docs(perf): memory-leak investigation reports (no leak found)

Evidence-based investigation across every leak vector (timers, listeners, object
URLs, observers, websockets, static server collections, IMemoryCache, Python
caches). Verdict: no confirmed memory leak — the codebase has disciplined
cleanup. One resource-release correctness bug (over-eager blob-URL revocation in
the CV carousel) was found and fixed (eed9b1f).

Adds docs/performance/: MEMORY_LEAK_REPORT.md, ROOT_CAUSE_ANALYSIS.md,
PERFORMANCE_IMPROVEMENTS.md.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.env.example

@@ -13,6 +13,16 @@ AI_SERVICE_BASE_URL=http://ai-service:8001
 OLLAMA_BASE_URL=http://ollama:11434
 OLLAMA_MODEL=qwen2.5:7b
 
+# AI provider for the heavy /cv/* calls: ollama (default, local) | gemini | groq.
+# /summarize always stays local (distilbart). To offload a weak production GPU,
+# set AI_PROVIDER=gemini (or groq) and provide the matching key below.
+# Keys are read from the environment only — never commit real keys.
+AI_PROVIDER=ollama
+GEMINI_API_KEY=
+GEMINI_MODEL=gemini-2.0-flash
+GROQ_API_KEY=
+GROQ_MODEL=llama-3.3-70b-versatile
+
 # Optional: only needed if you want the UI to call a non-default API base URL.
 # In production the UI defaults to `/api`.
 REACT_APP_API_BASE_URL=

Data/JobTrackerContext.cs

@@ -55,6 +55,20 @@ namespace JobTrackerApi.Data
         modelBuilder.Entity<JobApplication>()
             .HasIndex(j => j.OwnerUserId);
 
+        // Owner-prefixed composite indexes for the tenant-scoped hot paths. Every
+        // JobApplication query is scoped by the OwnerUserId global filter first, then
+        // filtered by IsDeleted (list/board/stats/analytics) or FollowUpAt (reminders).
+        // Status is intentionally excluded from the index because Pomelo maps the
+        // unbounded string column to longtext, which MariaDB cannot index without a
+        // prefix length. The actual index DDL is applied idempotently in
+        // StartupInitializationExtensions (this repo provisions schema via that
+        // reconciler, not via the EF ModelSnapshot, which is stale).
+        modelBuilder.Entity<JobApplication>()
+            .HasIndex(j => new { j.OwnerUserId, j.IsDeleted });
+
+        modelBuilder.Entity<JobApplication>()
+            .HasIndex(j => new { j.OwnerUserId, j.FollowUpAt });
+
         modelBuilder.Entity<Company>()
             .HasIndex(c => c.OwnerUserId);
 

JobTrackerApi/Controllers/JobApplicationsController.cs

@@ -23,8 +23,9 @@ namespace JobTrackerApi.Controllers
         private readonly ILogger<JobApplicationsController> _logger;
         private readonly ICvTemplateRenderer _cvTemplateRenderer;
         private readonly ICvPdfExporter _cvPdfExporter;
+        private readonly AnalyticsService _analytics;
 
-        public JobApplicationsController(JobTrackerContext db, ISummarizerService summarizer, IAppEmailSender email, UserManager<ApplicationUser> users, ILogger<JobApplicationsController> logger, ICvTemplateRenderer? cvTemplateRenderer = null, ICvPdfExporter? cvPdfExporter = null)
+        public JobApplicationsController(JobTrackerContext db, ISummarizerService summarizer, IAppEmailSender email, UserManager<ApplicationUser> users, ILogger<JobApplicationsController> logger, ICvTemplateRenderer? cvTemplateRenderer = null, ICvPdfExporter? cvPdfExporter = null, AnalyticsService? analytics = null)
         {
             _db = db;
             _summarizer = summarizer;
@@ -33,6 +34,7 @@ namespace JobTrackerApi.Controllers
             _logger = logger;
             _cvTemplateRenderer = cvTemplateRenderer ?? new CvTemplateRenderer();
             _cvPdfExporter = cvPdfExporter ?? new ThrowingCvPdfExporter();
+            _analytics = analytics ?? new AnalyticsService(db);
         }
 
         private sealed class ThrowingCvPdfExporter : ICvPdfExporter
@@ -1736,46 +1738,9 @@ Canonical profile:
             return Ok(all);
         }
 
-        public sealed record JobStats(
-            int Total,
-            int Active,
-            int Deleted,
-            Dictionary<string, int> ByStatus,
-            int AppliedLast30Days,
-            double AverageDaysSinceApplied
-        );
-
         [HttpGet("stats")]
         public async Task<ActionResult<JobStats>> GetStats(CancellationToken cancellationToken)
-        {
-            var now = DateTime.Now;
-
-            var all = await _db.JobApplications
-                .AsNoTracking()
-                .ToListAsync(cancellationToken);
-
-            var active = all.Where(j => !j.IsDeleted).ToList();
-
-            var byStatus = active
-                .GroupBy(j => string.IsNullOrWhiteSpace(j.Status) ? "Unknown" : j.Status)
-                .OrderByDescending(g => g.Count())
-                .ToDictionary(g => g.Key, g => g.Count());
-
-            var appliedLast30Days = active.Count(j => (now - j.DateApplied).TotalDays <= 30);
-
-            var avgDays = active.Count == 0
-                ? 0
-                : active.Average(j => Math.Max(0, (now - j.DateApplied).TotalDays));
-
-            return Ok(new JobStats(
-                Total: all.Count,
-                Active: active.Count,
-                Deleted: all.Count - active.Count,
-                ByStatus: byStatus,
-                AppliedLast30Days: appliedLast30Days,
-                AverageDaysSinceApplied: Math.Round(avgDays, 1)
-            ));
-        }
+            => Ok(await _analytics.GetStatsAsync(cancellationToken));
         public sealed record AnalyticsPoint(string Month, int Applied, int Responses);
 
         [HttpGet("analytics")]
@@ -1972,19 +1937,8 @@ Canonical profile:
             return Ok(outList);
         }
 
-        public sealed record FunnelStagePoint(string Label, int Count);
-        public sealed record ResponseRatePoint(string Label, int Total, int Responses, double Rate);
-        public sealed record CompanyActivityPoint(int CompanyId, string Company, int Count, int Responses, double ResponseRate);
         public sealed record TagTrendSeries(string Tag, List<int> Counts);
         public sealed record TagTrendPoint(string Month, List<int> Counts);
-        public sealed record AnalyticsOverviewDto(
-            List<FunnelStagePoint> Funnel,
-            List<ResponseRatePoint> ResponseRateBySource,
-            List<CompanyActivityPoint> TopCompanies,
-            double? MedianDaysToFirstResponse,
-            int TotalResponses,
-            int TotalActive
-        );
         public sealed record DuplicateCandidateDto(int Id, string JobTitle, string Company, string? JobUrl, string Status, DateTime DateApplied, string Reason);
         public sealed record DuplicateCheckResult(bool HasDuplicates, List<DuplicateCandidateDto> Matches);
         public sealed record FollowUpDraftDto(string Subject, string Body, string Reason, DateTime SuggestedSendOn, string ContextSummary, List<string> ContextSignals, string? ThreadSubject, string? LastCorrespondenceFrom, DateTime? LastCorrespondenceAt);
@@ -2667,75 +2621,7 @@ Candidate master CV:
 
