feat: research pipeline + speed-v1 experiment results#19
Merged
Conversation
- --extract-tokens: add video extraction (<video>, <source>, poster), lazy-load data-* attributes, URL decoding on filenames, inline script library detection (GSAP/p5/Three.js/Lottie/etc.), raise element cap to 50k, new VideoAsset type, detectedLibraries field on DesignTokens - --rip: full site download via Playwright network interception. Captures every request/response, rewrites HTML/CSS references to local paths, auto-scrolls for lazy loading, reveals hidden content (accordions/tabs/ carousels), multi-page crawl, self-contained output directory - --design-compare: comprehensive side-by-side comparison of two URLs. Token extraction + pixel diff (pixelmatch) + structural diff. Interactive content reveal before screenshots (expands accordions, clicks tabs, scrolls carousels, opens mobile menus, dismisses modals). Captures per- interaction-state screenshots. HTML + JSON report output New module: src/design/ (page-interaction, rip, compare, types, index)
Add automated hypothesis testing pipeline (run-research-pipeline.mjs) with two-stage screening, cost estimation, and parallel execution. Speed-v1 results (10 hypotheses, 3 reps each): - Promoted: micro-plan-2 (global default), retry-1 + block-images-media (benchmark profiles) - Rejected: compact-first-turn, scout-enabled, vision-auto-escalation, combo-speed - Neutral: llm-timeout-30s, supervisor-early - Candidate: micro-plan-3 (positive signal, needs larger case set) Config changes: - microPlan enabled by default (maxActionsPerTurn: 2) - resourceBlocking explicitly set in DEFAULTS (blockImages/blockMedia off globally) - Benchmark profiles block images/media for both webbench and webbench-stealth - A/B runner supports arbitrary arm IDs and spec.modes
drewstone
added a commit
that referenced
this pull request
Mar 19, 2026
Unpublished since 0.10.0: - feat: screenX/screenY CDP fix for Cloudflare Turnstile (#29) - fix: boost output tokens near max turns (#28) - feat: canvas fingerprint noise + stealth patches (#27) - fix: headless UA override — platform-agnostic Akamai bypass (#26) - fix: nightly CI — Xvfb headed stealth + system Chrome (#25) - feat: retry malformed JSON with minimal context (#24) - feat: three-tier history compression -22% cost (#23) - feat: headless passthrough + Docker benchmark runner (#22) - feat: WebVoyager + WebArena benchmark adapters (#20) - fix: graceful recovery from execute wall-clock timeouts (#21) - feat: showcase command for marketing asset capture (#18) - feat: research pipeline + speed-v1 experiment results (#19) - feat: design rip, compare, and extract-tokens overhaul (#17) - feat: CDP connection, browser profiles, and asset downloader (#16)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
scripts/run-research-pipeline.mjs): automated hypothesis testing with two-stage screening, cost estimation, parallel execution, and decision classificationbench/research/speed-v1.json): 10 hypotheses tested, all annotated with resultsspec.modespassthrough,.startsWith('webbench')mode detectionExperiment Results
Pipeline Features
--two-stage: screen (1 rep) → validate candidates (5 reps), ~40% cheaper--estimate: cost estimation before running--hypothesis-concurrency N: parallel hypothesis execution--max-priority N,--hypothesis <id>,--resume: filtering and resumptionTest plan
pnpm lintpassespnpm check:boundariespassespnpm testpasses (629 tests)pnpm research:pipeline --queue bench/research/speed-v1.json --estimateshows cost estimatepnpm research:pipeline --queue bench/research/speed-v1.json --resumeskips completed (all p99)