diff --git a/.env.example b/.env.example
index 4b5c96de..26746e0a 100644
--- a/.env.example
+++ b/.env.example
@@ -21,6 +21,8 @@ REMEMBERSTACK_SELFHOST_API_PORT=8000
 REMEMBERSTACK_OPENROUTER_API_KEY=replace-before-real-use
 # Optional exact OpenRouter embedding-provider slug; unset keeps automatic routing.
 # REMEMBERSTACK_OPENROUTER_EMBEDDING_PROVIDER=nebius
+# 32k covers reasoning plus content; the provider account cap remains the monetary boundary.
+REMEMBERSTACK_OPENROUTER_MAX_COMPLETION_TOKENS=32000
 # Optional global reasoning effort for every chat-generation call. Allowed:
 # none, minimal, low, medium, high, xhigh, max. Unset uses the model default;
 # "none" reduces extraction latency but can reduce adjudication quality, and
diff --git a/compose.yaml b/compose.yaml
index 037a6f32..4ebba275 100644
--- a/compose.yaml
+++ b/compose.yaml
@@ -17,6 +17,7 @@ x-app: &app
     REMEMBERSTACK_MINIO_SECRET_KEY: ${REMEMBERSTACK_MINIO_SECRET_KEY}
     REMEMBERSTACK_OPENROUTER_API_KEY: ${REMEMBERSTACK_OPENROUTER_API_KEY}
     REMEMBERSTACK_OPENROUTER_EMBEDDING_PROVIDER: ${REMEMBERSTACK_OPENROUTER_EMBEDDING_PROVIDER:-}
+    REMEMBERSTACK_OPENROUTER_MAX_COMPLETION_TOKENS: ${REMEMBERSTACK_OPENROUTER_MAX_COMPLETION_TOKENS:-}
     REMEMBERSTACK_OPENROUTER_REASONING_EFFORT: ${REMEMBERSTACK_OPENROUTER_REASONING_EFFORT:-}
     REMEMBERSTACK_OPENROUTER_REASONING_EFFORT_MAP: ${REMEMBERSTACK_OPENROUTER_REASONING_EFFORT_MAP:-}
     REMEMBERSTACK_SELFHOST_DEPLOYMENT_ID: ${REMEMBERSTACK_SELFHOST_DEPLOYMENT_ID}
diff --git a/plan/designs/locomo_benchmark_design.md b/plan/designs/locomo_benchmark_design.md
index 2a382266..236860e0 100644
--- a/plan/designs/locomo_benchmark_design.md
+++ b/plan/designs/locomo_benchmark_design.md
@@ -162,6 +162,12 @@ reliably honour it. Two distinct violations were observed in real runs against
   appending a sentence period inside the `arguments_json` string, after the
   closing brace. The agent loop parses the first complete JSON object and
   records the trailing text on the trace row (§7).
+- **2026-07-29 — `z-ai/glm-5.2` cap exhaustion and mitigation.** OpenRouter
+  requests sent no `max_tokens`, so production E1/E2 calls exhausted the
+  provider default on reasoning: `finish_reason='length'`, `content=null`, and
+  `reasoning_present=True`, or truncated JSON (`Unterminated string`), produced
+  dead-letters. The adapter now sends a 32,000-token reasoning-plus-content
+  budget by default; the provider account cap remains the monetary boundary.
 
 Neither is reproducible on demand: the same fact-label prompt succeeded on 24
 consecutive retries after failing once, and three prompt variants each returned
diff --git a/src/rememberstack/adapters/openrouter.py b/src/rememberstack/adapters/openrouter.py
index cae61b2e..2cc76b38 100644
--- a/src/rememberstack/adapters/openrouter.py
+++ b/src/rememberstack/adapters/openrouter.py
@@ -38,6 +38,7 @@
 _ALLOWED_REASONING_EFFORTS: Final[frozenset[str]] = frozenset(
     ("none", "minimal", "low", "medium", "high", "xhigh", "max")
 )
+_DEFAULT_MAX_COMPLETION_TOKENS: Final[int] = 32_000
 
