mac 端审查 Python 项目失败

[hardy1970]

OpenCodeReview Version

v1.1.10

Operating System

macOS (Apple Silicon)

Installation Method

GitHub Release binary

LLM Provider

Anthropic (Claude)

Bug Description

❯ ./ocr r
[ocr] 1 file(s) changed, reviewing 1 in /xxxx/project/xxx
[ocr] Skipping ocr (~13694460 tokens exceeds 80% of max_tokens(58888))
[ocr] Pre-filtered 1 file(s) exceeding 80% of max_tokens
Error: review failed: all diffs filtered out by token size
❯ pwd
/xxxx/project/groven
❯ ./ocr r --preview

Preview: 1 file(s) changed | +167965 -0

Will review (1):
[B] ocr +167965 -0

Steps to Reproduce

LLM 模型使用 minimax 2.7 ，审查 Python 项目报错

Expected Behavior

生成审查结果

Logs / Error Output

❯ ./ocr r
[ocr] 1 file(s) changed, reviewing 1 in /xxxxx/project/xxx
[ocr] Skipping ocr (~13694460 tokens exceeds 80% of max_tokens(58888))
[ocr] Pre-filtered 1 file(s) exceeding 80% of max_tokens
Error: review failed: all diffs filtered out by token size
❯ ./ocr r --preview

Preview: 1 file(s) changed  |  +167965  -0

Will review (1):
  [B]  ocr                  +167965 -0

Additional Context

No response

[stay-foolish-forever]

Hi @hardy1970, thanks for reporting this!

Why you're seeing the error

The error you're seeing is expected behavior, though the outcome (no review produced) clearly isn't what you want. Before invoking the LLM, our tool checks whether each diff exceeds a predefined token threshold (currently 58888), and skips any diff that exceeds it. In your case there's only a single diff in the review set, and that diff is far above the threshold (~13694460 tokens), so it gets filtered out and the review ends with all diffs filtered out by token size.

A likely cause: the diff is from a binary file

Looking at the preview output you shared:

[B]  ocr                  +167965 -0

the file being reviewed is literally named ocr, marked as [B] (binary), with 167,965 lines added and 0 removed. This strongly suggests the diff comes from a binary file — possibly the ocr executable itself that was accidentally added to the repository. Could you confirm whether this is the case? Running something like file ./ocr or checking the file's contents would help us verify.

What this issue reveals about the tool

Your report points to two real shortcomings on our side:

1. The diff filtering threshold is not configurable. The current value (58888) comes from our internal experience and represents a balance between review quality and cost. We agree this isn't ideal for every project. If you (or anyone else watching) believe a configurable threshold would better fit your scenarios, please open a follow-up Issue or Discussion — we'd be glad to work through the design with you.

2. Binary files are not filtered out before the size check. They should never have been candidates for LLM review in the first place. We'll fix this in an upcoming release so binary files are excluded up front instead of getting caught by the token-size filter.

If the file really is a large text file

If it turns out you're genuinely trying to review a very large source file, the most practical mitigation today is to split the file into smaller, more cohesive units. The benefit there goes beyond our tool — large monolithic files are hard for human reviewers and downstream tooling alike, and splitting them generally improves the reviewability of the change.

Thanks again for the detailed report, and we'll follow up here once the binary-file filtering fix lands.

[hardy1970]

二进制 使用 release 版本 https://github.com/alibaba/open-code-review/releases/download/v1.1.12/opencodereview-darwin-arm64

