + 52 location or link only | 11,982 | 5,804 + 2,575 location or link only |
The verdicts that read them: comment_verdicts (AACR-Bench, a grade per
comment) and defect_verdicts (the comments that found each defect);
martian_verdicts (per golden, matched or not) with
martian_reviews.unmatched_candidates, the candidates extracted from a review
that matched no golden; qodo_verdicts (per issue, the comment that matched
it, comment_index).
Other files. figures/ holds the papers' charts and their data, and
keys/ the answer key rendered as Markdown. validation.json holds the
checks below (see "How to verify" for how they relate to the claims
verifier's release checks). LICENSES/ holds the MIT licence texts of Martian's and
Qodo's data (see "Licence"). The working report the papers grew from is not
included: the papers supersede it.
Second judge. second_judges/ holds another judge's verdicts on reviews
these tables already hold, one folder per judge, each with its README, the
scorers' output (scores/) and the scripts that made it (code/). No config
reads it, so load_dataset returns the same rows with or without it. Its
verdict tables:
| File | Rows |
|---|---|
second_judges/deepseek-v4.1-flash-1017/aacr/defect_verdicts.parquet |
3,640 |
second_judges/deepseek-v4.1-flash-1017/aacr/comment_verdicts.parquet |
3,941 |
second_judges/deepseek-v4.1-flash-1017/aacr/judgements.parquet |
1,168 |
second_judges/deepseek-v4.1-flash-1017/martian/martian_verdicts.parquet |
1,211 |
second_judges/deepseek-v4.1-flash-1017/martian/martian_reviews.parquet |
350 |
second_judges/deepseek-v4.1-flash-1017/qodo/qodo_verdicts.parquet |
4,640 |
validation.json checks the folder file by file against the repository's
bench/aacr/hf/second-judges.sha256, checks that its rows are exactly the
published judge's rows of the reviewers it names, and re-runs the repository's
scorers on it (see "How to verify").
Judges
Which judge scored what, exactly. The judge column of each table uses the
labels in the first column.
| Judge | Model | What it scored |
|---|---|---|
gpt-6-sol-high |
GPT-6 Sol, high effort | the primary judge of the papers: the frozen build, the final comparison, the earlier study, and every other AACR arm except the open-model ones below (base models, stability, selection, development runs, the re-baseline, the refactor smoke). |
claude-opus-5-5 |
Claude Opus 5.5 | the earlier study's second judge. Only the earlier gpt-6-sol study: its seven arms (listed under arms above), 3,640 reviewer-defect pairs. Claude Code's run is shared by both comparisons, so the final comparison's Claude Code row carries both judges; no other final-comparison arm does. |
gpt-6-sol-high-repeat |
GPT-6 Sol, high effort | the primary judge run again, on the same seven arms and 30 PRs, for self-consistency. |
gpt-6.1-sol-high |
GPT-6.1 Sol, high effort | the open-model runs (GLM-5.3 Flash, DeepSeek V4.1 Flash, the DigitalOcean smokes) and the GPT-6.1 Sol baselines copied beside them. |
| AACR-Bench's harness | GPT-6 Sol, low effort, three rounds | aacr_matches for every arm the papers report. Early development arms (prompt iterations, smokes, end-to-end checks; none in the papers) ran it with GPT-5.6 Terra at low effort, or with a single round. arms.aacr_judge names the model and rounds per arm. |
gpt-6.1-sol |
GPT-6.1 Sol, low effort, medium verbosity (ext_judge.py's default) |
every row of martian_verdicts and qodo_verdicts. One judge, served by the ChatGPT Codex backend through bench/aacr/ext/ext_judge.py. The backend rejects a sampling temperature, so the bridge drops the harnesses' temperature 0 (and, on one Martian smoke, some requests' max_completion_tokens). On Martian the same model also extracts and merges the candidates (steps 2 and 2.5). Each Martian scoring dir's runs.judges row states the model, effort, backend and dropped request keys from its request log (one dir has none, and says so), and the verbosity as ext_judge.py's default, which no log records. |
Elo: gpt-6.1-sol and claude-opus-5-5 |
GPT-6.1 Sol and Claude Opus 5.5, both at xhigh | the pairwise judgments, 298 each in the ratings, pooled into 596. The two pick the same overall winner in 56% of the comparisons both judged. |
deepseek-v4.1-flash-1017 (in second_judges/, not in the tables) |
DeepSeek V4.1 Flash; on AACR-Bench high effort, no tools and no repository; on Martian and Qodo low effort, as gpt-6.1-sol |
the reviewers of the repository README's full tables, again: the frozen build and the final comparison's six third-party reviewers on AACR-Bench, and the README's Martian and Qodo tools. See "Second judge". |
The primary AACR judge's prompt is judgePrompt() in src/bench/defects.ts.
The repository's public main builds exactly the prompts behind the published
verdicts from commit 34164fe0e to 839e26454; later commits render a few
reviews differently. bench/aacr/recompute/judge_prompt.py prints that prompt
for any arm and PR from these tables. review_hash in the archive's judge
files is not a public check: it was taken before the name rewrite. The Qodo
judge was asked twice on the 10 PRs of ext-r6-qodo-smoke10, every tool's
review: 222 of the 223 matches of the first pass recur in the second (the
run's scores.md). In the tables, the final comparison, the frozen build, the open-model rows, Martian and
Qodo each rest on one judge from the GPT family; second_judges/ re-judges
the frozen build, the final comparison's third-party reviewers and the
README's Martian and Qodo tools with a judge from a third family. The two
judges of the tables agree closely
on defect verdicts where both ran (3,593 of 3,640 pairs) and much less on
pairwise preference. validation.json checks this coverage against the
tables: the second judge and the re-run hold rows for the seven arms only,
arms.judges names exactly the judges with rows for each arm, and the
Martian, Qodo and Elo judges are the ones named here.
