Regression in Homebrew Codex CLI 0.144.1: GPT-5.6 Code Mode fails with code-mode host exited during handshake, while ChatGPT bundled CLI works #32696

Activity

1. skyforevers

   added

   bugSomething isn't workingSomething isn't working

   on Jul 13, 2026
2. github-actions

   added

   CLIIssues related to the Codex CLIIssues related to the Codex CLI

   execIssues related to the `codex exec` subcommandIssues related to the `codex exec` subcommand

   tool-callsIssues related to tool callingIssues related to tool calling

   on Jul 13, 2026
3. github-actions commented on Jul 13, 2026

   

   github-actionsbot

   on Jul 13, 2026 – with GitHub Actions

   Contributor

   More actions

   Potential duplicates detected. Please review them and close your issue if it is a duplicate.

   • codex 0.144.0 homebrew cask is missing codex-code-mode-host every command is failing with "failed to spawn code-mode host" #31906
   • Tool calling fails with brew #31907

   Powered by Codex Action
4. bbingz commented on Jul 25, 2026

   

   bbingz

   on Jul 25, 2026

   More actions

   Analysis (community)

   Residual analysis for #32696 (Homebrew 0.144.1: code-mode host exited during handshake while the ChatGPT-bundled CLI works) under docs/contributing.md — invitation-only. No unsolicited PR.

   Root-cause hypothesis

   Your A/B result (same version, Homebrew build fails, bundled build works) has a precise OSS-side structural explanation: the code-mode host resolver treats the CLI binary's own invocation path as the source of truth for locating the host binary, and the handshake-failure class bypasses the one safety net that was added for the missing-binary class. Verified at HEAD 4c43465133:

   1. The host path is resolved as a sibling of an un-canonicalized current_exe(), latched once, with no validation. default_host_program() (codex-rs/code-mode/src/remote_session.rs:572-577) calls resolve_host_program(var_os(CODEX_CODE_MODE_HOST_PATH), current_exe()); the resolver (:579-597) returns the env override verbatim if set (no exists() check, no canonicalize), otherwise current_exe().parent().join("codex-code-mode-host") — and there is no .canonicalize() anywhere in the code-mode crate (grep = 0 hits). On macOS, std::env::current_exe() returns the symlink path, not the link target (Linux is immune: /proc/self/exe is already resolved — this is consistent with the Darwin-only host-symlink logic the installer grew in the 0.144.1 packaging fix). The result is then latched once in ProcessOwnedCodeModeSessionProvider::default() (remote_session.rs:81-85), so a stale path survives for the process lifetime even if you repair the install mid-session. A channel-specific path divergence (Homebrew formula layout vs. ChatGPT bundle layout) therefore selects a different — possibly stale or mismatched — host binary per channel, which is exactly your A/B.

   2. The 0.144.1 in-process fallback covers only Spawn + NotFound — handshake failures fall straight through. Commit 8cf9a1b1f8 (#31899) added the degrade-to-embedded-V8 net at remote_session.rs:99-113, gated by error.host_program_not_found() (code-mode/src/remote_session/connection.rs:74-81), which matches only ConnectionError::Spawn with io::ErrorKind::NotFound. Your error string — "code-mode host exited during handshake" — is produced at connection.rs:291 and wrapped as ConnectionError::Other (:304-308; handshake timeout takes the same Other path). Other bypasses the NotFound check and propagates via the bare Err(error) => return Err(error.to_string()) arm (remote_session.rs:113), so when the host binary exists but exits during the V1 handshake (version/channel mismatch between the resolved host and the CLI — precisely what a stale sibling resolution produces), there is no degradation path: every code-mode tool call fails wholesale, and the model falls back to describing the environment instead of executing — matching your report verbatim.

   3. Two amplifier gaps make this invisible until it breaks. (a) codex doctor never validates the host binary's existence/executability, so it reports the install consistent while the resolved host path is wrong (also reported on #31831). (b) The flip to the in-process fallback (remote_session.rs:99-113) emits zero user-visible notice, so the NotFound variant degrades silently while the handshake variant fails loudly — two failure classes with opposite visibility and no telemetry on either.

   What this analysis deliberately does NOT cover: why the Homebrew 0.144.1 host binary exits during handshake while the bundled one does not is packaging-level (channel co-location/signing/layout) and cannot be pinned from this repo — the OSS anchor is the error surface and the missing fallback, not the binary divergence itself. Adjacent spawn-family reports: #31831 (missing host, 0.144.0 — the NotFound class; the packaging half was fixed in 0.144.1, the resolver gaps above remain), #32495 (stale ~/.local/bin/codex symlink after the Codex.app → ChatGPT.app migration re-triggering sibling resolution against a removed bundle). The code-mode token-billing axis (the wait tool's unclamped yield_time_ms) is a different mechanism on the same subsystem, analyzed separately on #32309.

   Fix outline (all OSS, codex-rs/code-mode/): (a) canonicalize current_exe() before parent().join(...) in resolve_host_program and add an exists() precheck with an ordered fallback chain (env override → canonicalized sibling → raw sibling → bare-name PATH lookup), warning when an env override names a nonexistent file; (b) resolve lazily in create_session (or invalidate on failure) instead of latching in Default, so a repaired install takes effect without a restart; (c) extend the fallback classifier beyond Spawn+NotFound to handshake ConnectionError::Other — a present-but-incompatible host should degrade to the embedded V8 session exactly like an absent one, or at minimum enrich the error with a remediation hint ("host/CLI channel mismatch; reinstall or point CODEX_CODE_MODE_HOST_PATH at the matching binary"); (d) emit a one-time user-visible notice on any fallback flip and add a host-binary existence check to codex doctor.

   Fail-first tests I'd want before any merge: (1) invoke via a symlink whose directory lacks the host binary → resolver canonicalizes and finds the real sibling (today: resolves against the symlink's parent); (2) host binary present but written to exit immediately → session degrades to in-process or returns an actionable error (today: bare ConnectionError::Other propagates); (3) a stale env override plus a valid sibling → warning + sibling wins; (4) repair-after-start (host binary appears mid-session) → next create_session succeeds without a process restart (today: latched path sticks). Happy to draft the test patch if a maintainer wants it. No unsolicited PR from me.