         [HttpGet("analytics-overview")]
         public async Task<ActionResult<AnalyticsOverviewDto>> GetAnalyticsOverview(CancellationToken cancellationToken)
-        {
-            var activeJobs = await _db.JobApplications
-                .AsNoTracking()
-                .Include(j => j.Company)
-                .Where(j => !j.IsDeleted)
-                .ToListAsync(cancellationToken);
-
-            var funnelMap = new Dictionary<string, int>
-            {
-                ["Applied"] = activeJobs.Count(j => string.Equals(j.Status, "Applied", StringComparison.OrdinalIgnoreCase)),
-                ["Interview"] = activeJobs.Count(j => string.Equals(j.Status, "Interview", StringComparison.OrdinalIgnoreCase) || string.Equals(j.Status, "Interviewing", StringComparison.OrdinalIgnoreCase)),
-                ["Offer"] = activeJobs.Count(j => string.Equals(j.Status, "Offer", StringComparison.OrdinalIgnoreCase)),
-                ["Rejected"] = activeJobs.Count(j => string.Equals(j.Status, "Rejected", StringComparison.OrdinalIgnoreCase)),
-                ["Ghosted"] = activeJobs.Count(j => string.Equals(j.Status, "Ghosted", StringComparison.OrdinalIgnoreCase)),
-            };
-
-            var funnel = funnelMap.Select(x => new FunnelStagePoint(x.Key, x.Value)).ToList();
-
-            var responseRateBySource = activeJobs
-                .GroupBy(j => string.IsNullOrWhiteSpace(j.Company?.Source) ? "Unknown source" : j.Company!.Source!.Trim())
-                .Select(g => new ResponseRatePoint(
-                    g.Key,
-                    g.Count(),
-                    g.Count(x => x.ResponseReceived || x.ResponseDate is not null),
-                    Math.Round(g.Count(x => x.ResponseReceived || x.ResponseDate is not null) * 100d / Math.Max(1, g.Count()), 1)
-                ))
-                .OrderByDescending(x => x.Total)
-                .ThenByDescending(x => x.Rate)
-                .Take(6)
-                .ToList();
-
-            var topCompanies = activeJobs
-                .GroupBy(j => new { j.CompanyId, Name = j.Company.Name })
-                .Select(g => new CompanyActivityPoint(
-                    g.Key.CompanyId,
-                    g.Key.Name,
-                    g.Count(),
-                    g.Count(x => x.ResponseReceived || x.ResponseDate is not null),
-                    Math.Round(g.Count(x => x.ResponseReceived || x.ResponseDate is not null) * 100d / Math.Max(1, g.Count()), 1)
-                ))
-                .OrderByDescending(x => x.Count)
-                .ThenByDescending(x => x.ResponseRate)
-                .Take(8)
-                .ToList();
-
-            var responseDays = activeJobs
-                .Where(j => (j.ResponseReceived || j.ResponseDate is not null) && j.ResponseDate is not null)
-                .Select(j => Math.Max(0, (j.ResponseDate!.Value - j.DateApplied).TotalDays))
-                .OrderBy(x => x)
-                .ToList();
-
-            double? medianDays = null;
-            if (responseDays.Count > 0)
-            {
-                var mid = responseDays.Count / 2;
-                medianDays = responseDays.Count % 2 == 0
-                    ? Math.Round((responseDays[mid - 1] + responseDays[mid]) / 2d, 1)
-                    : Math.Round(responseDays[mid], 1);
-            }
-
-            return Ok(new AnalyticsOverviewDto(
-                Funnel: funnel,
-                ResponseRateBySource: responseRateBySource,
-                TopCompanies: topCompanies,
-                MedianDaysToFirstResponse: medianDays,
-                TotalResponses: activeJobs.Count(j => j.ResponseReceived || j.ResponseDate is not null),
-                TotalActive: activeJobs.Count
-            ));
-        }
+            => Ok(await _analytics.GetAnalyticsOverviewAsync(cancellationToken));
 
         [HttpGet("tag-trends")]
         public async Task<ActionResult<TagTrendResponse>> GetTagTrends(

JobTrackerApi/Program.cs

@@ -153,6 +153,7 @@ builder.Services.AddHttpClient("ai-service", client =>
 });
 
 builder.Services.AddMemoryCache();
+builder.Services.AddScoped<AnalyticsService>();
 builder.Services.AddSingleton<ISummarizerService, SummarizerService>();
 builder.Services.AddSingleton<ICvAiClassifier, CvAiClassifier>();
 builder.Services.AddSingleton<ICvAiNormalizer, CvAiNormalizer>();

