search-with-proof: name the exhausted limit, keep partial evidence, tighter reservation

budget_spent carries end_detail (spend, task_spend, reason_rounds, ...); a task
whose allowance cannot cover the next call finishes with its facts; per-task
metrics in pilot records; input reservation bytes/2 + 1024 per message (measured
3-4x over-reservation before).

Co-Authored-By: Claude Fable 5.1 <[email protected]>
This commit is contained in:
Marius Ciepluch
2026-10-05 20:57:56 +02:00
co-authored by Claude Fable 5.1
parent 4be2063d40
commit c33b053b15
11 changed files with 388 additions and 39 deletions
+4 -2
View File
@@ -6,8 +6,10 @@ from dataclasses import dataclass
@dataclass(frozen=True)
class SearchBudget:
spend_usd: float = 0.50
max_call_usd: float = 0.05
# Recalibrated 2026-10-05 on the first three measured searches: settled cost was ~$0.003 per model call,
# but worst-case reservations (~$0.027 per call) exhausted the old $0.05 task allowance after 2-4 calls.
spend_usd: float = 2.00
max_call_usd: float = 0.40
reason_rounds: int = 4
intents_per_round: int = 3
demonstrations: int = 3