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 same os.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)) and readSenseFile (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)) and looksLikePlainText (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.md on 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, where zero.md is a link to /dev/zero, refuses at once and allocates nothing (run it under ulimit -v 500000).
- Unit:
read (bare and the session wrapper) and read_document return the refusal within 100 ms for a FIFO, a link to /dev/zero and a link to /dev/tty, and still read an ordinary file and a link to one.
- Unit (law): no
os.Open / os.ReadFile on a model-supplied path remains in the three files above outside the shared opener.
- The manual page that covers
read and read_document (internal/manual/chat/) states the refusal in its exact words.
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/zeroor/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@f101ad2e7and unchanged onsantos/dev2(#1410). #1410's review fixed the same class of fault in skill discovery and the Claude Code settings readers (one opener,OpenRegular, ininternal/skills, which checks the file type, opens withO_NONBLOCKand 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_skillchecks that aSKILL.mdis 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:codeaf doc real.mdon an ordinary file prints it and exits 0 at once.Replication
Deterministic (no model).
The
readhand has no shell door, so its twin is a Go test ininternal/exec/bare: make a FIFO int.TempDir(), callReadTool(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'sreadwrapper blocks before that, insniffSense'sos.Open(a.txtname 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/zerowithout 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, runcodeaf exec "read notes.txt and tell me what it says" --timeout 2mwithOPENROUTER_API_KEYset. The model'sreadcall 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/ReadRegularHeadis the shape to reuse or move somewhere both packages can import.Acceptance
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.codeaf doc zero.md, wherezero.mdis a link to/dev/zero, refuses at once and allocates nothing (run it underulimit -v 500000).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.os.Open/os.ReadFileon a model-supplied path remains in the three files above outside the shared opener.readandread_document(internal/manual/chat/) states the refusal in its exact words.