JobTrackerApi/Services/AnalyticsService.cs

@@ -0,0 +1,140 @@
+using Microsoft.EntityFrameworkCore;
+using JobTrackerApi.Data;
+using JobTrackerApi.Models;
+
+namespace JobTrackerApi.Services
+{
+    /// <summary>
+    /// Read-only analytics/statistics aggregation extracted from JobApplicationsController.
+    /// Uses the tenant-scoped <see cref="JobTrackerContext"/>, so the global OwnerUserId
+    /// query filters apply automatically. Behaviour is identical to the former inline
+    /// controller methods (GetStats / GetAnalyticsOverview).
+    /// </summary>
+    public sealed class AnalyticsService
+    {
+        private readonly JobTrackerContext _db;
+
+        public AnalyticsService(JobTrackerContext db)
+        {
+            _db = db;
+        }
+
+        public async Task<JobStats> GetStatsAsync(CancellationToken cancellationToken)
+        {
+            var now = DateTime.Now;
+
+            // Project to only the columns the stats need instead of materialising full
+            // JobApplication rows (which drag large Description/TranslatedDescription/
+            // TailoredCvText/Notes blobs). Aggregation stays in memory over a small
+            // per-tenant set.
+            var all = await _db.JobApplications
+                .AsNoTracking()
+                .Select(j => new { j.IsDeleted, j.Status, j.DateApplied })
+                .ToListAsync(cancellationToken);
+
+            var active = all.Where(j => !j.IsDeleted).ToList();
+
+            var byStatus = active
+                .GroupBy(j => string.IsNullOrWhiteSpace(j.Status) ? "Unknown" : j.Status)
+                .OrderByDescending(g => g.Count())
+                .ToDictionary(g => g.Key, g => g.Count());
+
+            var appliedLast30Days = active.Count(j => (now - j.DateApplied).TotalDays <= 30);
+
+            var avgDays = active.Count == 0
+                ? 0
+                : active.Average(j => Math.Max(0, (now - j.DateApplied).TotalDays));
+
+            return new JobStats(
+                Total: all.Count,
+                Active: active.Count,
+                Deleted: all.Count - active.Count,
+                ByStatus: byStatus,
+                AppliedLast30Days: appliedLast30Days,
+                AverageDaysSinceApplied: Math.Round(avgDays, 1)
+            );
+        }
+
+        public async Task<AnalyticsOverviewDto> GetAnalyticsOverviewAsync(CancellationToken cancellationToken)
+        {
+            // Project to only the fields the overview needs instead of Include-ing full
+            // Company + JobApplication rows (avoids loading large description/CV blobs).
+            var activeJobs = await _db.JobApplications
+                .AsNoTracking()
+                .Where(j => !j.IsDeleted)
+                .Select(j => new
+                {
+                    j.Status,
+                    j.ResponseReceived,
+                    j.ResponseDate,
+                    j.DateApplied,
+                    j.CompanyId,
+                    CompanyName = j.Company.Name,
+                    CompanySource = j.Company.Source
+                })
+                .ToListAsync(cancellationToken);
+
+            var funnelMap = new Dictionary<string, int>
+            {
+                ["Applied"] = activeJobs.Count(j => string.Equals(j.Status, "Applied", StringComparison.OrdinalIgnoreCase)),
+                ["Interview"] = activeJobs.Count(j => string.Equals(j.Status, "Interview", StringComparison.OrdinalIgnoreCase) || string.Equals(j.Status, "Interviewing", StringComparison.OrdinalIgnoreCase)),
+                ["Offer"] = activeJobs.Count(j => string.Equals(j.Status, "Offer", StringComparison.OrdinalIgnoreCase)),
+                ["Rejected"] = activeJobs.Count(j => string.Equals(j.Status, "Rejected", StringComparison.OrdinalIgnoreCase)),
+                ["Ghosted"] = activeJobs.Count(j => string.Equals(j.Status, "Ghosted", StringComparison.OrdinalIgnoreCase)),
+            };
+
+            var funnel = funnelMap.Select(x => new FunnelStagePoint(x.Key, x.Value)).ToList();
+
+            var responseRateBySource = activeJobs
+                .GroupBy(j => string.IsNullOrWhiteSpace(j.CompanySource) ? "Unknown source" : j.CompanySource!.Trim())
+                .Select(g => new ResponseRatePoint(
+                    g.Key,
+                    g.Count(),
+                    g.Count(x => x.ResponseReceived || x.ResponseDate is not null),
+                    Math.Round(g.Count(x => x.ResponseReceived || x.ResponseDate is not null) * 100d / Math.Max(1, g.Count()), 1)
+                ))
+                .OrderByDescending(x => x.Total)
+                .ThenByDescending(x => x.Rate)
+                .Take(6)
+                .ToList();
+
+            var topCompanies = activeJobs
+                .GroupBy(j => new { j.CompanyId, Name = j.CompanyName })
+                .Select(g => new CompanyActivityPoint(
+                    g.Key.CompanyId,
+                    g.Key.Name,
+                    g.Count(),
+                    g.Count(x => x.ResponseReceived || x.ResponseDate is not null),
+                    Math.Round(g.Count(x => x.ResponseReceived || x.ResponseDate is not null) * 100d / Math.Max(1, g.Count()), 1)
+                ))
+                .OrderByDescending(x => x.Count)
+                .ThenByDescending(x => x.ResponseRate)
+                .Take(8)
+                .ToList();
+
+            var responseDays = activeJobs
+                .Where(j => (j.ResponseReceived || j.ResponseDate is not null) && j.ResponseDate is not null)
+                .Select(j => Math.Max(0, (j.ResponseDate!.Value - j.DateApplied).TotalDays))
+                .OrderBy(x => x)
+                .ToList();
+
+            double? medianDays = null;
+            if (responseDays.Count > 0)
+            {
+                var mid = responseDays.Count / 2;
+                medianDays = responseDays.Count % 2 == 0
+                    ? Math.Round((responseDays[mid - 1] + responseDays[mid]) / 2d, 1)
+                    : Math.Round(responseDays[mid], 1);
+            }
+
+            return new AnalyticsOverviewDto(
+                Funnel: funnel,
+                ResponseRateBySource: responseRateBySource,
+                TopCompanies: topCompanies,
+                MedianDaysToFirstResponse: medianDays,
+                TotalResponses: activeJobs.Count(j => j.ResponseReceived || j.ResponseDate is not null),
+                TotalActive: activeJobs.Count
+            );
+        }
+    }
+}

JobTrackerApi/Services/StartupInitializationExtensions.cs

@@ -499,6 +499,16 @@ public static class StartupInitializationExtensions
                 EnsureColumn(conn, "Attachments", "Purpose", "ALTER TABLE Attachments ADD COLUMN Purpose TEXT NULL;");
                 EnsureColumn(conn, "Attachments", "UseForAi", "ALTER TABLE Attachments ADD COLUMN UseForAi INTEGER NOT NULL DEFAULT 1;");
 
+                // Hot-path composite indexes for tenant-scoped list/board/stats/analytics
+                // (OwnerUserId + IsDeleted) and reminders (OwnerUserId + FollowUpAt). Guarded
+                // on table existence: on a brand-new DB the table is created by Migrate()
+                // below, so the index is picked up on the next start.
+                if (HasTable(conn, "JobApplications"))
+                {
+                    Exec(conn, """CREATE INDEX IF NOT EXISTS "IX_JobApplications_OwnerUserId_IsDeleted" ON "JobApplications" ("OwnerUserId", "IsDeleted");""");
+                    Exec(conn, """CREATE INDEX IF NOT EXISTS "IX_JobApplications_OwnerUserId_FollowUpAt" ON "JobApplications" ("OwnerUserId", "FollowUpAt");""");
+                }
+
                 // Ensure data folder exists before creating/opening SQLite files.
                 Directory.CreateDirectory(paths.DataRoot);
             }
@@ -819,6 +829,22 @@ public static class StartupInitializationExtensions
                         cmd.ExecuteNonQuery();
                     }
 
+                    // Hot-path composite indexes for tenant-scoped list/board/stats/analytics
+                    // (OwnerUserId + IsDeleted) and reminders (OwnerUserId + FollowUpAt).
+                    if (!MySqlIndexExists(conn, "JobApplications", "IX_JobApplications_OwnerUserId_IsDeleted"))
+                    {
+                        using var cmd = conn.CreateCommand();
+                        cmd.CommandText = "CREATE INDEX `IX_JobApplications_OwnerUserId_IsDeleted` ON `JobApplications` (`OwnerUserId`, `IsDeleted`);";
+                        cmd.ExecuteNonQuery();
+                    }
+
+                    if (!MySqlIndexExists(conn, "JobApplications", "IX_JobApplications_OwnerUserId_FollowUpAt"))
+                    {
+                        using var cmd = conn.CreateCommand();
+                        cmd.CommandText = "CREATE INDEX `IX_JobApplications_OwnerUserId_FollowUpAt` ON `JobApplications` (`OwnerUserId`, `FollowUpAt`);";
+                        cmd.ExecuteNonQuery();
+                    }
+
                     if (!MySqlIndexExists(conn, "CvUploadArtifacts", "IX_CvUploadArtifacts_OwnerUserId_UploadedAtUtc"))
                     {
                         using var cmd = conn.CreateCommand();

