fix: re-read each message for a live file reference before downloading it

The walk collects every item's file reference up front, and Telegram expires
those. The lifetime is undocumented; a run was seen to outlast an hour and fail
before two, which is less than a large chat spends downloading. So every fetch
past that point returned FILE_REFERENCE_EXPIRED, five in a row tripped the
breaker, and the run abandoned the rest of its todo list.

Telegram's contract is to cache a reference together with the source it came
from and re-read that source when it expires, so each message is re-read for a
live reference immediately before its own download. That is one round trip per
transfer and needs no guess at how long a reference lives. It runs after the
staging acquire rather than before: that acquire is the run's brake and blocks
for as long as the destination is slow, which on a stalled remote is long enough
to expire a fresh reference all over again.

A re-read that describes a different file — different name or different length —
is refused, because archiving it under the name this run recorded would store
the wrong bytes. That case, a deleted message, and one that no longer holds a
file are all skipped: nothing can be archived for them, and one must not strand
the rest of the chat.

Any other re-read failure is recorded as a failed download instead, because that
is what it is. The connection that serves the re-read is the one that serves the
file, so these arrive exactly when transfers are failing too. An item that never
reaches a worker never reaches OnAdd or OnDone, so on the skip path it would
move no counter, feed no streak and produce no outcome to retry — a source
outage would walk the whole list one dead round trip at a time and still exit as
merely incomplete.

tutil.GetSingleMessage is not reused for the re-read: it reports a deleted
message three different ways, two of them wrapping a nil error, and only one is
distinguishable — so deletions would be misclassified as transient and spend the
breaker's streak.

README.md

@@ -187,6 +187,28 @@ verifier, which matched on name and non-zero size — on the archive this was
 built for it found six objects that had been counted complete for months, one
 of them 221 MiB standing in for a 2 GiB video. Re-running repairs them.
 
+**File references.** Telegram hands out a short-lived token with every file
+location, and a run over a large chat outlives the ones its walk collected — the
+lifetime is undocumented, but an observed run stopped just under two hours in
+with `FILE_REFERENCE_EXPIRED` on every remaining file. So each message is re-read
+for a live token immediately before its own download, which costs one round trip
+per transfer and needs no guess at how long a token lasts. Retry passes go
+through the same path, so a token that dies during a single very large transfer
+is replaced rather than replayed.
+
+The re-read has to describe the same file — same name, same length — or it is not
+the file this run recorded, and archiving it under that name would store the
+wrong bytes. A message that fails that check, or that has been deleted, or that
+no longer holds a file at all, is skipped and reported: nothing can be archived
+for it, one of them must not strand the rest of the chat, and the closing
+`verify` still lists it as outstanding.
+
+A re-read that fails for any other reason is counted as a failed download, not a
+skip, because that is what it is — the connection that serves the re-read is the
+one that serves the file. So it is retried with the rest, it shows up in the
+progress report and the closing failure list, and enough of them in a row trip
+the same breaker below.
+
 **Failures.** A failed download is retried inside the run: up to three passes
 over whatever is still missing, spaced a minute and then two apart. The failure
 this exists for is the transient one — a connection that dies takes every

cmd/tgexport/sync.go

@@ -176,6 +176,10 @@ func syncCmd(ctx context.Context, args []string) error {
 			Confirm: *confirm,
 			Takeout: *takeout,
 			Events:  rep,
+			// The walk above collected every item's file reference up front, and
+			// a run this size spends longer downloading than one of those stays
+			// valid, so each is re-read immediately before its own download.
+			Refresh: tgsource.Refresher(api, peer),
 			// Re-checked during the run, not only before it: an archive of this
 			// size runs for hours, and the destination can fill in the middle.
 			FreeBytes: func(ctx context.Context) (int64, bool) {

internal/pipeline/download.go

@@ -29,6 +29,11 @@ type DownloadOptions struct {
 	// concurrently.
 	Events Events
 
+	// Refresh, when set, re-mints an item's file reference just before its
+	// download starts. Unset means the reference the walk minted is used as-is,
+	// which only holds up for a run shorter than a reference's lifetime.
+	Refresh tgsource.Refresh
+
 	// acquire reserves staging space before a download starts, blocking until
 	// there is room. Unset means no bound. release hands a reservation back for
 	// an item that never reaches a download.
@@ -87,6 +92,7 @@ func Download(ctx context.Context, seq iter.Seq2[tgsource.Item, error], o Downlo
 	ctx = captureCauses(ctx)
 
 	it := newElemIter(seq, o.Staging, o.Takeout)
+	it.refresh = o.Refresh
 	it.acquire = o.acquire
 	it.release = o.release
 	if o.stop != nil {
@@ -113,6 +119,9 @@ func Download(ctx context.Context, seq iter.Seq2[tgsource.Item, error], o Downlo
 	prog.stats = o.seed
 	prog.maxStreak = o.maxFailures
 	prog.onTrip = func() { it.stopped.Store(true) }
+	// Wired after prog exists, so an item the iterator cannot even hand over is
+	// counted, reported and retried like any other failed transfer.
+	it.refused = prog.refused
 
