diff --git a/CHANGELOG.md b/CHANGELOG.md --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,32 @@ ## [Unreleased] +## [0.1.1.0] - 2026-08-23 + +### Changed + +- Widen dependency bounds for GHC 9.8–9.12 (`base`, `bytestring`, `containers`). + `tested-with`: GHC 9.6.7, 9.8.4, 9.10.2, 9.12.2. GitHub Actions CI runs + `cabal test` and `cabal haddock` on that matrix. +- Example `model-catalog.json`: default Gemini entry is `gemini_lite` (`gemini-3.1-flash-lite`); + removed deprecated `gemini-2.5-flash` config; raised example `maxTokens` to 4096; added + optional `gpt_5_6_terra` (`gpt-5.6-terra`); corrected `gemini_lite` and `deepseek4flash` + pricing (DeepSeek rates are peak cache-miss; off-peak is half). +- `get_history` tool description now documents the `"(no earlier history)"` / + `"(no more history)"` sentinels instead of claiming an empty result. + +### Fixed + +- Claude and Gemini gateways honor catalog `baseUrl` / `baseUrlEnv` (previously + ignored; always hit the public Anthropic/Google hosts). New + `claudeGatewayWith` / `geminiGatewayWith` constructors; optional + `CLAUDE_BASE_URL` and `GEMINI_BASE_URL` env overrides. +- `get_history` no longer hangs when the visible context window has zero user + turns (page size 0); `chunkBackward` treats `n <= 0` as a single unpaged chunk. +- `generateObject` / `generateObjectUntyped` never advertise tools (including + auto-injected `get_history`), even when `agContextWindow` is set. New + `createGenRequestNoTools` helper; windowing still truncates messages. + ## [0.1.0.2] - 2026-07-15 ### Added @@ -59,6 +85,7 @@ DeepSeek), single-shot generation with fallbacks, agent tool loops, structured output, JSON model catalog loading, and workspace-scoped filesystem tools. +[0.1.1.0]: https://github.com/aische/llm-simple/compare/v0.1.0.2...v0.1.1.0 [0.1.0.2]: https://github.com/aische/llm-simple/compare/v0.1.0.1...v0.1.0.2 [0.1.0.1]: https://github.com/aische/llm-simple/compare/v0.1.0.0...v0.1.0.1 [0.1.0.0]: https://github.com/aische/llm-simple/releases/tag/v0.1.0.0 diff --git a/Readme.md b/Readme.md --- a/Readme.md +++ b/Readme.md @@ -21,7 +21,7 @@ cabal build ``` -Requires GHC 9.6+ with `GHC2021` (see `llm-simple.cabal`). +Requires GHC 9.6–9.12 with `GHC2021` (see `tested-with` in `llm-simple.cabal`). ## Quick start @@ -84,7 +84,7 @@ | ------------- | --------------------------------------------------------------------------- | | `providerName`| Key referenced by `providerName` in model catalog entries | | `protocol` | `"openai"`, `"claude"`, `"gemini"`, `"ollama"`, or `"deepseek"` | -| `baseUrl` | Provider host/base URL (OpenAI-compatible protocols append `/v1/chat/completions`) | +| `baseUrl` | Provider origin (scheme+host[+port[+optional path prefix]]). OpenAI/DeepSeek/Ollama append `/v1/chat/completions`; Claude appends `/v1/messages`; Gemini appends `/v1beta/models/{model}:generateContent` | | `apiKeyEnv` | Environment variable holding the API key (omit for keyless providers) | | `baseUrlEnv` | Optional env var that overrides `baseUrl` at runtime | @@ -93,8 +93,9 @@ `providerName` from your model catalog. Optional base URL overrides for built-in providers: `OPENAI_BASE_URL`, -`DEEPSEEK_BASE_URL`, `OLLAMA_BASE_URL`. - +`CLAUDE_BASE_URL`, `GEMINI_BASE_URL`, `DEEPSEEK_BASE_URL`, `OLLAMA_BASE_URL`. +Do not include the protocol path (`/v1/...` or `/v1beta/...