Practical use.
A selective packet should let the next agent act without rereading a long transcript. Its value comes from choosing the evidence needed for the next decision, not from making the packet short at any cost.
ShardStitch builds hot context instead of dumping the whole transcript: current goal, changed files, decisions, failed attempts, verification state, and next action.
A selective packet should let the next agent act without rereading a long transcript. Its value comes from choosing the evidence needed for the next decision, not from making the packet short at any cost.
This feature controls packet scope: include the evidence needed for the next action while leaving irrelevant history and abandoned branches behind.
A good packet answers five questions quickly: what was being attempted, which files changed, what was already tried, what is verified, and what is the smallest safe next action. It can omit unrelated transcript turns while retaining the decision or failure that the next AI must not repeat.
Review the selected files and claims before handing the packet to another tool. Selective context reduces noise, but it does not remove the need to verify the repository after the next edit.
When preparing a handoff, identify the next action first. Include only the current goal, files and diff relevant to that action, decisions that constrain it, failed approaches, and the checks already run.
For a one-file bug fix, the packet might contain the failing behavior, the exact file and diff, the reproduction command, what the command returned, one failed approach, and the next check. Leave unrelated setup discussion out. If an omitted design decision constrains the fix, add that decision and its source instead of relying on the next agent to infer it.
Selective context is not a license to discard a decision, safety constraint, or known failure just to reduce length. If the next action depends on broader architecture or a prior trade-off, include that evidence and explain why it matters.