 	err := downloader.New(downloader.Options{
 		Pool:     o.Pool,

internal/pipeline/download_test.go

@@ -3,6 +3,7 @@ package pipeline
 import (
 	"context"
 	"errors"
+	"fmt"
 	"os"
 	"path/filepath"
 	"strings"
@@ -533,3 +534,260 @@ func TestProgressWithoutABreakerNeverTrips(t *testing.T) {
 		t.Error("brokeCircuit() = true with no limit configured")
 	}
 }
+
+// The reference the walk minted is dead by the time a long run reaches the end
+// of its list, so the location the downloader is handed must be the one refresh
+// just produced. Without that substitution every fetch past the reference's
+// lifetime returns FILE_REFERENCE_EXPIRED and the breaker abandons the rest.
+func TestElemIterDownloadsTheRefreshedLocation(t *testing.T) {
+	staging := t.TempDir()
+	stale := testItem(t, 42, "clip.mp4", 10)
+
+	var asked []int
+	it := newElemIter(func(yield func(tgsource.Item, error) bool) { yield(stale, nil) },
+		staging, false)
+	it.refresh = func(_ context.Context, in tgsource.Item) (tgsource.Item, error) {
+		asked = append(asked, in.MessageID)
+		fresh := *in.Media
+		fresh.InputFileLoc = &tg.InputDocumentFileLocation{
+			ID:            in.Media.InputFileLoc.(*tg.InputDocumentFileLocation).ID,
+			FileReference: []byte("fresh"),
+		}
+		in.Media = &fresh
+		return in, nil
+	}
+	defer func() { _ = it.Close() }()
+
+	if !it.Next(t.Context()) {
+		t.Fatalf("Next produced nothing: %v", it.failure)
+	}
+	if len(asked) != 1 || asked[0] != 42 {
+		t.Fatalf("refresh calls = %v, want one for message 42", asked)
+	}
+
+	loc, ok := it.Value().File().Location().(*tg.InputDocumentFileLocation)
+	if !ok {
+		t.Fatalf("location is %T, want *tg.InputDocumentFileLocation", it.Value().File().Location())
+	}
+	if string(loc.FileReference) != "fresh" {
+		t.Errorf("downloader was handed file reference %q, want the refreshed one", loc.FileReference)
+	}
+}
+
+// A message that cannot be re-read — deleted, or now holding a different file —
+// must not strand every message behind it, which is what ending the walk would
+// mean. It is reported instead, so nothing vanishes silently.
+func TestElemIterSkipsItemsRefreshCannotFindAndKeepsGoing(t *testing.T) {
+	staging := t.TempDir()
+	gone := testItem(t, 7, "deleted.mp4", 10)
+	good := testItem(t, 8, "fine.mp4", 10)
+
+	seq := func(yield func(tgsource.Item, error) bool) {
+		if !yield(gone, nil) {
+			return
+		}
+		yield(good, nil)
+	}
+	it := newElemIter(seq, staging, false)
+	it.refresh = func(_ context.Context, in tgsource.Item) (tgsource.Item, error) {
+		if in.MessageID == 7 {
+			return tgsource.Item{}, fmt.Errorf("%w: message 7 was deleted", tgsource.ErrGone)
+		}
+		return in, nil
+	}
+	defer func() { _ = it.Close() }()
+
+	if !it.Next(t.Context()) {
+		t.Fatal("an unrefreshable item ended the walk; the item after it was never reached")
+	}
+	if got := it.current.item.MessageID; got != 8 {
+		t.Fatalf("Next yielded message %d, want the item after the missing one", got)
+	}
+	if it.failure != nil {
+		t.Errorf("a skipped item must not fail the run, got: %v", it.failure)
+	}
+	if len(it.skipped) != 1 {
+		t.Fatalf("skipped = %d, want 1", len(it.skipped))
+	}
+	if !strings.Contains(it.skipped[0].Error(), "message 7") {
+		t.Errorf("the skip should name the message, got: %v", it.skipped[0])
+	}
+}
+
+// A cancelled run ends the walk rather than skipping its way through the rest of
+// the list, one refused refresh at a time.
+func TestElemIterStopsWhenRefreshIsCancelled(t *testing.T) {
+	staging := t.TempDir()
+	seq := seqOf(testItems(3, 10))
+
+	ctx, cancel := context.WithCancel(t.Context())
+	it := newElemIter(seq, staging, false)
+	it.refused = func(tgsource.Item, error) {
+		t.Error("cancellation was recorded as an item failure, which the breaker would count")
+	}
+	it.refresh = func(rctx context.Context, in tgsource.Item) (tgsource.Item, error) {
+		cancel()
+		return tgsource.Item{}, fmt.Errorf("re-read message %d: %w", in.MessageID, rctx.Err())
+	}
+	defer func() { _ = it.Close() }()
+
+	if it.Next(ctx) {
+		t.Fatal("Next produced an item after cancellation")
+	}
+	if !errors.Is(it.failure, context.Canceled) {
+		t.Errorf("failure = %v, want context.Canceled", it.failure)
+	}
+	if len(it.skipped) != 0 {
+		t.Errorf("cancellation was recorded as %d skipped item(s), not a failure", len(it.skipped))
+	}
+}
+
+// A refresh that failed for any reason other than the message being gone is the
+// same event as a transfer dying — Telegram is not serving this file — and has
+// to be counted as one. An item that never reaches a worker never reaches OnAdd
+// or OnDone, so on the skip path a source outage would move no counter, trip no
+// breaker and produce no Outcome for Run to retry, and would walk the whole todo
+// list one dead round trip at a time.
+func TestElemIterRecordsRefreshFailuresAsFailuresNotSkips(t *testing.T) {
+	staging := t.TempDir()
+	items := []tgsource.Item{
+		testItem(t, 1, "a.mp4", 10),
+		testItem(t, 2, "b.mp4", 20),
+	}
+
+	it := newElemIter(seqOf(items), staging, false)