 
 class StrictSchemaError(ValueError):
@@ -57,6 +58,15 @@ class OpenRouterSettings(BaseSettings):
     api_key: str = Field(min_length=1)
     base_url: str = Field(default="https://openrouter.ai/api/v1")
     timeout_s: float = Field(default=120.0, gt=0)
+    max_completion_tokens: int | None = Field(
+        default=_DEFAULT_MAX_COMPLETION_TOKENS, ge=1
+    )
+    """Combined reasoning-and-content budget for chat completions.
+
+    The 32k default gives reasoning models deliberate generation headroom;
+    the provider account cap remains the deployment's monetary boundary.
+    Explicit ``None`` omits ``max_tokens`` from the provider payload.
+    """
     embedding_provider: str | None = None
     reasoning_effort: ReasoningEffort | None = None
     reasoning_effort_map: dict[str, ReasoningEffort] | None = None
@@ -76,6 +86,14 @@ def normalize_optional_string(cls, value: object) -> object:
             return value
         return value.strip() or None
 
+    @field_validator("max_completion_tokens", mode="before")
+    @classmethod
+    def default_empty_max_completion_tokens(cls, value: object) -> object:
+        """Treat an empty env value as the deliberate 32k default."""
+        if isinstance(value, str) and not value.strip():
+            return _DEFAULT_MAX_COMPLETION_TOKENS
+        return value
+
     @field_validator("reasoning_effort_map", mode="before")
     @classmethod
     def parse_reasoning_effort_map(cls, value: object) -> object:
@@ -158,6 +176,8 @@ def generate(
         }
         if request.temperature is not None:
             payload["temperature"] = request.temperature
+        if self._settings.max_completion_tokens is not None:
+            payload["max_tokens"] = self._settings.max_completion_tokens
         effort = self._reasoning_effort_for(model=request.model)
         if effort is not None:
             payload["reasoning"] = {"effort": effort}
diff --git a/src/rememberstack/profiles/selfhost.py b/src/rememberstack/profiles/selfhost.py
index 9a2ef9b5..4d7e28c5 100644
--- a/src/rememberstack/profiles/selfhost.py
+++ b/src/rememberstack/profiles/selfhost.py
@@ -630,6 +630,11 @@ def _model_bindings() -> dict[str, str]:
         "p1_embedding": p1.embedding_model,
         "fact_label": p1.label_model,
         "openrouter_embedding_provider": openrouter.embedding_provider or "auto",
+        "openrouter_max_completion_tokens": (
+            str(openrouter.max_completion_tokens)
+            if openrouter.max_completion_tokens is not None
+            else "unset"
+        ),
         "openrouter_reasoning_effort": openrouter.reasoning_effort or "auto",
         # Canonical (sorted-key) form so the effective per-model effort policy
         # is part of measurement provenance, not hidden behind the global pin.
diff --git a/src/tests/adapters/test_openrouter.py b/src/tests/adapters/test_openrouter.py
index 96744acb..bca6488a 100644
--- a/src/tests/adapters/test_openrouter.py
+++ b/src/tests/adapters/test_openrouter.py
@@ -66,6 +66,67 @@ def test_usage_fails_closed_when_required_accounting_is_unusable(
         _usage(body=body, requested_model="requested/model", latency_ms=1)
 
