feat: upload audio, get a private Swedish transcript, discuss it (#33, ADR-033) (#39)

Upload a recording (audio, or video with the camera off), get a private Swedish transcript, and chat about it.

- ADR-033: RLS-scoped private transcript store; ffprobe/ffmpeg intake; gateway transcription (iguana kb-whisper, berget fallback); in-process single-replica worker; chat over the transcript (local models only).
- ADR-034: optional two-engine mode (pianissimo + kb-whisper) reconciled by a LOCAL model; off unless TAPIR_RECONCILE_BASE_MODEL is set.
- ADR-035: real progress (phases, running/quiet/lost), phase logging, every judged spot stored, merged segments.
- Hardening from four adversarial reviews: errors logged and stored as kinds never text; per-engine deadlines; converted-length check; bounded queue; client hang-ups are not errors; strict local-alias check.
- Dockerfile: static ffmpeg/ffprobe 7.0.2 pinned by sha256 (#34). Built and inspected by hand; CI does not build the image on a PR, so its first CI build is this merge.

Uploads stay OFF until TAPIR_UPLOADS_ENABLED is set in the deployment.

Closes #33, #34, #42, #43, #44.

Co-authored-by: Mathias <mthbqv@gmail.com>

cmd/tapir/e2e_upload_test.go

@@ -0,0 +1,209 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"mime/multipart"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	embeddedpostgres "github.com/fergusstrange/embedded-postgres"
+	"github.com/jackc/pgx/v5/pgxpool"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/config"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/web"
+)
+
+// TestE2E_UploadTranscribeAndChat is the end-to-end check for tapir#33: upload a
+// real audio file, watch it become a transcript through the REAL ffprobe, ffmpeg
+// and gateway, then ask a question and get a grounded answer from a REAL model.
+// It is opt-in — it needs real binaries and the live gateway — and skipped unless
+// TAPIR_E2E_AUDIO is set:
+//
+//	TAPIR_E2E_AUDIO=/path/meeting.m4a TAPIR_E2E_FFMPEG=/path/ffmpeg \
+//	TAPIR_E2E_FFPROBE=/path/ffprobe TAPIR_E2E_QUESTION='Vem ansvarar för virket?' \
+//	TAPIR_E2E_EXPECT=anders op run --env-file ~/.op-env -- \
+//	  go test ./cmd/tapir -run E2E_Upload -v -count=1
+//
+// Use SYNTHETIC audio only: this repo is public, and a fallback to berget sends
+// the audio to a cloud endpoint (#33).
+func TestE2E_UploadTranscribeAndChat(t *testing.T) {
+	audioPath := os.Getenv("TAPIR_E2E_AUDIO")
+	if audioPath == "" {
+		t.Skip("set TAPIR_E2E_AUDIO (and friends) to run the live end-to-end check")
+	}
+	key := os.Getenv("LITELLM_MASTER_KEY")
+	require.NotEmpty(t, key, "LITELLM_MASTER_KEY must be set (via op run)")
+	ctx, cancel := context.WithCancel(context.Background())
+	t.Cleanup(cancel)
+
+	// A private embedded Postgres: nothing here may touch a real database.
+	port := uint32(55000 + os.Getpid()%1000)
+	rt := filepath.Join(os.TempDir(), fmt.Sprintf("tapir-epg-e2e-%d", os.Getpid()))
+	pg := embeddedpostgres.NewDatabase(embeddedpostgres.DefaultConfig().Port(port).
+		RuntimePath(rt).DataPath(filepath.Join(rt, "data")).BinariesPath(filepath.Join(rt, "bin")).
+		CachePath(filepath.Join(os.TempDir(), "tapir-epg-cache")))
+	require.NoError(t, pg.Start())
+	t.Cleanup(func() { _ = pg.Stop(); _ = os.RemoveAll(rt) })
+	dsn := fmt.Sprintf("postgres://postgres:postgres@localhost:%d/postgres?sslmode=disable", port)
+
+	st, err := store.New(ctx, dsn)
+	require.NoError(t, err)
+	t.Cleanup(st.Close)
+	raw, err := pgxpool.New(ctx, dsn)
+	require.NoError(t, err)
+	t.Cleanup(raw.Close)
+	const uid = "11111111-1111-1111-1111-111111111111"
+	_, err = raw.Exec(ctx, `INSERT INTO users (id) VALUES ($1)`, uid)
+	require.NoError(t, err)
+	_, err = raw.Exec(ctx, `INSERT INTO user_identities (dex_subject, user_id) VALUES ('e2e', $1)`, uid)
+	require.NoError(t, err)
+
+	cfg, err := config.Load()
+	require.NoError(t, err)
+	cfg.GatewayKey = key
+	cfg.UploadsEnabled = true
+	cfg.UploadTmpDir = filepath.Join(t.TempDir(), "spool")
+	if os.Getenv("TAPIR_E2E_RECONCILE") == "1" {
+		// Two engines and a local model, as ADR-034 describes.
+		cfg.ReconcileBaseModel = envOrDefault("TAPIR_E2E_BASE_MODEL", "iguana/pianissimo")
+		cfg.ReconcileModel = envOrDefault("TAPIR_E2E_RECONCILE_MODEL", "iguana/gemma4-26b")
+	}
+	cfg.FFmpegBin = envOrDefault("TAPIR_E2E_FFMPEG", "ffmpeg")
+	cfg.FFprobeBin = envOrDefault("TAPIR_E2E_FFPROBE", "ffprobe")
+
+	intake, worker, err := buildUploads(ctx, cfg, st, quietLog())
+	require.NoError(t, err)
+	go worker.Run(ctx)
+	app := &web.App{
+		Store: st, Identity: st, Auth: web.StubAuth{U: web.User{Subject: "e2e"}},
+		Chat: buildChat(cfg),
+	}
+	wireUploads(app, cfg, intake) // the same wiring cmdServe uses
+	srv := httptest.NewServer(app.Router())
+	t.Cleanup(srv.Close)