Second judge
second_judges/deepseek-v4.1-flash-1017/ holds a second judge's verdicts on the reviews of every
reviewer in the repository README's full tables: on AACR-Bench the frozen
build and the final comparison's six third-party reviewers
(1,168 reviews, every review of those arms the published judge was asked
about), on Martian the README's seven tools (350 reviews) and on Qodo
its eight (800 reviews). It is a judge of the final reviews; the
earlier study's second judge, claude-opus-5-5, is in the tables ("Judges").
The benchmark paper calls this one the third judge, after claude-opus-5-5.
The judge. DeepSeek V4.1 Flash (deepseek-v4.1-flash-1017), from a third
model family. On AACR-Bench it was given the published judge's prompt
(judge_prompt.py rebuilds it), but no tools and no repository: the
published gpt-6-sol-high read the code at the PR head while it judged; this
judge read only the message (the defects with their quoted evidence, the
rejected claims and the review). It ran at high effort, and at low effort when
an answer at high was cut off (1,056 and 112 of the accepted
answers). On Martian and Qodo it ran at low effort through the same
ext_judge.py bridge as the published gpt-6.1-sol. It is an unaudited
LLM judge: no person has checked its
verdicts. Its numbers show how much the published results depend on the
judge; they do not correct them. The folder's README gives the run, its cost
and every difference in full.
AACR-Bench. Each cell: this judge / the published gpt-6-sol-high, on the
same reviews. Agreement: the share of reviewer-defect pairs both judges gave
the same verdict (found, partial or missed), with Cohen's kappa.
| Reviewer | Found | Recall | High-severity recall | Comments correct | Agreement (kappa) |
|---|---|---|---|---|---|
| QuickReview (frozen) | 461.5 / 461 | 88.8% / 88.7% | 98.9% / 97.2% | 85.4% / 92.4% | 98.5% (0.928) |
| pi-review panel (high) | 348 / 347.5 | 66.9% / 66.8% | 88.9% / 88.9% | 81.4% / 89.4% | 99.0% (0.979) |
| Claude Code (Opus 5.5) | 324 / 324.5 | 62.3% / 62.4% | 81.1% / 80.6% | 80.8% / 80.6% | 99.0% (0.980) |
| open-code-review (high) | 296.5 / 298.5 | 57.0% / 57.4% | 70.6% / 70.6% | 76.8% / 86.1% | 99.2% (0.985) |
| Codex CLI (high) | 287.5 / 287 | 55.3% / 55.2% | 90.6% / 90.0% | 87.5% / 92.7% | 99.2% (0.985) |
| OpenCode (+ PR text) | 263.5 / 263.5 | 50.7% / 50.7% | 84.4% / 84.4% | 88.1% / 92.7% | 99.6% (0.992) |
| PR-Agent | 192.5 / 194 | 37.0% / 37.3% | 61.1% / 61.7% | 74.5% / 75.7% | 97.9% (0.956) |
Recall moves by at most 0.4 points: six of the seven reviewers' recall moves, none by more than that against its published recall, and the order of the seven is the same on recall, high-severity recall and weighted recall. Comments correct is not: it is lower under this judge for six of the seven reviewers, by 1.2 to 9.3 points, and level for Claude Code (+0.2). Most of the disagreements grade as incorrect a comment that the published judge, which could read the code, graded correct or partly correct. Comment grades agree less than defect verdicts: the two judges give the same grade to 84.0% to 92.1% of each reviewer's comments (kappa 0.59 to 0.79). Two pairs swap on comments correct: OpenCode passes Codex CLI, and Claude Code passes open-code-review.
Martian (Core; this judge / the published gpt-6.1-sol)
| Tool | Precision | Recall | F1 |
|---|---|---|---|
| QuickReview (frozen) | 16.9 / 16.2 | 66.5 / 65.8 | 27.0 / 26.1 |
| qodo-extended-v2 | 63.3 / 56.2 | 60.1 / 60.1 | 61.7 / 58.1 |
| cubic-v2 | 57.6 / 54.3 | 62.0 / 60.1 | 59.8 / 57.1 |
| open-code-review | 28.8 / 26.7 | 62.0 / 58.9 | 39.4 / 36.8 |
| PR-Agent | 48.0 / 41.9 | 38.0 / 36.1 | 42.4 / 38.8 |
| claude-code | 42.0 / 36.5 | 43.0 / 41.1 | 42.5 / 38.7 |
| OpenCode | 36.1 / 32.7 | 35.4 / 33.5 | 35.8 / 33.1 |
Qodo PR-Review-Bench (this judge / the published gpt-6.1-sol; the same
location rule)
| Tool | Precision | Recall | F1 |
|---|---|---|---|
| qodo-exhaustive | 58.4 / 58.0 | 46.7 / 46.4 | 51.9 / 51.5 |
| qodo-precise | 68.6 / 68.0 | 36.2 / 35.9 | 47.4 / 47.0 |
| QuickReview (frozen) | 38.5 / 37.3 | 46.9 / 45.3 | 42.3 / 40.9 |
| augment | 67.5 / 67.5 | 29.7 / 29.7 | 41.2 / 41.2 |
| OpenCode | 65.6 / 64.9 | 29.3 / 29.0 | 40.5 / 40.0 |
| copilot | 48.2 / 46.8 | 34.7 / 33.6 | 40.3 / 39.1 |
| open-code-review | 48.0 / 47.1 | 28.6 / 28.1 | 35.9 / 35.2 |
| PR-Agent | 48.1 / 47.2 | 27.8 / 27.2 | 35.2 / 34.5 |
On Martian and Qodo this judge is more lenient. Core precision is 0.7 to 7.1 points higher for every Martian tool, and Core F1 is 0.9 to 3.8 points higher; the smallest gains are QuickReview's (+0.7 precision, +0.9 F1), so every other tool's Martian precision and F1 lead over it widens. Qodo F1 is 0 to 1.4 points higher (augment's precision, recall and F1 are the same under both judges). Every change of order on the scorers' headline metrics:
- Martian. Precision keeps its order, and QuickReview keeps the highest
recall and the lowest precision. Recall: open-code-review moves from fourth
to a tie for second with cubic-v2, and cubic-v2 leaves its published tie
with qodo-extended-v2 to lead it. F1: claude-code passes PR-Agent. F2
(
scores/martian.txt): cubic-v2 passes qodo-extended-v2, and claude-code passes QuickReview. - Qodo. Functional and rule recall keep their order. Recall: QuickReview passes qodo-exhaustive, by 0.2 points, a tie. F1: QuickReview passes augment. Precision: copilot passes PR-Agent and open-code-review.
