# Review: PR #48 (fix/issue-45) @ 35935f9

- Role: Reviewer (grok)
- Identity: loop `loop-20261005T064728Z-a4e9c8`, base `c5c7eff6f758c76ac37a138dbc3c301f3f7e10cf` (main, v4.8.4), head `35935f9e8a052ca9e9d9b806a62d17eb72afea78`, tree `dff9db13f3e018528d04a024acda942f501513ff`, 10 commits, clean snapshot
- Verified read-only at `/var/folders/q_/68sxgw3j073dgjg8wl9s08vw0000gn/T/mam-review/loop-20261005T064728Z-a4e9c8/1`. Main stayed `c5c7eff`. No product edit, no commit, no push.
- Diff vs base: 25 files, `tests/test_issue10_13_purge.py` absent. Interpreter for the required pytest: `pr-48/.venv` Python 3.9.6. The shim's `python3` is `/usr/bin/python3`, also 3.9.6. CI (`deploy/gitea-ci.yml:94`) is Python 3.10. `tomllib` is stdlib only from 3.11. `requirements-dev.txt` is `pytest>=8.0`. `/usr/bin/python3` has no `tomli`. The venv has `tomli` and no `tomllib`.

The required command exits at collection. With `--continue-on-collection-errors` added only so the rest of the list could run: **5 failed, 192 passed, 1 skipped, 1 error** in 254.58s (`--basetemp=/tmp/667follow.YrMavO`). The first command, as written in the brief, was **1 collection error in 0.22s, exit 2**, and ran no test.

## What fails

### 1. The merge tests do not load on the Python this repo runs

`tests/test_issue45_herdr_config.py:16` is a top-level `import tomllib`. Collection raises `ModuleNotFoundError`. `tests/test_issue9_server_env.py:157` imports it inside `test_bootstrap_hands_the_server_a_no_restore_config` and fails the same way.

`lib_py/herdr_config.py:54-57` catches that import and sets `tomllib = None`. `server_config_text` then returns `MINIMAL` plus a warning (`:146-147`). A throwaway `HOME=/tmp/m45probe-min` run of `/usr/bin/python3 -m lib_py.herdr_config server-config` printed that warning and wrote only:

```toml
# generated by multi-agent-mux (Issue #45) - do not edit
# source: /tmp/m45probe-min/home/.config/herdr/config.toml
[session]
resume_agents_on_restore = false
```

The throwaway user file's checksum was unchanged. That is the accepted "no parser, restore still off" path. It is also the path the shim takes on this host and on CI 3.10.

`VERSIONS.md:49` and `multi-agent-mux-create/SKILL.md:120` describe the generated file as the user's config with that one key forced false. `test_issue9_server_env.py:171` asserts the merged document (`ui.confirm_close` kept). On this `python3` the file is `MINIMAL`, so fixing only the test import still fails `:171`.

Alternative: in `herdr_config.py:54-57` and in both test imports, use `tomli` when `tomllib` is missing (`tomli.loads` / `tomli.TOMLDecodeError` match). Add `tomli` to `requirements-dev.txt`, and install it for the `python3` that `lib.sh:241` actually runs (`/usr/bin/python3` here, not the venv). Keep the `MINIMAL` warning only when neither library imports. With `tomli` assigned onto the module, the merge kept inline tables, dotted keys, quoted keys, arrays, and comments semantically, and a broken document or a non-table `session` fell back to `MINIMAL`. `herdr config check` on `confirm_close = "no"` in a throwaway file exited 1 with `parse error` and `using defaults`, and `herdr_rejects` (`:169-180`) reported it.

### 2. The #45 positive control never runs on macOS

`tests/test_issue45_restore.py:227-228` reads `/proc/<pid>/environ`. Both parametrizations of `test_a_mam_server_does_not_relaunch_saved_agents` raised `FileNotFoundError` there. The asserts above that line had already passed: the fake claude did not log `--resume`, the restored pane had no agent, and the foreground pid was the shell. The positive control at `:230-236` (plain server must relaunch) did not execute.

Alternative: build the `HERDR_CONFIG_PATH=...` needle, read `/proc/<pid>/environ` when that file exists, and otherwise use `ps eww -p <pid> -o command=` and search the raw output. Run `:230-236` even when the environ read fails, so a missing `/proc` cannot skip the control that shows restore-on still relaunches.

### 3. Restored cwd is the resolved path

`tests/test_issue45_restore.py:263`:

```text
AssertionError: assert '/private/tmp/m45swco8dxx' == '/tmp/m45swco8dxx'
```

`mkdtemp(dir="/tmp")` returns `/tmp/...`. On this Mac `/tmp` is `/private/tmp`, and herdr's pane `cwd` is the resolved path. The close calls start at `:267`, so the selective-close asserts (live agent Z, busy `sleep` X2, user pane U, close only X and Y) did not run. `test_b_resume_leaves_no_restored_shell_behind` got create rc 0 and `status == running`, then `StopIteration` at `:325` on `p.get("cwd") == str(ws)` with that same unresolved `/tmp` path.

