Imported from hamzah2304/messageboardauditbench (AGENTS.md). Install upstream with npx skills add hamzah2304/messageboardauditbench. Copyright stays with the author.

Shared agent working agreement

Codex and Claude Code share this repository with the user, and more than one agent may be working at the same time. CLAUDE.md imports this file so both tools follow the same agreement.

Worktrees

• The primary checkout stays on main. Never switch it to a task branch. Small, stable edits may be made directly on main; anything substantive, multi-file, or likely to overlap with someone else's work goes in a task branch and linked worktree.

• Worktrees live inside the repo under .worktrees/ (gitignored), named <tool>-<task> on branch <tool>/<task>, created from main:

  scripts/worktree_add.sh <task> [claude|codex]   # -> .worktrees/claude-<task> on claude/<task>

  Claude Code's EnterWorktree goes through the same helper via the tracked WorktreeCreate hook in .claude/settings.json. Codex should call the helper rather than a bare git worktree add.

• A worktree contains committed HEAD, not the primary checkout's dirty changes. Read those in place; ask only when incorporating them is a real choice.

• The gitignored inputs and outputs are shared, not copied: the helper symlinks data/, runs/, logs/ and .env to the primary checkout, so a trial launched from any worktree reads the one built dataset and lands in the one runs/ archive. Never build a second data/ inside a worktree. Each worktree gets its own .venv (uv sync).

• .worktrees/inspect-eval is the long-lived checkout of the inspect-eval branch that the Inspect report exporter publishes to; it is merged into main periodically. Do not use it for task work.

• Remove a worktree only when its branch is merged and it has no uncommitted or untracked work: git worktree remove .worktrees/<name> then git branch -d <tool>/<name>. git worktree list is the inventory.

Merging to main

• Merge a small change without asking: a few files, behaviour preserved or narrowly fixed, no new dependency, prompt bytes, stored format, or grading change.
• Stop at a ready-to-integrate branch for anything else: new or removed surface area, a changed prompt or rubric, a changed run or metadata format, a new dependency, or results from a paid run. Say what changed, how it was verified (uv run ruff check . && uv run pytest -q), and what risk remains, then ask. An instruction to merge, integrate, finish, or ship in the original request is that approval already given.
• A dirty main is not a reason to stop. If the dirty paths and the incoming diff are disjoint, merge; if they overlap, reconcile with a path-scoped stash and say so. Another agent's uncommitted work is that agent's to commit; never discard it.
• Push main after merging. Collaborators commit to main directly; fetch before merging and never force-push.

Runs and results

• runs/ and data/ are gitignored and exist only in the primary checkout.
• Never edit a script that a running trial is executing in place (the shell reads it incrementally); write a copy and mv it over.
• Reports and grades that feed the README numbers are tracked under reports/ and benchmark/. Regenerate through the scripts that produced them and commit the outputs with the code change that changed them.