@warmdrift/kgauto-check 0.2.0 → 0.4.0

package/README.md

@@ -48,7 +48,7 @@ npx -y @warmdrift/kgauto-check --extract-contract   # extraction instructions +
 npx -y @warmdrift/kgauto-check --extract extract.json   # prints the findings card, fully offline
 
 # Standing-value one-shots (also offline):
-npx -y @warmdrift/kgauto-check --model-facts claude-sonnet-5   # one model → status, verified prices, limits, cliffs
+npx -y @warmdrift/kgauto-check --model-facts claude-sonnet-5   # one model → the provider's word, verified prices, limits, cliffs
 npx -y @warmdrift/kgauto-check --observe-plan                  # rung 2: the one-day observe run
 ```
 
@@ -65,12 +65,51 @@ npx -y @warmdrift/kgauto-check --observe-plan                  # rung 2: the one
 | `get_model_facts` | One model id → status, verified prices, limits, capabilities, known cliffs. Honest miss on unknown models. |
 | `get_observe_day_plan` | Rung 2: what a one-day observe-only run on real traffic would measure, and the two honesty requirements any setup must carry. |
 
+## An example card
+
+This is the output of `kgauto-check 0.4.0` on a small illustrative extract (two call sites, one on the provider's own API and one on Vertex AI, and a model picker), printed by the real build — not a mock-up:
+
+````markdown
+# Communicating Vessels — the check: example-app
+
+2 AI call site(s), 3 model id(s): `anthropic/claude-sonnet-4-20250514` (looked up as `claude-sonnet-4-20250514`), `gemini-2.5-flash`, `gpt-4o-mini` (3 flagged by its provider)
+
+## Certain — read from the providers’ pages and your code
+
+- 🔴 **On Anthropic's own API, `anthropic/claude-sonnet-4-20250514` (looked up as `claude-sonnet-4-20250514`) was retired on 2026-06-15; its page names `claude-sonnet-5-5`** — Anthropic's deprecations page (https://platform.claude.com/docs/en/about-claude/model-deprecations.md): “The model is no longer available for use. Requests to retired models fail.” The catalog has no entry for it, so its limits and price were not checked.
+  <sub>app/api/chat/route.ts:41 · basis: Anthropic's page, read 2026-11-12: https://platform.claude.com/docs/en/about-claude/model-deprecations.md</sub>
+- 🟡 **On OpenAI's own API, `gpt-4o-mini` is deprecated and shuts down on 2027-01-14; its page names `gpt-6-luna`** — The page's line: “| January 14, 2027 | GPT-4o mini (`gpt-4o-mini-2024-07-18`) | `gpt-6-luna` |”. The catalog has no entry for it, so its limits and price were not checked.
+  <sub>components/ModelPicker.tsx:12 · basis: OpenAI's page, read 2026-11-12: https://developers.openai.com/api/docs/deprecations.md (and https://developers.openai.com/api/docs/models/gpt-4o-mini.md)</sub>
+
+## Likely — priced from your estimates, or a catalog value no provider page is cited for
+
+- 🟡 **On Google's own API, `gemini-2.5-flash` is deprecated; from 2026-12-01 its requests route to `gemini-3.5-flash-lite`** — The page's line: “`gemini-2.5-flash` | June 17, 2025 | December 1, 2026 | `gemini-3.5-flash-lite` (`gemini-2.5-flash` is no longer available in Google AI Studio; effective December 1, 2026, API requests to `gemini-2.5-flash` will automatically route to `gemi”. This site runs on Vertex AI, and Vertex AI runs its own schedule (its page: https://cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions).
+  <sub>lib/summarize.ts:18 · basis: Google's page, read 2026-11-12: https://aistudio.google.com/docs/deprecations; Vertex AI: https://cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions</sub>
+- ℹ️ **~$0.0046/call on `gemini-2.5-flash`** — From your token estimates and doc-verified prices ($0.3/M in, $2.5/M out, verified 2026-08-14).
+  <sub>lib/summarize.ts:18 · basis: arithmetic on your estimates × verified catalog prices</sub>
+- 🟡 **97% of this site's tokens are input** — Above ~85% input share, spend is dominated by context rather than answers — the single strongest signal of a leaking prompt. What's actually IN that context, and how much of it each call needed, is measurable, not guessable.
+  <sub>lib/summarize.ts:18 · basis: your token estimates; the >85% threshold is a measured portfolio heuristic</sub>
+
+## Only measurable — what the observe day answers
+
+- ℹ️ **1 model(s) in the catalog price under half of `gemini-2.5-flash` on summarize work** — Cheapest: `claude-haiku-5-5` at ~$0.0014/call. The price is arithmetic; whether any of them holds YOUR quality bar on summarize work is exactly what a one-day observe run with a judged eval measures. The check never ranks a model on price alone.
+  <sub>lib/summarize.ts:18 · basis: verified catalog prices; quality deliberately unclaimed</sub>
+- ℹ️ **what a code read cannot tell you (and a one-day observe run can)** — Your actual failure rate per model (walkaway failures look like green health checks); your actual input-waste ratio per call; which call sites dominate spend; whether a cheaper model holds quality on your real traffic. These need one day of observed calls — observe-only, your keys, no routing changes.
+  <sub>app/api/chat/route.ts:41, lib/summarize.ts:18, components/ModelPicker.tsx:12 · basis: the honest limit of static analysis</sub>
+
+---
+*This check ran on your machine, against the catalog bundled with it and the providers’ deprecation pages it was built with. Nothing was uploaded; no rows were written anywhere.*
+<sub>@warmdrift/kgauto-check 0.4.0 · catalog @warmdrift/kgauto-compiler 2.0.0-alpha.132 · deprecation pages read 2026-11-12</sub>
+````
+
 ## The confidence tiers (why the card is grouped)
 
-- **certain** — a catalog or code fact. Stated flatly.
-- **likely** — priced from your token estimates against verified prices.
+- **certain** — the provider's own page (with its URL and the day it was read), or a fact from your code. Stated flatly.
+- **likely** — priced from your token estimates against verified prices, or a catalog value no provider page is cited for.
 - **measurable-only** — honestly unknowable from a code read: actual failure rates, actual input-waste, spend distribution, quality-at-cheaper-price. Each of these is exactly what the one-day observe run answers.
 
+**The provider's word first (0.4.0).** A model id is matched as you wrote it (`openai/gpt-5`, `openrouter/anthropic/…`, a dated snapshot, a Bedrock or Vertex form) and looked up in a lifecycle ledger read from the four providers' own deprecation pages (Anthropic, OpenAI, Google, DeepSeek) before the catalog. A retired, shut-down, deprecated or legacy id says so, with the date, the replacement the page names, and the page's URL.
+
 kgauto refuses to fabricate verdicts: unknown models get an honest "no measured evidence", and cheaper candidates are named with their price arithmetic only — never with an unmeasured quality claim.
 
 ## Client setup variants

package/dist/bin.js

@@ -7,7 +7,7 @@ import {
   renderModelFacts,
   runCheck,
   startStdio
-} from "./chunk-BWPYT7JO.js";
+} from "./chunk-XWIGHIJP.js";
 
 // src/bin.ts
 import { readFileSync } from "fs";