fix(buzz-agent): recover from 400-shaped image rejections; unbound benchmark agent rounds

Two failure modes from the tb21-glm52-crusoe-1 benchmark run wedged or
killed 13 of 89 trials without the model being at fault.

**Conversation poisoning on text-only endpoints.** Crusoe's serverless
`crusoeai/GLM-5.2-NVFP4` rejects any request whose history contains an
image with `400: ... is not a multimodal model`. The recovery path for
exactly this situation already exists (`AgentError::UnsupportedImageInput`
→ `replace_unsupported_images` strips the image blocks and continues the
turn with a text placeholder), but classification only matched
OpenRouter's 404 body ("no endpoints found that support image input") and
was only consulted on the 404 arms. The Crusoe 400 fell through to
terminal `AgentError::Llm`; the image stayed in history, every subsequent
call failed identically, buzz-acp rode its 10-retry ladder (~40 min), and
the trial idled to budget death. Measured: 8 trials, 12.7h aggregate idle.

Fix: extend `is_unsupported_image_input_error()` to also match the
verbatim "is not a multimodal model" body, and consult it on the 400 arms
of both the shared `post()` ladder and `openrouter_post()` (a BYOK/
passthrough upstream can surface the provider's own 400). The matcher
stays deliberately tight, same doctrine as `is_context_length_error`:
misclassifying a generic 400 as recoverable would mutate history for an
error that removing images cannot fix.

**Bounded agent rounds in benchmark trials.** The harness default
`DEFAULT_MAX_AGENT_ROUNDS = 32` ended solo trials mid-work when a turn
rotated (max_tokens rotation on thinking-heavy models burned rounds
fast). Benchmark trials already have a wall-clock budget as the real
limit, so the round cap only converts recoverable rotation into trial
death. Default is now 0 (unbounded — `BUZZ_AGENT_MAX_ROUNDS=0` is the
agent's own documented unbounded value); per-agent `budget.max_calls`
in manifests still overrides.

Tests: 400-body classification asserted through `complete()` on the
shared-post path and at the `openrouter_post` terminal, both proving
single-attempt (no retry of a deterministic capability rejection);
python tests updated for 0-is-legal with a negative arm at -1.

Signed-off-by: Eva <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz>
Co-authored-by: Eva <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz>

benchmarks/harbor-buzz-orchestra/src/harbor_buzz_orchestra/container_runtime.py

@@ -24,7 +24,7 @@ from .manifest import AgentClass, ExperimentManifest
 from .provisioning import AgentCredential, TrialHandle
 from .runtime import RuntimeResult
 
-DEFAULT_MAX_AGENT_ROUNDS = 32
+DEFAULT_MAX_AGENT_ROUNDS = 0  # 0 = unbounded (BUZZ_AGENT_MAX_ROUNDS=0); the trial budget is the clock
 # Container-side layout for the uploaded Buzz stack.
 REMOTE_ROOT = "/opt/buzz"
 REMOTE_BIN = f"{REMOTE_ROOT}/bin"
@@ -80,8 +80,8 @@ class BuzzContainerRuntime:
         readiness_timeout_seconds: float = 60.0,
         poll_seconds: float = 1.0,
     ) -> None:
-        if max_agent_rounds <= 0:
-            raise ValueError("max_agent_rounds must be positive")
+        if max_agent_rounds < 0:
+            raise ValueError("max_agent_rounds must be >= 0 (0 = unbounded)")
         if readiness_timeout_seconds <= 0:
             raise ValueError("readiness_timeout_seconds must be positive")
         self.logs_dir = Path(logs_dir)

benchmarks/harbor-buzz-orchestra/tests/test_container_runtime.py

@@ -247,7 +247,7 @@ async def test_forwarder_bridges_the_canonical_relay_address(tmp_path):
         rt._ws_authority("http://relay")
 
 
-@pytest.mark.parametrize(("configured", "expected"), [(None, "32"), (7, "7")])
+@pytest.mark.parametrize(("configured", "expected"), [(None, "0"), (7, "7")])
 async def test_launch_wires_the_desktop_environment(tmp_path, configured, expected):
     manifest = write_manifest(tmp_path)
     agent_class = manifest.roster[0]
@@ -290,9 +290,12 @@ async def test_launch_wires_the_desktop_environment(tmp_path, configured, expect
     )
 
 
-def test_runtime_rejects_unbounded_agent_rounds(tmp_path):
-    with pytest.raises(ValueError, match="positive"):
-        runtime(tmp_path, max_agent_rounds=0)
+def test_runtime_validates_construction_bounds(tmp_path):
+    # 0 is legal and means unbounded (BUZZ_AGENT_MAX_ROUNDS=0); the trial
+    # budget is the clock. Only negatives are rejected.
+    runtime(tmp_path, max_agent_rounds=0)
+    with pytest.raises(ValueError, match="unbounded"):
+        runtime(tmp_path, max_agent_rounds=-1)
     with pytest.raises(ValueError, match="positive"):
         runtime(tmp_path, readiness_timeout_seconds=0)
 