+	noRedirect := &http.Client{CheckRedirect: func(*http.Request, []*http.Request) error { return http.ErrUseLastResponse }}
+
+	// 1. Upload.
+	f, err := os.Open(audioPath)
+	require.NoError(t, err)
+	defer func() { _ = f.Close() }()
+	pr, pw := io.Pipe()
+	mw := multipart.NewWriter(pw)
+	go func() {
+		_ = mw.WriteField("title", "E2E recording")
+		fw, _ := mw.CreateFormFile("file", filepath.Base(audioPath))
+		_, _ = io.Copy(fw, f)
+		_ = mw.Close()
+		_ = pw.Close()
+	}()
+	req, err := http.NewRequest(http.MethodPost, srv.URL+"/uploads", pr)
+	require.NoError(t, err)
+	req.Header.Set("Content-Type", mw.FormDataContentType())
+	resp, err := noRedirect.Do(req)
+	require.NoError(t, err)
+	_ = resp.Body.Close()
+	require.Equal(t, http.StatusSeeOther, resp.StatusCode, "an accepted upload redirects to its page")
+	loc := resp.Header.Get("Location")
+	require.True(t, strings.HasPrefix(loc, "/uploads/"), loc)
+	t.Logf("accepted -> %s", loc)
+
+	// 2. Wait for the transcript: unfinished states poll, terminal states do not.
+	var page string
+	require.Eventually(t, func() bool {
+		r, err := http.Get(srv.URL + loc + "/status")
+		if err != nil {
+			return false
+		}
+		b, _ := io.ReadAll(r.Body)
+		_ = r.Body.Close()
+		page = string(b)
+		return !strings.Contains(page, "hx-trigger")
+	}, 12*time.Minute, 3*time.Second, "the upload never reached a terminal state")
+	require.Contains(t, page, "Transcribed", "terminal state was not 'done': %.300s", page)
+
+	var status, endpoint, content string
+	var secs float64
+	var hadVideo bool
+	require.NoError(t, raw.QueryRow(ctx,
+		`SELECT status, endpoint, content, audio_seconds, had_video FROM uploaded_transcripts WHERE user_id = $1`, uid).
+		Scan(&status, &endpoint, &content, &secs, &hadVideo))
+	t.Logf("done: endpoint=%s audio_s=%.1f transcript_chars=%d had_video=%v", endpoint, secs, len(content), hadVideo)
+	if os.Getenv("TAPIR_E2E_RECONCILE") == "1" {
+		var recJSON string
+		require.NoError(t, raw.QueryRow(ctx,
+			`SELECT COALESCE(reconciliation::text, 'NULL') FROM uploaded_transcripts WHERE user_id = $1`, uid).Scan(&recJSON))
+		require.NotEqual(t, "NULL", recJSON, "a two-engine run must record its reconciliation")
+		t.Logf("reconciliation: %.600s", recJSON)
+		assert.Contains(t, endpoint, "reconciled by", "every contributor is named in the endpoint")
+		assert.Contains(t, page, "Reconciled from two transcripts")
+	}
+	if os.Getenv("TAPIR_E2E_HAD_VIDEO") == "1" {
+		// A Teams meeting with the camera off: a video file whose audio is what matters.
+		assert.True(t, hadVideo, "the file had a video track and the row must say so")
+		assert.Contains(t, page, "Only its audio was used", "and the user must be told")
+	}
+	assert.Equal(t, "done", status)
+	assert.NotEmpty(t, endpoint)
+	assert.NotEmpty(t, content)
+
+	// ADR-035: a finished job's phases are all closed, and it was reporting while it ran.
+	var progJSON string
+	var progressAt, aliveAt *time.Time
+	require.NoError(t, raw.QueryRow(ctx,
+		`SELECT COALESCE(progress::text, '[]'), progress_at, alive_at FROM uploaded_transcripts WHERE user_id = $1`, uid).
+		Scan(&progJSON, &progressAt, &aliveAt))
+	var steps []domain.ProgressStep
+	require.NoError(t, json.Unmarshal([]byte(progJSON), &steps))
+	require.NotEmpty(t, steps, "the job recorded no phases")
+	for _, st := range steps {
+		assert.Contains(t, []domain.StepState{domain.StepDone, domain.StepSkipped}, st.State,
+			"phase %q was left %q after the job finished", st.Key, st.State)
+	}
+	assert.Equal(t, "save", steps[len(steps)-1].Key)
+	assert.NotNil(t, progressAt, "progress_at must have been stamped")
+	t.Logf("phases: %d, heartbeat stamped: %v", len(steps), aliveAt != nil)
+	entries, err := os.ReadDir(cfg.UploadTmpDir)
+	require.NoError(t, err)
+	assert.Empty(t, entries, "the audio must be deleted once transcribed")
+
+	// 3. Ask a question; the answer must come from the transcript.
+	q := envOrDefault("TAPIR_E2E_QUESTION", "Who is responsible for the timber order?")
+	chatResp, err := http.PostForm(srv.URL+loc+"/chat", url.Values{"question": {q}})
+	require.NoError(t, err)
+	answerPage, _ := io.ReadAll(chatResp.Body)
+	_ = chatResp.Body.Close()
+	require.Equal(t, http.StatusOK, chatResp.StatusCode)
+	t.Logf("question: %s", q)
+	if want := os.Getenv("TAPIR_E2E_EXPECT"); want != "" {
+		assert.Contains(t, strings.ToLower(string(answerPage)), strings.ToLower(want),
+			"the answer should be grounded in the transcript")
+	}
+	t.Logf("answer page excerpt: %.600s", afterMarker(string(answerPage), `class="chat-turn chat-a"`))
+}
+
+func envOrDefault(k, d string) string {
+	if v := os.Getenv(k); v != "" {
+		return v
+	}
+	return d
+}
+
+func afterMarker(s, m string) string {
+	if i := strings.Index(s, m); i >= 0 {
+		return s[i:]
+	}
+	return "(no answer turn found)"
+}

