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>

internal/adapters/audio/ffmpeg.go

@@ -0,0 +1,62 @@
+package audio
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"os/exec"
+	"strconv"
+	"strings"
+)
+
+// FFmpeg converts any audio container to the strict format whisper wants: 16 kHz
+// mono signed 16-bit PCM WAV (ADR-033). It ALWAYS transcodes, even when the input
+// already is that format — a Teams export often is, and the transcode is then
+// nearly free; assuming it would be the bug. This replaces whisper-server's
+// --convert, whose --tmp-dir is created at process start and rots under macOS
+// periodic cleanup.
+type FFmpeg struct {
+	// Bin is the ffmpeg executable ("ffmpeg" resolves via PATH).
+	Bin string
+	// MaxSeconds caps the converted length when positive. The duration limit is a claim
+	// in the file's header and can be lied about (a 2 h file claiming 10 s converted to a
+	// 120-minute WAV); this bounds what a lie can cost. A minute of slack is added so a
+	// file at the limit is never cut, and the worker measures the result.
+	MaxSeconds int
+}
+
+// lengthSlackSeconds is added to MaxSeconds so ffmpeg's rounding never clips a file that
+// is really at the limit.
+const lengthSlackSeconds = 60
+
+// ToWAV writes src, converted, to dst. Video streams are dropped (-vn): by now the
+// file has already been probed as audio-only, but a cover-art stream must not end
+// up in the WAV.
+func (f *FFmpeg) ToWAV(ctx context.Context, src, dst string) error {
+	bin := f.Bin
+	if bin == "" {
+		bin = "ffmpeg"
+	}
+	// Input restrictions go BEFORE -i: placed after it they would apply to the output.
+	args := append([]string{"-y", "-v", "error"}, inputRestrictions...)
+	args = append(args, "-i", src, "-vn", "-ac", "1", "-ar", "16000", "-c:a", "pcm_s16le")
+	if f.MaxSeconds > 0 {
+		// After -i: an OUTPUT cap. Before it, it would limit what is read, not written.
+		args = append(args, "-t", strconv.Itoa(f.MaxSeconds+lengthSlackSeconds))
+	}
+	args = append(args, "-f", "wav", dst)
+	cmd := exec.CommandContext(ctx, bin, args...)
+	cmd.WaitDelay = killGrace
+	if out, err := cmd.CombinedOutput(); err != nil {
+		if ctxErr := ctx.Err(); ctxErr != nil {
+			return fmt.Errorf("audio: ffmpeg: %w", ctxErr)
+		}
+		return fmt.Errorf("audio: ffmpeg: %w: %s", err, strings.TrimSpace(string(out)))
+	}
+	// A zero exit that wrote nothing is still a failure: never hand whisper a
+	// missing file and let the error surface three layers later.
+	if info, err := os.Stat(dst); err != nil || info.Size() == 0 {
+		return fmt.Errorf("audio: ffmpeg wrote no output at %s", dst)
+	}
+	return nil
+}

internal/adapters/audio/ffmpeg_test.go

@@ -0,0 +1,111 @@
+package audio_test
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/audio"
+)
+
+func TestFFmpegToWAV_AsksForSixteenKilohertzMonoPCMAlways(t *testing.T) {
+	// Always transcode, even when the input already is 16 kHz mono: a Teams export
+	// often is, and the point is to not ASSUME (ADR-033). Records the argv.
