tree 3d4f3c2911390428c0c89aa8dcd13dbfb035ce31
parent f18aafd177a90b928b1c1babbe5e763f85eab56d
author Josh Bleecher Snyder <josharian@gmail.com> 1753151294 +0000
committer Josh Bleecher Snyder <josharian@gmail.com> 1753240227 -0700

loop: do slug generation outside the agent loop

[This commit message written entirely by a human; it is all useful.]

We can make a slug based on the first message.
It's good enough.
That keeps it--and the slug tool--out of the agent's context.
It's also one fewer step for extremely short Sketch runs,
which is the straw that broke this particular camel's back.

This is a mild UI regression, in that there's a slight stall
after the user types their first message, during which
the slug is being generated. See (2) below.

While we're here, add handling of compaction agent messages.

This leaves two big TODOs outstanding:

1.

Untangle the awful rats nest that is slug and branch management;
we have callbacks back and forth and layers and locking and it's all confusing.
One visible for that this ugliness takes is that every time the agent tries out a slug,
the top bar in the webui updates, even if we then reject that slug as a duplicate.
there are other forms of ugliness, just less visible.

2.

Make slug generation concurrent with the rest of the agent,
to avoid a short stall right after the user's first request (ick).

When we make slug setting concurrent, we'll likely need to resuscitate
the bashPermissionCheck, except it'll be "silently block and wait for
background slug generation to complete", rather than "reject the tool call".
Ditto for about_sketch, and any other tool call that expects
the slug or branch name to be set.

Generally, before undertaking this effort, we should fix (1) above,
make convos generally concurrency safe (maybe COW?), and
figure out to get race-enabled innie builds.

Co-Authored-By: sketch <hello@sketch.dev>
Change-ID: s8ac5f6a9faa611ebk