TOKENCRAFT·notebook·v1.4.7·$0.05/promote⌘ ⏎ run cell

Version your prompts the way you version your libraries — and let the eval block the promote.

TokenCraft is a Lab-Notebook IDE for prompt ops. Every cell is a semver version (v1.4.6 → v1.4.7). Every token carries a violet USD subscript. The right rail rebalances four provider costs as you type. The promote button stays locked until HumanEval clears HEAD − 5%.

open /studio$0.05 per promote
cell #4·v1.4.7HEADcustomer-support-tier1.prompt
You are a {{role}}$.00012 agent.
Reply in <= {{max_chars}}$.00009 chars.
<jinja>{{policy}}$.00018</jinja>
47 tokens·$0.00031/call·context 4.7%PASS 91.2%drift +0.4σ
providers · projected per call · v1.4.7
  • openai$0.00084openai gpt-4o-mini · per call
  • anthropic$0.00064anthropic claude-haiku · per call
  • google$0.00071google gemini-flash · per call
  • llama-70b$0.00018fireworks llama-70b · per call
K1.8

Semver promote with eval gate

Patch/minor/major bumps per change-type. Promotion is blocked when HumanEval drops more than 5 percent versus HEAD. Override requires a typed reason — it does not slip through.

K1.4

Token-budget enforcer mid-stream

Set --max-cost-cents per cell. The SSE stream terminates the moment projected cost crosses the cap. No $1,000-surprise-bill scenarios.

K6.2

OTLP traces to Honeycomb

Every cell-run emits a distributed trace: tokenizer → eval API → 4-provider projector → ledger rebalance. Deep link in the verdict block — no copy-paste needed.

persona · prompt engineer · ships releases, not chats

You speak in semver. You read drift in sigmas. You think in BPE tokens. You will not promote v1.4.7 if HumanEval drops more than five points against HEAD. TokenCraft is the IDE that respects that discipline by enforcing it.