Can your coding agent recover the project state, not just read the code?

Most coding agents can reopen a repository. That is not the same as resuming a project.

After a long run, context compression, an interruption, or an agent handoff, the agent can inspect the code and retrieve old notes. But it still has to reconstruct questions such as:

• What did the user actually accept?

• Which part of the project is active, and why?

• What is blocked by what?

• What is finished in reality, rather than merely planned?

• What evidence is still missing before the work can be called complete?

The repository mostly preserves implementation results. Retrieval memory returns relevant fragments. Neither necessarily provides a coherent view of the project's current operational state.

That gap is why I built TreeWork.

TreeWork is an open-source Codex plugin that guides development through three stages: align the real intent, build a branch tree with explicit dependencies, then work through the tree deliberately. Each branch carries its Spec, Plan, Progress, Findings, Verification, and lifecycle state. When context is lost or another agent takes over, it can recall the accepted project state instead of reconstructing it from scattered evidence.

The goal is not to prescribe every coding step. It is to give long-running coding agents a durable sense of direction: where they are, what they are building, what must happen next, and what has actually been verified.

TreeWork is available here:

https://github.com/Johnny-xuan/TreeWork

I would love to hear from people using coding agents on real, long-running projects: where does your agent usually lose the thread? Is the missing piece better retrieval, or a persistent model of the project's current state?