trail-code 0.0.0-stage → 0.28.0-semantic.3

package/dist/ai/llm/anthropic.js

@@ -0,0 +1,144 @@
+/**
+ * Native Anthropic transport. Wraps `@anthropic-ai/sdk` (the Messages API), so
+ * Claude is a first-class provider rather than something tunneled through an
+ * OpenAI-compatible endpoint.
+ *
+ * The Messages API differs from Chat Completions in ways this adapter absorbs so
+ * the rest of graft never sees them:
+ *   - `system` is a top-level parameter, not a message.
+ *   - tool results ride inside a USER turn, and all results answering one
+ *     assistant turn must be coalesced into a single user message.
+ *   - `max_tokens` is required (callers get a default).
+ *   - `temperature` is rejected by current models, so it is never forwarded.
+ *   - structured output is a forced tool (there is no `response_format`).
+ *   - assistant turns are replayed verbatim (via `providerRaw`) so thinking-block
+ *     signatures survive a multi-turn loop.
+ */
+import Anthropic from "@anthropic-ai/sdk";
+import { transportRetries } from "./types.js";
+const PROVIDER = "anthropic";
+const JSON_TOOL = "emit_json";
+const DEFAULT_MAX_TOKENS = 4096;
+const cc = (on) => (on ? { cache_control: { type: "ephemeral" } } : {});
+export class AnthropicChatModel {
+    label;
+    client;
+    model;
+    constructor(opts) {
+        this.model = opts.model;
+        this.label = opts.label ?? `${PROVIDER}:${opts.model}`;
+        this.client =
+            opts.client ??
+                new Anthropic({ apiKey: opts.apiKey, baseURL: opts.baseUrl, maxRetries: transportRetries() });
+    }
+    async create(req) {
+        const system = [];
+        const messages = [];
+        for (const m of req.messages) {
+            if (m.role === "system") {
+                system.push({ type: "text", text: m.content, ...cc(m.cacheBreakpoint) });
+            }
+            else if (m.role === "user") {
+                messages.push({ role: "user", content: [{ type: "text", text: m.content, ...cc(m.cacheBreakpoint) }] });
+            }
+            else if (m.role === "tool") {
+                // Tool results live in a user turn; coalesce consecutive results together.
+                const block = {
+                    type: "tool_result",
+                    tool_use_id: m.toolCallId ?? "",
+                    content: [{ type: "text", text: m.content }],
+                    ...cc(m.cacheBreakpoint),
+                };
+                const last = messages[messages.length - 1];
+                if (last?.role === "user" && Array.isArray(last.content)) {
+                    last.content.push(block);
+                }
+                else {
+                    messages.push({ role: "user", content: [block] });
+                }
+            }
+            else {
+                messages.push(this.assistantParam(m));
+            }
+        }
+        const tools = req.tools?.map(toAnthropicTool);
+        const params = {
+            model: this.model,
+            max_tokens: req.maxTokens ?? DEFAULT_MAX_TOKENS,
+            messages,
+            ...(system.length ? { system } : {}),
+        };
+        // temperature is intentionally NOT forwarded — current models reject it.
+        const fmt = req.responseFormat ?? { kind: "text" };
+        if (fmt.kind === "json") {
+            params.tools = [
+                ...(tools ?? []),
+                { name: JSON_TOOL, description: "Return the answer as a JSON object.", input_schema: { type: "object" } },
+            ];
+            params.tool_choice = { type: "tool", name: JSON_TOOL };
+        }
+        else if (fmt.kind === "tool") {
+            params.tools = tools;
+            params.tool_choice = { type: "tool", name: fmt.name };
+        }
+        else if (tools) {
+            params.tools = tools;
+        }
+        const resp = await this.client.messages.create(params);
+        return this.fromResponse(resp, fmt.kind);
+    }
+    /** Reconstruct an assistant turn, replaying the original blocks when we made them. */
+    assistantParam(m) {
+        if (m.providerRaw?.provider === PROVIDER)
+            return m.providerRaw.raw;
+        const content = [];
+        if (m.content)
+            content.push({ type: "text", text: m.content });
+        for (const tc of m.toolCalls ?? []) {
+            content.push({ type: "tool_use", id: tc.id, name: tc.name, input: tc.args });
+        }
+        return { role: "assistant", content };
+    }
+    fromResponse(resp, format) {
+        let text = "";
+        let toolCalls = [];
+        for (const block of resp.content) {
+            if (block.type === "text")
+                text += block.text;
+            else if (block.type === "tool_use")
+                toolCalls.push({ id: block.id, name: block.name, args: block.input });
+        }
+        if (format === "json") {
+            const jsonCall = toolCalls.find((c) => c.name === JSON_TOOL);
+            if (jsonCall)
+                text = JSON.stringify(jsonCall.args);
+            toolCalls = toolCalls.filter((c) => c.name !== JSON_TOOL);
+        }
+        return {
+            text,
+            toolCalls,
+            usage: normalizeUsage(resp.usage),
+            stopReason: resp.stop_reason ?? null,
+            // Replay the raw content verbatim so thinking-block signatures survive.
+            assistant: {
+                role: "assistant",
+                content: text,
+                toolCalls: toolCalls.length ? toolCalls : undefined,
+                providerRaw: { provider: PROVIDER, raw: { role: "assistant", content: resp.content } },
+            },
+        };
+    }
+}
+function toAnthropicTool(t) {
+    return { name: t.name, description: t.description, input_schema: t.parameters };
+}
+/** Anthropic reports uncached input in `input_tokens` and cache tokens separately. */
+function normalizeUsage(u) {
+    return {
+        input: u.input_tokens ?? 0,
+        output: u.output_tokens ?? 0,
+        cacheRead: u.cache_read_input_tokens ?? 0,
+        cacheCreate: u.cache_creation_input_tokens ?? 0,
+    };
+}
+//# sourceMappingURL=anthropic.js.map