- AACR-Bench. Only the two comments-correct swaps above.
The scorers make no model call. From a clone of the repository with this
dataset downloaded as quickreview-bench (see "How to verify"), each command
prints this judge's numbers beside the published judge's on the same reviews
and writes, under work/ (gitignored), the scores/ file of the same name:
D=../quickreview-bench
V=$D/second_judges/deepseek-v4.1-flash-1017
L=deepseek-v4.1-flash-1017
run() { uv run --no-project --with pyarrow python "$@"; }
mkdir -p work
run bench/aacr/recompute/score_verdicts.py --data $D --verdicts $V/aacr --judge $L \
--arms quickreview-frozen,pireview-panel-high-g61,claude-code-opus55,ocr-high-g61,codex-high-g61,opencode-native-pr-g61,pragent-g61 \
--json work/aacr.json > work/aacr.txt
run bench/aacr/recompute/martian_scores.py --data $D --verdicts $V/martian --judge $L --json work/martian.json > work/martian.txt
run bench/aacr/recompute/qodo_scores.py --data $D --verdicts $V/qodo --judge $L --json work/qodo.json > work/qodo.txt
validation.json runs the same three and compares their output with
scores/ (see "How to verify"), and the claims verifier recomputes every
number of this section from the folder (second_judge in
bench/aacr/claims.json).
What is not here, and why
- QuickReview reviewing the authors' own repositories. A self-review run and the follow-up-round experiments on the authors' private product repository are withheld whole. They are not benchmark runs, and they quote private code.
- A post-cleanup check of the default build (
pc2-*, public AACR-Bench and SWE-PRBench PRs) is withheld by the authors. It is a regression check of a development build and backs no published number. - Raw model session logs of every kind are not published: the reviewers'
pi event streams and session files, lens, Codex, Claude Code and OpenCode
transcripts, the judges' event streams, the review and judge processes' own
output logs, the sandbox command logs, and the prompt files each session was
given. They hold the prompt text QuickReview's sessions were built from during
development, not benchmark data. Each JSONL session log ships instead as a
numbers-only extract (
*.pruned.jsonl: timestamps, token usage, cost, stop reasons and lens outcomes, no prompt or model text), so the speed, cost, token and lens claims stay verifiable:compare.pyprices a case from its extract exactly as from the full stream. Session text inside the files that are kept is rewritten field by field: a harness result'sstdout(the Codex event stream) andstderrbecome numbers-only extracts, a prompt in a driver's recorded argv or a tool's prompt template becomes a hash, and Claude Code's denied commands become a count. The set is defined by file kind and fails closed, and a kept data file that still holds session text is left out whole; the archive's README lists every kind and its count. - Aborted and interrupted attempts (
aborted-*, the open-model driver'sinterrupted*) are left out. Most were re-run under their normal name, and the re-run ships. The attempts at a full Martian run of QuickReview on Opus 5.5 (aborted-c4-opus55-martian50,aborted-ratelimit-opus55-martian50) were stopped and never completed: only that configuration's smoke,opus55-martian-smoke10, ships. The partial base-model probe (aacr-astra-probe) ships as the partial run it is;runs.purposesays where it stopped. - The Elo
packs/are left out. They are whole files of the reviewed repositories, cut to each PR, and some are under AGPL or source-available licences.bench/aacr/elo/elo.py packrebuilds them from the public repositories at the commits of the plan's set (expect_merge_baseandhead_commit), read from bare clones the archive does not ship (bench/aacr/elo/README.mdsays how to make them), andpacks.jsonllists the files each pack held.recompute/elo_ratings.pyrefits the ratings from the judgments alone, with no pack. - Third-party text is linked, not copied: the vendor review comments
Martian publishes (and its
published-benchmark_data.json), the reviews of Qodo's nine vendors, the upstream Copilot reviews, SWE-PRBench's human review comments, and PR authors' descriptions wherever a file copied one (the answer key'smeta.json, PR-Agent's describe output, the open-model case sets). A Qodo or Copilot row carries the source URL and the comment locations; a Martian vendor row carries none (Martian's own file lists them). An archived file carries[third-party text omitted, sha256:<prefix>; see <link>], so equal texts stay equal and every score recomputes unchanged. To restore the text, fetch it and match the prefix (the sha256 of the comment's text): Martian's vendor comments are in Martian's repository ate616e84(offline/results/benchmark_data.json); Qodo's are on the vendors' public fork PRs (qodo_reviews.source_url), and the archive's<run>.qodo/judgements.jsonlholds each comment's prefix; Copilot's are on the upstream PRs: fetchpulls/{n}/reviewsandpulls/{n}/commentsfrom GitHub's API into<id>.reviews.jsonand<id>.comments.json, andbench/aacr/harvest_upstream.py --api DIR --out RUNS --prefix upstream-copilot2025 --fetchrebuilds the run dirs. Third-party text that does ship, under its own licence: Martian's golden comments and published tool-level judgements and Qodo's issues (MIT, see "Licence"), AACR-Bench's reference comments (Apache-2.0), and code quoted by reviewers and judges (each project's licence). Two copies of vendor text also ship verbatim: the Martian judge's extractions of the re-judged vendors' comments (the vendor candidates in the archived and slimcandidates.jsonandevaluations.json) and the vendor candidates of the blind annotation's candidate list (the archive'sanalysis/…/fair/secret.json). That text belongs to its authors, the vendors' review bots that posted it on public pull requests, and is not covered by this dataset's licence. The tables keep it out:martian_reviewsrows of a vendor carry no comment or candidate text. - Encrypted reasoning blobs are blanked: the value of every OpenAI
encrypted_contentand every AnthropicthinkingSignatureandsignatureis the empty string, and the record around it is kept. They are ciphertext under the providers' keys. The usage records still carry the reasoning-token counts. - Data anyone can regenerate: the reviewed repositories' code (
repo.git,base/,head/,pr.diff; refetch from the repo and commits in each case'smeta.jsonfor AACR-Bench, themartian_prstable for Martian; for Qodo and SWE-PRBench the run files carry the harness's replay commit ids, made so the reviewer could not read the PR's history and never on GitHub: the real commits are in each case set'sleakproofblock,source_headandsource_merge_baseinbench/aacr/sets/<set>.jsonof the repository, and inqodo_prs, and the replayed trees equal those commits' trees), the external benchmarks' cached data and harnesses (pinned: withmartian/code-review-benchmark@e616e84, Qodo/PR-Review-Bench@a73957c, foundry-ai/swe-prbench@b87f579), scratch directories, scheduler state and release build outputs. - The authors' private working notes.