`) in `baseUrl`; the +library appends it. `loadGateways` and `loadGatewaysWithDotenv` always use built-in provider defaults. Custom providers from a co-located `providers.json` apply when loading models via `loadModelOrThrow` or `loadModelsOrThrow`. @@ -109,7 +110,7 @@ main = do loadFile defaultConfig `catch` \(_ :: SomeException) -> pure () (gemini, deepseek) <- - loadModelsOrThrow "./model-catalog.json" ("gemini_2_5_flash", "deepseek4flash") + loadModelsOrThrow "./model-catalog.json" ("gemini_lite", "deepseek4flash") ... ``` @@ -135,7 +136,7 @@ loadFile defaultConfig `catch` \(_ :: SomeException) -> pure () (gemini, deepseek) <- - loadModelsOrThrow "./model-catalog.json" ("gemini_2_5_flash", "deepseek4flash") + loadModelsOrThrow "./model-catalog.json" ("gemini_lite", "deepseek4flash") let models = ModelWithFallbacks { mwfModel = gemini, mwfFallbacks = [deepseek] } diff --git a/app/Main.hs b/app/Main.hs --- a/app/Main.hs +++ b/app/Main.hs @@ -30,7 +30,7 @@ (_gpt, llama, _haiku, gemini, mistral, deepseek) <- loadModelsOrThrow "./model-catalog.json" - ("gpt_4_1", "llama_3_2", "haiku_4_5", "gemini_2_5_flash", "mistral", "deepseek4flash") + ("gpt_4_1", "llama_3_2", "haiku_4_5", "gemini_lite", "mistral", "deepseek4flash") let _models1 = ModelWithFallbacks {mwfModel = llama, mwfFallbacks = []} _models2 = ModelWithFallbacks {mwfModel = mistral, mwfFallbacks = []} diff --git a/llm-simple.cabal b/llm-simple.cabal --- a/llm-simple.cabal +++ b/llm-simple.cabal @@ -2,7 +2,7 @@ name: llm-simple -version: 0.1.0.2 +version: 0.1.1.0 synopsis: Multi-provider LLM library with agent tool loops and filesystem tools @@ -44,6 +44,9 @@ tested-with: GHC ==9.6.7 + , GHC ==9.8.4 + , GHC ==9.10.2 + , GHC ==9.12.2 source-repository head type: git @@ -142,9 +145,9 @@ aeson-pretty >=0.8 && <0.9, autodocodec >=0.5 && < 0.6, autodocodec-schema >=0.2 && <0.3, - base ^>=4.18, - bytestring >=0.11 && <0.12, - containers ^>=0.6.7, + base >=4.18 && <4.22, + bytestring >=0.11 && <0.13, + containers >=0.6.7 && <0.8, directory >=1.3 && <1.4, dotenv >=0.9 && <0.13, filepath >=1.4 && <1.6, @@ -169,9 +172,9 @@ aeson >=2.0 && <2.3, autodocodec >=0.5 && < 0.6, autodocodec-schema >=0.2 && <0.3, - base ^>=4.18, - bytestring >=0.11 && <0.12, - containers ^>=0.6.7, + base >=4.18 && <4.22, + bytestring >=0.11 && <0.13, + containers >=0.6.7 && <0.8, dotenv >=0.9 && <0.13, directory >=1.3 && <1.4, filepath >=1.4 && <1.6, @@ -193,9 +196,9 @@ aeson >=2.0 && <2.3, aeson-pretty >=0.8 && <0.9, autodocodec >=0.5 && < 0.6, - base ^>=4.18, - bytestring >=0.11 && <0.12, - containers ^>=0.6.7, + base >=4.18 && <4.22, + bytestring >=0.11 && <0.13, + containers >=0.6.7 && <0.8, dotenv >=0.9 && <0.13, directory >=1.3 && <1.4, filepath >=1.4 && <1.6, @@ -228,8 +231,8 @@ LLM.WeatherTool hs-source-dirs: test build-depends: - base ^>=4.18, - containers ^>=0.6.7, + base >=4.18 && <4.22, + containers >=0.6.7 && <0.8, directory >=1.3 && <1.4, filepath >=1.4 && <1.6, temporary >=1.3 && <1.4, @@ -237,7 +240,7 @@ aeson >=2.0 && <2.3, aeson-pretty >=0.8 && <0.9, autodocodec >=0.5 && < 0.6, - bytestring >=0.11 && <0.12, + bytestring >=0.11 && <0.13, hspec >=2.10 && <2.12, heptapod >=1.1 && <1.2, llm-simple, diff --git a/model-catalog.json b/model-catalog.json --- a/model-catalog.json +++ b/model-catalog.json @@ -7,7 +7,7 @@ "pricePerMillionInput": 2.0, "pricePerMillionOutput": 8.0 }, - "maxTokens": 1024, + "maxTokens": 4096, "temperature": 0.5, "requestTimeout": 10000, "throttleDelay": 1000, @@ -15,6 +15,21 @@ "jitterBackoff": 1000 }, { + "modelConfigName": "gpt_5_6_terra", + "providerName": "openai", + "modelName": "gpt-5.6-terra", + "pricing": { + "pricePerMillionInput": 2.0, + "pricePerMillionOutput": 12.0 + }, + "maxTokens": 4096, + "temperature": 0.5, + "requestTimeout": 10000, + "throttleDelay": 1000, + "retryCount": 3, + "jitterBackoff": 1000 + }, + { "modelConfigName": "llama_3_2", "providerName": "ollama", "modelName": "llama3.2:latest", @@ -22,7 +37,7 @@ "pricePerMillionInput": 0.0, "pricePerMillionOutput": 0.0 }, - "maxTokens": 1024, + "maxTokens": 4096, "temperature": 0.5, "requestTimeout": 3000, "throttleDelay": 1000, @@ -37,7 +52,7 @@ "pricePerMillionInput": 