package/dist/ai/llm/anthropic.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"anthropic.js","sourceRoot":"","sources":["../../../src/ai/llm/anthropic.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AACH,OAAO,SAAS,MAAM,mBAAmB,CAAC;AAC1C,OAAO,EAAE,gBAAgB,EAAE,MAAM,YAAY,CAAC;AAG9C,MAAM,QAAQ,GAAG,WAAW,CAAC;AAC7B,MAAM,SAAS,GAAG,WAAW,CAAC;AAC9B,MAAM,kBAAkB,GAAG,IAAI,CAAC;AAYhC,MAAM,EAAE,GAAG,CAAC,EAAuB,EAAgB,EAAE,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,aAAa,EAAE,EAAE,IAAI,EAAE,WAAW,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;AAE3G,MAAM,OAAO,kBAAkB;IACpB,KAAK,CAAS;IACf,MAAM,CAAY;IAClB,KAAK,CAAS;IAEtB,YAAY,IAA+B;QACzC,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC;QACxB,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,KAAK,IAAI,GAAG,QAAQ,IAAI,IAAI,CAAC,KAAK,EAAE,CAAC;QACvD,IAAI,CAAC,MAAM;YACT,IAAI,CAAC,MAAM;gBACX,IAAI,SAAS,CAAC,EAAE,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE,UAAU,EAAE,gBAAgB,EAAE,EAAE,CAAC,CAAC;IAClG,CAAC;IAED,KAAK,CAAC,MAAM,CAAC,GAAgB;QAC3B,MAAM,MAAM,GAA+B,EAAE,CAAC;QAC9C,MAAM,QAAQ,GAA6B,EAAE,CAAC;QAE9C,KAAK,MAAM,CAAC,IAAI,GAAG,CAAC,QAAQ,EAAE,CAAC;YAC7B,IAAI,CAAC,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;gBACxB,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC,OAAO,EAAE,GAAG,EAAE,CAAC,CAAC,CAAC,eAAe,CAAC,EAAE,CAAC,CAAC;YAC3E,CAAC;iBAAM,IAAI,CAAC,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;gBAC7B,QAAQ,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC,OAAO,EAAE,GAAG,EAAE,CAAC,CAAC,CAAC,eAAe,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC;YAC1G,CAAC;iBAAM,IAAI,CAAC,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;gBAC7B,2EAA2E;gBAC3E,MAAM,KAAK,GAAmC;oBAC5C,IAAI,EAAE,aAAa;oBACnB,WAAW,EAAE,CAAC,CAAC,UAAU,IAAI,EAAE;oBAC/B,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC;oBAC5C,GAAG,EAAE,CAAC,CAAC,CAAC,eAAe,CAAC;iBACzB,CAAC;gBACF,MAAM,IAAI,GAAG,QAAQ,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;gBAC3C,IAAI,IAAI,EAAE,IAAI,KAAK,MAAM,IAAI,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC;oBACxD,IAAI,CAAC,OAAyC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;gBAC9D,CAAC;qBAAM,CAAC;oBACN,QAAQ,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;gBACpD,CAAC;YACH,CAAC;iBAAM,CAAC;gBACN,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,cAAc,CAAC,CAAC,CAAC,CAAC,CAAC;YACxC,CAAC;QACH,CAAC;QAED,MAAM,KAAK,GAAG,GAAG,CAAC,KAAK,EAAE,GAAG,CAAC,eAAe,CAAC,CAAC;QAC9C,MAAM,MAAM,GAA8C;YACxD,KAAK,EAAE,IAAI,CAAC,KAAK;YACjB,UAAU,EAAE,GAAG,CAAC,SAAS,IAAI,kBAAkB;YAC/C,QAAQ;YACR,GAAG,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;SACrC,CAAC;QACF,yEAAyE;QAEzE,MAAM,GAAG,GAAG,GAAG,CAAC,cAAc,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC;QACnD,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YACxB,MAAM,CAAC,KAAK,GAAG;gBACb,GAAG,CAAC,KAAK,IAAI,EAAE,CAAC;gBAChB,EAAE,IAAI,EAAE,SAAS,EAAE,WAAW,EAAE,qCAAqC,EAAE,YAAY,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,EAAE;aAC1G,CAAC;YACF,MAAM,CAAC,WAAW,GAAG,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,SAAS,EAAE,CAAC;QACzD,CAAC;aAAM,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YAC/B,MAAM,CAAC,KAAK,GAAG,KAAK,CAAC;YACrB,MAAM,CAAC,WAAW,GAAG,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,GAAG,CAAC,IAAI,EAAE,CAAC;QACxD,CAAC;aAAM,IAAI,KAAK,EAAE,CAAC;YACjB,MAAM,CAAC,KAAK,GAAG,KAAK,CAAC;QACvB,CAAC;QAED,MAAM,IAAI,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;QACvD,OAAO,IAAI,CAAC,YAAY,CAAC,IAAI,EAAE,GAAG,CAAC,IAAI,CAAC,CAAC;IAC3C,CAAC;IAED,sFAAsF;IAC9E,cAAc,CAAC,CAAU;QAC/B,IAAI,CAAC,CAAC,WAAW,EAAE,QAAQ,KAAK,QAAQ;YAAE,OAAO,CAAC,CAAC,WAAW,CAAC,GAA6B,CAAC;QAC7F,MAAM,OAAO,GAAkC,EAAE,CAAC;QAClD,IAAI,CAAC,CAAC,OAAO;YAAE,OAAO,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC;QAC/D,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,SAAS,IAAI,EAAE,EAAE,CAAC;YACnC,OAAO,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,UAAU,EAAE,EAAE,EAAE,EAAE,CAAC,EAAE,EAAE,IAAI,EAAE,EAAE,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC,IAA+B,EAAE,CAAC,CAAC;QAC1G,CAAC;QACD,OAAO,EAAE,IAAI,EAAE,WAAW,EAAE,OAAO,EAAE,CAAC;IACxC,CAAC;IAEO,YAAY,CAAC,IAAgC,EAAE,MAAgC;QACrF,IAAI,IAAI,GAAG,EAAE,CAAC;QACd,IAAI,SAAS,GAAe,EAAE,CAAC;QAC/B,KAAK,MAAM,KAAK,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YACjC,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM;gBAAE,IAAI,IAAI,KAAK,CAAC,IAAI,CAAC;iBACzC,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU;gBAAE,SAAS,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,KAAK,CAAC,EAAE,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,IAAI,EAAE,KAAK,CAAC,KAAK,EAAE,CAAC,CAAC;QAC5G,CAAC;QAED,IAAI,MAAM,KAAK,MAAM,EAAE,CAAC;YACtB,MAAM,QAAQ,GAAG,SAAS,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,SAAS,CAAC,CAAC;YAC7D,IAAI,QAAQ;gBAAE,IAAI,GAAG,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;YACnD,SAAS,GAAG,SAAS,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,SAAS,CAAC,CAAC;QAC5D,CAAC;QAED,OAAO;YACL,IAAI;YACJ,SAAS;YACT,KAAK,EAAE,cAAc,CAAC,IAAI,CAAC,KAAK,CAAC;YACjC,UAAU,EAAE,IAAI,CAAC,WAAW,IAAI,IAAI;YACpC,wEAAwE;YACxE,SAAS,EAAE;gBACT,IAAI,EAAE,WAAW;gBACjB,OAAO,EAAE,IAAI;gBACb,SAAS,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,SAAS;gBACnD,WAAW,EAAE,EAAE,QAAQ,EAAE,QAAQ,EAAE,GAAG,EAAE,EAAE,IAAI,EAAE,WAAW,EAAE,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE,EAAE;aACvF;SACF,CAAC;IACJ,CAAC;CACF;AAED,SAAS,eAAe,CAAC,CAAW;IAClC,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,WAAW,EAAE,CAAC,CAAC,WAAW,EAAE,YAAY,EAAE,CAAC,CAAC,UAAwC,EAAE,CAAC;AAChH,CAAC;AAED,sFAAsF;AACtF,SAAS,cAAc,CAAC,CAAkB;IACxC,OAAO;QACL,KAAK,EAAE,CAAC,CAAC,YAAY,IAAI,CAAC;QAC1B,MAAM,EAAE,CAAC,CAAC,aAAa,IAAI,CAAC;QAC5B,SAAS,EAAE,CAAC,CAAC,uBAAuB,IAAI,CAAC;QACzC,WAAW,EAAE,CAAC,CAAC,2BAA2B,IAAI,CAAC;KAChD,CAAC;AACJ,CAAC"}

