The last post on the Redwood City project said the City had produced 200 schedule files. That was already a lot. It was the second production: 201 schedule records, about 145 MB, including 154 PDFs and 47 Primavera P6 XER source files covering monthly updates from UD04 through UD44.
That was not the whole record. The first production had 33 PDFs: staff reports, agreements, amendments, and the contract trail. Then the delayed June production landed after a follow-up to the City Clerk: 504 more files, 1,303,180,620 bytes on disk, with zero download failures after validation. That batch included the Thompson change-order log, C03 project-fund disbursement requests, monthly progress reports, weekly OAC / owner meeting minutes from 2021 through 2025, and another full schedule set through UD45. Two files were too large for normal GitHub push limits, so even archiving the source material became its own workflow.
The hard part is that these are not just documents. The PDFs are readable, but the XER files are database exports from a scheduling system. A useful answer has to preserve activity IDs, WBS structure, calendars, predecessors, successors, data dates, baseline dates, critical-path flags, float, and milestone movement across dozens of monthly updates. You cannot paste that into a chatbot and ask what happened. You have to parse the files, turn the tables into something joinable, compare update to update, then reconcile that against the narrative PDFs and meeting minutes when the stated cause changes.
That changed the job. This is no longer “ask an AI to summarize some PDFs.” It is document operations: download before links expire, validate every file, build a per-file guide, track duplicates, keep oversized records recoverable, parse XER schedules, reconcile meeting minutes against schedule narratives, and maintain an open-questions ledger so the analysis does not drift. The active frontier is still the OAC minutes: 195 weekly files, with 72 summarized and 123 still pending when the document-guide push began.
The local Claude logs are not a clean invoice, but they show the scale. In the same ai-projects work window, the de-duplicated Claude counters I could safely aggregate show roughly 1,774 model calls, 11.6 million non-cached tokens, and 242.8 million cache-read tokens. Codex does not expose the same clean token counters in the local logs I checked, but its project trail is visible in the commits: downloader scaffolding, 504-file archive work, document-guide batches, change-order summaries, disbursement summaries, monthly-report summaries, OAC-minute batches, and the staff-report backfill.
The $20 plan is fine for a clever answer, a draft, or one evening of research. It is not enough for sustained analysis of a messy public-records production where the useful work is iterative: read, extract, compare, notice a contradiction, update the question tracker, go back to the sources, and run the next batch. The constraint is not just model intelligence. It is throughput.
That is the honest status: the AI work is making progress, but only because it is being treated like infrastructure. The cheap plan gets you a taste of the workflow. It does not carry the workflow.