+	rec := filepath.Join(t.TempDir(), "argv")
+	bin := stub(t, `printf '%s\n' "$@" > `+rec+`
+for a; do last="$a"; done
+printf data > "$last"`)
+
+	err := (&audio.FFmpeg{Bin: bin}).ToWAV(context.Background(), "/spool/in", filepath.Join(t.TempDir(), "out.wav"))
+
+	require.NoError(t, err)
+	raw, err := os.ReadFile(rec)
+	require.NoError(t, err)
+	argv := strings.Split(strings.TrimSpace(string(raw)), "\n")
+	joined := " " + strings.Join(argv, " ") + " "
+	for _, want := range []string{" -ac 1 ", " -ar 16000 ", " -c:a pcm_s16le ", " -vn ", " -i /spool/in "} {
+		assert.Contains(t, joined, want)
+	}
+	assert.True(t, strings.HasSuffix(strings.TrimSpace(joined), "out.wav"), "the output path is the last argument")
+}
+
+func TestFFmpegToWAV_NonzeroExitIsAnErrorCarryingStderr(t *testing.T) {
+	bin := stub(t, `echo "moov atom not found" >&2; exit 1`)
+
+	err := (&audio.FFmpeg{Bin: bin}).ToWAV(context.Background(), "/in", filepath.Join(t.TempDir(), "o.wav"))
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "moov atom not found")
+}
+
+func TestFFmpegToWAV_ZeroExitWithNoOutputIsAnError(t *testing.T) {
+	bin := stub(t, `exit 0`)
+
+	err := (&audio.FFmpeg{Bin: bin}).ToWAV(context.Background(), "/in", filepath.Join(t.TempDir(), "never-written.wav"))
+
+	require.Error(t, err)
+}
+
+func TestFFmpegToWAV_MissingBinaryIsAnError(t *testing.T) {
+	err := (&audio.FFmpeg{Bin: filepath.Join(t.TempDir(), "no-ffmpeg")}).ToWAV(context.Background(), "/in", "/out.wav")
+
+	require.Error(t, err)
+	assert.False(t, errors.Is(err, context.Canceled))
+}
+
+func TestFFmpegToWAV_RestrictsWhatTheFileMayMakeFFmpegOpen(t *testing.T) {
+	rec := filepath.Join(t.TempDir(), "argv")
+	bin := stub(t, `printf '%s\n' "$@" > `+rec+`
+for a; do last="$a"; done
+printf data > "$last"`)
+
+	err := (&audio.FFmpeg{Bin: bin}).ToWAV(context.Background(), "/spool/in", filepath.Join(t.TempDir(), "o.wav"))
+
+	require.NoError(t, err)
+	args := argvOf(t, rec)
+	pw, fw, in := indexOf(args, "-protocol_whitelist"), indexOf(args, "-format_whitelist"), indexOf(args, "-i")
+	require.GreaterOrEqual(t, pw, 0, "argv: %v", args)
+	require.GreaterOrEqual(t, fw, 0, "argv: %v", args)
+	assert.Equal(t, "file,pipe", args[pw+1])
+	assert.Equal(t, audio.AllowedDemuxers, args[fw+1])
+	assert.Less(t, pw, in, "input options must come BEFORE -i or ffmpeg applies them to the output")
+	assert.Less(t, fw, in)
+}
+
+func TestFFmpegToWAV_PassesAnOutputLengthCapAfterTheInputWhenAMaximumIsSet(t *testing.T) {
+	// The duration limit is a CLAIM in the file's header; a file that lies about it can
+	// convert to hours of audio (adversarial review F2: a 2 h file claiming 10 s became a
+	// 120-minute WAV). The cap is an OUTPUT option, after -i, with a minute of slack.
