The next session should know what is still unfinished: three ways to contribute

[OthmanAdi]

The next session should know what is still unfinished

Try the interactive reset demo.

The image illustrates the workflow. It is not a benchmark recording.

The agent writes the patch, reports completion, and leaves the boundary tests pending. Then the conversation resets. Where does the next session find the unfinished work?

Planning with Files keeps a plan, findings and progress in the project. Supported integrations can bring selected planning state back into the session. Optional completion checks can catch unfinished recorded phases, with different controls depending on the host. The files still need accurate updates and real test evidence.

I maintain this project and want contributors to help test that boundary. Here are a few useful places to start.

Contribution	Concrete result	Start here
A first documentation fix	Remove stale instructions that choose the newest plan when several are active. The resolver already refuses ambiguous selection.	Issue #300
A focused Codex documentation fix	Explain that the disabled registered hook exits silently rather than emitting an allow decision. Behavior is already verified.	Issue #302
A current-version recovery example	Save a task midway, reset the conversation, then show which next step the new session recovers from the files. Include the missed steps too.	Issue #303
A large-plan example	Put the active phase deep in a long plan. Compare default and smart injection, and show what reaches the session.	Discussion #218

For a recovery report, please include:

• Host, host version, Planning with Files version and install route.
• The saved planning files, with private information removed.
• The next step you expected after the reset, and what actually happened.
• Whether transcript recovery was enabled. The current automatic default recovers from project files.
• Commands or a small reproducible project, plus the relevant test result.

A successful run helps. A reproducible failure is just as useful. Please check the linked thread before starting so two people do not unknowingly take the same task.

The earlier recovery measurements used a different transcript-recovery default. We need fresh examples for the current behavior before making broader performance claims. A checked-off plan is also not proof that the implementation is correct. Tests and review remain part of the work.

Installation · Contributing guide · Source

If the workflow helps you, a star makes the project easier to find. If it fails in your setup, show us the smallest example that reproduces it.