0.0, "pricePerMillionOutput": 0.0 }, - "maxTokens": 1024, + "maxTokens": 4096, "temperature": 0.5, "requestTimeout": 30000, "throttleDelay": 1000, @@ -45,29 +60,14 @@ "jitterBackoff": 1000 }, { - "modelConfigName": "gemini_2_5_flash", - "providerName": "gemini", - "modelName": "gemini-2.5-flash", - "pricing": { - "pricePerMillionInput": 0.1, - "pricePerMillionOutput": 0.4 - }, - "maxTokens": 1024, - "temperature": 0.5, - "requestTimeout": 10000, - "throttleDelay": 1000, - "retryCount": 3, - "jitterBackoff": 1000 - }, - { "modelConfigName": "gemini_lite", "providerName": "gemini", "modelName": "gemini-3.1-flash-lite", "pricing": { - "pricePerMillionInput": 0.1, - "pricePerMillionOutput": 0.4 + "pricePerMillionInput": 0.25, + "pricePerMillionOutput": 1.5 }, - "maxTokens": 1024, + "maxTokens": 4096, "temperature": 0.5, "requestTimeout": 10000, "throttleDelay": 1000, @@ -82,7 +82,7 @@ "pricePerMillionInput": 1, "pricePerMillionOutput": 5 }, - "maxTokens": 1024, + "maxTokens": 4096, "temperature": 0.5, "requestTimeout": 30000, "throttleDelay": 5000, @@ -94,10 +94,10 @@ "providerName": "deepseek", "modelName": "deepseek-v4-flash", "pricing": { - "pricePerMillionInput": 0.14, - "pricePerMillionOutput": 0.28 + "pricePerMillionInput": 0.44, + "pricePerMillionOutput": 1.32 }, - "maxTokens": 1024, + "maxTokens": 4096, "temperature": 0.5, "requestTimeout": 20000, "throttleDelay": 3000, diff --git a/providers.json b/providers.json --- a/providers.json +++ b/providers.json @@ -10,13 +10,15 @@ "providerName": "claude", "protocol": "claude", "baseUrl": "https://api.anthropic.com", - "apiKeyEnv": "CLAUDE_API_KEY" + "apiKeyEnv": "CLAUDE_API_KEY", + "baseUrlEnv": "CLAUDE_BASE_URL" }, { "providerName": "gemini", "protocol": "gemini", "baseUrl": "https://generativelanguage.googleapis.com", - "apiKeyEnv": "GEMINI_API_KEY" + "apiKeyEnv": "GEMINI_API_KEY", + "baseUrlEnv": "GEMINI_BASE_URL" }, { "providerName": "deepseek", diff --git a/src/LLM.hs b/src/LLM.hs --- a/src/LLM.hs +++ b/src/LLM.hs @@ -98,6 +98,7 @@ Tool (..), ToolContext (..), createGenRequest, + createGenRequestNoTools, toTool, GenerateEvent (..), GenerateEventDetail (..), @@ -125,6 +126,7 @@ Tool (..), ToolContext (..), createGenRequest, + createGenRequestNoTools, generateText, noEventObserver, streamText, diff --git a/src/LLM/Agent.hs b/src/LLM/Agent.hs --- a/src/LLM/Agent.hs +++ b/src/LLM/Agent.hs @@ -26,6 +26,7 @@ noEventObserver, toTool, createGenRequest, + createGenRequestNoTools, ) where diff --git a/src/LLM/Agent/GenerateObject.hs b/src/LLM/Agent/GenerateObject.hs --- a/src/LLM/Agent/GenerateObject.hs +++ b/src/LLM/Agent/GenerateObject.hs @@ -1,8 +1,7 @@ module LLM.Agent.GenerateObject where import Data.Aeson (Value) -import Data.Map qualified as Map -import LLM.Agent.ToolUtils (createGenRequest) +import LLM.Agent.ToolUtils (createGenRequestNoTools) import LLM.Agent.Types ( Agent (..), RuntimeArgs (..), @@ -19,7 +18,8 @@ -- | Generate a typed Haskell value from the model via Autodocodec. -- --- Tools are not used; conversation context is taken from @turns@ only. +-- Tools are never sent: @grTools@ is always empty. 'agContextWindow' still +-- truncates messages, but does not inject @get_history@. generateObject :: (GeneratableObject t) => Agent -> @@ -27,9 +27,12 @@ RuntimeArgs -> [Turn] -> IO (Either GenerateErrorResult (t, Usage)) -generateObject a m r t = genObject (createGenRequest id a Map.empty r t) m +generateObject a m r t = genObject (createGenRequestNoTools a r t) m -- | Generate a JSON 'Value' from the model using a caller-supplied schema. +-- +-- Same tool policy as 'generateObject': windowing may truncate, tools are not +-- advertised or executed. generateObjectUntyped :: Agent -> ModelWithFallbacks -> @@ -37,4 +40,4 @@ [Turn] -> Value -> IO (Either GenerateErrorResult (Value, Usage)) -generateObjectUntyped a m r t = genObjectUntyped (createGenRequest id a Map.empty r t) m +generateObjectUntyped a m r t = genObjectUntyped (createGenRequestNoTools a r t) m diff --git a/src/LLM/Agent/ToolUtils.hs b/src/LLM/Agent/ToolUtils.hs --- a/src/LLM/Agent/ToolUtils.hs +++ b/src/LLM/Agent/ToolUtils.hs @@ -9,6 +9,7 @@ getResolvedTools, windowOffset, createGenRequest, + createGenRequestNoTools, embedTextTool, ) where @@ -165,6 +166,23 @@ in GenRequest { grSystemPrompt = agent.agSystemPrompt, grTools = map (\x -> x.toolDef) tools, + grMessages = drop offset messages, + grAbortSignal = rt.rtAbortSignal, + grLLMHooks = rt.rtLLMHooks, + grHooks = rt.rtHooks + } + +-- | Like 'createGenRequest', but never attaches tool definitions. +-- +-- Still applies 'agContextWindow' truncation to 'grMessages'. Use this for +-- structured-output paths ('generateObject' / 'generateObjectUntyped'), which +-- cannot execute tools and must not advertise @get_history@. +createGenRequestNoTools :: Agent -> RuntimeArgs -> [Turn] -> GenRequest +createGenRequestNoTools agent rt messages = + let offset = windowOffset agent.agContextWindow messages + in GenRequest + { grSystemPrompt = agent.agSystemPrompt, + grTools = [], grMessages = drop offset messages, grAbortSignal = rt.rtAbortSignal, grLLMHooks = rt.rtLLMHooks, diff --git a/src/LLM/Agent/Tools/HistoryTool.hs b/src/LLM/Agent/Tools/HistoryTool.hs --- a/src/LLM/Agent/Tools/HistoryTool.hs +++ b/src/LLM/Agent/Tools/HistoryTool.hs @@ -42,7 +42,8 @@ ttoolDescription = "Retrieve earlier conversation history that is not in your current context window. " <> "Pass chunk=0 for the most recent hidden history, chunk=1 for the one before that, etc. " - <> "Returns an empty result when there is no more history.", + <> "Returns \"(no earlier history)\" when nothing is hidden, and \"(no more history)\" " + <> "when the chunk index is out of range.", ttoolReadonly = True, ttoolExecute = getHistoryExecTyped } @@ -59,15 +60,23 @@ -- backward from the end. Each page starts at a 'UserTurn'. -- Chunk 0 is the most recent page, chunk 1 the one before, etc. -- The oldest chunk (highest index) may contain fewer than @n@ user messages. +-- +-- When @n <= 0@ (e.g. the visible window has no user turns), paging is +-- disabled and the whole conversation is returned as a single chunk. +-- This avoids an infinite loop in the page walker when @start == end@. chunkBackward :: Int -> [Turn] -> [[Turn]] chunkBackward _ [] = [] -chunkBackward n conv = reverse (go (length conv) []) +chunkBackward n conv + | n <= 0 = [conv] + | otherwise = reverse (go (length conv) []) where go 0 acc = acc go end acc = let start = findNthUserBack n (take end conv) page = slice start end conv - in go start (page : acc) + in if start >= end + then page : acc -- should not happen for n > 0; refuse to loop + else go start (page : acc) -- | Find the start index for a page containing @n@ user messages, -- scanning backward from the end of the given prefix. diff --git a/src/LLM/Generate/GenerateUtils.hs b/src/LLM/Generate/GenerateUtils.hs --- a/src/LLM/Generate/GenerateUtils.hs +++ b/src/LLM/Generate/GenerateUtils.hs @@ -19,7 +19,7 @@ ) import LLM.Core.Usage (Usage (..), estimateCost) import LLM.Core.Utils (withRetry, withTimeout) -import LLM.Generate.Logger (Hooks (..), LogLevel (..), onLog) +import LLM.Generate.Logger (Hooks (..), LogLevel (..)) import LLM.Generate.ModelConfig ( ModelConfig (..), ModelWithFallbacks (..), diff --git a/src/LLM/Load/LoadGateways.hs b/src/LLM/Load/LoadGateways.hs --- a/src/LLM/Load/LoadGateways.hs +++ b/src/LLM/Load/LoadGateways.hs @@ -29,9 +29,9 @@ defaultProviderCatalogMap, ) import LLM.Load.Types (LoadConfigError (LoadModelConfigError)) -import