+	rec := filepath.Join(t.TempDir(), "argv")
+	bin := stub(t, `printf '%s\n' "$@" > `+rec+`
+for a; do last="$a"; done
+printf data > "$last"`)
+
+	err := (&audio.FFmpeg{Bin: bin, MaxSeconds: 5400}).ToWAV(context.Background(), "/spool/in", filepath.Join(t.TempDir(), "out.wav"))
+
+	require.NoError(t, err)
+	raw, err := os.ReadFile(rec)
+	require.NoError(t, err)
+	joined := " " + strings.Join(strings.Split(strings.TrimSpace(string(raw)), "\n"), " ") + " "
+	assert.Contains(t, joined, " -t 5460 ", "limit plus a minute of slack")
+	assert.Less(t, strings.Index(joined, " -i /spool/in "), strings.Index(joined, " -t 5460 "), "-t after -i is an output cap; before it, an input one")
+}
+
+func TestFFmpegToWAV_PassesNoLengthCapWhenNoMaximumIsSet(t *testing.T) {
+	rec := filepath.Join(t.TempDir(), "argv")
+	bin := stub(t, `printf '%s\n' "$@" > `+rec+`
+for a; do last="$a"; done
+printf data > "$last"`)
+
+	require.NoError(t, (&audio.FFmpeg{Bin: bin}).ToWAV(context.Background(), "/spool/in", filepath.Join(t.TempDir(), "out.wav")))
+
+	raw, err := os.ReadFile(rec)
+	require.NoError(t, err)
+	assert.NotContains(t, " "+strings.ReplaceAll(string(raw), "\n", " ")+" ", " -t ")
+}

internal/adapters/audio/ffprobe.go

@@ -0,0 +1,109 @@
+// Package audio holds the adapters that inspect and convert user-supplied audio
+// with ffmpeg's tools (ADR-033). They sit behind small ports so the use case is
+// tested with fakes and only this package knows a subprocess is involved.
+package audio
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os/exec"
+	"strconv"
+	"time"
+
+	"git.d-ma.be/mathias/tapir/internal/uploads"
+)
+
+// FFprobe implements uploads.Prober by running ffprobe on the spooled file.
+type FFprobe struct {
+	// Bin is the ffprobe executable ("ffprobe" resolves via PATH).
+	Bin string
+}
+
+// AllowedDemuxers are the input formats an upload may be. The input is
+// attacker-supplied, and a file's CONTENT can tell ffmpeg to open other files or
+// URLs (HLS and concat playlists, image2 sequences, network demuxers). Recent
+// ffmpeg refuses the ones tried (7.0.2: HLS without a .m3u8 name is not sniffed,
+// concat rejects unsafe paths), but that is version luck; this list is not. It
+// holds the containers real recordings use and none of the playlist or network
+// formats. "mov" covers .mp4/.m4a/.3gp; "matroska" covers .webm.
+const AllowedDemuxers = "wav,mov,mp4,m4a,3gp,mp3,ogg,matroska,webm,flac,aac,opus,aiff,amr,asf"
+
+// inputRestrictions are ffmpeg/ffprobe INPUT options and must precede the input.
+// Local files and pipes only: no network, no other protocols.
+var inputRestrictions = []string{"-protocol_whitelist", "file,pipe", "-format_whitelist", AllowedDemuxers}
+
+// killGrace bounds how long Probe waits for a cancelled ffprobe's pipes to close.
+const killGrace = 2 * time.Second
+
+// Probe runs ffprobe and classifies the outcome. A non-zero exit means ffprobe
+// looked at the file and could not read it: that is the FILE's fault and returns
+// uploads.ErrUnreadable. A missing binary, a cancelled context or unparsable
+// output is OURS and is returned as an ordinary error, so it is never reported to
+// the user as their file being bad.
+func (f *FFprobe) Probe(ctx context.Context, path string) (uploads.Probe, error) {
+	bin := f.Bin
+	if bin == "" {
+		bin = "ffprobe"
+	}
+	args := append([]string{"-v", "error"}, inputRestrictions...)
+	args = append(args, "-print_format", "json", "-show_format", "-show_streams", path)
+	cmd := exec.CommandContext(ctx, bin, args...)
