Add a native Anthropic provider and stop sending Claude malformed requests

Two auto-reported session failures traced to the same place: Autohand sends
Claude requests that Anthropic rejects. The first named a model OpenRouter does
not serve; the second named a valid one and still failed, because OpenRouter
forwarded a payload Anthropic refused. Anthropic validates strictly and rejects
unsupported parameters rather than ignoring them, so anything we send loosely
comes back as an opaque 400.

Adds `anthropic` as a first-class provider on the official TypeScript SDK, with
the Messages API contract handled at the provider boundary: system prompts,
images, tool calls and results, structured output, adaptive thinking, and
Anthropic usage fields. The bundled catalog gains the current Claude lineup for
both the native provider and OpenRouter, and drops two OpenRouter IDs that never
existed. Setup, /model, /settings, ACP, and the answer-only RPC surface all know
about the provider.

Every optional request parameter is now gated on the target model. Temperature
is dropped for the Claude 5 family and Opus 4.7/4.8, output effort is clamped or
omitted where unsupported, and the thinking parameter is omitted entirely for
models that always think. Sending temperature unconditionally had been breaking
the default OpenRouter model on every request.

Thinking blocks are captured from each response and replayed unchanged on the
next one. Thinking level defaults to normal, which omits the parameter, but the
current models think adaptively regardless — so blocks come back every turn, and
dropping them risks ordering and signature rejections inside the tool loop.
`reasoningBlocks` is a provider-opaque passthrough on the message type, so other
providers needing the same replay contract can adopt it later.

Conversation history is repaired on the way out rather than trusted. A tool-only
assistant turn now sends null content instead of an empty string, which
aggregators translate into an empty text block; blank turns are dropped; blank
tool results become a placeholder; and tool calls or results orphaned by context
compaction are backfilled or removed. This is shared by OpenRouter and LLM
Gateway, and the equivalent repair is applied to Bedrock's Converse payload.

Vertex AI's native Claude branch was posting OpenAI-shaped messages, system role
included, straight to streamRawPredict. It could not have worked with a system
prompt. It now uses the shared Messages translation.

OpenRouter's error.metadata.raw carries the upstream provider's actual
complaint, while its own message is often the wrapper "Provider returned error".
Discarding it is why both reports were undiagnosable; it is now included, which
also lets a prompt-too-long rejection classify as context overflow and trigger
auto-compaction instead of failing the session.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>

Author: igorcosta

.github/ISSUE_TEMPLATE/model_catalog.yml

@@ -16,6 +16,7 @@ body:
       options:
         - autohandai
         - openrouter
+        - anthropic
         - ollama
         - openai
         - llmgateway

.github/scripts/generate-model-catalog.mjs

@@ -5,6 +5,7 @@ import { readFileSync, writeFileSync } from "node:fs";
 const PROVIDER_DEFAULTS = {
   autohandai: { api: "openai-completions", baseUrl: "https://api.autohand.ai/v1", contextWindow: 64000 },
   openrouter: { api: "openai-completions", baseUrl: "https://openrouter.ai/api/v1", contextWindow: 131072 },
+  anthropic: { api: "anthropic-messages", baseUrl: "https://api.anthropic.com", contextWindow: 1000000 },
   ollama: { api: "openai-completions", baseUrl: "http://127.0.0.1:11434/v1", contextWindow: 131072 },
   llamacpp: { api: "openai-completions", baseUrl: "http://127.0.0.1:8080/v1", contextWindow: 131072 },
   openai: { api: "openai-responses", baseUrl: "https://api.openai.com/v1", contextWindow: 400000 },

bun.lock

@@ -188,6 +188,7 @@ Active LLM provider to use.
 | -------------- | ---------------------------- |
 | `"autohandai"` | Autohand AI Cloud or Local   |
 | `"openrouter"` | OpenRouter API (default)     |
+| `"anthropic"`  | Anthropic Messages API       |
 | `"ollama"`     | Local Ollama instance        |
 | `"llamacpp"`   | Local llama.cpp server       |
 | `"openai"`     | OpenAI API directly          |
@@ -280,6 +281,29 @@ OpenRouter provider configuration.
 | `model`         | string | Yes      | -                              | Model identifier (e.g., `your-modelcard-id-here`)                           |
 | `contextWindow` | number | No       | Auto                           | Exact model context window. Autohand fills this from OpenRouter when known. |
 
