Skip to content

CHANGELOG.md

Latest commit

 

History

History
27 lines (19 loc) · 2.5 KB

CHANGELOG.md

File metadata and controls

27 lines (19 loc) · 2.5 KB
Raw

Changelog

All notable changes to this project are documented here. Format follows Keep a Changelog; versioning follows SemVer.

0.0.1 — 2026-07-09

First public release.

The harness

  • Kernel — the agent loop as a state machine: assemble context → stream the model → dispatch tools through guardrails → feed results back, within a per-turn tool budget, abortable at every await.
  • Context — layered assembly: base system prompt, fresh per-turn environment (current tab, open tabs, time, locale), persisted session history.
  • Tools — registry + built-in packs (page / tabs / browser data / skills) + MCP mount (Streamable HTTP, no SDK). One ToolDefinition file = one tool, spec derived from its Zod schema.
  • Engine (page_act) — whole-page semantic perception (shadow roots, same-origin iframes, occlusion), fingerprint grounding across re-renders, adapter-aware execution (dom / cdp channels), post-condition verification against the live DOM, diagnosis-driven healing under hard budgets.
  • Observed-facts reconciliation — pre/post snapshots diffed into facts that overrule model predictions in both directions; list_count_delta page-truth baseline catches fake success (toast shown, nothing persisted).
  • Guardrails — risk tiers (read / act / dangerous), per-tool authorization memory, site allow/blocklists, dangerous-action confirmation, sensitive-value redaction, full audit log; optional Chrome permissions requested on first use.
  • Skills & batch — verified runs extracted into parameterized skills; compiled deterministic replay with per-step verification and per-row delivery cross-checked against the page.
  • Model port — OpenAI-compatible / OpenAI Responses / Anthropic providers with native streaming tool calls, automatic prompted-JSON degradation for endpoints without native tools, deterministic scripted mock for tests.
  • UI — chat-first side panel with the live page_act step timeline and per-check evidence; options and onboarding surfaces; zh-CN + en-US i18n; black-and-white flat + frosted-glass design system.

Tests and benchmark

  • Reproducible benchmark versus a deliberate replica of the viewport-agent architecture (pnpm bench → docs/benchmark.md).
  • Headless engine checks (pnpm test:engine), full-loop extension e2e on a scripted mock model (pnpm test:e2e), screenshot sweep (pnpm shots).