plan.md drifted away from what actually changed
Project notes help until they become stale. Then the next AI treats an old plan like current truth.
Why this hurts.
If a plan file is not reconciled, it can hide completed work, missed files, or decisions that changed during implementation.
How ShardStitch solves it.
Compare planned files with actual changes
Read notes and compare them with current repo state.
Classify each plan item
Mark each plan item done, pending, stale, or unknown using current files and check results.
Refresh claims from current evidence
Build the next packet from latest disk truth, not the note alone.
Keep the remaining work explicit
Show both the plan and the evidence that confirms or contradicts it.
What the next AI receives.
- Plan state
- Diff reality
- Stale warnings
- Updated next action
What stays out.
- Outdated checklist items
- Completed tasks marked open
- Old assumptions
- Notes with no repo evidence
Reconcile the plan with the checkout
FAQ.
Should plan.md be treated as the source of truth?
Treat it as a planning record, not proof of the current code state. Compare its claims with the checkout, commits, and test results before relying on them.
Can ShardStitch determine that every plan item is complete?
A handoff can compare saved notes with available project evidence, but completion still depends on acceptance criteria and current verification. Review ambiguous items yourself.