+### `anthropic`
+
+Native Anthropic Messages API configuration. Autohand sends system prompts, tool definitions, tool uses, and tool results using Anthropic's native request format.
+
+```json
+{
+  "anthropic": {
+    "apiKey": "sk-ant-xxx",
+    "baseUrl": "https://api.anthropic.com",
+    "model": "claude-sonnet-5",
+    "contextWindow": 1000000
+  }
+}
+```
+
+| Field           | Type   | Required | Default                     | Description                                      |
+| --------------- | ------ | -------- | --------------------------- | ------------------------------------------------ |
+| `apiKey`        | string | Yes      | -                           | Your Anthropic Console API key                   |
+| `baseUrl`       | string | No       | `https://api.anthropic.com` | Anthropic API origin; `/v1/messages` is appended |
+| `model`         | string | Yes      | `claude-sonnet-5`           | Native Anthropic model identifier                |
+| `contextWindow` | number | No       | Catalog value               | Model context window used by Autohand accounting |
+| `reasoningEffort` | string | No     | Provider default            | Native output effort (`low` through `xhigh`)      |
+
 ### `zai`
 
 Z.ai provider configuration.

docs/config-reference.md

@@ -8,6 +8,7 @@ Autohand supports multiple LLM providers, giving you flexibility to choose betwe
 - [Provider Comparison](#provider-comparison)
 - [Cloud Providers](#cloud-providers)
   - [OpenRouter](#openrouter)
+  - [Anthropic](#anthropic)
   - [OpenAI](#openai)
   - [LLM Gateway](#llm-gateway)
   - [DeepSeek](#deepseek)
@@ -54,6 +55,7 @@ Bundled provider model choices live in `src/providers/models.json` and packaged
 | --------------- | ----- | ----------- | ------- | ----------------------------------------------- |
 | **Autohand AI** | Cloud/Local | Account, API key, or local | Low | Fantail ultra-fast coding, Moa thinking, or local MLX coding models |
 | **OpenRouter**  | Cloud | Pay-per-use | Low     | Access to 100+ models, recommended default      |
+| **Anthropic**   | Cloud | Pay-per-use | Low     | Native Claude Messages API and tool use         |
 | **OpenAI**      | Cloud | Pay-per-use | Low     | Direct OpenAI access, GPT-5, o3 models          |
 | **LLM Gateway** | Cloud | Pay-per-use | Low     | Unified API for multiple providers              |
 | **DeepSeek**    | Cloud | Pay-per-use | Low     | DeepSeek V4 Flash and V4 Pro models             |
@@ -150,21 +152,67 @@ OpenRouter provides a unified API to access 100+ models from various providers (
 }
 ```
 
-**Popular Models:**
+**Current Anthropic Models:**
 | Model | Description |
 |-------|-------------|
-| `your-modelcard-id-here` | Best balance of speed and capability |
-| `anthropic/claude-5-opus` | Most capable Claude model |
+| `anthropic/claude-sonnet-5` | Claude Sonnet 5 |
+| `anthropic/claude-opus-5` | Claude Opus 5 |
+| `anthropic/claude-fable-5` | Claude Fable 5 |
+| `anthropic/claude-opus-4.8` | Claude Opus 4.8 |
+| `anthropic/claude-haiku-4.5` | Claude Haiku 4.5 |
 | `openai/gpt-5` | OpenAI's flagship model |
 | `google/gemini-3.0-pro` | Google's latest model |
 | `meta-llama/llama-3.1-70b-instruct` | Open-source alternative |
 
 **Switching Models:**
 
