Survey venus for shared-mutable-file contention between agents (incl. CLAUDE.md, commons .md, pm/)
HOST: venus. Elazar GO 2026-08-31, child of OPS-156. SURVEY YOUR OWN HOST for the contention pattern OPS-156 describes, then FILE YOUR FINDINGS AS CHILD ITEMS UNDER OPS-156 (one child per distinct contended surface, not one child for the whole survey) so each gets tracked and closed on its own. WHAT TO LOOK FOR: any file or directory that more than one agent writes, with no isolation between them. The signal is agents coordinating ABOUT the file instead of the file being safe to edit - DMs saying "do not touch X while I am in it", "pull before you edit", "I left changes in Y", stray uncommitted entries belonging to another lane, or a documented convention that exists only to stop two agents colliding. SURFACES EXPLICITLY IN SCOPE, including pure-.md ones - a docs collision is the same defect and costs the same coordination: - global CLAUDE.md and the ~/.claude/*.md rule files (note: on venus these are TRACKED SYMLINKS into sh.git, so editing one is editing the shared repo) - per-project commons/shared .md, design docs, guides - pm/ directories and PM-owned planning files - any shared checkout, shared state dir, shared sqlite, or shared config that two agents write - shared systemd unit dirs and .llmmsg-env style overlays FOR EACH FINDING RECORD: the path; which agents write it; how collisions are currently prevented (convention, DM, lock, luck); whether a collision has actually occurred and the evidence; and whether the fix is isolation (separate trees/files per agent), single-ownership (one named writer, everyone else read-only), or generation (the file is built from a source, not hand-edited). A NIL RETURN IS A RESULT, BUT IT MUST BE EVIDENCED. "No contention on this host" needs to say what you looked at and what you did NOT exercise. Absence of a reported collision is not absence of contention - most of these are invisible until two agents happen to write in the same minute. Do not fix anything under this item without saying so on its child first; some of these have an owner who has not been asked.
Sub-tickets — 0/3 done
Activity
-
parent=#3095
-
Filed 3 children: OPS-163 (~/.codex/config.toml, confirmed collision, no git tracking, 10 live -ca agents), OPS-164 (maintainer-agents.sqlite, convention-only write protection, no incident evidenced), OPS-165 (/opt/llmmsg-srv, same shape as sh.git but already mitigated via strict+hunk-gate, out of my maintenance boundary, no fix taken). Also checked: ~/.claude/*.md rule-file symlinks into sh.git confirmed true on venus (verified per instruction, not assumed) — not filed separately, subsumed by OPS-157/158 since it's the same tree. NOT exercised: exhaustive pm/ dir + per-project commons .md sweep, jobreg.tsv's own collision history beyond the sh.git parent case, SQL write-pattern audit of maintainer-agents.sqlite's 6+ consumer scripts. No fixes taken on any finding per instruction.