Project instructions
Repo rules, branch discipline, sensitive files, test expectations, and your definition of done — the entry point for the whole stack.
AGENTS.mdInstall layer
Exactly what lands in your repo.
No prompt pack, no hidden service. CodexFlow writes a signed operating stack into your project as human-readable files — every one visible in your Git diff before you commit.
The install layer
Six folders, all human-readable.
This is the entire footprint — it lives under your repo root and nowhere else on your machine.
your-repo/
├─ AGENTS.mdProject rules, branch discipline, definition of done
├─ .agents/
└─ skills/2,050+ task-specific workflows · 35 categories
└─ .codexflow/
├─ agents/11 specialist roles — planning, review, safety, memory, release
├─ memory/Persistent facts, decisions, handoffs, learned rules
├─ commands/12 safe, repeatable procedures
└─ checks/9 deterministic safety checks
What each layer does
One connected operating stack.
Specialist skills
2,050+ repeatable workflows across 35 domains. Each one has a purpose, inputs, process, checks, and success criteria.
.agents/skillsAgent roles
11 focused operating modes — planning, review, safety, memory, tests, release, dependencies, scope, and onboarding.
.codexflow/agentsPersistent memory
Schema-versioned facts, decisions, handoffs, incidents, and learned rules kept in the repo — not buried in chat history.
.codexflow/memorySafe command layer
Repeated rituals — start, plan, build, review, ship, rollback — turned into reusable procedures that live with the repo.
.codexflow/commandsAutomated verification
Deterministic checks for branches, secrets, destructive commands, payload integrity, memory freshness, and public leaks.
.codexflow/checksAGENTS.md defines
- What the project is
- How the repo is structured
- Which conventions matter
- Which files need extra care
- How to plan before editing
- How testing should be handled
- How handoffs are written
- What “done” means
Memory keeps
Project factsArchitectural decisionsOpen threadsWork handoffsIncidentsLearned rulesRelease notesOperating assumptions
12 safe commands
cf-startcf-onboardcf-plancf-buildcf-reviewcf-testcf-debugcf-shipcf-handoffcf-memorycf-rollbackcf-learn
9 automated checks
Branch guardSecret guardDestructive-command guardPaid-payload verifierMemory freshnessSkill integrityPublic-leak scanClaims consistencyPackage exclusion
.codexflow/agents
The 11 agent roles.
Each is one file with a purpose, trigger conditions, allowed context, output contract, and guardrails.
Task PlannerImplementerCode ReviewerSafety SentinelError WhispererMemory KeeperTest ArchitectRelease CaptainDependency StewardScope SentinelOnboarding Mentor new
Local-first by design
Read every line in a diff.
The payload is signature- and checksum-verified before anything is written. The CLI sends no usage telemetry. Everything it emits is human-readable, versionable, and easy to remove — so you always review before you commit.
$ git status
$ git diff
# keep only the files you want, then commit
Honest boundaries
What CodexFlow does not install.
It gives your repo a reusable operating stack. You still review, test, and own the code you ship.
- A replacement for Codex
- A hosted coding agent
- A background service or daemon
- A production monitoring agent
- A team permissions system
- A guarantee your AI code is correct
See it land in your next commit.
Install the signed CodexFlow layer once, read the diff, and keep only what you want.