diff --git a/internal/agent/guardrails.go b/internal/agent/guardrails.go
index 089dc818e..5657aa3c4 100644
--- a/internal/agent/guardrails.go
+++ b/internal/agent/guardrails.go
@@ -2,6 +2,7 @@ package agent
 
 import (
 	"encoding/json"
+	"fmt"
 	"strconv"
 	"strings"
 
@@ -284,12 +285,67 @@ const (
 	noOutputStopSuffix = "to avoid consuming tokens without making progress."
 )
 
+// providerEmptyStream reports that a cleanly-completed turn carried NOTHING —
+// no visible text, no tool calls (real or dropped), and no reasoning signal.
+// That shape is a transport/backend fault (e.g. the ollama cloud relay
+// answering 200 with an instantly-empty SSE under load), not model behavior:
+// a model that is generating always leaves SOME trace (text, a tool call
+// attempt, or reasoning deltas).
+func providerEmptyStream(collected zeroruntime.CollectedStream) bool {
+	return strings.TrimSpace(collected.Text) == "" &&
+		len(collected.ToolCalls) == 0 &&
+		collected.DroppedToolCalls == 0 &&
+		len(collected.ReasoningBlocks) == 0 &&
+		!collected.ReasoningEmitted
+}
+
+// allEmptyTurnsProviderEmpty reports that every strike toward the no-output
+// guard was a provider-empty stream, so the stop message should name the
+// backend fault rather than generic agent no-progress.
+func (state *guardState) allEmptyTurnsProviderEmpty() bool {
+	return state.emptyTurns > 0 && state.providerEmptyTurns == state.emptyTurns
+}
+
 // noOutputStopAnswer is the final answer returned when the no-output guard
 // stops the run. The turn count is interpolated at the call site.
 func noOutputStopAnswer(turns int) string {
 	return noOutputStopPrefix + strconv.Itoa(turns) + " turns " + noOutputStopMarker + " " + noOutputStopSuffix
 }
 
+// providerEmptyStopAnswer is the final answer when every guard strike was a
+// provider-empty stream: the backend repeatedly answered with contentless
+// completions (already retried with backoff inside each turn). It tells the
+// user the truth — the provider failed, not the agent — and what to do next.
+// detail carries wire-level diagnostics from the last empty attempt
+// (finish reason, token counts) so the condition is debuggable straight from
+// the session log — this exact failure class previously took a live traffic
+// capture to diagnose because nothing about the response shape was recorded.
+func providerEmptyStopAnswer(turns int, detail string) string {
+	answer := providerEmptyStopPrefix +
+		strconv.Itoa(turns) + " consecutive attempts, each already retried with backoff. " +
+		"The backend is likely rate-limiting or degraded right now — try again shortly, or switch providers/models (zero providers use <name>)."
+	if detail != "" {
+		answer += " [last attempt: " + detail + "]"
+	}
+	return answer
+}
+
+// emptyStreamDetail renders the wire-level shape of an empty attempt for
+// providerEmptyStopAnswer.
+func emptyStreamDetail(collected zeroruntime.CollectedStream) string {
+	finish := collected.FinishReason
+	if finish == "" {
+		finish = "stop"
+	}
+	return fmt.Sprintf("finish_reason=%s, output_tokens=%d, reasoning_tokens=%d",
+		finish, collected.Usage.OutputTokens, collected.Usage.ReasoningTokens)
+}
+
+// providerEmptyStopPrefix is the stable head of providerEmptyStopAnswer, used
+// by IsNoProgressStop so session titling / resume filtering treat the
+// provider-empty stop like the other guard answers (not real content).
+const providerEmptyStopPrefix = "The model provider returned an empty response (no text, no tool calls, no reasoning) on "
+
 // IsNoProgressStop reports whether content IS the no-output guardrail stop answer
 // (a run that produced no visible text and no tool calls). It matches the EXACT
 // structure noOutputStopAnswer emits — prefix + "<int> turns " + marker + " " +
@@ -300,6 +356,15 @@ func noOutputStopAnswer(turns int) string {
 // and skip its title generation.
 func IsNoProgressStop(content string) bool {
 	trimmed := strings.TrimSpace(content)
+	// The provider-empty variant: stable prefix + a bare integer count.
+	if rest, ok := strings.CutPrefix(trimmed, providerEmptyStopPrefix); ok {
+		if sep := strings.Index(rest, " consecutive attempts"); sep > 0 {
+			if _, err := strconv.Atoi(rest[:sep]); err == nil {
+				return true
+			}
+		}
+		return false
+	}
 	if !strings.HasPrefix(trimmed, noOutputStopPrefix) {
 		return false
 	}
@@ -352,6 +417,7 @@ func toolOnlyProgressReminder(turns int) string {
 // purely from tool-call names and per-turn output, matching what the loop holds.
 type guardState struct {
 	emptyTurns               int
+	providerEmptyTurns       int
 	totalToolCalls           int
 	toolCallsSincePlanUpdate int
 	planEverCalled           bool
@@ -417,8 +483,18 @@ func (state *guardState) observeTurn(collected zeroruntime.CollectedStream) (sto
 