package/dist/ai/llm/factory.d.ts

@@ -0,0 +1,28 @@
+/**
+ * One place that turns resolved config into a {@link ChatModel}. `provider` names
+ * the WIRE FORMAT, not a vendor: `openai` speaks the OpenAI-compatible API (point
+ * `baseUrl` at OpenRouter, Fireworks, a LiteLLM proxy, Groq, a local server, …),
+ * `anthropic` speaks the native Messages API. Adding a vendor is a base URL, not
+ * a code change; adding a wire format is one new adapter here.
+ *
+ * `litellm` is a convenience over `openai`: same wire format, but pointed at a
+ * LiteLLM proxy by default and paired with `/v1/models` auto-discovery
+ * (see litellm.ts), so one endpoint reaches 100+ providers.
+ *
+ * `orcarouter` is the same kind of convenience over `openai`, pointed at the
+ * OrcaRouter AI gateway by default (see orcarouter.ts) so its users get the
+ * gateway's routing, failover, and guardrails behind a named provider instead
+ * of a bare custom base URL.
+ */
+import type { ChatModel } from "./types.js";
+export type ProviderKind = "openai" | "anthropic" | "litellm" | "orcarouter";
+export interface ChatModelConfig {
+    provider: ProviderKind;
+    apiKey: string;
+    model: string;
+    baseUrl?: string;
+    /** Extra default headers for OpenAI-compatible endpoints (e.g. OpenRouter `X-Title`). */
+    headers?: Record<string, string>;
+}
+export declare function createChatModel(cfg: ChatModelConfig): ChatModel;
+//# sourceMappingURL=factory.d.ts.map

package/dist/ai/llm/factory.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"factory.d.ts","sourceRoot":"","sources":["../../../src/ai/llm/factory.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AACH,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,YAAY,CAAC;AAM5C,MAAM,MAAM,YAAY,GAAG,QAAQ,GAAG,WAAW,GAAG,SAAS,GAAG,YAAY,CAAC;AAE7E,MAAM,WAAW,eAAe;IAC9B,QAAQ,EAAE,YAAY,CAAC;IACvB,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;IACd,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,yFAAyF;IACzF,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CAClC;AAED,wBAAgB,eAAe,CAAC,GAAG,EAAE,eAAe,GAAG,SAAS,CA8B/D"}

package/dist/ai/llm/factory.js

@@ -0,0 +1,36 @@
+import { OpenAIChatModel } from "./openai.js";
+import { AnthropicChatModel } from "./anthropic.js";
+import { LiteLLMChatModel } from "./litellm.js";
+import { OrcaRouterChatModel } from "./orcarouter.js";
+export function createChatModel(cfg) {
+    switch (cfg.provider) {
+        case "anthropic":
+            return new AnthropicChatModel({ apiKey: cfg.apiKey, model: cfg.model, baseUrl: cfg.baseUrl });
+        case "openai":
+            return new OpenAIChatModel({
+                apiKey: cfg.apiKey,
+                model: cfg.model,
+                baseUrl: cfg.baseUrl,
+                headers: cfg.headers,
+            });
+        case "litellm":
+            return new LiteLLMChatModel({
+                apiKey: cfg.apiKey,
+                model: cfg.model,
+                baseUrl: cfg.baseUrl,
+                headers: cfg.headers,
+            });
+        case "orcarouter":
+            return new OrcaRouterChatModel({
+                apiKey: cfg.apiKey,
+                model: cfg.model,
+                baseUrl: cfg.baseUrl,
+                headers: cfg.headers,
+            });
+        default: {
+            const _exhaustive = cfg.provider;
+            throw new Error(`unknown provider: ${String(_exhaustive)}`);
+        }
+    }
+}
+//# sourceMappingURL=factory.js.map