+	defer func() { _ = it.Close() }()
+
+	prog := newProgress(func(*elem, error) error { return nil }, nil)
+	prog.maxStreak = 2
+	var tripped bool
+	prog.onTrip = func() { tripped = true; it.stopped.Store(true) }
+	it.refused = prog.refused
+	it.refresh = func(context.Context, tgsource.Item) (tgsource.Item, error) {
+		return tgsource.Item{}, errors.New("dial telegram: connection refused")
+	}
+
+	for it.Next(t.Context()) {
+		t.Fatal("Next handed over an item whose reference could not be refreshed")
+	}
+
+	if it.failure != nil {
+		t.Errorf("failure = %v, want none — a refused item ends the pass, not the program", it.failure)
+	}
+	if len(it.skipped) != 0 {
+		t.Errorf("skipped = %d, want 0 — a transient re-read failure is not a skip: %v",
+			len(it.skipped), it.skipped)
+	}
+
+	outcomes, stats := prog.results()
+	if len(outcomes) != 2 {
+		t.Fatalf("outcomes = %d, want 2 — Run builds its retry list from these", len(outcomes))
+	}
+	for _, oc := range outcomes {
+		if oc.Err == nil {
+			t.Errorf("message %d recorded without an error", oc.Item.MessageID)
+		}
+	}
+	if got := failedItems(outcomes); len(got) != 2 {
+		t.Errorf("failedItems = %d, want 2 — these have to reach the next pass", len(got))
+	}
+	if stats.Started != 2 || stats.Failed != 2 || stats.Done != 0 {
+		t.Errorf("Stats = %+v, want 2 started, 2 failed, 0 done", stats)
+	}
+	if stats.BytesTotal != 30 {
+		t.Errorf("BytesTotal = %d, want 30 — the report's target must include them", stats.BytesTotal)
+	}
+	if !tripped {
+		t.Error("two consecutive refused items did not trip the breaker")
+	}
+}
+
+// A message that is gone is the one refresh failure that is genuinely permanent,
+// so it stays on the skip path and must not spend the breaker's streak.
+func TestElemIterDoesNotCountAMissingMessageAsAFailure(t *testing.T) {
+	staging := t.TempDir()
+	it := newElemIter(seqOf([]tgsource.Item{testItem(t, 1, "a.mp4", 10)}), staging, false)
+	defer func() { _ = it.Close() }()
+
+	prog := newProgress(func(*elem, error) error { return nil }, nil)
+	prog.maxStreak = 1
+	prog.onTrip = func() { t.Error("a gone message tripped the source breaker") }
+	it.refused = func(tgsource.Item, error) { t.Error("a gone message was recorded as a failure") }
+	it.refresh = func(_ context.Context, in tgsource.Item) (tgsource.Item, error) {
+		return tgsource.Item{}, fmt.Errorf("%w: message %d was deleted", tgsource.ErrGone, in.MessageID)
+	}
+
+	for it.Next(t.Context()) {
+	}
+	if len(it.skipped) != 1 {
+		t.Fatalf("skipped = %d, want 1", len(it.skipped))
+	}
+	if outcomes, _ := prog.results(); len(outcomes) != 0 {
+		t.Errorf("outcomes = %d, want 0 — a gone message is not worth retrying", len(outcomes))
+	}
+}
+
+// The refresh has to come after the reservation, not before it. That acquire is
+// the run's brake: it blocks for as long as the destination is slow, and a
+// reference minted before it would spend that whole wait ageing — which on a
+// stalled remote is long enough to expire it again, reproducing the failure the
+// refresh exists to prevent.
+func TestElemIterRefreshesAfterReservingSpace(t *testing.T) {
+	staging := t.TempDir()
+	it := newElemIter(seqOf([]tgsource.Item{testItem(t, 1, "a.mp4", 4096)}), staging, false)
+	defer func() { _ = it.Close() }()
+
+	var acquired int64
+	it.acquire = func(_ context.Context, n int64) error { acquired += n; return nil }
+	it.release = func(n int64) { acquired -= n }
+
+	var heldAtRefresh int64
+	it.refresh = func(_ context.Context, in tgsource.Item) (tgsource.Item, error) {
+		heldAtRefresh = acquired
+		return in, nil
+	}
+
+	if !it.Next(t.Context()) {
+		t.Fatalf("Next produced nothing: %v", it.failure)
+	}
+	if heldAtRefresh != 4096 {
+		t.Errorf("staging held %d bytes when the reference was refreshed, want 4096 — "+
+			"the refresh ran before the acquire it is supposed to follow", heldAtRefresh)
+	}
+}
+
+// Whichever way a refresh fails, the bytes it reserved go back. Without that the
+// budget shrinks by that much for the rest of the run.
+func TestElemIterReturnsReservationWhenRefreshFails(t *testing.T) {
+	staging := t.TempDir()
+
+	for _, tc := range []struct {
+		name string
+		err  error
+	}{
+		{"gone", fmt.Errorf("%w: deleted", tgsource.ErrGone)},
+		{"transient", errors.New("dial telegram: connection refused")},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			it := newElemIter(seqOf([]tgsource.Item{testItem(t, 1, "a.mp4", 4096)}), staging, false)
+			defer func() { _ = it.Close() }()
+
+			var acquired, released int64
+			it.acquire = func(_ context.Context, n int64) error { acquired += n; return nil }
+			it.release = func(n int64) { released += n }
+			it.refresh = func(context.Context, tgsource.Item) (tgsource.Item, error) {
+				return tgsource.Item{}, tc.err
+			}
+
+			for it.Next(t.Context()) {
+			}
+			if acquired == 0 {
+				t.Fatal("nothing was reserved, so the test proves nothing")
+			}
+			if acquired != released {
+				t.Errorf("acquired %d bytes but released %d — the reservation leaked",
+					acquired, released)
+			}
+		})
+	}
+}