- SWE-PRBench runs are in the archive only, with no table. They back no
published number. The eval100 run's judged comments
(
quickreview-answers.json) differ from the posted ones on 8 of its 100 PRs: 5 add a finding the review posted in its body, and 3 leave out a deleted-line comment. Both sets ship.
Names rewritten
Before export, every string in every table passed through one rewrite, the
same one the run archive applies. Apart from it, the only changes to the
text are the third-party omissions above and the qcr_commit remap (see
runs).
| What | Replaced with | Occurrences in the tables |
|---|---|---|
| credential value | <redacted-secret> |
0 |
| old product name | quickreview |
0 |
| old product name (other case) | quickreview |
0 |
| review bot | <review-bot> |
3,874 |
| review bot (bare) | <review-bot> |
0 |
| owner | <owner> |
2 |
| owner (other case) | <owner> |
0 |
| private repo | <private-repo> |
0 |
| host path: runs dir | <runs> |
0 |
| host path: repo checkout | <repo> |
0 |
| host path: scratchpad | <scratchpad> |
0 |
| host path: Claude Code project | .claude/projects/<project> |
0 |
| host path: home | <home> |
0 |
| host path: home (dashed) | -<home> |
0 |
| host path: home (JSON-escaped) | <home> |
0 |
| host volume: pi login | <pi-login-volume> |
0 |
| private token | <private> |
0 |
| private host | <private-host> |
0 |
| study rename | ccvq |
2,318 |
| tracker path | <ticket> |
0 |
| ticket id | <ticket> |
40 |
| public-repository stand-ins (content rules) | the public repository's stand-in |
3,952 |
<owner>is the operator the reviewer's production prompt addresses.<review-bot>is the reviewer's GitHub App persona.quickreviewis the public name of the product. Its earlier internal name appears in prompts and footers, in any case.<private-repo>is the authors' private product repository. A few model texts name it.<redacted-secret>replaces every value of the bench's credential files, wherever one occurs. A provider API key that a reviewer could read from its container environment appeared in a few open-model transcripts; those transcripts are not published.<ticket>replaces the ids of the authors' tracker, which is not public, in any spelling (with or without the hyphen, in any case, as in an image tagqcr-bench:<ticket>), and any path into it. No ticket number is kept: the runs of the Claude Code vs QuickReview study, whose directory names began with one, are namedccvq-*throughout (with the Elo planmartian-hard20-ccvq), as the public repository's Elo plans name them.<private>and<private-host>replace a personal handle, a private commit id, an account id and private host names.<runs>,<repo>,<home>and<pi-login-volume>replace host paths and a docker volume name; they occur almost only in the archive's logs (the table above counts the few in the tables).- The public repository's stand-ins: where QuickReview's prompts, as built at run time, carried examples from private work, the same stand-ins the public repository's history carries replace them, by the same rules.
The archive's README gives the count of every rewrite over the whole archive.
How to verify
Needs: git, Python 3.11 or newer, uv, zstd (for
tar --zstd); the dataset is public, so no Hugging Face login. Run hf as
uvx hf when Hugging Face's CLI is not installed, and load_dataset (see
"Configs") with uv run --no-project --with datasets python. Nothing in this
section needs docker, a key or a model call. Disk: about 3 GB covers the
download, the clone, the unpacked slim bundle, the archive volumes the two
relabel scripts unpack and uv's caches.
Download the dataset and the QuickReview repository at the ref this release was verified against, then recompute:
hf download quickchatai/quickreview-bench --repo-type dataset --local-dir quickreview-bench
git clone --branch bench-data-v1 https://github.com/quickchatai/quickreview
cd quickreview
D=../quickreview-bench
# the tables: bootstrap intervals, paired differences, held-out results
uv run --no-project --with pyarrow python bench/aacr/recompute/aacr_bootstrap.py --data $D
uv run --no-project --with pyarrow python bench/aacr/recompute/heldout.py --data $D
# the slim bundle: every final table and chart, the open-model rows,
# the DigitalOcean table, the Elo refit
mkdir slim && tar --zstd -xf $D/slim/quickreview-bench-slim.tar.zst -C slim
(cd slim && awk '{print $1" "$3}' FILES.sha256 | sha256sum -c --quiet) # every file's sha256 (hash, size, path)
QCR_BENCH_RUNS_DIR=$PWD/slim/runs QCR_CHARTS_OUT=$PWD/charts-out bench/aacr/final/score.sh
python3 bench/aacr/recompute/open_models.py --runs-dir slim/open-models/runs
python3 bench/aacr/recompute/do_speed.py --runs-dir slim/open-models/runs --gpt-runs-dir slim/runs
python3 bench/aacr/recompute/elo_ratings.py slim/elo/martian-hard20-all-gpt --check-ratings \
--runs-dir slim/runs --runs-dir slim/open-models/runs
bench-data-v1 is the release's tag. It is made on the repository's main
branch when the release's code is merged, and its bench/aacr/ is the code
this release was built and verified with, file for file.
