What broke. What got fixed. What it cost.

The system's first version was a working but tangled set of scripts. The v2 rewrite split the pipeline into a real dual-engine architecture — Hermes Agent as orchestrator/lead analyst, Codex as the independent checker that re-derives every key number — with an explicit fallback chain (GLM-5.2 → GPT-5.6 → DeepSeek V4 Pro → Qwen3 14B local) so no single point of failure could take the whole run down. The result is what's documented on the System page today.

Growing from a handful of manual scripts to 100+ scheduled workflows covering 82 tracked tickers meant building real scheduling discipline: per-session timing (pre-market, intraday snapshots, after-hours wrap-up), retry logic, and a quality gate that holds back anything that doesn't meet the bar rather than shipping broken output. The full schedule is public on the All-Weather Run.

The budget envelope is a design constraint, not an afterthought: a headline cap of $55/month, with actual spend running closer to ~$7/month (~13.5% of the cap) thanks to model routing, the local Qwen3 14B fallback, and a budget-watchdog pattern that tracks multi-model spend against the cap in real time. This pattern is one of the seven packs in the Engineering Patterns Bundle.