@plumbus/ai-decision-laya 0.2.1 → 0.3.0-beta.0

package/instructions/README.md

@@ -5,8 +5,13 @@
 | [framework.md](framework.md) | Installation, public exports, adapter boundary, usage rules |
 | [testing.md](testing.md) | Offline validation and live environment setup |
 
-Keep app business logic in Plumbus primitives and `ctx.*`. These packages do not
-add `ctx.ai.decide()` or automatic provider discovery yet. Never route model
-predictions directly around capability access checks or confirmation policy.
+**Classification, categorization, provider/model selection, TypeSafe/Jev or Laya:**
+read `node_modules/@plumbus/core/instructions/ai-classification.md` first (core
+**0.8.0-beta.6+**), then [framework.md](framework.md) for this package.
+
+Core 0.8.0-beta.6+ provides `ctx.ai.decide()` and also routes `ctx.ai.classify()`
+to explicitly registered decision adapters. Keep app business logic in Plumbus
+primitives and `ctx.*`. Never route predictions around capability access checks
+or confirmation policy. Provider registration is explicit in `app/server.ts`.
 
 [Package README](../README.md) · [Monorepo guide](https://github.com/plumbus-framework/plumbus/blob/main/docs/ai/decision-providers.md)

package/instructions/framework.md

@@ -4,7 +4,7 @@ Laya provider for self-hosted typed decision inference, with a persistent Python
 
 ## Install and ownership
 
-Install `@plumbus/ai-decision-laya` explicitly. Its core peer is exactly `"0.7.x"`.
+Install `@plumbus/ai-decision-laya` explicitly. Its core peer is exactly `">=0.8.0-beta.0 <0.9.0"` (core 0.8 beta family).
 Consumers reuse Zod/testing utilities from core; do not install them separately.
 Use this package for bounded typed decisions. Use core generation and embeddings
 for text responses and vector inference.
@@ -14,7 +14,19 @@ Public exports: `createLayaDecisionAdapter` — construct an explicitly configur
 File map: `src/index.ts` is the public barrel, `src/__tests__/` exercises the public
 API, `instructions/` contains these recipes. Laya alone also ships `service/`.
 
-## Minimal usage
+## Classification and model selection (core 0.8.0-beta.6+)
+
+Read `node_modules/@plumbus/core/instructions/ai-classification.md` for the complete
+registration and usage recipe. Call `ctx.ai.classify({ text, labels, provider, model,
+threshold })` inside capabilities/flows. `provider` is your registered decision map
+key; `model` is an optional endpoint-supported ID. Native `threshold` defaults to
+0.5 and returns all matching labels. Use `decide()` for richer questions.
+
+Core records one `classify` cost row, including dispatched failures, with actual
+model/usage and identity. Omitted `provider` still uses the text default. Keep
+decision adapters under `decisions.providers`, separate from text providers.
+
+## Minimal usage (infrastructure / smoke testing)
 
 ```ts
 import { createLayaDecisionAdapter } from '@plumbus/ai-decision-laya';
@@ -33,7 +45,7 @@ console.log(result.answers.refund.probability);
 1. Keep capabilities, flows, access policy and `ctx.*` as the application implementation path.
 2. Construct adapters at a server-owned infrastructure boundary and inject them. Keep keys on the server.
 3. Treat direct adapter usage as protocol integration, not a replacement for core's security/budget/audit lifecycle.
-4. With core 0.7.3+, use `ctx.ai.decide()` in capabilities/flows. Export `decisions = { providers, defaultProvider }` from `app/server.ts`; API and workers share it. Do not register decision adapters as completion providers.
+4. With core 0.8.0-beta.6+, use `ctx.ai.decide()` in capabilities/flows. Export `decisions = { providers, defaultProvider }` from `app/server.ts`; API and workers share it. Do not register decision adapters as completion providers.
 5. Preserve distributions and provider confidence separately. A high confidence value is not authorization or proof of correctness.
 6. Validate thresholds on the actual task/model/language. Unknown cost is `null`, not free.
 7. Pass `signal: ctx.signal` and `timeoutMs` where needed. Flow steps supply their signal by default.

package/instructions/testing.md

@@ -18,7 +18,12 @@ Validate choices, scores, probabilities, actual model/routing, usage and unknown
 cost handling. Measure live accuracy and calibration separately from connection
 success. Do not assert a universal confidence cutoff or guaranteed live label.
 
-For core 0.7.3+ integration tests, configure `mockAI({ decide: result })` or a real
+For core 0.8.0-beta.6+ classification, use `mockAI({ classify: ['billing'] })` and test
+through `runCapability()` / `simulateFlow()`. Integration tests should assert
+provider/model forwarding, zero/multiple matches, inclusive thresholds, and one
+`classify` cost row. Use injected fetch; routine tests do not require live Laya.
+
+For core 0.8.0-beta.6+ integration tests, configure `mockAI({ decide: result })` or a real
 `createAIService({ decisions, costTracker, onAICostRecorded })` with injected
 provider fetch. Assert exactly one cost row, identity, actual model/usage, unknown
 cost handling, billed validation failures, and budget rejection before dispatch.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@plumbus/ai-decision-laya",
-  "version": "0.2.1",
+  "version": "0.3.0-beta.0",
   "description": "Self-hosted Laya typed decision adapter and Python reference service for Plumbus.",
   "type": "module",
   "main": "./dist/index.js",
@@ -36,17 +36,17 @@
     "node": ">=20.6.0"
   },
   "peerDependencies": {
-    "@plumbus/core": "0.7.x"
+    "@plumbus/core": ">=0.8.0-beta.0 <0.9.0"
   },
   "devDependencies": {
     "@biomejs/biome": "^2.4.6",
     "@types/node": "^25.4.0",
     "typescript": "^5.7.0",
     "vitest": "^3.2.4",
-    "@plumbus/core": "0.7.3"
+    "@plumbus/core": "0.8.0-beta.6"
   },
   "dependencies": {
-    "@plumbus/ai-decision": "~0.2.1"
+    "@plumbus/ai-decision": "~0.3.0-beta.0"
   },
   "scripts": {
     "build": "tsc -b",

package/service/test_audit_round_two.py

@@ -43,6 +43,14 @@ class ServiceRoundTwo(unittest.TestCase):
         finally:
             conn.close()
 
+    def refused(self, address):
+        # Over capacity the server answers at accept without reading a request, so a client
+        # still sending one can be reset before it reads the 503. Connect and only read.
+        with socket.create_connection(address, timeout=2) as sock:
+            response = http.client.HTTPResponse(sock)
+            response.begin()
+            return response.status, json.loads(response.read())
+
     def test_L23_missing_length_cannot_start_inference(self):
         self.assertEqual(self.send(length=None)[0], 400)
         self.assertEqual(self.backend.calls, [])
@@ -108,7 +116,7 @@ class ServiceRoundTwo(unittest.TestCase):
         first.start()
         try:
             self.assertTrue(entered.wait(1))
-            self.assertEqual(self.send(address=server.server_address)[0], 503)
+            self.assertEqual(self.refused(server.server_address)[0], 503)
             finish.set()
             first.join(2)
             self.assertEqual(results, [200])