Every command from here on runs in the clone and keeps its scratch there:
slim/, work/, charts-out/, martian/ and qodo/. The repository's
.gitignore ignores all five, so git status in the clone stays clean.
Two tables in the papers rest on a labelling pass as well as the runs: the
benchmark paper's tab:martiantiers and the QuickReview paper's lost-credit
figures. The label files ship in the archive's analysis volume and the slim
bundle under analysis/relabel/, each row with the model that wrote it
(gpt-6.1-sol) and the sha256 of the prompt shipped beside it; LABELS.json
names the script, model, effort, prompt hash and row count of each pass. The
two scripts turn the labels and the archive into the published numbers with no
model call (each unpacks the archive volumes it reads into work/), and refuse
a row whose prompt hash is not the shipped prompt's:
uv run --no-project --with pyarrow python3 bench/aacr/relabel/martian_tiers.py \
--archive $D/archive --data $D --work work --compute
uv run --no-project --with pyarrow python3 bench/aacr/relabel/lost_credit.py \
--archive $D/archive --data $D --work work --compute
QCR_CHARTS_OUT keeps the re-drawn charts out of the clone (a re-rendered
JPEG differs byte for byte from the committed one).
bench/aacr/recompute/README.md maps each published number to its script.
The claims verifier checks every published number listed in
bench/aacr/claims.json against this data in one command. Two of its
inputs are third-party and never in this dataset: a checkout of Martian's
benchmark at the pinned commit (its scorer and published judgements) and
Qodo's ground-truth file (the dataset's
git_code_review_bench_100_w_open_prs.jsonl):
git clone https://github.com/withmartian/code-review-benchmark martian
git -C martian checkout e616e849755441da38f18bf3adba2c9583b03803
hf download Qodo/PR-Review-Bench --repo-type dataset --revision a73957c450a70693a743260e5637fffc44625f16 --local-dir qodo
uv run --no-project --with pyarrow python bench/aacr/verify_claims.py --data $D \
--martian martian --qodo-gt qodo/git_code_review_bench_100_w_open_prs.jsonl
It fails on a mismatch, an error, a missing source, a claim with no
recompute script and a claim whose data the release does not hold; left
without an input above, it fails on the claims that need it.
On this release: all 6,337 claims pass, with 0 mismatch, 0 error, 0 missing source and no claim left unchecked. The release's own checks pass 285 of 285: 280 from validation.json, re-run on the dataset (271 on the tables and 9 on its second_judges/ folder), and 5 the slim bundle's. --partial without --martian and --qodo-gt re-runs 283 of them (the slim bundle's Martian and Qodo re-scores need those inputs), and with the tables alone the 271 from validation.json.
Its release checks are the checks of validation.json (below) that need only
the tables and the second_judges/ folder, re-run on your download, and the
slim bundle's own (every file's hash, no transcript text, every Martian and
Qodo scoring dir re-scored, the Elo refit). validation.json also holds the checks the export ran against its
own inputs (the source run directories, the commit remap, the archive's index,
the figures): that every run family has rows in the tables it should, the
runs table's build ids, each Elo plan's refit against its run's
ratings.json. A download cannot re-run those, so its count is the larger.
--partial checks what you have. verify_claims.py --data $D --partial
skips, and does not fail, a claim whose source is not under --data or whose
only missing input is Martian's checkout or Qodo's file (with the slim
bundle's Martian and Qodo re-scores), and exits 0 when every other claim
passes. Its result line counts the skipped claims and names the flag that
would include them (--martian, --qodo-gt). With the tables alone (no slim
bundle, no archive) it recomputes every claim they hold. Without --partial
a missing input fails.
validation.json recomputes, from these tables alone, the row counts (520
defects, 3,093 exclusions, 185 PRs per full arm, 1,418 references, 173
Martian goldens, 580 Qodo issues), the papers' headline numbers (every arm's
defects found by severity, comments correct, cost and time; the second judge;
judge agreement; the AACR metric; Martian and Qodo scores), the open-model
rows, the Elo ratings (refit from elo_judgments and compared with the
run's own ratings.json), the judge coverage above, and that every shipped
run family has rows in the tables it should. This release passes 504 of 504
checks.
Re-judge with your own judge
Every QuickReview review behind a published number is in the tables, so you
can judge it with your own judge and score that judge's answers the way we
scored ours, with no model call of ours. Write your answers as tables shaped
like ours (.parquet or .jsonl, in a directory DIR, no judge column),
pick a label no published row carries (LABEL), and run the scorer. It
prints your judge's numbers beside our judge's on the same reviews: when your
judge judged only some of a run's PRs, both columns are counted over those
PRs, and the header says how many (--whole-run counts every PR instead, a PR
your judge did not judge counting as missed, so the two columns are then not
comparable). Each
script's docstring gives the columns and what it refuses. D is this
dataset, and the scripts are in the repository's bench/aacr/recompute/;
run them as uv run --no-project --with pyarrow python <script>.