当我将 二进制重命名并 放到 mv ocr /usr/local/bin/ 下之后，显示没有支持的
❯ mv ocr /usr/local/bin/
❯ ocr review --preview
No files changed.
❯ ocr review
[ocr] 0 file(s) changed, reviewing 0 in /xxx/project/xxx
[ocr] No supported files changed. Skipping review.
[ocr] Summary: 0 file(s) reviewed, 0 comment(s), ~0 token(s) used, 0s elapsed
No comments generated. Looks good to me.
❯ ls -al
total 64
drwxr-xr-x@ 17 asa staff 544 Jun 3 13:42 .
drwxr-xr-x 66 asa staff 2112 Jun 2 16:21 ..
drwxr-xr-x@ 4 asa staff 128 Jun 2 16:19 .claude
drwxr-xr-x@ 15 asa staff 480 Jun 3 13:48 .git
-rw-r--r--@ 1 asa staff 49 Jun 2 16:19 .gitattributes
-rw-r--r--@ 1 asa staff 347 Jun 3 10:57 .gitignore
drwxr-xr-x@ 10 asa staff 320 Jun 3 10:57 .idea
drwxr-xr-x@ 4 asa staff 128 Jun 3 10:57 .vscode
-rw-r--r--@ 1 asa staff 5339 Jun 2 16:19 AGENTS.md
-rw-r--r--@ 1 asa staff 5339 Jun 2 16:19 CLAUDE.md
-rw-r--r--@ 1 asa staff 3861 May 12 10:24 README.md
drwxr-xr-x@ 17 asa staff 544 Jun 2 16:19 backend
drwxr-xr-x@ 7 asa staff 224 May 12 10:25 docs
drwxr-xr-x@ 18 asa staff 576 Jun 2 16:19 frontend
drwxr-xr-x@ 15 asa staff 480 May 12 10:25 samples
drwxr-xr-x@ 3 asa staff 96 Jun 2 16:19 scripts
-rwxr-xr-x@ 1 asa staff 365 Jun 2 16:19 setup-hooks.sh

[stay-foolish-forever]

二进制 使用 release 版本 https://github.com/alibaba/open-code-review/releases/download/v1.1.12/opencodereview-darwin-arm64

当我将 二进制重命名并 放到 mv ocr /usr/local/bin/ 下之后，显示没有支持的 ❯ mv ocr /usr/local/bin/ ❯ ocr review --preview No files changed. ❯ ocr review [ocr] 0 file(s) changed, reviewing 0 in /xxx/project/xxx [ocr] No supported files changed. Skipping review. [ocr] Summary: 0 file(s) reviewed, 0 comment(s), ~0 token(s) used, 0s elapsed No comments generated. Looks good to me. ❯ ls -al total 64 drwxr-xr-x@ 17 asa staff 544 Jun 3 13:42 . drwxr-xr-x 66 asa staff 2112 Jun 2 16:21 .. drwxr-xr-x@ 4 asa staff 128 Jun 2 16:19 .claude drwxr-xr-x@ 15 asa staff 480 Jun 3 13:48 .git -rw-r--r--@ 1 asa staff 49 Jun 2 16:19 .gitattributes -rw-r--r--@ 1 asa staff 347 Jun 3 10:57 .gitignore drwxr-xr-x@ 10 asa staff 320 Jun 3 10:57 .idea drwxr-xr-x@ 4 asa staff 128 Jun 3 10:57 .vscode -rw-r--r--@ 1 asa staff 5339 Jun 2 16:19 AGENTS.md -rw-r--r--@ 1 asa staff 5339 Jun 2 16:19 CLAUDE.md -rw-r--r--@ 1 asa staff 3861 May 12 10:24 README.md drwxr-xr-x@ 17 asa staff 544 Jun 2 16:19 backend drwxr-xr-x@ 7 asa staff 224 May 12 10:25 docs drwxr-xr-x@ 18 asa staff 576 Jun 2 16:19 frontend drwxr-xr-x@ 15 asa staff 480 May 12 10:25 samples drwxr-xr-x@ 3 asa staff 96 Jun 2 16:19 scripts -rwxr-xr-x@ 1 asa staff 365 Jun 2 16:19 setup-hooks.sh

这似乎是一个正常现象，这说明当前的仓库的工作区没有做任何的改动（即没有任何的 staged、unstaged、untracked 的文件改动），或许你可以尝试 ocr review --commit <commit_id> 来直接评审最近一次 commit 的改动，或者你也可以参考 README 来获得进一步的关于 CLI 使用的帮助 :)

[hardy1970]