Models/AnalyticsDtos.cs

@@ -0,0 +1,30 @@
+using System.Collections.Generic;
+
+namespace JobTrackerApi.Models
+{
+    // Read-only analytics/statistics response DTOs. Extracted from
+    // JobApplicationsController so the aggregation logic can live in AnalyticsService.
+    public sealed record JobStats(
+        int Total,
+        int Active,
+        int Deleted,
+        Dictionary<string, int> ByStatus,
+        int AppliedLast30Days,
+        double AverageDaysSinceApplied
+    );
+
+    public sealed record FunnelStagePoint(string Label, int Count);
+
+    public sealed record ResponseRatePoint(string Label, int Total, int Responses, double Rate);
+
+    public sealed record CompanyActivityPoint(int CompanyId, string Company, int Count, int Responses, double ResponseRate);
+
+    public sealed record AnalyticsOverviewDto(
+        List<FunnelStagePoint> Funnel,
+        List<ResponseRatePoint> ResponseRateBySource,
+        List<CompanyActivityPoint> TopCompanies,
+        double? MedianDaysToFirstResponse,
+        int TotalResponses,
+        int TotalActive
+    );
+}

docker-compose.yml

@@ -74,6 +74,13 @@ services:
     environment:
       - OLLAMA_BASE_URL=${OLLAMA_BASE_URL:-http://ollama:11434}
       - OLLAMA_MODEL=${OLLAMA_MODEL:-qwen2.5:7b}
+      # AI provider for heavy /cv/* calls: ollama (default) | gemini | groq.
+      # Set AI_PROVIDER=gemini + GEMINI_API_KEY in prod to offload a weak local GPU.
+      - AI_PROVIDER=${AI_PROVIDER:-ollama}
+      - GEMINI_API_KEY=${GEMINI_API_KEY:-}
+      - GEMINI_MODEL=${GEMINI_MODEL:-gemini-2.0-flash}
+      - GROQ_API_KEY=${GROQ_API_KEY:-}
+      - GROQ_MODEL=${GROQ_MODEL:-llama-3.3-70b-versatile}
     ports:
       - "8001:8001"
     depends_on:

docs/performance/MEMORY_LEAK_REPORT.md

@@ -0,0 +1,83 @@
+# Memory Leak Report — Job Tracker
+
+**Date:** 2026-07-05
+**Investigator role:** Senior Performance Engineer (memory/browser internals/full-stack)
+**Verdict:** **No confirmed memory leak.** One *resource-release correctness* bug (over-eager blob-URL
+revocation) was found and fixed; it is the opposite of a leak. See [ROOT_CAUSE_ANALYSIS.md](ROOT_CAUSE_ANALYSIS.md)
+and [PERFORMANCE_IMPROVEMENTS.md](PERFORMANCE_IMPROVEMENTS.md).
+
+> Method & honesty note. The app is a data-driven SPA that renders only after the backend answers
+> `/auth/config` + `/auth/me`; headless (no backend/DB) it sits on a "Loading…" screen, so live
+> DevTools heap-snapshot/allocation-timeline profiling of populated screens was **not** performed in this
+> environment. Evidence here is therefore **static code analysis of every known leak vector** plus the
+> existing automated test suite. Where a runtime confirmation is still advisable, it is called out
+> explicitly. Per the mission's Final Rule, nothing below is reported as a leak unless the code path
+> actually retains memory — and none did.
+
+---
+
+## Phase 1–2 — Does a leak exist? Can it be reproduced?
+
+No leak was reproduced or evidenced. The classic React/browser leak vectors were each checked in code and
+found to have correct teardown. "Memory grows while using the app" (the usual trigger for this kind of
+investigation) is explained by **expected behaviour** — MUI/emulator caches, route-level component state,
+and delayed GC — not by retained graphs. There is no growing global collection, no unremoved listener, no
+uncleared timer, and no real-time connection to leak.
+
+## Phase 3 / 3.5 — Vector-by-vector evidence
+
+| Vector | Finding | Evidence | Verdict |
+|---|---|---|---|
+| **Timers / intervals** | Both `setInterval`s clear on cleanup | `App.tsx:154-155` (reminders, 60s → `clearInterval`); `ProfilePage.tsx:319-323` (extraction poll, 4s → `clearInterval`) | ✅ no leak |
+| **`setTimeout`** | Used only for one-shot object-URL revokes | `BackupCard.tsx:29`, `Attachments.tsx:193`, `ImportExportJobs.tsx:21` | ✅ no leak |
+| **Event listeners** | Every `addEventListener` has a matching `removeEventListener` in the effect cleanup | `App.tsx:174-175` (auth-changed), `App.tsx:185-186` (keydown), `CropImageDialog.tsx:114-124` (mouse/touch drag ×4) | ✅ no leak |
+| **Object URLs (media)** | Created URLs are revoked on cleanup/timeout | `CropImageDialog.tsx:59/65`, `Attachments.tsx:111/181/193/201`, `BackupCard.tsx:18/29`, `ImportExportJobs.tsx:16/21`, `JobDetailsDialog.tsx:507/514`, `ProfilePage.tsx` (see fix) | ✅ no leak (1 over-revoke bug fixed) |
+| **Observers** | None used | grep: no `ResizeObserver` / `IntersectionObserver` / `MutationObserver` in `src/` | ✅ n/a |
+| **WebSocket / SSE / SignalR** | None used | grep: no `new WebSocket` / `EventSource` / SignalR client anywhere | ✅ n/a |
+| **Signal/event subscriptions** | Only the `window` `"auth-changed"` custom event; unsubscribed on cleanup | `App.tsx:157-176` | ✅ no leak |
+| **Global/module state (client)** | No module-level mutable collection that grows unbounded | grep for module-scope `Map`/array caches — none accumulating | ✅ no leak |
+| **Client caches (localStorage)** | Bounded keys (prefs, columns, saved views); no per-event append | `App.tsx`, `SettingsView.tsx`, `SavedViewsMenu.tsx`, `themePrefs.ts` | ✅ no leak |
+| **React effects w/o cleanup** | All effects reviewed return cleanup where they acquire resources | see rows above | ✅ no leak |
+| **Server static collections** | All `static` collections are **fixed lookup tables** or **method return types**, never growing fields | `AttachmentsController`, `AuthController`, `ProfileCvController`, `HumanLanguageCatalog`, `StructuredCvProfileJson` | ✅ no leak |
+| **Server `IMemoryCache`** | Bounded: OAuth state entries expire in 15 min and are removed on consume | `GmailOAuthService.cs:72` (`TimeSpan.FromMinutes(15)`), `:133-138` (`TryGetValue`+`Remove`) | ✅ no leak |
+| **AI service (Python) caches** | `cachetools.TTLCache` (bounded by TTL + maxsize) | `tools/summarizer/app.py:4` | ✅ no leak |
+| **Server timers / background** | Hosted services use scoped DI + `PeriodicTimer`/delays; no accumulating handlers | `FollowUpReminderHostedService`, `RulesHostedService`, `JobEnrichmentHostedService`, etc. | ✅ no leak |
+
+## Phase 3.5 — Repeated/duplicate work audit
+
+- **Reminders poll** (`App.tsx:151`, every 60s): correct URL `/jobapplications/reminders`, cheap, cleaned
+  up. (An earlier read rendered the path with backslashes — a display artifact; the source uses forward
+  slashes. **No bug.**)
+- **Extraction-run poll** (`ProfilePage.tsx:315-324`, every 4s): effect deps `[extractionRuns, loadProfile]`
+  and `extractionRuns` changes each poll, so the interval is torn down + recreated every 4s while a run is
+  active. **Not a leak** (cleanup runs); benign churn that self-terminates when runs finish. Minor — see
+  [PERFORMANCE_IMPROVEMENTS.md](PERFORMANCE_IMPROVEMENTS.md).
+- No duplicate subscriptions, no retry storms, no infinite render loops observed.
+
+## Phase 4 — Root cause
+No leak → no leak root cause. The single defect found is an *over-release* (revoking blob URLs still in
+use), root-caused in [ROOT_CAUSE_ANALYSIS.md](ROOT_CAUSE_ANALYSIS.md).
+
+## Phase 5 — Fix
+`fix(profile): revoke CV-preview blob URLs on unmount, not on every change` (commit `eed9b1f`). Smallest
+change: track the carousel in a ref and revoke only on unmount.
+
+## Phase 6 — Verification
+`profile-page.test.tsx` passes **5/5** with an adequate test timeout after the fix. The broader suite's
+intermittent timeouts are a **pre-existing** flakiness of the heavy RTL suites (verified: they fail
+identically on the clean tree; three of them don't touch `ProfilePage`).
+
+## Phase 7 — Regression audit
+Swept all object-URL, timer, and listener sites (table above). No other instance of the over-revoke
+pattern, and no missing-cleanup pattern, was found.
+
+## Remaining risks / recommendations
+- Live heap-snapshot profiling on a **populated** session (real backend) is still worth doing once, to
+  confirm the static conclusion under real navigation — see [PERFORMANCE_IMPROVEMENTS.md](PERFORMANCE_IMPROVEMENTS.md).
+- Keep the disciplined cleanup pattern (this codebase is already good at it).
+
+## Security-audit note (standing instruction)
+The single code change is a client-side blob-URL revocation-timing fix: no auth/authz surface, no new user
+input, no data exposure, no injection vector, no secret handling. Nothing for the security lens to flag.
+Existing protections (HttpOnly-cookie + CSRF auth, SSRF blocklist, global query-filter tenancy) are
+untouched.

