What you get when you fork it

Five architectural choices baked into the template. Each one is documented in tech_design.md and enforced by the code in this repository.

• §1 · §3

  Telegram-native reporting with deterministic triage

  Reports arrive by Telegram bot with inline-keyboard categories. A fixed Aho-Corasick lexicon sets priority at ingestion — priority is a rule, not a model output. The LLM only writes headlines and routing hints.

• §5

  Priority-aware egress queue with an isolated emergency lane

  Compound index on (status, priority_tier, created_at). Emergency Express dispatches Tier-1 tickets one at a time; Standard Batch claims up to 25 Tier-2 rows with per-request AbortSignals so one hung connection can't block the batch.

• §4

  No-human-review image moderation

  Cloudflare edge hash-matching, then a quantized NSFW/violence model on ONNX Runtime WASM. Images resolve to broadcast or delete — there is no pending_review state, and no human ever views a flagged image.

• §5 · §6 · §7

  Claim-and-lease workers with a per-ticket SLA scheduler

  One-off ctx.scheduler.runAfter(60_000, …) at ingestion — not a periodic cron — because cron minimum granularity would turn a 60-second SLA into a 60-120s window. A reaper reverts stuck rows and dead-letters at retry_count >= 3.

• §6

  Public dashboard with honest time-to-resolution

  True TTR includes queue delay; System Backpressure Latency isolates transport time. Every broadcast is prefixed with an immutable ticket ID so at-least-once retries are recognisable as duplicates rather than second incidents.