selftune 0.2.29 → 0.2.30

package/cli/selftune/dashboard-server.ts

@@ -19,12 +19,14 @@
  */
 
 import type { Database } from "bun:sqlite";
-import { existsSync, readFileSync, unwatchFile, watchFile } from "node:fs";
+import { existsSync, readFileSync, statSync, unwatchFile, watchFile } from "node:fs";
 import { dirname, extname, isAbsolute, join, relative, resolve } from "node:path";
 
 import type { BadgeFormat } from "./badge/badge-data.js";
-import { LOG_DIR, SELFTUNE_CONFIG_DIR } from "./constants.js";
+import { getCachedUpdateStatus } from "./auto-update.js";
+import { DASHBOARD_ACTION_STREAM_LOG, LOG_DIR, SELFTUNE_CONFIG_DIR } from "./constants.js";
 import type {
+  DashboardActionEvent,
   HealthResponse,
   OverviewResponse,
   SkillReportResponse,
@@ -53,6 +55,7 @@ import {
 import type { StatusResult } from "./status.js";
 import { computeStatus } from "./status.js";
 import type { EvolutionAuditEntry, EvolutionEvidenceEntry } from "./types.js";
+import { readJsonlFrom } from "./utils/jsonl.js";
 
 export interface DashboardServerOptions {
   port?: number;
@@ -72,6 +75,13 @@ interface DashboardSocketData {
   upstreamUrl?: string;
 }
 
+interface ActionEventHistoryEntry {
+  eventId: string;
+  updatedAt: number;
+  finished: boolean;
+  events: DashboardActionEvent[];
+}
+
 /** Read selftune version from package.json (fresh on each call to pick up auto-updates). */
 const VERSION_PKG_PATH = join(import.meta.dir, "..", "..", "package.json");
 function getSelftuneVersion(): string {
@@ -189,7 +199,10 @@ async function serveSpaShell(spaDir: string | null): Promise<Response> {
   if (!spaDir) {
     return new Response("Dashboard build not found. Run `bun run build:dashboard` first.", {
       status: 503,
-      headers: { "Content-Type": "text/plain; charset=utf-8", ...corsHeaders() },
+      headers: {
+        "Content-Type": "text/plain; charset=utf-8",
+        ...corsHeaders(),
+      },
     });
   }
 
@@ -260,9 +273,11 @@ function withCors(response: Response): Response {
   });
 }
 
-export async function startDashboardServer(
-  options?: DashboardServerOptions,
-): Promise<{ server: ReturnType<typeof Bun.serve>; stop: () => void; port: number }> {
+export async function startDashboardServer(options?: DashboardServerOptions): Promise<{
+  server: ReturnType<typeof Bun.serve>;
+  stop: () => void;
+  port: number;
+}> {
   const port = options?.port ?? 3141;
   const hostname = options?.host ?? "localhost";
   const openBrowser = options?.openBrowser ?? true;
@@ -321,12 +336,60 @@ export async function startDashboardServer(
 
   // -- SSE (Server-Sent Events) live update layer -----------------------------
   const sseClients = new Set<ReadableStreamDefaultController>();
+  const actionEventHistory = new Map<string, ActionEventHistoryEntry>();
+  const MAX_ACTION_HISTORY_RUNS = 24;
+  const MAX_ACTION_HISTORY_EVENTS_PER_RUN = 320;
+
+  function trimActionEventHistory(): void {
+    if (actionEventHistory.size <= MAX_ACTION_HISTORY_RUNS) return;
+
+    const staleEntries = [...actionEventHistory.values()].sort((left, right) => {
+      if (left.finished !== right.finished) {
+        return left.finished ? -1 : 1;
+      }
+      return left.updatedAt - right.updatedAt;
+    });
 
