read and read_document block forever on a pipe, and read without end from a link to a device
メンテナーはふだん 1 日以内に返信
まだ誰も着手していません。
評価
調査の方向性
Start with OpenRegular and ReadRegularHead in internal/skills, then trace readTool in internal/exec/bare/tools.go, sniffSense and readSenseFile in internal/session/tools_sense.go, and readDocument in internal/session/tools_doc.go. Run the FIFO reproduction and the ordinary-file command first. Done means the listed unit and e2e cases return promptly, preserve ordinary-file reads, and the manual page in internal/manual/chat/ states the exact refusal.
索引モデルが issue の本文から書いたものです。
説明
What happened
The worker's file-reading hands open whatever path they are given with a blocking open and read it with an unbounded os.ReadFile. A path that is a FIFO (named pipe), or a link to a device such as /dev/zero or /dev/tty, therefore either blocks at the open for as long as nothing writes to the pipe, or reads without end until memory runs out. The turn holding that tool call cannot finish.
This is on dev@f101ad2e7 and unchanged on santos/dev2 (#1410). #1410's review fixed the same class of fault in skill discovery and the Claude Code settings readers (one opener, OpenRegular, in internal/skills, which checks the file type, opens with O_NONBLOCK and bounds the read), but the worker-facing read doors do not use it. The committed path a model is most likely to be pointed at is a skill: use_skill checks that a SKILL.md is a regular file when it hands out the path, and the file can be swapped for a pipe or a link after that check.
The shell door that runs the same read as read_document, today:
$ mkfifo notes.md
$ timeout 6 codeaf doc notes.md; echo "exit=$?"
exit=124
codeaf doc real.md on an ordinary file prints it and exits 0 at once.
Replication
Deterministic (no model).
d=$(mktemp -d) && cd "$d"
mkfifo notes.md
echo '# hi' > real.md
timeout 6 codeaf doc real.md; echo "exit=$?" # prints "# hi", exit=0
timeout 6 codeaf doc notes.md; echo "exit=$?" # prints nothing, exit=124: blocked in the open
The read hand has no shell door, so its twin is a Go test in internal/exec/bare: make a FIFO in t.TempDir(), call ReadTool(dir, 0).Execute(ctx, json.RawMessage({"path":"notes.txt"})) with a 2-second context, and observe it has not returned when the context ends. The session's read wrapper blocks before that, in sniffSense's os.Open (a .txt name is not in its extension table, so it opens the file to sniff magic bytes).
Do not point the reproduction at a link to /dev/zero without a memory limit (ulimit -v): the unbounded read allocates until the process is killed.
Field (real models). In a scratch repository with mkfifo notes.txt, run codeaf exec "read notes.txt and tell me what it says" --timeout 2m with OPENROUTER_API_KEY set. The model's read call never returns and the run ends on the wall. Budget: under a cent.
Where
internal/exec/bare/tools.go:readTool(data, err := os.ReadFile(absPath)), and the sameos.ReadFile(absPath)further down the file in the edit hand's read of the old content.internal/session/tools_sense.go:sniffSense(file, err := os.Open(absolute)) andreadSenseFile(os.ReadFile(absolute)after a size check that a device or a pipe passes with size 0).internal/session/tools_doc.go:readDocument's local rung (data, err := os.ReadFile(absolute)) andlooksLikePlainText(file, err := os.Open(absolute)).
The fix
A path that is not a regular file, or a link to one that is not, is refused at once in the hand's own words, the way skill discovery already refuses one. For example: notes.md is not a regular file (a pipe), so it was not read. Reads are bounded at each hand's existing size ceiling. All of these doors should use one opener; internal/skills' OpenRegular / ReadRegularHead is the shape to reuse or move somewhere both packages can import.
Acceptance
- e2e:
codeaf doc notes.mdon a FIFO exits within a second with exit 1 and the refusal sentence on stderr; the same command on an ordinary file still prints it and exits 0. - e2e:
codeaf doc zero.md, wherezero.mdis a link to/dev/zero, refuses at once and allocates nothing (run it underulimit -v 500000). - Unit:
read(bare and the session wrapper) andread_documentreturn the refusal within 100 ms for a FIFO, a link to/dev/zeroand a link to/dev/tty, and still read an ordinary file and a link to one. - Unit (law): no
os.Open/os.ReadFileon a model-supplied path remains in the three files above outside the shared opener. - The manual page that covers
readandread_document(internal/manual/chat/) states the refusal in its exact words.
- 主要言語
- Go
- スター
- 115
- フォーク
- 14
- 平均マージ
- 9時間 38分
- マージ済み PR(30日)
- 749
環境構築
このプロジェクトの環境構築ファイルはまだ確認していません。まず README を読み、一般的な手順ははじめてのコントリビューションガイドを参照してください。
はじめの一歩
- issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
- 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
- リポジトリをフォークし、ブランチを切って変更します。
- issue 番号を参照したプルリクエストを送ります。
Agent-Field/CodeAF のほかの issue
-
area:chat feature
難易度 2/5 1〜3時間 初心者へのやさしさ 88/100
Agent-Field/CodeAF#1510 ·
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 82/100
Agent-Field/CodeAF#1489 ·
メンテナーはふだん 1 日以内に返信
-
area:chat bug good first issue sev:papercut
難易度 2/5 1〜3時間 初心者へのやさしさ 78/100
Agent-Field/CodeAF#1470 ·
メンテナーはふだん 1 日以内に返信
-
area:chat bug good first issue sev:papercut
難易度 2/5 1〜3時間 初心者へのやさしさ 85/100
Agent-Field/CodeAF#1469 ·
メンテナーはふだん 1 日以内に返信
-
area:chat bug sev:papercut
難易度 2/5 1〜3時間 初心者へのやさしさ 88/100
Agent-Field/CodeAF#1468 ·
メンテナーはふだん 1 日以内に返信
Agent-Field/CodeAF の issue をすべて見る
似ている issue
-
enhancement low priority
難易度 2/5 1〜3時間 初心者へのやさしさ 85/100
eugenioenko/ttt#674 ·
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 72/100
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 84/100
メンテナーはふだん 1 日以内に返信
-
難易度 1/5 1時間未満 初心者へのやさしさ 92/100
GoogleCloudPlatform/k8s-config-connector#13462 ·
メンテナーはふだん 1 日以内に返信
-
bug good first issue
難易度 2/5 1〜3時間 初心者へのやさしさ 85/100
vavallee/bindery#2793 · コメント 1 件 ·
メンテナーはふだん 1 日以内に返信