multi-agent-mux/.agents/reports/reviewer-agy-01/report-db46a0da.md

Pre-Merge Independent Review Report: PR #37 (fix/issue-36)

• Job ID: db46a0da
• Target Agent: agy (Session: herdr:reviewer-agy-01)
• Role: Reviewer
• Target Worktree Context:
  • Loop ID: loop-20260924T080014Z-2b3df0
  • Base Commit: 1debc7d92444385b9200e821efa0ff0f6d82023e (v4.7.2)
  • Head Commit: 59be1c89209a232059aec0c5aeecc2329929cb49 (6 commits: +765 / −30)
  • Tree SHA: a009c9062d9818f343ac57e6c72cda932887fab7
  • Verification Path: /var/folders/q_/68sxgw3j073dgjg8wl9s08vw0000gn/T/mam-review/loop-20260924T080014Z-2b3df0/1

---

1. Executive Summary

PR #37 resolves Issue #36 ("위임 프롬프트를 argv 가 아닌 파일로 전달하고, 위임 실패를 종료 코드 1 로 보고", Closes #36) across 6 commits and 11 files (+765 / −30):

• 59be1c8 fix(issue-36): create_session hands the agent a brief path, not the task; a prompt file that cannot be written is removed
• `f26952c fix(issue-36): the legacy shell loop and lib.sh's submit helpers pass prompts by file
• 14d11bf fix(issue-36): a job that cannot be registered after a rebuttal ends the loop with rc 1
• f5cadd9 test(issue-36): prompts over the argv cap go through the real wrapper, registry and loop; a failed delegation ends the loop with rc 1 and no traceback
• c77806b fix(issue-36): the loop passes every delegated prompt as a file; a wrapper that cannot start is a failed submit, on the first try and on the retry
• 68a5d31 fix(issue-36): registry and wrapper accept --prompt-file; the wrapper hands the same path on

The core architecture correctly eliminates the 131,072-byte MAX_ARG_STRLEN barrier on Linux by passing prompts via temporary files (--prompt-file) across the Python loop runner, delegate-job wrapper, registry.py, lib.sh, and create_session.sh. All legacy CLI contracts and --prompt backward compatibility remain intact.

However, during execution in the verification environment on macOS, test test_s9_shell_loop_removes_a_partial_prompt_file_and_says_why in tests/test_issue36_prompt_transport.py failed with an AssertionError at line 602 (assert "File too large" in result.stderr). This is a platform-dependent test expectation: GNU bash / glibc on Linux prints write error: File too large when ulimit -f is exceeded, whereas macOS /bin/bash (3.2.57) does not emit that string. Because the test suite must pass cleanly across supported platforms (macOS/Linux), a verdict of NOT PASS is issued.

---

2. Blocking Defect Details

[Defect 1] Platform-Specific Assertion Failure on macOS in test_s9

• File & Line: tests/test_issue36_prompt_transport.py:602
• Code:

  598:     harness.write_text('set -euo pipefail\nREPO_ROOT="$PWD"\nMAM_REAL_ROOT="$PWD"\n'
  599:                        'log_error() { printf "%s\\n" "$1"; }\n' + function +
  600:                        '\ntrap "" XFSZ\nulimit -f 1\n'
  601:                        'output=$(delegate_job_safe submit --prompt "$(cat input.md)")\n')
  602:     result = subprocess.run(["bash", str(harness)], cwd=tmp_path, capture_output=True, text=True, timeout=60)
  603:     assert result.returncode != 0
  604:     assert "File too large" in result.stderr
  605:     leftovers = list((tmp_path / ".mam").glob("prompt-shloop-*"))
  606:     assert not leftovers, [(p.name, p.stat().st_size) for p in leftovers]
  607:     assert "delegate_job_safe: cannot write the prompt file" in result.stderr
  

• Error Trace:

  FAILED tests/test_issue36_prompt_transport.py::test_s9_shell_loop_removes_a_partial_prompt_file_and_says_why
  AssertionError: assert 'File too large' in 'xxxx...xxxxdelegate_job_safe: cannot write the prompt file under /private/tmp/.../.mam\n'
  

• Root Cause: When ulimit -f 1 is exceeded during printf '%s' "$2" > "$prompt_file":
  • On Linux (GNU bash 5.x / glibc), the shell writes bash: line X: printf: write error: File too large to stderr.
  • On macOS (Darwin / bash 3.2.57), the redirection write fails and printf exits non-zero without printing the exact literal "File too large".
  • The actual implementation behavior is completely correct on macOS: result.returncode != 0 (Line 603 passes), partial files are cleaned up (assert not leftovers Line 606 passes), and MAM's error message delegate_job_safe: cannot write the prompt file is printed to stderr (Line 607 passes).
  • Line 604 strictly demands "File too large" unconditionally, causing the test to fail on macOS.
• Remediation: Guard the assertion with sys.platform == "linux", or allow MAM's error string to satisfy the failure check:

  if sys.platform == "linux":
      assert "File too large" in result.stderr
  

  or

  assert ("File too large" in result.stderr or "cannot write the prompt file" in result.stderr)
  

