Add max tokens to llm providers

Author: ericdallo

src/eca/db.clj

@@ -21,18 +21,22 @@
                   :output-token-cost (/ 8.0 one-million)}
             "gpt-4.1" {:tools true
                        :web-search true
+                       :max-output-tokens 32000
                        :input-token-cost (/ 2.0 one-million)
                        :output-token-cost (/ 8.0 one-million)}
             "claude-sonnet-4-0" {:tools true
                                  :web-search true
+                                 :max-output-tokens 8196
                                  :input-token-cost (/ 3.0 one-million)
                                  :output-token-cost (/ 15.0 one-million)}
             "claude-opus-4-0" {:tools true
                                :web-search true
+                               :max-output-tokens 8196
                                :input-token-cost (/ 15.0 one-million)
                                :output-token-cost (/ 75.0 one-million)}
             "claude-3-5-haiku-latest" {:tools true
                                        :web-search true
+                                       :max-output-tokens 4096
                                        :input-token-cost (/ 0.8 one-million)
                                        :output-token-cost (/ 4.0 one-million)}} ;; + ollama local models + custom provider models
    :mcp-clients {}})

src/eca/llm_api.clj

@@ -86,6 +86,7 @@
         tools (when (:tools model-config)
                 (mapv tool->llm-tool tools))
         web-search (:web-search model-config)
+        max-output-tokens (:max-output-tokens model-config)
         custom-providers (:customProviders config)
         custom-models (set (mapcat (fn [[k v]]
                                      (map #(str (name k) "/" %) (:models v)))
@@ -103,6 +104,7 @@
        {:model model
         :instructions instructions
         :user-prompt user-prompt
+        :max-output-tokens max-output-tokens
         :past-messages past-messages
         :tools tools
         :web-search web-search
@@ -117,6 +119,7 @@
        {:model model
         :instructions instructions
         :user-prompt user-prompt
+        :max-output-tokens max-output-tokens
         :past-messages past-messages
         :tools tools
         :web-search web-search
@@ -148,6 +151,7 @@
          {:model model
           :instructions instructions
           :user-prompt user-prompt
+          :max-output-tokens max-output-tokens
           :past-messages past-messages
           :web-search web-search
           :tools tools

src/eca/llm_providers/anthropic.clj

@@ -76,17 +76,16 @@
    #(assoc-in % [:content 0 :cache_control] {:type "ephemeral"})))
 
 (defn completion!
-  [{:keys [model user-prompt temperature instructions max-tokens
+  [{:keys [model user-prompt temperature instructions max-output-tokens
            api-url api-key past-messages tools web-search]
-    :or {max-tokens 8192
-         temperature 1.0}}
+    :or {temperature 1.0}}
    {:keys [on-message-received on-error on-prepare-tool-call on-tool-called]}]
   (let [messages (conj (past-messages->messages past-messages)
                        {:role "user" :content [{:type :text
                                                 :text user-prompt}]})
         body {:model model
               :messages (add-cache-to-last-message messages)
-              :max_tokens max-tokens
+              :max_tokens max-output-tokens
               :temperature temperature
               ;; TODO support :thinking
               :stream true

src/eca/llm_providers/openai.clj

@@ -4,6 +4,7 @@
    [clojure.java.io :as io]
    [eca.llm-util :as llm-util]
    [eca.logger :as logger]
+   [eca.shared :refer [assoc-some]]
    [hato.client :as http]))
 
 (set! *warn-on-reflection* true)
@@ -54,7 +55,8 @@
             msg))
         past-messages))
 
-(defn completion! [{:keys [model user-prompt instructions temperature api-key api-url past-messages tools web-search]
+(defn completion! [{:keys [model user-prompt instructions temperature api-key api-url
+                           max-output-tokens past-messages tools web-search]
                     :or {temperature 1.0}}
                    {:keys [on-message-received on-error on-prepare-tool-call on-tool-called on-reason]}]
   (let [input (conj (past-messages->input past-messages)
@@ -67,7 +69,8 @@
               :instructions instructions
               :temperature temperature
               :tools tools
-              :stream true}
+              :stream true
+              :max_completion_tokens max-output-tokens}
         mcp-call-by-item-id* (atom {})
         on-response-fn
         (fn handle-response [event data]