Summarize: LAN primary with local-Ollama rescue + in-app failure guidance

- Engine: when the primary summarizer is out of retries (or misconfigured),
  run_summarize now finishes the job on the configured rescue provider
  (SHONAR_LLM_FALLBACK_*), tags the summary with the provider that wrote it,
  and stores a human note in job.error; success clears stale notes.
- App (auto/lan): passes Ollama as the rescue provider when it is up.
- Detail screen: shows the rescue-swap note in plain words, a red
  'Summary failed' line with Settings -> Summarizer fix instructions and a
  Retry summary button on hard failure.
- Settings copy explains the fallback. 3 new pytest cases (14/14 pass);
  live E2E on 2026-09-18: sarcastic summary v6 via LAN on attempt 2.

app/src/main/kotlin/com/shonar/desktop/DesktopState.kt

@@ -358,6 +358,15 @@ class DesktopState(private val appDir: File = defaultAppDir()) {
                     env["SHONAR_LLM_BASE_URL"] = LAN_LLM_BASE_URL
                     env["SHONAR_LLM_MODEL"] = LAN_LLM_MODEL
                     env["SHONAR_LLM_API_KEY"] = lanKey
+                    // Rescue summarizer: the LAN GPU intermittently replies
+                    // with garbage; when a summarize job exhausts its retries
+                    // the engine finishes it on local Ollama instead of
+                    // failing, and the detail screen names the swap.
+                    if (ollamaUp()) {
+                        env["SHONAR_LLM_FALLBACK_PROVIDER"] = "ollama"
+                        env["SHONAR_LLM_FALLBACK_BASE_URL"] = "http://127.0.0.1:11434"
+                        env["SHONAR_LLM_FALLBACK_MODEL"] = ollamaModel()
+                    }
                 } else if (ollamaUp()) {
                     env["SHONAR_LLM_PROVIDER"] = "ollama"
                     env["SHONAR_LLM_BASE_URL"] = "http://127.0.0.1:11434"

app/src/main/kotlin/com/shonar/desktop/Screens.kt

@@ -796,6 +796,35 @@ fun DetailScreen(state: DesktopState) {
                 )
             }
         }
+        // Rescue-swap note: the engine succeeded but had to switch
+        // summarizers mid-job. Tell the user in plain words, not the raw
+        // provider ids.
+        sJob?.takeIf { it.status == "succeeded" && it.error != null }?.let {
+            val note = it.error!!
+                .replace("openai_compat", "LAN server (H200)")
+                .replace("ollama", "this laptop (Ollama)")
+            Text(note, style = MaterialTheme.typography.bodySmall,
+                color = MaterialTheme.colorScheme.tertiary)
+            Text("You can change the summarizer in Settings → Summarizer.",
+                style = MaterialTheme.typography.bodySmall,
+                color = MaterialTheme.colorScheme.onSurfaceVariant)
+        }
+        // Summarize failed even after any rescue: say what broke and how to
+        // fix it from inside the app.
+        sJob?.takeIf { it.status == "failed" }?.let {
+            Text("Summary failed: " + (it.error ?: "unknown error")
+                    .replace("openai_compat", "LAN server (H200)")
+                    .replace("ollama", "this laptop (Ollama)"),
+                style = MaterialTheme.typography.bodySmall,
+                color = MaterialTheme.colorScheme.error)
+            Text("Fix: open Settings → Summarizer and pick \"This laptop " +
+                "(Ollama)\", then press Retry summary below.",
+                style = MaterialTheme.typography.bodySmall,
+                color = MaterialTheme.colorScheme.onSurfaceVariant)
+            if (detail.busy == null) {
+                Button({ state.summarize() }) { Text("Retry summary") }
+            }
+        }
         detail.error?.let {
             Text(it, color = MaterialTheme.colorScheme.error)
             if (tJob?.status == "failed") {
@@ -1167,11 +1196,14 @@ fun SettingsScreen(state: DesktopState) {
         }
         Text(when (summarizer) {
                 "lan" -> "Best quality; shares the H200 with other clients — " +
-                    "summaries may wait in its queue before starting."
+                    "summaries may wait in its queue before starting. If it " +
+                    "keeps failing, the job finishes on this laptop's Ollama " +
+                    "and this screen says so."
                 "local" -> "Private and never queued behind other users; " +
                     "smaller model, so summaries are plainer."
-                else -> "Uses the LAN server when its key is available, " +
-                    "otherwise this laptop's Ollama."
+                else -> "Uses the LAN server when its key is available " +
+                    "(falling back to this laptop's Ollama on repeated " +
+                    "failures), otherwise Ollama straight away."
             },
             style = MaterialTheme.typography.bodySmall,
             color = MaterialTheme.colorScheme.onSurfaceVariant)

backend/shonar/core/config.py

@@ -90,6 +90,14 @@ class Settings(BaseSettings):
     llm_base_url: str = ""
     llm_api_key: str = Field(default="", repr=False)
 