package/dist/ai/llm/factory.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"factory.js","sourceRoot":"","sources":["../../../src/ai/llm/factory.ts"],"names":[],"mappings":"AAiBA,OAAO,EAAE,eAAe,EAAE,MAAM,aAAa,CAAC;AAC9C,OAAO,EAAE,kBAAkB,EAAE,MAAM,gBAAgB,CAAC;AACpD,OAAO,EAAE,gBAAgB,EAAE,MAAM,cAAc,CAAC;AAChD,OAAO,EAAE,mBAAmB,EAAE,MAAM,iBAAiB,CAAC;AAatD,MAAM,UAAU,eAAe,CAAC,GAAoB;IAClD,QAAQ,GAAG,CAAC,QAAQ,EAAE,CAAC;QACrB,KAAK,WAAW;YACd,OAAO,IAAI,kBAAkB,CAAC,EAAE,MAAM,EAAE,GAAG,CAAC,MAAM,EAAE,KAAK,EAAE,GAAG,CAAC,KAAK,EAAE,OAAO,EAAE,GAAG,CAAC,OAAO,EAAE,CAAC,CAAC;QAChG,KAAK,QAAQ;YACX,OAAO,IAAI,eAAe,CAAC;gBACzB,MAAM,EAAE,GAAG,CAAC,MAAM;gBAClB,KAAK,EAAE,GAAG,CAAC,KAAK;gBAChB,OAAO,EAAE,GAAG,CAAC,OAAO;gBACpB,OAAO,EAAE,GAAG,CAAC,OAAO;aACrB,CAAC,CAAC;QACL,KAAK,SAAS;YACZ,OAAO,IAAI,gBAAgB,CAAC;gBAC1B,MAAM,EAAE,GAAG,CAAC,MAAM;gBAClB,KAAK,EAAE,GAAG,CAAC,KAAK;gBAChB,OAAO,EAAE,GAAG,CAAC,OAAO;gBACpB,OAAO,EAAE,GAAG,CAAC,OAAO;aACrB,CAAC,CAAC;QACL,KAAK,YAAY;YACf,OAAO,IAAI,mBAAmB,CAAC;gBAC7B,MAAM,EAAE,GAAG,CAAC,MAAM;gBAClB,KAAK,EAAE,GAAG,CAAC,KAAK;gBAChB,OAAO,EAAE,GAAG,CAAC,OAAO;gBACpB,OAAO,EAAE,GAAG,CAAC,OAAO;aACrB,CAAC,CAAC;QACL,OAAO,CAAC,CAAC,CAAC;YACR,MAAM,WAAW,GAAU,GAAG,CAAC,QAAQ,CAAC;YACxC,MAAM,IAAI,KAAK,CAAC,qBAAqB,MAAM,CAAC,WAAW,CAAC,EAAE,CAAC,CAAC;QAC9D,CAAC;IACH,CAAC;AACH,CAAC"}

package/dist/ai/llm/litellm.d.ts

@@ -0,0 +1,34 @@
+/**
+ * LiteLLM transport. A LiteLLM proxy speaks the OpenAI-compatible wire format,
+ * so this reuses the OpenAI adapter's request/response translation and simply
+ * points it at the proxy. The value over a bare `openai` provider + `baseUrl`
+ * is a first-class, self-describing provider: a sensible default proxy address
+ * and `/v1/models` auto-discovery (see {@link listLiteLLMModels}), so callers
+ * can enumerate whatever models the proxy is configured to serve — reaching
+ * 100+ providers (OpenAI, Anthropic, Gemini, Bedrock, Vertex, …) through one
+ * endpoint without a per-vendor code path.
+ */
+import OpenAI from "openai";
+import { OpenAIChatModel, type OpenAIChatModelOptions } from "./openai.js";
+/** Conventional local address of a self-hosted LiteLLM proxy. */
+export declare const DEFAULT_LITELLM_BASE_URL = "http://localhost:4000";
+export type LiteLLMChatModelOptions = OpenAIChatModelOptions;
+/**
+ * ChatModel backed by a LiteLLM proxy. Inherits every translation detail from
+ * {@link OpenAIChatModel} (the proxy is OpenAI-compatible) and only changes the
+ * defaults: the proxy base URL and a `litellm:<model>` label.
+ */
+export declare class LiteLLMChatModel extends OpenAIChatModel {
+    constructor(opts: LiteLLMChatModelOptions);
+}
+/**
+ * Auto-discover the models a LiteLLM proxy is serving via `GET /v1/models`.
+ * Returns the model ids so a caller can present/validate them instead of
+ * hardcoding a model string. A stub `client` may be injected for tests.
+ */
+export declare function listLiteLLMModels(opts: {
+    apiKey: string;
+    baseUrl?: string;
+    client?: OpenAI;
+}): Promise<string[]>;
+//# sourceMappingURL=litellm.d.ts.map

package/dist/ai/llm/litellm.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"litellm.d.ts","sourceRoot":"","sources":["../../../src/ai/llm/litellm.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AACH,OAAO,MAAM,MAAM,QAAQ,CAAC;AAC5B,OAAO,EAAE,eAAe,EAAE,KAAK,sBAAsB,EAAE,MAAM,aAAa,CAAC;AAG3E,iEAAiE;AACjE,eAAO,MAAM,wBAAwB,0BAA0B,CAAC;AAEhE,MAAM,MAAM,uBAAuB,GAAG,sBAAsB,CAAC;AAE7D;;;;GAIG;AACH,qBAAa,gBAAiB,SAAQ,eAAe;gBACvC,IAAI,EAAE,uBAAuB;CAO1C;AAED;;;;GAIG;AACH,wBAAsB,iBAAiB,CAAC,IAAI,EAAE;IAC5C,MAAM,EAAE,MAAM,CAAC;IACf,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,MAAM,CAAC,EAAE,MAAM,CAAC;CACjB,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,CAUpB"}

