@mikode13/harness 2.1.0 → 2.2.0

package/dist/agent/domain/agent.d.ts

@@ -44,6 +44,13 @@ export type ProgressEvent = {
 } | {
     type: 'agentMessage';
     message: string;
+}
+/** `id` pairs a call's end with its start when one step calls the same tool more than once. */
+ | {
+    type: 'tool';
+    id: string;
+    name: string;
+    status: 'in_progress' | 'completed' | 'error';
 } | {
     type: 'todoList';
     items: {

package/dist/agent/domain/agent.js.map

@@ -1 +1 @@
-{"version":3,"file":"agent.js","sourceRoot":"","sources":["../../../src/agent/domain/agent.ts"],"names":[],"mappings":"","sourcesContent":["import type { Tokens } from '#src/shared/domain/tokens';\n\n/**\n * How a run ended when it did not fail. Every token a run spends travels with its end: here\n * when it finishes, in `RecoverableError.tokens` or `UnrecoverableError.tokens` when it fails.\n */\nexport interface AgentResponse {\n\t/** The final text. Empty when the run produced none, for example a turn that only edited files. */\n\tresponse: string;\n\t/**\n\t * Missing when any call in the run completed without reporting usage, which is not the same\n\t * as zero: a partial sum would look complete.\n\t */\n\ttokens?: Tokens;\n\t/** Wall-clock seconds the consumer waited for the run, retries and every role included. */\n\tduration: number;\n}\n\nexport type Callback = (item: ProgressEvent) => void;\n\n/**\n * Live activity during a run, for narration only: the result travels in `AgentResponse`.\n * New event types can arrive in a minor release, so render the types you know and ignore\n * the rest instead of switching over them exhaustively with a `never` check.\n */\nexport type ProgressEvent =\n\t| { type: 'command'; command: string; exitCode?: number }\n\t| { type: 'reasoning'; message: string }\n\t| { type: 'search'; query: string }\n\t| { type: 'fileChange'; changes: { path: string; kind: 'add' | 'update' | 'delete' }[] }\n\t| { type: 'mcpTool'; server: string; tool: string; status: string }\n\t| { type: 'agentMessage'; message: string }\n\t| { type: 'todoList'; items: { text: string; completed: boolean }[] }\n\t| { type: 'turnStarted' }\n\t| { type: 'turnEnded' };\n\n/**\n * The contract every engine (CodexAgent, ClaudeAgent, future providers) and every\n * decorator (RetryingAgent, OrchestratorAgent) is built against.\n *\n * Implementers MUST only ever reject with `RecoverableError` or `UnrecoverableError`\n * (see ./errors.ts) — never a raw SDK error, a plain `Error`, or anything else leaked\n * unclassified. Every consumer of `Agent` (RetryingAgent's retry decision,\n * OrchestratorAgent's failure handling) `instanceof`-checks against those two types to\n * decide what to do next; a leaked, unclassified error bypasses that decision\n * entirely — it gets retried when it shouldn't be, or crashes a run that a retry\n * would have recovered. Wrap every call into the underlying SDK so nothing escapes\n * unclassified, including failures the SDK itself doesn't model as a domain error\n * (network errors, malformed responses, etc.); `classifyProviderFailure` does this.\n *\n * Cancellation is the one exception: an `AbortError` must propagate unchanged, because\n * consumers check for it before either error type.\n */\nexport interface Agent {\n\trun(prompt: string, signal: AbortSignal, callback: Callback): Promise<AgentResponse>;\n}\n"]}
+{"version":3,"file":"agent.js","sourceRoot":"","sources":["../../../src/agent/domain/agent.ts"],"names":[],"mappings":"","sourcesContent":["import type { Tokens } from '#src/shared/domain/tokens';\n\n/**\n * How a run ended when it did not fail. Every token a run spends travels with its end: here\n * when it finishes, in `RecoverableError.tokens` or `UnrecoverableError.tokens` when it fails.\n */\nexport interface AgentResponse {\n\t/** The final text. Empty when the run produced none, for example a turn that only edited files. */\n\tresponse: string;\n\t/**\n\t * Missing when any call in the run completed without reporting usage, which is not the same\n\t * as zero: a partial sum would look complete.\n\t */\n\ttokens?: Tokens;\n\t/** Wall-clock seconds the consumer waited for the run, retries and every role included. */\n\tduration: number;\n}\n\nexport type Callback = (item: ProgressEvent) => void;\n\n/**\n * Live activity during a run, for narration only: the result travels in `AgentResponse`.\n * New event types can arrive in a minor release, so render the types you know and ignore\n * the rest instead of switching over them exhaustively with a `never` check.\n */\nexport type ProgressEvent =\n\t| { type: 'command'; command: string; exitCode?: number }\n\t| { type: 'reasoning'; message: string }\n\t| { type: 'search'; query: string }\n\t| { type: 'fileChange'; changes: { path: string; kind: 'add' | 'update' | 'delete' }[] }\n\t| { type: 'mcpTool'; server: string; tool: string; status: string }\n\t| { type: 'agentMessage'; message: string }\n\t/** `id` pairs a call's end with its start when one step calls the same tool more than once. */\n\t| { type: 'tool'; id: string; name: string; status: 'in_progress' | 'completed' | 'error' }\n\t| { type: 'todoList'; items: { text: string; completed: boolean }[] }\n\t| { type: 'turnStarted' }\n\t| { type: 'turnEnded' };\n\n/**\n * The contract every engine (CodexAgent, ClaudeAgent, future providers) and every\n * decorator (RetryingAgent, OrchestratorAgent) is built against.\n *\n * Implementers MUST only ever reject with `RecoverableError` or `UnrecoverableError`\n * (see ./errors.ts) — never a raw SDK error, a plain `Error`, or anything else leaked\n * unclassified. Every consumer of `Agent` (RetryingAgent's retry decision,\n * OrchestratorAgent's failure handling) `instanceof`-checks against those two types to\n * decide what to do next; a leaked, unclassified error bypasses that decision\n * entirely — it gets retried when it shouldn't be, or crashes a run that a retry\n * would have recovered. Wrap every call into the underlying SDK so nothing escapes\n * unclassified, including failures the SDK itself doesn't model as a domain error\n * (network errors, malformed responses, etc.); `classifyProviderFailure` does this.\n *\n * Cancellation is the one exception: an `AbortError` must propagate unchanged, because\n * consumers check for it before either error type.\n */\nexport interface Agent {\n\trun(prompt: string, signal: AbortSignal, callback: Callback): Promise<AgentResponse>;\n}\n"]}