AACR-Bench. Input: the reviews rows of an arm (arms.arm_id, or the arm
a claim's q key names): the review's verdict, body and comments (C1,
C2, …) as the judge read them. Answer key: the PR's defects, and its exclusions (the rejected
claims our judge was also shown). The code: prs.repo, diff
prs.merge_base..prs.head_commit, and the PR's title. Our judge's prompt
is judgePrompt() in src/bench/defects.ts; judge_prompt.py --data $D --arm ARM --pr ID prints the one our judge received, and --all --out-dir OUT every one of an arm that our judge received. Output: defect_verdicts (arm_id, instance_id,
defect_id, verdict: found, partial or missed) and, optionally,
comment_verdicts (arm_id, instance_id, comment_id, grade, defect_id,
new_defect). Score:
score_verdicts.py --data $D --verdicts DIR --judge LABEL [--arms a,b] # recall, comments correct, agreement
aacr_bootstrap.py --data $D --verdicts DIR --judge LABEL --arms a,b # the intervals
heldout.py --data $D --fixed-set --verdicts DIR --judge LABEL # the held-out PRs
If you re-ran our own judge instead (bench.sh judge over the archive's run
directories, as "Reproduce" in bench/aacr/README.md says), score its files
directly: score_verdicts.py --data $D --judge LABEL --from-judgements ARM --runs-dir <archive>/runs --judge-dir NAME reads each run's
judge/v<key version>-NAME/ in the arm's runs (NAME is the label the
judge wrote under, such as openai-gpt-6-sol-high on your own key), and
--save DIR writes the two tables for aacr_bootstrap.py and heldout.py.
Run on our judge's own files in the archive (--judge-dir gpt-6-sol-high),
it prints the published numbers.
Martian. Input: the martian_reviews rows of QuickReview, by work_dir
and tool as a claim's key names them (the frozen build: final1-martian50,
quickreview-frozen), or the archive's runs/<work_dir>.martian/results/benchmark_data.json,
which Martian's pipeline reads. Answer key: martian_goldens. Our judge is
Martian's own steps 2, 2.5 and 3 at e616e84, on GPT-6.1 Sol; their prompts
are in Martian's repository, and bench/aacr/ext/ext_judge.py's docstring
gives the commands to run them with no docker (through its bridge for a
Responses API model, or with MARTIAN_BASE_URL set straight to a Chat
Completions endpoint). Output: martian_verdicts (work_dir, tool, pr_url,
golden_index, matched) and martian_reviews (work_dir, tool, pr_url, fp: the
PR's unmatched candidates). Score:
martian_scores.py --data $D --verdicts DIR --judge LABEL [--runs WORK|TOOL] [--half martian-test30]
martian_scores.py --data $D --judge LABEL --work final1-martian50 \
--from-evaluations <dir>/results/<model>/evaluations.json --tool 'final1-martian50|quickreview=quickreview-frozen'
The second form reads Martian's pipeline output (results/<model>/evaluations.json)
as the export does: a golden is matched when a true positive's golden_comment
is its comment, fp is the PR's, and --tool gives the tool the table's
name. Run on our judge's own file in the archive, it prints the published
scores.
Without --verdicts it recomputes our judge's stored scores and fails on
any difference. The stored column holds whole-run scores only: on a sample of
a run's PRs, or with --half, it reads unknown, and the column to compare
yours with is our judge's recomputed one. To judge one PR, cut the archive's
benchmark_data.json to that PR and to QuickReview's review
(ext_judge.py's docstring has the command; ext/martian_ingest.py --only
writes such a file from a run directory): step 3 judges every tool the file
holds.
Qodo. Input: qodo_reviews with work_dir and tool
final1-qodo-pr100 and quickreview-frozen (the frozen build), or
ext-r6-qodo-pr100 and quickreview-2026-10-02; comments holds what our
judge was sent. Answer key: qodo_issues. Our judge's prompt is SYSTEM in
bench/aacr/ext/qodo_judge.py, and ask() there builds the user message
(it cuts long comment text and issue code); the code, not the judge, applies
the location rule (located()). Output: qodo_verdicts (work_dir, tool,
pr_url, issue_index, matched, comment_index). Score:
qodo_scores.py --data $D --verdicts DIR --judge LABEL [--runs WORK|TOOL,...]
qodo_scores.py --data $D --judge LABEL --work final1-qodo-pr100 \
--from-judgements <dir>/judgements.jsonl --tool 'final1-qodo-pr100|quickreview=quickreview-frozen'
The second form reads qodo_judge.py's judgements.jsonl as the export
does (per PR and tool the first-pass row its own scoring reads, one verdict per
issue) and refuses a judgement of comments other than the published review's;
give --from-judgements and --work once per work dir to score both builds.
Run on our judge's own files in the archive, it prints the published scores.
To run our judge's prompt on your model, with no docker: start
ext_judge.py on your provider (its docstring), then judge the published rows
(--only ID[,ID] judges those cases of the set only, as an instance id,
Ghost#6 or a PR URL; leave it out for all of them):
uv run --no-project --with pyarrow python bench/aacr/ext/qodo_judge.py \
--set bench/aacr/sets/qodo-pr100.json --gt qodo/git_code_review_bench_100_w_open_prs.jsonl \
--reviews $D/data/qodo_reviews.parquet --review-key 'final1-qodo-pr100|quickreview-frozen' \
--endpoint http://127.0.0.1:8765/v1/chat/completions --model MODEL --judge-label LABEL \
--out qodo-judge --only Ghost#6
and score qodo-judge/judgements.jsonl with the second form above.
It also scores the held-out PRs (sets/qodo-heldout31.json) and the
frozen build's held-out gain over the 2026-10-02 build.
Copilot 2025 (tab:copilot2025). Input: the arms in the table under
"What is in here", on the PRs of copilot-2025-upstream. Judge and output
as for AACR-Bench. Score:
score_verdicts.py --data $D --verdicts DIR --judge LABEL --select arm=copilot-2025-upstream \
--arms quickreview-default-sol,quickreview-default-sol-rerun,quickreview-default-g61,quickreview-low-sol
Copilot's own text is link-only, and judge_prompt.py refuses a link-only
review: re-judging Copilot's own reviews needs them restored ("What is not
here") and our pipeline run on the restored run directories (Tier 2 in
bench/aacr/README.md).
