Greptile Summary

This PR fixes issue #40179 by preventing shell startup files from overriding daemon-inherited environment variables during exec command execution. It adds a new helper resolvePosixShellArgs that selects shell-specific flags (zsh -f -c, bash --noprofile --norc -c, all others -c) and wires them through every branch of getShellConfig.

Key changes:

• New resolvePosixShellArgs(shellPath) function that maps shell name to the appropriate startup-suppression flags
• All getShellConfig return paths now call this helper instead of hardcoding ["-c"]
• Tests extended to assert the correct args for each shell branch (bash, sh, zsh, fish fallback, unset SHELL)

Observation:

• Fish shell (--no-config) is intentionally omitted from the startup-suppression logic per the PR description. When fish is the only available shell and is used as the final fallback, ~/.config/fish/config.fish can still override the inherited daemon environment. If this edge case matters, adding "fish" to resolvePosixShellArgs with ["--no-config", "-c"] would close the gap.

Confidence Score: 4/5

• This PR is safe to merge; the changes are narrowly scoped and the fix is technically correct.
• The implementation correctly uses -f for zsh (suppresses all startup files, equivalent to --no-rcs) and --noprofile --norc for bash, which are the standard flags for deterministic non-interactive invocations. Tests cover the new behavior across all relevant branches. The only intentional omission is fish's --no-config, which is a known, documented trade-off in the PR rather than an oversight.
• No files require special attention; src/agents/shell-utils.ts is the only changed production file and the diff is small and well-tested.

_{Last reviewed commit: acb1b7b}

Addressed the fish fallback gap noted in feedback.

Changes pushed to fix/issue-40179-exec-launchd-env:

• resolvePosixShellArgs now returns ["--no-config", "-c"] for fish.
• Updated the fish fallback test in src/agents/shell-utils.test.ts to assert the new args.

Validation run locally on this branch:

• pnpm build
• pnpm format:check
• pnpm tsgo
• pnpm lint
• pnpm test

Investigated the failing macOS job.

This branch is 6,973 commits behind origin/main and the failure burst is repo-wide, not in the shell-utils diff from this PR. The job is failing across unrelated suites (memory, outbound channels, ACP, Slack/Telegram/Discord, etc.), so I’m not going to push a speculative change here and pretend it fixes the PR.

Current blocker:

• stale branch / unrelated CI avalanche
• touched files in this PR: src/agents/shell-utils.ts, src/agents/shell-utils.test.ts
• failing areas are far outside that scope

Recommended next step: rebase this branch onto current main (or recreate it from fresh main) and rerun CI. That should tell us whether there is still a real shell-utils regression to fix.

Contributor cycle update:\n\n- Rebased onto current and resolved the merge conflict in .\n- Verified PR scope is still limited to:\n - \n - \n- Verified fish startup suppression is present () alongside existing zsh/bash suppression paths.\n\nCurrent head: \nChecks have restarted on the rebased head.

Correction (previous comment had shell-escaping issues):

Contributor cycle update:

• Rebased fix/issue-40179-exec-launchd-env onto current origin/main and resolved the merge conflict in src/agents/shell-utils.ts.
• Verified PR scope is still limited to:
  • src/agents/shell-utils.ts
  • src/agents/shell-utils.test.ts
• Verified fish startup suppression is present (["--no-config", "-c"]) alongside existing zsh/bash suppression paths.

Current head: 8d1841b2f715f2bfc4b2ca03994baa64cc374bf5
Checks have restarted on the rebased head.

Codex automated review: keeping this open.

Keep #40200 open. Current main has related PATH/login-shell handling for host exec, and #40179 was closed after a Codex review, but main still does not implement this PR's requested command-shell startup suppression for zsh, bash, or fish. The PR remains a plausible focused fix for exec commands whose invoked shell startup files override the sanitized daemon environment.