package/dist/engines/domain/model/llmAgent.d.ts

@@ -1,17 +1,34 @@
 import type { Agent, AgentResponse, Callback } from '#src/agent/domain/agent';
 import type { LLMClient } from '#src/llm/domain/llm';
 import type { Message } from '#src/llm/domain/message';
+import type { Tool } from '#src/llm/domain/tool';
 /**
  * An agent whose conversation is owned by MiKode: every run sends the whole context to a
- * stateless `LLMClient`. Turn boundaries are the consumer's to emit, as with the other
- * engines; this agent only narrates what the model produced.
+ * stateless `LLMClient`. Within a run the model may call the agent's tools; the agent runs them
+ * and sends their results back until the model answers without calling one. Turn boundaries
+ * are the consumer's to emit, as with the other engines; this agent only narrates what the
+ * model produced and what its tools did.
  */
 export declare class LLMAgent implements Agent {
     private readonly llmClient;
     private readonly conversation;
-    constructor({ llmClient, messages }: {
+    private readonly tools;
+    private readonly toolDefinitions;
+    private readonly maxSteps;
+    constructor({ llmClient, messages, tools, maxSteps, }: {
         llmClient: LLMClient;
         messages?: Message[];
+        tools?: Tool[];
+        /** How many calls to the model one run may make before it fails. */
+        maxSteps?: number;
     });
+    private llmCall;
+    /**
+     * Always answers the call: a missing tool or a failing one becomes an error result, so the
+     * model can correct itself. Only a cancellation escapes, whatever error the tool turned it into.
+     */
+    private runTool;
+    /** One at a time and in order, so their effects and their narration never interleave. */
+    private runTools;
     run(prompt: string, signal: AbortSignal, callback: Callback): Promise<AgentResponse>;
 }

package/dist/engines/domain/model/llmAgent.js

@@ -1,30 +1,76 @@
-import { UnrecoverableError, withSpentTokens } from '#src/agent/domain/errors';
-import { classifyHostFailure, classifyProviderFailure, treatErrors, } from '#src/agent/domain/providerFailure';
+import { InvalidAgentConfigError, RecoverableError, UnrecoverableError, withSpentTokens, } from '#src/agent/domain/errors';
+import { classifyHostFailure, classifyProviderFailure, describeFailure, treatErrors, } from '#src/agent/domain/providerFailure';
 import { Conversation } from '#src/llm/domain/conversation';
+import { addTokens } from '#src/shared/domain/tokens';
 function describePart(part) {
-    return part.type === 'text'
-        ? { type: 'agentMessage', message: part.text }
-        : { type: 'reasoning', message: part.text };
+    switch (part.type) {
+        case 'text':
+            return { type: 'agentMessage', message: part.text };
+        case 'reasoning':
+            return { type: 'reasoning', message: part.text };
+        case 'toolCall':
+            return { type: 'tool', id: part.id, name: part.name, status: 'in_progress' };
+        case 'toolResult':
+            return {
+                type: 'tool',
+                id: part.callId,
+                name: part.name,
+                status: part.isError ? 'error' : 'completed',
+            };
+        case 'providerData':
+            // Opaque by design: its readable side, if any, arrives as a reasoning part.
+            return undefined;
+    }
+}
+function narrate(part, callback) {
+    const event = describePart(part);
+    if (!event) {
+        return;
+    }
+    treatErrors(() => {
+        callback(event);
+    }, classifyHostFailure, 'LLM agent progress callback failed');
 }
 /**
  * An agent whose conversation is owned by MiKode: every run sends the whole context to a
- * stateless `LLMClient`. Turn boundaries are the consumer's to emit, as with the other
- * engines; this agent only narrates what the model produced.
+ * stateless `LLMClient`. Within a run the model may call the agent's tools; the agent runs them
+ * and sends their results back until the model answers without calling one. Turn boundaries
+ * are the consumer's to emit, as with the other engines; this agent only narrates what the
+ * model produced and what its tools did.
  */
 export class LLMAgent {
     llmClient;
     conversation;
-    constructor({ llmClient, messages }) {
+    tools;
+    toolDefinitions;
+    maxSteps;
+    constructor({ llmClient, messages, tools = [], maxSteps = 25, }) {
+        if (!Number.isInteger(maxSteps) || maxSteps < 1) {
+            throw new InvalidAgentConfigError(`maxSteps must be a positive integer; got ${String(maxSteps)}`);
+        }
         this.llmClient = llmClient;
         this.conversation = new Conversation(messages);
+        this.maxSteps = maxSteps;
+        this.tools = new Map(tools.map(tool => [tool.name, tool]));
+        if (this.tools.size !== tools.length) {
+            // The model calls tools by name, so a second one with the same name could never run.
+            throw new InvalidAgentConfigError('Two tools share a name; each name must be unique');
+        }
+        // The client only describes the tools to the model; running them stays with the agent.
+        this.toolDefinitions = tools.map(({ name, description, inputSchema }) => ({
+            name,
+            description,
+            inputSchema,
+        }));
     }
-    async run(prompt, signal, callback) {
-        const start = Date.now();
-        const userMessage = { role: 'user', content: [{ type: 'text', text: prompt }] };
+    async llmCall(runMessages, signal) {
         let response;
         try {
-            // The prompt is sent as a copy so the one recorded below is the one the agent built.
-            response = await this.llmClient.send([...this.conversation.getContext(), structuredClone(userMessage)], signal);
+            // Sent as a copy, so what the run records is what the agent built, whatever the client does.
+            response = await this.llmClient.send({
+                context: [...this.conversation.getContext(), ...structuredClone(runMessages)],
+                tools: this.toolDefinitions,
+            }, signal);
         }
         catch (error) {
             throw classifyProviderFailure(error, 'The LLM call failed');
@@ -37,27 +83,107 @@ export class LLMAgent {
                 usageUnreported: !response.usage,
             });
         }
-        this.conversation.addExchange(userMessage, response.message);
+        return response;
+    }
+    /**
+     * Always answers the call: a missing tool or a failing one becomes an error result, so the
+     * model can correct itself. Only a cancellation escapes, whatever error the tool turned it into.
+     */
+    async runTool(call, signal) {
+        const result = { type: 'toolResult', callId: call.id, name: call.name };
+        const tool = this.tools.get(call.name);
+        if (!tool) {
+            return {
+                ...result,
+                output: `There is no tool named "${call.name}". Available tools: ${[...this.tools.keys()].join(', ') || 'none'}.`,
+                isError: true,
+            };
+        }
+        // Checked again at the last moment: the consumer may cancel on the announcement itself.
+        signal.throwIfAborted();
         try {
-            for (const part of response.message.content) {
-                treatErrors(() => {
-                    callback(describePart(part));
-                }, classifyHostFailure, 'LLM agent progress callback failed');
+            return { ...result, output: await tool.execute(call.input, signal), isError: false };
+        }
+        catch (error) {
+            signal.throwIfAborted();
+            return { ...result, output: describeFailure(error), isError: true };
+        }
+    }
+    /** One at a time and in order, so their effects and their narration never interleave. */
+    async runTools(calls, signal, callback) {
+        const results = [];
+        for (const call of calls) {
+            // A call before this one may have ignored the cancellation; this one must not start,
+            // nor be announced as running.
+            signal.throwIfAborted();
+            // Announced only as its turn comes, so a consumer never shows a call as running early.
+            narrate(call, callback);
+            const result = await this.runTool(call, signal);
+            narrate(result, callback);
+            results.push(result);
+        }
+        return { role: 'tool', content: results };
+    }
+    async run(prompt, signal, callback) {
+        const start = Date.now();
+        const runMessages = [{ role: 'user', content: [{ type: 'text', text: prompt }] }];
+        let tokens;
+        // Once one call went unreported the run's total is unknown, however many report later.
+        let unreported = false;
+        let toolRan = false;
+        try {
+            for (let step = 1; step <= this.maxSteps; step++) {
+                const response = await this.llmCall(runMessages, signal);
+                tokens = addTokens(tokens, response.usage);
+                unreported ||= !response.usage;
+                runMessages.push(response.message);
+                // Tool calls are narrated as they start, in `runTools`, and never if they do not run.
+                for (const part of response.message.content) {
+                    if (part.type !== 'toolCall')
+                        narrate(part, callback);
+                }
+                const calls = response.message.content.filter(part => part.type === 'toolCall');
+                if (calls.length === 0) {
+                    // Recorded only now, so a failed run leaves no tool call without its result.
+                    this.conversation.addRun(runMessages);
+                    return {
+                        // Only the final answer: text from earlier steps was narrated as it came.
+                        response: response.message.content
+                            .filter(part => part.type === 'text')
+                            .map(part => part.text)
+                            .join('\n'),
+                        tokens: unreported ? undefined : tokens,
+                        duration: (Date.now() - start) / 1000,
+                    };
+                }
+                // On the last step no call is left to send the results to, so the tools do not run.
+                if (step < this.maxSteps) {
+                    // Only a call to a tool the agent has reaches `execute`; a missing one has no
+                    // effect a retry could repeat.
+                    toolRan ||= calls.some(call => this.tools.has(call.name));
+                    runMessages.push(await this.runTools(calls, signal, callback));
+                }
             }
+            throw new UnrecoverableError('The LLM did not finish within its step limit', {
+                cause: `The model was still calling tools after ${String(this.maxSteps)} steps.`,
+            });
         }
         catch (error) {
-            // The answer already arrived, so the run spent its tokens whatever the consumer did.
-            throw withSpentTokens(error, response.usage, !response.usage);
-        }
-        const text = response.message.content
-            .filter(part => part.type === 'text')
-            .map(part => part.text)
-            .join('\n');
-        return {
-            response: text,
-            tokens: response.usage,
-            duration: (Date.now() - start) / 1000,
-        };
+            let failure = error;
+            if (toolRan && error instanceof RecoverableError) {
+                // RetryingAgent would run the whole prompt again, and with it every tool this run
+                // already executed, so a failure after one ran must not invite a retry.
+                const unrecoverable = new UnrecoverableError(error.message, {
+                    cause: error.cause,
+                    tokens: error.tokens,
+                    usageUnreported: error.usageUnreported,
+                });
+                unrecoverable.stack = error.stack;
+                failure = unrecoverable;
+            }
+            // The run spent the tokens of every step before the one that failed.
+            throw withSpentTokens(failure, tokens, unreported);
+        }
     }
 }
 //# sourceMappingURL=llmAgent.js.map

