AI agents waste tokens rereading the same repo
Every fresh chat starts with amnesia. Without a compact project state, the agent spends expensive tokens reconstructing the basics.
Why this hurts.
Repeated file reads are slow, costly, and still miss why previous decisions were made.
How ShardStitch solves it.
Scope the unfinished task
Keep a reusable project view with changed files, dependency graph, and hot context.
Point to the relevant project evidence
Point the next AI at the relevant slice instead of the whole repo.
Carry discovery that changed a decision
Carry the file map or dependency finding only if it changes the next action; omit broad repeated exploration.
Leave a check against current files
Refresh from disk so the view does not go stale.
What the next AI receives.
- Relevant file set
- Dependency risks
- Current diff
- Known decisions
What stays out.
- Full-repo dumps
- Irrelevant files
- Repeated import wandering
- Context with no current task value
Reduce repeated discovery without skipping verification
FAQ.
Will a handoff guarantee fewer model tokens?
No. Token use depends on the tool, model, task, and how much context must be rechecked. A focused packet can reduce repeated discovery, but savings are not guaranteed.
Should the next AI skip reading files named in the packet?
No. The packet helps locate relevant files; the next session should inspect their current contents before relying on a summary.