LLM.Providers.Claude (claudeGateway) +import LLM.Providers.Claude (claudeGatewayWith) import LLM.Providers.DeepSeek (deepSeekGatewayWith) -import LLM.Providers.Gemini (geminiGateway) +import LLM.Providers.Gemini (geminiGatewayWith) import LLM.Providers.Ollama (ollamaGatewayWith) import LLM.Providers.OpenAI (openAIGatewayWithName) import Network.HTTP.Client qualified as HC @@ -91,8 +91,12 @@ buildGateway :: ProviderCatalogItem -> ProviderBaseUrl -> Text -> LLMGateway buildGateway item baseUrl apiKey = case item.protocol of - ClaudeProtocol -> claudeGateway apiKey - GeminiProtocol -> geminiGateway apiKey + ClaudeProtocol -> case baseUrl of + HttpsProviderBase url opts -> claudeGatewayWith url opts apiKey + HttpProviderBase url opts -> claudeGatewayWith url opts apiKey + GeminiProtocol -> case baseUrl of + HttpsProviderBase url opts -> geminiGatewayWith url opts apiKey + HttpProviderBase url opts -> geminiGatewayWith url opts apiKey OpenAIProtocol -> case baseUrl of HttpsProviderBase url opts -> openAIGatewayWithName item.providerName url opts apiKey HttpProviderBase url opts -> openAIGatewayWithName item.providerName url opts apiKey diff --git a/src/LLM/Load/ProviderCatalog.hs b/src/LLM/Load/ProviderCatalog.hs --- a/src/LLM/Load/ProviderCatalog.hs +++ b/src/LLM/Load/ProviderCatalog.hs @@ -85,7 +85,7 @@ protocol = ClaudeProtocol, baseUrl = "https://api.anthropic.com", apiKeyEnv = Just "CLAUDE_API_KEY", - baseUrlEnv = Nothing + baseUrlEnv = Just "CLAUDE_BASE_URL" } geminiItem = ProviderCatalogItem @@ -93,7 +93,7 @@ protocol = GeminiProtocol, baseUrl = "https://generativelanguage.googleapis.com", apiKeyEnv = Just "GEMINI_API_KEY", - baseUrlEnv = Nothing + baseUrlEnv = Just "GEMINI_BASE_URL" } deepseekItem = ProviderCatalogItem diff --git a/src/LLM/Providers.hs b/src/LLM/Providers.hs --- a/src/LLM/Providers.hs +++ b/src/LLM/Providers.hs @@ -6,11 +6,15 @@ -- * Gemini geminiProvider, + geminiProviderWith, geminiGateway, + geminiGatewayWith, -- * Claude claudeProvider, + claudeProviderWith, claudeGateway, + claudeGatewayWith, -- * Ollama ollamaProvider, @@ -22,8 +26,8 @@ ) where -import LLM.Providers.Claude (claudeGateway, claudeProvider) +import LLM.Providers.Claude (claudeGateway, claudeGatewayWith, claudeProvider, claudeProviderWith) import LLM.Providers.DeepSeek (deepSeekGateway, deepSeekProvider) -import LLM.Providers.Gemini (geminiGateway, geminiProvider) +import LLM.Providers.Gemini (geminiGateway, geminiGatewayWith, geminiProvider, geminiProviderWith) import LLM.Providers.Ollama (ollamaGateway, ollamaProvider) import LLM.Providers.OpenAI (openAIGateway, openAIProvider) diff --git a/src/LLM/Providers/Claude.hs b/src/LLM/Providers/Claude.hs --- a/src/LLM/Providers/Claude.hs +++ b/src/LLM/Providers/Claude.hs @@ -1,6 +1,8 @@ module LLM.Providers.Claude ( claudeGateway, + claudeGatewayWith, claudeProvider, + claudeProviderWith, parseClaudeResponse, parseClaudeUsage, ) @@ -53,7 +55,6 @@ ( Option, POST (POST), ReqBodyJson (ReqBodyJson), - Scheme (Https), Url, header, https, @@ -66,51 +67,57 @@ (/:), ) --- | Create a LLMGateway for the Claude provider. Takes the API key as a parameter. +-- | Create a LLMGateway for the Claude provider at api.anthropic.com. claudeGateway :: Text -> LLMGateway claudeGateway apiKey = toGateway $ claudeProvider apiKey --- | Create a LLMProvider for the Claude provider. Takes the API key as a parameter. +-- | Create a Claude-compatible client with a custom base URL (origin). +-- The library appends @/v1/messages@. +claudeGatewayWith :: Url scheme -> Option scheme -> Text -> LLMGateway +claudeGatewayWith baseUrl baseOpts apiKey = toGateway (claudeProviderWith baseUrl baseOpts apiKey) + +-- | Create a LLMProvider for the Claude provider at api.anthropic.com. claudeProvider :: Text -> LLMProvider -claudeProvider apiKey = +claudeProvider = claudeProviderWith (https "api.anthropic.com") mempty + +-- | Claude-compatible provider with a custom base URL (origin). +-- The library appends @/v1/messages@. +claudeProviderWith :: Url scheme -> Option scheme -> Text -> LLMProvider +claudeProviderWith baseUrl baseOpts apiKey = LLMProvider { providerName = "claude", buildBody = claudeBuildBody, sendRequest = sendRequest, sendStreamRequest = \body callback -> - runReq lenientConfig $ - reqBr POST claudeUrl (ReqBodyJson body) (claudeOpts apiKey) $ \resp -> + runReq lenientConfig $ do + let url = baseUrl /: "v1" /: "messages" + opts = baseOpts <> claudeAuthOpts apiKey + reqBr POST url (ReqBodyJson body) opts $ \resp -> handleStreamResponse resp (`parseClaudeStream` callback), parseResponse = pure . parseClaudeResponse, - -- buildObjectBody _ r schema = claudeBuildBody False (r {reqConversation = reqConversation r <> Conversation [UserTurn ("Generate a JSON object matching this schema: " <> T.pack (show schema))]}) buildObjectBody = \r schema -> let schemaText = TL.toStrict . decodeUtf8 $ encode schema instruction = "Respond with a raw JSON object matching this schema. No markdown, no explanation, no code fences:\n" <> schemaText conv' = r.reqConversation <> [UserTurn instruction] in claudeBuildBody False (r {reqConversation = conv'}), sendObjectRequest = sendRequest, - -- parseObjectResponse _ = parseClaudeObjectResponse parseObjectResponse = parseClaudeObjectResponse } where sendRequest body = runReq lenientConfig $ do - resp <- req POST claudeUrl (ReqBodyJson body) jsonResponse (claudeOpts apiKey) + let url = baseUrl /: "v1" /: "messages" + opts = baseOpts <> claudeAuthOpts apiKey + resp <- req POST url (ReqBodyJson body) jsonResponse opts pure (responseStatusCode resp, responseBody resp) -- Internal helpers -claudeUrl :: Url 'Https -claudeUrl = https "api.anthropic.com" /: "v1" /: "messages" - -claudeOpts :: Text -> Option 'Https -claudeOpts apiKey = +claudeAuthOpts :: Text -> Option scheme +claudeAuthOpts apiKey = header "x-api-key" (encodeUtf8 apiKey) <> header "anthropic-version" "2023-06-01" --- | Create an LLMClient from Claude credentials --- claudeGateway :: Text -> LLMGateway --- claudeGateway apiKey = toGateway (Claude apiKey) parseClaudeStream :: HC.BodyReader -> (StreamEvent -> IO ()) -> IO LLMTextResult parseClaudeStream reader callback = do blocksRef <- newIORef ([] :: [ContentBlock]) diff --git a/src/LLM/Providers/Gemini.hs b/src/LLM/Providers/Gemini.hs --- a/src/LLM/Providers/Gemini.hs +++ b/src/LLM/Providers/Gemini.hs @@ -1,6 +1,8 @@ module LLM.Providers.Gemini ( geminiGateway, + geminiGatewayWith, geminiProvider, + geminiProviderWith, parseGeminiResponse, parseGeminiUsage, ) @@ -53,8 +55,10 @@ import LLM.Core.Usage (Usage (..)) import Network.HTTP.Client qualified as HC import Network.HTTP.Req - ( POST (POST), + ( Option, + POST (POST), ReqBodyJson (ReqBodyJson), + Url, header, https, jsonResponse, @@ -67,13 +71,23 @@ (=:), ) --- | Create a LLMGateway for the Gemini provider. Takes the API key as a parameter. +-- | Create a LLMGateway for the Gemini provider at generativelanguage.googleapis.com. geminiGateway :: Text -> LLMGateway geminiGateway apiKey = toGateway (geminiProvider apiKey) --- | Create a LLMProvider for the Gemini provider. Takes the API key as a parameter. +-- | Create a Gemini-compatible client with a custom base URL (origin). +-- The library appends @/v1beta/models/{model}:generateContent@ (and the stream variant). +geminiGatewayWith :: Url scheme -> Option scheme -> Text -> LLMGateway +geminiGatewayWith baseUrl baseOpts apiKey = toGateway (geminiProviderWith baseUrl baseOpts apiKey) + +-- | Create a LLMProvider for the Gemini provider at