internal/pipeline/elem.go

@@ -3,6 +3,7 @@ package pipeline
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"io"
 	"iter"
@@ -76,6 +77,14 @@ type elemIter struct {
 	staging  string
 	takeout  bool
 
+	// refresh re-mints an item's file reference just before its download; see
+	// reminted. Unset leaves the walk's own reference in place.
+	refresh tgsource.Refresh
+	// refused records an item that failed before any transfer could start, so it
+	// counts as a failure rather than disappearing. Unset means such an item is
+	// only skipped.
+	refused func(tgsource.Item, error)
+
 	// acquire reserves staging space for the next item. Blocking here is what
 	// makes backpressure work: core's Download calls Next from its dispatch
 	// loop (downloader.go:38), so a blocked Next stops new downloads starting
@@ -122,7 +131,6 @@ func newElemIter(seq iter.Seq2[tgsource.Item, error], staging string, takeout bo
 }
 
 func (i *elemIter) Next(ctx context.Context) bool {
-	var item tgsource.Item
 	for {
 		if i.failure != nil || i.stopped.Load() {
 			return false
@@ -132,11 +140,7 @@ func (i *elemIter) Next(ctx context.Context) bool {
 			return false
 		}
 
-		var (
-			err error
-			ok  bool
-		)
-		item, err, ok = i.next()
+		item, err, ok := i.next()
 		if !ok {
 			return false
 		}
@@ -151,35 +155,110 @@ func (i *elemIter) Next(ctx context.Context) bool {
 		// is either a second caller or a gap there; in both cases one unwritable
 		// name must not strand the rest of the chat behind it.
 		if err := naming.Safe(item.Name); err != nil {
-			if len(i.skipped) < maxRecordedErrors {
-				i.skipped = append(i.skipped, fmt.Errorf("message %d: %w", item.MessageID, err))
+			i.skip(item, err)
+			continue
+		}
+
+		if i.acquire != nil {
+			if err := i.acquire(ctx, item.Size()); err != nil {
+				i.failure = err
+				return false
+			}
+		}
+
+		// The refreshed item replaces the original only on success. A failed
+		// refresh reports a zero Item, and handing that to the reservation or to
+		// the failure record would release nothing and blame message 0.
+		fresh, err := i.reminted(ctx, item)
+		if err == nil {
+			item = fresh
+		} else {
+			// Nothing was staged and nothing will be, so the reservation goes
+			// back before anything else; otherwise the cap ends the run
+			// over-committed by every item that failed here.
+			i.giveBack(item)
+			if ctx.Err() != nil {
+				i.failure = err
+				return false
+			}
+			// A message that is gone is skipped, for the same reason an
+			// unwritable name is: it is a permanent, per-message fact, and one
+			// of them must not strand every message behind it. Nothing is
+			// archived for it, so the run's closing verify still reports it as
+			// outstanding.
+			//
+			// Anything else is a failure and is recorded as one. Routing it
+			// there rather than onto the skip path is what keeps the breaker,
+			// the run's retry passes and the progress report working: a re-read
+			// fails because the connection did, which is exactly when every
+			// download is failing too, and a skip is invisible to all three — a
+			// source outage would otherwise walk the whole todo list one dead
+			// round trip at a time and still exit as merely incomplete.
+			if errors.Is(err, tgsource.ErrGone) {
+				i.skip(item, err)
+			} else if i.refused != nil {
+				i.refused(item, err)
 			}
 			continue
 		}
-		break
-	}
 