 	if hasToolCalls || hasVisibleText {
 		state.emptyTurns = 0
+		state.providerEmptyTurns = 0
 	} else {
 		state.emptyTurns++
+		// Track separately whether the empty turn was a PROVIDER-empty stream
+		// (nothing at all came back — not even reasoning) vs a behavioral empty
+		// (the model thought/streamed but committed no answer). When every
+		// strike was provider-empty, the stop message names the backend fault
+		// instead of blaming the agent — the difference between the user
+		// retrying/switching providers and filing a "the agent gives up" bug.
+		if providerEmptyStream(collected) {
+			state.providerEmptyTurns++
+		}
 	}
 	if hasToolCalls && !hasVisibleText {
 		state.toolOnlyTurns++
diff --git a/internal/agent/guardrails_test.go b/internal/agent/guardrails_test.go
index fd7758df0..e8e564fe4 100644
--- a/internal/agent/guardrails_test.go
+++ b/internal/agent/guardrails_test.go
@@ -9,11 +9,23 @@ import (
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
-// emptyTurn is a stream that produces no visible text and no tool calls.
+// emptyTurn is a PROVIDER-empty stream: it completes cleanly carrying nothing
+// at all (the ollama-cloud 200+[DONE] shape). The loop retries these in-turn
+// before counting a no-output strike.
 func emptyTurn() []zeroruntime.StreamEvent {
 	return []zeroruntime.StreamEvent{{Type: zeroruntime.StreamEventDone}}
 }
 
+// reasoningOnlyTurn is a BEHAVIORAL empty turn: the model streamed reasoning
+// (so the provider clearly worked) but committed no text and no tool calls.
+// These are NOT retried in-turn; they count directly toward the no-output guard.
+func reasoningOnlyTurn() []zeroruntime.StreamEvent {
+	return []zeroruntime.StreamEvent{
+		{Type: zeroruntime.StreamEventReasoning, Content: "thinking…"},
+		{Type: zeroruntime.StreamEventDone},
+	}
+}
+
 // textTurn produces a turn with visible assistant text.
 func textTurn(content string) []zeroruntime.StreamEvent {
 	return []zeroruntime.StreamEvent{
@@ -45,9 +57,9 @@ func countUserMessagesContaining(messages []zeroruntime.Message, needle string)
 func TestRunStopsAfterConsecutiveEmptyTurns(t *testing.T) {
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
-			emptyTurn(),
-			emptyTurn(),
-			emptyTurn(),
+			reasoningOnlyTurn(),
+			reasoningOnlyTurn(),
+			reasoningOnlyTurn(),
 			// A 4th turn exists but must never be requested.
 			textTurn("should never reach here"),
 		},
@@ -77,10 +89,10 @@ func TestRunStopsAfterConsecutiveEmptyTurns(t *testing.T) {
 func TestRunResetsEmptyTurnCounterOnVisibleOutput(t *testing.T) {
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
-			emptyTurn(),
-			emptyTurn(),
+			reasoningOnlyTurn(),
+			reasoningOnlyTurn(),
 			textTurn("here is real progress"), // resets the counter and is the final answer
-			emptyTurn(),
+			reasoningOnlyTurn(),
 		},
 	}
 
@@ -109,11 +121,11 @@ func TestRunResetsEmptyTurnCounterOnToolCall(t *testing.T) {
 