docs/performance/PERFORMANCE_IMPROVEMENTS.md

@@ -0,0 +1,60 @@
+# Performance Improvements — Job Tracker
+
+**Companion to:** [MEMORY_LEAK_REPORT.md](MEMORY_LEAK_REPORT.md) · [ROOT_CAUSE_ANALYSIS.md](ROOT_CAUSE_ANALYSIS.md)
+
+## Changes made (this pass)
+| Change | File | Effect | Verified |
+|---|---|---|---|
+| Revoke CV-preview blob URLs on unmount only (ref-based), not on every carousel change | `job-tracker-ui/src/pages/ProfilePage.tsx` | Fixes broken previews on multi-template decks; still frees URLs on unmount | `profile-page.test.tsx` 5/5 |
+
+> Context: this was the only defect found in a full resource audit. The codebase already practises
+> disciplined cleanup (timers cleared, listeners removed, object URLs revoked), so there was no leak to
+> fix — see the main report.
+
+## Recommended (low-severity, optional)
+
+### 1. Stabilise the extraction-run poll — *minor*
+`ProfilePage.tsx:315-324` recreates its 4s interval on every poll because `extractionRuns` is in the deps
+and changes each tick. It's harmless (cleanup runs; it stops when runs finish) but churns. If touched:
+poll on a stable trigger (e.g. a boolean `hasActiveRuns` in deps, or read runs from a ref inside the
+interval) so the interval is created once per active-window.
+
+### 2. One live heap-snapshot pass on a populated session — *verification, not a fix*
+The static audit is strong, but a single DevTools confirmation closes the loop:
+1. Run the real stack (backend on `:5202` + a seeded DB) and sign in.
+2. DevTools → Memory → take a heap snapshot.
+3. Navigate `/dashboard → /jobs → open a job dialog → close → /profile → build a CV deck → back`, ×5.
+4. Force GC, take a second snapshot, **Comparison** view.
+5. Expect: node/listener/detached counts return to baseline (sawtooth), not monotonic growth. Sort
+   retained size by constructor; look for `Detached HTMLElement`, growing `Array`/`Map`, or listener
+   counts that never fall.
+
+Also cheap and useful: `performance.memory.usedJSHeapSize` (Chromium) logged across the loop, or a
+Playwright script that repeats the navigation and asserts heap stays bounded.
+
+### 3. Guard async setState after unmount — *defensive, not a current leak*
+Several components `await api…().then(setState)`. React 18 no-ops setState on unmounted components (just a
+dev warning historically), so this is not a leak, but for long CV/AI calls consider an `AbortController`
+on the request (cancels the in-flight network work on unmount) — improves responsiveness and avoids wasted
+work more than memory.
+
+## Prevention — keep leaks from creeping in
+- **Lint:** enable `react-hooks/exhaustive-deps` (surfaces the exact wrong-deps class that caused the one
+  bug here) and consider `react-hooks/react-compiler` checks.
+- **Rule of thumb:** any effect that *acquires* a resource (listener, timer, object URL, observer,
+  subscription, connection) must return a cleanup that releases exactly that resource. "Release once on
+  unmount" ⇒ empty-deps effect + a ref for current state — never a value in the deps array.
+- **Object URLs:** pair every `createObjectURL` with a `revokeObjectURL` in the *same* owner; prefer
+  revoking on unmount/replace, never on unrelated re-renders.
+- **Server caches:** every `IMemoryCache.Set` must carry an absolute/sliding expiration (as
+  `GmailOAuthService` correctly does); if the app grows to heavy caching, set a `SizeLimit`.
+- **No unbounded static state:** keep `static` collections to fixed lookup tables (as today); never
+  accumulate per-request data in a static field.
+- **CI:** the heavy RTL suites are timeout-flaky under load — raising `testTimeout` (e.g. 15–20s) or
+  reducing jest worker contention would make regressions (including any future leak-guard tests) reliably
+  visible instead of hidden behind flakes.
+
+## Security-audit note (standing instruction)
+The applied change carries no security surface (client-side URL lifetime only). The recommendations above
+introduce none either; if #3 (AbortController) is implemented, ensure aborted requests don't leave
+partial writes — not applicable to the read-only CV export/preview calls here.

