|
|
79c15cd4df
|
feat: cost-aware trimming + auto-escalation (T1.4 PR 2)
Behavior changes for cache-first prompt discipline:
- PromptAssembly::trim_to_budget(CostMode): trims volatile scratch
first, then oldest appendable log entries, to fit within cost mode
budget. Prefix is never trimmed. Returns count of items removed.
- EscalationTrigger enum: BudgetExceeded + CompactionInsufficient
variants for auto-escalation decisions.
- auto_escalate(): returns Some(next_mode) if trigger warrants
escalation, None if already at ceiling or trigger doesn't apply.
- 11 new tests: trim budget scenarios (under/over/deterministic/
prefix-preserved), escalation chain (fast→smart→max→ceiling),
compaction triggers.
No scheduler injection, no cache warming — PR 3 follows.
Build: pass | Tests: 51/51 green | Clippy: clean | Fmt: clean
|
2026-05-31 16:13:11 +02:00 |
|