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
-			emptyTurn(),
-			emptyTurn(),
+			reasoningOnlyTurn(),
+			reasoningOnlyTurn(),
 			toolTurn("call-1", "read_file", `{"path":"notes.txt"}`), // resets counter
-			emptyTurn(),
-			emptyTurn(),
+			reasoningOnlyTurn(),
+			reasoningOnlyTurn(),
 			textTurn("done"),
 		},
 	}
@@ -455,3 +467,96 @@ func TestRunInjectsToolFailureHintWithSchema(t *testing.T) {
 		t.Fatalf("expected a tool-failure hint on the 3rd turn, messages: %+v", provider.requests[2].Messages)
 	}
 }
+
+// A provider-empty stream (clean completion carrying nothing at all) is
+// retried in-turn with backoff before it may count as a no-output strike; when
+// the backend keeps answering empty, the run stops with a message that names
+// the PROVIDER fault — not the generic "agent made no progress" text. This is
+// the exact live failure observed on the ollama cloud relay (HTTP 200 in ~1s,
+// SSE straight to [DONE]), which silently killed 56% of sessions on the
+// reporting machine.
+func TestRunRetriesProviderEmptyStreamsThenStopsWithProviderMessage(t *testing.T) {
+	turns := make([][]zeroruntime.StreamEvent, 0, 9)
+	for i := 0; i < 9; i++ { // 3 strikes × (1 initial + 2 in-turn retries)
+		turns = append(turns, emptyTurn())
+	}
+	provider := &mockProvider{turns: turns}
+
+	result, err := Run(context.Background(), "go", provider, Options{
+		Registry: tools.NewRegistry(),
+		MaxTurns: 12,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	wantRequests := maxEmptyTurns * (1 + maxEmptyStreamRetries)
+	if len(provider.requests) != wantRequests {
+		t.Fatalf("expected %d requests (%d strikes × %d attempts), got %d",
+			wantRequests, maxEmptyTurns, 1+maxEmptyStreamRetries, len(provider.requests))
+	}
+	if !strings.Contains(result.FinalAnswer, "provider returned an empty response") {
+		t.Fatalf("expected the provider-empty stop message, got %q", result.FinalAnswer)
+	}
+	if strings.Contains(result.FinalAnswer, noOutputStopMarker) {
+		t.Fatalf("provider fault must not be reported as generic agent no-progress: %q", result.FinalAnswer)
+	}
+	if !IsNoProgressStop(result.FinalAnswer) {
+		t.Fatal("the provider-empty stop answer must be recognized by IsNoProgressStop (titling/resume filters)")
+	}
+	// Wire-level diagnostics must land in the message (and therefore the
+	// session log) so this failure class is debuggable without a traffic capture.
+	if !strings.Contains(result.FinalAnswer, "output_tokens=0") || !strings.Contains(result.FinalAnswer, "finish_reason=") {
+		t.Fatalf("stop answer must carry last-attempt diagnostics, got %q", result.FinalAnswer)
+	}
+}
+
+// A transient empty response recovers on the in-turn retry: no strike, no stop,
+// the retried turn's real output is the answer.
+func TestRunRecoversWhenEmptyStreamRetrySucceeds(t *testing.T) {
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			emptyTurn(),           // initial attempt: backend hiccup
+			textTurn("recovered"), // in-turn retry gets the real answer
+		},
+	}
+
+	result, err := Run(context.Background(), "go", provider, Options{
+		Registry: tools.NewRegistry(),
+		MaxTurns: 12,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(provider.requests) != 2 {
+		t.Fatalf("expected 2 requests (initial + one retry), got %d", len(provider.requests))
+	}
+	if result.FinalAnswer != "recovered" {
+		t.Fatalf("expected the retried turn's text as final answer, got %q", result.FinalAnswer)
+	}
+	if result.Turns != 1 {
+		t.Fatalf("an in-turn retry is the SAME turn, want 1 turn, got %d", result.Turns)
+	}
+}
+
+// Mixed strikes (provider-empty + behavioral reasoning-only) must NOT claim a
+// provider fault — the generic no-output message stays.
+func TestRunMixedEmptyTurnsKeepGenericStopMessage(t *testing.T) {
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			reasoningOnlyTurn(),                   // strike 1: behavioral
+			emptyTurn(), emptyTurn(), emptyTurn(), // strike 2: provider-empty ×(1+2 retries)
+			reasoningOnlyTurn(), // strike 3: behavioral
+		},
+	}
+
+	result, err := Run(context.Background(), "go", provider, Options{
+		Registry: tools.NewRegistry(),
+		MaxTurns: 12,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(result.FinalAnswer, noOutputStopMarker) {
+		t.Fatalf("mixed strikes must use the generic no-output message, got %q", result.FinalAnswer)
+	}
+}
diff --git a/internal/agent/loop.go b/internal/agent/loop.go
index 29c99a7d3..b969993cf 100644
--- a/internal/agent/loop.go
+++ b/internal/agent/loop.go
@@ -25,7 +25,20 @@ const maxTurnsFinalAnswerPrompt = "You have reached the tool-turn limit. Do not
 // WITH NO OUTPUT yet is re-issued on a fresh connection before giving up. Only
 // the no-output case is retried (a partial turn would duplicate), so this is a
 // safe recovery for a stalled/dead pooled connection.
-const maxStreamStallRetries = 2
+// maxStreamStallRetries: 3 in-turn re-issues (4 attempts total) approaches
+// Codex's stream_max_retries=5 posture while staying hard-capped (opencode's
+// unbounded session retry has produced infinite loops on this same failure).
+// Affordable now that a mid-stream death is detected at the gap timeout
+// (~2m) instead of the full 5m idle window.
+const maxStreamStallRetries = 3
+
+// maxEmptyStreamRetries caps in-turn re-issues of a request whose stream
+// completed cleanly but carried nothing at all (providerEmptyStream). Observed
+// live on the ollama cloud relay under load: HTTP 200 in ~1s with an SSE that
+// goes straight to [DONE]. The condition is transient backend state, so a
+// short backoff-retry usually recovers; without it three such responses
+// arrived inside two seconds and killed the run via the no-output guard.
+const maxEmptyStreamRetries = 2
 
