docs/superpowers/specs/2026-07-04-mi50-summary-cap-fallback-design.md

@@ -72,6 +72,18 @@ Fallback is per-`_summarize_text` call (i.e. per chunk), so the long-session
 chunk/merge path in `_summarize_transcript` is unaffected. The old `_RETRIES = 4`
 loop is replaced by this structure.
 
+### 3. Degenerate-output guard (added 2026-07-04)
+
+A wedged local backend — observed live when the MI50 overheated to 99°C junction —
+returns a single character repeated (`"?????"`) as a *successful* 200 response,
+which neither the timeout nor the exception path catches. So each `_call()`
+validates its output: `_looks_degenerate(text)` flags output (≥24 non-space chars)
+whose most-common non-whitespace character exceeds 50% of the text, and raises
+`DegenerateOutput` — which the retry/fallback loop treats exactly like any other
+failure (retry the primary, then fall back to cloud). Real gists are diverse prose
+(top char well under 20%), so the threshold won't false-positive; short outputs are
+exempt. If cloud *also* returns junk, it raises and stops — no infinite loop.
+
 ## Testing
 
 Unit (pytest, `tests/test_summary_fallback.py`), monkeypatching `llm.complete`:
@@ -95,6 +107,9 @@ llama.cpp slot shows bounded `n_decoded ≈ 768`.
 
 ## Out of scope (YAGNI)
 
+- The degenerate-output guard (§3) targets the *observed* failure — one char
+  repeated. It does not try to detect subtler degeneration (repeated phrases,
+  off-topic rambling); that's fuzzy and unmotivated until seen.
 - No change to `chat_call`/reflect/think or `config.summary_backend`.
 - No change to profile/era/narrative rebuild calls (separate, and not the loop
   culprit); can adopt the same `max_tokens` later if they show the same rambling.

lyra/summary.py

@@ -12,6 +12,7 @@ from __future__ import annotations
 import sys
 import threading
 import time
+from collections import Counter
 from concurrent.futures import ThreadPoolExecutor, as_completed
 
 from lyra import config, llm, logbus, memory
@@ -28,6 +29,25 @@ MI50_ATTEMPTS = 2
 # MI50; 150s is headroom but bails a hung call fast so fallback isn't slow.
 SUMMARY_TIMEOUT = 150
 
+# Degenerate-output guard. A wedged local model (e.g. an overheated GPU) returns
+# a single character repeated ("?????") as a *successful* 200, which no timeout or
+# exception catches — so validate the text and treat junk as a failure. Real gists
+# are diverse prose; flag output whose most-common non-space char dominates. Short
+# outputs are exempt (nothing meaningful to judge).
+_DEGENERATE_MIN_CHARS = 24
+_DEGENERATE_CHAR_RATIO = 0.5
+
+
+class DegenerateOutput(RuntimeError):
+    """A backend returned junk (e.g. one char repeated) as a successful response."""
+
+
+def _looks_degenerate(text: str) -> bool:
+    stripped = "".join(text.split())
+    if len(stripped) < _DEGENERATE_MIN_CHARS:
+        return False
+    return max(Counter(stripped).values()) / len(stripped) > _DEGENERATE_CHAR_RATIO
+
 # Re-summarize a session once it has accumulated this many new raw exchanges.
 SUMMARIZE_AFTER = 20
 # Transcript budget per LLM call; longer sessions are chunked + merged. Cloud has
@@ -72,8 +92,11 @@ def _summarize_text(text: str, backend: Backend) -> str:
     ]
 
     def _call(be: Backend) -> str:
-        return llm.complete(messages, backend=be,
+        out = llm.complete(messages, backend=be,
                            max_tokens=SUMMARY_MAX_TOKENS, timeout=SUMMARY_TIMEOUT)
+        if _looks_degenerate(out):
+            raise DegenerateOutput(f"{be} returned degenerate output ({len(out)} chars)")
+        return out
 
     # Try the primary backend a bounded number of times (each call fast-fails via
     # SUMMARY_TIMEOUT), with a short backoff for a transient blip / restarting GPU.

tests/test_summary_fallback.py

@@ -95,3 +95,48 @@ def test_happy_path_uses_primary_only(calls, monkeypatch):
 
     assert out == "mi50-gist"
     assert [c["backend"] for c in calls.recorded] == ["mi50"]  # no retries, no fallback
+
+
+# --- degenerate ("?" garbage) output guard: a wedged local model returns junk as
+# a successful 200, so treat it as a failure and fall back to cloud. ---
+
+def test_looks_degenerate_flags_repeated_char():
+    assert summary._looks_degenerate("?" * 60) is True
+    assert summary._looks_degenerate("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!") is True
+
+
+def test_looks_degenerate_passes_real_prose():
+    gist = ("Brian sat down at the Meadows 1/3 in seat 6 with two straddles active; "
+            "he tagged a seat-3 calling station and finished the session up 240.")
+    assert summary._looks_degenerate(gist) is False
+
+
+def test_looks_degenerate_ignores_short_output():
+    # Too short to judge — don't false-positive a terse-but-valid reply.
+    assert summary._looks_degenerate("ok") is False
+
+
+def test_degenerate_mi50_output_falls_back_to_cloud(calls, monkeypatch):
+    _set_key(monkeypatch)
+
+    def responder(backend):
+        if backend == "mi50":
+            return "?" * 200          # garbage-as-200, not an exception
+        return "a real cloud gist of the session, diverse and coherent."
+    calls.fake.responder = responder
+
+    out = summary._summarize_text("transcript", "mi50")
+
+    assert "cloud gist" in out
+    assert [c["backend"] for c in calls.recorded] == ["mi50", "mi50", "cloud"]
+
+
+def test_degenerate_cloud_output_raises_no_infinite_loop(calls, monkeypatch):
+    _set_key(monkeypatch)
+    calls.fake.responder = lambda backend: "?" * 200  # every backend returns garbage
+
+    with pytest.raises(Exception):
+        summary._summarize_text("t", "mi50")
+
+    # mi50 x2, then one cloud fallback that's also garbage -> give up, no loop.
+    assert [c["backend"] for c in calls.recorded] == ["mi50", "mi50", "cloud"]