-	if i.acquire != nil {
-		if err := i.acquire(ctx, item.Size()); err != nil {
-			i.failure = err
+		f, err := os.OpenFile(partPath(i.staging, item), os.O_CREATE|os.O_RDWR, 0o600)
+		if err != nil {
+			// The reservation is handed back here because this item will never
+			// reach a download, so no OnDone will ever release it for us.
+			i.giveBack(item)
+			i.failure = fmt.Errorf("open destination for message %d: %w", item.MessageID, err)
 			return false
 		}
-	}
+		i.opened = append(i.opened, f)
 
-	f, err := os.OpenFile(partPath(i.staging, item), os.O_CREATE|os.O_RDWR, 0o600)
-	if err != nil {
-		// The reservation is handed back here because this item will never
-		// reach a download, so no OnDone will ever release it for us.
-		if i.release != nil {
-			i.release(item.Size())
-		}
-		i.failure = fmt.Errorf("open destination for message %d: %w", item.MessageID, err)
-		return false
+		i.current = &elem{item: item, file: f, takeout: i.takeout}
+		return true
 	}
-	i.opened = append(i.opened, f)
+}
 
-	i.current = &elem{item: item, file: f, takeout: i.takeout}
-	return true
+// reminted replaces an item's file reference with one Telegram has just issued.
+//
+// The reference in Media.InputFileLoc was minted when the walk saw the message,
+// and Telegram expires those. The lifetime is undocumented and was observed to
+// outlast an hour but not two, which is less than a run over a large chat spends
+// downloading — so by the time the dispatch loop reaches an item near the end of
+// the list its reference is dead, every remaining fetch returns
+// FILE_REFERENCE_EXPIRED, and the breaker reads that as a dead source and
+// abandons everything still to come.
+//
+// It is called as late as Next can leave it, and in particular after the budget
+// acquire rather than before. core's downloader reads Elem.File().Location()
+// once per attempt, inside the worker (downloader.go:89), so the only wait left
+// between here and there is for a free worker slot, which one download bounds.
+// Re-minting before the acquire would reintroduce the very failure this exists
+// to prevent: that acquire is the run's brake and blocks for as long as the
+// destination is slow, which on a stalled remote is long enough to expire a
+// fresh token all over again.
+//
+// One round trip per transfer, and no guess at how long a reference lives.
+// Retries get it for free: Run feeds a failed item straight back through
+// Download, so a reference that died mid-transfer — a multi-gigabyte file can
+// outlive its own token — is re-minted on the next pass rather than replayed.
+func (i *elemIter) reminted(ctx context.Context, item tgsource.Item) (tgsource.Item, error) {
+	if i.refresh == nil {
+		return item, nil
+	}
+	return i.refresh(ctx, item)
+}
+
+// skip records an item no run will ever fetch, without ending this one.
+func (i *elemIter) skip(item tgsource.Item, err error) {
+	if len(i.skipped) < maxRecordedErrors {
+		i.skipped = append(i.skipped, fmt.Errorf("message %d: %w", item.MessageID, err))
+	}
+}
+
+// giveBack returns an item's staging reservation.
+func (i *elemIter) giveBack(item tgsource.Item) {
+	if i.release != nil {
+		i.release(item.Size())
+	}
 }
 
 func (i *elemIter) Value() downloader.Elem { return i.current }

internal/pipeline/pipeline.go

@@ -46,6 +46,12 @@ type Options struct {
 	// Events, when set, receives the run's per-item lifecycle: which files are
 	// downloading, which are uploading, and how far along each one is.
 	Events Events
+
+	// Refresh, when set, re-reads a message for a live file reference just
+	// before its download starts. A run over a large chat outlasts the
+	// references its walk collected, so without this every fetch past that point
+	// fails with FILE_REFERENCE_EXPIRED; see elemIter.Next.
+	Refresh tgsource.Refresh
 }
 
 // Result is what a run achieved.
@@ -235,6 +241,7 @@ func Run(ctx context.Context, seq iter.Seq2[tgsource.Item, error], o Options) (R
 			Limit:       o.Limit,
 			Takeout:     o.Takeout,
 			Events:      o.Events,
+			Refresh:     o.Refresh,
 			acquire:     budget.acquire,
 			release:     budget.release,
 			maxFailures: o.MaxFailures,

internal/pipeline/progress.go

@@ -122,6 +122,46 @@ func (p *progress) OnDone(e downloader.Elem, err error) {
 	p.events.Stats(stats)
 }
 