 
+@pytest.mark.parametrize(
+    ("settings_override", "expected"),
+    (({}, 32_000), ({"max_completion_tokens": None}, None)),
+)
+def test_generation_forwards_configured_max_completion_tokens(
+    monkeypatch: pytest.MonkeyPatch,
+    settings_override: dict[str, object],
+    expected: int | None,
+) -> None:
+    """The 32k default is sent, while explicit None leaves provider defaults."""
+    provider = OpenRouterModelProvider(
+        settings=OpenRouterSettings.model_validate(
+            {"api_key": "test-key", **settings_override}
+        )
+    )
+    observed: dict[str, object] = {}
+
+    def post(*, path: str, payload: dict[str, object]) -> dict[str, object]:
+        observed.update(payload)
+        assert path == "/chat/completions"
+        return {
+            "model": "z-ai/glm-5.2",
+            "usage": {"prompt_tokens": 3, "completion_tokens": 1, "cost": "0"},
+            "choices": [{"message": {"content": '{"answer":"Prague"}'}}],
+        }
+
+    monkeypatch.setattr(provider, "_post", post)
+    try:
+        provider.generate(
+            request=ModelRequest(model="z-ai/glm-5.2", prompt="Where?"),
+            response_type=_Answer,
+        )
+    finally:
+        provider._client.close()
+
+    assert ("max_tokens" in observed) is (expected is not None)
+    assert observed.get("max_tokens") == expected
+
+
+def test_max_completion_tokens_empty_env_uses_default(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Compose's empty forwarded value keeps the deliberate 32k default."""
+    monkeypatch.setenv("REMEMBERSTACK_OPENROUTER_MAX_COMPLETION_TOKENS", "")
+
+    settings = OpenRouterSettings(api_key="test-key")
+
+    assert settings.max_completion_tokens == 32_000
+
+
+@pytest.mark.parametrize("invalid", (0, -1, "not-an-integer"))
+def test_max_completion_tokens_rejects_invalid_values(invalid: object) -> None:
+    """Zero, negative, and malformed caps cannot silently reach OpenRouter."""
+    with pytest.raises(ValidationError) as raised:
+        OpenRouterSettings.model_validate(
+            {"api_key": "test-key", "max_completion_tokens": invalid}
+        )
+
+    assert raised.value.errors()[0]["loc"] == ("max_completion_tokens",)
+
+
 @pytest.mark.parametrize(("temperature", "present"), ((None, False), (0.0, True)))
 def test_generation_forwards_temperature_only_when_declared(
     monkeypatch: pytest.MonkeyPatch, temperature: float | None, present: bool
diff --git a/src/tests/conftest.py b/src/tests/conftest.py
index b651b7c1..ba48d792 100644
--- a/src/tests/conftest.py
+++ b/src/tests/conftest.py
@@ -24,6 +24,7 @@
 #: Optional pins that alter provider request payloads when present.
 _LEAKY_OPTIONAL_PINS = (
     "REMEMBERSTACK_OPENROUTER_EMBEDDING_PROVIDER",
+    "REMEMBERSTACK_OPENROUTER_MAX_COMPLETION_TOKENS",
     "REMEMBERSTACK_OPENROUTER_REASONING_EFFORT",
     "REMEMBERSTACK_OPENROUTER_REASONING_EFFORT_MAP",
 )
diff --git a/src/tests/profiles/test_selfhost_profile.py b/src/tests/profiles/test_selfhost_profile.py
index bf0b8c8e..37a2add7 100644
--- a/src/tests/profiles/test_selfhost_profile.py
+++ b/src/tests/profiles/test_selfhost_profile.py
@@ -87,6 +87,26 @@ def test_model_bindings_report_embedding_provider_without_secrets(
     assert "test-key" not in bindings.values()
 
 
+@pytest.mark.parametrize(
+    ("configured", "reported"), (("64000", "64000"), ("", "32000"))
+)
+def test_compose_and_model_bindings_expose_max_completion_tokens(
+    monkeypatch: pytest.MonkeyPatch, configured: str, reported: str
+) -> None:
+    """Compose forwards the cap and readiness fingerprints its effective value."""
+    compose = (_ROOT / "compose.yaml").read_text(encoding="utf-8")
+    assert (
+        "REMEMBERSTACK_OPENROUTER_MAX_COMPLETION_TOKENS:"
+        " ${REMEMBERSTACK_OPENROUTER_MAX_COMPLETION_TOKENS:-}"
+    ) in compose
+    monkeypatch.setenv("REMEMBERSTACK_OPENROUTER_API_KEY", "test-key")
+    monkeypatch.setenv("REMEMBERSTACK_OPENROUTER_MAX_COMPLETION_TOKENS", configured)
+
+    bindings = _model_bindings()
+
+    assert bindings["openrouter_max_completion_tokens"] == reported
+
+
 @pytest.mark.parametrize(("configured", "reported"), (("none", "none"), ("", "auto")))
 def test_model_bindings_report_reasoning_effort(
     monkeypatch: pytest.MonkeyPatch, configured: str, reported: str