 const (
 	toolResultMetaControl       = "control"
@@ -324,6 +337,35 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			}
 			collected = zeroruntime.CollectStreamWithOptions(ctx, retryStream, forwardingOpts)
 		}
+		// Provider-empty retry: the stream completed CLEANLY but carried nothing
+		// — no text, no tool calls, no reasoning. That is not model behavior (a
+		// generating model always leaves some trace); it is a backend answering
+		// with a contentless completion (seen on the ollama cloud relay under
+		// load: 200 + instant [DONE]). Transient, so re-issue with backoff before
+		// letting it count as a no-output strike. Nothing from the empty stream
+		// was forwarded or committed to history, so the retry is clean.
+		for attempt := 1; attempt <= maxEmptyStreamRetries &&
+			collected.Error == "" && !forwardedVisibleText &&
+			providerEmptyStream(collected); attempt++ {
+			if notify := emptyRetryNoticeFor(options); notify != nil {
+				notify(attempt, maxEmptyStreamRetries)
+			}
+			if err := sleepWithContext(ctx, backoffFor(attempt)); err != nil {
+				result.Messages = copyMessages(messages)
+				return result, err
+			}
+			retryRequest := zeroruntime.CompletionRequest{
+				Messages:        copyMessages(messages),
+				Tools:           exposed,
+				ReasoningEffort: options.ReasoningEffort,
+			}
+			retryStream, retryErr := streamWithReconnect(ctx, provider, retryRequest, reconnectNoticeFor(options))
+			if retryErr != nil {
+				result.Messages = copyMessages(messages)
+				return result, retryErr
+			}
+			collected = zeroruntime.CollectStreamWithOptions(ctx, retryStream, forwardingOpts)
+		}
 		if collected.Error != "" {
 			// Route a reissued stream's non-stall error through the SAME recovery as
 			// the initial stream (image-rejection wrapping / context-limit compaction)
@@ -370,6 +412,15 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			// A truly-empty turn (no text, no tool calls, no dropped calls) is
 			// counted toward the runaway cap so we stop before burning maxTurns.
 			if guards.observeTurn(collected) {
+				// Every strike being a provider-empty stream means the BACKEND
+				// failed repeatedly (each attempt already retried with backoff
+				// above) — say so, instead of the generic no-progress message
+				// that reads as the agent giving up.
+				if guards.allEmptyTurnsProviderEmpty() {
+					result.FinalAnswer = providerEmptyStopAnswer(result.Turns, emptyStreamDetail(collected))
+					result.Messages = copyMessages(messages)
+					return result, nil
+				}
 				result.FinalAnswer = noOutputStopAnswer(result.Turns)
 				result.Messages = copyMessages(messages)
 				return result, nil
diff --git a/internal/agent/reconnect.go b/internal/agent/reconnect.go
index 641366a95..b0a3053a5 100644
--- a/internal/agent/reconnect.go
+++ b/internal/agent/reconnect.go
@@ -55,6 +55,22 @@ func stallRetryNoticeFor(options Options) reconnectNotifier {
 	}
 }
 