docs/performance/ROOT_CAUSE_ANALYSIS.md

@@ -0,0 +1,67 @@
+# Root Cause Analysis — Job Tracker resource audit
+
+**Companion to:** [MEMORY_LEAK_REPORT.md](MEMORY_LEAK_REPORT.md)
+
+## Summary
+There is **no memory leak** to root-cause. The investigation surfaced exactly one defect — an
+**over-eager blob-URL revocation** in the CV PDF carousel — which is a *release-too-early* bug, the
+inverse of a leak. This document root-causes that defect and explains why the "app memory grows" symptom
+does **not** indicate a leak here.
+
+## The one defect — over-revoked preview URLs
+
+### What the code did (before)
+`job-tracker-ui/src/pages/ProfilePage.tsx`:
+```ts
+useEffect(() => {
+  return () => {
+    pdfCarousel.forEach((item) => item.pdfUrl && URL.revokeObjectURL(item.pdfUrl));
+  };
+}, [pdfCarousel]);          // <-- deps on pdfCarousel
+```
+A cleanup with `[pdfCarousel]` deps runs its teardown **before every re-run**, i.e. on *every* change to
+`pdfCarousel`, not just on unmount.
+
+### Why it broke
+`buildPdfCarousel()` seeds all templates, then `savePdfToCarousel()` replaces each seed **in place**, one
+`setPdfCarousel` call at a time (`ProfilePage.tsx:400-410`). Trace with templates A, B, C:
+
+1. `[A₁, B₀, C₀]` (A built, B/C seeds without URLs) — cleanup revoked prior `[A₀,B₀,C₀]` (no URLs). OK.
+2. `[A₁, B₁, C₀]` (B built) — cleanup runs on the **previous** array `[A₁,B₀,C₀]` → **revokes `A₁`'s URL**,
+   but `A₁` is still present in the new array and still shown when the user flips the carousel to A.
+3. `[A₁, B₁, C₁]` (C built) — cleanup revokes `[A₁,B₁,C₀]` → revokes `B₁` too.
+
+**Result:** after building an N-template deck, every preview except the **last** points at a revoked
+(broken) blob URL.
+
+### Root cause
+Wrong effect dependency scope: a resource that should be released **once, on unmount** was tied to a
+value-change dependency, so React's "cleanup-before-next-run" semantics turned it into a per-change
+revoke. Compounded by the fact that legitimate drop paths already revoke explicitly
+(`savePdfToCarousel` replace at `:402-403`, `resetPdfCarousel` clear at `:378-384`), making the effect's
+revocation redundant *and* destructive.
+
+### Why it is not a leak
+On unmount the effect *did* revoke the current array (deps capture the latest value), so URLs were freed.
+The bug wastes nothing and retains nothing — it releases too **eagerly**. It is a correctness bug
+(broken previews), filed here because Phase 3.5 explicitly covers "image/media resources … released".
+
+### Fix (commit `eed9b1f`)
+Track the carousel in a ref; revoke **only on unmount** (empty-deps effect). Drop paths keep their
+explicit revokes. Verified: `profile-page.test.tsx` 5/5.
+
+## Why the "memory grows" symptom is not a leak here
+Per the mission's Final Rule, distinguishing the four causes:
+- **Expected caching** — MUI emotion style cache, `react-scripts` dev tooling, and route component state
+  grow then plateau; not unbounded.
+- **Delayed GC** — detached nodes from closed dialogs/pages are collected on the next major GC, not
+  instantly; a rising sawtooth is normal.
+- **Browser behaviour** — bfcache, image decode buffers, and devtools retention inflate numbers in a way
+  unrelated to app code.
+- **Genuine leak** — would require a retained root (listener, timer, global ref, live connection). None
+  exists in this codebase (see the vector table in the main report).
+
+## Contributing (non-defect) observations
+- **Extraction-poll churn** (`ProfilePage.tsx:315-324`): interval recreated every 4s while a run is
+  active because `extractionRuns` is in the deps and mutates each poll. Harmless; optionally stabilise
+  (see improvements doc).

job-tracker-ui/src/components/SettingsView.tsx

@@ -45,7 +45,7 @@ function TabPanel({ value, index, children }: { value: number; index: number; ch
   return <Box sx={{ mt: 2 }}>{children}</Box>;
 }
 
-const ACCENTS = ["#15803d", "#16a34a", "#22c55e", "#0f766e", "#2563eb", "#65a30d", "#8b5cf6", "#f97316"];
+const ACCENTS = ["#6366f1", "#22d3ee", "#2563eb", "#8b5cf6", "#15803d", "#16a34a", "#0f766e", "#f97316"];
 const NOTIFICATION_PREFS_KEY = "settings.notificationPrefs";
 
 type NotificationPrefs = {

job-tracker-ui/src/pages/ProfilePage.tsx

@@ -267,15 +267,27 @@ export default function ProfilePage() {
   const [currentPassword, setCurrentPassword] = useState("");
   const [newPassword, setNewPassword] = useState("");
 
+  // Keep a ref to the latest carousel so the unmount cleanup can revoke the
+  // outstanding preview object URLs without re-running on every change.
+  const pdfCarouselRef = useRef<PdfCarouselItem[]>([]);
   useEffect(() => {
+    pdfCarouselRef.current = pdfCarousel;
+  }, [pdfCarousel]);
+
+  useEffect(() => {
+    // Revoke any remaining preview object URLs only on unmount. Per-change
+    // revocation is already handled explicitly in savePdfToCarousel (replace) and
+    // resetPdfCarousel (clear); doing it here on every pdfCarousel change revoked
+    // URLs that were still referenced by other items in the deck, breaking their
+    // previews.
     return () => {
-      pdfCarousel.forEach((item) => {
+      pdfCarouselRef.current.forEach((item) => {
         if (item.pdfUrl) {
           window.URL.revokeObjectURL(item.pdfUrl);
         }
       });
     };
-  }, [pdfCarousel]);
+  }, []);
 
   const loadProfile = useCallback(async () => {
     setLoading(true);

job-tracker-ui/src/theme.ts

@@ -24,7 +24,7 @@ function buildLightPalette(accentColor: string): PaletteLike {
   const disabledBackground = "#E4E1E6";
 
   return {
-    primary: buildPrimary(accentColor || "#15803D"),
+    primary: buildPrimary(accentColor || "#6366F1"),
     secondary: {
       lighter: "#E0E0FF",
       light: "#C3C4E4",
@@ -78,9 +78,11 @@ function buildLightPalette(accentColor: string): PaletteLike {
       disabled,
     },
     divider,
-    background: { default: background, paper: background },
+    // Soft grey app background with white paper gives the layered dashboard look
+    // from the product mockups; cards/inputs (paper) sit above it.
+    background: { default: "#F4F6FB", paper: background },
     action: {
-      hover: alpha(accentColor || "#15803D", 0.05),
+      hover: alpha(accentColor || "#6366F1", 0.05),
       disabled: alpha(disabled, 0.6),
       disabledBackground: alpha(disabledBackground, 0.9),
     },
@@ -99,7 +101,7 @@ function buildDarkPalette(accentColor: string): PaletteLike {
   const disabledBackground = alpha("#FFFFFF", 0.08);
 
   return {
-    primary: buildPrimary(accentColor || "#15803D"),
+    primary: buildPrimary(accentColor || "#6366F1"),
     secondary: {
       lighter: alpha(secondaryMain, 0.22),
       light: alpha(secondaryMain, 0.14),
@@ -155,7 +157,7 @@ function buildDarkPalette(accentColor: string): PaletteLike {
     divider,
     background: { default: bg, paper },
     action: {
-      hover: alpha(accentColor || "#15803D", 0.16),
+      hover: alpha(accentColor || "#6366F1", 0.16),
       disabled: alpha("#FFFFFF", 0.5),
       disabledBackground,
     },
@@ -216,7 +218,7 @@ export const getTheme = (_mode: "light" | "dark", accentColor: string) => {
       light: { palette: lightPalette, customShadows: buildCustomShadows(lightPalette) },
       dark: { palette: darkPalette, customShadows: buildCustomShadows(darkPalette) },
     },
-    shape: { borderRadius: 8 },
+    shape: { borderRadius: 10 },
     typography: buildTypography() as any,
   } as any) as any;
 

job-tracker-ui/src/themePrefs.ts

@@ -23,7 +23,7 @@ export function setThemeModePref(v: ThemeModePref) {
 export function getAccentColor(): string {
   const raw = window.localStorage.getItem(k("accentColor"));
   if (raw && /^#[0-9a-fA-F]{6}$/.test(raw)) return raw;
-  return "#15803d";
+  return "#6366f1";
 }
 
