Skip to content

perception.md

Latest commit

 

History

History
31 lines (27 loc) · 2.67 KB

perception.md

File metadata and controls

31 lines (27 loc) · 2.67 KB
Raw

Task: perception + similarity layer (argus/browser/*, argus/healing/similarity.py)

Owner: you own ONLY argus/browser/__init__.py, argus/browser/snapshot.js, argus/browser/snapshot.py, argus/browser/effects.py, argus/healing/__init__.py, argus/healing/similarity.py, tests/fixtures/*.html, tests/test_snapshot.py, tests/test_effects.py, tests/test_similarity.py. Read AGENTS.md, docs/SPEC.md §2 and §3.1, and argus/models.py first. Implement exactly those interfaces.

This is the heart of the self-healing engine, so correctness and robustness matter more than speed of delivery:

  1. snapshot.js + snapshot.py per SPEC §2.1/§2.2. Accessible-name computation must handle: aria-labelledby, aria-label, <label for>, wrapping <label>, placeholder, title, button innerText, input[type=submit] value, img alt inside links/buttons. context must disambiguate identical buttons in different table rows / cards (e.g. two "View" buttons → context "Falcon-1 DJI M350 92% Idle" vs "Hawk-7 …").
  2. effects.py per SPEC §2.3, including wait_for_settle (network quiet + DOM quiet via an injected MutationObserver counter) and dialog auto-accept with message capture. Only same-origin fetch/xhr/document.
  3. similarity.py per SPEC §3.1 (pure functions, rapidfuzz). Include the synonym table, hash-class stripping, effective_weights, changed_attributes, compatible, rank.

Tests (no network): build fixtures in tests/fixtures/ — e.g. v1.html (a login form + a table with 3 rows each with a "View" button + a wizard-like form) and v2.html (same page after a refactor: ids renamed, classes hashed, testids removed, wrappers added, nav moved, "Log in"→"Sign in", "Next"→"Continue", a button turned into a role=button). Load them with page.set_content(...). Required test assertions:

  • snapshot finds all controls with correct role/name/label/context; runs < 150 ms on the fixture.
  • for EVERY target fingerprint taken from v1, rank() on the v2 snapshot puts the true counterpart first with score ≥ 0.72 and margin ≥ 0.12 (the row "View" buttons must map to the same row). Print a table of scores.
  • identical page → top score ≥ 0.95.
  • a deleted element (exists in v1, removed in v2) → best score < 0.5 (must NOT be confidently healed).
  • effects: clicking a button that does fetch('/api/x', {method:'POST'}) (use page.route to fulfil it with 201) records NetCall(method="POST", path="/api/x", status=201); a thrown error is captured in page_errors; confirm() is auto-accepted and recorded; url change is detected. Tune weights/thresholds only inside similarity.py if the tests show a real need; explain any change.