---

3. Detailed Verification of Other Checklist Items

(1) Completeness of Prompt Transport Pathways

• No Residual Raw --prompt Exec Invocations:
  • lib_py/loop/runner.py: _run_cmd creates a temporary file under self.mam_dir (prompt-{loop_id}-*.md) and invokes the wrapper with --prompt-file <path>. The temporary file is safely cleaned up in a finally: block.
  • lib_py/delegation/registry.py: Supports mutually exclusive --prompt and --prompt-file in both register and update commands. Evaluated via _prompt_from_args() during CLI parsing before writing to registry JSON.
  • lib_py/delegation/cli.py: Supports --prompt-file, forwarding to registry.py.
  • lib.sh: delegate_submit_job() and _delegate_submit_job_legacy() both utilize _delegate_prompt_file and clean up via rm -f "$prompt_file".
  • multi-agent-mux-delegate-job: Iteration brief provisions write to $job_dir/prompt.md and invoke registry.py update --prompt-file "$job_dir/prompt.md".
  • run_loop.sh: delegate_job_safe intercepts --prompt and converts it into --prompt-file "$prompt_file", unlinking it upon return.
  • create_session.sh: Injected pane text no longer includes the full prompt text; instead, writes $SUBMIT_JOB_PROMPT to $task_brief (.mam/jobs/<id>/brief.md) and passes: Task: read the complete task in $task_brief and follow it..
  • Relative paths passed to --prompt-file in cmd_submit are converted to absolute paths against caller pwd before changing directory to WORKDIR (multi-agent-mux-delegate-job:384), preventing directory-shift bugs.

(2) Error and Failure Handling

• OSError Handling on Wrapper Execution:
  • In runner.py:

    except OSError as e:
        msg = (f"cannot start the wrapper: {e} "
               f"(prompt {len(p_text.encode('utf-8'))} bytes, wrapper {wrapper_script})")
        return subprocess.CompletedProcess(cmd, 126, "", msg)
    

    Wraps execution failure cleanly as returncode 126. The loop logs the failure and aborts with rc 1 without tracebacks (test_s3b, test_s4).
• Fail-Closed Unregistered / Failed Jobs:
  • Re-adjudication failure (if not adj_jid:): terminates with error and rc 1.
  • Arbitration failure (if not arb_jid:): terminates with error and rc 1.
  • Compliance failure (if not comp_jid or not self.wait_for_job(comp_jid):): terminates with error and rc 1.
  • Changing compliance failure from continue to return 1 is an intentional, sound hardening: if the compliance job fails or times out after an overruled rebuttal, the Creator did not implement the required changes; proceeding to review un-updated code would be erroneous.

(3) Evidence vs. Delivery Sizing (MAM_DIFF_MAX_BYTES / MAM_DIFF_MAX_LINES)

• In diff_collect.sh: Comments clarify that MAM_DIFF_MAX_BYTES (default 200,000) and MAM_DIFF_MAX_LINES (default 4000) bound the evidence shown to reviewers, while delivery itself is unbound due to --prompt-file.
• test_s5_a_diff_in_the_former_blind_spot_is_delivered_whole confirms a 150,000-byte diff (between 131,072 and 200,000 bytes) is not truncated and delivered intact.

(4) create_session.sh Onboarding & Brief Alignment

• Task text is written to .mam/jobs/<id>/brief.md.
• Injected instruction provides the absolute path to brief.md.
• Handshake protocol (delegate_publish_cmd, --event started, token verification) remains 100% compliant with Issue #23. Verified by test_issue23_24_startup_path.py (19/19 tests passed).

(5) Existing Test Contracts & Backwards Compatibility

• test_issue5_phase1_delegation.py: 66 / 66 passed.
• test_version_consistency.py: 3 / 3 passed.
• test_loop_cli.py: 9 / 9 passed.
• test_issue14_delivery_ack.py: 12 / 12 passed.
• test_issue21_phase2_loop_a.py: 29 / 29 passed.
• Existing --prompt syntax remains fully supported for backward compatibility across all CLIs.

---

4. Test Execution Summary

Test Suite	Result	Duration	Notes
tests/test_issue36_prompt_transport.py	20 PASS, 1 FAIL, 1 SKIP	39.9s	test_s9 failed on line 602 (File too large assertion on macOS); test_s1b skipped (linux_only)
tests/test_issue5_phase1_delegation.py	66 / 66 PASS	12.1s	Registry CLI, backward compatibility with --prompt
tests/test_issue23_24_startup_path.py (S7/S8)	19 / 19 PASS	63.1s	create_session.sh brief task injection & handshake
tests/test_loop_cli.py	9 / 9 PASS	39.2s	Loop CLI regressions
tests/test_issue14_delivery_ack.py	12 / 12 PASS	98.2s	Delivery acknowledgement contracts
tests/test_issue21_phase2_loop_a.py	29 / 29 PASS	155.3s	Isolated worktree loop review regressions
tests/test_version_consistency.py	3 / 3 PASS	0.3s	Version consistency checks
git diff 1debc7d...HEAD --check	0 errors	0.05s	Clean formatting and no whitespace errors

---

[VERDICT: NOT PASS]