diff --git a/CHANGELOG.md b/CHANGELOG.md index 56940783..4ae57c68 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -2,6 +2,8 @@ ## Unreleased +- (OpenAI Chat) - Configurable reasoning history via `keepHistoryReasoning` (model-level, default: prune) + ## 0.92.1 - Add `x-llm-application-name: eca` to prompt requests, useful to track and get metrics when using LLM gateways. diff --git a/docs/configuration.md b/docs/configuration.md index bf55c215..fdd15611 100644 --- a/docs/configuration.md +++ b/docs/configuration.md @@ -673,7 +673,8 @@ To configure, add your OTLP collector config via `:otlp` map following [otlp aut thinkTagEnd?: string; models: {[key: string]: { modelName?: string; - extraPayload?: {[key: string]: any} + extraPayload?: {[key: string]: any}; + keepHistoryReasoning?: boolean; }}; }}; defaultModel?: string; diff --git a/docs/models.md b/docs/models.md index cc8da224..de142768 100644 --- a/docs/models.md +++ b/docs/models.md @@ -61,19 +61,20 @@ You just need to add your provider to `providers` and make sure add the required Schema: -| Option | Type | Description | Required | -|-------------------------------|--------|--------------------------------------------------------------------------------------------------------------|----------| -| `api` | string | The API schema to use (`"openai-responses"`, `"openai-chat"`, or `"anthropic"`) | Yes | -| `url` | string | API URL (with support for env like `${env:MY_URL}`) | No* | -| `key` | string | API key (with support for `${env:MY_KEY}` or `{netrc:api.my-provider.com}` | No* | -| `completionUrlRelativePath` | string | Optional override for the completion endpoint path (see defaults below and examples like Azure) | No | -| `thinkTagStart` | string | Optional override the think start tag tag for openai-chat (Default: "") api | No | -| `thinkTagEnd` | string | Optional override the think end tag for openai-chat (Default: "") api | No | -| `httpClient` | map | Allow customize the http-client for this provider requests, like changing http version | No | -| `models` | map | Key: model name, value: its config | Yes | -| `models extraPayload` | map | Extra payload sent in body to LLM | No | -| `models modelName` | string | Override model name, useful to have multiple models with different configs and names that use same LLM model | No | -| `fetchModels` | boolean | Enable automatic model discovery from `/models` endpoint (OpenAI-compatible providers) | No | +| Option | Type | Description | Required | +|---------------------------------------|---------|--------------------------------------------------------------------------------------------------------------|----------| +| `api` | string | The API schema to use (`"openai-responses"`, `"openai-chat"`, or `"anthropic"`) | Yes | +| `url` | string | API URL (with support for env like `${env:MY_URL}`) | No* | +| `key` | string | API key (with support for `${env:MY_KEY}` or `{netrc:api.my-provider.com}` | No* | +| `completionUrlRelativePath` | string | Optional override for the completion endpoint path (see defaults below and examples like Azure) | No | +| `thinkTagStart` | string | Optional override the think start tag tag for openai-chat (Default: "") api | No | +| `thinkTagEnd` | string | Optional override the think end tag for openai-chat (Default: "") api | No | +| `httpClient` | map | Allow customize the http-client for this provider requests, like changing http version | No | +| `models` | map | Key: model name, value: its config | Yes | +| `models extraPayload` | map | Extra payload sent in body to LLM | No | +| `models modelName` | string | Override model name, useful to have multiple models with different configs and names that use same LLM model | No | +| `models keepHistoryReasoning` | boolean | Keep `reason` messages in conversation history. Default: `false` | No | +| `fetchModels` | boolean | Enable automatic model discovery from `/models` endpoint (OpenAI-compatible providers) | No | _* url and key will be searched as envs `_API_URL` and `_API_KEY`, they require the env to be found or config to work._ @@ -120,6 +121,33 @@ Examples: This way both will use gpt-5 model but one will override the reasoning to be high instead of the default. +=== "History reasoning" + `keepHistoryReasoning` - Determines whether the model's internal reasoning chain is persisted in the conversation history for subsequent turns. + + - **Standard Behavior**: Most models expect reasoning blocks (e.g., `` tags or `reasoning_content`) to be removed in subsequent requests to save tokens and avoid bias. + - **Usage**: Enable this for models that explicitly support "preserved thinking," or if you want to experiment with letting the model see its previous thought process (with XML-based reasoning). + - **Example**: See [GLM-4.7 with Preserved thinking](https://docs.z.ai/guides/capabilities/thinking-mode#preserved-thinking). + + ```javascript title="~/.config/eca/config.json" + { + "providers": { + "z-ai": { + "api": "openai-chat", + "url": "https://api.z.ai/api/paas/v4/", + "key": "your-api-key", + "models": { + "GLM-4.7": { + "keepHistoryReasoning": true, // Preserves reasoning + "extraPayload": {"clear_thinking": false} // Preserved thinking (see https://docs.z.ai/guides/capabilities/thinking-mode#preserved-thinking) + } + } + } + } + } + ``` + + Default: `false`. + === "Dynamic model discovery" For OpenAI-compatible providers, set `fetchModels: true` to automatically discover available models: @@ -211,7 +239,7 @@ Notes: 3. Type the chosen method 4. Authenticate in your browser, copy the code. 5. Paste and send the code and done! - + === "Codex / Openai" 1. Login to Openai via the chat command `/login`. diff --git a/integration-test/integration/chat/github_copilot_test.clj b/integration-test/integration/chat/github_copilot_test.clj index f8ba132c..a85d26f5 100644 --- a/integration-test/integration/chat/github_copilot_test.clj +++ b/integration-test/integration/chat/github_copilot_test.clj @@ -168,7 +168,7 @@ (match-content chat-id "system" {:type "progress" :state "finished"}) (is (match? {:input [{:role "user" :content [{:type "input_text" :text "hello!"}]} - {:role "assistant" :content [{:type "output_text" :text "I should say hello\nhello there!"}]} + {:role "assistant" :content [{:type "output_text" :text "hello there!"}]} {:role "user" :content [{:type "input_text" :text "how are you?"}]}] :instructions (m/pred string?)} (llm.mocks/get-req-body :reasoning-1))))))) diff --git a/integration-test/integration/chat/google_test.clj b/integration-test/integration/chat/google_test.clj index 5f34d341..30222616 100644 --- a/integration-test/integration/chat/google_test.clj +++ b/integration-test/integration/chat/google_test.clj @@ -167,7 +167,7 @@ (match-content chat-id "system" {:type "progress" :state "finished"}) (is (match? {:input [{:role "user" :content [{:type "input_text" :text "hello!"}]} - {:role "assistant" :content [{:type "output_text" :text "I should say hello\nhello there!"}]} + {:role "assistant" :content [{:type "output_text" :text "hello there!"}]} {:role "user" :content [{:type "input_text" :text "how are you?"}]}] :instructions (m/pred string?)} (llm.mocks/get-req-body :reasoning-1))))))) diff --git a/src/eca/llm_api.clj b/src/eca/llm_api.clj index 96536a8d..571b4e33 100644 --- a/src/eca/llm_api.clj +++ b/src/eca/llm_api.clj @@ -206,6 +206,7 @@ (let [url-relative-path (:completionUrlRelativePath provider-config) think-tag-start (:thinkTagStart provider-config) think-tag-end (:thinkTagEnd provider-config) + keep-history-reasoning (:keepHistoryReasoning model-config) http-client (:httpClient provider-config)] (handler {:model real-model @@ -221,6 +222,7 @@ :url-relative-path url-relative-path :think-tag-start think-tag-start :think-tag-end think-tag-end + :keep-history-reasoning keep-history-reasoning :http-client http-client :api-url api-url :api-key api-key} diff --git a/src/eca/llm_providers/openai_chat.clj b/src/eca/llm_providers/openai_chat.clj index a6cfb164..18b1ce82 100644 --- a/src/eca/llm_providers/openai_chat.clj +++ b/src/eca/llm_providers/openai_chat.clj @@ -384,19 +384,24 @@ (reset! reasoning-state* {:id nil :type nil :content "" :buffer ""}))) (defn ^:private prune-history - "Ensure DeepSeek-style reasoning_content is discarded from history but kept for the active turn. - Only drops 'reason' messages WITH :delta-reasoning? before the last user message. - Think-tag based reasoning (without :delta-reasoning?) is preserved and transformed to assistant messages." - [messages] - (if-let [last-user-idx (llm-util/find-last-user-msg-idx messages)] - (->> messages - (keep-indexed (fn [i m] - (when-not (and (= "reason" (:role m)) - (get-in m [:content :delta-reasoning?]) - (< i last-user-idx)) - m))) - vec) - messages)) + "Discard reasoning messages from history. + Reasoning with :delta-reasoning? is preserved in the same turn (as required by Deepseek). + This corresponds to the implementation standard. However, it can be change it at the model level configuration. + Parameters: + - messages: the conversation history + - keep-history-reasoning: if true, preserve all reasoning in history" + [messages keep-history-reasoning] + (if keep-history-reasoning + messages + (if-let [last-user-idx (llm-util/find-last-user-msg-idx messages)] + (->> messages + (keep-indexed (fn [i m] + (when-not (and (= "reason" (:role m)) + (or (< i last-user-idx) + (not (get-in m [:content :delta-reasoning?])))) + m))) + vec) + messages))) (defn chat-completion! "Primary entry point for OpenAI chat completions with streaming support. @@ -406,14 +411,14 @@ 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? - think-tag-start think-tag-end http-client]} + think-tag-start think-tag-end keep-history-reasoning 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-tag-end (or think-tag-end "") stream? (boolean callbacks) system-messages (when instructions [{:role "system" :content instructions}]) ;; Pipeline: prune history -> normalize -> merge adjacent assistants -> filter - all-messages (prune-history (vec (concat past-messages user-messages))) + all-messages (prune-history (vec (concat past-messages user-messages)) keep-history-reasoning) messages (vec (concat system-messages (normalize-messages all-messages supports-image? think-tag-start think-tag-end))) @@ -473,7 +478,7 @@ tool-calls)) on-tools-called-wrapper (fn on-tools-called-wrapper [tools-to-call on-tools-called handle-response] (when-let [{:keys [new-messages]} (on-tools-called tools-to-call)] - (let [pruned-messages (prune-history new-messages) + (let [pruned-messages (prune-history new-messages keep-history-reasoning) new-messages-list (vec (concat system-messages (normalize-messages pruned-messages supports-image? think-tag-start think-tag-end))) diff --git a/test/eca/llm_providers/openai_chat_test.clj b/test/eca/llm_providers/openai_chat_test.clj index 67c9edf8..eb1b642f 100644 --- a/test/eca/llm_providers/openai_chat_test.clj +++ b/test/eca/llm_providers/openai_chat_test.clj @@ -259,7 +259,7 @@ {:role "assistant" :reasoning_content "Thinking..."}]))))) (deftest prune-history-test - (testing "Drops reason messages WITH :delta-reasoning? before the last user message (DeepSeek)" + (testing "Drops all reason messages before the last user message by default" (is (match? [{:role "user" :content "Q1"} {:role "assistant" :content "A1"} @@ -272,15 +272,14 @@ {:role "assistant" :content "A1"} {:role "user" :content "Q2"} {:role "reason" :content {:text "r2" :delta-reasoning? true}} - {:role "assistant" :content "A2"}])))) + {:role "assistant" :content "A2"}] + false)))) - (testing "Preserves reason messages WITHOUT :delta-reasoning? (think-tag based)" + (testing "Preserves reason messages (without :delta-reasoning?) before last user message" (is (match? [{:role "user" :content "Q1"} - {:role "reason" :content {:text "thinking..."}} {:role "assistant" :content "A1"} {:role "user" :content "Q2"} - {:role "reason" :content {:text "more thinking..."}} {:role "assistant" :content "A2"}] (#'llm-providers.openai-chat/prune-history [{:role "user" :content "Q1"} @@ -288,12 +287,30 @@ {:role "assistant" :content "A1"} {:role "user" :content "Q2"} {:role "reason" :content {:text "more thinking..."}} - {:role "assistant" :content "A2"}])))) + {:role "assistant" :content "A2"}] + false)))) + + (testing "Preserves all reasoning when keep-history-reasoning is true (Bedrock)" + (is (match? + [{:role "user" :content "Q1"} + {:role "reason" :content {:text "r1"}} + {:role "assistant" :content "A1"} + {:role "user" :content "Q2"} + {:role "reason" :content {:text "r2"}} + {:role "assistant" :content "A2"}] + (#'llm-providers.openai-chat/prune-history + [{:role "user" :content "Q1"} + {:role "reason" :content {:text "r1"}} + {:role "assistant" :content "A1"} + {:role "user" :content "Q2"} + {:role "reason" :content {:text "r2"}} + {:role "assistant" :content "A2"}] + true)))) (testing "No user message leaves list unchanged" (let [msgs [{:role "assistant" :content "A"} {:role "reason" :content {:text "r"}}]] - (is (= msgs (#'llm-providers.openai-chat/prune-history msgs)))))) + (is (= msgs (#'llm-providers.openai-chat/prune-history msgs false)))))) (deftest valid-message-test (testing "Tool messages are always kept"