+    # Rescue summarizer tried when the PRIMARY one fails mid-job (desktop:
+    # LAN GPU server primary, local Ollama fallback). "none" disables the
+    # fallback; the primary's error then stands as today.
+    llm_fallback_provider: str = "none"  # none | openai_compat | ollama
+    llm_fallback_model: str = ""
+    llm_fallback_base_url: str = ""
+    llm_fallback_api_key: str = Field(default="", repr=False)
+
     # Search needs no setting: services/search.py picks its path by SQL
     # dialect (SQLite substring scan / Postgres tsvector). A Meilisearch or
     # OpenSearch backend would replace that module, not add a config knob.

backend/shonar/services/ai/__init__.py

@@ -160,3 +160,32 @@ def get_llm_provider(settings: Settings) -> LlmProvider | None:
             model=settings.llm_model,
         )
     raise ProviderConfigError(f"Unknown LLM provider: {settings.llm_provider!r}")
+
+
+def get_llm_fallback_provider(settings: Settings) -> LlmProvider | None:
+    """Rescue provider tried when the primary summarize fails mid-job.
+
+    None means "no fallback configured" — the primary's error stands.
+    Same kinds as get_llm_provider, built from the llm_fallback_* settings.
+    """
+    kind = settings.llm_fallback_provider.strip().lower()
+    if kind in ("", "none"):
+        return None
+    if kind == "openai_compat":
+        from shonar.services.ai.openai_compat import OpenAICompatProvider
+
+        return OpenAICompatProvider(
+            base_url=settings.llm_fallback_base_url,
+            model=settings.llm_fallback_model,
+            api_key=settings.llm_fallback_api_key,
+        )
+    if kind == "ollama":
+        from shonar.services.ai.ollama import OllamaProvider
+
+        return OllamaProvider(
+            base_url=settings.llm_fallback_base_url,
+            model=settings.llm_fallback_model,
+        )
+    raise ProviderConfigError(
+        f"Unknown LLM fallback provider: {settings.llm_fallback_provider!r}"
+    )

backend/shonar/services/processing.py

@@ -39,7 +39,9 @@ from shonar.db.models import (
 )
 from shonar.services.ai import (
     AIError,
+    ProviderConfigError,
     ProviderTransientError,
+    get_llm_fallback_provider,
     get_llm_provider,
     get_transcription_provider,
 )
@@ -479,19 +481,48 @@ async def run_summarize(ctx: dict, recording_id: str) -> None:
         job.stage = "summarizing"
         rec.processing_status = ProcessingStatus.processing
         await session.commit()  # visible before the long LLM call
+        fallback_note: str | None = None
+        result_provider = provider.name  # overwritten only on the rescue path
         try:
             result = await provider.summarize(
                 text, title=rec.title, tone=job.tone,
                 on_progress=_async_progress_writer(job.id))
         except AIError as e:
+            # Transient failures keep their normal retry budget first (the
+            # LAN GPU often recovers on try 2); the rescue summarizer runs
+            # when the primary is OUT of tries or misconfigured, and only
+            # if one is configured. The summary then comes from the
+            # fallback and job.error carries a note for the UI.
+            retryable = (isinstance(e, ProviderTransientError)
+                         and _job_try(ctx) < MAX_TRIES)
+            # (result, provider_name, note) when the rescue succeeds.
+            rescue: tuple[object, str, str] | None = None
+            if not retryable:
+                fallback = get_llm_fallback_provider(settings)
+                if fallback is not None:
+                    try:
+                        fb_result = await fallback.summarize(
+                            text, title=rec.title, tone=job.tone,
+                            on_progress=_async_progress_writer(job.id))
+                    except AIError as fe:
+                        logger.warning(
+                            "summarize fallback (%s) also failed: %s",
+                            fallback.name, fe)
+                    else:
+                        rescue = (fb_result, fallback.name, (
+                            f'primary "{provider.name}" failed ({e}). '
+                            f'This summary was written by "{fallback.name}".'))
+            if rescue is None:
                 await _fail(session, rec, job, str(e), ctx, e)
                 await session.commit()
                 return
-        await store_summary(session, rec, result.to_dict(), provider.name,
+            result, result_provider, fallback_note = rescue
+        await store_summary(session, rec, result.to_dict(), result_provider,
                             result.model, tone=job.tone)
         job.status = JobStatus.succeeded
         job.stage = None
         job.progress = 100
+        job.error = fallback_note  # None on the clean path: clears stale notes
         job.finished_at = utcnow()
         rec.processing_status = ProcessingStatus.completed
         rec.processing_error = None

backend/tests/test_ai_pipeline.py

@@ -110,11 +110,13 @@ class FakeLlm:
 _UNSET = object()
 