cmd/tapir/main.go

@@ -207,6 +207,26 @@ func cmdServe(ctx context.Context, log *slog.Logger) error {
 		log.Info("web chat enabled (stored-transcript only)", "models", chatModels(cfg))
 	}
 
+	// Audio upload -> private transcript (ADR-033). Off unless TAPIR_UPLOADS_ENABLED:
+	// the image must carry ffmpeg first. The worker is IN-PROCESS and one job at a
+	// time, so it inherits the single-replica assumption below (its queue is
+	// process-local); it stops with the same ctx as the server.
+	intake, worker, err := buildUploads(ctx, cfg, st, log)
+	if err != nil {
+		return err
+	}
+	wireUploads(app, cfg, intake)
+	if intake != nil {
+		go worker.Run(ctx)
+		log.Info("audio upload enabled", "model", cfg.TranscribeModel, "fallback", cfg.TranscribeFallbackModel,
+			"max_mb", cfg.UploadMaxBytes>>20, "max_minutes", int(cfg.UploadMaxDuration.Minutes()))
+		log.Warn("audio upload assumes a SINGLE replica: the transcription queue is process-local (ADR-033, ADR-018)")
+		if cfg.TranscribeFallbackModel != "" {
+			log.Warn("audio upload fallback is a CLOUD endpoint: if the primary fails, private audio leaves the homelab",
+				"fallback", cfg.TranscribeFallbackModel)
+		}
+	}
+
 	// User onboarding is handled by the IdP (Authentik invite flow), not Tapir —
 	// the Dex local-password provisioning path was removed (ADR-019). An
 	// authenticated subject with no Tapir user is routed to /register.

cmd/tapir/processor.go

@@ -134,9 +134,30 @@ func chatModels(cfg config.Config) []string {
 	add(cfg.SummarizerModel)
 	add(cfg.FallbackModel)
 	add(cfg.CloudFallbackModel)
+	// The model chat over an uploaded recording opens on must be one the shared service can
+	// call: it re-maps a model it does not offer, which silently swapped a configured 31B for
+	// phi4-mini, whose context cannot hold an hour of speech. Local only, and only when
+	// uploads are on, so the video chat is unchanged otherwise (second adversarial review).
+	if cfg.UploadsEnabled && config.IsLocalModel(cfg.UploadChatModel) {
+		add(cfg.UploadChatModel)
+	}
 	return models
 }
 
+// uploadChatModels is the chat switcher for a PRIVATE uploaded recording: the same
+// ordered set as chatModels, local models only. The whole transcript (up to 90,000
+// characters) is sent to the one chosen, so a cloud model the video chat offers must not
+// be offered here; with none local, none is offered (ADR-035 review).
+func uploadChatModels(cfg config.Config) []string {
+	var local []string
+	for _, m := range chatModels(cfg) {
+		if config.IsLocalModel(m) {
+			local = append(local, m)
+		}
+	}
+	return local
+}
+
 // buildChat wires the per-video chat service (ADR-027): a Completer factory over
 // the SAME LiteLLM gateway the summarizer uses (a different alias per model, not a
 // second client config) and the same transcript-truncation budget. It returns nil

