Codex — the default execution door

Codex natively reads AGENTS.md — MEMIA projects the full canonical instruction there from the OPERA workspace. Codex is also the harness's default door: the 218 agents run through codex exec, with an always-explicit sandbox.

Install

cd ~/MEMIA
memia projections write     # AGENTS.md carries the full canonical instruction
codex                       # launch Codex INSIDE the workspace: it reads AGENTS.md

Run agents

# see what would be launched, without launching anything
memia agents run core-search-navigator --prompt "Which engagements have no date?" --dry-run

# actually run (read-only sandbox: this agent cannot write)
memia agents run core-search-navigator --prompt "Which engagements have no date?" --porte codex

# the board review: 6 roles, recommendation, risks, disagreements, OPERA trace
memia board run --sujet "Activate the scheduled operations?" --porte codex

• Read-only agent → --sandbox read-only. Writer agent → --sandbox workspace-write, workspace-bound.
• Never danger-full-access in the public package — tested, not declared.
• Every run leaves a trace: a sourced OPERA operation (agent-run), replayable and auditable.