fix: replace removed --full-auto with --sandbox workspace-write

[zengchang233]

Summary

Fixes #237.

This PR fixes the bug reported in #237 and changes the two call sites listed there. It does not use the fix suggested in the issue (probing for --approve-for-me). The reasons are in the next section.

Codex CLI 0.154.0 removed --full-auto. Every codex exec call that scripts/ask-codex.sh or hooks/loop-codex-stop-hook.sh makes now fails immediately:

error: unexpected argument '--full-auto' found

This breaks /humanize:ask-codex, both Codex passes in /humanize:gen-plan, and the summary review in the RLCR stop hook. I hit it on codex-cli 0.155.1.

This PR replaces the default flag in both places with --sandbox=workspace-write.

Why --sandbox workspace-write and not --approve-for-me or a help-text probe

#237 suggests probing codex exec --help for --approve-for-me, falling back to --full-auto. I went with a plain --sandbox workspace-write instead:

• It matches the old behavior. In codex exec, --full-auto only selected the workspace-write sandbox. The approval policy stayed at the headless default, never. See codex-rs/exec/src/lib.rs at rust-v0.114.0: full_auto maps to SandboxMode::WorkspaceWrite, and the default approval is AskForApproval::Never.
• --approve-for-me is more permissive. It also sets the approval policy to on-request and sends escalation requests to an automatic reviewer (codex-rs/exec/tests/suite/approval_policy.rs at rust-v0.155.1). A command the reviewer approves could run outside the sandbox. That is more than a review step needs.
• No probing needed. -s/--sandbox exists on codex exec in rust-v0.114.0 (codex-rs/exec/src/cli.rs), which is the minimum version the Codex install path requires. It is also still present in 0.155.1 (codex exec --help). So the fix does not need to probe for version-specific flags the way the --disable hook-feature probe does.

HUMANIZE_CODEX_BYPASS_SANDBOX=true|1 still replaces the flag with --dangerously-bypass-approvals-and-sandbox, which is unchanged in 0.155.1.

Changes

• scripts/ask-codex.sh, hooks/loop-codex-stop-hook.sh: the default CODEX_AUTO_FLAG is now --sandbox=workspace-write. It is kept as a single token so the existing array expansion stays the same.
• docs/usage.md: the HUMANIZE_CODEX_BYPASS_SANDBOX default now describes --sandbox workspace-write.
• Tests:
  • tests/test-ask-codex.sh: the default run passes --sandbox=workspace-write and not --full-auto. HUMANIZE_CODEX_BYPASS_SANDBOX=1 still replaces the sandbox flag.
  • tests/test-disable-nested-codex-hooks.sh: the implementation-phase stop hook passes --sandbox=workspace-write and not --full-auto.

Deliberately left alone:

• The mock help text in tests/test-ask-codex.sh (Test B of the --disable probe) mentions --full-auto. It stands in for an older CLI and does not affect the flag under test.
• tests/test-bitlesson-select-routing.sh already asserts that the BitLesson selector never passes --full-auto, and that still holds.

Testing

• bash tests/test-ask-codex.sh: 39 passed, 0 failed. This includes the 2 new tests, which fail against dev without the fix.
• bash tests/test-disable-nested-codex-hooks.sh: 7 passed, 0 failed. This includes 1 new test, which also fails without the fix and shows ... --full-auto -C ... in the captured argv.
• bash tests/run-all-tests.sh: 2261 passed, 4 failed. The same 4 fail on unmodified dev on my machine, for reasons unrelated to this change:
  • test-unified-codex-config.sh checks the built-in gpt-5.5/high fallback, but my ~/.config/humanize/config.json overrides it. With XDG_CONFIG_HOME pointed at an empty dir, the file passes 69/69 on this branch.
  • test-viz.sh and test-streaming.sh need the Python yaml module, which is not installed locally.
• Manual check: with the patched ask-codex.sh, a real codex exec call on codex-cli 0.155.1 succeeds (exit_code=0). Before the patch it failed with unexpected argument '--full-auto'.

🤖 Generated with Claude Code