package/dist/ai/llm/litellm.js

@@ -0,0 +1,45 @@
+/**
+ * LiteLLM transport. A LiteLLM proxy speaks the OpenAI-compatible wire format,
+ * so this reuses the OpenAI adapter's request/response translation and simply
+ * points it at the proxy. The value over a bare `openai` provider + `baseUrl`
+ * is a first-class, self-describing provider: a sensible default proxy address
+ * and `/v1/models` auto-discovery (see {@link listLiteLLMModels}), so callers
+ * can enumerate whatever models the proxy is configured to serve — reaching
+ * 100+ providers (OpenAI, Anthropic, Gemini, Bedrock, Vertex, …) through one
+ * endpoint without a per-vendor code path.
+ */
+import OpenAI from "openai";
+import { OpenAIChatModel } from "./openai.js";
+import { transportRetries } from "./types.js";
+/** Conventional local address of a self-hosted LiteLLM proxy. */
+export const DEFAULT_LITELLM_BASE_URL = "http://localhost:4000";
+/**
+ * ChatModel backed by a LiteLLM proxy. Inherits every translation detail from
+ * {@link OpenAIChatModel} (the proxy is OpenAI-compatible) and only changes the
+ * defaults: the proxy base URL and a `litellm:<model>` label.
+ */
+export class LiteLLMChatModel extends OpenAIChatModel {
+    constructor(opts) {
+        super({
+            ...opts,
+            baseUrl: opts.baseUrl ?? DEFAULT_LITELLM_BASE_URL,
+            label: opts.label ?? `litellm:${opts.model}`,
+        });
+    }
+}
+/**
+ * Auto-discover the models a LiteLLM proxy is serving via `GET /v1/models`.
+ * Returns the model ids so a caller can present/validate them instead of
+ * hardcoding a model string. A stub `client` may be injected for tests.
+ */
+export async function listLiteLLMModels(opts) {
+    const client = opts.client ??
+        new OpenAI({
+            apiKey: opts.apiKey,
+            baseURL: opts.baseUrl ?? DEFAULT_LITELLM_BASE_URL,
+            maxRetries: transportRetries(),
+        });
+    const res = await client.models.list();
+    return res.data.map((m) => m.id);
+}
+//# sourceMappingURL=litellm.js.map

package/dist/ai/llm/litellm.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"litellm.js","sourceRoot":"","sources":["../../../src/ai/llm/litellm.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AACH,OAAO,MAAM,MAAM,QAAQ,CAAC;AAC5B,OAAO,EAAE,eAAe,EAA+B,MAAM,aAAa,CAAC;AAC3E,OAAO,EAAE,gBAAgB,EAAE,MAAM,YAAY,CAAC;AAE9C,iEAAiE;AACjE,MAAM,CAAC,MAAM,wBAAwB,GAAG,uBAAuB,CAAC;AAIhE;;;;GAIG;AACH,MAAM,OAAO,gBAAiB,SAAQ,eAAe;IACnD,YAAY,IAA6B;QACvC,KAAK,CAAC;YACJ,GAAG,IAAI;YACP,OAAO,EAAE,IAAI,CAAC,OAAO,IAAI,wBAAwB;YACjD,KAAK,EAAE,IAAI,CAAC,KAAK,IAAI,WAAW,IAAI,CAAC,KAAK,EAAE;SAC7C,CAAC,CAAC;IACL,CAAC;CACF;AAED;;;;GAIG;AACH,MAAM,CAAC,KAAK,UAAU,iBAAiB,CAAC,IAIvC;IACC,MAAM,MAAM,GACV,IAAI,CAAC,MAAM;QACX,IAAI,MAAM,CAAC;YACT,MAAM,EAAE,IAAI,CAAC,MAAM;YACnB,OAAO,EAAE,IAAI,CAAC,OAAO,IAAI,wBAAwB;YACjD,UAAU,EAAE,gBAAgB,EAAE;SAC/B,CAAC,CAAC;IACL,MAAM,GAAG,GAAG,MAAM,MAAM,CAAC,MAAM,CAAC,IAAI,EAAE,CAAC;IACvC,OAAO,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;AACnC,CAAC"}

