Version 0.1.2
- No eager HTML push → prevents 128k context overflows.
- Per-turn chunk quota (default 2) → model can fetch more in the next turn.
- Heuristic history trimming to reduce token usage.
- New tools: getDigest, getHtmlBySelector, getTextBySelector.
- Watch mode (SPA/DOM changes): resnapshots meta/digest after route changes or big DOM mutations.
- chrome://extensions → Developer mode → Load unpacked → select this folder.
- Open the Side Panel from the toolbar button.
- Click Snapshot DOM → sends META + DIGEST (no full HTML).
- Ask ChatGPT for actions; it will request
getSourceChunkor selector-based tools as needed. - Toggle Watch: On to auto-resume on URL/DOM changes (SPAs).
- Keep conversations focused; long unrelated chat increases tokens.
- Let the model fetch small, relevant windows (selectors) instead of whole-page chunks.
- Increase chunk quota carefully if needed (sidepanel.js → MAX_CHUNKS_THIS_TURN).