package/dist/engines/domain/model/llmAgent.js.map

@@ -1 +1 @@
-{"version":3,"file":"llmAgent.js","sourceRoot":"","sources":["../../../../src/engines/domain/model/llmAgent.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,kBAAkB,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAC;AAC/E,OAAO,EACN,mBAAmB,EACnB,uBAAuB,EACvB,WAAW,GACX,MAAM,mCAAmC,CAAC;AAC3C,OAAO,EAAE,YAAY,EAAE,MAAM,8BAA8B,CAAC;AAI5D,SAAS,YAAY,CAAC,IAAiB;IACtC,OAAO,IAAI,CAAC,IAAI,KAAK,MAAM;QAC1B,CAAC,CAAC,EAAE,IAAI,EAAE,cAAc,EAAE,OAAO,EAAE,IAAI,CAAC,IAAI,EAAE;QAC9C,CAAC,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,OAAO,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC;AAC9C,CAAC;AAED;;;;GAIG;AACH,MAAM,OAAO,QAAQ;IACH,SAAS,CAAY;IACrB,YAAY,CAAe;IAE5C,YAAY,EAAE,SAAS,EAAE,QAAQ,EAAkD;QAClF,IAAI,CAAC,SAAS,GAAG,SAAS,CAAC;QAC3B,IAAI,CAAC,YAAY,GAAG,IAAI,YAAY,CAAC,QAAQ,CAAC,CAAC;IAChD,CAAC;IAED,KAAK,CAAC,GAAG,CAAC,MAAc,EAAE,MAAmB,EAAE,QAAkB;QAChE,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QACzB,MAAM,WAAW,GAAY,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC,EAAE,CAAC;QAEzF,IAAI,QAAqB,CAAC;QAC1B,IAAI,CAAC;YACJ,qFAAqF;YACrF,QAAQ,GAAG,MAAM,IAAI,CAAC,SAAS,CAAC,IAAI,CACnC,CAAC,GAAG,IAAI,CAAC,YAAY,CAAC,UAAU,EAAE,EAAE,eAAe,CAAC,WAAW,CAAC,CAAC,EACjE,MAAM,CACN,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,MAAM,uBAAuB,CAAC,KAAK,EAAE,qBAAqB,CAAC,CAAC;QAC7D,CAAC;QAED,IAAI,QAAQ,CAAC,UAAU,KAAK,WAAW,EAAE,CAAC;YACzC,MAAM,IAAI,kBAAkB,CAAC,8CAA8C,EAAE;gBAC5E,KAAK,EAAE,2BAA2B,QAAQ,CAAC,UAAU,IAAI;gBACzD,0DAA0D;gBAC1D,MAAM,EAAE,QAAQ,CAAC,KAAK;gBACtB,eAAe,EAAE,CAAC,QAAQ,CAAC,KAAK;aAChC,CAAC,CAAC;QACJ,CAAC;QAED,IAAI,CAAC,YAAY,CAAC,WAAW,CAAC,WAAW,EAAE,QAAQ,CAAC,OAAO,CAAC,CAAC;QAE7D,IAAI,CAAC;YACJ,KAAK,MAAM,IAAI,IAAI,QAAQ,CAAC,OAAO,CAAC,OAAO,EAAE,CAAC;gBAC7C,WAAW,CACV,GAAG,EAAE;oBACJ,QAAQ,CAAC,YAAY,CAAC,IAAI,CAAC,CAAC,CAAC;gBAC9B,CAAC,EACD,mBAAmB,EACnB,oCAAoC,CACpC,CAAC;YACH,CAAC;QACF,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,qFAAqF;YACrF,MAAM,eAAe,CAAC,KAAK,EAAE,QAAQ,CAAC,KAAK,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;QAC/D,CAAC;QAED,MAAM,IAAI,GAAG,QAAQ,CAAC,OAAO,CAAC,OAAO;aACnC,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,KAAK,MAAM,CAAC;aACpC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC;aACtB,IAAI,CAAC,IAAI,CAAC,CAAC;QAEb,OAAO;YACN,QAAQ,EAAE,IAAI;YACd,MAAM,EAAE,QAAQ,CAAC,KAAK;YACtB,QAAQ,EAAE,CAAC,IAAI,CAAC,GAAG,EAAE,GAAG,KAAK,CAAC,GAAG,IAAI;SACrC,CAAC;IACH,CAAC;CACD","sourcesContent":["import type { Agent, AgentResponse, Callback, ProgressEvent } from '#src/agent/domain/agent';\nimport { UnrecoverableError, withSpentTokens } from '#src/agent/domain/errors';\nimport {\n\tclassifyHostFailure,\n\tclassifyProviderFailure,\n\ttreatErrors,\n} from '#src/agent/domain/providerFailure';\nimport { Conversation } from '#src/llm/domain/conversation';\nimport type { LLMClient, LLMResponse } from '#src/llm/domain/llm';\nimport type { Message, MessagePart } from '#src/llm/domain/message';\n\nfunction describePart(part: MessagePart): ProgressEvent {\n\treturn part.type === 'text'\n\t\t? { type: 'agentMessage', message: part.text }\n\t\t: { type: 'reasoning', message: part.text };\n}\n\n/**\n * An agent whose conversation is owned by MiKode: every run sends the whole context to a\n * stateless `LLMClient`. Turn boundaries are the consumer's to emit, as with the other\n * engines; this agent only narrates what the model produced.\n */\nexport class LLMAgent implements Agent {\n\tprivate readonly llmClient: LLMClient;\n\tprivate readonly conversation: Conversation;\n\n\tconstructor({ llmClient, messages }: { llmClient: LLMClient; messages?: Message[] }) {\n\t\tthis.llmClient = llmClient;\n\t\tthis.conversation = new Conversation(messages);\n\t}\n\n\tasync run(prompt: string, signal: AbortSignal, callback: Callback): Promise<AgentResponse> {\n\t\tconst start = Date.now();\n\t\tconst userMessage: Message = { role: 'user', content: [{ type: 'text', text: prompt }] };\n\n\t\tlet response: LLMResponse;\n\t\ttry {\n\t\t\t// The prompt is sent as a copy so the one recorded below is the one the agent built.\n\t\t\tresponse = await this.llmClient.send(\n\t\t\t\t[...this.conversation.getContext(), structuredClone(userMessage)],\n\t\t\t\tsignal,\n\t\t\t);\n\t\t} catch (error) {\n\t\t\tthrow classifyProviderFailure(error, 'The LLM call failed');\n\t\t}\n\n\t\tif (response.stopReason !