package/dist/ai/llm/openai.d.ts

@@ -0,0 +1,145 @@
+/**
+ * OpenAI-compatible transport. Wraps the `openai` SDK pointed at any
+ * OpenAI-compatible endpoint (OpenAI, OpenRouter, Fireworks, a LiteLLM proxy,
+ * Groq, Together, DeepSeek, a local server, …) — the user picks the endpoint
+ * with `baseUrl` and authenticates with their own key.
+ *
+ * This adapter reproduces graft's historical wire behavior exactly: temperature
+ * is forwarded, cache breakpoints become `cache_control` content parts (which
+ * OpenRouter forwards to Anthropic), and cached tokens are subtracted out of the
+ * input count so {@link Usage.input} is uncached-only.
+ *
+ * Past that, everything here is tolerance rather than preference. An
+ * OpenAI-compatible endpoint is a gateway whose own upstream calls can fail
+ * while the HTTP status stays 200, and whose support for a forced `tool_choice`
+ * — the one structured-output mechanism graft can rely on across providers —
+ * ranges from exact to none. Each way that goes wrong gets its own narrow
+ * recovery: {@link ProviderResponseError} for a 200 that is not a completion,
+ * and a bounded rung ladder in {@link OpenAIChatModel.structured} for a
+ * structured reply the endpoint answers some other way. None of it changes the
+ * bytes on the wire for an endpoint that behaves.
+ */
+import OpenAI from "openai";
+import type { ChatModel, ChatRequest, ChatResponse } from "./types.js";
+export interface OpenAIChatModelOptions {
+    apiKey: string;
+    model: string;
+    baseUrl?: string;
+    /** Stable manifest label; defaults to `openai:<model>`. */
+    label?: string;
+    /** Extra default headers (e.g. OpenRouter's `X-Title`). */
+    headers?: Record<string, string>;
+    /** Inject a pre-built client (tests pass a stub; production omits it). */
+    client?: OpenAI;
+    /** Wait between retries of a 200-that-is-an-error (tests pass a no-op). */
+    sleep?: (ms: number) => Promise<void>;
+}
+/**
+ * A 200 that is not a completion.
+ *
+ * A proxy in front of a model can answer HTTP 200 with
+ * `{"error": {...}}` and no `choices` when the upstream call fails, because the
+ * failure belongs to a server the proxy itself did not talk to. The SDK's retry
+ * policy only ever sees the status, so it returns that body as a success, and
+ * the first `choices[0]` used to turn it into a TypeError naming nothing a user
+ * could act on. Retryable by construction — the request was never served — and
+ * carrying the provider's own message and code so the failure that does reach
+ * the caller says what the endpoint said.
+ */
+export declare class ProviderResponseError extends Error {
+    /** The provider's own error code, when the body carried one. */
+    readonly code?: string | number | undefined;
+    /** The status the provider actually sent. 200: that is the whole surprise. */
+    readonly status = 200;
+    /** Always true — the same request can succeed on a later attempt. */
+    readonly retryable = true;
+    constructor(message: string, 
+    /** The provider's own error code, when the body carried one. */
+    code?: string | number | undefined);
+}
+export declare class OpenAIChatModel implements ChatModel {
+    readonly label: string;
+    private client;
+    private model;
+    private sleep;
+    /** Key into {@link forcedToolChoiceIgnored}: one entry per endpoint + model. */
+    private choiceKey;
+    private forcedIgnored;
+    constructor(opts: OpenAIChatModelOptions);
+    create(req: ChatRequest): Promise<ChatResponse>;
+    /**
+     * A structured reply, however the endpoint is willing to produce one.
+     *
+     * A forced `tool_choice` is the only structured-output mechanism graft can
+     * rely on across providers, and it is the least portable corner of the
+     * OpenAI-compatible spec: an endpoint can reject the object form (handled one
+     * layer down), answer 200 with an error object, or quietly answer in prose.
+     * So walk a bounded ladder — the object form, then `"required"`, then
+     * `"auto"` with the tool named in the prompt — and take the payload from
+     * `content` if that is where it landed. `"required"` only stands in for the
+     * object form when one tool was offered, the same rule the 400 fallback
+     * applies: with several tools it would let the model pick freely rather than
+     * honour the caller's choice.
+     *
+     * An endpoint that honours the forced form never sees rung two — the first
+     * attempt answers the question and the ladder ends.
+     *
+     * An errored body is a weaker signal than a served-but-wrong reply, since it
+     * may belong to an upstream that a retry would reach, so it never downgrades
+     * on its own: it is counted instead. A call where every forced-family rung
+     * failed that way and the final "auto" rung then answered is a rescue, and
+     * the second rescue with no successful forced call in between is the
+     * endpoint's pattern, not its upstream's — that, and only that, switches
+     * later calls straight to "auto".
+     */
+    private structured;
+    /** One rung: the forced object form, its equivalent string form, or "auto" plus a prompt. */
+    private attempt;
+    /** Did the reply carry the payload the caller forced a tool for? */
+    private delivered;
+    /**
+     * One more rescued call for this endpoint + model. When the count reaches
+     * the threshold with no successful forced call in between, the endpoint is
+     * marked the same way a served-but-wrong reply marks it.
+     */
+    private noteForcedRescue;
+    /**
+     * Remember, for this process, that the endpoint did not answer a forced
+     * `tool_choice` with a tool call, so later requests go straight to the rung
+     * that works. Logged once per endpoint + model — the set is shared by every
+     * instance, so a second adapter for the same endpoint never repeats it: it
+     * changes the bytes on the wire, which is exactly what someone reading a
+     * failing build needs told.
+     */
+    private rememberForcedIgnored;
+    /**
+     * One completion, with the transport's retry policy extended to a 200 that
+     * carried an error instead of one. The SDK retries 429, 5xx and connection
+     * failures but never a successful status, and that body is the endpoint
+     * reporting a failed upstream call — the same judgement, so the same
+     * `GRAFT_LLM_RETRIES` budget and backoff apply, and the typed error is what
+     * reaches the caller once that budget is spent.
+     */
+    private complete;
+    /**
+     * Wraps `chat.completions.create` with a narrow, safe fallback: if the
+     * server rejects an object-form `tool_choice` and exactly one tool was
+     * offered, "required" is behaviorally identical (the model has nothing
+     * else to pick), so retry once with the string form instead of failing
+     * the whole build. Left alone when more than one tool is offered, since
+     * "required" would then let the model choose freely instead of the
+     * caller-specified tool — that ambiguity isn't safe to paper over
+     * automatically.
+     */
+    private createChatCompletion;
+    /**
+     * Normalize provider message content to the plain text representation
+     * expected by the rest of Graft. Some OpenAI-compatible providers
+     * (observed with GLM-5.3 hosted on Mistral's platform) return content as
+     * structured content parts, e.g. `[{ type: "text", text: "..." }]`,
+     * instead of a plain string — this path in Graft assumed the latter.
+     */
+    private static contentToText;
+    private fromResponse;
+}
+//# sourceMappingURL=openai.d.ts.map