-  function broadcastSSE(eventType: string): void {
-    const payload = `event: ${eventType}\ndata: ${JSON.stringify({ type: eventType, ts: Date.now() })}\n\n`;
+    while (actionEventHistory.size > MAX_ACTION_HISTORY_RUNS) {
+      const next = staleEntries.shift();
+      if (!next) break;
+      actionEventHistory.delete(next.eventId);
+    }
+  }
+
+  function rememberActionEvent(event: DashboardActionEvent): void {
+    const existing = actionEventHistory.get(event.event_id);
+    if (existing) {
+      existing.updatedAt = event.ts;
+      existing.finished = event.stage === "finished" ? true : existing.finished;
+      existing.events.push(event);
+      existing.events = existing.events.slice(-MAX_ACTION_HISTORY_EVENTS_PER_RUN);
+      return;
+    }
+
+    actionEventHistory.set(event.event_id, {
+      eventId: event.event_id,
+      updatedAt: event.ts,
+      finished: event.stage === "finished",
+      events: [event],
+    });
+    trimActionEventHistory();
+  }
+
+  function recentActionEventsForBackfill(): DashboardActionEvent[] {
+    return [...actionEventHistory.values()]
+      .sort((left, right) => left.updatedAt - right.updatedAt)
+      .flatMap((entry) => entry.events);
+  }
+
+  function broadcastSSE(eventType: string, payload: Record<string, unknown>): void {
+    if (eventType === "action") {
+      rememberActionEvent(payload as DashboardActionEvent);
+    }
+    const message = `event: ${eventType}\ndata: ${JSON.stringify(payload)}\n\n`;
     for (const controller of sseClients) {
       try {
-        controller.enqueue(new TextEncoder().encode(payload));
+        controller.enqueue(new TextEncoder().encode(message));
       } catch {
         sseClients.delete(controller);
       }
@@ -347,9 +410,16 @@ export async function startDashboardServer(
   // -- SQLite WAL watcher for push-based updates ------------------------------
   const walPath = `${DB_PATH}-wal`;
   let walWatcherActive = false;
+  const actionStreamPath =
+    process.env.SELFTUNE_DASHBOARD_ACTION_STREAM_LOG || DASHBOARD_ACTION_STREAM_LOG;
+  let actionStreamWatcherActive = false;
+  let actionStreamOffset = existsSync(actionStreamPath) ? statSync(actionStreamPath).size : 0;
 
   let fsDebounceTimer: ReturnType<typeof setTimeout> | null = null;
+  let actionStreamDebounceTimer: ReturnType<typeof setTimeout> | null = null;
   const FS_DEBOUNCE_MS = 500;
+  const ACTION_STREAM_DEBOUNCE_MS = 100;
+  const ACTION_STREAM_POLL_MS = 250;
   const proxiedSpaSockets = new Map<unknown, WebSocket>();
 
   function onWALChange(): void {
@@ -357,15 +427,36 @@ export async function startDashboardServer(
     fsDebounceTimer = setTimeout(() => {
       fsDebounceTimer = null;
       refreshV2DataImmediate();
-      broadcastSSE("update");
+      broadcastSSE("update", { type: "update", ts: Date.now() });
     }, FS_DEBOUNCE_MS);
   }
 
   watchFile(walPath, { interval: 500 }, onWALChange);
   walWatcherActive = true;
 
+  function flushActionStream(): void {
+    if (actionStreamDebounceTimer) return;
+    actionStreamDebounceTimer = setTimeout(() => {
+      actionStreamDebounceTimer = null;
+      const { records, newOffset } = readJsonlFrom<DashboardActionEvent>(
+        actionStreamPath,
+        actionStreamOffset,
+      );
+      actionStreamOffset = newOffset;
+      for (const record of records) {
+        broadcastSSE("action", record);
+      }
+    }, ACTION_STREAM_DEBOUNCE_MS);
+  }
+
+  const actionStreamPoller = setInterval(() => {
+    flushActionStream();
+  }, ACTION_STREAM_POLL_MS);
+  actionStreamWatcherActive = true;
+
   function getWatcherMode(): HealthResponse["watcher_mode"] {
-    return walWatcherActive ? "wal" : "none";
+    if (walWatcherActive && actionStreamWatcherActive) return "wal";
+    return walWatcherActive || actionStreamWatcherActive ? "wal" : "none";
   }
 
   let cachedStatusResult: StatusResult | null = null;
@@ -454,10 +545,15 @@ export async function startDashboardServer(
 
