tell Codex when to use the tools it already has

[ayushcodes10]

Asked to analyse Eigon, a Codex session with all four MCP tools connected made zero calls.

Measured, not inferred:

codex echo-memory server   0:00.67 CPU over 32:43 elapsed   (0.01s in the window)
read_event                 no query_memory rows at all
facts by agent_id=codex    0

Nothing was broken

Codex had the tools. The server was healthy — I drove the full MCP handshake through ~/.codex/config.toml verbatim and it served all four. agent_id was correct. The write_episode description had just been fixed to fit inside the 2048-char window with its example intact.

What Codex did not have was any instruction, anywhere, about when to call them.

	claude-code	codex
SessionStart injection	✅	✗ no hook system
Prompt-time recall	✅	✗
Stop gate	✅	✗
Project instruction	✅ SKILL.md	✗ nothing
MCP tool description	✅	✅

install wrote SKILL.md for Claude and a rule for Cursor, and nothing Codex reads. adopt registered its MCP entry and stopped there — which is exactly the "the tools being available is not the same as an agent knowing to call them" gap that install's own docstring names for every other client.

The change

echo-memory install --for codex writes an AGENTS.md section — merged into whatever the project already has, replaced in place on re-run rather than appended. --for all covers claude, cursor and codex together.

The honest caveat

This session's own evidence is that instructions lose to structure: the SessionStart briefing asked 16 times and was ignored 16 times, because Claude Code's file-based memory is a permanent system-prompt section and the briefing was one reminder.

Codex is not that case. There is no competing memory system — there was simply nothing. This moves it from nothing to something, and whether that is enough is now a testable question rather than an unexamined gap.

Testing

420 pass, 5 added: the section is written, existing project text survives, a re-run replaces rather than stacks, hand-edits after install survive a re-run, and --for claude still writes no AGENTS.md.