Patterns for routing inbound tickets, messages, or requests to the right queue or action path. Related paid starter: T1 Ticket routing. Cases #1 (Cloudflare) and #16 (Vercel Gateway) both teach department Choice — keep both sources; avoid treating them as conflicting products.
Case card detail fields stay in English (source language). Titles and chapter shells are localized.
Case cards
1 · Title
#1 Support department routing (account / billing / technical / other)
2 · When to use / When not
When to use
Route a single inbound support message to one department when labels are fixed and mutually exclusive.
When not
Do not use alone when you also need urgency or frustration in the same pass — prefer the parallel multi-question pattern (#2). Avoid inventing departments not in your org chart.
3 · Question stem
Which team should handle this support request?
4 · Options
account
billing
technical
other
5 · Criteria
account
Login, password, profile, or security issues
billing
Charges, invoices, refunds, or subscriptions
technical
Product bugs, outages, or integrations
other
Requests that do not fit the other departments
6 · Example in → out
In
I cannot log in after changing my password, and the reset email never arrives.
Out
department → account (confidence 1.0 in CF sample response)
From Cloudflare Workers AI Jev docs “Support department routing” example. Same Choice teaching as Gateway #16 (billing/shipping/technical) — keep both sources; do not duplicate long prose. Secondary: DEV practical guide (maps original #1) — short link only, not a re-recorded B card.
1 · Title
#2 Parallel urgency + department + frustration
2 · When to use / When not
When to use
Score several independent judgments on one ticket in a single call (Noul + Choice + Score).
When not
Skip when you only need one department label (#1). Do not hide urgency inside department criteria.
3 · Question stem
Multiple: Does this convey urgency? Which team should handle this? How frustrated is the customer?
4 · Options
is_urgent (noul)
department: billing | technical | sales
frustration Score: Calm → Frustrated → Very angry
5 · Criteria
is_urgent.true
Explicitly time-sensitive
is_urgent.false
No urgency expressed
billing
Payments, invoicing, refunds
technical
Bugs, outages, integrations
sales
Pricing, upgrades, new accounts
frustration
Calm; Frustrated; Very angry (ordered Score levels)
Cloudflare docs Usage example — three questions answered in one parallel pass. Secondary: jamilxt DEV post (parallel department/tone/urgency-shaped example) — short link only.
1 · Title
#5 Confidence-gated action routing
2 · When to use / When not
When to use
Gate side effects by per-action confidence: low bar for read-only actions; high bar before money-moving or irreversible actions.
When not
Do not use a single global threshold for every action. Do not treat confidence as a guarantee on one sample.
3 · Question stem
What is the user trying to do? (intent Choice)
4 · Options
check_balance
approve_transfer
(else → human)
5 · Criteria
check_balance
View the account balance (read-only, low bar once above floor)
approve_transfer
Approve the pending withdrawal / transfer (high confidence bar in source examples)
_routing
If confidence < 0.5 → human; check_balance may auto; approve_transfer only if confidence above high bar (DEV ~0.85 / Flavio ~0.9) else confirm
6 · Example in → out
In
(pattern) user message classified as check_balance vs approve_transfer
Out
Low confidence → route_to_human; high-confidence check_balance → show_balance; approve_transfer only above high bar else ask to confirm
DEV “Pattern 2: Confidence-gated routing”. Option labels and threshold numbers taken from the article’s code samples — calibrate on your labeled traffic. Secondary: puffball1567 DEV comparison. NO automatic money transfer — confidence gates suggestions only; code + humans own irreversible/money actions.
1 · Title
#6 Cascade intent → code / LLM / human
2 · When to use / When not
When to use
Front a support flow with cheap Jev classification, then branch to pure code, a specialist LLM, or a human.
When not
Not a replacement for frontier models on hard reasoning; not for generating replies.
3 · Question stem
Primary intent of this message + how complex is this to resolve
DEV “Pattern 1: speculative fan-out”. Cookbook speed/cost multiples are TypeSafe’s reported figures. Secondary: Cloudflare support-routing docs (related family) — short link, not a duplicate long card.
1 · Title
#15 Sponsorship form sorting
2 · When to use / When not
When to use
Triage structured sponsor / newsletter inquiry forms: real sponsor ask?, product category, how specific the request is.
When not
Do not auto-send rate cards or accept sponsors from model output alone — Flavio’s own plan starts in shadow / email annotation mode.
3 · Question stem
Does `description` ask to sponsor the site or newsletter? What kind of product? How specific is the request?