+// refused records an item that failed before any transfer could start.
+//
+// It exists so that "the message could not be re-read" counts as the same kind
+// of event as "the transfer died", because operationally it is: both mean
+// Telegram is not serving this file right now, and both are usually the same
+// outage. An item that never reaches a worker never reaches OnAdd or OnDone, so
+// without this it would move no counter, feed no streak, produce no Outcome for
+// Run to retry, and leave the live report on a still frame while the pass walked
+// the rest of the list.
+//
+// Everything OnDone does for a failure it does here, minus the file: the item is
+// counted as started and failed, its bytes join the total the report is working
+// towards, the streak advances so a dead source still trips the breaker, and the
+// Outcome is what makes Run try the item again on the next pass.
+func (p *progress) refused(it tgsource.Item, err error) {
+	p.mu.Lock()
+	p.stats.Started++
+	p.stats.BytesTotal += it.Size()
+	p.stats.Failed++
+	p.streak++
+	p.outcomes = append(p.outcomes, Outcome{Item: it, Err: err})
+	stats := p.stats
+	// Decided under the lock and acted on outside it, as in OnDone.
+	trip := p.maxStreak > 0 && p.streak >= p.maxStreak && !p.tripped
+	if trip {
+		p.tripped = true
+	}
+	p.mu.Unlock()
+
+	if trip && p.onTrip != nil {
+		p.onTrip()
+	}
+	// Reported as a start and an immediate end rather than as an end alone: a
+	// reporter tracking which files are in flight is entitled to see every item
+	// it is told about finish, and one it never saw start would be a stray.
+	p.events.DownloadStart(it)
+	p.events.DownloadDone(it, err)
+	p.events.Stats(stats)
+}
+
 // brokeCircuit reports whether the failure streak ended this pass.
 func (p *progress) brokeCircuit() bool {
 	p.mu.Lock()

internal/pipeline/run_test.go

@@ -419,3 +419,35 @@ func TestRunDoesNotReportASourceThatRecovered(t *testing.T) {
 		t.Errorf("Failed() = %d, want 0", got)
 	}
 }
+
+// Every pass gets the refresher, retries included. Run feeds a failed item back
+// through Download as the same struct it failed with, so without this a
+// reference that expired mid-transfer would be replayed dead on every attempt.
+func TestRunRefreshesOnEveryPassIncludingRetries(t *testing.T) {
+	var passes int
+	refresh := func(_ context.Context, it tgsource.Item) (tgsource.Item, error) { return it, nil }
+
+	runFake(t, func(ctx context.Context, seq iter.Seq2[tgsource.Item, error], o DownloadOptions) ([]Outcome, Stats, error) {
+		passes++
+		if o.Refresh == nil {
+			t.Errorf("pass %d was given no refresher", passes)
+		}
+		return stageItems(map[int]bool{2: true})(ctx, seq, o)
+	})
+
+	o, staging, _ := runOpts(t, 0)
+	o.Refresh = refresh
+
+	res, err := Run(t.Context(), seqOf(testItems(2, 10)), o)
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if passes != downloadAttempts {
+		t.Fatalf("passes = %d, want %d — the retry passes are where a stale reference would be replayed",
+			passes, downloadAttempts)
+	}
+	if got := len(res.Failed()); got != 1 {
+		t.Errorf("Failed() = %d, want 1", got)
+	}
+	assertEmpty(t, staging)
+}