== 'completed') {\n\t\t\tthrow new UnrecoverableError('The LLM stopped before completing its answer', {\n\t\t\t\tcause: `The model stopped with \"${response.stopReason}\".`,\n\t\t\t\t// The call was billed even though its answer is unusable.\n\t\t\t\ttokens: response.usage,\n\t\t\t\tusageUnreported: !response.usage,\n\t\t\t});\n\t\t}\n\n\t\tthis.conversation.addExchange(userMessage, response.message);\n\n\t\ttry {\n\t\t\tfor (const part of response.message.content) {\n\t\t\t\ttreatErrors(\n\t\t\t\t\t() => {\n\t\t\t\t\t\tcallback(describePart(part));\n\t\t\t\t\t},\n\t\t\t\t\tclassifyHostFailure,\n\t\t\t\t\t'LLM agent progress callback failed',\n\t\t\t\t);\n\t\t\t}\n\t\t} catch (error) {\n\t\t\t// The answer already arrived, so the run spent its tokens whatever the consumer did.\n\t\t\tthrow withSpentTokens(error, response.usage, !response.usage);\n\t\t}\n\n\t\tconst text = response.message.content\n\t\t\t.filter(part => part.type === 'text')\n\t\t\t.map(part => part.text)\n\t\t\t.join('\\n');\n\n\t\treturn {\n\t\t\tresponse: text,\n\t\t\ttokens: response.usage,\n\t\t\tduration: (Date.now() - start) / 1000,\n\t\t};\n\t}\n}\n"]}
+{"version":3,"file":"llmAgent.js","sourceRoot":"","sources":["../../../../src/engines/domain/model/llmAgent.ts"],"names":[],"mappings":"AACA,OAAO,EACN,uBAAuB,EACvB,gBAAgB,EAChB,kBAAkB,EAClB,eAAe,GACf,MAAM,0BAA0B,CAAC;AAClC,OAAO,EACN,mBAAmB,EACnB,uBAAuB,EACvB,eAAe,EACf,WAAW,GACX,MAAM,mCAAmC,CAAC;AAC3C,OAAO,EAAE,YAAY,EAAE,MAAM,8BAA8B,CAAC;AAI5D,OAAO,EAAE,SAAS,EAAe,MAAM,2BAA2B,CAAC;AAEnE,SAAS,YAAY,CAAC,IAAiB;IACtC,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;QACnB,KAAK,MAAM;YACV,OAAO,EAAE,IAAI,EAAE,cAAc,EAAE,OAAO,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC;QACrD,KAAK,WAAW;YACf,OAAO,EAAE,IAAI,EAAE,WAAW,EAAE,OAAO,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC;QAClD,KAAK,UAAU;YACd,OAAO,EAAE,IAAI,EAAE,MAAM,EAAE,EAAE,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,MAAM,EAAE,aAAa,EAAE,CAAC;QAC9E,KAAK,YAAY;YAChB,OAAO;gBACN,IAAI,EAAE,MAAM;gBACZ,EAAE,EAAE,IAAI,CAAC,MAAM;gBACf,IAAI,EAAE,IAAI,CAAC,IAAI;gBACf,MAAM,EAAE,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,WAAW;aAC5C,CAAC;QACH,KAAK,cAAc;YAClB,4EAA4E;YAC5E,OAAO,SAAS,CAAC;IACnB,CAAC;AACF,CAAC;AAED,SAAS,OAAO,CAAC,IAAiB,EAAE,QAAkB;IACrD,MAAM,KAAK,GAAG,YAAY,CAAC,IAAI,CAAC,CAAC;IACjC,IAAI,CAAC,KAAK,EAAE,CAAC;QACZ,OAAO;IACR,CAAC;IAED,WAAW,CACV,GAAG,EAAE;QACJ,QAAQ,CAAC,KAAK,CAAC,CAAC;IACjB,CAAC,EACD,mBAAmB,EACnB,oCAAoC,CACpC,CAAC;AACH,CAAC;AAED;;;;;;GAMG;AACH,MAAM,OAAO,QAAQ;IACH,SAAS,CAAY;IACrB,YAAY,CAAe;IAC3B,KAAK,CAAoB;IACzB,eAAe,CAAmB;IAClC,QAAQ,CAAS;IAElC,YAAY,EACX,SAAS,EACT,QAAQ,EACR,KAAK,GAAG,EAAE,EACV,QAAQ,GAAG,EAAE,GAOb;QACA,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,QAAQ,CAAC,IAAI,QAAQ,GAAG,CAAC,EAAE,CAAC;YACjD,MAAM,IAAI,uBAAuB,CAChC,4CAA4C,MAAM,CAAC,QAAQ,CAAC,EAAE,CAC9D,CAAC;QACH,CAAC;QAED,IAAI,CAAC,SAAS,GAAG,SAAS,CAAC;QAC3B,IAAI,CAAC,YAAY,GAAG,IAAI,YAAY,CAAC,QAAQ,CAAC,CAAC;QAC/C,IAAI,CAAC,QAAQ,GAAG,QAAQ,CAAC;QACzB,IAAI,CAAC,KAAK,GAAG,IAAI,GAAG,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC;QAC3D,IAAI,IAAI,CAAC,KAAK,CAAC,IAAI,KAAK,KAAK,CAAC,MAAM,EAAE,CAAC;YACtC,qFAAqF;YACrF,MAAM,IAAI,uBAAuB,CAAC,kDAAkD,CAAC,CAAC;QACvF,CAAC;QACD,uFAAuF;QACvF,IAAI,CAAC,eAAe,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,WAAW,EAAE,EAAE,EAAE,CAAC,CAAC;YACzE,IAAI;YACJ,WAAW;YACX,WAAW;SACX,CAAC,CAAC,CAAC;IACL,CAAC;IAEO,KAAK,CAAC,OAAO,CAAC,WAAsB,EAAE,MAAmB;QAChE,IAAI,QAAqB,CAAC;QAC1B,IAAI,CAAC;YACJ,6FAA6F;YAC7F,QAAQ,GAAG,MAAM,IAAI,CAAC,SAAS,CAAC,IAAI,CACnC;gBACC,OAAO,EAAE,CAAC,GAAG,IAAI,CAAC,YAAY,CAAC,UAAU,EAAE,EAAE,GAAG,eAAe,CAAC,WAAW,CAAC,CAAC;gBAC7E,KAAK,EAAE,IAAI,CAAC,eAAe;aAC3B,EACD,MAAM,CACN,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,MAAM,uBAAuB,CAAC,KAAK,EAAE,qBAAqB,CAAC,CAAC;QAC7D,CAAC;QAED,IAAI,QAAQ,CAAC,UAAU,KAAK,WAAW,EAAE,CAAC;YACzC,MAAM,IAAI,kBAAkB,CAAC,8CAA8C,EAAE;gBAC5E,KAAK,EAAE,2BAA2B,QAAQ,CAAC,UAAU,IAAI;gBACzD,0DAA0D;gBAC1D,MAAM,EAAE,QAAQ,CAAC,KAAK;gBACtB,eAAe,EAAE,CAAC,QAAQ,CAAC,KAAK;aAChC,CAAC,CAAC;QACJ,CAAC;QAED,OAAO,QAAQ,CAAC;IACjB,CAAC;IAED;;;OAGG;IACK,KAAK,CAAC,OAAO,CAAC,IAAkB,EAAE,MAAmB;QAC5D,MAAM,MAAM,GAAG,EAAE,IAAI,EAAE,YAAqB,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC;QACjF,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACvC,IAAI,CAAC,IAAI,EAAE,CAAC;YACX,OAAO;gBACN,GAAG,MAAM;gBACT,MAAM,EAAE,2BAA2B,IAAI,CAAC,IAAI,uBAAuB,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,MAAM,GAAG;gBACjH,OAAO,EAAE,IAAI;aACb,CAAC;QACH,CAAC;QAED,wFAAwF;QACxF,MAAM,CAAC,cAAc,EAAE,CAAC;QACxB,IAAI,CAAC;YACJ,OAAO,EAAE,GAAG,MAAM,EAAE,MAAM,EAAE,MAAM,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,KAAK,EAAE,MAAM,CAAC,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QACtF,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,MAAM,CAAC,cAAc,EAAE,CAAC;YACxB,OAAO,EAAE,GAAG,MAAM,EAAE,MAAM,EAAE,eAAe,CAAC,KAAK,CAAC,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;QACrE,CAAC;IACF,CAAC;IAED,yFAAyF;IACjF,KAAK,CAAC,QAAQ,CACrB,KAAqB,EACrB,MAAmB,EACnB,QAAkB;QAElB,MAAM,OAAO,GAAkB,EAAE,CAAC;QAClC,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;YAC1B,qFAAqF;YACrF,+BAA+B;YAC/B,MAAM,CAAC,cAAc,EAAE,CAAC;YACxB,uFAAuF;YACvF,OAAO,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC;YACxB,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,OAAO,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;YAChD,OAAO,CAAC,MAAM,EAAE,QAAQ,CAAC,CAAC;YAC1B,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QACtB,CAAC;QAED,OAAO,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,EAAE,CAAC;IAC3C,CAAC;IAED,KAAK,CAAC,GAAG,CAAC,MAAc,EAAE,MAAmB,EAAE,QAAkB;QAChE,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QACzB,MAAM,WAAW,GAAc,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC,EAAE,CAAC,CAAC;QAC7F,IAAI,MAA0B,CAAC;QAC/B,uFAAuF;QACvF,IAAI,UAAU,GAAG,KAAK,CAAC;QACvB,IAAI,OAAO,GAAG,KAAK,CAAC;QAEpB,IAAI,CAAC;YACJ,KAAK,IAAI,IAAI,GAAG,CAAC,EAAE,IAAI,IAAI,IAAI,CAAC,QAAQ,EAAE,IAAI,EAAE,EAAE,CAAC;gBAClD,MAAM,QAAQ,GAAG,MAAM,IAAI,CAAC,OAAO,CAAC,WAAW,EAAE,MAAM,CAAC,CAAC;gBACzD,MAAM,GAAG,SAAS,CAAC,MAAM,EAAE,QAAQ,CAAC,KAAK,CAAC,CAAC;gBAC3C,UAAU,KAAK,CAAC,QAAQ,CAAC,KAAK,CAAC;gBAC/B,WAAW,CAAC,IAAI,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;gBAEnC,sFAAsF;gBACtF,KAAK,MAAM,IAAI,IAAI,QAAQ,CAAC,OAAO,CAAC,OAAO,EAAE,CAAC;oBAC7C,IAAI,IAAI,CAAC,IAAI,KAAK,UAAU;wBAAE,OAAO,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC;gBACvD,CAAC;gBAED,MAAM,KAAK,GAAG,QAAQ,CAAC,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,KAAK,UAAU,CAAC,CAAC;gBAChF,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;oBACxB,6EAA6E;oBAC7E,IAAI,CAAC,YAAY,CAAC,MAAM,CAAC,WAAW,CAAC,CAAC;oBAEtC,OAAO;wBACN,0EAA0E;wBAC1E,QAAQ,EAAE,QAAQ,CAAC,OAAO,CAAC,OAAO;6BAChC,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,KAAK,MAAM,CAAC;6BACpC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC;6BACtB,IAAI,CAAC,IAAI,CAAC;wBACZ,MAAM,EAAE,UAAU,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,MAAM;wBACvC,QAAQ,EAAE,CAAC,IAAI,CAAC,GAAG,EAAE,GAAG,KAAK,CAAC,GAAG,IAAI;qBACrC,CAAC;gBACH,CAAC;gBAED,oFAAoF;gBACpF,IAAI,IAAI,GAAG,IAAI,CAAC,QAAQ,EAAE,CAAC;oBAC1B,8EAA8E;oBAC9E,+BAA+B;oBAC/B,OAAO,KAAK,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC;oBAC1D,WAAW,CAAC,IAAI,CAAC,MAAM,IAAI,CAAC,QAAQ,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,CAAC,CAAC,CAAC;gBAChE,CAAC;YACF,CAAC;YAED,MAAM,IAAI,kBAAkB,CAAC,8CAA8C,EAAE;gBAC5E,KAAK,EAAE,2CAA2C,MAAM,CAAC,IAAI,CAAC,QAAQ,CAAC,SAAS;aAChF,CAAC,CAAC;QACJ,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,IAAI,OAAO,GAAG,KAAK,CAAC;YACpB,IAAI,OAAO,IAAI,KAAK,YAAY,gBAAgB,EAAE,CAAC;gBAClD,kFAAkF;gBAClF,wEAAwE;gBACxE,MAAM,aAAa,GAAG,IAAI,kBAAkB,CAAC,KAAK,CAAC,OAAO,EAAE;oBAC3D,KAAK,EAAE,KAAK,CAAC,KAAK;oBAClB,MAAM,EAAE,KAAK,CAAC,MAAM;oBACpB,eAAe,EAAE,KAAK,CAAC,eAAe;iBACtC,CAAC,CAAC;gBACH,aAAa,CAAC,KAAK,GAAG,KAAK,CAAC,KAAK,CAAC;gBAClC,OAAO,GAAG,aAAa,CAAC;YACzB,CAAC;YAED,qEAAqE;YACrE,MAAM,eAAe,CAAC,OAAO,EAAE,MAAM,EAAE,UAAU,CAAC,CAAC;QACpD,CAAC;IACF,CAAC;CACD","sourcesContent":["import type { Agent, AgentResponse, Callback, ProgressEvent } from '#src/agent/domain/agent';\nimport {\n\tInvalidAgentConfigError,\n\tRecoverableError,\n\tUnrecoverableError,\n\twithSpentTokens,\n} from '#src/agent/domain/errors';\nimport {\n\tclassifyHostFailure,\n\tclassifyProviderFailure,\n\tdescribeFailure,\n\ttreatErrors,\n} from '#src/agent/domain/providerFailure';\nimport { Conversation } from '#src/llm/domain/conversation';\nimport type { LLMClient, LLMResponse } from '#src/llm/domain/llm';\nimport type { Message, MessagePart, ToolCallPart, ToolResultPart } from '#src/llm/domain/message';\nimport type { Tool, ToolDefinition } from '#src/llm/domain/tool';\nimport { addTokens, type Tokens } from '#src/shared/domain/tokens';\n\nfunction describePart(part: MessagePart): ProgressEvent | undefined {\n\tswitch (part.type) {\n\t\tcase 'text':\n\t\t\treturn { type: 'agentMessage', message: part.text };\n\t\tcase 'reasoning':\n\t\t\treturn { type: 'reasoning', message: part.text };\n\t\tcase 'toolCall':\n\t\t\treturn { type: 'tool', id: part.id, name: part.name, status: 'in_progress' };\n\t\tcase 'toolResult':\n\t\t\treturn {\n\t\t\t\ttype: 'tool',\n\t\t\t\tid: part.callId,\n\t\t\t\tname: part.name,\n\t\t\t\tstatus: part.isError ? 