 
-def use_fakes(monkeypatch, stt=_UNSET, llm=_UNSET):
+def use_fakes(monkeypatch, stt=_UNSET, llm=_UNSET, fallback=None):
     tprov = FakeTranscriber() if stt is _UNSET else stt
     lprov = FakeLlm() if llm is _UNSET else llm
     monkeypatch.setattr(processing, "get_transcription_provider", lambda settings: tprov)
     monkeypatch.setattr(processing, "get_llm_provider", lambda settings: lprov)
+    monkeypatch.setattr(processing, "get_llm_fallback_provider",
+                        lambda settings: fallback)
 
 
 async def jobs_for(recording_id):
@@ -409,3 +411,56 @@ async def test_ai_endpoints_enforce_ownership(client, monkeypatch):
     for path in ("transcript", "summary", "jobs"):
         r = await client.get(f"/api/v1/recordings/{rec['id']}/{path}", headers=h)
         assert r.status_code == 404, path
+
+
+# --- summarize fallback to the rescue summarizer --------------------------------
+
+async def test_summarize_fallback_rescues_after_primary_exhausted(client, monkeypatch):
+    primary = FakeLlm(fail=ProviderTransientError("LLM reply was not JSON."))
+    fb = FakeLlm()
+    fb.name = "fake-ollama"
+    use_fakes(monkeypatch, FakeTranscriber(), primary, fallback=fb)
+    token = await user_tokens(client, email="m7-fb1@example.com")
+    rec = await upload_recording(client, token, client_id="m7-fb-1")
+    await processing.run_transcribe({}, rec["id"])
+    # Last try: primary out of budget -> rescue runs and the note lands in job.error.
+    await processing.run_summarize({"job_try": 3}, rec["id"])
+    job = (await jobs_for(rec["id"]))[JobType.summarize]
+    assert job.status == JobStatus.succeeded
+    assert "fake-llm" in (job.error or "") and "fake-ollama" in (job.error or "")
+    h = {"Authorization": f"Bearer {token}"}
+    s = await client.get(f"/api/v1/recordings/{rec['id']}/summary", headers=h)
+    assert s.status_code == 200
+    assert s.json()["provider"] == "fake-ollama"
+    assert fb.seen == ["hello world from the meeting"]
+
+
+async def test_summarize_fallback_also_fails_marks_job_failed(client, monkeypatch):
+    err = ProviderTransientError("LLM reply was not JSON.")
+    fb = FakeLlm(fail=ProviderTransientError("ollama down"))
+    fb.name = "fake-ollama"
+    use_fakes(monkeypatch, FakeTranscriber(), FakeLlm(fail=err), fallback=fb)
+    token = await user_tokens(client, email="m7-fb2@example.com")
+    rec = await upload_recording(client, token, client_id="m7-fb-2")
+    await processing.run_transcribe({}, rec["id"])
+    await processing.run_summarize({"job_try": 3}, rec["id"])
+    job = (await jobs_for(rec["id"]))[JobType.summarize]
+    assert job.status == JobStatus.failed
+    assert "not JSON" in job.error  # the primary's error is what we surface
+    assert fb.seen == ["hello world from the meeting"]  # fallback was tried
+
+
+async def test_summarize_keeps_retry_budget_before_fallback(client, monkeypatch):
+    # Try 1 of 3 transient: normal arq retry wins, fallback stays untouched.
+    fb = FakeLlm()
+    fb.name = "fake-ollama"
+    use_fakes(monkeypatch, FakeTranscriber(),
+              FakeLlm(fail=ProviderTransientError("boom")), fallback=fb)
+    token = await user_tokens(client, email="m7-fb3@example.com")
+    rec = await upload_recording(client, token, client_id="m7-fb-3")
+    await processing.run_transcribe({}, rec["id"])
+    with pytest.raises(ProviderTransientError):
+        await processing.run_summarize({"job_try": 1}, rec["id"])
+    job = (await jobs_for(rec["id"]))[JobType.summarize]
+    assert job.status == JobStatus.queued
+    assert fb.seen == []