Question
How is a back end chosen for a batch, and what does the user see when it splits?
The charting session settled that the edits file becomes an interface with two interchangeable back ends behind it: a deterministic local writer and the Claude skill. It did not settle who picks, or when.
Resolve:
- What selects a back end. Always try the writer and fall back on refusal, user picks at Apply time, a project-level preference, or something else.
- Whether the choice is per-batch or per-patch. Per-patch is the honest granularity, since a batch will routinely contain nine patches the writer handles and two it cannot - but it means a single Apply produces a partial write plus a hand-off.
- What the Overlay says in that split case. "9 of 11 applied, 2 need Claude" is honest; it is also a worse experience than either back end alone, and the wording decides whether the tool feels capable or broken.
- What happens to the
pending/reconciled Batch flip when only part of a Batch was applied. The current model has two states and no concept of partial.
- Whether the reconcile badge still reads correctly, given it reflects the edits file's own view of what is pending.
- Whether someone with no LLM at all ever sees a dead end, and what the tool tells them when they do.
Blocked on the writer's uncertainty policy, which decides how often the split case actually happens.
Question
How is a back end chosen for a batch, and what does the user see when it splits?
The charting session settled that the edits file becomes an interface with two interchangeable back ends behind it: a deterministic local writer and the Claude skill. It did not settle who picks, or when.
Resolve:
pending/reconciledBatch flip when only part of a Batch was applied. The current model has two states and no concept of partial.Blocked on the writer's uncertainty policy, which decides how often the split case actually happens.