Lean sessions: where long-session requests go, and flag-gated prototypes

[solvyxtech]

Measurement study of where a long session's request characters go, plus prototypes behind MAAT_LEAN_* env flags. No default changes; every flag is off unless set. Full write-up: reports/lean-sessions-study.md (local reports dir).

Breakdown (flags off, scripted providers, no paid calls)

Scenarios: lean-budget d, plus two real Mistral Large 4 bench runs (80 and 93 requests) replayed call for call from their exuviae. The duration-bug replay sends 3.01M estimated tokens; the real run billed 3.06M.

	d	replay duration-bug	replay perf-pairs
total chars	7,262,564	12,049,129	11,607,938
tool results	89.9%	76.3%	71.9%
results older than newest six	60.8%	67.8%	64.0%
tool calls	2.1%	13.0%	17.8%
system prompt + tool schemas	5.8%	5.3%	4.7%
cacheable prefix	94.2%	97.5%	97.6%

The system prompt and tool schemas are a small, cached cost, and nothing changing sits early in the prompt. The session's size is set by the 60k auto-shed threshold.

Prototypes (total chars / cache-weighted, vs baseline)

flag	d	duration-bug	perf-pairs
MAAT_LEAN_SHED_AT=30000	−39% / −23%	−44% / −34%	−50% / −39%
MAAT_LEAN_SHED_AT=40000	−25% / −15%	−32% / −27%	−41% / −33%
MAAT_LEAN_AGE=1 + _ARGS=1 (batched ageing)	−27% / +13%	−30% / −17%	−38% / −23%
MAAT_LEAN_AGE=at=30000,batch=0 + _ARGS=1 (soft shed)	−31% / −9%	−38% / −28%	−48% / −35%
MAAT_LEAN_SHED_MINFREE=0.25	0	0	0 (fixes the shed defects below)
MAAT_LEAN_SUPERSEDE=1	0	0	0 (fixes the elision bug below)

Cache-weighted counts cached prefix chars at 10%.

Defects found

• Wasted shed: Maat's own notes are user messages, so a shed can cut near the start of the session. The real run shed 60,788 → 60,387 tokens (0.7%), which threw away the whole cache, and then shed again one step later. With a tiny first exchange, planShed returns null and auto-shed never fires again. SHED_MINFREE fixes both.
• Elision drops unrelated results: superseded reads are keyed by step, so writing dur.py also elided test_dur.py, which had been read in the same step. SUPERSEDE makes elision per call.

Checks

• lean-budget passes 7/7 with each flag, except batched AGE. That fails only d's "the long session never shed" mechanism check: ageing keeps the history under 60k. Its numbers are well under the ceilings.
• npm run check passes with each flag on. Batched AGE fails the same single test; e2e and self-check pass when run separately.