multi-agent-mux/.agents/reports/reviewer-grok-01/report-d6ae03a9.md

Review d6ae03a9 — PR #58, second pass on the four Mac findings

Reviewed only the snapshot /var/folders/q_/68sxgw3j073dgjg8wl9s08vw0000gn/T/mam-review/loop-20261015T123502Z-1b9df6/1 (loop-20261015T123502Z-1b9df6). Head a7cb90e63abf454ccf22f6ff887cc25393755402, tree 96c5c5f82cb89e4383011d52fb1dcb21d4e787e6, base e32c9574644a9185813a4dc881812694d62b6204 is an ancestor. Fifteen commits, 23 files, +908/−131. The four commits after the first-round head 63d8164 are 75c0042, 151af99, 1c598bb, and a7cb90e. git diff --check is clean. Snapshot porcelain is empty and the snapshot has no .mam. Control-plane main is still e32c9574644a9185813a4dc881812694d62b6204, porcelain empty. No product file was edited.

Host tools: /usr/bin/python3 3.9.6 has no paho. The control-plane venv is also 3.9.6 and does have paho. /bin/bash is 3.2.57. The changed modules parse with ast on 3.9.6. bash -n accepts the wrapper.

F1 — path rules import without paho

lib_py/delegation/logroot.py is standard library only. sibling_logs_dir, get_logs_dir, and default_registry_dir live there. mqtt_common.py:33-40 imports paho inside try/except and re-exports the logroot names, so mqtt_common.sibling_logs_dir is the same function. publisher.py:17-20 does the same for its own paho import. control_plane.py:28 and both status.sh copies (status.sh:171, and the text-mode copy) import sibling_logs_dir from logroot, not from mqtt_common.

/usr/bin/python3 imports lib_py.delegation, lib_py.control_plane, and logroot in one process. Sourcing lib.sh with AGENT_PYTHON_BIN=/usr/bin/python3 and mam_control_plane_json on a temporary workspace exits 0 and returns a delegate_job_logs_dir under that workspace's .mam. make_client still raises ImportError when a broker is actually required (mqtt_common.py:286-288).

lib_py.delegation.logroot is not named in eager.CORE_MODULES. Importing lib_py.delegation.mqtt_common, which eager load does, loads logroot as a dependency. On this interpreter that module is in sys.modules after that import.

F2 — logs keeps the workspace default

cmd_logs (multi-agent-mux-delegate-job:1096-1113) passes --registry-dir only when the caller gave it. The empty array uses ${reg[@]+"${reg[@]}"}. On bash 3.2 with set -u, an empty array expands to nothing and a two-element array expands to both words. registry.py then applies its own default: MAM_JOBS_DIR, then WORKSPACE_ROOT, then the cwd.

With WORKSPACE_ROOT set to a temporary workspace and the cwd set to a different directory, registry.py register wrote the job under the workspace .mam/jobs and the audit log under the workspace .mam/delegate_job_logs. logs --list from that other cwd printed the job id and exited 0. The other directory gained no .mam.

R1 — a later dispatch is not the earlier one's success

_check_disk_fallback (job_subscriber.py:121-138) re-reads the registry record after it has read the audit status and before it adopts that status. It adopts only when dispatch_token, attempt_id (falling back to iteration), and transition_seq are unchanged. A changed record leaves the job pending for the next poll, which reads both again. The extra read is one load_job per non-terminal job on the audit path. The token check below still uses the first snapshot, which this comparison has just required to match the second read, and there is no further I/O between them.

The regression test interleaves error, then redispatch to attempt 2, then completed, inside read_logged_status. The caller waiting on attempt 1's token gets an empty terminal set. The next call, waiting on attempt 2's token, gets completed. A third call with attempt 1's token stays pending. On an archive of 63d8164 the same test fails with B's completion was attributed to A.

R2 — a symlinked registry has one audit log

sibling_logs_dir (logroot.py:51) uses os.path.realpath of the registry, then the sibling delegate_job_logs. A relative registry still resolves from the cwd. forwarded_keys (control_plane.py:82-84) compares _abs of both sides, so a /tmp path and its /private/tmp real path still count as the same sibling and the derivable root stays off the pane.

A probe under /tmp (/tmp is a symlink to /private/tmp): from_env with no MAM_JOBS_DIR equals defaults; get_logs_dir() is that same directory; a jobs path given through /tmp is not forwarded; a jobs directory that is itself a symlink to another parent pairs with the target's sibling, and the binding matches that directory and is not forwarded.

submit --dry-run --registry-dir <tmp>/alias/jobs, where that path links to <tmp>/actual/jobs, wrote meta.json under <tmp>/actual/delegate_job_logs and printed that directory. <tmp>/alias/delegate_job_logs was not created, and neither was the workspace audit root. logs --registry-dir of the alias printed registered.

Tests

Pytest used the control-plane venv, PYTHONDONTWRITEBYTECODE=1, -p no:cacheprovider, and --basetemp /tmp/mamd6-d6ae03a9-pytest. Inherited HERDR_*, MAM_*, MQTT_*, SKS_*, WORKSPACE_ROOT, DELEGATE_JOB_LOGS_DIR, and PYTHONPATH were dropped. AGENT_PYTHON_BIN pointed at that venv.

On the snapshot, 25 tests passed in 50.76s: all of tests/test_issue57_logs_follow_registry.py, including the four new repros (control-plane without paho, status json and text without paho, logs from another cwd, the interleaved redispatch, and the symlinked registry), plus test_t3_0_control_plane_module, test_logs_dir_follows_the_same_order, and test_b9_logs_dir_follows_cwd_changes.

The six new test cases copied onto an archive of 63d8164 failed in 2.55s: the control-plane import traceback, both status modes, logs reading <cwd>/.mam/delegate_job_logs, the redispatch attributed to the earlier attempt, and ImportError for logroot. Scratch directories were removed. Snapshot porcelain stayed empty.

The full suite was left to the orchestrator. SKILL.md's Audit Logs sentence still names .mam/delegate_job_logs/<job_id>/, which is the directory this rule produces when the registry is .mam/jobs. The design-doc status row now lists the sibling lookup (docs/design/worktree-creator-model.md:128).

The pane herdr config file is newer than the mtime seen at the end of the previous review. This review did not start herdr and did not write that file. ~/.config/herdr/config.toml is unchanged. That mtime is not part of the verdict.

[VERDICT: PASS]