From c90886955d63cf2e4d58273559666575d748eec5 Mon Sep 17 00:00:00 2001
From: T <t@t.com>
Date: Mon, 14 Sep 2026 22:24:42 -0400
Subject: [PATCH] fix(AI-CLI-if5u): resolve real tty device before
 backgrounding agent stdin

commit 6f0127b redirected a backgrounded agent's stdin from the literal
/dev/tty alias. On macOS, kqueue refuses to poll an fd opened through that
alias path (ttyname_r resolves it to /dev/tty again rather than a real
device) -- Bun-compiled claude then crashes with EINVAL on launch. Resolve
the controlling terminal's actual device path via `tty` first and redirect
from that instead, falling back to the literal /dev/tty when resolution
fails. Root cause corroborated externally: oven-sh/bun#41504.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PFi83LbqMTznL3DW2sVm9d
---
 .../bugs/ai-cli-if5u-kqueue-dev-tty-einval.md | 120 ++++++++++++++++++
 src/ai_cli/session_script.py                  |  10 +-
 tests/test_session_launch_shell_resolution.py |  14 ++
 3 files changed, 142 insertions(+), 2 deletions(-)
 create mode 100644 docs/bugs/ai-cli-if5u-kqueue-dev-tty-einval.md

diff --git a/docs/bugs/ai-cli-if5u-kqueue-dev-tty-einval.md b/docs/bugs/ai-cli-if5u-kqueue-dev-tty-einval.md
new file mode 100644
index 00000000..a59bf145
--- /dev/null
+++ b/docs/bugs/ai-cli-if5u-kqueue-dev-tty-einval.md
@@ -0,0 +1,120 @@
+---
+title: "Claude Code launch fails when stdin is reopened through /dev/tty"
+category: bug
+tags: [bug, session-launch, macos, terminal, kqueue]
+status: fixed
+severity: P1
+template_version: "bug-1.0.0"
+---
+
+# Claude Code launch fails when stdin is reopened through /dev/tty
+
+## Summary
+
+On macOS, `ai c 1` could repeatedly crash Claude Code with `EINVAL: invalid
+argument, kqueue`. After three exits, the launcher's consecutive-failure circuit
+breaker stopped retrying, so the requested session could not be launched or
+resumed.
+
+## Environment and reproduction
+
+The failure was observed with Bun 1.4.3 on macOS arm64 after commit `6f0127b`
+changed supervised agents to retain terminal stdin. Running `ai c 1` from a
+registered repository printed the Bun crash and then the launcher's
+three-consecutive-exits diagnostic.
+
+The implementation worker's sandbox cannot create tmux sockets or access a real
+TTY. The live symptom and causal boundary were therefore confirmed externally;
+the repository regression test covers the generated shell contract.
+
+## Root cause analysis
+
+The generated `run_agent` function backgrounded Claude Code with stdin reopened
+from the literal controlling-terminal alias:
+
+```bash
+"$@" </dev/tty &
+```
+
+Claude Code is a Bun-compiled executable. On macOS, `ttyname_r()` on a descriptor
+opened through `/dev/tty` can return that alias again, and kqueue rejects polling
+that device with `EINVAL`. The resulting stdin stream failure surfaced from
+`pull()`, terminated Claude Code, and let the supervisor's new circuit breaker
+turn three identical exits into the final launcher diagnostic.
+
+[Bun PR #41504](https://github.com/oven-sh/bun/pull/41504) documents the same
+kernel boundary: a pseudo-terminal slave opened by its resolved device path is
+pollable, while the `/dev/tty` alias is not. That upstream change concerns Bun's
+own non-stdio file polling; this launcher created the problematic stdin
+descriptor before Bun started, so the caller must supply the resolved device.
+
+### Hypothesis ledger
+
+| Hypothesis | Evidence | Result |
+| --- | --- | --- |
+| The consecutive-exit counter caused the crash | It only observes exits after Claude Code emits the Bun error. | Rejected; it exposes and bounds the failure. |
+| Repeated direnv evaluation exhausted kqueue resources | The environment loaded before each visible crash, but the Bun stack identifies stdin polling and the alias behavior is independently documented. | Rejected. |
+| Reopening stdin through `/dev/tty` supplied an unpollable descriptor | The regression began with that redirect, the stack failed in stream `pull()`, and Bun's macOS analysis describes the matching `EINVAL`. | Confirmed. |
+
+## Prior fix attempts
+
+The first investigation correctly stopped without a patch because its sandbox
+could not reproduce the real tmux/TTY boundary. No speculative fix was applied.
+
+## Fix
+
+`run_agent` now calls `tty` before backgrounding and, when it returns a readable