+	cmd.WaitDelay = killGrace
+	out, err := cmd.Output()
+	if err != nil {
+		if ctxErr := ctx.Err(); ctxErr != nil {
+			return uploads.Probe{}, fmt.Errorf("audio: ffprobe: %w", ctxErr)
+		}
+		var exit *exec.ExitError
+		if errors.As(err, &exit) {
+			return uploads.Probe{}, fmt.Errorf("%w: ffprobe: %s", uploads.ErrUnreadable, exit.Stderr)
+		}
+		return uploads.Probe{}, fmt.Errorf("audio: run ffprobe: %w", err)
+	}
+	p, err := ParseProbe(out)
+	if err != nil {
+		return uploads.Probe{}, fmt.Errorf("audio: ffprobe output: %w", err)
+	}
+	return p, nil
+}
+
+// probeJSON is the subset of `ffprobe -print_format json` output that matters.
+type probeJSON struct {
+	Streams []struct {
+		CodecType   string `json:"codec_type"`
+		Disposition struct {
+			AttachedPic int `json:"attached_pic"`
+		} `json:"disposition"`
+	} `json:"streams"`
+	Format struct {
+		Duration string `json:"duration"`
+	} `json:"format"`
+}
+
+// ParseProbe turns ffprobe's JSON into a uploads.Probe. A video stream that is
+// only attached cover art (an mp3/m4a with an album picture) is not video. A
+// missing or "N/A" duration yields zero, which uploads.Validate rejects as
+// unreadable rather than accepting an empty recording.
+func ParseProbe(data []byte) (uploads.Probe, error) {
+	var j probeJSON
+	if err := json.Unmarshal(data, &j); err != nil {
+		return uploads.Probe{}, fmt.Errorf("decode: %w", err)
+	}
+	var p uploads.Probe
+	for _, s := range j.Streams {
+		switch s.CodecType {
+		case "audio":
+			p.HasAudio = true
+		case "video":
+			if s.Disposition.AttachedPic == 0 {
+				p.HasVideo = true
+			}
+		}
+	}
+	if secs, err := strconv.ParseFloat(j.Format.Duration, 64); err == nil && secs > 0 {
+		p.Duration = time.Duration(secs * float64(time.Second))
+	}
+	return p, nil
+}

internal/adapters/audio/ffprobe_test.go

@@ -0,0 +1,186 @@
+package audio_test
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/audio"
+	"git.d-ma.be/mathias/tapir/internal/uploads"
+)
+
+func TestParseProbe(t *testing.T) {
+	tests := []struct {
+		name string
+		json string
+		want uploads.Probe
+	}{
+		{
+			name: "audio-only m4a",
+			json: `{"streams":[{"codec_type":"audio","disposition":{"attached_pic":0}}],"format":{"duration":"2701.5"}}`,
+			want: uploads.Probe{Duration: 2701*time.Second + 500*time.Millisecond, HasAudio: true},
+		},
+		{
+			name: "mp4 with a real video stream",
+			json: `{"streams":[{"codec_type":"video","disposition":{"attached_pic":0}},{"codec_type":"audio"}],"format":{"duration":"240.0"}}`,
+			want: uploads.Probe{Duration: 240 * time.Second, HasAudio: true, HasVideo: true},
+		},
+		{
+			name: "mp3 whose only video stream is attached cover art is NOT video",
+			json: `{"streams":[{"codec_type":"audio"},{"codec_type":"video","disposition":{"attached_pic":1}}],"format":{"duration":"60"}}`,
+			want: uploads.Probe{Duration: 60 * time.Second, HasAudio: true},
+		},
+		{
+			name: "no streams at all",
+			json: `{"streams":[],"format":{"duration":"10"}}`,
+			want: uploads.Probe{Duration: 10 * time.Second},
+		},
+		{
+			name: "duration absent means zero, which Validate rejects as unreadable",
+			json: `{"streams":[{"codec_type":"audio"}],"format":{}}`,
+			want: uploads.Probe{HasAudio: true},
+		},
+		{
+			name: "duration reported as N/A means zero",
+			json: `{"streams":[{"codec_type":"audio"}],"format":{"duration":"N/A"}}`,
+			want: uploads.Probe{HasAudio: true},
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got, err := audio.ParseProbe([]byte(tt.json))
+
+			require.NoError(t, err)
+			assert.Equal(t, tt.want, got)
+		})
+	}
+}
+
+func TestParseProbe_GarbageIsAnError(t *testing.T) {
+	_, err := audio.ParseProbe([]byte("not json"))
+	assert.Error(t, err)
+}
+
+// stub writes an executable shell script standing in for ffprobe. This is a fake
+// at the process boundary: the real binary is not on every host (it is not on
+// koala), and what is under test is how Probe drives and interprets a process.
