Optimization flow UI, schema relaxation, and in-memory KV

.env.example

@@ -22,6 +22,12 @@ TELEMETRY_LLM_PREVIEW_CHARS=800
 # Comma-separated list of allowed origins for the API (set in production).
 CORS_ALLOWED_ORIGINS=http://localhost:8000
 
+# ─────────────────────────────────────────────────
+# Data
+# ─────────────────────────────────────────────────
+# Limit number of epics loaded (useful for fast test runs).
+# EPICS_LIMIT=1
+
 # ─────────────────────────────────────────────────
 # Generation Settings
 # ─────────────────────────────────────────────────

.github/workflows/deploy-deno.yml

@@ -40,6 +40,7 @@ jobs:
           deployctl deploy \
             --project="${{ vars.DENO_DEPLOY_PROJECT }}" \
             --prod \
+            --include=src/ui/dist \
             --exclude=node_modules \
             --exclude=.git \
             --exclude=.github \

.github/workflows/deploy-staging.yml

@@ -40,6 +40,7 @@ jobs:
           deployctl deploy \
             --project="${{ vars.DENO_DEPLOY_PROJECT_STAGING }}" \
             --prod \
+            --include=src/ui/dist \
             --exclude=node_modules \
             --exclude=.git \
             --exclude=.github \

deno.json

@@ -2,6 +2,7 @@
   "imports": {
     "@std/assert": "jsr:@std/assert@^1.0.16",
     "@std/expect": "jsr:@std/expect@^1.0.17",
+    "@std/dotenv": "jsr:@std/dotenv@0.225.6",
     "@std/dotenv/load": "jsr:@std/dotenv@0.225.6/load",
     "@std/http/file-server": "jsr:@std/http@^1.0.23/file-server",
     "@std/path": "jsr:@std/path@^1.1.4",

playwright.config.ts

@@ -1,5 +1,34 @@
 import { defineConfig, devices } from "@playwright/test";
+import { parse } from "@std/dotenv";
 
+const readEnvFile = (): Record<string, string> => {
+  try {
+    const raw = Deno.readTextFileSync(".env");
+    return parse(raw);
+  } catch (e) {
+    if (e instanceof Deno.errors.NotFound) {
+      return {};
+    }
+    throw e;
+  }
+};
+
+const envFromFile = readEnvFile();
+const getEnv = (key: string, fallback?: string) =>
+  Deno.env.get(key) ?? envFromFile[key] ?? fallback ?? "";
+
+const llmBaseUrl = getEnv("LLM_BASE_URL", "https://openrouter.ai/api/v1");
+const llmApiKey = getEnv("LLM_API_KEY");
+const llmModel = getEnv("LLM_MODEL", "openai/gpt-4o-mini");
+
+// Warn when targeting a remote LLM provider without an API key
+const isLocalhost = /^https?:\/\/(localhost|127\.0\.0\.1)(:\d+)?/.test(llmBaseUrl);
+if (!llmApiKey && !isLocalhost) {
+  console.warn(
+    `[playwright] LLM_API_KEY is empty while LLM_BASE_URL points to a remote provider (${llmBaseUrl}). ` +
+    "LLM calls will likely fail with 401. Set LLM_API_KEY in .env or environment.",
+  );
+}
 export default defineConfig({
   testDir: "./tests/e2e",
   timeout: 60_000,
@@ -18,10 +47,16 @@ export default defineConfig({
     reuseExistingServer: !Deno.env.get("CI"),
     timeout: 120_000,
     env: {
-      LMSTUDIO_BASE_URL: "http://127.0.0.1:1234/v1",
-      LMSTUDIO_API_KEY: "lm-studio",
-      LMSTUDIO_MODEL: "gpt-oss-120b",
-      LMSTUDIO_JUDGE_MODEL: "gpt-oss-120b",
+      ...Deno.env.toObject(),
+      LLM_BASE_URL: llmBaseUrl,
+      LLM_API_KEY: llmApiKey,
+      LLM_MODEL: llmModel,
+      POLL_ENABLED: "false",
+      EVAL_REPLICATES: "1",
+      OPT_CONCURRENCY: "3",
+      OPT_ITERATIONS: "1",
+      OPT_PATCH_CANDIDATES: "1",
+      EPICS_LIMIT: "1",
     },
   },
   projects: [

prompts/champion.base.md

@@ -8,13 +8,14 @@ Rules:
 3. Each story MUST include:
    - title (short, action-oriented)
    - asA / iWant / soThat
-   - acceptanceCriteria: >= 2 items, objectively testable
+   - acceptanceCriteria: >= 1 item, objectively testable
 4. Prefer acceptance criteria in Given/When/Then style.
-5. Do NOT invent requirements. If something is unclear, put it in assumptions or
+5. Keep output compact; omit optional fields unless needed.
+6. Do NOT invent requirements. If something is unclear, put it in assumptions or
    followUps.
-6. Reflect constraints/nonFunctional/outOfScope from the Epic.
+7. Reflect constraints/nonFunctional/outOfScope from the Epic.
 