generativelanguage.googleapis.com. geminiProvider :: Text -> LLMProvider -geminiProvider apiKey = +geminiProvider = geminiProviderWith (https "generativelanguage.googleapis.com") mempty + +-- | Gemini-compatible provider with a custom base URL (origin). +-- The library appends @/v1beta/models/{model}:generateContent@ (and the stream variant). +geminiProviderWith :: Url scheme -> Option scheme -> Text -> LLMProvider +geminiProviderWith baseUrl baseOpts apiKey = LLMProvider { providerName = "gemini", buildBody = const geminiBuildBody, @@ -82,11 +96,12 @@ runReq lenientConfig $ do let model = extractModel body url = - https "generativelanguage.googleapis.com" + baseUrl /: "v1beta" /: "models" /: (model <> ":streamGenerateContent") - reqBr POST url (ReqBodyJson (stripBoundsAndComments $ stripModel body)) (header "x-goog-api-key" (encodeUtf8 apiKey) <> "alt" =: ("sse" :: Text)) $ \resp -> + opts = baseOpts <> geminiAuthOpts apiKey <> "alt" =: ("sse" :: Text) + reqBr POST url (ReqBodyJson (stripBoundsAndComments $ stripModel body)) opts $ \resp -> handleStreamResponse resp (`parseGeminiStream` callback), parseResponse = parseGeminiResponse, buildObjectBody = \r schema -> @@ -118,12 +133,16 @@ -- We extract it from the request body JSON since the LLMProvider only passes Value. let model = extractModel body url = - https "generativelanguage.googleapis.com" + baseUrl /: "v1beta" /: "models" /: (model <> ":generateContent") - resp <- req POST url (ReqBodyJson (stripBoundsAndComments $ stripModel body)) jsonResponse (header "x-goog-api-key" (encodeUtf8 apiKey)) + opts = baseOpts <> geminiAuthOpts apiKey + resp <- req POST url (ReqBodyJson (stripBoundsAndComments $ stripModel body)) jsonResponse opts pure (responseStatusCode resp, responseBody resp) + +geminiAuthOpts :: Text -> Option scheme +geminiAuthOpts apiKey = header "x-goog-api-key" (encodeUtf8 apiKey) -- | Extract model name stashed in the request body by geminiBuildBody. extractModel :: Value -> Text diff --git a/test/LLM/GenerateObjectSpec.hs b/test/LLM/GenerateObjectSpec.hs --- a/test/LLM/GenerateObjectSpec.hs +++ b/test/LLM/GenerateObjectSpec.hs @@ -3,13 +3,14 @@ module LLM.GenerateObjectSpec (spec) where import Data.Aeson (Value, object, (.=)) +import Data.IORef (IORef, newIORef, readIORef, writeIORef) import Data.Text (Text) import Data.Text qualified as T import Heptapod (generate) import LLM.Agent.GenerateObject (generateObject, generateObjectUntyped) import LLM.Agent.Types (Agent (..), RuntimeArgs (..)) import LLM.Core.Abort (AbortSignal, abort, newAbortSignal) -import LLM.Core.Types (ChatResponse (..), LLMError (..), LLMGateway (..), LLMHooks (..), Turn (..)) +import LLM.Core.Types (ChatRequest (..), ChatResponse (..), LLMError (..), LLMGateway (..), LLMHooks (..), Turn (..)) import LLM.Core.Usage (PricingInfo (..), Usage (..)) import LLM.Generate.Logger (noHooks) import LLM.Generate.ModelConfig @@ -88,6 +89,26 @@ T.unpack msg `shouldContain` "Can't decode object" _ -> expectationFailure "expected GErrParseObjectError" + it "does not advertise tools even when agContextWindow would inject get_history" $ do + captured <- newIORef Nothing + let gw = capturingObjectGateway captured (object ["location" .= ("Paris" :: Text)]) (Usage 1 0 0) + models = ModelWithFallbacks (mockModel gw) [] + agent = defaultAgent {agContextWindow = Just 1} + conv = + [ UserTurn "first", + AssistantTurn "a1" Nothing [], + UserTurn "second" + ] + rt <- mkRuntime Nothing + result <- generateObjectUntyped agent models rt conv (object ["type" .