+func stub(t *testing.T, body string) string {
+	t.Helper()
+	p := filepath.Join(t.TempDir(), "ffprobe")
+	require.NoError(t, os.WriteFile(p, []byte("#!/bin/sh\n"+body+"\n"), 0o755))
+	return p
+}
+
+func TestFFprobeProbe_ParsesWhatTheProcessPrints(t *testing.T) {
+	bin := stub(t, `echo '{"streams":[{"codec_type":"audio"}],"format":{"duration":"90"}}'`)
+
+	got, err := (&audio.FFprobe{Bin: bin}).Probe(context.Background(), "/ignored")
+
+	require.NoError(t, err)
+	assert.Equal(t, uploads.Probe{Duration: 90 * time.Second, HasAudio: true}, got)
+}
+
+func TestFFprobeProbe_PassesTheFilePathAsTheLastArgument(t *testing.T) {
+	out := filepath.Join(t.TempDir(), "args")
+	bin := stub(t, `for a; do last="$a"; done; printf '%s' "$last" > `+out+`
+echo '{"streams":[{"codec_type":"audio"}],"format":{"duration":"1"}}'`)
+
+	_, err := (&audio.FFprobe{Bin: bin}).Probe(context.Background(), "/spool/upload-123")
+
+	require.NoError(t, err)
+	got, err := os.ReadFile(out)
+	require.NoError(t, err)
+	assert.Equal(t, "/spool/upload-123", string(got))
+}
+
+func TestFFprobeProbe_NonzeroExitMeansTheFileIsUnreadable(t *testing.T) {
+	bin := stub(t, `echo "Invalid data found when processing input" >&2; exit 1`)
+
+	_, err := (&audio.FFprobe{Bin: bin}).Probe(context.Background(), "/x")
+
+	assert.ErrorIs(t, err, uploads.ErrUnreadable)
+}
+
+func TestFFprobeProbe_MissingBinaryIsOurFaultNotAnUnreadableFile(t *testing.T) {
+	_, err := (&audio.FFprobe{Bin: filepath.Join(t.TempDir(), "no-such-ffprobe")}).Probe(context.Background(), "/x")
+
+	require.Error(t, err)
+	assert.False(t, errors.Is(err, uploads.ErrUnreadable),
+		"a missing ffprobe must not be reported to the user as their file being bad")
+}
+
+func TestFFprobeProbe_StopsWhenTheContextIsCancelled(t *testing.T) {
+	bin := stub(t, `sleep 30`)
+	ctx, cancel := context.WithTimeout(context.Background(), 100*time.Millisecond)
+	defer cancel()
+
+	start := time.Now()
+	_, err := (&audio.FFprobe{Bin: bin}).Probe(ctx, "/x")
+
+	require.Error(t, err)
+	assert.Less(t, time.Since(start), 5*time.Second)
+	assert.False(t, errors.Is(err, uploads.ErrUnreadable), "a timeout is not a verdict on the file")
+}
+
+// argvStub records the arguments it was called with into a file and then behaves as
+// a successful ffprobe, so a test can assert exactly what the process was asked.