cmd/tapir/uploads.go

@@ -0,0 +1,142 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"log/slog"
+	"net/http"
+	"os"
+	"path/filepath"
+	"time"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/audio"
+	"git.d-ma.be/mathias/tapir/internal/adapters/llm"
+	"git.d-ma.be/mathias/tapir/internal/adapters/transcribe"
+	"git.d-ma.be/mathias/tapir/internal/config"
+	"git.d-ma.be/mathias/tapir/internal/reconcile"
+	"git.d-ma.be/mathias/tapir/internal/uploads"
+	"git.d-ma.be/mathias/tapir/internal/web"
+)
+
+// uploadsStore is the store surface the upload wiring needs. *store.Store
+// satisfies it.
+type uploadsStore interface {
+	uploads.Store
+	uploads.WorkerStore
+	FailInterruptedUploads(ctx context.Context) (int, error)
+}
+
+// buildUploads wires audio upload and private transcription (ADR-033). It returns
+// nils when uploads are disabled — never a typed-nil Intake, which would mount the
+// /uploads routes over nothing. When enabled it, in order: prepares a private spool
+// directory, deletes any audio a crashed run left in it, fails the rows such a run
+// left queued or transcribing (the audio is not retained, so they can never
+// resume), and only then builds the intake and the worker. The caller runs the
+// worker. The fallback endpoint is a cloud service; an empty
+// TAPIR_TRANSCRIBE_FALLBACK_MODEL keeps private audio on the homelab.
+func buildUploads(ctx context.Context, cfg config.Config, st uploadsStore, log *slog.Logger) (*uploads.Intake, *uploads.Worker, error) {
+	if !cfg.UploadsEnabled {
+		return nil, nil, nil
+	}
+	if cfg.GatewayURL == "" {
+		return nil, nil, fmt.Errorf("uploads: TAPIR_UPLOADS_ENABLED is set but there is no TAPIR_GATEWAY_URL to transcribe through")
+	}
+
+	if err := os.MkdirAll(cfg.UploadTmpDir, 0o700); err != nil {
+		return nil, nil, fmt.Errorf("uploads: spool dir: %w", err)
+	}
+	removeSpooledAudio(cfg.UploadTmpDir, log)
+
+	n, err := st.FailInterruptedUploads(ctx)
+	if err != nil {
+		return nil, nil, fmt.Errorf("uploads: fail interrupted rows: %w", err)
+	}
+	if n > 0 {
+		log.Warn("uploads: marked interrupted by a restart (audio is not retained; they must be re-uploaded)", "count", n)
+	}
+
+	worker := uploads.NewWorker(st, newConverter(cfg), buildTranscriber(cfg),
+		cfg.UploadTmpDir, cfg.TranscribeTimeout, uploads.WithMaxDuration(cfg.UploadMaxDuration))
+	intake := uploads.NewIntake(st, &audio.FFprobe{Bin: cfg.FFprobeBin}, worker,
+		uploadLimits(cfg), cfg.UploadTmpDir)
+	return intake, worker, nil
+}
+
+func uploadLimits(cfg config.Config) uploads.Limits {
+	return uploads.Limits{MaxBytes: cfg.UploadMaxBytes, MaxDuration: cfg.UploadMaxDuration}
+}
+
+// removeSpooledAudio deletes the spool files this process family creates
+// (upload-*, wav-*) and nothing else.
+func removeSpooledAudio(dir string, log *slog.Logger) {
+	for _, pat := range []string{"upload-*", "wav-*"} {
+		matches, _ := filepath.Glob(filepath.Join(dir, pat))
+		for _, m := range matches {
+			if err := os.Remove(m); err != nil {
+				log.Warn("uploads: remove leftover audio", "path", m, "err", err)
+			}
+		}
+	}
+}
+
+// wireUploads connects a built Intake to the web App: the routes, the limits the
+// form states, and the chat settings for an uploaded recording. A nil intake
+// leaves the App untouched — assigning a nil *Intake to the Uploader interface
+// would produce a non-nil interface and mount the routes over nothing.
+func wireUploads(app *web.App, cfg config.Config, in *uploads.Intake) {
+	if in == nil {
+		return
+	}
+	app.Uploads = in
+	app.UploadLimits = uploadLimits(cfg)
+	app.UploadChatMaxChars = cfg.MaxUploadedTranscriptChars
+	app.UploadChatModel = cfg.UploadChatModel
+	app.UploadChatModels = uploadChatModels(cfg)
+}
+
+// newConverter is the ffmpeg adapter capped at the configured maximum duration, so a file
+// whose header lies about its length cannot become hours of WAV (ADR-035 review).
+func newConverter(cfg config.Config) *audio.FFmpeg {
+	return &audio.FFmpeg{Bin: cfg.FFmpegBin, MaxSeconds: int(cfg.UploadMaxDuration.Seconds())}
+}
+
+// adjudicatorTimeout bounds one adjudication call. A call takes ~2 s on gemma4-26b and ~10 s
+// on 31b; this is the wait before an unresponsive model is counted as an error.
+const adjudicatorTimeout = 2 * time.Minute
+
+// buildTranscriber builds what the worker transcribes with. By default that is the
+// single-engine chain (the primary, then the optional cloud fallback). When
+// TAPIR_RECONCILE_BASE_MODEL is set it is the two-engine ensemble (ADR-034): the base
+// engine and the primary run on the same WAV and a LOCAL model reconciles them. Config
+// has already refused any non-local model, so nothing here can send private audio or
+// text to a cloud model for reconciliation; the cloud fallback is unchanged.
+func buildTranscriber(cfg config.Config) uploads.Transcriber {
+	// No overall client timeout: a 90 minute recording takes minutes. The deadline is
+	// the worker's per-job context.
+	client := &http.Client{}
+	// Each LOCAL engine call gets its own deadline: a wedged engine does not answer, and
+	// without one it used the whole job before the fallback could run.
+	local := func(model string) transcribe.Backend {
+		return transcribe.WithTimeout(transcribe.NewGateway(cfg.GatewayURL, cfg.GatewayKey, model, client), cfg.TranscribeEngineTimeout)
+	}
+	primary := local(cfg.TranscribeModel)
+	var fallback transcribe.Backend
+	if cfg.TranscribeFallbackModel != "" && cfg.TranscribeFallbackModel != cfg.TranscribeModel {
+		fallback = transcribe.NewGateway(cfg.GatewayURL, cfg.GatewayKey, cfg.TranscribeFallbackModel, client)
+	}
+	if cfg.ReconcileBaseModel == "" {
+		return transcribe.NewChain(primary, fallback)
+	}
+
+	base := local(cfg.ReconcileBaseModel)
+	// Temperature 0 and a short reply: the adjudicator's accuracy was measured that way.
+	chat := llm.New(cfg.GatewayURL, cfg.GatewayKey, cfg.ReconcileModel, adjudicatorTimeout,
+		llm.WithTemperature(0), llm.WithMaxTokens(200))
+	var adj reconcile.Adjudicator = reconcile.NewLLMAdjudicator(chat)
+	if cfg.ReconcileCheckOrder {
+		adj = reconcile.NewOrderChecked(adj)
+	}
+	return transcribe.NewEnsemble(base, primary, fallback, adj, reconcile.Options{
+		BaseEngine: cfg.ReconcileBaseModel, AltEngine: cfg.TranscribeModel, Reconciler: cfg.ReconcileModel,
+	})
+}

