Fusion pairs a lead agent with a persistent worker. The lead owns the user request,
architecture, acceptance criteria, and final review. The worker handles bounded code
localization through lf-grep and settled execution through sidekick.
/fusion on enable for this session
/fusion off disable for this session
/fusion status show this session's setting
Enabling Fusion disables Boost in the same session. Other sessions keep their own settings.
Prompting the lead to delegate is not always enough. The private routing advisor actively recommends when substantial work should move to the worker. It observes the lead's initial evidence and subsequent execution or delegation boundaries. Recommendations remain visible to the lead while relevant, without becoming user-facing chat.
The advisor does not design the solution or replace the lead's judgment. A failed routing consultation leaves Fusion tools available so the lead can continue.
lf-grep answers a specific repository question with file locations and evidence. It should
stop once it has the requested answer instead of turning adjacent discoveries into a full
repository audit. Each search pass has a smaller runtime allowance than execution work.
If it cannot close a requested edge, it returns a partial map with that uncertainty named.
The lead can commission another focused pass using the worker's existing repository context.
A Sidekick package carries the settled approach, constraints, deliverable, and bounded verification. The worker can discover local implementation details within that boundary. It reports a structured result; failed or skipped required checks cannot be labeled DONE. The lead reviews the result and decides whether further work is needed.
LF-grep and Sidekick share one persistent worker session. New assignments retain useful repository knowledge while replacing obsolete task intent. Recovery of a missing handoff is restricted to reporting, rather than reopening execution.
A new message can return control to the lead while a delegated task continues. The lead can send a targeted follow-up to the running worker using the same task identity. Completion returns to the lead automatically; it should avoid duplicating work the worker still owns.
Delegation adds coordination and model calls, but can displace expensive lead execution. The useful comparison is total cost and completion quality for the whole task. Small, well-localized edits may be better completed directly; substantial implementation and search can benefit from delegation. Actual savings depend on the model pair and workload.