chat-branch

chat-branch — fork heavy work, merge back the summary (a jusTokenMax module)

Every file read and tool result in the main conversation stays in its context for the rest of the session, compounding cost. The fix is to branch: run self-contained sub-tasks in a subagent whose context is thrown away afterward, returning only what matters.

When to branch

• Research/exploration sweeps ("how does X work across these 12 files?").
• Investigations over large files or logs where you need a conclusion, not the raw content.
• Broad searches that surface a lot but resolve to a short answer.
• Any step whose intermediate output is large but whose result is small.

How

1. Launch a subagent (Task tool) with a tight, self-contained brief.
2. Tell it to return ONLY a compact digest: the result/decision, relevant file:line references, and a few lines of summary — never transcripts or full file dumps.
3. Integrate just that digest into the main thread. Do not re-read what the subagent read; trust its digest (and its cited locations if you need detail).

This is the cheapest big lever: the main conversation grows by a paragraph instead of by tens of thousands of tokens. Combine with code-index (have the subagent query the index) and justokenmax (have it digest logs) for compounding savings.

The /justokenmax:branch command is a shortcut for this pattern.