Best possible solution:

Keep the PR open for maintainer review. The best path is to rebase it onto current main, preserve the existing sh/bash PATH fallback and non-interactive shell placeholder handling, add/keep regression tests for zsh -f, bash --noprofile --norc, and fish --no-config, then run the changed exec/agents lanes. If maintainers decide startup files should remain allowed, that should be an explicit product/security decision with docs or a narrower follow-up, not an automated close.

What I checked:

• Current main still uses plain POSIX -c args: On current main, getShellConfig returns ["-c"] for fish fallback shells, for envShell including /bin/zsh or /bin/bash, and for the sh/bash fallback. There is no resolvePosixShellArgs helper or zsh/bash/fish startup-suppression behavior on main. (src/agents/shell-utils.ts:70, 724e92505adf)
• Host exec uses getShellConfig for command execution: runExecProcess builds childArgv from getShellConfig(), so the shell args from shell-utils are the args used for non-PTY host exec commands. (src/agents/bash-tools.exec-runtime.ts:708, 724e92505adf)
• PTY exec also uses getShellConfig: The process supervisor's PTY path also calls getShellConfig and passes shellArgs to createPtyAdapter, so the missing suppression affects both child and PTY command paths. (src/process/supervisor/supervisor.ts:134, 724e92505adf)
• Current tests do not cover the PR behavior: Existing shell-utils tests mainly assert shell selection; they do not assert zsh -f, bash --noprofile --norc, or fish --no-config args, except for one placeholder-shell ["-c"] case. The PR diff adds those missing assertions. (src/agents/shell-utils.test.ts:56, 724e92505adf)
• Related main work is PATH-specific, not command-shell startup suppression: Current main imports login-shell PATH for host=gateway and sanitizes env for that PATH probe, but the actual exec command still runs through getShellConfig. This is related to [Bug]: exec tool does not inherit gateway's environment variables (launchd) #40179 but does not replace the PR's zsh/bash/fish shell-arg change. (src/agents/bash-tools.exec.ts:1590, 724e92505adf)
• Upstream shell behavior supports the concern: zsh documentation says user startup files such as $ZDOTDIR/.zshenv are read by default after /etc/zshenv, and the RCS option is controlled by +f/NO_RCS; fish documents --no-config as the flag to skip config files. Bash --help lists --noprofile and --norc.

Remaining risk / open question:

• The PR likely needs a current-main rebase and targeted validation because main has since added PATH/login-shell handling around the same exec surface.
• Before merge, maintainers should decide whether zsh -f's unavoidable /etc/zshenv behavior is acceptable and whether bash --noprofile --norc is intentionally redundant for non-interactive bash.

Codex Review notes: model gpt-5.5, reasoning high; reviewed against 724e92505adf.

Re-verified this against current main before acting.

What I checked:

• Current origin/main still does not have the shell startup-suppression behavior from this PR. src/agents/shell-utils.ts still returns plain ["-c"] for env-provided zsh/bash/fish paths, while this PR keeps the change scoped to src/agents/shell-utils.ts and src/agents/shell-utils.test.ts.
• I rebased the PR branch locally onto current origin/main in a clean worktree. The rebase itself was clean.
• I then ran the full local gate on the rebased tree: pnpm build, pnpm tsgo, and pnpm lint passed; pnpm test failed.
• The failing tests were outside this PR scope:
  • src/agents/skills.sherpa-onnx-tts-bin.test.ts
  • src/cli/gateway-cli/run.option-collisions.test.ts
• I reproduced those same two failures on a separate clean origin/main worktree, so this is current-main baseline red, not a shell-utils regression from this PR.

I did not force-push the rebased branch because that would rewrite the PR while the unrelated current-main test failures are still present. Once baseline is green again, this still looks like a focused fix worth rebasing and pushing cleanly.

ProjectClownfish could not safely update this branch, so it opened a narrow replacement PR instead.

Replacement PR: #73969
Source PR: #40200
Contributor credit is preserved in the replacement PR body and changelog plan.

Thanks @NewdlDewdl. The focused startup-file suppression fix from this PR has landed via the replacement PR #73969 in ea9f172, with the contributor credit preserved in the changelog.