feat: oversize handling, exact <facts>, calibration with real engine timings, llama 5.0.4 - #130

Merged

bernardladenthin merged 18 commits into

mainfrom

claude_20260630_00

Jul 2, 2026

Commits on Jun 30, 2026

feat: configurable onOversize handling for files beyond the model window

bernardladenthin
and
claude
committed
fix(config): map-reduce huge files over a small fast window (~70 min/file)

bernardladenthin
and
claude
committed

Commits on Jul 1, 2026

feat(config): default the oversize fallback to Granite 4.0 H-1B (1.5B, CPU-fast)

bernardladenthin
and
claude
committed
feat(mapReduce): hierarchical reduce + token-safety headroom (fix cpt overflow)

bernardladenthin
and
claude
committed
feat(facts): exact whole-file counts prepended to the oversize body

bernardladenthin
and
claude
committed
feat(facts): apply to every file (not just oversize) + curated fact sets

bernardladenthin
and
claude
committed
feat: shared <factDefinitions> registry + retry-on-overflow for the general path

bernardladenthin
and
claude
committed
feat: ai-index:calibrate goal + per-machine <calibration> for the plan estimate

bernardladenthin
and
claude
committed
fix: use <calibration> for the live per-file ETA too, not just the plan

bernardladenthin
and
claude
committed
test: end-to-end real-model mapReduce+facts smoke test (SmolLM2, guarded)

bernardladenthin
and
claude
committed

Commits on Jul 2, 2026