src/humomni/core/context.py

# context.py — build the VLM input under a token/cost budget (§4).
# Recent frames at full detail, older history as sparse downscaled keyframes, plus an
# explicit memory of what we already said (structural anti-repeat). Causal by construction:
# `history` only ever contains frames with timestamp <= now.
# Frame budget is env-tunable (CTX_RECENT_S / CTX_HIST_STRIDE_S) so the perception lever can be
# swept without code edits; defaults reproduce the shipped behavior.
import os

RECENT_S = float(os.environ.get("CTX_RECENT_S", "6.0"))        # last N s at full res (~2N frames @ 2 fps)
HIST_STRIDE_S = float(os.environ.get("CTX_HIST_STRIDE_S", "2.0"))  # older history: 1 keyframe / N s
STEP = 0.5


def build_context(question, history, said):
    """history: ascending list of (t, frame_path); said: list of emitted strings."""
    now = history[-1][0]
    recent = [(t, p) for (t, p) in history if now - t <= RECENT_S]
    older = [(t, p) for (t, p) in history if now - t > RECENT_S]
    keyframes = older[::max(1, int(HIST_STRIDE_S / STEP))]   # subsample older frames
    return {
        "question": question,
        "now": now,
        "recent_frames": recent,       # sent at native 480x270
        "history_frames": keyframes,   # downscaled in vlm_client
        "already_said": said,          # anti-repeat memory
    }