[dsyme]

Summary

The fortnight assessment found successful runs skipping required investigations, losing blocker memory during schema recovery, and describing queued publication as completed. Make those outcomes harder to hide behind a green run.

• Gate normal-mode noops against the actual task selection and recorded successful shell commands, with concrete per-task file/API evidence. Missing required work must be reported incomplete.
• Require monthly reporting for substantive public work, including labels; reject mutation-plus-noop batches and reporting-only bookkeeping used to bypass assessment.
• Expose the exact memory schema and useful field-level errors; require sequential, backed-up edits and validation without clearing useful state.
• Use the pre-authenticated read-only gh proxy, excluding bookkeeping issues from task weights.
• Clarify information-only publication intent, fresh human feedback, actual test evidence, complete PR-head delta/bundle preflight, and queued-versus-confirmed publication.
• Document behavior and add regression scenarios exercising the executable guards.

The gate verifies recorded assessment coverage, not the semantic quality of every conclusion or a requirement to manufacture PRs. Explicit commands and intentional stop requests remain exempt from normal-mode reporting/assessment. Existing scheduled PR capacity protections remain in place.

Validation

• PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s tests -p 'test_repo_assist_*.py': 55 tests pass.
• gh aw compile workflows/repo-assist.md --strict --no-check-update: succeeds with the memory-validator review warning described below.
• git diff --check: passes.
• The generated shared sample lock remains intentionally gitignored; this PR does not change deployed workflow locks or enable/dispatch any workflows.
• No live agent trial was dispatched. Reassess real run compliance after merge and deployment.

Security/update review

• No new secrets, write permissions, credential scopes, or redirect destinations.
• The new task-selection artifact uses the existing pinned upload/download-artifact actions and contains deterministic task-selection data. The outcome gate parses artifacts as data; it does not execute agent-authored code.
• The gh proxy remains read-only, replacing the broad local MCP tool inventory and its unauthenticated-CLI/fork-enrichment failure paths. All GitHub writes still use safe outputs.
• Protected-file handling for existing-PR updates is tightened from allowed to request-review. The explicit 4096 KiB cap matches the existing default; no publication protection is relaxed.
• The compiler flags the changed memory validation script for review. Those edits improve error messages only; accepted schema and validator capabilities are unchanged. Reviewed without --approve.

Repository-specific Fantomas startup and calendar-month scheduling are fixed separately in fsprojects/fantomas#3574. Other repository schedules and intentional disablements are unchanged.