'error' : 'completed',\n\t\t\t};\n\t\tcase 'providerData':\n\t\t\t// Opaque by design: its readable side, if any, arrives as a reasoning part.\n\t\t\treturn undefined;\n\t}\n}\n\nfunction narrate(part: MessagePart, callback: Callback): void {\n\tconst event = describePart(part);\n\tif (!event) {\n\t\treturn;\n\t}\n\n\ttreatErrors(\n\t\t() => {\n\t\t\tcallback(event);\n\t\t},\n\t\tclassifyHostFailure,\n\t\t'LLM agent progress callback failed',\n\t);\n}\n\n/**\n * An agent whose conversation is owned by MiKode: every run sends the whole context to a\n * stateless `LLMClient`. Within a run the model may call the agent's tools; the agent runs them\n * and sends their results back until the model answers without calling one. Turn boundaries\n * are the consumer's to emit, as with the other engines; this agent only narrates what the\n * model produced and what its tools did.\n */\nexport class LLMAgent implements Agent {\n\tprivate readonly llmClient: LLMClient;\n\tprivate readonly conversation: Conversation;\n\tprivate readonly tools: Map<string, Tool>;\n\tprivate readonly toolDefinitions: ToolDefinition[];\n\tprivate readonly maxSteps: number;\n\n\tconstructor({\n\t\tllmClient,\n\t\tmessages,\n\t\ttools = [],\n\t\tmaxSteps = 25,\n\t}: {\n\t\tllmClient: LLMClient;\n\t\tmessages?: Message[];\n\t\ttools?: Tool[];\n\t\t/** How many calls to the model one run may make before it fails. */\n\t\tmaxSteps?: number;\n\t}) {\n\t\tif (!Number.isInteger(maxSteps) || maxSteps < 1) {\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`maxSteps must be a positive integer; got ${String(maxSteps)}`,\n\t\t\t);\n\t\t}\n\n\t\tthis.llmClient = llmClient;\n\t\tthis.conversation = new Conversation(messages);\n\t\tthis.maxSteps = maxSteps;\n\t\tthis.tools = new Map(tools.map(tool => [tool.name, tool]));\n\t\tif (this.tools.size !== tools.length) {\n\t\t\t// The model calls tools by name, so a second one with the same name could never run.\n\t\t\tthrow new InvalidAgentConfigError('Two tools share a name; each name must be unique');\n\t\t}\n\t\t// The client only describes the tools to the model; running them stays with the agent.\n\t\tthis.toolDefinitions = tools.map(({ name, description, inputSchema }) => ({\n\t\t\tname,\n\t\t\tdescription,\n\t\t\tinputSchema,\n\t\t}));\n\t}\n\n\tprivate async llmCall(runMessages: Message[], signal: AbortSignal): Promise<LLMResponse> {\n\t\tlet response: LLMResponse;\n\t\ttry {\n\t\t\t// Sent as a copy, so what the run records is what the agent built, whatever the client does.\n\t\t\tresponse = await this.llmClient.send(\n\t\t\t\t{\n\t\t\t\t\tcontext: [...this.conversation.getContext(), ...structuredClone(runMessages)],\n\t\t\t\t\ttools: this.toolDefinitions,\n\t\t\t\t},\n\t\t\t\tsignal,\n\t\t\t);\n\t\t} catch (error) {\n\t\t\tthrow classifyProviderFailure(error, 'The LLM call failed');\n\t\t}\n\n\t\tif (response.stopReason !== 'completed') {\n\t\t\tthrow new UnrecoverableError('The LLM stopped before completing its answer', {\n\t\t\t\tcause: `The model stopped with \"${response.stopReason}\".`,\n\t\t\t\t// The call was billed even though its answer is unusable.\n\t\t\t\ttokens: response.usage,\n\t\t\t\tusageUnreported: !response.usage,\n\t\t\t});\n\t\t}\n\n\t\treturn response;\n\t}\n\n\t/**\n\t * Always answers the call: a missing tool or a failing one becomes an error result, so the\n\t * model can correct itself. Only a cancellation escapes, whatever error the tool turned it into.\n\t */\n\tprivate async runTool(call: ToolCallPart, signal: AbortSignal): Promise<ToolResultPart> {\n\t\tconst result = { type: 'toolResult' as const, callId: call.id, name: call.name };\n\t\tconst tool = this.tools.get(call.name);\n\t\tif (!tool) {\n\t\t\treturn {\n\t\t\t\t...result,\n\t\t\t\toutput: `There is no tool named \"${call.name}\". Available tools: ${[...this.tools.keys()].join(', ') || 'none'}.