二进制 使用 release 版本 https://github.com/alibaba/open-code-review/releases/download/v1.1.12/opencodereview-darwin-arm64
当我将 二进制重命名并 放到 mv ocr /usr/local/bin/ 下之后，显示没有支持的 ❯ mv ocr /usr/local/bin/ ❯ ocr review --preview No files changed. ❯ ocr review [ocr] 0 file(s) changed, reviewing 0 in /xxx/project/xxx [ocr] No supported files changed. Skipping review. [ocr] Summary: 0 file(s) reviewed, 0 comment(s), ~0 token(s) used, 0s elapsed No comments generated. Looks good to me. ❯ ls -al total 64 drwxr-xr-x@ 17 asa staff 544 Jun 3 13:42 . drwxr-xr-x 66 asa staff 2112 Jun 2 16:21 .. drwxr-xr-x@ 4 asa staff 128 Jun 2 16:19 .claude drwxr-xr-x@ 15 asa staff 480 Jun 3 13:48 .git -rw-r--r--@ 1 asa staff 49 Jun 2 16:19 .gitattributes -rw-r--r--@ 1 asa staff 347 Jun 3 10:57 .gitignore drwxr-xr-x@ 10 asa staff 320 Jun 3 10:57 .idea drwxr-xr-x@ 4 asa staff 128 Jun 3 10:57 .vscode -rw-r--r--@ 1 asa staff 5339 Jun 2 16:19 AGENTS.md -rw-r--r--@ 1 asa staff 5339 Jun 2 16:19 CLAUDE.md -rw-r--r--@ 1 asa staff 3861 May 12 10:24 README.md drwxr-xr-x@ 17 asa staff 544 Jun 2 16:19 backend drwxr-xr-x@ 7 asa staff 224 May 12 10:25 docs drwxr-xr-x@ 18 asa staff 576 Jun 2 16:19 frontend drwxr-xr-x@ 15 asa staff 480 May 12 10:25 samples drwxr-xr-x@ 3 asa staff 96 Jun 2 16:19 scripts -rwxr-xr-x@ 1 asa staff 365 Jun 2 16:19 setup-hooks.sh

这似乎是一个正常现象，这说明当前的仓库的工作区没有做任何的改动（即没有任何的 staged、unstaged、untracked 的文件改动），或许你可以尝试 ocr review --commit <commit_id> 来直接评审最近一次 commit 的改动，或者你也可以参考 README 来获得进一步的关于 CLI 使用的帮助 :)

我想要针对整个仓库进行 review 呢，类似于 sast 的功能，暂时不支持是吗

[stay-foolish-forever]

当前 OCR 走的是经典的面向 diff 的代码评审流程，三种模式都建立在变更之上：

• 工作区模式：staged + unstaged + untracked
• --commit：单提交 vs 父提交
• --from/--to：分支区间（merge-base）

每文件独立跑 plan + 工具循环，聚焦"本次改了什么"。全量评审目前不直接支持——它的 token 成本、并发/超时假设、跨文件视角都和当前 per-file diff 流程不一致，简单加个开关并不合适。但是我们认为全量评审是一个非常重要的需求，或许你可以开启一个单独的 Issue 或者 Discussion 来讨论全量评审的需求，我们会根据社区反馈来判断这个需求的优先级，后续打算针对全量场景特定优化一个 Agent 来做这件事。

虽然我们当前不直接支持全量评审，但是你可以采用一个临时的 workaround 来实现和全量评审相同的效果：

临时 workaround

利用 workspace 模式把 untracked 文件按"整文件新增"渲染的特性，可以在仓库之外临时跑一次：

# 1. 把仓库整个拷出来到临时目录
cp -a /path/to/your/project /tmp/full-review
cd /tmp/full-review

# 2. 抹掉原 git 历史,重新 init
rm -rf .git
git init

# 4. 先 preview 确认范围,再正式跑
ocr review --preview
ocr review

原理：workspace 模式下，untracked 文件会被包成 --- /dev/null → +++ b/<path> 的整文件 hunk，等价于"全部新增"。

几点提醒

1. 保护本地仓库：务必在临时目录里操作。不要在原仓库执行 git rm -r --cached .、git rebase -i --root 这类改写索引/历史的命令——一旦这么做将很难干净复原。
2. token 成本：全量评审 ≈ 文件数 × 每文件多轮 LLM 调用，对单价高的模型成本会显著放大。建议先用 --preview 确认命中清单，对评审规模建立初步预期，同时时刻关注 token 账单，避免产生大额支出。
3. 超大文件/生成物：你这次遇到的 exceeds 80% of max_tokens 就是单文件 token 预过滤兜底。请在 .gitignore 或 rule.json 里把二进制、*.min.js、protobuf 生成代码、超大 fixture 等排除掉，避免把上下文撑爆。