Codex CLI crashes with SIGABRT when invoked through AutoAide shell spawn

[SparkssL]

Bug Description

When sending a message through AutoAide's CLI shell, the underlying Codex CLI process crashes with SIGABRT, even though the exact same command works perfectly when run directly in the terminal.

Environment

• OS: macOS (Apple Silicon / MacBook Air)
• Node.js: v22+
• Codex CLI: codex-cli 0.118.0
• AutoAide: latest main branch
• Shell: zsh (with conda base environment active)

Steps to Reproduce

1. Install AutoAide: git clone && npm install && npm link
2. Run autoaide and complete the bootstrap flow
3. Type any message, e.g. hello
4. Observe:

[autoaide> hello
Running on [main]...
autoaide>
Codex interrupted (SIGABRT).

Direct CLI Test (Works Fine)

Running the same underlying command manually succeeds:

$ echo "hello" | codex exec --skip-git-repo-check --json -
{"type":"thread.started","thread_id":"019d4a05-cbb3-7232-86a4-82d2bfe4b1a5"}
{"type":"turn.started"}
{"type":"item.completed","item":{"id":"item_0","type":"agent_message","text":"Hello. What do you need?"}}
{"type":"turn.completed","usage":{"input_tokens":17226,"cached_input_tokens":3456,"output_tokens":70}}

Suspected Cause

In src/codex-runner.mjs, the command is spawned via zsh -lc:

spawn(shellSpec.command, [...shellSpec.args, command], {
  cwd,    // ~/.autoaide/workspace
  env: process.env,
  stdio: ["pipe", "pipe", "pipe"],
});

Possible causes:

1. Workspace context injection — buildWorkspacePrompt() prepends SOUL.md / IDENTITY.md / USER.md / TOOLS.md content via stdin. A malformed or overly large prompt may trigger the abort.
2. CWD mismatch — ~/.autoaide/workspace may interact differently with Codex's internal path resolution vs a normal directory.
3. Login shell interference — zsh -lc loads the full login profile (conda, nvm, etc.) which may set environment variables or aliases that conflict with Codex when spawned as a child process.
4. Signal handling — The daemon or parent process management may inadvertently send signals to the Codex child process.

Suggested Investigation

• Add error logging for stderr output of the spawned Codex process in codex-runner.mjs
• Test with zsh -c instead of zsh -lc to rule out login shell profile issues
• Log the exact prompt being piped to stdin to check for encoding or formatting issues

[Moshiii]

Fixed on 9f43d5f (fix: avoid login-shell codex crashes).

What changed:

• src/codex-runner.mjs: default Codex child-process shell changed from login shell (zsh -lc) to non-login shell (zsh -c)
• added opt-in fallback via AUTOAIDE_LOGIN_SHELL=true if someone explicitly needs login-shell behavior
• src/cli.mjs: interrupted runs now include stderr, so future signal failures are easier to diagnose
• added unit coverage in test/unit/codex-runner.test.mjs

I could not reproduce the SIGABRT reliably on the current machine, but this removes the most likely source of interference: full login-shell profile loading (conda, nvm, shell init hooks, etc.) during Codex spawn.

I also re-ran a local CLI smoke test (hello) after the change and it completed normally.