-
Notifications
You must be signed in to change notification settings - Fork 1
History
Project Architect 2.0 ran its whole method inside one long Claude Code session: the plan, the rules registry, the in-progress phase file, the cookbook and the ops notes all sat in or next to the load order, and every request re-sent them. Measured on two real 2.0 projects in September 2026, TerrainDiffusion7DTD and Vantage, the average context per request was 459k and 547k tokens. The living files had grown without bound: phase files of 161 to 302 KB, rules registries of 103 KB and 324 KB, cookbooks of 226 KB, 306 KB and 3.6 MB. What Claude Code injected into those sessions (prompt snapshots, reminders, instructions) came to 79k and 140k tokens, re-sent with every later request. By Claude Code's own cost figures, at API list prices, the three projects' 2.0 sessions from late August came to $4,197 (Vantage, 38 sessions), $4,000 (TerrainDiffusion7DTD, 28) and $14,400 (bfm-decomp, 52).
3.0 moves each role into its own small context and measures itself. The acceptance record compares the first three 3.0 phases, run through the router, with that baseline:
| Measure | 2.0 baseline | 3.0: phase 3.1 / 3.2 / 3.3 | Target |
|---|---|---|---|
| Average context per request (experts) | 459k / 547k | 47.1k / 68.6k / 48.0k | ≤150k |
| An expert's starting context (median) | 100–300k | 13.9k / 14.5k / 15.1k | ≤40k |
| Bash output per expert run (median) | ~140k per session | 6.0k / 14.1k / 5.4k | ≤30k |
| Task-tool reminders | 35–50 per session | 0 / 0 / 0 | 0 |
| Cost per request | $0.35–0.42 | $0.110 / $0.102 / $0.076 | ≤$0.10 |
| Cache prefix rewrites after a run's first request | 1–4 | 30 / 2 / 6 | 0 |
| A retriever's starting context (median) | 25–50k | 8.2k / 8.1k / 9.2k | ≤6k |
The last two rows missed their targets, as did the router's context at the phase end and the locking of every live savings row, and they became later phases' work: 3.8 cut the retrievers' bodies and the router's own outputs, and 3.9.5's warmer keeps caches from expiring while agents wait. The full record, every row with its method, is PA3's verification record, kept with its build history rather than in the package. Its sharpest contrast needs no baseline at all: while 3.0 was being built, this repository's own sessions also worked the 2.0 way, by hand, in the same weeks and by the same developer, and averaged 505k to 591k tokens of context per request.
2.0 is retired: it is no longer supported or distributed.