       // ---- GET /api/health ----
       if (url.pathname === "/api/health" && req.method === "GET") {
+        const updateStatus = getCachedUpdateStatus();
         const healthResponse: HealthResponse = {
           ok: true,
           service: "selftune-dashboard",
           version: getSelftuneVersion(),
+          latest_version: updateStatus.latestVersion,
+          update_available: updateStatus.updateAvailable,
+          auto_update_supported: updateStatus.autoUpdateSupported,
+          update_hint: updateStatus.updateHint,
           pid: process.pid,
           spa: Boolean(spaDir || spaProxyUrl),
           spa_mode: spaMode,
@@ -503,6 +599,11 @@ export async function startDashboardServer(
           start(controller) {
             sseClients.add(controller);
             controller.enqueue(new TextEncoder().encode(": connected\n\n"));
+            for (const event of recentActionEventsForBackfill()) {
+              controller.enqueue(
+                new TextEncoder().encode(`event: action\ndata: ${JSON.stringify(event)}\n\n`),
+              );
+            }
           },
           cancel(controller) {
             sseClients.delete(controller);
@@ -533,7 +634,10 @@ export async function startDashboardServer(
             `Dashboard SPA proxy unavailable at ${spaProxyUrl.toString()}: ${message}`,
             {
               status: 502,
-              headers: { "Content-Type": "text/plain; charset=utf-8", ...corsHeaders() },
+              headers: {
+                "Content-Type": "text/plain; charset=utf-8",
+                ...corsHeaders(),
+              },
             },
           );
         }
@@ -544,7 +648,10 @@ export async function startDashboardServer(
         const filePath = resolve(spaDir, `.${url.pathname}`);
         const rel = relative(spaDir, filePath);
         if (rel.startsWith("..") || isAbsolute(rel)) {
-          return new Response("Not Found", { status: 404, headers: corsHeaders() });
+          return new Response("Not Found", {
+            status: 404,
+            headers: corsHeaders(),
+          });
         }
         const bunFile = Bun.file(filePath);
         if (await bunFile.exists()) {
@@ -558,7 +665,10 @@ export async function startDashboardServer(
             },
           });
         }
-        return new Response("Not Found", { status: 404, headers: corsHeaders() });
+        return new Response("Not Found", {
+          status: 404,
+          headers: corsHeaders(),
+        });
       }
 
       // ---- GET / ---- Serve SPA shell
@@ -597,7 +707,10 @@ export async function startDashboardServer(
             { status: 400, headers: corsHeaders() },
           );
         }
-        return withCors(await handleAction(action, body, executeAction));
+        const emitActionEvent = (event: DashboardActionEvent) => {
+          broadcastSSE("action", event);
+        };
+        return withCors(await handleAction(action, body, executeAction, emitActionEvent));
       }
 
       // ---- GET /badge/:skillName ----
@@ -634,7 +747,9 @@ export async function startDashboardServer(
       // ---- GET /api/v2/overview ----
       if (url.pathname === "/api/v2/overview" && req.method === "GET") {
         if (getOverviewResponse) {
-          return Response.json(getOverviewResponse(), { headers: corsHeaders() });
+          return Response.json(getOverviewResponse(), {
+            headers: corsHeaders(),
+          });
         }
         if (!db) {
           return Response.json(
@@ -737,6 +852,7 @@ export async function startDashboardServer(
   const shutdownHandler = () => {
     unwatchFile(walPath, onWALChange);
     clearInterval(sseKeepaliveTimer);
+    clearInterval(actionStreamPoller);
     for (const c of sseClients) {
       try {
         c.close();
@@ -754,6 +870,7 @@ export async function startDashboardServer(
     }
     proxiedSpaSockets.clear();
     if (fsDebounceTimer) clearTimeout(fsDebounceTimer);
+    if (actionStreamDebounceTimer) clearTimeout(actionStreamDebounceTimer);
     closeSingleton();
     server.stop();
   };

