TaskState Vault: a local project-state layer for long-running Codex work

[zhaoyeyu]

I have been using a small local project called TaskState Vault alongside Codex, and it has solved part of a problem I kept running into on longer work: the useful project state tends to be scattered between chat history, files, and whatever I remember to mention again.

TaskState Vault keeps that state outside the conversation. It gives Codex a local structure for:

• stable project intent and a mutable task objective
• a dependency-aware task graph and execution queue
• account, domain, project, task, and run-level information
• task state, next action, evidence, artifacts, errors, and objective-change history
• SQLite/JSONL indexes and tsv:// pointers
• copying relevant long-lived information into the current task scope
• a read-only local web UI for inspecting the state without browsing every file
• CLI, Python SDK, and an MCP-ready adapter

It is not another coding agent and it does not replace Codex execution. The idea is to give Codex a durable project-state layer so a complex project does not have to be reconstructed from the current chat every time.

I am already using it myself, but it is still early. I would genuinely like feedback from people who use Codex on multi-stage or long-running projects: does this model match the problems you see, and where does it become too heavy or miss important state? If anyone finds it useful, contributions and critical reviews are very welcome.