crates/buzz-agent/src/llm.rs

@@ -1932,9 +1932,24 @@ fn classify_body_read_error(
     }
 }
 
+/// Provider bodies that mean "this model cannot accept image input", the
+/// signal the agent loop uses to strip rejected images from history and
+/// continue the turn (see `replace_unsupported_images`).
+///
+/// Deliberately tight, same doctrine as [`is_context_length_error`]: each
+/// phrase is a verbatim capability rejection observed live. Misclassifying a
+/// generic 400 as recoverable would mutate history for an error that removing
+/// images cannot fix.
 fn is_unsupported_image_input_error(body: &str) -> bool {
-    body.to_ascii_lowercase()
-        .contains("no endpoints found that support image input")
+    let b = body.to_ascii_lowercase();
+    // OpenRouter 404: no provider endpoint accepts images for this model.
+    b.contains("no endpoints found that support image input")
+        // OpenAI-compatible 400 from text-only single-model deployments,
+        // e.g. Crusoe serverless GLM: `"crusoeai/GLM-5.2-NVFP4 is not a
+        // multimodal model"`. Without this arm the 400 is terminal, the image
+        // stays in history, and every subsequent request in the session fails
+        // identically — the turn wedges until the harness/user gives up.
+        || b.contains("is not a multimodal model")
 }
 
 /// Build the terminal `AgentError::Llm` for a `post()` exit that has given up
@@ -2129,6 +2144,13 @@ where
                     "{status}: {body}"
                 ))));
             }
+            // Image-capability rejection is equally recoverable and equally
+            // deterministic: a text-only deployment 400s the same request
+            // forever. Typed here (not just on the 404 arm) because
+            // OpenAI-compatible providers report it as a 400.
+            if status == 400 && is_unsupported_image_input_error(&body) {
+                return Err(PostError::Agent(AgentError::UnsupportedImageInput(body)));
+            }
             return Err(PostError::Agent(AgentError::Llm(format!(
                 "{status}: {body}"
             ))));
@@ -2535,6 +2557,12 @@ async fn openrouter_post(
             if status == 400 && is_context_length_error(&body) {
                 return Err(AgentError::LlmContextExceeded(format!("{status}: {body}")));
             }
+            // Same 400-shaped image rejection as the shared `post()` terminal:
+            // OpenRouter normally reports this as a 404 (handled above), but a
+            // BYOK/passthrough upstream can surface the provider's own 400.
+            if status == 400 && is_unsupported_image_input_error(&body) {
+                return Err(AgentError::UnsupportedImageInput(body));
+            }
             return Err(AgentError::Llm(format!("{status}: {body}")));
         }
         if let Some(len) = resp.content_length() {
@@ -7544,6 +7572,73 @@ mod tests {
         );
     }
 
+    /// OpenAI-compatible text-only deployments report the image rejection as a
+    /// 400, not OpenRouter's 404 — Crusoe serverless GLM answers
+    /// `"crusoeai/GLM-5.2-NVFP4 is not a multimodal model"` to every request
+    /// whose history contains an image. Before the 400 arm existed, this fell
+    /// through to terminal `AgentError::Llm`: the image stayed in history and
+    /// every later call in the session failed identically (measured live:
+    /// 8 wedged benchmark trials, 40 min of doomed retries each). Asserted
+    /// through `complete()` so the arm's return path into the convergence
+    /// mapper is covered, same doctrine as the context-400 tests above.
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    async fn openai_400_unsupported_image_is_typed_through_complete() {
+        let (base_url, captured) = spawn_sequence_stub(vec![StubHttpResponse {
+            status: 400,
+            body: json!({"error":{"message":"crusoeai/GLM-5.2-NVFP4 is not a multimodal model","type":"invalid_request_error"}}),
+        }])
+        .await;
+        let mut c = cfg(Provider::OpenAi);
+        c.base_url = base_url;
+        let llm = Llm::new(&c).unwrap();
+        let err = complete_model(&llm, &c, "gpt-probe-model")
+            .await
+            .unwrap_err();
+        assert!(
+            matches!(&err, AgentError::UnsupportedImageInput(s) if s.contains("not a multimodal model")),
+            "a text-only deployment's 400 must reach the history-recovery path: got {err:?}"
+        );
+        assert_eq!(
+            captured.lock().await.len(),
+            1,
+            "a deterministic capability rejection must not be retried"
+        );
+    }
+
+    /// Same 400-shaped rejection at the OpenRouter terminal, which has its own
+    /// status ladder: a BYOK/passthrough upstream can surface the provider's
+    /// own 400 body instead of OpenRouter's 404 routing error.
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    async fn openrouter_post_400_unsupported_image_is_typed_and_not_retried() {
+        let (url, _captured, attempts) = spawn_openrouter_stub(vec![CannedResponse::new(
+            400,
+            r#"{"error":{"message":"crusoeai/GLM-5.2-NVFP4 is not a multimodal model"}}"#,
+        )])
+        .await;
+        let http = Client::builder()
+            .timeout(Duration::from_secs(5))
+            .build()
+            .unwrap();
+        let err = openrouter_post(
+            &http,
+            &format!("{url}/x"),
+            &json!({}),
+            "key",
+            Duration::from_secs(5),
+        )
+        .await
+        .unwrap_err();
+        assert!(
+            matches!(&err, AgentError::UnsupportedImageInput(s) if s.contains("not a multimodal model")),
+            "image rejection must reach the history-recovery path: got {err:?}"
+        );
+        assert_eq!(
+            attempts.load(std::sync::atomic::Ordering::SeqCst),
+            1,
+            "a deterministic capability rejection must not be retried"
+        );
+    }
+
     /// Every other 404 still maps to `LlmModelNotFound`, including one that
     /// shares the `No endpoints found` prefix but is about the model rather than
     /// the parameters — the discriminator is narrow enough that a genuinely