+// emptyRetryNoticeFor builds a notifier for the loop's provider-empty retry
+// (a stream that completed cleanly but carried NOTHING — no text, no tool
+// calls, no reasoning). Distinct wording from the stall notice because nothing
+// timed out: the backend answered instantly with an empty completion (observed
+// on the ollama cloud relay under load, and possible on any gateway having a
+// bad moment). Surfaced through OnReasoning, the non-content channel. Nil when
+// there is no reasoning sink (the retry still happens silently).
+func emptyRetryNoticeFor(options Options) reconnectNotifier {
+	if options.OnReasoning == nil {
+		return nil
+	}
+	return func(attempt, max int) {
+		options.OnReasoning(fmt.Sprintf("\n[provider returned an empty response — retrying %d/%d…]\n", attempt, max))
+	}
+}
+
 // streamWithReconnect issues request via provider.StreamCompletion and, on a
 // transient disconnect error, retries the connect up to maxStreamReconnects
 // times with exponential backoff. It returns the live stream on success, or the
diff --git a/internal/providers/openai/provider.go b/internal/providers/openai/provider.go
index 49de6a0a4..2bd1d7d27 100644
--- a/internal/providers/openai/provider.go
+++ b/internal/providers/openai/provider.go
@@ -276,6 +276,16 @@ func (provider *Provider) emitPayload(ctx context.Context, data string, state *t
 	return true
 }
 
+// firstNonEmptyString returns the first argument that is not empty.
+func firstNonEmptyString(values ...string) string {
+	for _, value := range values {
+		if value != "" {
+			return value
+		}
+	}
+	return ""
+}
+
 func (provider *Provider) emitChunk(
 	ctx context.Context,
 	chunk streamChunk,
@@ -283,10 +293,12 @@ func (provider *Provider) emitChunk(
 	events chan<- zeroruntime.StreamEvent,
 ) {
 	for _, choice := range chunk.Choices {
-		if choice.Delta.ReasoningContent != "" {
+		// Two dialects for the same thing (see streamDelta): prefer
+		// `reasoning_content`, fall back to `reasoning`. No backend sends both.
+		if reasoning := firstNonEmptyString(choice.Delta.ReasoningContent, choice.Delta.Reasoning); reasoning != "" {
 			sendEvent(ctx, events, zeroruntime.StreamEvent{
 				Type:    zeroruntime.StreamEventReasoning,
-				Content: choice.Delta.ReasoningContent,
+				Content: reasoning,
 			})
 		}
 		if choice.Delta.Content != "" {
diff --git a/internal/providers/openai/provider_test.go b/internal/providers/openai/provider_test.go
index 69d7e9027..01cc87b40 100644
--- a/internal/providers/openai/provider_test.go
+++ b/internal/providers/openai/provider_test.go
@@ -227,6 +227,33 @@ func TestStreamCompletionEmitsReasoningContentDeltas(t *testing.T) {
 	}
 }
 