Elo. Input: the plan's run directory in the archive (elo/martian-hard20-all-gpt/:
plan.json, units.jsonl, judgments/) and the reviews it pairs, which are
also in martian_reviews. Our two judges' prompt is built by
bench/aacr/elo/elo.py pack from bare clones of the reviewed repositories,
which the archive does not ship (bench/aacr/elo/README.md says how to make
them), and each elo_judgments row carries its prompt_sha256. Give your
judge the same prompts: its prompt_sha256 must equal the row's, or the
latest judgment sets the unit's prompt and the others' judgments of it go
stale (elo_ratings.py warns). Point the plan's run_dirs
at the unpacked archive and add your judge under the plan's judges
(bench/aacr/elo/README.md says how). Output: the judgments elo.py run
writes. Score:
python3 bench/aacr/elo/elo.py pack D # D: the plan's run directory, then: run D --judge NAME, report D
elo_ratings.py D --runs-dir <archive>/runs --runs-dir <archive>/open-models/runs
Both rate every judge the plan names, and give each pair's agreement.
--runs-dir prices a configuration only when its run directory is whole:
run bench/aacr/ext/restore_copies.py over the unpacked volumes first,
the open-model volumes included.
To re-run our own judge instead, see bench/aacr/README.md ("Reproduce",
Tier 2) and bench/aacr/ext/ext.sh; set EXT_JUDGE_MODEL to judge Martian
and Qodo with another model. Those runs build and start docker containers:
they need a host where docker info succeeds, with Docker 25 or newer.
Which provider each judge can be pointed at, and how, is in one place:
bench/aacr/README.md, "Judges on any provider".
Versions
This is one release. The tables of the papers' first data build (2026-10-05,
never published) were rebuilt inside it from the same run directories, and
every run held since was added. Building the claims manifest against this
data found published numbers that differ from it; each was corrected to the
data in the papers (both rebuilt), the README and bench/aacr/docs/ before
this release. The pages as previously published on the repository's main
branch printed the old values:
- Paper tables. The stability run
stab-r4: comments correct 77% → 73% (tab:stab-sol,tab:stab). The Martian development table (tab:martianraw) and the posting-policy table: the default'stest30F2 38.8 → 38.7, round 4's issues per PR 10.6 → 10.5, and issues per PR as posted 9.0 → 8.9 and under the frozen posting policy 2.0 → 1.9 (179 and 39 comments on 20 PRs, printed the way every other table prints a half-way value). - Paper prose. The share of Martian's real goldens the three third-party harnesses find, 54--78% → 53--78%; the frozen build's cost against the third-party tools, 4 to 29 → 4 to 31 times; the pi-review panel's cost, 2.7 → 2.8 times one high reviewer; Codex at high effort, $0.05 → $0.06 more per PR; the frozen build against the default, 20% → 21% less; the default's lead on T40, 18.2 → 18.3 points; PR-Agent's comments mapping to a defect, 77% → 76%; AACR-Bench's "human" references, 362 → 361; their AI-written share, 74% → 75%; the key's reviewer comments, 2,216 → 2,221, and its excluded candidates, 691 → 692; the Opus-shifted Martian projection, 43.5 → 43.6; the seven smaller sub-lenses' cost share, 2.8% → 2.7%; defects raised by six or more lenses, 127 → 126, and by five, 53 → 54. Ratios and differences of costs use the unrounded costs, so a few read one unit off a ratio of the rounded figures printed beside them.
- README. QuickReview on Opus 5.5 in the Elo pool costs $4.61 per PR on average (was "about $4"); the open-model rows' lost lens and validator passes are given per kind (was "missing a lens pass on 7").
docs/DIGITALOCEAN.md. The GPT-6.1 Sol row of the speed table, 16 min, 37k output tokens, 78 replies, $2.27 (was 8, 21k, 71, $2.30); the DeepSeek row, 42 min, 230 replies, 13 s, $0.69 (was 41, 223, 14 s, about $0.57); 2.5–3× the steps and 7–9× the output (was about 3× and 12–15×); about 16 minutes a PR for GPT-6.1 Sol (was about 8); GLM's slowest AACR reviews 55–72 minutes (was 54–66); the GLM smoke's median 26.5 minutes (was "about 28", the mean) and its total $3.32 (was $3.31, a sum of per-set cents); the Qwen3.8-Max projection, about $6.29 a PR and ~$2,100 a full run at GLM's full-run token profile (was $3.09 and ~$1,030, a smoke's profile); the count of timed-out PRs in the first full attempt is dropped (no run file records it).docs/THREATS-TO-VALIDITY.mdandMETRICS.md. 361 "human" references (was 362).- Checked against the data in the final pass. The halves bootstrap
interval, −6.3 to +8.9 → −6.3 to +9.1 (benchmark paper, THREATS); the Opus
judge's API calls per judgement, about 1.7 → 2.0 (benchmark paper, THREATS,
METHODOLOGY); the code-quality lenses in the technical report's tuning log,
about 3 more defects found → 2, AACR maintainability matches 1 → 4 of 30 →
1 → 2 of 30, +29% → +28% cost; the re-judge's distance below Martian's
published scores, 1.5–6.3 → 0.9–6.3 F1 points; the Martian posting-policy
sweep, "F2 between 42 and 44 at best" → "42.0 as posted to 46.2 at best";
smoke10's backup PRs in
bench/aacr/README.md, 17 → 16. Loose ranges ("about 1–2 defects per 10 PRs", "about 10–15 lens passes", "about 10× the tokens", "$2–3 of review") now print the measured figure. - Removed, because no published file holds them. The stopped blind
hunt's 32 drafts; the line-gate sweep's run-time judge-call counts (3,227
and 10,983), its 2 failed calls and its ~$18 cost; the GPT judge's 2
retried error files; the critics' correction counts (6, 4 and 12; 16
added, 1 removed: the critics fixed the key in place); T40's v2 check of
the judge's new-defect candidates (44 / 12 / 1 / 2 / 41); the hand checks
of 20 parsed Claude Code reviews (45 / 44) and of 40 unmatched Qodo
comments (12 / 19 / 9); the lens instructions' 32% shortening; 45 of 219
flagged comments; Copilot's 6.9% and the 0.4-point precision line; the 46
bot-reviewed PRs; PR-Agent pruning 4 of 40 diffs; and the timings and
smoke-test figures in
bench/aacr/README.md's development notes. - Recomputed. Run-to-run agreement on fresh10, 78–89% → 81–93%, now
over two named pairs of published runs (
fresh10-onwithlean,new-r1withnew-r2). - Relabelled. The Martian severity table and the technical report's
lost-credit figures are recomputed from two labelling passes that ship in
analysis/relabel/(gpt-6.1-sol; see "How to verify"). The table's candidate-to-comment mapping is rebuilt (word overlap, a label where overlap cannot settle it), which moves a few candidates between tiers and changes the per-subset Martian scores. The lost-credit attribution now counts a nearby comment or dropped finding only when the label says it is about the defect: lens generation's share rises, weaker comments' falls, and fewer of the default's misses had a dropped finding stating them. Every changed cell and sentence prints what the labels give. The labels are one model's (gpt-6.1-sol, effort xhigh), not checked by hand; the lost-credit labeller was not told which findings were posted.
