Updates August 3
PGrid
- found a reddit post asking for feedback on a pcpartpicker build. found that an agent could search alternative pricing for similar components in our db. created a skill for this so i can iterate on how the agent works and potentially turn it into a new feature
- expanded crawler coverage
- started crawling mobile phones for a few sites
- added a Costco AU listing crawler for laptops, monitors, phones, smart watches and headphones
- added more headphone crawling. catalog and classification support for headphones is still work in progress
- continued work on laptop details pages
- added direct-access detail pages for laptop models that aren’t listed publicly yet, while keeping them out of the sitemap and laptop list
- added configuration filtering for processor, memory, storage and touch, so users can narrow the listing table to configurations that actually exist
- kept iterating on the model vs variant ux, screen-size navigation, configuration controls, colours and spacing
- did a lot more production classification and specs search work
- got agents to work through and review the laptop classification queue, finding repeated issues around touch being omitted, approximate screen sizes, model generations and similar Lenovo families
- ran supervised laptop model and variant specs searches and fixed agent schema inspection failures and inconsistent rules between the different laptop prompts
- fixed classification review corrections so they can safely recover the matched model and brand after an earlier reclassification
- corrected all active Apple laptop variants in prod to use CPU and GPU core-qualified configuration names. this included splitting a mixed M5 MacBook Air variant and verifying postgres/clickhouse data afterwards
- continued work on catalog images and cutouts
- added an internal catalog image dashboard for comparing source images, selected images and cutouts
- built an ai cutout audit that can review new or existing cutouts, keep resumable progress and stop when images are flagged
- created a new r2 bucket for public images and migrated public images to there from current private bucket
- added an internal dashboard page for viewing prod browser debug info
- improved crawler reliability
- found JB Hi-Fi had a default pagination cap of one page, so phone crawls were only getting the first 100 products. raised the safety cap and tightened the phone filters
- fixed pagination failures across centre com, computer alliance, mwave, newegg and officeworks
- changed the crawling feature flag into a real queue pause. automatic crawl jobs now stay queued instead of being claimed and completed as no-ops
- completed the move from NordVPN to Surfshark for normal crawler traffic
- moved all six proxy allocations to Surfshark across Australia and the US
- added all five Australian Surfshark locations
- kept NordVPN and PureVPN available as supported providers even though they aren’t currently selected
- built the first version of a remote development workflow for codex tasks
- can publish local committed and uncommitted state to hosthatch2, promote it into an isolated worktree and fetch the result back for local review
- still iterating on fetch/recovery safety and keeping large local
.pgrid artifacts out of transfers
- continued work on the buildapc knowledge base
- imported the latest reddit data and updated the wiki with the new synthesis
- added a refinement log to record why prompts and rules change instead of losing that history between iterations
- continued work on the ai deal discovery agent
- added better price context and direct pgrid links to alerts
- reviewed its production suggestions and adjusted the available tool budget
- made RAM speed filtering more precise with exact-value and range controls