+// Ollama's OpenAI-compat endpoint, OpenRouter, and most gateways stream
+// thinking under `reasoning` (not DeepSeek's `reasoning_content`). Dropping it
+// made every thinking token on those backends vanish: the model looked frozen
+// for minutes and reasoning-only turns were indistinguishable from a dead
+// provider. The exact shape below is a live capture from ollama glm-5.2:cloud.
+func TestStreamCompletionEmitsOllamaStyleReasoningDeltas(t *testing.T) {
+	provider := newTestProvider(t, func(w http.ResponseWriter, r *http.Request) {
+		writeSSE(w, `{"choices":[{"index":0,"delta":{"role":"assistant","content":"","reasoning":"The user"},"finish_reason":null}]}`)
+		writeSSE(w, `{"choices":[{"index":0,"delta":{"role":"assistant","content":"","reasoning":" wants hello"},"finish_reason":null}]}`)
+		writeSSE(w, `{"choices":[{"index":0,"delta":{"role":"assistant","content":"hello"},"finish_reason":"stop"}]}`)
+		writeSSE(w, `[DONE]`)
+	})
+
+	events := collectProviderEvents(t, provider)
+	reasoning := eventsOfType(events, zeroruntime.StreamEventReasoning)
+	if len(reasoning) != 2 {
+		t.Fatalf("reasoning events = %#v, want two reasoning deltas (ollama `reasoning` field regression)", reasoning)
+	}
+	if reasoning[0].Content != "The user" || reasoning[1].Content != " wants hello" {
+		t.Fatalf("unexpected reasoning events: %#v", reasoning)
+	}
+	text := eventsOfType(events, zeroruntime.StreamEventText)
+	if len(text) != 1 || text[0].Content != "hello" {
+		t.Fatalf("content after thinking must still stream as text, got %#v", text)
+	}
+}
+
 func TestStreamCompletionEmitsReasoningBeforeRegularContent(t *testing.T) {
 	provider := newTestProvider(t, func(w http.ResponseWriter, r *http.Request) {
 		writeSSE(w, `{"choices":[{"delta":{"reasoning_content":"Thinking. ","content":"Answer."}}]}`)
diff --git a/internal/providers/openai/types.go b/internal/providers/openai/types.go
index f093ce30b..8be080f21 100644
--- a/internal/providers/openai/types.go
+++ b/internal/providers/openai/types.go
@@ -76,8 +76,16 @@ type streamChoice struct {
 }
 
 type streamDelta struct {
-	Content          string                `json:"content"`
+	Content string `json:"content"`
+	// Reasoning/thinking deltas arrive under DIFFERENT keys depending on the
+	// backend dialect: DeepSeek-style servers emit `reasoning_content`, while
+	// ollama's OpenAI-compat endpoint, OpenRouter, and most gateways emit
+	// `reasoning`. Parse both — dropping `reasoning` made every thinking token
+	// on those backends silently vanish (the model looked frozen for minutes,
+	// its thinking was never replayed into the next turn's context, and a
+	// reasoning-only turn was indistinguishable from a dead provider).
 	ReasoningContent string                `json:"reasoning_content"`
+	Reasoning        string                `json:"reasoning"`
 	ToolCalls        []streamToolCallDelta `json:"tool_calls"`
 }
 
diff --git a/internal/providers/providerio/providerio.go b/internal/providers/providerio/providerio.go
index d3199876d..6d97b830b 100644
--- a/internal/providers/providerio/providerio.go
+++ b/internal/providers/providerio/providerio.go
@@ -82,6 +82,53 @@ func ContentStallTimeout(idleTimeout time.Duration) time.Duration {
 	return idleTimeout + extra
 }
 