cmd/tapir/uploads_chat_models_test.go

@@ -0,0 +1,85 @@
+package main
+
+import (
+	"reflect"
+	"testing"
+
+	"git.d-ma.be/mathias/tapir/internal/config"
+)
+
+// The chat over an uploaded recording sends the whole private transcript to the model
+// the user picks, so the switcher for it may offer local models only (ADR-035 review).
+
+func TestUploadChatModels_NeverOffersACloudModelButKeepsTheLocalOrder(t *testing.T) {
+	var cfg config.Config
+	cfg.SummarizerModel = "koala/phi4-mini"
+	cfg.FallbackModel = "iguana/gemma4-26b"
+	cfg.CloudFallbackModel = "berget/mistral-small"
+
+	got := uploadChatModels(cfg)
+
+	want := []string{"koala/phi4-mini", "iguana/gemma4-26b"}
+	if !reflect.DeepEqual(got, want) {
+		t.Errorf("uploadChatModels = %v, want %v: the cloud model that the video chat offers must not be offered for a private recording", got, want)
+	}
+}
+
+func TestUploadChatModels_ACloudPrimaryIsDroppedToo(t *testing.T) {
+	var cfg config.Config
+	cfg.SummarizerModel = "berget/mistral-small"
+	cfg.FallbackModel = "iguana/gemma4-26b"
+
+	got := uploadChatModels(cfg)
+
+	if !reflect.DeepEqual(got, []string{"iguana/gemma4-26b"}) {
+		t.Errorf("uploadChatModels = %v", got)
+	}
+}
+
+func TestUploadChatModels_NoLocalModelMeansNoneOfferedNotACloudDefault(t *testing.T) {
+	var cfg config.Config
+	cfg.SummarizerModel = "berget/mistral-small"
+
+	if got := uploadChatModels(cfg); len(got) != 0 {
+		t.Errorf("uploadChatModels = %v, want none", got)
+	}
+}
+
+// A valid local TAPIR_UPLOAD_CHAT_MODEL that is not in the video chat's chain used to be
+// accepted by config and then silently ignored: the shared chat service re-maps a model it
+// does not offer, so the user got koala/phi4-mini, whose 8,192-token context cannot hold an
+// hour of speech (second adversarial review).
+
+func TestChatModels_AnUploadChatModelOutsideTheChainIsOfferedWhenUploadsAreOn(t *testing.T) {
+	var cfg config.Config
+	cfg.SummarizerModel = "koala/phi4-mini"
+	cfg.FallbackModel = "iguana/gemma4-26b"
+	cfg.UploadsEnabled = true
+	cfg.UploadChatModel = "iguana/gemma4-31b"
+
+	service := chatModels(cfg)
+	upload := uploadChatModels(cfg)
+
+	if !reflect.DeepEqual(service, []string{"koala/phi4-mini", "iguana/gemma4-26b", "iguana/gemma4-31b"}) {
+		t.Errorf("chatModels = %v: the shared service must be able to call the configured upload model", service)
+	}
+	if !reflect.DeepEqual(upload, service) {
+		t.Errorf("uploadChatModels = %v, want %v", upload, service)
+	}
+}
+
+func TestChatModels_NothingIsAddedWhenUploadsAreOffOrTheModelIsNotLocal(t *testing.T) {
+	var cfg config.Config
+	cfg.SummarizerModel = "koala/phi4-mini"
+	cfg.UploadChatModel = "iguana/gemma4-31b"
+
+	if got := chatModels(cfg); !reflect.DeepEqual(got, []string{"koala/phi4-mini"}) {
+		t.Errorf("uploads off: chatModels = %v, want the video chain unchanged", got)
+	}
+
+	cfg.UploadsEnabled = true
+	cfg.UploadChatModel = "berget/mistral-small"
+	if got := chatModels(cfg); !reflect.DeepEqual(got, []string{"koala/phi4-mini"}) {
+		t.Errorf("a cloud upload chat model was added to the service: %v", got)
+	}
+}

