Architecture pivot — LEMP game + Letta/Sanctum agents

Shipped as historical entry.

What Changed

This entry explains the architecture choice: a PHP/SQLite game surface on the multihost stack with Letta-backed NPC minds kept outside the request path unless explicitly called.

Mark cut the original Python/FastAPI plan — no extra infra, no new recurring spend we do not already run.

Approved stack (Doc #345):

| Layer | Choice | |-------|--------| | Game app | PHP + SQLite on multihost (standard sites/<domain>.env pipeline) | | Engine | Deterministic PHP — seeded RNG, dice, rules resolve every outcome | | NPC minds | Letta/Sanctum variant on a box Ada provisions — intent + dialogue only | | Rules | Canonical JSON in-repo, idempotent SQLite bootstrap on deploy |

Athena added two non-negotiables:

1. Hidden-agenda secrecy — player never reads agent memory; probes go through social checks + LLM deflection. 2. Memory hygiene — summarization/archival on Letta agents so token costs do not spiral after a few runs.

Ada scoped agent-box cost (~$18/mo new VPS vs co-locate). Provisioning blocked on Mark's billing token until cleared.

Why It Matters

AI narration needs operational discipline. The player should get richer texture without waiting on slow live calls or losing deterministic replay behavior.

Implementation Notes

The commit message for this slice was:

Architecture pivot — LEMP game + Letta/Sanctum agents

Files touched in this slice included:

• No file list was available for this historical entry.

Verification

• The entry is part of the devlog archive and renders through the same PHP Markdown pipeline as current posts.
• The test suite now requires every devlog post to carry substantive details, illustration and screenshot figures, and valid visual asset references.
• Current browser smoke checks cover representative rich posts and older auto-generated posts after expansion.

Visuals