4 · Options
is_sponsor_inquiry (noul)
product_category: dev_tool | course | unrelated
message_quality Score (3 levels)
5 · Criteria
is_sponsor_inquiry
Does `description` ask to sponsor the site or newsletter?
dev_tool
Developer tools, hosting, APIs, SaaS for developers
course
Courses, books, or training
unrelated
Anything not aimed at developers
message_quality
Generic template, no reference to this site → Mentions the site but no concrete ask → Concrete ask with a timeframe or product named
6 · Example in → out
In
name: Managed Postgres; description: We make a managed PostgreSQL hosting product and would like to sponsor the newsletter in October.
Out
is_sponsor_inquiry high; product_category=dev_tool; message_quality near top level (illustrative shape in article)
Flavio Copes deep-dive opening example. Article notes sample answer numbers are illustrative; shape is exact. Short note: Flavio’s Playground includes helpdesk-ticket routing walkthroughs — companion pointer, not a second long card.
Vercel AI Gateway Evaluation docs Choice example. Shares department-Choice teaching with CF #1; keep both sources, avoid duplicated long prose. Secondary: Vercel changelog (Jev on AI Gateway).
1 · Title
#N3 Model routing Choice: fast vs powerful
2 · When to use / When not
When to use
Use Jev Choice middleware to pick a cheaper/faster model vs a more capable one from the latest user message criteria you define.
When not
Not for generating the answer itself — only selecting which generative model should run. Calibrate criteria to your tasks.
3 · Question stem
Choose the least costly model that can complete the task. (ModelRouterMiddleware instructions in article)
4 · Options
fast
powerful
5 · Criteria
fast
Direct lookups, extraction, and localized changes.
Same LangChain harness article as N2 — Model routing section. Option labels and criteria strings copied from the published sample; not a duplicate long B card.
Plain-language web search where Jev chooses sources/time/query shape and scores each result for subject relevance; app merges and ranks — no generated answers.
When not
Independent Search1API project, not an official TypeSafe product. Engine coverage and model choices can vary; failed engines warn rather than zero-count.
3 · Question stem
Does the request ask for recent results, and if so how recent? + per-source Nouls (e.g. Would Reddit threads fit…?) + Which candidate is the best keyword query…? + Is results[i] about the subject the user asked for…?
superagents-lab/jev-search. Source ask strings live in src/lib/sources.ts; rerank Noul in typesafe.ts. Try the live app; self-host needs Search1API + TypeSafe keys.
Local-first router that turns models, subagents, skills, MCP tools, CLIs, and plugins into one candidate set, asks Jev one typed Choice, then applies hard policy (availability, permissions, risk, confirmation).
When not
Jev owns probabilities; JevRouter owns policy — filtered candidates are never re-normalized. Decision-only by default; MCP adapter exposes jev_route and does not execute the selected capability implicitly.
3 · Question stem
Which single capability should handle this request? Choose only from the supplied options.
4 · Options
Choice over capability ids (criteria = capability descriptions from the candidate set)
5 · Criteria
_default
DEFAULT_QUESTION_INSTRUCTIONS in provider: Which single capability should handle this request? Choose only from the supplied options.
_two_stage
If candidates exceed single_stage_max_candidates: coarse Top-K then final Choice; raw_jev_stages keeps both responses
BillionsBobby/JevRouter (+ jevrouter.co). Benchmark snapshot in README is author Toolathlon routing-accuracy experiment — not end-to-end task completion.
1 · Title
#N11 kody v2026.09.18 — Jev Score search rerank
2 · When to use / When not
When to use
After hybrid lexical+vector recall in kody MCP search, optionally Score-rerank skinny candidate cards with Workers AI typesafe/jev; fall back to hybrid order on failure or low mean confidence.
When not
Feature-flagged (jev-search-rerank). Offline/deterministic embedding paths never call Jev. Dropping below keep score must not empty the result set.
3 · Question stem
How relevant is state.candidates[${String(index)}] to state.query for the agent's next hop (open detail or execute)? Use state.intent only as context.
4 · Options
per-candidate Score (4 levels)
5 · Criteria
0
Unrelated or misleading for this query
1
Tangentially related
2
Clearly relevant next hop
3
Best primary match for this query
_policy
jevSearchMinKeepScore 1.5; jevSearchMinMeanConfidence 0.45; candidate cap 40; wide recall ≥50 when flag on
Stems/criteria from packages/worker/.../search-jev-rerank.ts on tag v2026.09.18. Kody is a full assistant platform — this card is only the Jev rerank slice.