multi-agent-mux/.agents/reports/planner-reviewer-codex-01/report-70d302c2.md

Issue #17 follow-ups — Codex review

Job: 70d302c2 Loop: loop-20261007T224638Z-cf1efb Base: 56d0bfda54 Head: 8241bd53b1 Tree: efef640d36 Validation path: /var/folders/q_/68sxgw3j073dgjg8wl9s08vw0000gn/T/mam-review/loop-20261007T224638Z-cf1efb/1

Finding R1 — subscriber-up decision failures can still report successful submission (P2)

Evidence: .agents/skills/multi-agent-mux-delegate-job/multi-agent-mux-delegate-job:384-395 (direct), :556-567 (loop/discuss); .agents/skills/lib_py/delegation/wrapper.py:332-333.

The new failure handling only protects printed answers. An execution failure in the subscriber-up predicate is consumed as a normal false answer. If the subscriber has already exited 0, the error branch maps that code to completed, finalize-begin keeps it, and _exit_for exits 0. This violates N2's requirement that a failed decision call ends the wrapper with 1, never 0.

I read the exclusion rationale in commit a670ef5: it states that subscriber-up failure treats the subscriber as dead and reaches a fail-closed sub-status/exit-code path. Those later calls can succeed and report completed, so the rationale does not cover a failure confined to the earlier process invocation. This is the same per-decision failure-injection model used by the added regression tests, not an assumption that every later Python invocation must also fail.

Reproduction (no broker or agent session):

1. Use tests/test_issue17_item4_decisions.py::_failing_py with decision="subscriber-up", mode="error" (or "gone").
2. Set SUB_STANDIN=1 and SUB_RC=0. The shim makes only job_subscriber.py return 0 and only the targeted predicate fail; other Python calls remain real.
3. Call _human_submit with a temporary HOME/workspace/registry and that shim as $PY.
4. Expected: submit exits 1 with a decision-failure diagnostic. Actual: module traceback rc 1 -> submit rc 0; missing-interpreter rc 127 -> submit rc 0. The wrapper prints "subscriber died early (... exit=0 ...)" and a log path, then returns success.

Scripts: /tmp/70d302c2-probes.py and /tmp/70d302c2-predicate-errors.py. All created job records are in temporary probe directories, not the real registry.

Fix direction: distinguish a successfully computed boolean from inability to execute the decision. A printed boolean through the existing _decided machinery (possibly a new flag/subcommand retaining the old CLI contract) can return execution status 0 for both valid answers while preserving the normal true/false branches in Bash. Treat nonzero execution status or missing/invalid output as _decision_failed. Do not simply fail on every current rc 1, since false legitimately uses that code. At minimum, rc values outside 0/1 must never be silently treated as false, but that alone cannot fix the reproduced import-error rc 1 case. Add tests for both true and false predicates, plus selectively failing subscriber-up with subscriber rc 0, in both direct and loop/discuss call sites.

The same ambiguity exists at :718 for stall-watch: a probe returning 127 yielded caller rc 0 and continued execution ("survived"), silently disabling the watcher. The creator explicitly describes the watcher as optional/notify-only; this observation is related evidence, while the submission-success counterexample above is sufficient for NOT PASS.

Other requirements

• Restore premise: tests/test_issue45_restore.py now accepts HOME or the original removed directory. The source docstring distinguishes herdr 0.9.1/0.9.3. The additional production handling for restore_error + pane_not_found is explained by commit f19f940: merely relaxing cwd exposed a placeholder with no process, so closing previously raised before reaching other candidates. The matching session, no-live-agent, non-default-server and normal foreground-shell filters remain.
• Restore regression negative control: executed the new placeholder unit-test function against the base herdr_restore implementation using a stubbed _herdr; it raises CalledProcessError at base and passes at head.
• Printed decisions: the introduced _decided/_exit_for checks cover exit-code, sub-status, finalize-begin, wait-rc and finalize-end process failures. Allowing empty finalize-end output is intentional for OwnerGone, whose CLI handler returns normally without a printed decision. Bash 3.2 dynamic scoping makes printf -v assign the intended caller variables.
• Exit-code negative control: the exact base wrapper placed in a temporary dependency layout with the same failure shim exits 255, while head exits 1. This confirms the added failure test actually detects the original bug.
• N1 accurately documents the package import path to paho; execution code did not change.
• N3 accurately documents PATH python3 and missing-interpreter behavior; executable code did not change.
• Six item-scoped commits, seven changed files, no version bump or VERSIONS entry. No further code defect was identified in the reviewed normal-path handling.

Verification

• Python 3.9 from the designated issue17-followups/.venv; native /bin/bash 3.2.57.
• With --noconftest, cache disabled, external basetemp and per-test temporary WORKSPACE_ROOT: tests/test_issue17_item4_decisions.py + tests/test_issue45_restore_unit.py -> 281 passed, 4 failed in 292.91s. All newly added failure-handling and restore unit tests passed. Log: /tmp/70d302c2-selected.log.
• The four failures are existing zombie-owner tests: test_u_an_exited_unreaped_owner_is_not_alive, both test_c_finalize_ends_with_an_unreaped_wrapper parameters, and test_c_an_unreaped_wrapper_leaves_the_real_record_alone. A separate owned-child probe confirmed ps execution raises PermissionError(1, Operation not permitted), causing the existing _alive fallback to return True for an unreaped exited child. The wrapper.py executable AST is identical to base (only its module docstring changed). These sandbox-caused failures are not a new finding against this PR.
• Budget tests with --noconftest and selection "not test_b3 and not test_r3 and not test_b7": 13 passed, 7 deselected in 28.88s. Log: /tmp/70d302c2-budget.log. This exercises budget propagation, child wait caps, remote deadline labels and closed-unconfirmed classification without needing a reachable broker.
• Explicit loopback bind probe failed with PermissionError(1, Operation not permitted). Consequently no full broker suite result is claimed. Real-herdr restore integration was not run in this sandbox; the orchestrator retains that validation responsibility. Static inspection of its rig confirms disposable HOME/HOME_DIR, mamt45-* session names, scrubbed inherited server selectors and fake agent commands, rather than the user's server/config.
• Both shell files pass /bin/bash -n. All five changed Python files pass AST parsing. git diff --check BASE HEAD passes.
• HEAD and tree match the brief; final git status --porcelain is empty.
• Independent probes reproduced subscriber-up failures returning submit 0 for both module-error rc 1 and interpreter-error rc 127, plus swallowed stall-watch rc 127.
• Negative controls confirmed exit-code failure base 255 -> head 1 and placeholder regression base exception -> head pass.
• The environment restrictions alone do not determine the verdict. R1 is an independently reproduced N2 requirement violation.

Isolation and lifecycle

All verification ran from the specified review tree. Inherited HERDR_, MAM_, MQTT_, SKS_ and control-plane bindings were cleared; temporary HOME/HOME_DIR/WORKSPACE_ROOT, XDG_CONFIG_HOME and an absent MAM_ENV_FILE were supplied. Tests sourcing lib.sh received a per-test temporary WORKSPACE_ROOT via an environment-only pytest plugin. Direct probes used synthetic jobs and temporary registries. No .mam.env was exported, no live herdr or agent session was used, and no creator/source files, commits or branches were changed. Base comparison artifacts and logs were outside the review tree.

The requested started publication was attempted first; MQTT exhausted three attempts with Operation not permitted and exited 2. The requested completed command was also executed after saving the report; all three MQTT attempts failed with Operation not permitted and it exited 2. The report exists, but broker delivery of started/completed is not confirmed.

[VERDICT: NOT PASS]