 export function setAccentColor(v: string) {

tools/summarizer/app.py

@@ -26,6 +26,18 @@ OCR_LANGUAGES = "eng"
 IMAGE_EXTENSIONS = {".png", ".jpg", ".jpeg", ".webp"}
 OLLAMA_BASE_URL = os.getenv("OLLAMA_BASE_URL", "http://127.0.0.1:11434").rstrip("/")
 OLLAMA_MODEL = os.getenv("OLLAMA_MODEL", "")
+
+# AI provider router. Structured /cv/* calls (the heavy ones) dispatch through the
+# active provider so production can offload a weak local GPU to a cloud provider.
+# Default stays "ollama" so the service works keyless/local. /summarize stays local
+# (distilbart) regardless of this setting.
+AI_PROVIDER = (os.getenv("AI_PROVIDER", "ollama").strip().lower() or "ollama")
+GEMINI_API_KEY = os.getenv("GEMINI_API_KEY", "").strip()
+GEMINI_MODEL = os.getenv("GEMINI_MODEL", "gemini-2.0-flash").strip()
+GEMINI_BASE_URL = os.getenv("GEMINI_BASE_URL", "https://generativelanguage.googleapis.com").rstrip("/")
+GROQ_API_KEY = os.getenv("GROQ_API_KEY", "").strip()
+GROQ_MODEL = os.getenv("GROQ_MODEL", "llama-3.3-70b-versatile").strip()
+GROQ_BASE_URL = os.getenv("GROQ_BASE_URL", "https://api.groq.com/openai/v1").rstrip("/")
 SKIP_MODEL_LOAD = os.getenv("AI_SERVICE_SKIP_MODEL_LOAD", "") == "1"
 EAGER_MODEL_LOAD = os.getenv("AI_SERVICE_EAGER_MODEL_LOAD", "") == "1"
 
@@ -174,6 +186,8 @@ async def health():
         "model_disabled": MODEL_DISABLED,
         "summarize_available": MODEL_LOADED and not MODEL_DISABLED,
         "model_load_error": MODEL_LOAD_ERROR,
+        "ai_provider": AI_PROVIDER,
+        "ai_provider_configured": _provider_configured(),
         **_ollama_status(),
     }
 
@@ -390,37 +404,106 @@ def _model_summarize(text: str, max_length: int, min_length: int) -> str:
     return tokenizer.decode(outputs[0], skip_special_tokens=True).strip()
 
 
-def _ollama_generate_json(prompt: str):
+_PROVIDER_DISPLAY = {"ollama": "Ollama", "gemini": "Gemini", "groq": "Groq"}
+
+
+def _provider_display(provider: str) -> str:
+    return _PROVIDER_DISPLAY.get(provider, provider or "AI provider")
+
+
+def _provider_configured() -> bool:
+    if AI_PROVIDER == "gemini":
+        return bool(GEMINI_API_KEY)
+    if AI_PROVIDER == "groq":
+        return bool(GROQ_API_KEY)
+    return bool(OLLAMA_MODEL)
+
+
+def _http_post_json(url: str, payload: dict, headers: dict, timeout: int) -> dict:
+    data = json.dumps(payload).encode("utf-8")
+    req = urllib_request.Request(
+        url,
+        data=data,
+        headers={"Content-Type": "application/json", **headers},
+        method="POST",
+    )
+    with urllib_request.urlopen(req, timeout=timeout) as response:
+        return json.loads(response.read().decode("utf-8"))
+
+
+def _ollama_generate(prompt: str, *, json_mode: bool, temperature: float, timeout: int) -> str:
     if not OLLAMA_MODEL:
         raise HTTPException(status_code=503, detail="OLLAMA_MODEL is not configured.")
-
-    payload = json.dumps({
+    payload = {
         "model": OLLAMA_MODEL,
         "prompt": prompt,
         "stream": False,
-        "format": "json",
-        "options": {"temperature": 0.1}
-    }).encode("utf-8")
+        "options": {"temperature": temperature},
+    }
+    if json_mode:
+        payload["format"] = "json"
+    body = _http_post_json(f"{OLLAMA_BASE_URL}/api/generate", payload, {}, timeout)
+    return (body.get("response") or "").strip()
 
-    req = urllib_request.Request(
-        f"{OLLAMA_BASE_URL}/api/generate",
-        data=payload,
-        headers={"Content-Type": "application/json"},
-        method="POST",
-    )
 
+def _gemini_generate(prompt: str, *, json_mode: bool, temperature: float, timeout: int) -> str:
+    if not GEMINI_API_KEY:
+        raise HTTPException(status_code=503, detail="GEMINI_API_KEY is not configured.")
+    generation_config = {"temperature": temperature}
+    if json_mode:
+        generation_config["responseMimeType"] = "application/json"
+    payload = {
+        "contents": [{"role": "user", "parts": [{"text": prompt}]}],
+        "generationConfig": generation_config,
+    }
+    # Pass the key via header (not the URL query string, which can leak into logs).
+    url = f"{GEMINI_BASE_URL}/v1beta/models/{GEMINI_MODEL}:generateContent"
+    body = _http_post_json(url, payload, {"x-goog-api-key": GEMINI_API_KEY}, timeout)
+    candidates = body.get("candidates") or []
+    if not candidates:
+        return ""
+    parts = (candidates[0].get("content") or {}).get("parts") or []
+    return "".join(part.get("text", "") for part in parts).strip()
+
+
+def _groq_generate(prompt: str, *, json_mode: bool, temperature: float, timeout: int) -> str:
+    if not GROQ_API_KEY:
+        raise HTTPException(status_code=503, detail="GROQ_API_KEY is not configured.")
+    payload = {
+        "model": GROQ_MODEL,
+        "messages": [{"role": "user", "content": prompt}],
+        "temperature": temperature,
+    }
+    if json_mode:
+        payload["response_format"] = {"type": "json_object"}
+    url = f"{GROQ_BASE_URL}/chat/completions"
+    body = _http_post_json(url, payload, {"Authorization": f"Bearer {GROQ_API_KEY}"}, timeout)
+    choices = body.get("choices") or []
+    if not choices:
+        return ""
+    return ((choices[0].get("message") or {}).get("content") or "").strip()
+
+
+def _provider_generate(prompt: str, *, json_mode: bool, temperature: float, timeout: int) -> str:
+    provider = AI_PROVIDER
     try:
-        with urllib_request.urlopen(req, timeout=120) as response:
-            body = json.loads(response.read().decode("utf-8"))
+        if provider == "gemini":
+            return _gemini_generate(prompt, json_mode=json_mode, temperature=temperature, timeout=timeout)
+        if provider == "groq":
+            return _groq_generate(prompt, json_mode=json_mode, temperature=temperature, timeout=timeout)
+        return _ollama_generate(prompt, json_mode=json_mode, temperature=temperature, timeout=timeout)
+    except HTTPException:
+        raise
     except HTTPError as ex:
-        raise HTTPException(status_code=502, detail=f"Ollama request failed with {ex.code}.")
+        raise HTTPException(status_code=502, detail=f"{_provider_display(provider)} request failed with {ex.code}.")
     except URLError as ex:
-        raise HTTPException(status_code=503, detail=f"Ollama is unreachable: {ex.reason}.")
+        raise HTTPException(status_code=503, detail=f"{_provider_display(provider)} is unreachable: {ex.reason}.")
 