package/cli/selftune/eval/hooks-to-evals.ts

@@ -25,6 +25,10 @@ import { parseArgs } from "node:util";
 
 import { PUBLIC_COMMAND_SURFACES, renderCommandHelp } from "../command-surface.js";
 import { GENERIC_NEGATIVES, QUERY_LOG, SKILL_LOG, TELEMETRY_LOG } from "../constants.js";
+import {
+  createDashboardLlmObserver,
+  emitDashboardStepProgress,
+} from "../dashboard-action-instrumentation.js";
 import { getDb } from "../localdb/db.js";
 import {
   queryQueryLog,
@@ -615,16 +619,49 @@ export async function cliMain(): Promise<void> {
     const maxPerSide = Number.parseInt(values.max ?? "50", 10);
     const effectiveMax = Number.isNaN(maxPerSide) || maxPerSide <= 0 ? 50 : maxPerSide;
 
+    emitDashboardStepProgress({
+      current: 1,
+      total: 4,
+      status: "started",
+      phase: "load_skill",
+      label: "Load skill content",
+    });
     console.log(`Generating synthetic evals for skill '${values.skill}'...`);
     const evalSet = await generateSyntheticEvals(values["skill-path"], values.skill, agent, {
       maxPositives: effectiveMax,
       maxNegatives: effectiveMax,
       modelFlag: values.model,
+      llmObserverFactory: createDashboardLlmObserver,
+    });
+    emitDashboardStepProgress({
+      current: 1,
+      total: 4,
+      status: "finished",
+      phase: "load_skill",
+      label: "Load skill content",
+      passed: true,
+      evidence: values["skill-path"],
     });
 
     const outputPath = values.output ?? values.out ?? `${values.skill}_trigger_eval.json`;
+    emitDashboardStepProgress({
+      current: 4,
+      total: 4,
+      status: "started",
+      phase: "write_eval_set",
+      label: "Write eval set",
+    });
     writeFileSync(outputPath, JSON.stringify(evalSet, null, 2), "utf-8");
     const canonicalPath = writeCanonicalEvalSet(values.skill, evalSet);
+    emitDashboardStepProgress({
+      current: 4,
+      total: 4,
+      status: "finished",
+      phase: "write_eval_set",
+      label: "Write eval set",
+      passed: true,
+      evidence: outputPath,
+    });
 
     const pos = evalSet.filter((e) => e.should_trigger);
     const neg = evalSet.filter((e) => !e.should_trigger);
@@ -666,6 +703,13 @@ export async function cliMain(): Promise<void> {
   const hasCustomQueryLog = queryLogPath !== QUERY_LOG;
   const hasCustomTelemetryLog = telemetryLogPath !== TELEMETRY_LOG;
 
+  emitDashboardStepProgress({
+    current: 1,
+    total: values.blend ? 5 : 3,
+    status: "started",
+    phase: "load_records",
+    label: "Load telemetry and query records",
+  });
   const db = hasCustomSkillLog && hasCustomQueryLog && hasCustomTelemetryLog ? undefined : getDb();
   skillRecords = hasCustomSkillLog
     ? readJsonl<SkillUsageRecord>(skillLogPath)
@@ -676,6 +720,15 @@ export async function cliMain(): Promise<void> {
   telemetryRecords = hasCustomTelemetryLog
     ? readJsonl<SessionTelemetryRecord>(telemetryLogPath)
     : (querySessionTelemetry(db!) as SessionTelemetryRecord[]);
+  emitDashboardStepProgress({
+    current: 1,
+    total: values.blend ? 5 : 3,
+    status: "finished",
+    phase: "load_records",
+    label: "Load telemetry and query records",
+    passed: true,
+    evidence: `${skillRecords.length} skill rows · ${queryRecords.length} query rows`,
+  });
 
