# Step 3 plan: the Wildcard Pipeline bench pack Status: built. The pack is merged into the Wildcard Pipeline's `dev` (`bench/`, #141), and the check on the bench PC passed on 2026-10-27: S1, W1 to W4, W13, X1, C1 with C12 and P1 ran on qwen3.5-9b with AI times close to the robot's q9 runs. It found one false pass in the pack (a problem's Fix check), fixed in the same PR. P12 didn't run there because the app offered no **Fix all with AI** after P1's build, as with the robot. [docs/plan.md](../plan.md) holds the base design and [step-2.md](step-2.md) the web app. This step ported the bench robot into a Watchwright test pack, added the few engine pieces the port showed were missing, and checked on the PC that the pack reproduces the robot's q9 runs. ## 1. What step 3 delivers The bench robot (`scripts/bench-robot/` in the Wildcard Pipeline repo: `run.js` 437 lines, `lib.js` 606, `tests.js` 673, `tests-meta.json` with 126 tests) clicks through the Wildcard Pipeline AI bench and saves screenshots, answer texts and the app's AI debug log for every run. At the end of step 3: 1. **The bench is a Watchwright pack** in the Wildcard Pipeline repo, `bench/`, next to the `data-test` ids it clicks (research note, section 5). Adding that folder in Watchwright's Packs page shows all 126 tests in 14 sections, with their prompts and what a good answer looks like. 2. **Starting a bench is a New run**: pick the model (q9, q35, q27, haiku, sonnet, opus, fable), optionally reset the library, include the web tests or not, pick sections or tests. Attempts per test follow the bench page's recommended runs for that model; Fable gets its subset. 3. **Everything the robot saved is in the attempt**: screenshots (panel, chat, page, scroll sets), answer texts, the AI debug log per call (Markdown and JSON, viewable in the attempt page), times as metrics (`Plan s`, `Build s`, `Fix s`, …), flags with the right kind. 4. **The robot's control flow maps onto the engine**: pause on a robot problem with Retry or Skip, hot reload of a fix, Stop, resume after a crash, redo broken (with what it needs). No hand-kept `NEEDS` table or `chain` field: tests declare `needs` and `inside`. 5. **Review** uses a bench rubric, and the model matrix is a Compare of the runs for each model. 6. **The pack is checked twice**: in the cloud sandbox against the real Wildcard Pipeline manager with a fake model (robot side only, every test), and on the PC against ComfyUI and LM Studio with q9 (a subset of every section, compared with the robot's last q9 results). Not in step 3: YAML tests, the editor, Record and a second pack (step 4), CLI and MCP (step 5), cost and budgets for cloud models (step 6). The pack only labels which model the app is set to; switching the app's model stays a manual step, as today. ## 2. What the port needs from Watchwright The rule for this step (Dumi, 2026-10-27): Watchwright is a general tester that Wildcard Pipeline uses, not a tool built for it. Everything below is a general feature any pack can use, with no Wildcard names or rules in Watchwright. What is specific to Wildcard Pipeline (its `data-test` ids, its AI debug log, its backup API) stays in the pack. Read side by side, most of the robot maps onto what step 1 and 2 built: `Bot.goto`/`click`/`fill`/`has`/`shot`/`scrollShots`/`keepText`/`note` are core `h` helpers, the pause loop is the engine's broken pause, `--resume` and `--redo` are actions, the live page is the run page. These parts don't map yet; each becomes a general feature, not a Wildcard special case. | Robot | Gap | Watchwright change | |---|---|---| | `saveDebugs(before)` after each run: fetches the app's debug records made during the run and writes `debug-NN-