FM Phase 3: Tiered payload assembly + context budget + shrinking

Build payloads from the storage hierarchy instead of flat "last N turns".

L1 within verbatim_turns/verbatim_minutes
L3 within configured window
L2/L4/L5/L6 via memory search (relevance-gated); L6 filtered to current
Token estimation per section, percentage-based allocation from max_context (budget_persona_pct, budget_session_pct, budget_summaries_pct, budget_search_pct, budget_reserve_pct)
Shrinking: cut from priority 9 upward (see design doc table); never cut new user message or previous raw turn

Files: chat_sessions.{h,cpp}, config.{h,cpp}

FM #23 (step 1): payload assembly budget + window config keys

FM #23 (step 2): budget-aware payload assembler + priority shrinking

FM #23 (step 3a): recipe-ingredient fetch methods

FM #23 (step 3b): wire tiered payload into the chat REPL

Merge #23: tiered payload assembly + budget/shrinking (CLI path)

Payload assembly can be triggered before it's triggered by the agent framework and be aware of cache validity.

Navigation Menu