   if (values["list-skills"]) {
     listSkills(skillRecords, queryRecords, telemetryRecords);
@@ -701,6 +754,13 @@ export async function cliMain(): Promise<void> {
   const searchDirs = getEvalSkillSearchDirs();
   const detectedSkillPath = findInstalledSkillPath(values.skill, searchDirs);
 
+  emitDashboardStepProgress({
+    current: 2,
+    total: values.blend ? 5 : 3,
+    status: "started",
+    phase: "build_eval_set",
+    label: "Build eval set",
+  });
   const evalSet = buildEvalSet(
     skillRecords,
     queryRecords,
@@ -710,6 +770,15 @@ export async function cliMain(): Promise<void> {
     seed,
     annotateTaxonomy,
   );
+  emitDashboardStepProgress({
+    current: 2,
+    total: values.blend ? 5 : 3,
+    status: "finished",
+    phase: "build_eval_set",
+    label: "Build eval set",
+    passed: true,
+    evidence: `${evalSet.length} entries`,
+  });
 
   const positiveCount = evalSet.filter((entry) => entry.should_trigger).length;
   if (positiveCount === 0 && values["auto-synthetic"]) {
@@ -731,6 +800,13 @@ export async function cliMain(): Promise<void> {
       );
     }
 
+    emitDashboardStepProgress({
+      current: 1,
+      total: 4,
+      status: "started",
+      phase: "load_skill",
+      label: "Load skill content",
+    });
     console.log(
       `No trusted triggers found for '${values.skill}'. Falling back to synthetic cold-start eval generation...`,
     );
@@ -739,10 +815,36 @@ export async function cliMain(): Promise<void> {
       maxPositives: effectiveMax,
       maxNegatives: effectiveMax,
       modelFlag: values.model,
+      llmObserverFactory: createDashboardLlmObserver,
+    });
+    emitDashboardStepProgress({
+      current: 1,
+      total: 4,
+      status: "finished",
+      phase: "load_skill",
+      label: "Load skill content",
+      passed: true,
+      evidence: skillPath,
     });
     const outputPath = values.output ?? values.out ?? `${values.skill}_trigger_eval.json`;
+    emitDashboardStepProgress({
+      current: 4,
+      total: 4,
+      status: "started",
+      phase: "write_eval_set",
+      label: "Write eval set",
+    });
     writeFileSync(outputPath, JSON.stringify(syntheticEvalSet, null, 2), "utf-8");
     const canonicalPath = writeCanonicalEvalSet(values.skill, syntheticEvalSet);
+    emitDashboardStepProgress({
+      current: 4,
+      total: 4,
+      status: "finished",
+      phase: "write_eval_set",
+      label: "Write eval set",
+      passed: true,
+      evidence: outputPath,
+    });
     const pos = syntheticEvalSet.filter((e) => e.should_trigger);
     const neg = syntheticEvalSet.filter((e) => !e.should_trigger);
 
@@ -789,23 +891,78 @@ export async function cliMain(): Promise<void> {
     }
 
     const effectiveMax = Number.isNaN(maxPerSide) || maxPerSide <= 0 ? 50 : maxPerSide;
+    emitDashboardStepProgress({
+      current: 1,
+      total: 5,
+      status: "started",
+      phase: "build_log_eval_set",
+      label: "Build log eval set",
+    });
+    emitDashboardStepProgress({
+      current: 1,
+      total: 5,
+      status: "finished",
+      phase: "build_log_eval_set",
+      label: "Build log eval set",
+      passed: true,
+      evidence: `${evalSet.length} entries`,
+    });
     console.log(`Generating synthetic evals for blending with '${values.skill}'...`);
     const syntheticEvalSet = await generateSyntheticEvals(skillPath, values.skill, agent, {
       maxPositives: effectiveMax,
       maxNegatives: effectiveMax,
       modelFlag: values.model,
+      llmObserverFactory: ({ current, total, phase, label }) =>
+        createDashboardLlmObserver({
+          current: current + 1,
+          total: total + 1,
+          phase,
+          label,
+        }),
     });
 