`test_b_default_server_is_never_touched` passed: `HERDR_SESSION_NAME=default` returns before invoking `--herdr /nonexistent/herdr` (`herdr_restore.py:54-55`).

Alternative: compare `Path(cwd).resolve()` with `Path(rig.root).resolve()` at `:263`, and `Path(p["cwd"]).resolve() == Path(ws).resolve()` at `:325`.

## What holds

1. User file bytes are not the output. `output_path` (`herdr_config.py:192-205`) skips the source (including `samefile`) and a file whose marker source is not this source. Both names taken raises `OSError`; the fallback directory goes through the same helper (`:270-276`). Writes are `mkstemp` + `os.replace` (`:208-216`). A generated `HERDR_CONFIG_PATH` is followed one hop via the `# source:` line (`:79-87`). The trace beside the user config is `mam-server-config.toml` (or `mam-server-config-<sha256(realpath)[:12]>.toml`), marker plus `# source: <path>`. A `nan` value makes `tomllib.loads(text) == expected` false (`:159`, `nan != nan`) and falls back to `MINIMAL` with a warning. The user file stays.

2. The only server spawn is the shim heredoc `lib.sh:212-261` (`$REAL_HERDR --session <name> server`). create (`create_session.sh:27` sources `lib.sh`, `:264` `_herdr new-session`) and resume (`resume_session.sh:8`, `:245`) go through that shim. The bootstrap runs only when the name is not `default`, `MAM_NO_AUTOSTART` is not `1`, and `session list` says the server is down (`lib.sh:193-212`). `status.sh:13` sets `MAM_NO_AUTOSTART=1` before sourcing. `reconcile.sh:28` sets it before herdr calls; the bootstrap is in the shim, not at source time. An already-running server is left as it was. `create/SKILL.md:120-121` and `VERSIONS.md:50` say the file is rebuilt the next time MAM starts the server, and a server the user starts with `herdr session attach` does not get it.

3. A pane is closed only when the session name is set and not `default`, the pane has no live agent, `agent_session.value` equals the id being resumed, and `shell_pid` equals `foreground_process_group_id` (`herdr_restore.py:54-65`). Empty id and `default` return `[]` before herdr. `main` maps `OSError`, `SubprocessError`, and `ValueError` to exit 1 (`:81-83`). `resume_session.sh:240-242` calls that and on failure only prints a warning; dry-run has already exited at `:225-228`. Bash does not decide which pane. There is a gap between `process-info` and `pane close` in which a command can start; a failed query still closes nothing further.

5. `MAM_VERSION` and `FRAMEWORK_VERSION` are `4.8.5`. The eight `SKILL.md` frontmatters match. `tests/data/skill_code_blocks.json` changes line numbers and the resume script anchors (`:237/:241/:247-250` → `:245/:249/:255-258`). No class, kind, or sha field changes. `test_version_consistency` and `test_issue42_skill_blocks` are in the 192 passes. S11's base moved to `811b080` (`tests/test_issue23_24_startup_path.py`) because the heredoc change is intentional.

6. `tests/test_issue10_13_purge.py` is not in the diff. Those claimed full-suite failures are not produced by this PR. This review did not run that file.

7. New judgment is in `herdr_config.py` and `herdr_restore.py`. The new bash is "config path empty → warning, start anyway" (`lib.sh:243-245`) and the resume warning above. The spawn snippet copies `MAM_SERVER_CFG` into the child as `HERDR_CONFIG_PATH` and drops `MAM_SERVER_CFG` from the child (`lib.sh:253-256`).

## Isolation

Restore tests use `mkdtemp(prefix="m45", dir="/tmp")`, session `mamt45-<pid>-<uuid4>`, and strip `HERDR_WORKSPACE_ID`, `HERDR_PANE_ID`, `HERDR_TAB_ID`, `HERDR_SOCKET_PATH`, `HERDR_SESSION`, `HERDR_ENV`, `HERDR_CONFIG_PATH`, and `XDG_CONFIG_HOME` (`test_issue45_restore.py:41-42`, `:160-164`). After the run, no `mamt45` process remained. `~/.config/herdr/config.toml` mtime stayed `1784456394`, the directory mtime stayed `1791183576`, and no `mam-server-config.toml` appeared there. The live `multi-agent-mux` server log grew on its own during the window. Snapshot `git status` was empty afterward. Throwaway pytest dirs were removed.

[VERDICT: NOT PASS]