Updates June 29

PGrid

• Started scraping smartwatches from MSY, JB Hi-Fi, Mwave and Officeworks
• Continued adding support for monitors
  • added new CLI commands to allow agents to do more classification tasks:
    • new command to allow agents to clear specs search queues and classification queues for quicker iteration
    • started an AGENTS.md for the CLI
    • new command for agents to run specs search queue dedupe
    • new command to allow agents to better run specs searches
  • updated verifiers for monitors to not verify identity-defining parameters of a monitor but verify specs like refresh rate
  • ran lots of monitor specs searches; updated the CLI to stop on LLM failures, since these were a common, repeated issue from Vertex
  • ran into normalisation issues similar to those with laptops
• Started a new project to create a knowledge base of common advice and questions that people ask in PC-building subreddits
  • started with agents that take in threads and create takeaways from each thread, then a synthesis agent that takes those notes and can make edits to wiki pages
  • added a SQLite-backed CLI workflow to ingest, synthesise and create wiki pages based on OpenCode agents
  • added the ability to upload and store the wiki and inputs in R2
  • started supporting stacked prepared batches, ingestion retries, cancellation and per-run progress reporting
• Continued work on laptops
  • ran a laptop model expansion specs search; found a bug where some validator calls could result in errors and cause partial data to be saved for a given model
  • updated verifiers to get original listing information
  • changed how verifiers work slightly, removed duplicate raw result keys and removed fallback evidence provided to them.
  • added the ability for laptop model specs search to not expand on a given input due to insufficient source information
  • refactored verifier keys across all categories to be less ambiguous
  • even still after all this work the baml workflow for laptop model specs search kept having accuracy issues. so i added codex sdk agent mode to laptop model specs search. iterated so its 2 agents, one for research and one for normalization and validation
  • with agent specs search, extracted a shared agent runner that is used by laptop and monitors
  • started crawling amazon and ebay for laptops by model family
• Updated the classification audit prompt to be stricter and only try to flag obvious issues
• Removed Mysterium VPN support from the proxy after getting a refund for it
• For live browser debugging, deployed a live browser debugging capability and storage for browser sessions that could be looked up later. It is controlled by a feature flag and limited to item enrichment tasks to start with. Haven’t tested it yet
• Investigated bug with lower scorptec results, fixed bug with some outdated urls to scrape
• Added observability cli commands for agents to use to help find bugs
• Work in progress on a deal email alert interface