[bompus]

The ACP panel client (hr-code-review) exits 4 and discards a review when git status differs between launch and the end of a seat. It compared the whole --porcelain text, so any new file counted, including the bundled skills Codex writes under skills/system/ when a session starts. A Terra-max review of setupLinuxHost#1810 was discarded this way on 2026-11-03 because Codex wrote an untracked skills/system/run-workflow/SKILL.md.

This ports the fix from chief-of-staff c72a0fc1 (#836) into the single-file client:

• The status is read with -z --untracked-files=all and compared per path. Untracked files under skills/system/ in the launch directory (resolved with git rev-parse --show-prefix) are not counted. A tracked change there, and a file that existed at launch and is gone at the end, still count.
• The log names up to ten changed paths.
• A status that fails inside a checkout is a change, not "nothing to compare"; a failure at launch exits 1.
• Differences from the chief-of-staff version, found by an independent read (Sol xhigh) of the first draft: the rename or copy source path is part of the comparison (two renames that swap their sources were invisible), the listing is compared as bytes (names that differ only in bytes that are not valid UTF-8 were the same string), a failed final check or prefix lookup is reported as a change instead of passing, and the client stays one file because it is copied around standalone.

A seat's own new file under skills/system/ still goes undetected; references/acp-client.md says so.

Checks: bun run check, and GIT_CONFIG_GLOBAL=/dev/null GIT_CONFIG_SYSTEM=/dev/null bun test test skills (969 pass, 19 skip, 0 fail). The nine new tests in acp-client.test.mjs use real Git; the three for the review's findings fail on the first draft. The Windows job was not run locally: the new file names avoid characters Windows rejects, and the invalid-UTF-8 test runs on Linux only.

The setupLinuxHost pin adoption of the new revision is separate.

[bompus-ai-review]

Bompus AI Review (GPT-6.1 Terra, max reasoning; SWE-2 Max, max reasoning): 4 finding(s).

[bompus-ai-review]

Bompus AI Review (GPT-6.1 Terra, max reasoning; SWE-2 Max, max reasoning): 4 finding(s).

[bompus-ai-review]

Bompus AI Review (GPT-6.1 Terra, max reasoning; SWE-2 Max, max reasoning): 0 finding(s).

Review 3: only high findings and above are posted from here on.

[bompus]

Pushed 874419b. Review 3 posted no findings, but from review 3 on the bot reports only high findings, and this client handles file names a seat controls, so I read the final diff by hand and found two defects: a missing git program had become a hard launch failure (it again means nothing to compare, as before the change; every other git failure still exits 1), and a control character such as a newline in a changed file's name could forge a log line (control characters are now written as \xNN). Both have tests that fail on the previous head. bun run check and bun test test skills pass (974 pass, 19 skip, 0 fail).

[bompus-ai-review]

Bompus AI Review (GPT-6.1 Terra, max reasoning; SWE-2 Max, max reasoning): 0 finding(s).

Review 4: only high findings and above are posted from here on.