+func argvStub(t *testing.T) (bin, argvFile string) {
+	t.Helper()
+	argvFile = filepath.Join(t.TempDir(), "argv")
+	bin = stub(t, `printf '%s\n' "$@" > `+argvFile+`
+echo '{"streams":[{"codec_type":"audio"}],"format":{"duration":"1"}}'`)
+	return bin, argvFile
+}
+
+func argvOf(t *testing.T, file string) []string {
+	t.Helper()
+	raw, err := os.ReadFile(file)
+	require.NoError(t, err)
+	return strings.Split(strings.TrimSpace(string(raw)), "\n")
+}
+
+func indexOf(args []string, want string) int {
+	for i, a := range args {
+		if a == want {
+			return i
+		}
+	}
+	return -1
+}
+
+func TestFFprobeProbe_RestrictsWhatTheFileMayMakeFFmpegOpen(t *testing.T) {
+	// The input is attacker-supplied. ffmpeg can be told by a file's CONTENT to open
+	// other files or URLs (HLS and concat playlists); recent versions refuse by
+	// default, but that is version luck. Pin it: local files and pipes only, and
+	// only the demuxers real audio recordings use.
+	bin, argvFile := argvStub(t)
+
+	_, err := (&audio.FFprobe{Bin: bin}).Probe(context.Background(), "/spool/upload-1")
+
+	require.NoError(t, err)
+	args := argvOf(t, argvFile)
+	pw := indexOf(args, "-protocol_whitelist")
+	require.GreaterOrEqual(t, pw, 0, "argv: %v", args)
+	assert.Equal(t, "file,pipe", args[pw+1])
+	fw := indexOf(args, "-format_whitelist")
+	require.GreaterOrEqual(t, fw, 0, "argv: %v", args)
+	assert.Equal(t, audio.AllowedDemuxers, args[fw+1])
+	assert.Equal(t, "/spool/upload-1", args[len(args)-1], "the path stays the last argument")
+}
+
+func TestAllowedDemuxers_ExcludeThePlaylistFormats(t *testing.T) {
+	for _, bad := range []string{"concat", "hls", "applehttp", "dash", "rtsp", "sdp", "image2"} {
+		assert.NotContains(t, strings.Split(audio.AllowedDemuxers, ","), bad,
+			"%s can make ffmpeg open other files or URLs named inside the upload", bad)
+	}
+	for _, good := range []string{"wav", "mov", "mp4", "m4a", "mp3", "ogg", "matroska", "flac"} {
+		assert.Contains(t, strings.Split(audio.AllowedDemuxers, ","), good)
+	}
+}

internal/adapters/audio/hostile_live_test.go

@@ -0,0 +1,32 @@
+package audio_test
+
+import (
+	"context"
+	"os"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/audio"
+	"git.d-ma.be/mathias/tapir/internal/uploads"
+)
+
+// TestFFprobeLive_HostilePlaylistsAreRefusedAndRealAudioIsNot runs the REAL ffprobe
+// (opt-in: TAPIR_LIVE_FFPROBE) on crafted concat/HLS files and on a real m4a.
+func TestFFprobeLive_HostilePlaylistsAreRefusedAndRealAudioIsNot(t *testing.T) {
+	bin := os.Getenv("TAPIR_LIVE_FFPROBE")
+	if bin == "" {
+		t.Skip("set TAPIR_LIVE_FFPROBE (and TAPIR_LIVE_HOSTILE_DIR, TAPIR_LIVE_M4A) to run against the real ffprobe")
+	}
+	p := &audio.FFprobe{Bin: bin}
+	for _, name := range []string{"up-concat", "up-hls"} {
+		_, err := p.Probe(context.Background(), os.Getenv("TAPIR_LIVE_HOSTILE_DIR")+"/"+name)
+		assert.ErrorIs(t, err, uploads.ErrUnreadable, name)
+	}
+	got, err := p.Probe(context.Background(), os.Getenv("TAPIR_LIVE_M4A"))
+	require.NoError(t, err)
+	assert.True(t, got.HasAudio)
+	assert.False(t, got.HasVideo)
+	assert.InDelta(t, 82.5, got.Duration.Seconds(), 1)
+}