Add openai-ads domain skill; fill uploads interaction skill stub

domain-skills/openai-ads/ads-manager.md

@@ -0,0 +1,44 @@
+# OpenAI Ads Manager (ads.openai.com)
+
+Beta product for ads shown in ChatGPT. Account-scoped URLs: `/manage/campaigns?act=<ad account id>`.
+
+## Structure
+
+Campaign → Ad group → Ad. Settings live at different levels:
+
+- **Campaign**: name, type (Standard/product feed), objective (Clicks), locations, daily budget (defaults to `100.00`).
+- **Ad group**: name, max CPC bid (defaults to `4.00`, shows a live "Strong/Weak Delivery" competitiveness signal), Default Ad Destination URL, context hints textarea (semantic conversation matching, not exact-match keywords).
+- **Ad**: name, headline (50 max), description (100 max), destination URL (inherits ad group default), one square image (PNG/JPG ≥256×256).
+
+## Create flows
+
+- `+ Create` button (top right of Campaigns page) → dropdown: Create campaign / Create ad group / Create ad / Upload bulk.
+- "Create campaign" opens a 3-step wizard (Campaign → Ad Group & Ads → Review) in a full-screen modal. Final button is **Publish** with an "agree to the Advertising Tool Terms" note; publishing shows "Applying changes / Creating ads… (0/1)" then a toast: "Campaign created successfully with 1 ad group and 1 ad."
+- "Create ad" alone opens a small modal that requires picking parent campaign + ad group — no budget/CPC controls there.
+
+## Locating fields
+
+Inputs have no stable ids; locate by placeholder / value:
+
+- Campaign name: placeholder `New traffic campaign`
+- Budget: input with value `100.00` (select-all + retype)
+- Ad group name: placeholder `New ad group`
+- Context hints: textarea whose placeholder contains `context hints`
+- Ad name: placeholder `New ad`
+- Description: placeholder `Enter ad description`
+
+Headlines >~40 chars trigger "Copy may be truncated in some placements." (soft warning, non-blocking). Same warning appears for long descriptions.
+
+## Trap: stuck image attachment
+
+The "Ad images" slot can show a grey square tile with an X above it. That tile is a **stuck upload** (`_ImageAttachment_` with a frozen `_TrackProgress_` spinner), not an upload button:
+
+- While it exists there is **no `input[type=file]` in the DOM**, and clicking the tile does nothing (no `Page.fileChooserOpened` even with interception enabled).
+- Fix: click the tile's X (small button absolutely positioned at its top-right corner). A real `input[type=file]` plus an "Upload image" label appear; then `DOM.setFileInputFiles` on that input works (`upload_file()` helper). Uploaded state: tile shows the image, preview card thumbnail updates.
+
+## Misc
+
+- Table cells (Status, Impressions, …) load async as skeleton placeholders; poll row text until `Serving` appears before reading state.
+- Account display name, timezone, and advertiser logo: Settings → General → Account Info (Edit). The logo is the small avatar shown next to the advertiser name on ad preview cards.
+- A yellow "Complete business verification to keep ads serving" banner persists account-wide until verification is done; ads serve during the grace period.
+- Ad previews render the image as a tiny (~72px) square thumbnail next to the copy — creative is closer to a search-ad favicon than a Meta-style banner.

interaction-skills/uploads.md

@@ -1 +1,24 @@
 # Uploads
+
+Order of attack for file uploads:
+
+1. **Existing `input[type=file]`** — even hidden ones. `upload_file('input[type=file]', '/abs/path')` (CDP `DOM.setFileInputFiles`) works without any dialog appearing.
+2. **No file input in the DOM** (input created on demand): enable interception *before* clicking the trigger, then feed the chooser from the event:
+
+   ```python
+   cdp("Page.setInterceptFileChooserDialog", enabled=True)
+   click(x, y)  # the Upload button/tile
+   wait(1)
+   evs = [e for e in drain_events() if e.get("method") == "Page.fileChooserOpened"]
+   if evs:
+       cdp("DOM.setFileInputFiles", files=["/abs/path"],
+           backendNodeId=evs[-1]["params"]["backendNodeId"])
+   cdp("Page.setInterceptFileChooserDialog", enabled=False)
+   ```
+
+   Without interception the native macOS chooser opens and CDP cannot drive it.
+3. **Click produced no `fileChooserOpened` event**: the thing you clicked is not the trigger. Common cause: a stale/stuck attachment preview occupying the slot — remove it first (look for a small X button) and a real `input[type=file]` often appears (see `domain-skills/openai-ads/ads-manager.md` for a field example).
+
+Verify after upload: re-query the preview element (`img.naturalWidth > 0`, `complete === true`) or re-screenshot — do not assume the set call worked.
+
+Drag-and-drop-only dropzones: see `drag-and-drop.md`; synthesize a `DataTransfer` via JS or fall back to `DOM.setFileInputFiles` on the dropzone's hidden input.