Mainline: Git-native intent memory for Codex agents

[atlas-comstock]

I’ve been exploring a workflow problem that shows up more as Codex agents start making non-trivial repo changes:

Codex can read the code, but it does not automatically know the repo’s historical intent.

Example failure mode:

• A repo still has a half-built Redis queue.
• There is a redis.go, TODOs, and Redis is still in docker-compose.yml.
• Codex reasonably tries to finish the Redis implementation.
• But the team had already abandoned Redis because replication lag caused duplicate billing events.

The code points in one direction; the repo history points in another.

I open-sourced Mainline as one experiment around this:
https://github.com/mainline-org/mainline

Mainline stores engineering intent in Git refs/notes so Codex and other coding agents can retrieve context before editing:

mainline context --current --json
mainline start "Add JWT auth"
mainline append "Preserved legacy OAuth callback path"
mainline seal

The goal is to capture small durable intent records:

• why the work exists
• what was decided
• what was rejected
• what risks remain
• what future agents should not repeat

This is not meant to replace docs, AGENTS.md, or review. I see it as complementary:

• AGENTS.md / project instructions: stable behavior and conventions
• Mainline: historical decisions and repo-specific intent that future agents should retrieve before touching related code

I’d love feedback from Codex users:

1. Do you see this failure mode in real repos?
2. Should this kind of memory live in AGENTS.md, Git metadata, docs, or Codex session memory?
3. What would make this feel natural in a Codex workflow?