Datasheet (summary)
- Motivation. AACR-Bench scores a review by how closely it reproduces the original PR's comments; on 145 PRs checked in full, only 12.8% of those comments belong to a real defect. This key asks instead whether a review found the real defects the PR introduces.
- Composition. 185 PRs from 50 open-source repositories in ten languages (every AACR-Bench v1.0 PR whose head GitHub still serves), 520 defects, 3,093 rejected candidates, the reviews and verdicts above, and the run directories behind them.
- Collection. Per PR: a blind expert hunt from the diff and code; then every candidate (the hunt, AACR's references, every comment any reviewer posted in our runs) checked against the code by agents that could not see who raised it, merged where it is one defect, graded high / medium / low, and every rejection recorded with its reason; then an adversarial critic per set. Reviews were produced in a locked-down local sandbox (no GitHub writes); the judge reads the whole review and grades every defect.
- Severity. high: would block a merge; medium: should be fixed before merge; low: real, fine as a follow-up. Weighted recall: 3 / 2 / 1.
- Uses. Scoring code reviewers on recall of real defects, with comment accuracy, cost and time beside it; studying judge agreement; studying reference-comment benchmarks. Not for training a reviewer you then evaluate on these PRs.
- Known limits. No human check yet (above); one judge family for the
final comparison, the frozen build, Martian and Qodo; one run per reviewer;
90 high-severity defects limit how small a difference can be resolved;
reviewed-repo code is not included (
repoand commits identify it). - Maintenance. Changes to the key bump
key_version; verdicts are stored per key version, so old and new truths never mix.
Licence
Apache-2.0 (LICENSE), with NOTICE carrying AACR-Bench's attribution and the
modified-work statement. Two parts stay under their own MIT licences, whose
copyright and permission notices travel with them in LICENSES/: Martian's
golden comments and published tool-level judgements (martian_*,
LICENSES/MIT-martian.txt, Copyright (c) 2025 Martian) and Qodo
PR-Review-Bench's issues (qodo_*, LICENSES/MIT-qodo.txt, the Qodo team).
Code quoted anywhere in the release stays under its own project's licence,
copyleft and source-available ones included (prs.repo_license,
martian_prs.repo_license, qodo_prs.repo_license). The vendor text the
run archive and the slim bundle copy verbatim (the Martian judge's
extractions of the re-judged vendors' comments and the blind annotation's
vendor candidates) belongs to its authors, the vendors' review bots that
posted it on public pull requests, and is not covered by this licence. The
run archive and the slim bundle carry the same licence files and their own
NOTICE; the archive's also carries the SWE-PRBench attribution.
Citation
@misc{jakubczak2026realdefects,
title = {Measuring whether a code review finds the real defects: a verified defect benchmark on AACR-Bench, applied to eight reviewers at their best-shot configurations},
author = {Mateusz Jakubczak},
year = {2026},
note = {Quickchat AI. arXiv <arXiv pending>}
}
@misc{jakubczak2026quickreview,
title = {QuickReview: a multi-pass automated code reviewer, its design, its cost, and how it performs on a verified defect benchmark},
author = {Mateusz Jakubczak},
year = {2026},
note = {Quickchat AI. arXiv <arXiv pending>}
}
@misc{jakubczak2026quickreviewbench,
title = {Real defects in code review: QuickReview benchmark data},
author = {Mateusz Jakubczak},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/quickchatai/quickreview-bench}},
note = {Quickchat AI. DOI 10.57967/hf/11598}
}
@misc{zhang2026aacrbenchevaluatingautomaticcode,
title = {AACR-Bench: Evaluating Automatic Code Review with Holistic Repository-Level Context},
author = {Lei Zhang and Yongda Yu and Minghui Yu and Xinxin Guo and Zhengqi Zhuang and Guoping Rong and Dong Shao and Haifeng Shen and Hongyu Kuang and Zhengfeng Li and Boge Wang and Guoan Zhang and Bangyu Xiang and Xiaobin Xu},
year = {2026},
eprint = {2601.19494},
archivePrefix = {arXiv},
primaryClass = {cs.SE}
}
Martian's benchmark: https://github.com/withmartian/code-review-benchmark. Qodo PR-Review-Bench: https://huggingface.co/datasets/Qodo/PR-Review-Bench.
- Downloads last month
- 70