cmd/tapir/uploads_converter_test.go

@@ -0,0 +1,26 @@
+package main
+
+import (
+	"testing"
+	"time"
+
+	"git.d-ma.be/mathias/tapir/internal/config"
+)
+
+// The converter is told the configured maximum, so a file whose header lies about its
+// length cannot become hours of WAV (adversarial review F2).
+
+func TestNewConverter_IsCappedAtTheConfiguredMaximumDuration(t *testing.T) {
+	var cfg config.Config
+	cfg.FFmpegBin = "/usr/local/bin/ffmpeg"
+	cfg.UploadMaxDuration = 90 * time.Minute
+
+	c := newConverter(cfg)
+
+	if c.Bin != "/usr/local/bin/ffmpeg" {
+		t.Errorf("Bin = %q", c.Bin)
+	}
+	if c.MaxSeconds != 5400 {
+		t.Errorf("MaxSeconds = %v, want 5400", c.MaxSeconds)
+	}
+}

cmd/tapir/uploads_test.go

@@ -0,0 +1,196 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/transcribe"
+	"git.d-ma.be/mathias/tapir/internal/config"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/web"
+)
+
+// fakeUploadsStore satisfies the store surface buildUploads needs, and counts the
+// start-up sweep.
+type fakeUploadsStore struct {
+	sweeps      int
+	sweepErr    error
+	failedKinds []domain.UploadErrorKind
+}
+
+func (f *fakeUploadsStore) CreateUpload(context.Context, string, string, bool) (string, error) {
+	return "id", nil
+}
+
+func (f *fakeUploadsStore) FailUpload(_ context.Context, _, _ string, k domain.UploadErrorKind) error {
+	f.failedKinds = append(f.failedKinds, k)
+	return nil
+}
+
+func (f *fakeUploadsStore) StartTranscribing(context.Context, string, string) error { return nil }
+
+func (f *fakeUploadsStore) CompleteUpload(context.Context, string, string, domain.Transcription) error {
+	return nil
+}
+
+func (f *fakeUploadsStore) SaveProgress(context.Context, string, string, []domain.ProgressStep, time.Time) error {
+	return nil
+}
+
+func (f *fakeUploadsStore) Heartbeat(context.Context, string, string) error { return nil }
+
+func (f *fakeUploadsStore) FailInterruptedUploads(context.Context) (int, error) {
+	f.sweeps++
+	return 2, f.sweepErr
+}
+
+func uploadCfg(t *testing.T, enabled bool) config.Config {
+	t.Helper()
+	var cfg config.Config
+	cfg.GatewayURL = "http://gateway.invalid/v1"
+	cfg.UploadsEnabled = enabled
+	cfg.UploadMaxBytes = 200 << 20
+	cfg.UploadTmpDir = filepath.Join(t.TempDir(), "spool")
+	cfg.TranscribeModel = "iguana/kb-whisper"
+	cfg.TranscribeFallbackModel = "berget/kb-whisper"
+	return cfg
+}
+
+func TestBuildUploads_DisabledMountsNothingAndSweepsNothing(t *testing.T) {
+	st := &fakeUploadsStore{}
+
+	in, w, err := buildUploads(context.Background(), uploadCfg(t, false), st, quietLog())
+
+	require.NoError(t, err)
+	assert.Nil(t, in, "a typed-nil Intake must never reach web.App.Uploads: it would mount the routes over nothing")
+	assert.Nil(t, w)
+	assert.Zero(t, st.sweeps)
+}
+
+func TestBuildUploads_EnabledFailsInterruptedRowsBeforeAnythingElseStarts(t *testing.T) {
+	st := &fakeUploadsStore{}
+
+	in, w, err := buildUploads(context.Background(), uploadCfg(t, true), st, quietLog())
+
+	require.NoError(t, err)
+	require.NotNil(t, in)
+	require.NotNil(t, w)
+	assert.Equal(t, 1, st.sweeps,
+		"audio is not retained, so rows left queued/transcribing by a restart can never resume: fail them at start-up, not never")
+}
+
+func TestBuildUploads_SweepFailureStopsStartup(t *testing.T) {
+	st := &fakeUploadsStore{sweepErr: errors.New("db down")}
+
+	_, _, err := buildUploads(context.Background(), uploadCfg(t, true), st, quietLog())
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "db down")
+}
+
+func TestBuildUploads_CreatesTheSpoolDirectoryPrivately(t *testing.T) {
+	cfg := uploadCfg(t, true)
