codeSplitter: agent budgets per path, hand off to fresh eyes with a caveman-terse note #1077

Agents get worse per dollar as their context grows. Every turn re-reads the whole session, and long sessions also drift. codeSplitter already runs agents (conflict agent, self-heal fix tasks, agent sandbox) and has a per-path policy table (#1050, #1051). Give that table a budget, and have agents hand off to fresh eyes instead of grinding on.

What to build

1. Budget columns in the per-path policy table: a context or spend threshold per task, plus what happens when it's crossed: handoff (fresh agent), escalate (person), or downshift (a cheaper model for mechanical work such as merges, renames and retitles).
2. Triggers that cut a run short before the budget runs out:
   • context past the threshold (transcript size; for Claude Code harnesses, a hook can read it)
   • the same failure twice (the same test, or the same error text)
   • a third conflict round on one branch
3. A handoff note, written caveman-terse. The next agent reads it cold, so it should be short and blunt:

   DONE: x, y
   NEXT: z (file:line)
   BROKE: test foo, why
   DONT: tried a, failed because b
   

   It lands on the branch (or as a review comment), and the relay starts a fresh agent from it with nothing else.
4. Shown on the review page: each agent run's spend, plus "handed off at N tokens because …".

Why

On 2026-10-12, the slow and costly parts of a ~90-PR day were long-lived agents and rework rounds (conflicts, refixes), not task size. Fresh agents with a short note beat one agent grinding through round three. Prior art: Claude Code's auto-compact and /compact do this within one session; this does it across agents, with a budget.

Related: #938, #1050, #1051, #1015 (self-healing CI).

🤖 Generated with Claude Code