feat(browser-test): persistence + screenshots + smarter AI loop (v0.7.9) #19

Closed
Muyue wants to merge 2 commits from release/v0.7.9 into develop

2 Commits

Author SHA1 Message Date
Muyue
83dce00265 feat(browser-test): persistent token + auto-reconnect + screenshot + smarter strategy (v0.7.9)
All checks were successful
PR Check / check (pull_request) Successful in 1m3s
Four user-reported issues with the AI-driven browser test feature:

1. WS dies on every page reload / navigation (token was single-use,
   5 min TTL → AI loses session permanently).
2. AI can't see new pages opened by its actions (same root cause).
3. No screenshot capability — AI cannot capture page state visually.
4. AI burns ~150 tool calls for what's 5 human actions, mostly
   list_clickables loops.

Fixes:

- Token sliding TTL (60 min): ConsumeToken no longer deletes the
  token; it refreshes its expiration on each successful WS connect.
  Same token survives reload / re-paste / navigation as long as
  there's no 60-min idle gap.

- Snippet auto-reconnect: WS onclose schedules reconnect with
  500ms × attempt backoff (max ~2.5s). Handles transient drops,
  server restarts, and WS hiccups without user intervention. Full
  navigation kills the JS context and is unrecoverable from JS — but
  the user just re-pastes the snippet, same token works.

- New 'screenshot' action: snippet captures via SVG foreignObject +
  canvas → base64 PNG → sent back over the existing WS reply
  channel. Server decodes and writes to ~/.muyue/screenshots/
  <filename>.png (sanitized name, timestamp default). Filename
  characters limited to a safe charset to prevent path escape.
  Best-effort: external CSS / cross-origin images / iframes won't
  inline.

- Studio system prompt rewritten <browser_test_strategy>:
  - Explicit rule: don't list_clickables after every click
  - Action cost table (cheap vs expensive)
  - When to re-list (URL change, dialog, click-not-found only)
  - Standard final report format ✓ / ✗ / ⚠ / 📸

Also bundles v0.7.8 (cherry-picked): unsafe.Pointer(uintptr(hPC))
instead of unsafe.Pointer(&hPC) in UpdateProcThreadAttribute, so
PROC_THREAD_ATTRIBUTE_PSEUDOCONSOLE is correctly applied and the
spawned shell attaches to the embedded xterm.js instead of opening
a separate external console window (regression fix from v0.7.6).

- internal/api/browsertest.go: token sliding, screenshot save,
  param schema, snippet rewrite, helpers
- internal/agent/prompts/studio_system.md: strategy rewrite
- internal/version/version.go: 0.7.7 → 0.7.9
- CHANGELOG.md: v0.7.9 entry covering all fixes
2026-04-27 15:07:36 +02:00
Muyue
5ad0ec6662 fix(windows/conpty): pass HPCON value, not &hPC (v0.7.8)
User reported regression introduced in v0.7.6: PowerShell / cmd open
in a separate external console window instead of attaching to the
xterm.js tab (v0.7.5 worked).

Root cause: the ConPTY wiring used
  attrList.Update(PROC_THREAD_ATTRIBUTE_PSEUDOCONSOLE,
                  unsafe.Pointer(&hPC),    // ← wrong
                  unsafe.Sizeof(hPC))

The PSEUDOCONSOLE attribute is a Win32 API quirk: lpValue must be
the HPCON *value* (cast to PVOID), not a pointer to the local
variable holding the handle. With &hPC the kernel reads garbage,
silently drops the attribute, and CreateProcessW spawns the child
with a fresh console — hence the external window.

Fix is one line:
  unsafe.Pointer(uintptr(hPC))

Confirmed against Microsoft's EchoCon sample and Go libraries that
work in production (UserExistsError/conpty, aymanbagabas/go-pty).

- internal/version/version.go: 0.7.7 → 0.7.8
- CHANGELOG.md: v0.7.8 entry with the diagnostic write-up
2026-04-27 15:05:11 +02:00