 ```
-/model anthropic/claude-5-opus
+/model anthropic/claude-opus-5
 ```
 
+**Claude models through OpenRouter:** OpenRouter forwards requests to Anthropic
+largely unchanged, so Anthropic's stricter validation applies. Autohand omits
+`temperature` for the Claude 5 family and Opus 4.7/4.8 (which reject sampling
+parameters), requests extended thinking through OpenRouter's `reasoning` field,
+and repairs turn shapes that Anthropic rejects — blank turns, blank tool
+results, and tool calls or results orphaned by context compaction. When an
+upstream provider does reject a request, the provider's own error text is now
+included rather than only OpenRouter's generic "Provider returned error".
+
+---
+
+### Anthropic
+
+The native Anthropic provider uses the official TypeScript SDK and the Messages API rather than routing Claude requests through the OpenAI-compatible provider layer. Native system prompts, images, tool calls, tool results, structured output, adaptive thinking, and Anthropic usage fields are translated at the provider boundary.
+
+1. Create a key in the [Anthropic Console](https://console.anthropic.com/settings/keys).
+2. Run `autohand --setup` or select Anthropic from `/model`.
+3. Choose a model from the bundled and remotely updated catalog.
+
+```json
+{
+  "provider": "anthropic",
+  "anthropic": {
+    "apiKey": "sk-ant-your-key-here",
+    "baseUrl": "https://api.anthropic.com",
+    "model": "claude-sonnet-5",
+    "contextWindow": 1000000
+  }
+}
+```
+
+Current catalog choices include `claude-sonnet-5`, `claude-opus-5`, `claude-fable-5`, `claude-opus-4-8`, `claude-sonnet-4-6`, and `claude-haiku-4-5`. Use the unprefixed model IDs for the native provider; use `anthropic/...` IDs only with OpenRouter.
+
+**Request behavior:**
+
+- Anthropic rejects unsupported parameters with a `400` rather than ignoring them, so Autohand gates each optional parameter on the selected model. `temperature` is dropped for the Claude 5 family and Opus 4.7/4.8, `output_config.effort` is clamped to `high` on Opus 4.5/4.6 and omitted for models without effort support, and `thinking` is omitted entirely for models where thinking is always on.
+- Prompt caching is on by default. The stable prefix (tools, system prompt, and prior turns) is cached automatically; check `/usage` for cache read and write counts.
+- Thinking blocks are captured from each response and replayed unchanged on the following request. Anthropic validates their ordering and signatures, so they are never edited or reconstructed.
+- Only the leading system messages become the top-level `system` prompt. Later system notes stay in conversation order as `<system-reminder>` blocks so the cached prefix is not invalidated mid-session.
+- Requests use a 10-minute floor for the client timeout because the Anthropic SDK applies `network.timeout` to the whole request, not just the response headers. A larger `network.timeout` is honored; a smaller one is not.
+
+See Anthropic's [model overview](https://platform.claude.com/docs/en/about-claude/models/overview) and [TypeScript Messages API](https://platform.claude.com/docs/en/api/typescript/messages/create) for the upstream contracts.
+
 ---
 
 ### OpenAI
@@ -238,8 +286,8 @@ LLM Gateway provides a unified API for multiple LLM providers with a single inte
 | `gpt-5` | OpenAI |
 | `gpt-5-mini` | OpenAI |
 | `gpt-4-turbo` | OpenAI |
-| `claude-5-sonnet` | Anthropic |
-| `claude-5-haiku` | Anthropic |
+| `claude-sonnet-5` | Anthropic |
+| `claude-haiku-4-5` | Anthropic |
 | `gemini-3.0-pro` | Google |
 | `gemini-3.0-flash` | Google |
 
@@ -608,7 +656,7 @@ Use the `/model` command to switch providers or models:
 ```
 /model                           # List available models
 /model gpt-5                     # Switch to GPT-5
-/model anthropic/claude-5-opus   # Switch to Claude Opus
+/model anthropic/claude-opus-5   # Switch OpenRouter to Claude Opus 5
 ```
 
 When you pick `openai`, Autohand now lets you choose between `API key` and `ChatGPT account` authentication.
@@ -627,10 +675,10 @@ Update `~/.autohand/config.json`:
 
