Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/bitwikiorg/continuity/llms.txt

Use this file to discover all available pages before exploring further.

OpenAI Codex uses AGENTS.md as its primary instruction file and composes instructions across global, repository, and directory scopes automatically. Because Continuity’s default entry file is already AGENTS.md, Codex requires no renaming or delegation step — copy the templates, declare the state chain in AGENTS.md, and Codex picks it up on the next run.

​
Setup

1

Copy Continuity templates to your workspace

Copy the files matching your chosen tier from the Continuity repo into your project root:
cp continuity/{AGENTS,IDENTITY,SELF,SOUL,STATE,INIT,TODO,SNAPSHOT}.md /your/workspace/
Replace all {{UPPER_CASE}} placeholders before running. Use the Minimum tier to start and add files as your agent’s loop requires them.
2

Place AGENTS.md in the project root

Codex composes AGENTS.md from three scopes in order: global (user home), repository root, and the directory the agent is currently working in. Place your Continuity AGENTS.md at the repository root. Codex will merge it with any global instructions automatically.You do not need a separate CLAUDE.md or wrapper file — AGENTS.md is the native entry point for Codex.
3

Declare the state file load order in AGENTS.md

Edit AGENTS.md to include an explicit load order. Codex reads this as part of its instruction composition:
## Boot sequence

On every session, load in order:
1. IDENTITY.md — instance identity
2. STATE.md    — current operational truth
3. TODO.md     — active task queue
4. SNAPSHOT.md — last known good state

Read STATE.md before acting.
After meaningful work, update STATE.md, TODO.md, and SNAPSHOT.md.
Write output updates as structured file changes.
Add further files (MEMORY.md, TOOLS.md, PLAN.md) only when your agent’s loop actually needs them.
4

Replace placeholders and confirm READY state

Search all copied files for {{UPPER_CASE}} markers and fill them in. Prompt Codex to read the state chain and confirm readiness:
Read STATE.md and TODO.md. If READY, implement the next valid action.
Update continuity files after work is complete.

​
Key difference from Claude Code

The main distinction between the Claude Code and Codex adapters is which file the runtime auto-loads:
Claude CodeCodex
Auto-loaded fileCLAUDE.mdAGENTS.md
Scope compositionNested CLAUDE.md per directoryGlobal → repo → directory AGENTS.md
Entry fileCLAUDE.md (or delegates to AGENTS.md)AGENTS.md directly
Auto-injectionAGENTS.promptinclude.md (Claude-native)Not supported natively
Because Codex does not support native auto-injection via AGENTS.promptinclude.md, reference all state files explicitly in AGENTS.md rather than relying on a promptinclude list. This also means you should be intentional about load order — Codex will not pull in IDENTITY.md or SNAPSHOT.md unless AGENTS.md instructs it to.

​
Writeback contract

Codex writes file updates as part of its result output. The expected writeback shape after each meaningful action is: 
{
  "status": "READY",
  "file_updates": {
    "STATE.md": "..updated content..",
    "TODO.md": "..updated content..",
    "SNAPSHOT.md": "..updated content.."
  }
}
At minimum, Codex must update STATE.md (new operational truth), TODO.md (task queue with completed items marked), and SNAPSHOT.md (refreshed restoration anchor) after every meaningful session. Move completed tasks to completed/ with evidence entries.
Codex produces file updates as part of its output — if your wrapper or calling code does not apply those updates to disk, the state chain will drift. Ensure your tooling writes the file_updates payload back to the filesystem after each run.

​
Minimal Codex configuration

A minimal AGENTS.md for Codex looks like this: 
# AGENTS.md

## Identity
Role: {{AGENT_ROLE}}
Version: {{AGENT_VERSION}}

## Boot sequence
On every session:
1. Read IDENTITY.md
2. Read STATE.md — current operational truth
3. Read TODO.md  — active task queue
4. Read SNAPSHOT.md — restoration anchor

## Governance
- Act only within defined scope.
- Check WARNING.md hard limits before irreversible actions.
- Never store secrets in markdown files.

## Writeback
After meaningful work:
- Update STATE.md with new reality.
- Update TODO.md — move completed items to completed/.
- Refresh SNAPSHOT.md.
- Commit all changes.
If you need to chain Codex output into a subsequent Claude Code or Hermes session, no prompt recap is needed. Once Codex writes STATE.md and TODO.md, any runtime that reads those files can continue from the same workspace — that handoff is the core value of the file-based state model.

​
Wrapper script (optional)

For automated pipelines, a thin wrapper script can drive Codex with a consistent prompt: 
python run_with_continuity.py \
  --workspace . \
  --task "Read TODO.md, implement the next valid action, update continuity files"
The wrapper is optional. It is useful when Codex runs as part of a CI pipeline or a multi-agent orchestration loop where the calling process needs to manage state file writes programmatically.

Build docs developers (and LLMs) love

Get started for freeTalk to us