+    emitDashboardStepProgress({
+      current: 4,
+      total: 5,
+      status: "started",
+      phase: "blend_eval_sets",
+      label: "Blend log and synthetic evals",
+    });
     finalEvalSet = blendEvalSets(evalSet, syntheticEvalSet);
     const stats = computeEvalSourceStats(finalEvalSet);
+    emitDashboardStepProgress({
+      current: 4,
+      total: 5,
+      status: "finished",
+      phase: "blend_eval_sets",
+      label: "Blend log and synthetic evals",
+      passed: true,
+      evidence: `${stats.total} total entries`,
+    });
     console.log(
       `Blended: ${stats.log} log + ${stats.blended} synthetic gap-fillers = ${stats.total} total`,
     );
   }
 
   const outputPath = values.output ?? values.out ?? `${values.skill}_trigger_eval.json`;
+  emitDashboardStepProgress({
+    current: values.blend ? 5 : 3,
+    total: values.blend ? 5 : 3,
+    status: "started",
+    phase: "write_eval_set",
+    label: "Write eval set",
+  });
   writeFileSync(outputPath, JSON.stringify(finalEvalSet, null, 2), "utf-8");
   const canonicalPath = writeCanonicalEvalSet(values.skill, finalEvalSet);
+  emitDashboardStepProgress({
+    current: values.blend ? 5 : 3,
+    total: values.blend ? 5 : 3,
+    status: "finished",
+    phase: "write_eval_set",
+    label: "Write eval set",
+    passed: true,
+    evidence: outputPath,
+  });
   printEvalStats(
     finalEvalSet,
     values.skill,

package/cli/selftune/eval/synthetic-evals.ts

@@ -10,6 +10,7 @@ import { readFileSync } from "node:fs";
 
 import type { EvalEntry, InvocationType, SkillUsageRecord } from "../types.js";
 import { callLlm, stripMarkdownFences } from "../utils/llm-call.js";
+import type { LlmCallObserver } from "../utils/llm-call.js";
 import { findInstalledSkillNames } from "../utils/skill-discovery.js";
 import { classifyInvocation } from "./invocation-classifier.js";
 
@@ -21,6 +22,12 @@ export interface SyntheticEvalOptions {
   maxPositives?: number;
   maxNegatives?: number;
   modelFlag?: string;
+  llmObserverFactory?: (step: {
+    current: number;
+    total: number;
+    phase: string;
+    label: string;
+  }) => LlmCallObserver | undefined;
 }
 
 interface RawSyntheticEntry {
@@ -484,7 +491,19 @@ export async function generateSyntheticEvals(
     siblingSkills,
   );
 
-  const raw = await callLlm(system, user, agent, options.modelFlag);
+  const raw = await callLlm(
+    system,
+    user,
+    agent,
+    options.modelFlag,
+    undefined,
+    options.llmObserverFactory?.({
+      current: 2,
+      total: 4,
+      phase: "draft_eval_set",
+      label: "Draft synthetic eval set",
+    }),
+  );
   const firstPass = dedupeEvalEntries(parseSyntheticResponse(raw, skillName));
 