-    raw = (body.get("response") or "").strip()
+
+def _ollama_generate_json(prompt: str):
+    raw = _provider_generate(prompt, json_mode=True, temperature=0.1, timeout=120)
     if not raw:
-        raise HTTPException(status_code=502, detail="Ollama returned an empty response.")
-
+        raise HTTPException(status_code=502, detail=f"{_provider_display(AI_PROVIDER)} returned an empty response.")
     try:
         return json.loads(raw)
     except json.JSONDecodeError:
@@ -428,39 +511,13 @@ def _ollama_generate_json(prompt: str):
         end = raw.rfind("}")
         if start >= 0 and end > start:
             return json.loads(raw[start:end + 1])
-        raise HTTPException(status_code=502, detail="Ollama did not return valid JSON.")
+        raise HTTPException(status_code=502, detail=f"{_provider_display(AI_PROVIDER)} did not return valid JSON.")
 
 
 def _ollama_generate_text(prompt: str) -> str:
-    if not OLLAMA_MODEL:
-        raise HTTPException(status_code=503, detail="OLLAMA_MODEL is not configured.")
-
-    payload = json.dumps({
-        "model": OLLAMA_MODEL,
-        "prompt": prompt,
-        "stream": False,
-        "options": {"temperature": 0.2}
-    }).encode("utf-8")
-
-    req = urllib_request.Request(
-        f"{OLLAMA_BASE_URL}/api/generate",
-        data=payload,
-        headers={"Content-Type": "application/json"},
-        method="POST",
-    )
-
-    try:
-        with urllib_request.urlopen(req, timeout=180) as response:
-            body = json.loads(response.read().decode("utf-8"))
-    except HTTPError as ex:
-        raise HTTPException(status_code=502, detail=f"Ollama request failed with {ex.code}.")
-    except URLError as ex:
-        raise HTTPException(status_code=503, detail=f"Ollama is unreachable: {ex.reason}.")
-
-    raw = (body.get("response") or "").strip()
+    raw = _provider_generate(prompt, json_mode=False, temperature=0.2, timeout=180)
     if not raw:
-        raise HTTPException(status_code=502, detail="Ollama returned an empty rewrite.")
-
+        raise HTTPException(status_code=502, detail=f"{_provider_display(AI_PROVIDER)} returned an empty rewrite.")
     return raw
 
 

tools/summarizer/tests/test_app.py

@@ -1,4 +1,5 @@
 import importlib
+import json
 import sys
 from pathlib import Path
 
@@ -141,3 +142,106 @@ def test_classify_block_defaults_missing_section_to_other(monkeypatch):
     assert payload["bullets"] == []
     assert payload["summary"] == []
     assert payload["skills"] == []
+
+
+# --- AI provider router -------------------------------------------------------
+
+class _FakeResponse:
+    def __init__(self, payload):
+        self._data = json.dumps(payload).encode("utf-8")
+
+    def read(self):
+        return self._data
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *exc):
+        return False
+
+
+def _install_fake_urlopen(monkeypatch, module, response_payload, captured):
+    def fake_urlopen(req, timeout=None):
+        captured["url"] = req.full_url
+        captured["headers"] = {k.lower(): v for k, v in req.header_items()}
+        captured["body"] = json.loads(req.data.decode("utf-8"))
+        return _FakeResponse(response_payload)
+
+    monkeypatch.setattr(module.urllib_request, "urlopen", fake_urlopen)
+
+
+def test_provider_defaults_to_ollama_and_is_unchanged(monkeypatch):
+    monkeypatch.delenv("AI_PROVIDER", raising=False)
+    monkeypatch.setenv("OLLAMA_BASE_URL", "http://ollama-host:11434")
+    module = load_app_module(monkeypatch, ollama_model="qwen2.5:7b")
+    assert module.AI_PROVIDER == "ollama"
+
+    captured = {}
+    _install_fake_urlopen(monkeypatch, module, {"response": '{"score": 7}'}, captured)
+
+    assert module._ollama_generate_json("hi") == {"score": 7}
+    assert captured["url"] == "http://ollama-host:11434/api/generate"
+    assert captured["body"]["model"] == "qwen2.5:7b"
+    assert captured["body"]["format"] == "json"
+    assert captured["body"]["options"]["temperature"] == 0.1
+
+
+def test_provider_gemini_dispatch(monkeypatch):
+    monkeypatch.setenv("AI_PROVIDER", "gemini")
+    monkeypatch.setenv("GEMINI_API_KEY", "test-key")
+    monkeypatch.setenv("GEMINI_MODEL", "gemini-2.0-flash")
+    module = load_app_module(monkeypatch)
+
+    captured = {}
+    payload = {"candidates": [{"content": {"parts": [{"text": '{"score": 9}'}]}}]}
+    _install_fake_urlopen(monkeypatch, module, payload, captured)
+
+    assert module._ollama_generate_json("hi") == {"score": 9}
+    assert "generativelanguage" in captured["url"]
+    assert "gemini-2.0-flash:generateContent" in captured["url"]
+    assert "key=" not in captured["url"]  # key must not be in the URL
+    assert captured["headers"].get("x-goog-api-key") == "test-key"
+    assert captured["body"]["generationConfig"]["responseMimeType"] == "application/json"
+
+
+def test_provider_groq_dispatch(monkeypatch):
+    monkeypatch.setenv("AI_PROVIDER", "groq")
+    monkeypatch.setenv("GROQ_API_KEY", "test-key")
+    module = load_app_module(monkeypatch)
+
+    captured = {}
+    payload = {"choices": [{"message": {"content": "rewritten CV text"}}]}
+    _install_fake_urlopen(monkeypatch, module, payload, captured)
+
+    assert module._ollama_generate_text("rewrite this") == "rewritten CV text"
+    assert captured["url"].endswith("/chat/completions")
+    assert captured["headers"].get("authorization") == "Bearer test-key"
+    assert captured["body"]["messages"][0]["content"] == "rewrite this"
+
+
+def test_provider_missing_cloud_key_raises_503(monkeypatch):
+    monkeypatch.setenv("AI_PROVIDER", "gemini")
+    monkeypatch.delenv("GEMINI_API_KEY", raising=False)
+    module = load_app_module(monkeypatch)
+
+    from fastapi import HTTPException
+
+    try:
+        module._ollama_generate_json("hi")
+    except HTTPException as ex:
+        assert ex.status_code == 503
+        assert "GEMINI_API_KEY" in ex.detail
+    else:
+        raise AssertionError("expected HTTPException for missing GEMINI_API_KEY")
+
+
+def test_health_reports_active_provider(monkeypatch):
+    monkeypatch.setenv("AI_PROVIDER", "gemini")
+    monkeypatch.setenv("GEMINI_API_KEY", "test-key")
+    module = load_app_module(monkeypatch)
+    client = TestClient(module.app)
+
+    payload = client.get("/health").json()
+
+    assert payload["ai_provider"] == "gemini"
+    assert payload["ai_provider_configured"] is True