package/dist/ai/llm/openai.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"openai.d.ts","sourceRoot":"","sources":["../../../src/ai/llm/openai.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AACH,OAAO,MAAM,MAAM,QAAQ,CAAC;AAG5B,OAAO,KAAK,EAAE,SAAS,EAAE,WAAW,EAAE,YAAY,EAAsC,MAAM,YAAY,CAAC;AAsD3G,MAAM,WAAW,sBAAsB;IACrC,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;IACd,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,2DAA2D;IAC3D,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,2DAA2D;IAC3D,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACjC,0EAA0E;IAC1E,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,2EAA2E;IAC3E,KAAK,CAAC,EAAE,CAAC,EAAE,EAAE,MAAM,KAAK,OAAO,CAAC,IAAI,CAAC,CAAC;CACvC;AAuGD;;;;;;;;;;;GAWG;AACH,qBAAa,qBAAsB,SAAQ,KAAK;IAO5C,gEAAgE;IAChE,QAAQ,CAAC,IAAI,CAAC,EAAE,MAAM,GAAG,MAAM;IAPjC,8EAA8E;IAC9E,QAAQ,CAAC,MAAM,OAAO;IACtB,qEAAqE;IACrE,QAAQ,CAAC,SAAS,QAAQ;gBAExB,OAAO,EAAE,MAAM;IACf,gEAAgE;IACvD,IAAI,CAAC,EAAE,MAAM,GAAG,MAAM,YAAA;CAKlC;AAiGD,qBAAa,eAAgB,YAAW,SAAS;IAC/C,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,OAAO,CAAC,MAAM,CAAS;IACvB,OAAO,CAAC,KAAK,CAAS;IACtB,OAAO,CAAC,KAAK,CAAgC;IAC7C,gFAAgF;IAChF,OAAO,CAAC,SAAS,CAAS;IAC1B,OAAO,CAAC,aAAa,CAAU;gBAEnB,IAAI,EAAE,sBAAsB;IAgBlC,MAAM,CAAC,GAAG,EAAE,WAAW,GAAG,OAAO,CAAC,YAAY,CAAC;IA6BrD;;;;;;;;;;;;;;;;;;;;;;;;OAwBG;YACW,UAAU;IA2CxB,6FAA6F;IAC7F,OAAO,CAAC,OAAO;IAMf,oEAAoE;IACpE,OAAO,CAAC,SAAS;IAKjB;;;;OAIG;IACH,OAAO,CAAC,gBAAgB;IAOxB;;;;;;;OAOG;IACH,OAAO,CAAC,qBAAqB;IAS7B;;;;;;;OAOG;YACW,QAAQ;IAgBtB;;;;;;;;;OASG;YACW,oBAAoB;IAkClC;;;;;;OAMG;IACH,OAAO,CAAC,MAAM,CAAC,aAAa;IAe5B,OAAO,CAAC,YAAY;CAgDrB"}