   try {
@@ -496,7 +515,19 @@ export async function generateSyntheticEvals(
       maxNegatives,
       siblingSkills,
     );
-    const refinedRaw = await callLlm(refinement.system, refinement.user, agent, options.modelFlag);
+    const refinedRaw = await callLlm(
+      refinement.system,
+      refinement.user,
+      agent,
+      options.modelFlag,
+      undefined,
+      options.llmObserverFactory?.({
+        current: 3,
+        total: 4,
+        phase: "refine_eval_set",
+        label: "Refine synthetic eval set",
+      }),
+    );
     const refined = dedupeEvalEntries(parseSyntheticResponse(refinedRaw, skillName));
     const selected = selectBalancedEvalEntries(refined, maxPositives, maxNegatives, siblingSkills);
     if (

package/cli/selftune/eval/unit-test-cli.ts

@@ -13,13 +13,17 @@
  *   --model <m>       Model flag for LLM calls
  */
 
-import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
+import { existsSync, mkdirSync, readFileSync } from "node:fs";
 import { join } from "node:path";
 import { parseArgs } from "node:util";
 
 import { SELFTUNE_CONFIG_DIR } from "../constants.js";
+import {
+  createDashboardLlmObserver,
+  emitDashboardStepProgress,
+} from "../dashboard-action-instrumentation.js";
 import type { EvalEntry } from "../types.js";
-import { writeUnitTestRunResult } from "../testing-readiness.js";
+import { writeCanonicalUnitTests, writeUnitTestRunResult } from "../testing-readiness.js";
 import { CLIError } from "../utils/cli-error.js";
 import { callLlm, detectLlmAgent } from "../utils/llm-call.js";
 import { generateUnitTests } from "./generate-unit-tests.js";
@@ -69,6 +73,13 @@ export async function cliMain(): Promise<void> {
     }
 
     let skillContent = `Skill: ${skillName}`;
+    emitDashboardStepProgress({
+      current: 1,
+      total: 3,
+      status: "started",
+      phase: "load_generation_inputs",
+      label: "Load skill and failure context",
+    });
     if (values["skill-path"] && existsSync(values["skill-path"])) {
       skillContent = readFileSync(values["skill-path"], "utf-8");
     } else if (values["skill-path"]) {
@@ -85,10 +96,31 @@ export async function cliMain(): Promise<void> {
         console.warn("[WARN] Failed to parse eval set. Proceeding without failure context.");
       }
     }
+    emitDashboardStepProgress({
+      current: 1,
+      total: 3,
+      status: "finished",
+      phase: "load_generation_inputs",
+      label: "Load skill and failure context",
+      passed: true,
+      evidence: `${evalFailures.length} eval failures`,
+    });
 
     const modelFlag = values.model;
     const llmCaller = (systemPrompt: string, userPrompt: string) =>
-      callLlm(systemPrompt, userPrompt, agent, modelFlag);
+      callLlm(
+        systemPrompt,
+        userPrompt,
+        agent,
+        modelFlag,
+        undefined,
+        createDashboardLlmObserver({
+          current: 2,
+          total: 3,
+          phase: "generate_tests",
+          label: "Generate unit tests",
+        }),
+      );
 
     console.log(`Generating unit tests for skill '${skillName}'...`);
     const tests = await generateUnitTests(skillName, skillContent, evalFailures, llmCaller);
@@ -98,9 +130,25 @@ export async function cliMain(): Promise<void> {
     }
 
     // Ensure output directory exists
+    emitDashboardStepProgress({
+      current: 3,
+      total: 3,
+      status: "started",
+      phase: "write_tests",
+      label: "Write generated tests",
+    });
     mkdirSync(unitTestDir, { recursive: true });
-    writeFileSync(testsPath, JSON.stringify(tests, null, 2), "utf-8");
-    console.log(`Generated ${tests.length} unit tests -> ${testsPath}`);
+    const storedPath = writeCanonicalUnitTests(skillName, tests, testsPath);
+    emitDashboardStepProgress({
+      current: 3,
+      total: 3,
+      status: "finished",
+      phase: "write_tests",
+      label: "Write generated tests",
+      passed: true,
+      evidence: storedPath,
+    });
+    console.log(`Generated ${tests.length} unit tests -> ${storedPath}`);
     return;
   }