feat: control/test comparison strategy + QA backlog doc

- Agent system prompt now instructs Claude to demonstrate BOTH working
  (control) and broken (test) states when bug is triggered by a setting
- Added docs/qa/backlog.md with future improvements: Type B/C comparisons,
  TTS, pre-seeding, cost optimization, environment-dependent issues

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
snomiao
2026-03-27 07:24:10 +00:00
parent fce31cf0bf
commit cf54ddb6d3
2 changed files with 73 additions and 0 deletions

59
docs/qa/backlog.md Normal file
View File

@@ -0,0 +1,59 @@
# QA Pipeline Backlog
## Comparison Modes
### Type A: Same code, different settings (IMPLEMENTED)
Agent demonstrates both working (control) and broken (test) states in one session by toggling settings. E.g., Nodes 2.0 OFF → drag works, Nodes 2.0 ON → drag broken.
### Type B: Different commits
For regressions reported as "worked in vX.Y, broken in vX.Z":
- `qa-analyze-pr.ts` detects regression markers ("since v1.38", "after PR #1234")
- Pipeline checks out the old commit, records control video
- Records test video on current main
- Side-by-side comparison on report page (reuses PR before/after infra)
### Type C: Different browsers
For browser-specific bugs ("works on Chrome, broken on Firefox"):
- Run recording with different Playwright browser contexts
- Compare behavior across browsers in one report
## Agent Improvements
### TTS Narration
- OpenAI TTS (`tts-1`, nova voice) generates audio from agent reasoning
- Merged into video via ffmpeg at correct timestamps
- Currently in qa-record.ts but needs wiring into hybrid agent path
### Image/Screenshot Reading
- `qa-analyze-pr.ts` already downloads and sends images from issue bodies to Gemini
- Could also send them to the Claude agent as context ("the reporter showed this screenshot")
### Placeholder Page
- Deploy a status page immediately when CI starts
- Auto-refreshes every 30s until final report replaces it
- Shows spinner, CI link, badge
### Pre-seed Assets
- Upload test images via ComfyUI API before recording
- Enables reproduction of bugs requiring assets (#10424 zoom button)
### Environment-Dependent Issues
- #7942: needs custom TestNode — could install a test custom node pack in CI
- #9101: needs completed generation — could run with a tiny model checkpoint
## Cost Optimization
### Lazy A11y Tree
- `inspect(selector)` searches tree for specific element (~20 tokens)
- `getUIChanges()` diffs against previous snapshot (~100 tokens)
- vs dumping full tree every turn (~2000 tokens)
### Gemini Video vs Images
- 30s video clip: ~7,700 tokens (258 tok/s)
- 15 screenshots: ~19,500 tokens (1,300 tok/frame)
- Video is 2.5x cheaper and shows temporal changes
### Model Selection
- Claude Sonnet 4.6: $3/$15 per 1M in/out — best reasoning
- Gemini 2.5 Flash: $0.10/$0.40 per 1M — best vision-per-dollar
- Hybrid uses each where it's strongest

View File

@@ -485,6 +485,20 @@ export async function runHybridAgent(opts: AgentOptions): Promise<{
5. Take screenshots at key moments for the video evidence.
6. When you've confirmed or ruled out the bug, call done().
## Control/Test Comparison (IMPORTANT)
When a bug is triggered by a specific setting, mode, or configuration:
1. **CONTROL phase**: First demonstrate the WORKING state. Disable the trigger (e.g., Nodes 2.0 OFF), perform the action, take a screenshot labeled "control-*", verify it works.
2. **TEST phase**: Then enable the trigger (e.g., Nodes 2.0 ON), reload if needed, perform the SAME action, take a screenshot labeled "test-*", verify it's broken.
3. In your done() summary, explicitly compare: "With X OFF, behavior was Y. With X ON, behavior was Z."
This contrast is critical evidence — it proves the bug is caused by the specific setting, not a general issue. Always try to show both states when possible.
Examples of control/test pairs:
- Nodes 2.0 OFF → ON (for node rendering, widget, drag bugs)
- Default theme → specific theme (for visual bugs)
- Single node → multiple overlapping nodes (for z-index bugs)
- Empty workflow → loaded workflow (for state bugs)
## ComfyUI Layout (1280×720 viewport)
- Canvas with node graph centered at ~(640, 400)
- Hamburger menu top-left (C logo)