+// DefaultStreamGapTimeout bounds the byte-silent gap AFTER a stream has
+// started delivering payloads. Once ANY payload (data or keep-alive) has
+// arrived, the backend has proven it heartbeats/streams — so a subsequent gap
+// with zero bytes is almost always a dead connection (live capture of the
+// reported "stuck for 33 minutes": ~119KB streamed, then the socket went
+// byte-frozen forever). Detecting that at 2 minutes instead of the 5-minute
+// idle default turns a ~15-minute blind wait (3 × 5m attempts) into a ~6-minute
+// bounded recovery, without touching the pre-first-byte window that slow cloud
+// proxies legitimately need (they may withhold output until the upstream model
+// produces its first token). Override with ZERO_STREAM_GAP_TIMEOUT. For
+// comparison: Codex ships the same 5m idle with 5 stream retries; opencode
+// suffers this exact freeze (idle HTTPS socket in "working" state) with no
+// mid-stream gap detection at all.
+const DefaultStreamGapTimeout = 2 * time.Minute
+
+// streamGapTimeoutEnv overrides the post-first-payload gap timeout. Accepts the
+// same forms as ZERO_STREAM_IDLE_TIMEOUT; "0"/"off" disables the tightened gap
+// (the plain idle timeout then applies for the whole stream).
+const streamGapTimeoutEnv = "ZERO_STREAM_GAP_TIMEOUT"
+
+// ResolveStreamGapTimeout selects the effective post-first-payload gap timeout
+// for a stream whose idle timeout resolved to idleTimeout. It is clamped to
+// never EXCEED the idle timeout (a gap larger than idle would be dead config),
+// and disabled entirely when the idle watchdog is disabled.
+func ResolveStreamGapTimeout(idleTimeout time.Duration) time.Duration {
+	if idleTimeout <= 0 {
+		return 0
+	}
+	gap := DefaultStreamGapTimeout
+	if raw := strings.TrimSpace(os.Getenv(streamGapTimeoutEnv)); raw != "" {
+		switch strings.ToLower(raw) {
+		case "0", "off", "none", "disabled":
+			return idleTimeout // no tightened gap: idle applies throughout
+		default:
+			if d, err := time.ParseDuration(raw); err == nil && d > 0 {
+				gap = d
+			} else if secs, err := strconv.Atoi(raw); err == nil && secs > 0 {
+				gap = time.Duration(secs) * time.Second
+			}
+		}
+	}
+	if gap > idleTimeout {
+		return idleTimeout
+	}
+	return gap
+}
+
 // streamIdleTimeoutEnv is the global override for the stream idle timeout. It
 // accepts a Go duration ("5m", "300s", "90s") or a bare number of seconds
 // ("300"). A value of "0", "off", "none", or "disabled" turns the watchdog off
@@ -317,7 +364,14 @@ func ScanSSEDataWithContext(
 		idle := time.NewTimer(idleTimeout)
 		defer idle.Stop()
 		idleC = idle.C
-		resetIdle = reset(idle, idleTimeout)
+		// The timer was armed above with the FULL idle window, which governs
+		// until the first payload (slow cloud proxies may withhold output until
+		// the upstream model produces its first token). resetIdle only runs when
+		// a payload HAS arrived — the backend has proven it streams/heartbeats —
+		// so every re-arm uses the tighter gap window: a byte-frozen socket
+		// mid-stream is a dead connection, and waiting the full idle for it just
+		// multiplies the user's blind wait per retry.
+		resetIdle = reset(idle, ResolveStreamGapTimeout(idleTimeout))
 