+character device, redirects stdin from that resolved path. The existing
+`/dev/tty` behavior remains the defensive fallback when resolution or validation
+fails; non-terminal launches continue to duplicate fd 0.
+
+This is a contained, reversible change to one generated launch boundary. It
+does not change a public API, introduce a shared abstraction, or cross a
+repository boundary, so no broader redesign is warranted.
+
+## Verification
+
+The frozen regression
+`test_given_terminal_stdin_when_agent_is_backgrounded_then_resolved_device_path_is_preferred`
+generates the real `run_agent` body and requires terminal resolution to precede
+the resolved-path redirect and the literal fallback.
+
+On the unfixed template it failed with:
+
+```text
+AssertionError: run_agent did not resolve the controlling terminal's device path
+1 failed in 0.07s
+```
+
+After the fix it passed:
+
+```text
+1 passed in 0.04s
+```
+
+Nearby launch coverage passed with `98 passed, 5 deselected in 1.63s`.
+Repository lint and formatting gates passed with `All checks passed!` and
+`297 files already formatted`. The full suite reached `2833 passed`, but 47
+tests failed because the sandbox denies tmux/Unix/network sockets, writes below
+the real user state directory, and several unrelated pre-existing portability
+and public-repository hygiene checks. Its final line was:
+
+```text
+47 failed, 2833 passed, 30 skipped, 5 warnings in 56.00s
+```
+
+A live `ai c 1` launch was not run in the implementation sandbox and remains a
+required external verification step.
+
+## Lessons learned
+
+Terminal retention tests must cover not only whether stdin is a TTY, but also
+which device path was used to open the descriptor. A generic terminal alias and
+its resolved pseudo-terminal slave are not interchangeable polling targets on
+macOS.
+
+## Fix log
+
+| Date | Commit | Notes |
+| --- | --- | --- |
+| 2026-09-14 | pending | Implement role, medium effort: resolve the terminal device before backgrounding and add frozen generator-level regression coverage. |
diff --git a/src/ai_cli/session_script.py b/src/ai_cli/session_script.py
index cdd499dd..ec029b76 100644
--- a/src/ai_cli/session_script.py
+++ b/src/ai_cli/session_script.py
@@ -465,9 +465,15 @@ def get_engine_script(
       # Non-interactive shells redirect a background job's stdin from /dev/null.
       # Bash honors <&0, but zsh performs that redirect after the fd duplication;
       # reopening the controlling terminal works for both while preserving the
-      # background PID used for signal forwarding.
+      # background PID used for signal forwarding. Prefer the resolved device:
+      # macOS kqueue cannot poll an fd opened through the /dev/tty alias.
       if [[ -t 0 && -r /dev/tty ]]; then
-        "$@" </dev/tty &
+        _agent_tty=$(tty 2>/dev/null)
+        if [[ -n "$_agent_tty" && -c "$_agent_tty" && -r "$_agent_tty" ]]; then
+          "$@" <"$_agent_tty" &
+        else
+          "$@" </dev/tty &
+        fi
       else
         "$@" <&0 &
       fi
diff --git a/tests/test_session_launch_shell_resolution.py b/tests/test_session_launch_shell_resolution.py
index 0f7718c3..872367bd 100644
--- a/tests/test_session_launch_shell_resolution.py
+++ b/tests/test_session_launch_shell_resolution.py
@@ -299,6 +299,20 @@ def _invoke_run_agent(tmp_path: Path, bin_dir: Path, agent_marker: Path) -> subp
     )
 
 
+def test_given_terminal_stdin_when_agent_is_backgrounded_then_resolved_device_path_is_preferred():
+    """The generated launcher must not hand agents a literal /dev/tty stdin fd."""
+    body = _run_agent_body()
+
+    resolve_tty = "_agent_tty=$(tty 2>/dev/null)"
+    resolved_redirect = '"$@" <"$_agent_tty" &'
+    literal_fallback = '"$@" </dev/tty &'
+
+    assert resolve_tty in body, "run_agent did not resolve the controlling terminal's device path"
+    assert resolved_redirect in body, "run_agent did not redirect stdin from the resolved terminal device"
+    assert literal_fallback in body, "run_agent must retain the existing fallback when tty resolution fails"
+    assert body.index(resolve_tty) < body.index(resolved_redirect) < body.index(literal_fallback)
+
+
 def test_given_no_direnv_when_run_agent_invoked_then_the_agent_command_still_runs(tmp_path):
     """direnv must be an enhancement, never a precondition for starting the agent.