internal/tgsource/refresh.go

@@ -0,0 +1,129 @@
+package tgsource
+
+import (
+	"context"
+	"errors"
+	"fmt"
+
+	"github.com/gotd/td/telegram/peers"
+	"github.com/gotd/td/telegram/query"
+	"github.com/gotd/td/tg"
+
+	"github.com/iyear/tdl/core/tmedia"
+
+	"github.com/tiennm99dev/telegram-exporter/internal/naming"
+)
+
+// ErrGone marks a message this run can no longer fetch: it was deleted, it no
+// longer carries media, or the file behind it has been replaced by a different
+// one.
+//
+// The distinction from an ordinary re-read failure is what the caller acts on.
+// There is nothing here to retry and nothing a later pass would do differently,
+// so one such message is left unarchived and the run carries on — whereas a
+// re-read that failed because the connection did must count as a failure like
+// any other, or a dead source stops looking like one.
+var ErrGone = errors.New("message is gone")
+
+// Refresh re-reads one message and returns the item with a live file reference.
+//
+// A function type rather than an interface because there is one implementation
+// and the download half only ever needs to call it; a func also lets a test
+// supply a canned reference without a Telegram connection.
+type Refresh func(context.Context, Item) (Item, error)
+
+// fetchMessage reads one message of a chat by id, reporting ErrGone if it is not
+// there any more. It is the seam Refresher is built on, so the checks that
+// decide whether a re-read describes the same file can be tested without a
+// Telegram client.
+type fetchMessage func(ctx context.Context, messageID int) (*tg.Message, error)
+
+// Refresher re-mints the file reference for an item from the chat it came from.
+//
+// The reference in Media.InputFileLoc is a short-lived token, and Telegram
+// deliberately does not document how long it lives — the documented contract is
+// the other way round: a client caches the reference *together with the source
+// it came from*, and re-reads that source when the reference expires
+// (core.telegram.org/api/file-references). This is that re-read. The source is
+// the message, so reading the message again yields a document carrying a new
+// reference.
+func Refresher(api *tg.Client, peer peers.Peer) Refresh {
+	return refresher(historyFetch(api, peer.InputPeer()))
+}
+
+func refresher(get fetchMessage) Refresh {
+	return func(ctx context.Context, it Item) (Item, error) {
+		msg, err := get(ctx, it.MessageID)
+		if err != nil {
+			if errors.Is(err, ErrGone) {
+				return Item{}, err
+			}
+			return Item{}, fmt.Errorf("re-read message %d: %w", it.MessageID, err)
+		}
+
+		media, ok := tmedia.GetMedia(msg)
+		if !ok {
+			// The walk only yields messages that had media, so this means the
+			// message was edited into something else.
+			return Item{}, fmt.Errorf("%w: message %d no longer carries a file",
+				ErrGone, it.MessageID)
+		}
+
+		// The re-read has to describe the same file, and both halves of that are
+		// checked because both are load-bearing. Name is what the remote index
+		// was diffed against and what the file will be written as, so a run that
+		// silently accepted a new one would archive different bytes under the
+		// old message's name. Size is what reserved the staging budget, what the
+		// finished download is measured against, and what the closing verify
+		// compares — a changed size would either fail that check every pass or
+		// be re-fetched forever as a size mismatch.
+		//
+		// A mismatch is therefore not a refresh failure but a different file,
+		// which this run has no name for and must leave alone.
+		if name := naming.For(it.DialogID, it.MessageID, media); name != it.Name || media.Size != it.Media.Size {
+			return Item{}, fmt.Errorf("%w: message %d now holds a different file (%q, %d bytes; was %q, %d bytes)",
+				ErrGone, it.MessageID, name, media.Size, it.Name, it.Media.Size)
+		}
+
+		it.Media = media
+		return it, nil
+	}
+}
+
+// historyFetch reads a single message out of a chat's history.
+//
+// This is a one-message GetHistory rather than channels.getMessages so that
+// channels, groups and users need no branching, and it is written here rather
+// than reusing tutil.GetSingleMessage because the classification is the point.
+// tutil reports a deleted message three different ways — its own
+// ErrMessageDeleted when an older message sits in the slot, "invalid message"
+// when that message is a service one, and a wrapped nil when the page comes back
+// empty — and only the first is distinguishable. This needs every one of them to
+// be exactly ErrGone, because the caller retries anything else.
+//
+// OffsetID is exclusive, so id+1 asks for the message itself first. Telegram
+// answers with the newest message at or below that offset, and gotd drops empty
+// slots while paging (query/messages/iter.go), so a returned id lower than the
+// one asked for means the message is not in the history any more.
+func historyFetch(api *tg.Client, peer tg.InputPeerClass) fetchMessage {
+	return func(ctx context.Context, messageID int) (*tg.Message, error) {
+		it := query.NewQuery(api).Messages().GetHistory(peer).
+			OffsetID(messageID + 1).BatchSize(1).Iter()
+
+		if !it.Next(ctx) {
+			// gotd ends the iterator with a nil error on an empty page, so a
+			// nil here is not success: it is "nothing at or below this id".
+			if err := it.Err(); err != nil {
+				return nil, err
+			}
+			return nil, fmt.Errorf("%w: no message at or before %d", ErrGone, messageID)
+		}
+
+		msg, ok := it.Value().Msg.(*tg.Message)
+		if !ok || msg.ID != messageID {
+			return nil, fmt.Errorf("%w: message %d is no longer in the chat's history",
+				ErrGone, messageID)
+		}
+		return msg, nil
+	}
+}