+
+	_, _, err := buildUploads(context.Background(), cfg, &fakeUploadsStore{}, quietLog())
+
+	require.NoError(t, err)
+	info, err := os.Stat(cfg.UploadTmpDir)
+	require.NoError(t, err)
+	assert.True(t, info.IsDir())
+	assert.Equal(t, os.FileMode(0o700), info.Mode().Perm(), "spooled audio is private third-party content")
+}
+
+func TestBuildUploads_EnabledWithoutAGatewayIsAnError(t *testing.T) {
+	cfg := uploadCfg(t, true)
+	cfg.GatewayURL = ""
+
+	_, _, err := buildUploads(context.Background(), cfg, &fakeUploadsStore{}, quietLog())
+
+	assert.Error(t, err, "uploads on with nowhere to transcribe must fail loudly at start-up, not at the first upload")
+}
+
+func TestBuildUploads_RemovesAudioLeftBehindByACrash(t *testing.T) {
+	cfg := uploadCfg(t, true)
+	require.NoError(t, os.MkdirAll(cfg.UploadTmpDir, 0o700))
+	leftovers := []string{"upload-123", "wav-456"}
+	for _, n := range leftovers {
+		require.NoError(t, os.WriteFile(filepath.Join(cfg.UploadTmpDir, n), []byte("audio"), 0o600))
+	}
+	unrelated := filepath.Join(cfg.UploadTmpDir, "someone-elses-file")
+	require.NoError(t, os.WriteFile(unrelated, []byte("keep"), 0o600))
+
+	_, _, err := buildUploads(context.Background(), cfg, &fakeUploadsStore{}, quietLog())
+
+	require.NoError(t, err)
+	for _, n := range leftovers {
+		_, err := os.Stat(filepath.Join(cfg.UploadTmpDir, n))
+		assert.True(t, os.IsNotExist(err), "%s: audio is never retained, so a spooled file from a crashed run must not survive start-up", n)
+	}
+	_, err = os.Stat(unrelated)
+	assert.NoError(t, err, "only our own spool files are removed")
+}
+
+func TestWireUploads_SetsEverythingTheWebAppNeedsForUploadsAndChat(t *testing.T) {
+	cfg := uploadCfg(t, true)
+	cfg.UploadMaxDuration = 90 * 60 * 1e9
+	cfg.MaxUploadedTranscriptChars = 90000
+	cfg.UploadChatModel = "iguana/gemma4-26b"
+	in, _, err := buildUploads(context.Background(), cfg, &fakeUploadsStore{}, quietLog())
+	require.NoError(t, err)
+	app := &web.App{}
+
+	wireUploads(app, cfg, in)
+
+	assert.NotNil(t, app.Uploads)
+	assert.Equal(t, cfg.UploadMaxBytes, app.UploadLimits.MaxBytes, "the form states the limits Intake enforces")
+	assert.Equal(t, cfg.UploadMaxDuration, app.UploadLimits.MaxDuration)
+	assert.Equal(t, 90000, app.UploadChatMaxChars,
+		"without this the chat would use the 18,000-char caption default and cut an hour to its first third")
+	assert.Equal(t, "iguana/gemma4-26b", app.UploadChatModel)
+}
+
+func TestWireUploads_NilIntakeLeavesTheAppUntouched(t *testing.T) {
+	app := &web.App{}
+
+	wireUploads(app, uploadCfg(t, false), nil)
+
+	assert.Nil(t, app.Uploads, "a nil *Intake must not become a non-nil interface that mounts the routes")
+	assert.Zero(t, app.UploadChatMaxChars)
+}
+
+func TestBuildTranscriber_DefaultIsTheSingleEngineChainAsBefore(t *testing.T) {
+	cfg := uploadCfg(t, true)
+
+	tr := buildTranscriber(cfg)
+
+	assert.IsType(t, &transcribe.Chain{}, tr, "with no base engine configured nothing changes")
+}
+
+func TestBuildTranscriber_ABaseEngineTurnsOnTheEnsemble(t *testing.T) {
+	cfg := uploadCfg(t, true)
+	cfg.ReconcileBaseModel = "iguana/pianissimo"
+	cfg.ReconcileModel = "iguana/gemma4-26b"
+
+	tr := buildTranscriber(cfg)
+
+	assert.IsType(t, &transcribe.Ensemble{}, tr)
+}
+
+func TestBuildUploads_WiresTheEnsembleWhenConfigured(t *testing.T) {
+	cfg := uploadCfg(t, true)
+	cfg.ReconcileBaseModel = "iguana/pianissimo"
+	cfg.ReconcileModel = "iguana/gemma4-26b"
+
+	in, w, err := buildUploads(context.Background(), cfg, &fakeUploadsStore{}, quietLog())
+
+	require.NoError(t, err)
+	assert.NotNil(t, in)
+	assert.NotNil(t, w)
+}

