v0.19.6 — cheaper skill-card injection
Esta es la nota de release tal como fue publicada, no una reescritura de ella. Las notas de release se publican en el idioma en que fueron escritas.
The self-evolution flywheel's card-reading loop (injecting learned reasoning cards into the agent's context) was kept opt-in mostly because injection cost +~349% tokens on short tasks. This release cuts that ~7–9×.
Changed
- Skill-card injection is ~7–9× cheaper. Three levers, measured over two runs on the hard suite:
- inject the single best card (
CHIMERA_SKILL_CARDS_K=1, was 3), - a relevance gate (
CHIMERA_SKILL_CARDS_MIN_OVERLAP=2) — a task with no strong match (< 2 shared query terms) injects nothing, so it pays no tokens and gets no misleading card, - a shorter render (
CHIMERA_SKILL_CARDS_MAX_LINES=3). Overhead dropped from +~349% to ~+37–48%, now under the registered +50% flip gate. These are the new defaults (they only apply when card reading is opted in).
- inject the single best card (
Honest note
The M19-A1 default (reading cards automatically) still stays OFF — the cost blocker is removed, but the accuracy lift isn't statistically significant at n=12 (it swung +16.7pp → −8.3pp between identical free-model runs). Flipping it now waits on a larger paired A/B, not on code. See bench/skillcard/RESULTS.md.
Gate: 1460 passed / 3 skipped, mypy + ruff green.
Upgrade: `pip install --upgrade 'chimera-agent[full]'`