CLI: code-mode host SIGTRAP on linux-aarch64 silently reports tool success

What version of Codex CLI is running?

codex-cli 0.144.1 (also confirmed as the latest npm version at submission time)

What subscription do you have?

ChatGPT Pro

Which model were you using?

Affected: gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna

Controls that execute the same tool successfully: gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.3-codex-spark

What platform is your computer?

Linux 4.19.294 aarch64 unknown

ASUSWRT-Merlin/Entware on a GT-BE98, using the official npm @openai/codex-linux-arm64 package and its aarch64-unknown-linux-musl binary.

What terminal emulator and version are you using?

Non-interactive SSH and Codex app-server over stdio. The failure reproduces in both codex exec and app-server.

Codex doctor report

codex doctor --json reports overallStatus: ok:

{
  "codexVersion": "0.144.1",
  "overallStatus": "ok",
  "runtime": "npm on linux-aarch64",
  "binary": "aarch64-unknown-linux-musl",
  "auth": "chatgpt",
  "providerReachability": "ok",
  "websocketReachability": "HTTP 101 Switching Protocols",
  "configLoad": "ok",
  "installation": "ok",
  "enabledRelevantFeatures": ["shell_tool", "unified_exec", "code_mode_host", "steer"]
}

What issue are you seeing?

Tool execution through code_mode_host terminates with SIGTRAP on linux-aarch64. More importantly, Codex exits with status 0 and the model answers DONE, even though the command was never executed. This creates a silent false-success condition for automation.

Observed output:

MODEL=gpt-5.6-sol RC=0 PROOF=MISSING
ERROR codex_core::tools::router: error=code-mode host exited with status signal: 5 (SIGTRAP)
ERROR codex_core::tools::router: error=code-mode host exited with status signal: 5 (SIGTRAP)
DONE

MODEL=gpt-5.5 RC=0 PROOF=CODEX_TOOL_OK
DONE

The same proof-file test through app-server produced:

gpt-5.5              PASS
gpt-5.6-sol          FAIL (reply DONE, proof missing)
gpt-5.6-terra        FAIL (reply DONE, proof missing)
gpt-5.6-luna         FAIL (reply DONE, proof missing)
gpt-5.4              PASS
gpt-5.4-mini         PASS
gpt-5.3-codex-spark  PASS

What steps can reproduce the bug?

rm -f /tmp/openai-codex-proof

codex exec \
  --skip-git-repo-check \
  --ephemeral \
  -m gpt-5.6-sol \
  -C /tmp \
  -s danger-full-access \
  "Run one shell command: printf CODEX_TOOL_OK > /tmp/openai-codex-proof . Then answer only DONE." \
  </dev/null

echo "codex_rc=$?"
test -f /tmp/openai-codex-proof && cat /tmp/openai-codex-proof || echo MISSING

Repeat with -m gpt-5.5 as a control. On this host, 5.6 exits 0 and prints DONE while the proof is missing; 5.5 creates the proof correctly.

What is the expected behavior?

The tool should execute successfully on the supported linux-arm64 package. If the tool host crashes, Codex must surface the failure and return a non-zero result instead of reporting successful completion.

Additional information

No proxy is configured, provider and WebSocket reachability checks pass, and the installation check reports healthy. No credentials, IP addresses, thread IDs, or private prompts are included in this report.