 		// Content watchdog: only real data lines reset it (keep-alives do not), so a
 		// stream that heartbeats without producing output is bounded instead of
diff --git a/internal/providers/providerio/providerio_test.go b/internal/providers/providerio/providerio_test.go
index 33f539a3f..1a029576f 100644
--- a/internal/providers/providerio/providerio_test.go
+++ b/internal/providers/providerio/providerio_test.go
@@ -371,3 +371,86 @@ func TestHTTPClientReturnsStallHardenedSharedClient(t *testing.T) {
 		t.Fatal("an explicit client must be returned unchanged")
 	}
 }
+
+// Once a stream has delivered its first payload, a byte-silent gap is bounded
+// by the TIGHTER gap timeout, not the full idle window — a mid-stream frozen
+// socket (live-captured "stuck" shape: bytes flowed, then nothing forever)
+// must be detected fast so retries aren't multiplied by five-minute waits.
+func TestScanSSETightensGapAfterFirstPayload(t *testing.T) {
+	t.Setenv("ZERO_STREAM_GAP_TIMEOUT", "80ms")
+	pr, pw := io.Pipe()
+	defer func() { _ = pw.Close() }()
+
+	go func() {
+		_, _ = io.WriteString(pw, "data: first\n\n")
+		// then: byte-frozen forever
+	}()
+
+	start := time.Now()
+	done := make(chan error, 1)
+	go func() {
+		done <- ScanSSEDataWithContext(context.Background(), func() {}, pr, 2*time.Second, func(string) bool { return true })
+	}()
+	select {
+	case err := <-done:
+		if !errors.Is(err, ErrStreamIdle) {
+			t.Fatalf("err = %v, want ErrStreamIdle", err)
+		}
+		if elapsed := time.Since(start); elapsed > time.Second {
+			t.Fatalf("gap abort took %s — the tightened window (80ms) did not apply after the first payload", elapsed)
+		}
+	case <-time.After(5 * time.Second):
+		t.Fatal("stream not aborted")
+	}
+}
+
+// BEFORE the first payload the FULL idle window applies: slow cloud proxies may
+// withhold output until the upstream model emits its first token, so the gap
+// timeout must NOT govern the wait for the first byte.
+func TestScanSSEKeepsFullIdleBeforeFirstPayload(t *testing.T) {
+	t.Setenv("ZERO_STREAM_GAP_TIMEOUT", "30ms")
+	pr, pw := io.Pipe()
+	defer func() { _ = pw.Close() }()
+	// never write anything
+
+	start := time.Now()
+	done := make(chan error, 1)
+	go func() {
+		done <- ScanSSEDataWithContext(context.Background(), func() {}, pr, 300*time.Millisecond, func(string) bool { return true })
+	}()
+	select {
+	case err := <-done:
+		if !errors.Is(err, ErrStreamIdle) {
+			t.Fatalf("err = %v, want ErrStreamIdle", err)
+		}
+		if elapsed := time.Since(start); elapsed < 250*time.Millisecond {
+			t.Fatalf("aborted after %s — the 30ms gap timeout must not apply before the first payload", elapsed)
+		}
+	case <-time.After(5 * time.Second):
+		t.Fatal("stream not aborted")
+	}
+}
+
+func TestResolveStreamGapTimeout(t *testing.T) {
+	t.Setenv("ZERO_STREAM_GAP_TIMEOUT", "")
+	if got := ResolveStreamGapTimeout(5 * time.Minute); got != DefaultStreamGapTimeout {
+		t.Fatalf("default gap = %s, want %s", got, DefaultStreamGapTimeout)
+	}
+	// Clamped: the gap never exceeds the idle window.
+	if got := ResolveStreamGapTimeout(time.Minute); got != time.Minute {
+		t.Fatalf("gap = %s, want clamp to the 1m idle", got)
+	}
+	// Disabled idle disables the gap too.
+	if got := ResolveStreamGapTimeout(0); got != 0 {
+		t.Fatalf("gap with idle off = %s, want 0", got)
+	}
+	// Env override wins; "off" falls back to the idle window for the whole stream.
+	t.Setenv("ZERO_STREAM_GAP_TIMEOUT", "45s")
+	if got := ResolveStreamGapTimeout(5 * time.Minute); got != 45*time.Second {
+		t.Fatalf("env gap = %s, want 45s", got)
+	}
+	t.Setenv("ZERO_STREAM_GAP_TIMEOUT", "off")
+	if got := ResolveStreamGapTimeout(5 * time.Minute); got != 5*time.Minute {
+		t.Fatalf("off gap = %s, want the idle window", got)
+	}
+}
diff --git a/internal/zeroruntime/helpers.go b/internal/zeroruntime/helpers.go
index 01e9188a1..3794c06ad 100644
--- a/internal/zeroruntime/helpers.go
+++ b/internal/zeroruntime/helpers.go
@@ -21,6 +21,12 @@ type CollectedStream struct {
 	// thinking blocks) that must be replayed on the next turn. Empty for providers
 	// or runs without extended thinking.
 	ReasoningBlocks []ReasoningBlock
+	// ReasoningEmitted reports that the stream carried ANY reasoning signal
+	// (live reasoning deltas or preserved blocks). It distinguishes "the model
+	// thought but produced no answer" (a behavioral empty turn) from "the
+	// provider returned a contentless stream" (a transport/backend fault worth
+	// retrying) — the two need different handling in the agent loop.
+	ReasoningEmitted bool
 }
 
 // Truncated reports whether the response ended for a non-normal reason (the
@@ -120,6 +126,7 @@ func CollectStreamWithOptions(ctx context.Context, events <-chan StreamEvent, op
 			// accumulate them regardless of type so they survive for replay.
 			if len(event.ReasoningBlocks) > 0 {
 				collected.ReasoningBlocks = append(collected.ReasoningBlocks, event.ReasoningBlocks...)
+				collected.ReasoningEmitted = true
 			}
 
 			switch event.Type {
@@ -132,6 +139,9 @@ func CollectStreamWithOptions(ctx context.Context, events <-chan StreamEvent, op
 					options.OnText(event.Content)
 				}
 			case StreamEventReasoning:
+				if event.Content != "" {
+					collected.ReasoningEmitted = true
+				}
 				if options.OnReasoning != nil {
 					options.OnReasoning(event.Content)
 				}