Full rewrite on mtcute (MTProto, no Bot API). TypeScript, pnpm workspaces.
| Package | Role |
|---|---|
packages/core |
Pure pipeline: signals → score (calibrated pSpam) → policy. Zero IO imports — replayable offline. |
packages/adapters |
mtcute gateway, TL→NormalizedMessage normalizer, enrichment (call budget), verdict executor with safety invariants. |
packages/data |
Mongo repos (byte-compatible with v1), signature/vector/moderation/LLM/velocity/session ports. |
packages/ui |
Typed locales (uk/en reference), compact notification views, PM-only settings. |
apps/bot |
Composition root. pnpm start (tsx). |
tools/replay |
Cutover gate: replays prod modevents through the v2 core offline. |
tools/spike |
Live MTProto bot-session capability verification. |
pnpm install
pnpm test # vitest, all packages
pnpm typecheck # tsc --build, strict
cd apps/bot && pnpm start # needs env (below)
cd tools/replay && pnpm replay -- --days 14 # needs MONGODB_URIRequired: API_ID, API_HASH, BOT_TOKEN, MONGODB_URI.
Optional (stage degrades gracefully when absent): QDRANT_URL, QDRANT_API_KEY,
OPENAI_API_KEY (embeddings + moderation), OPENROUTER_API_KEY,
LLM_MODEL, LLM_REQUIRE_SCHEMA, LLM_TEMPERATURE, SESSION_PATH.
LLM_MODEL is an OpenRouter slug — vendor/model, not a bare name. Leave
LLM_TEMPERATURE unset for a reasoning model: LLM_REQUIRE_SCHEMA makes every
parameter sent a routing requirement, so a temperature the model does not accept
404s every call (2026-08-07). See DEPLOY.md for the reason: "http" playbook.
LLM_CHEAP_MODEL / LLM_STRONG_MODEL are still read as fallbacks for
LLM_MODEL, so an existing deployment keeps working, but there is only one
classifier now — the two-tier escalation was removed after the strong tier was
found never to have answered in production (2026-08-07).
- Run
tools/replayover ≥2 weeks of production modevents — investigate every "v2 would act where v1 did not" line until zero unexplained FPs. - One week with the test bot in a live test group.
- 48h on conservative thresholds (
softpreset) after big-bang switch. - Keep v1 deployable for ≥1 month.
Same Mongo database and collections as v1 (additive-only). New collections:
pipeline_decisions (TTL 90d), pipeline_feedback (permanent), llm_cache
(TTL 7d). Signature hashing is a byte-compatible port of v1 — do NOT change
normalization without a migration.