Honor configured provider output limits

CHANGELOG.md

@@ -3,6 +3,7 @@
 ## Unreleased
 
 - Truncate oversized failed tool output, preserving full logs on disk instead of overflowing the model context.
+- Honor model output limits in OpenAI-compatible Chat Completions and Ollama requests while preserving `extraPayload` overrides.
 
 ## 0.147.3
 

src/eca/llm_api.clj

@@ -397,6 +397,7 @@
           :user-messages user-messages
           :past-messages past-messages
           :tools tools
+          :max-output-tokens max-output-tokens
           :extra-payload extra-payload
           :extra-headers extra-headers}
          callbacks)

src/eca/llm_providers/ollama.clj

@@ -138,19 +138,22 @@
                      }))))
         messages))
 
-(defn chat! [{:keys [model user-messages reason? instructions api-url past-messages tools extra-headers extra-payload]}
+(defn chat! [{:keys [model user-messages reason? instructions api-url past-messages tools max-output-tokens
+                     extra-headers extra-payload]}
              {:keys [on-message-received on-error on-prepare-tool-call on-tools-called
                      on-reason] :as callbacks}]
   (let [messages (concat
                   (normalize-messages (concat [{:role "system" :content instructions}] past-messages))
                   (normalize-messages user-messages))
         stream? (boolean callbacks)
         body (deep-merge
-              {:model model
-               :messages messages
-               :think reason?
-               :tools (->tools tools)
-               :stream stream?}
+              (cond-> {:model model
+                       :messages messages
+                       :think reason?
+                       :tools (->tools tools)
+                       :stream stream?}
+                max-output-tokens
+                (assoc :options {:num_predict max-output-tokens}))
               extra-payload)
         url (join-api-url api-url chat-path)
         tool-calls* (atom {})

src/eca/llm_providers/openai_chat.clj

@@ -523,7 +523,7 @@
    and message normalization. Supports both single and parallel tool execution.
    Compatible with OpenRouter and other OpenAI-compatible providers."
   [{:keys [model user-messages instructions temperature api-key api-url url-relative-path
-           past-messages tools extra-payload extra-headers supports-image?
+           max-output-tokens past-messages tools extra-payload extra-headers supports-image?
            think-tag-start think-tag-end reasoning-history http-client]}
    {:keys [on-message-received on-error on-prepare-tool-call on-tools-called on-reason on-usage-updated] :as callbacks}]
   (let [think-tag-start (or think-tag-start "<think>")
@@ -541,7 +541,7 @@
                {:model model
                 :messages messages
                 :stream stream?
-                :max_completion_tokens 32000}
+                :max_completion_tokens (or max-output-tokens 32000)}
                :temperature temperature
                :tools (when (seq tools) (->tools tools))
                ;; Required by the OpenAI streaming spec to receive a final

test/eca/llm_api_test.clj

@@ -5,6 +5,7 @@
    [eca.config :as config]
    [eca.llm-api :as llm-api]
    [eca.llm-providers.anthropic :as llm-providers.anthropic]
+   [eca.llm-providers.ollama :as llm-providers.ollama]
    [eca.llm-providers.openai :as llm-providers.openai]
    [eca.llm-providers.openai-chat :as llm-providers.openai-chat]
    [eca.secrets :as secrets]
@@ -220,6 +221,25 @@
       (is (= :openai-chat
              (:api (llm-api/provider->api-handler "github-copilot" "unknown-model" config)))))))
 
+(deftest prompt-forwards-max-output-tokens-to-ollama-test
+  (let [captured* (atom nil)]
+    (with-redefs [llm-providers.ollama/chat!
+                  (fn [opts _callbacks]
+                    (reset! captured* opts)
+                    {:output-text "ok"})]
+      (#'eca.llm-api/prompt!
+       {:provider "ollama"
+        :model "test-model"
+        :model-capabilities {:tools false
+                             :reason? false
+                             :model-name "test-model"
+                             :max-output-tokens 512}
+        :user-messages [{:role "user" :content [{:type :text :text "hello"}]}]
+        :past-messages []
+        :config {:providers {"ollama" {:url "http://localhost:11434"}}}
+        :sync? true}))
+    (is (= 512 (:max-output-tokens @captured*)))))
+
 (deftest prompt-test
   (testing "Custom OpenAI provider behavior and proper passing of httpClient options to the Hato client"
     (let [req* (atom nil)]

test/eca/llm_providers/ollama_test.clj

@@ -80,6 +80,28 @@
 
           (is (= {:output-text "Hello world"} result)))))))
 
+(deftest chat-request-enforces-max-output-tokens-test
+  (let [requests* (atom [])]
+    (with-client-proxied {}
+      (fn handler [req]
+        (swap! requests* conj req)
+        {:status 200
+         :body {:message {:content "ok"}}})
+      (let [base-opts {:model "test-model"
+                       :instructions "System prompt"
+                       :user-messages [{:role "user" :content "hello"}]
+                       :past-messages []
+                       :tools nil
+                       :api-url "http://localhost:1"
+                       :max-output-tokens 512}]
+        (llm-providers.ollama/chat! base-opts nil)
+        (llm-providers.ollama/chat!
+         (assoc base-opts :extra-payload {:options {:num_predict 99}}) nil)))
+    (let [bodies (mapv #(json/parse-string (:body %) true) @requests*)]
+      (is (= 512 (get-in bodies [0 :options :num_predict])))
+      (is (= 99 (get-in bodies [1 :options :num_predict]))
+          "Configured extraPayload must remain the final override"))))
+
 (deftest ->normalize-messages-test
   (testing "no previous history"
     (is (match?

test/eca/llm_providers/openai_chat_test.clj

@@ -51,6 +51,28 @@
           (is (= {:output-text "Hello there!"}
                  (select-keys response [:output-text]))))))))
 
+(deftest chat-request-enforces-max-output-tokens-test
+  (let [requests* (atom [])]
+    (with-client-proxied {}
+      (fn handler [req]
+        (swap! requests* conj req)
+        {:status 200
+         :body {:choices [{:message {:content "ok"}}]}})
+      (let [base-opts {:model "test-model"
+                       :instructions "System prompt"
+                       :user-messages [{:role "user" :content "hello"}]
+                       :past-messages []
+                       :tools nil
+                       :api-key "fake-key"
+                       :api-url "http://localhost:1"
+                       :max-output-tokens 512}]
+        (llm-providers.openai-chat/chat-completion! base-opts nil)
+        (llm-providers.openai-chat/chat-completion!
+         (assoc base-opts :extra-payload {:max_completion_tokens 99}) nil)))
+    (is (= 512 (get-in @requests* [0 :body :max_completion_tokens])))
+    (is (= 99 (get-in @requests* [1 :body :max_completion_tokens]))
+        "Configured extraPayload must remain the final override")))
+
 (deftest normalize-messages-test
   (testing "With tool_call history - assistant text and tool calls are merged"
     (is (match?