internal/tgsource/refresh_test.go

@@ -0,0 +1,167 @@
+package tgsource
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"strings"
+	"testing"
+
+	"github.com/gotd/td/tg"
+
+	"github.com/iyear/tdl/core/tmedia"
+
+	"github.com/tiennm99dev/telegram-exporter/internal/naming"
+)
+
+// These drive refresher through its fetch seam rather than through a Telegram
+// client, because what needs testing is not the round trip — it is the judgement
+// afterwards. A re-read that describes a different file must be refused, and a
+// re-read that failed because the message is gone must be distinguishable from
+// one that failed because the connection did: the download half retries the
+// second and skips the first.
+
+const testDialogID = 1234567890
+
+// docMessage builds a message carrying one document, the shape tmedia extracts.
+func docMessage(t *testing.T, id int, file string, size int64, ref string) *tg.Message {
+	t.Helper()
+	msg := &tg.Message{ID: id}
+	msg.SetMedia(&tg.MessageMediaDocument{
+		Document: &tg.Document{
+			ID:            1000000000000000001,
+			Size:          size,
+			DCID:          2,
+			MimeType:      "video/mp4",
+			FileReference: []byte(ref),
+			Attributes: []tg.DocumentAttributeClass{
+				&tg.DocumentAttributeFilename{FileName: file},
+			},
+		},
+	})
+	return msg
+}
+
+// walked is the item as the walk first saw it, carrying a reference that has
+// since expired.
+func walked(t *testing.T, id int, file string, size int64) Item {
+	t.Helper()
+	media, ok := tmedia.GetMedia(docMessage(t, id, file, size, "stale"))
+	if !ok {
+		t.Fatal("fixture message carries no media")
+	}
+	return Item{
+		DialogID:  testDialogID,
+		MessageID: id,
+		Name:      naming.For(testDialogID, id, media),
+		Media:     media,
+	}
+}
+
+func TestRefresherSubstitutesALiveReference(t *testing.T) {
+	it := walked(t, 4242, "clip.mp4", 5000)
+	name := it.Name
+
+	got, err := refresher(func(_ context.Context, id int) (*tg.Message, error) {
+		if id != 4242 {
+			t.Errorf("fetched message %d, want 4242", id)
+		}
+		return docMessage(t, 4242, "clip.mp4", 5000, "fresh"), nil
+	})(t.Context(), it)
+	if err != nil {
+		t.Fatalf("refresh: %v", err)
+	}
+
+	loc, ok := got.Media.InputFileLoc.(*tg.InputDocumentFileLocation)
+	if !ok {
+		t.Fatalf("location is %T, want *tg.InputDocumentFileLocation", got.Media.InputFileLoc)
+	}
+	if string(loc.FileReference) != "fresh" {
+		t.Errorf("file reference = %q, want the re-read one", loc.FileReference)
+	}
+	// The name is the remote's key: it is what the run diffed against the index
+	// and what the file will be written as, so a refresh must never move it.
+	if got.Name != name {
+		t.Errorf("name = %q, want it unchanged at %q", got.Name, name)
+	}
+	if got.Media.Size != 5000 {
+		t.Errorf("size = %d, want 5000", got.Media.Size)
+	}
+}
+
+// Every permanent reason a re-read cannot be used has to arrive as ErrGone,
+// because that is the only signal that says "skip this one and keep going"
+// rather than "the source is failing".
+func TestRefresherReportsGoneForEveryPermanentCase(t *testing.T) {
+	it := walked(t, 4242, "clip.mp4", 5000)
+
+	for _, tc := range []struct {
+		name string
+		get  fetchMessage
+		want string
+	}{
+		{
+			name: "message no longer in the history",
+			get: func(context.Context, int) (*tg.Message, error) {
+				return nil, fmt.Errorf("%w: no message at or before 4242", ErrGone)
+			},
+			want: "4242",
+		},
+		{
+			name: "message carries no media any more",
+			get: func(context.Context, int) (*tg.Message, error) {
+				return &tg.Message{ID: 4242, Message: "edited into text"}, nil
+			},
+			want: "no longer carries a file",
+		},
+		{
+			name: "a different file under the same message",
+			get: func(_ context.Context, _ int) (*tg.Message, error) {
+				return docMessage(t, 4242, "other.mp4", 5000, "fresh"), nil
+			},
+			want: "different file",
+		},
+		{
+			// Size is what reserved the staging budget, what the finished
+			// download is measured against and what the closing verify
+			// compares, so a same-name replacement of a different length is a
+			// different file too.
+			name: "same name, different length",
+			get: func(_ context.Context, _ int) (*tg.Message, error) {
+				return docMessage(t, 4242, "clip.mp4", 6000, "fresh"), nil
+			},
+			want: "different file",
+		},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			_, err := refresher(tc.get)(t.Context(), it)
+			if !errors.Is(err, ErrGone) {
+				t.Fatalf("err = %v, want ErrGone", err)
+			}
+			if !strings.Contains(err.Error(), tc.want) {
+				t.Errorf("err = %v, want it to mention %q", err, tc.want)
+			}
+		})
+	}
+}
+
+// A re-read that failed for any other reason is not permanent and must not be
+// dressed up as one: the caller counts these as download failures, retries them,
+// and gives up on the source once enough of them land in a row.
+func TestRefresherKeepsATransientFailureRetryable(t *testing.T) {
+	it := walked(t, 4242, "clip.mp4", 5000)
+
+	_, err := refresher(func(context.Context, int) (*tg.Message, error) {
+		return nil, errors.New("dial telegram: connection refused")
+	})(t.Context(), it)
+
+	if err == nil {
+		t.Fatal("a failed re-read reported success")
+	}
+	if errors.Is(err, ErrGone) {
+		t.Errorf("a connection failure was classified as ErrGone: %v", err)
+	}
+	if !strings.Contains(err.Error(), "4242") {
+		t.Errorf("err = %v, want it to name the message", err)
+	}
+}