`,\n\t\t\t\tisError: true,\n\t\t\t};\n\t\t}\n\n\t\t// Checked again at the last moment: the consumer may cancel on the announcement itself.\n\t\tsignal.throwIfAborted();\n\t\ttry {\n\t\t\treturn { ...result, output: await tool.execute(call.input, signal), isError: false };\n\t\t} catch (error) {\n\t\t\tsignal.throwIfAborted();\n\t\t\treturn { ...result, output: describeFailure(error), isError: true };\n\t\t}\n\t}\n\n\t/** One at a time and in order, so their effects and their narration never interleave. */\n\tprivate async runTools(\n\t\tcalls: ToolCallPart[],\n\t\tsignal: AbortSignal,\n\t\tcallback: Callback,\n\t): Promise<Message> {\n\t\tconst results: MessagePart[] = [];\n\t\tfor (const call of calls) {\n\t\t\t// A call before this one may have ignored the cancellation; this one must not start,\n\t\t\t// nor be announced as running.\n\t\t\tsignal.throwIfAborted();\n\t\t\t// Announced only as its turn comes, so a consumer never shows a call as running early.\n\t\t\tnarrate(call, callback);\n\t\t\tconst result = await this.runTool(call, signal);\n\t\t\tnarrate(result, callback);\n\t\t\tresults.push(result);\n\t\t}\n\n\t\treturn { role: 'tool', content: results };\n\t}\n\n\tasync run(prompt: string, signal: AbortSignal, callback: Callback): Promise<AgentResponse> {\n\t\tconst start = Date.now();\n\t\tconst runMessages: Message[] = [{ role: 'user', content: [{ type: 'text', text: prompt }] }];\n\t\tlet tokens: Tokens | undefined;\n\t\t// Once one call went unreported the run's total is unknown, however many report later.\n\t\tlet unreported = false;\n\t\tlet toolRan = false;\n\n\t\ttry {\n\t\t\tfor (let step = 1; step <= this.maxSteps; step++) {\n\t\t\t\tconst response = await this.llmCall(runMessages, signal);\n\t\t\t\ttokens = addTokens(tokens, response.usage);\n\t\t\t\tunreported ||= !response.usage;\n\t\t\t\trunMessages.push(response.message);\n\n\t\t\t\t// Tool calls are narrated as they start, in `runTools`, and never if they do not run.\n\t\t\t\tfor (const part of response.message.content) {\n\t\t\t\t\tif (part.type !== 'toolCall') narrate(part, callback);\n\t\t\t\t}\n\n\t\t\t\tconst calls = response.message.content.filter(part => part.type === 'toolCall');\n\t\t\t\tif (calls.length === 0) {\n\t\t\t\t\t// Recorded only now, so a failed run leaves no tool call without its result.\n\t\t\t\t\tthis.conversation.addRun(runMessages);\n\n\t\t\t\t\treturn {\n\t\t\t\t\t\t// Only the final answer: text from earlier steps was narrated as it came.\n\t\t\t\t\t\tresponse: response.message.content\n\t\t\t\t\t\t\t.filter(part => part.type === 'text')\n\t\t\t\t\t\t\t.map(part => part.text)\n\t\t\t\t\t\t\t.join('\\n'),\n\t\t\t\t\t\ttokens: unreported ? undefined : tokens,\n\t\t\t\t\t\tduration: (Date.now() - start) / 1000,\n\t\t\t\t\t};\n\t\t\t\t}\n\n\t\t\t\t// On the last step no call is left to send the results to, so the tools do not run.\n\t\t\t\tif (step < this.maxSteps) {\n\t\t\t\t\t// Only a call to a tool the agent has reaches `execute`; a missing one has no\n\t\t\t\t\t// effect a retry could repeat.\n\t\t\t\t\ttoolRan ||= calls.some(call => this.tools.has(call.name));\n\t\t\t\t\trunMessages.push(await this.runTools(calls, signal, callback));\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tthrow new UnrecoverableError('The LLM did not finish within its step limit', {\n\t\t\t\tcause: `The model was still calling tools after ${String(this.maxSteps)} steps.`,\n\t\t\t});\n\t\t} catch (error) {\n\t\t\tlet failure = error;\n\t\t\tif (toolRan && error instanceof RecoverableError) {\n\t\t\t\t// RetryingAgent would run the whole prompt again, and with it every tool this run\n\t\t\t\t// already executed, so a failure after one ran must not invite a retry.\n\t\t\t\tconst unrecoverable = new UnrecoverableError(error.message, {\n\t\t\t\t\tcause: error.cause,\n\t\t\t\t\ttokens: error.tokens,\n\t\t\t\t\tusageUnreported: error.usageUnreported,\n\t\t\t\t});\n\t\t\t\tunrecoverable.stack = error.stack;\n\t\t\t\tfailure = unrecoverable;\n\t\t\t}\n\n\t\t\t// The run spent the tokens of every step before the one that failed.\n\t\t\tthrow withSpentTokens(failure, tokens, unreported);\n\t\t}\n\t}\n}\n"]}