= ("object" :: Text)]) + case result of + Right _ -> do + mReq <- readIORef captured + case mReq of + Nothing -> expectationFailure "expected gwGenerateObject to be called" + Just req -> req.reqTools `shouldBe` [] + Left err -> expectationFailure $ show err + objectGateway :: Value -> Usage -> LLMGateway objectGateway value usage = LLMGateway @@ -95,6 +116,15 @@ gwGenerateText = \_ _ -> pure (Right (ChatResponse "" [] Nothing Nothing)), gwStreamText = \_ _ _ -> pure (Right (ChatResponse "" [] Nothing Nothing)), gwGenerateObject = \_ _ _ -> pure (Right (value, Just usage)) + } + +-- | Mock that records the 'ChatRequest' seen by 'gwGenerateObject'. +capturingObjectGateway :: IORef (Maybe ChatRequest) -> Value -> Usage -> LLMGateway +capturingObjectGateway ref value usage = + (objectGateway value usage) + { gwGenerateObject = \_ _ req -> do + writeIORef ref (Just req) + pure (Right (value, Just usage)) } errorObjectGateway :: LLMError -> LLMGateway diff --git a/test/LLM/HistoryToolSpec.hs b/test/LLM/HistoryToolSpec.hs --- a/test/LLM/HistoryToolSpec.hs +++ b/test/LLM/HistoryToolSpec.hs @@ -52,6 +52,22 @@ result <- runHistory windowedAgent sampleConversation 9 result `shouldBe` "(no more history)" + -- When the visible window has zero UserTurns, page size is 0. Previously + -- chunkBackward looped forever; it now returns the whole hidden prefix as + -- a single chunk. + it "returns the full hidden prefix when the visible window has no user turns" $ do + let conv = + [ UserTurn "hidden question", + AssistantTurn "hidden answer" Nothing [], + AssistantTurn "visible assistant only" Nothing [] + ] + -- Offset past both user+assistant hidden turns; visible slice is + -- assistant-only so countUserTurns == 0. + offset = 2 + result <- runHistoryAtOffset offset conv 0 + T.unpack result `shouldContain` "[User] hidden question" + T.unpack result `shouldContain` "[Assistant] hidden answer" + sampleConversation :: [Turn] sampleConversation = [ UserTurn "first question", @@ -76,10 +92,12 @@ windowedAgent = noWindowAgent {agContextWindow = Just 1} runHistory :: Agent -> [Turn] -> Int -> IO Text -runHistory agent conv chunk = do +runHistory agent conv = runHistoryAtOffset (windowOffset agent.agContextWindow conv) conv + +runHistoryAtOffset :: Int -> [Turn] -> Int -> IO Text +runHistoryAtOffset offset conv chunk = do genId <- generate - let offset = windowOffset agent.agContextWindow conv - ctx = + let ctx = ToolContext { tcConversation = conv, tcUsage = mempty, diff --git a/test/LLM/LoadSpec.hs b/test/LLM/LoadSpec.hs --- a/test/LLM/LoadSpec.hs +++ b/test/LLM/LoadSpec.hs @@ -95,6 +95,34 @@ LLMGateway {gwName = name} = gateway name `shouldBe` "openrouter" + it "builds a claude gateway from a custom baseUrl origin" $ do + let item = + ProviderCatalogItem + { providerName = "claude", + protocol = ClaudeProtocol, + baseUrl = "https://claude-proxy.example.com", + apiKeyEnv = Just "CLAUDE_API_KEY", + baseUrlEnv = Nothing + } + Right baseUrl <- pure $ parseProviderBaseUrl item.baseUrl + let gateway = buildGateway item baseUrl "test-key" + LLMGateway {gwName = name} = gateway + name `shouldBe` "claude" + + it "builds a gemini gateway from a custom baseUrl origin" $ do + let item = + ProviderCatalogItem + { providerName = "gemini", + protocol = GeminiProtocol, + baseUrl = "https://gemini-proxy.example.com", + apiKeyEnv = Just "GEMINI_API_KEY", + baseUrlEnv = Nothing + } + Right baseUrl <- pure $ parseProviderBaseUrl item.baseUrl + let gateway = buildGateway item baseUrl "test-key" + LLMGateway {gwName = name} = gateway + name `shouldBe` "gemini" + it "loads gateways from a provider catalog file" $ do result <- runExceptT $ loadProviderCatalog providersOllamaOnly case result of diff --git a/test/LLM/TypesSpec.hs b/test/LLM/TypesSpec.hs --- a/test/LLM/TypesSpec.hs +++ b/test/LLM/TypesSpec.hs @@ -13,10 +13,6 @@ addUsage, emptyUsage, estimateCost, - pricePerMillionInput, - pricePerMillionOutput, - usageInputTokens, - usageOutputTokens, ) import LLM.Core.Utils ( getToolCalls,