cmd/tapir/uploads_timeout_test.go

@@ -0,0 +1,78 @@
+package main
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// The wiring of the per-engine deadline: a gateway whose primary model never answers
+// must still reach the fallback model inside the job's own deadline (adversarial review
+// F4). The fake gateway answers per model from the multipart `model` field.
+
+func TestBuildTranscriber_AWedgedPrimaryReachesTheFallbackBeforeTheJobDeadline(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_ = r.ParseMultipartForm(1 << 20)
+		if r.FormValue("model") == "iguana/kb-whisper" {
+			<-r.Context().Done() // wedged: never answers
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"text":"hej","duration":1,"segments":[{"start":0,"end":1,"text":"hej"}]}`))
+	}))
+	defer srv.Close()
+	cfg := uploadCfg(t, true)
+	cfg.GatewayURL = srv.URL
+	cfg.TranscribeModel = "iguana/kb-whisper"
+	cfg.TranscribeFallbackModel = "berget/kb-whisper"
+	cfg.TranscribeEngineTimeout = 150 * time.Millisecond
+	wav := filepath.Join(t.TempDir(), "a.wav")
+	require.NoError(t, os.WriteFile(wav, []byte("RIFF....WAVEfmt "), 0o600))
+	job, cancel := context.WithTimeout(context.Background(), 20*time.Second)
+	defer cancel()
+
+	got, err := buildTranscriber(cfg).Transcribe(job, wav)
+
+	require.NoError(t, err, "the wedged primary used the whole job and the fallback never ran")
+	assert.Equal(t, "berget/kb-whisper", got.Endpoint)
+}
+
+func TestBuildTranscriber_AWedgedBaseEngineIsASoftFailureInTheEnsemble(t *testing.T) {
+	// The same deadline must guard the ensemble's BASE engine, not only the chain's primary
+	// (second adversarial review: the wiring was tested for the chain alone).
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_ = r.ParseMultipartForm(1 << 20)
+		if r.FormValue("model") == "iguana/pianissimo" {
+			<-r.Context().Done()
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"text":"hej","duration":1,"segments":[{"start":0,"end":1,"text":"hej"}]}`))
+	}))
+	defer srv.Close()
+	cfg := uploadCfg(t, true)
+	cfg.GatewayURL = srv.URL
+	cfg.TranscribeModel = "iguana/kb-whisper"
+	cfg.ReconcileBaseModel = "iguana/pianissimo"
+	cfg.ReconcileModel = "iguana/gemma4-31b"
+	cfg.TranscribeEngineTimeout = 150 * time.Millisecond
+	wav := filepath.Join(t.TempDir(), "a.wav")
+	require.NoError(t, os.WriteFile(wav, []byte("RIFF....WAVEfmt "), 0o600))
+	job, cancel := context.WithTimeout(context.Background(), 20*time.Second)
+	defer cancel()
+
+	got, err := buildTranscriber(cfg).Transcribe(job, wav)
+
+	require.NoError(t, err, "a wedged base engine must not use up the job")
+	assert.Equal(t, "hej", got.Text)
+	require.NotNil(t, got.Reconciliation)
+	assert.Equal(t, "single_engine", string(got.Reconciliation.Status))
+	assert.Contains(t, got.Reconciliation.Reason, "timeout")
+}

cmd/tapir/uploads_worker_limit_test.go

@@ -0,0 +1,39 @@
+package main
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/uploads"
+)
+
+// The worker buildUploads returns enforces the configured maximum on what was really
+// converted (adversarial review F2). A stub ffmpeg writes a sparse 100-minute WAV however
+// short the file claimed to be; the gateway is never reached because the job fails first.
+
+func TestBuildUploads_TheWorkerRefusesAudioLongerThanTheConfiguredMaximum(t *testing.T) {
+	dir := t.TempDir()
+	stub := filepath.Join(dir, "ffmpeg")
+	require.NoError(t, os.WriteFile(stub, []byte("#!/bin/sh\nfor a; do last=\"$a\"; done\ntruncate -s 192000044 \"$last\"\n"), 0o700))
+	cfg := uploadCfg(t, true)
+	cfg.FFmpegBin = stub
+	cfg.UploadMaxDuration = 90 * time.Minute
+	cfg.TranscribeTimeout = time.Minute
+	st := &fakeUploadsStore{}
+	_, w, err := buildUploads(context.Background(), cfg, st, quietLog())
+	require.NoError(t, err)
+	src := filepath.Join(cfg.UploadTmpDir, "upload-x")
+	require.NoError(t, os.WriteFile(src, []byte("audio"), 0o600))
+
+	w.Process(context.Background(), uploads.Job{UserID: "u", UploadID: "id", Path: src, Duration: 10 * time.Second})
+
+	assert.Equal(t, []domain.UploadErrorKind{domain.UploadErrTooLong}, st.failedKinds,
+		"a file that claimed 10 s but converted to 100 minutes must fail, not be sent to an engine")
+}