T-9VUY-desktop-console-capture-needs-a-human-pass
State: planning/draft · Impact: low · Complexity: small
Playwright cannot drive the Tauri webview, so the whole desktop surface was checked statically — capability files, generated IPC transport, capability matrices — never by opening the real app and watching its console. A human with a free dev session needs to run that pass once.
| Location | Role today |
|---|---|
solutions/determined/products/agent-pants/src-tauri/ | The Tauri app; a static-only read found IPC transport parity (471/471 methods) and three capability-file gaps (AP-99, AP-100, AP-101) |
Two survey attempts to launch it were blocked, not by the app itself:
one because the assignment’s own rules forbade starting a cargo build
of product-agent-pants, the other because the harness’s process-launch
attempts (bunx tauri dev, running the built binary directly) were
denied by the automation session’s own permission classifier.
Proposed
Section titled “Proposed”Someone launches the desktop build against a copy of the recommended vault fixture, exercises the same routes and actions the web survey covered (Home, a skill note, a plugin page, install/adopt/publish), and records what devtools’ console actually shows — confirming or refuting the statically-predicted gaps (the deep-link unhandled rejection in a focus/surface window, the denied directory picker) and catching anything static analysis could not predict.
Approach
Section titled “Approach”- Build or reuse a
product-agent-pantsdev binary outside any sandboxed session that would block launching it. - Point it at a disposable copy of the recommended vault fixture
(
solutions/determined/products/agent-pants/tests/serve.ts’s fixture is a reasonable source to copy). - Open devtools in the main window and in a focused vault window and a detached Settings (surface) window; exercise Browse (directory picker), a notification-triggering action, and normal navigation.
- Record every console error, unhandled rejection, and failed request, noting which window label it came from.
- File any new defect as its own task; if the findings only confirm AP-99/AP-100/AP-101, note that instead of filing duplicates.
| Area | Note |
|---|---|
solutions/determined/experiences/agent-pants/docs/current-limits.md | Record the desktop console-capture result once run |
Acceptance criteria
Section titled “Acceptance criteria”- AC-1: The desktop app has been launched at least once against a copy of the recommended vault, with devtools open in a main, focus-, and surface- window.
- AC-2: Every console error, unhandled rejection, and failed request observed is recorded, with the window label it came from.
- AC-3: The statically-predicted gaps (AP-99, AP-100, AP-101) are confirmed or refuted against the real console.
Out of scope
Section titled “Out of scope”- Fixing any defect this pass finds — that becomes its own task.
- Building Playwright/webdriver support for the Tauri webview.
Dependencies
Section titled “Dependencies”- none
Discovery context
Section titled “Discovery context”Found by the agent-pants bug survey, N-5LD4-bug-sweep-2026-09-12-agent-pants.
The survey’s two boot attempts and their exact block reasons are
recorded in the findings record (jq 'select(.needs_human==true and .surface=="desktop")' against the survey’s machine record).