-Azure DevOps mapping:
+Azure DevOps mapping (optional if requested):
 
 - System.Title: story title
 - System.Description: include As a / I want / So that in readable Markdown

prompts/champion.md

@@ -8,13 +8,14 @@ Rules:
 3. Each story MUST include:
    - title (short, action-oriented)
    - asA / iWant / soThat
-   - acceptanceCriteria: >= 2 items, objectively testable
+   - acceptanceCriteria: >= 1 item, objectively testable
 4. Prefer acceptance criteria in Given/When/Then style.
-5. Do NOT invent requirements. If something is unclear, put it in assumptions or
+5. Keep output compact; omit optional fields unless needed.
+6. Do NOT invent requirements. If something is unclear, put it in assumptions or
    followUps.
-6. Reflect constraints/nonFunctional/outOfScope from the Epic.
+7. Reflect constraints/nonFunctional/outOfScope from the Epic.
 
-Azure DevOps mapping:
+Azure DevOps mapping (optional if requested):
 
 - System.Title: story title
 - System.Description: include As a / I want / So that in readable Markdown

src/config.ts

@@ -102,6 +102,12 @@ const EnvSchema = z.object({
   // ─────────────────────────────────────────────────
   CORS_ALLOWED_ORIGINS: z.string().default(""),
 
+  // ─────────────────────────────────────────────────
+  // Data
+  // ─────────────────────────────────────────────────
+  /** Limit number of epics loaded (useful for fast test runs) */
+  EPICS_LIMIT: z.coerce.number().int().min(1).max(100).optional(),
+
   // ─────────────────────────────────────────────────
   // Generation Settings
   // ─────────────────────────────────────────────────

src/fpf/poll.ts

@@ -2,7 +2,7 @@
  * PoLL - Panel of LLM Evaluators
  *
  * Implements FPF B.3 Trust & Assurance Calculus with:
- * - 3 diverse judges (different temperatures for diversity)
+ * - N diverse judges (env-configured count and temperatures)
  * - Per-criterion evaluation (decomposed INVEST + GWT)
  * - WLNK aggregation: R_eff = max(0, min(R_i) - Φ(CL_min))
  * - Full SCR audit trail
@@ -38,26 +38,23 @@ import {
 // CONFIGURATION
 // ═══════════════════════════════════════════════════════════════
 
-const DEFAULT_JUDGES: JudgeConfig[] = [
-  {
-    id: "judge-1",
-    model: "gpt-4o-mini",
-    temperature: 0.3,
-    provider: "lmstudio",
-  },
-  {
-    id: "judge-2",
-    model: "gpt-4o-mini",
-    temperature: 0.5,
-    provider: "lmstudio",
-  },
-  {
-    id: "judge-3",
-    model: "gpt-4o-mini",
-    temperature: 0.7,
+const clampTemp = (value: number) => Math.max(0, Math.min(2, value));
+
+const buildDefaultJudges = (): JudgeConfig[] => {
+  const model = env.LMSTUDIO_JUDGE_MODEL ?? env.LMSTUDIO_MODEL;
+  const baseTemp = env.POLL_TEMP_BASE;
+  const spread = env.POLL_TEMP_SPREAD;
+  const count = env.POLL_NUM_JUDGES;
+
+  return Array.from({ length: count }, (_, index) => ({
+    id: `judge-${index + 1}`,
+    model,
+    temperature: clampTemp(baseTemp + index * spread),
     provider: "lmstudio",
-  },
-];
+  }));
+};
+
+const DEFAULT_JUDGES = buildDefaultJudges();
 
 const CRITERIA_WEIGHTS: Record<EvaluationCriterion, number> = {
   [EvaluationCriterion.CORRECTNESS]: 0.2,

src/generator.ts

@@ -9,7 +9,8 @@ import { parseAcceptanceCriteria } from "./utils/acceptanceCriteria.ts";
 export const baseStoryAgent = new Agent({
   id: "story-generator",
   name: "Story Generator",
-  instructions: "You generate Azure DevOps user stories from epics.",
+  instructions:
+    "You generate Azure DevOps user stories from epics. Always return JSON that matches the provided schema and keep the response compact.",
   model: makeGeneratorModel(),
 });
 
@@ -37,6 +38,14 @@ type ValidationFailure = {
   issues: string[];
 };
 
+const SCHEMA_GUARDRAILS = [
+  "Output MUST match the provided JSON schema.",
+  "Keep the response compact; omit optional fields unless needed.",
+  "Include `acceptanceCriteria` with at least 1 item per story.",
+  "If you include `ado.fields`, keep each value brief.",
+  "Do not omit or rename required schema fields.",
+].join("\n");
+
 /**
  * Helper to build provider-specific options including seed.
  * LM Studio accepts `seed` in the OpenAI-compatible API.
@@ -324,6 +333,8 @@ export async function generateStoryPack(
     },
   ];
 
+  const guardedInstructions = `${candidatePrompt}\n\nSchema guardrails:\n${SCHEMA_GUARDRAILS}`;
+
   let storyPack: StoryPack | null = null;
   let rawText = "";
   let error: string | undefined;
@@ -339,7 +350,7 @@ export async function generateStoryPack(
       { name: "story-generator", model: env.LMSTUDIO_MODEL },
       () =>
         baseStoryAgent.generate(messages, {
-          instructions: candidatePrompt,
+          instructions: guardedInstructions,
           structuredOutput: {
             schema: storyPackSchema,
             jsonPromptInjection: true,
@@ -441,7 +452,7 @@ export async function generateStoryPack(
   return {
     storyPack,
     rawText,
-    instructions: candidatePrompt,
+    instructions: guardedInstructions,
     trace,
     gammaTime: startedAt,
     seed,