[Fix] Read the browser screenshot from the capture's result object, not by pattern search

[chaimaerachdi]

Why

Follow-up to the review of #33 (P3). The screenshot an image-reading decision model (OmniJev) decides over was found with a regex over json.dumps of the whole run-code output. That output also carries the page state playwright-mcp appends after the result, so PNG-like text there could be taken for the picture, and a change in the wrapping could make it miss the real one.

How

_png_from_run_code (s1a/browser/decision_model.py) opens the known layers in turn: payload (compact RPC envelope), MCP content text, result/text/data, and JSON text after ### Result, possibly encoded twice (the same shapes the runtime's own _unwrap_mcp_text_result and extract_json_object handle). It reads the png field of the object SCREENSHOT_JS returns and accepts it only when it is valid base64 that decodes to a PNG signature. Anything else is "no screenshot", which takes the existing path: one retry, then the step goes on without a picture.

What

• s1a/browser/decision_model.py: _SCREENSHOT_RE is replaced by _png_from_run_code; no flag or caller change.
• tests/test_browser_screenshot.py: 4 tests. Every wrapping is opened; PNG-like page text outside the result object is not the picture (the old regex takes it); a png field that is not base64 of a PNG, or not a string, is no image; unknown shapes are no image.
• CHANGELOG.md: entry under Fixed.

Verification

• uv run ruff format --check . && uv run ruff check . && uv run ty check: all clean
• uv run pytest -q: 784 passed, 47 skipped; scripts/smoke.sh: ok
• CHANGELOG.md and the docs say what the code does now

The page-text case was checked against the regex from #33: it finds a picture there, and the new code returns none.

Demo / evidence

• Video link, or exemption/blocker with follow-up: N/A, minor-change exemption (parsing of one result field, no change in agent behaviour on a well-formed capture).
• Application/task → agent actions → independently checked result: N/A
• System1-Agents / System1-Omni commits, checkpoint and worker/frontend provenance: base main 320bb08, head 4236efb.

---

Note

Low Risk
Localized parsing change in browser screenshot extraction with broad unit tests; agent behavior is unchanged when captures return a well-formed png field.

Overview
Browser screenshot parsing for image-reading decision models (e.g. OmniJev) no longer uses a regex over the entire run-code JSON. It now unwraps the known result envelopes (compact RPC payload, MCP text blocks, result/text/data, and JSON after ### Result, including double-encoded cases) and reads only the png field from the object SCREENSHOT_JS returns.

Decoded bytes must be valid base64 and start with a PNG signature; otherwise the step treats the capture as missing (same retry-then-text-only behavior as before). That blocks false positives when playwright-mcp appends page state containing PNG-like text after a failed or empty capture.

Tests cover each wrapping shape, rejection of page-state decoys, invalid png values, and unknown shapes; CHANGELOG documents the fix under Fixed.

^{Reviewed by Cursor Bugbot for commit 4236efb. Bugbot is set up for automated code reviews on this repo. Configure here.}