Start with the current state
For a change, the agent needs the affected code and applicable decisions. I have it read the source. Older notes help with orientation but must be checked against the current state.
Keep uncertainty visible
If information is missing, the result should say so. I separate supported statements from assumptions. More context alone does not solve this: a detailed description can still be out of date.
Prepare the handoff
When the context window fills up, another agent needs to take over. context-parachute (CLI / hooks) records the task and work state. Open issues and completed checks belong in that handoff. The next agent checks dates and Git state before following it.
Find the history again
With lore (local, Python) I archive and search AI sessions across tools. Its MCP server also makes the history accessible to agents. The archive does not replace the current source. It helps me recover earlier decisions and their reasons.