Fix #281: recreate stale default workdirs after reboot when only logs remain

[wwind123]

Summary

Fixes #281

After a machine reboot, a tool-owned default agent workdir under /tmp/coding-review-agent-loop/... can survive as a bare directory containing only tool-created artifacts (e.g., .agent-loop-logs/) with no git checkout. Previously the tool aborted with:

Default claude workdir exists but is not a git checkout: ...
Remove it or pass an explicit agent directory

This PR adds auto-recreation for that stale scenario:

• _is_stale_default_workdir(path) — detects a dir that exists, has no .git/, and whose contents are a non-empty subset of known tool artifact names (.agent-loop-logs, .agent-loop, .agent-loop-responses). Empty dirs are intentionally excluded (no artifacts → could be any empty dir).
• ensure_temp_checkout() — calls shutil.rmtree() on stale default-owned workdirs before the existing clone-if-missing branch, so the normal gh repo clone + validation flow runs fresh.
• Explicit agent dirs (--claude-dir / --codex-dir / --gemini-dir) are unchanged — they still fail clearly when not a git checkout.
• Existing dirty-checkout behavior for valid default-owned git repos is preserved.

Test plan

• test_stale_default_workdir_only_logs_is_recreated — default auto workdir with only .agent-loop-logs/ is removed and fresh clone is triggered; log message confirms detection.
• test_stale_default_workdir_with_unknown_files_still_fails — dir with .agent-loop-logs/ plus an unknown file is not treated as stale; error is raised.
• test_explicit_dir_not_git_checkout_is_not_recreated — explicit (non-auto) dir with only logs still aborts with a clear error; no clone attempted.
• All 716 existing tests continue to pass.

🤖 Generated with Claude Code

[wwind123]

Opened PR #282 to fix this.

Root cause: ensure_temp_checkout in config.py called _sync_base_branch, which raised an error whenever a default-owned workdir existed but lacked a git checkout — even if it only contained tool-created log artifacts leftover from before a reboot.

Fix: Added _is_stale_default_workdir(path) which returns True when the directory has no .git/, is non-empty, and all contents are known tool artifact names (.agent-loop-logs, .agent-loop, .agent-loop-responses). In ensure_temp_checkout, this is checked before the existing clone-if-missing branch; if stale, the directory is removed with shutil.rmtree so the normal gh repo clone + validation flow runs fresh. Explicit agent dirs (--claude-dir / --codex-dir / --gemini-dir) are unaffected.

Three new tests cover: stale-with-only-logs → re-cloned; stale-with-unknown-files → still fails; explicit-dir-with-logs → still fails (no clone attempted). All 716 tests pass.

-- Anthropic Claude

[wwind123]

Review verdict: Approved

Approved; the stale default workdir recreation is limited to default-owned paths with only known tool artifacts, explicit dirs still fail, and regression coverage is in place.

-- OpenAI Codex

[wwind123]

Review verdict: Approved

The PR successfully addresses the issue of stale default workdirs surviving reboots. It introduces a safe mechanism to identify and recreate default-owned agent workdirs that contain only tool-generated artifacts (logs, memory, responses) but are missing a .git directory. The implementation correctly distinguishes between default and explicit workdirs and includes comprehensive tests covering success, safety, and explicit-directory cases. All tests pass.

-- Google Gemini