 ```json
 {
-  "provider": "llmgateway",
-  "llmgateway": {
-    "apiKey": "your-key",
-    "model": "claude-5-sonnet"
+  "provider": "anthropic",
+  "anthropic": {
+    "apiKey": "sk-ant-your-key",
+    "model": "claude-sonnet-5"
   }
 }
 ```

docs/providers.md

@@ -73,6 +73,7 @@
   },
   "dependencies": {
     "@agentclientprotocol/sdk": "1.3.0",
+    "@anthropic-ai/sdk": "^0.120.0",
     "@aws-sdk/client-bedrock": "^3.1106.0",
     "@aws-sdk/client-bedrock-runtime": "^3.1086.0",
     "@aws-sdk/credential-providers": "^3.1090.0",

package.json

@@ -59,6 +59,7 @@ const SETTING_KEY_ALIASES: Record<string, string> = {
 
 const CONFIG_PROVIDER_NAMES: readonly BuiltInProviderName[] = [
   'openrouter',
+  'anthropic',
   'ollama',
   'llamacpp',
   'openai',

src/commands/settings.ts

@@ -28,13 +28,14 @@ import { autoInitTheme, configureThemeSources, themeExists } from "./ui/theme/in
 import { loadLocalProjectSettings, type LocalProjectSettings } from "./permissions/localProjectPermissions.js";
 import { isAwsBedrockProviderEnabled } from "./features/featureRegistry.js";
 import { getCustomProviderConfig, isCustomProviderName } from "./providers/customProviders.js";
-import { getProviderDefaultModel, getProviderModelOptions, getProviderRuntimeDefaultModel } from "./providers/modelCatalog.js";
+import { getProviderDefaultModel, getProviderModelOptions, getProviderRuntimeDefaultModel, normalizeOpenRouterModelId } from "./providers/modelCatalog.js";
 
 const DEFAULT_CONFIG_PATH = AUTOHAND_FILES.configJson;
 const TOML_CONFIG_PATH = AUTOHAND_FILES.configToml;
 const YAML_CONFIG_PATH = AUTOHAND_FILES.configYaml;
 const YML_CONFIG_PATH = AUTOHAND_FILES.configYml;
 const DEFAULT_BASE_URL = "https://openrouter.ai/api/v1";
+const DEFAULT_ANTHROPIC_URL = "https://api.anthropic.com";
 const DEFAULT_OLLAMA_URL = "http://localhost:11434";
 const DEFAULT_LLAMACPP_URL = "http://localhost:8080";
 const DEFAULT_OPENAI_URL = "https://api.openai.com/v1";
@@ -85,6 +86,7 @@ function normalizeProviderName(provider: unknown): ProviderName | undefined {
   const validProviders: readonly BuiltInProviderName[] = [
     "autohandai",
     "openrouter",
+    "anthropic",
     "ollama",
     "llamacpp",
     "openai",
@@ -749,7 +751,18 @@ function normalizeConfig(
 
   if (isModernConfig(config)) {
     const provider = normalizeProviderName(config.provider) ?? "openrouter";
-    return { ...config, provider };
+    return {
+      ...config,
+      provider,
+      ...(config.openrouter
+        ? {
+            openrouter: {
+              ...config.openrouter,
+              model: normalizeOpenRouterModelId(config.openrouter.model),
+            },
+          }
+        : {}),
+    };
   }
 
   if (isLegacyConfig(config)) {
@@ -758,7 +771,7 @@ function normalizeConfig(
       openrouter: {
         apiKey: config.api_key ?? "replace-me",
         baseUrl: config.base_url ?? DEFAULT_BASE_URL,
-        model: getProviderDefaultModel("openrouter", "anthropic/claude-4-sonnet"),
+        model: getProviderDefaultModel("openrouter", "anthropic/claude-sonnet-5"),
       },
       workspace: {
         defaultRoot: process.cwd(),
@@ -784,6 +797,7 @@ function isModernConfig(
 ): config is AutohandConfig {
   return (
     typeof (config as AutohandConfig).openrouter === "object" ||
+    typeof (config as AutohandConfig).anthropic === "object" ||
     typeof (config as AutohandConfig).blueprintLocal === "object" ||
     typeof (config as AutohandConfig).autohandai === "object" ||
     typeof (config as AutohandConfig).ollama === "object" ||
@@ -1107,6 +1121,7 @@ export function getProviderConfig(
   const configByProvider: Record<BuiltInProviderName, ProviderSettings | undefined> = {
     autohandai: config.autohandai,
     openrouter: config.openrouter,
+    anthropic: config.anthropic,
     ollama: config.ollama,
     llamacpp: config.llamacpp,
     openai: config.openai,
@@ -1177,6 +1192,7 @@ export function getProviderConfig(
     }
   } else if (
     builtInProvider === "openrouter" ||
+    builtInProvider === "anthropic" ||
     builtInProvider === "llmgateway" ||
     builtInProvider === "zai" ||
     builtInProvider === "sakana" ||
@@ -1225,6 +1241,7 @@ function defaultBaseUrlFor(
   port?: number,
 ): string | undefined {
   if (provider === "openrouter") return DEFAULT_BASE_URL;
+  if (provider === "anthropic") return DEFAULT_ANTHROPIC_URL;
   if (provider === "autohandai") return DEFAULT_AUTOHAND_AI_URL;
   if (provider === "llmgateway") return DEFAULT_LLMGATEWAY_URL;
   if (provider === "zai") return DEFAULT_ZAI_URL;