multi-agent-mux/.agents/reports/reviewer-grok-01/report-c564d29d.md

Review: PR #46 round 2 @ 17eba09

• Role: Reviewer (grok)
• Identity: base 8d26106195a0a24621f140e76389d206332f84c0, head 17eba0904ebbf54c1d2243ad61a6f2ba17263889, tree a0f1ffede44be7c86267a56b25871e1d081feafb, 1 commit, clean worktree
• Verified read-only at /var/folders/q_/68sxgw3j073dgjg8wl9s08vw0000gn/T/mam-review/loop-20261003T230658Z-e7866a/1. Main stayed at 45d1621. No product file was edited.
• git diff --stat 8d26106..HEAD: 5 files, +93/−19. git diff --check is clean.

The commit is the three skill docs (create, resume, stop), tests/data/skill_code_blocks.json, and tests/test_issue42_skill_blocks.py. No script, no Python product module, no version line. Frontmatter stays 4.8.2.

Round 1 on 8d26106 was PASS from this reviewer. Codex R1–R3 are the items this commit answers. They are closed.

R1 — attach names the server

resume_session.sh is run as a child (SKILL.md:60). Its export HERDR_SESSION_NAME (resume_session.sh:137) ends with that process. Native herdr --help documents --session <name> as "Use or create a named persistent session" and lists one env var, HERDR_CONFIG_PATH. It does not list HERDR_SESSION_NAME. herdr agent attach --help is herdr agent attach <TARGET>. A fake herdr on PATH recorded herdr --session probe-srv agent attach probe-name as those five arguments, and HERDR_SESSION_NAME=probe-srv herdr agent attach probe-name as three arguments with no --session. The shim that does inject the variable is lib.sh:_real_herdr (255-268), and only after lib.sh is sourced: an explicit --session is passed through, otherwise a non-default HERDR_SESSION_NAME is prepended.

The skill now says that in Case 3 (SKILL.md:30), the workflow paragraph (:63), and When NOT (:139): herdr --session <server> agent attach <name>, with <server> the row's herdr_session. The verification fence reads that field and passes it:

SERVER=$(python3 -c "... [0].get('herdr_session') or 'default')")
herdr --session "$SERVER" agent get "$SESSION_NAME"
herdr --session "$SERVER" agent read "$SESSION_NAME" --source visible --lines 30

(SKILL.md:106-126). Against temp YAML, a row with herdr_session: isolated prints isolated. After a successful resume, update_yaml_resumed.sh:154-156 writes herdr_session (and herdr_server) from the server the script used, so this post-resume read is the server the resume ran on.

R2 — --dry-run is not "write nothing"

Both option rows now say: no spawn and no registry (YAML/DB) update; the checks before that exit still run; on macOS they clear the agent binary's quarantine attribute.

• create: xattr -d com.apple.quarantine is at create_session.sh:194-196, after preflight and before the dry-run exit at 280-282. spawn is called at 288, so dry-run neither spawns nor reaches the YAML write.
• resume: the same xattr is at resume_session.sh:167-169. The dry-run that prints the command exits at 225-227, before _herdr new-session (237) and before update_yaml_resumed.sh (247). The Darwin test is uname = Darwin and the path must be a file.

R3 — a deleted body stays deleted

fingerprint (test_issue42_skill_blocks.py:132-134) is sha256 of the body with all whitespace collapsed, first 12 hex digits. test_no_deleted_body_came_back (150-153) compares every current fence with DATA["removed"] and DATA["superseded"] by exact base_sha or that fingerprint. It does not read DATA["blocks"], so listing the returned body as class c does not hide it. test_a_returned_body_is_caught_reindented_and_with_the_inventory_updated (156-166) checks both: the exact sha hits, and a reindented body with a dummy sha still hits.

A different example does not match. The fingerprint is the whole fence, not a substring, so an ordinary edit of another example changes the words and stays off the list. The only edit that fails this test is putting a deleted body back, including reindented. The inventory sha lock still requires a JSON update when any fence's text changes; that is the other test, and it is unchanged.

test_superseded_blocks_are_recorded_and_replaced records the folded create call. Its base_sha ceb96a8c1293 and base_fingerprint b6a8a25279c5 match the fence at 45d1621 create SKILL.md:263-266 (the create_session.sh --workspace --agent --role --session call). The replacement bash .agents/skills/multi-agent-mux-create/scripts/create_session.sh is contained in the Workflow fence. Count: 33 current + 5 removed + 1 superseded = 39, and 39 − 2 fences that were not in the original 37 = 37. Current class counts stay 28 c and 5 b.

Inventory and the other doc lines

Pairing each block with the 8d26106 inventory by skill, note, class, and kind is equal. One span moved: resume verification 103-124 → 103-128, still class b, kind verification. No class or kind flipped.

create exit 42 now includes the live process with MAM_KILL_ON_READY_TIMEOUT=1 (create/SKILL.md:179). That is the else at create_session.sh:392-398: alive and the knob unset preserves the session; otherwise exit 42 and the trap rolls it back.

stop exit 0 now says the already-stopped no-op does not apply to --purge-conversation (stop/SKILL.md:61). The script no-ops only when purge is off (stop_session.sh:184-188). With purge and without --yes it exits 3 (192-198).

Codex C2 and claude O2/O3 are named in the brief as follow-ups and are not in this diff. The exit tables still match the scanners that landed in round 1; this commit does not change those scanners.

Checks

Snapshot, MQTT variables unset, no .mam.env exported, interpreter /Users/godopu16/PuKi/laa/canary_projects/multi-agent-mux-wt/pr-46/.venv/bin/python, --basetemp under /tmp/c564follow.* (removed after the run):

132 passed, 1 skipped in 293.02s — tests/test_issue42_skill_blocks.py, tests/test_tier2_component.py, tests/test_orc_onboard.py, tests/test_loop_cli.py, tests/test_o1_rebuttal.py, tests/test_version_consistency.py.

The skip is test_claude_skills_link_resolves_to_agents_skills: .claude/skills is absent in this worktree snapshot.

Notes that do not change the verdict

• No matching row: the SERVER one-liner raises IndexError and prints nothing. In a shell without set -e, SERVER is empty and the next command is herdr --session "". The block is the check after a resume that wrote the row.
• herdr_session missing, empty, or null prints default, including when the row has herdr_server: isolated. resolve_herdr_session (lib.sh:1601-1604) would use that alias, then the workspace mam-<slug>. A successful resume has already written herdr_session (update_yaml_resumed.sh:154-156), which is the case this fence is for.
• Resume's already-running dry-run exits at resume_session.sh:148-150, before the quarantine xattr. The option row describes the path that resolves the command and exits at 225-227.
• stop/SKILL.md:38 and :70 still say an already-stopped row is always exit 0. The table at :61 and the script at 184-188 keep the purge exception. The state machine at :84 labels the re-stop edge "stop default".

[VERDICT: PASS]