package/dist/llm/domain/conversation.d.ts

@@ -8,10 +8,11 @@ export declare class Conversation {
     private readonly messages;
     constructor(messages?: Message[]);
     /**
-     * Records a prompt and its answer together, once the call succeeded. A failed call
-     * records nothing, so a retry of the same prompt cannot appear twice.
+     * Records a whole run at once, once it succeeded: the prompt, every step the model took and
+     * every tool result. A failed run records nothing, so a retry cannot repeat its prompt, and
+     * no tool call is ever kept without its result.
      */
-    addExchange(prompt: Message, answer: Message): void;
+    addRun(messages: Message[]): void;
     /**
      * A deep copy, like everything the conversation takes in: a client that rewrites the
      * messages it was sent cannot change what the next turn sends.

package/dist/llm/domain/conversation.js

@@ -9,11 +9,12 @@ export class Conversation {
         this.messages = structuredClone(messages);
     }
     /**
-     * Records a prompt and its answer together, once the call succeeded. A failed call
-     * records nothing, so a retry of the same prompt cannot appear twice.
+     * Records a whole run at once, once it succeeded: the prompt, every step the model took and
+     * every tool result. A failed run records nothing, so a retry cannot repeat its prompt, and
+     * no tool call is ever kept without its result.
      */
-    addExchange(prompt, answer) {
-        this.messages.push(structuredClone(prompt), structuredClone(answer));
+    addRun(messages) {
+        this.messages.push(...structuredClone(messages));
     }
     /**
      * A deep copy, like everything the conversation takes in: a client that rewrites the

package/dist/llm/domain/conversation.js.map

@@ -1 +1 @@
-{"version":3,"file":"conversation.js","sourceRoot":"","sources":["../../../src/llm/domain/conversation.ts"],"names":[],"mappings":"AAEA;;;;GAIG;AACH,MAAM,OAAO,YAAY;IACP,QAAQ,CAAY;IAErC,YAAY,WAAsB,EAAE;QACnC,IAAI,CAAC,QAAQ,GAAG,eAAe,CAAC,QAAQ,CAAC,CAAC;IAC3C,CAAC;IAED;;;OAGG;IACH,WAAW,CAAC,MAAe,EAAE,MAAe;QAC3C,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,eAAe,CAAC,MAAM,CAAC,EAAE,eAAe,CAAC,MAAM,CAAC,CAAC,CAAC;IACtE,CAAC;IAED;;;OAGG;IACH,UAAU;QACT,OAAO,eAAe,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;IACvC,CAAC;CACD","sourcesContent":["import type { Message } from './message.ts';\n\n/**\n * What one agent remembers across its runs. The LLM never sees this object, only the\n * context built from it, so how the context is derived (compaction, later) can change\n * without touching either side.\n */\nexport class Conversation {\n\tprivate readonly messages: Message[];\n\n\tconstructor(messages: Message[] = []) {\n\t\tthis.messages = structuredClone(messages);\n\t}\n\n\t/**\n\t * Records a prompt and its answer together, once the call succeeded. A failed call\n\t * records nothing, so a retry of the same prompt cannot appear twice.\n\t */\n\taddExchange(prompt: Message, answer: Message): void {\n\t\tthis.messages.push(structuredClone(prompt), structuredClone(answer));\n\t}\n\n\t/**\n\t * A deep copy, like everything the conversation takes in: a client that rewrites the\n\t * messages it was sent cannot change what the next turn sends.\n\t */\n\tgetContext(): Message[] {\n\t\treturn structuredClone(this.messages);\n\t}\n}\n"]}
+{"version":3,"file":"conversation.js","sourceRoot":"","sources":["../../../src/llm/domain/conversation.ts"],"names":[],"mappings":"AAEA;;;;GAIG;AACH,MAAM,OAAO,YAAY;IACP,QAAQ,CAAY;IAErC,YAAY,WAAsB,EAAE;QACnC,IAAI,CAAC,QAAQ,GAAG,eAAe,CAAC,QAAQ,CAAC,CAAC;IAC3C,CAAC;IAED;;;;OAIG;IACH,MAAM,CAAC,QAAmB;QACzB,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,eAAe,CAAC,QAAQ,CAAC,CAAC,CAAC;IAClD,CAAC;IAED;;;OAGG;IACH,UAAU;QACT,OAAO,eAAe,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;IACvC,CAAC;CACD","sourcesContent":["import type { Message } from './message.ts';\n\n/**\n * What one agent remembers across its runs. The LLM never sees this object, only the\n * context built from it, so how the context is derived (compaction, later) can change\n * without touching either side.\n */\nexport class Conversation {\n\tprivate readonly messages: Message[];\n\n\tconstructor(messages: Message[] = []) {\n\t\tthis.messages = structuredClone(messages);\n\t}\n\n\t/**\n\t * Records a whole run at once, once it succeeded: the prompt, every step the model took and\n\t * every tool result. A failed run records nothing, so a retry cannot repeat its prompt, and\n\t * no tool call is ever kept without its result.\n\t */\n\taddRun(messages: Message[]): void {\n\t\tthis.messages.push(...structuredClone(messages));\n\t}\n\n\t/**\n\t * A deep copy, like everything the conversation takes in: a client that rewrites the\n\t * messages it was sent cannot change what the next turn sends.\n\t */\n\tgetContext(): Message[] {\n\t\treturn structuredClone(this.messages);\n\t}\n}\n"]}

package/dist/llm/domain/llm.d.ts

@@ -1,5 +1,6 @@
 import type { Tokens } from '#src/shared/domain/tokens';
 import type { Message } from './message.ts';
+import type { ToolDefinition } from './tool.ts';
 export type StopReason = 'completed' | 'truncated' | 'refused';
 export interface LLMResponse {
     message: Message & {
@@ -14,5 +15,8 @@ export interface LLMResponse {
     stopReason: StopReason;
 }
 export interface LLMClient {
-    send(context: Message[], signal: AbortSignal): Promise<LLMResponse>;
+    send({ context, tools }: {
+        context: Message[];
+        tools: ToolDefinition[];
+    }, signal: AbortSignal): Promise<LLMResponse>;
 }

package/dist/llm/domain/llm.js.map

@@ -1 +1 @@
-{"version":3,"file":"llm.js","sourceRoot":"","sources":["../../../src/llm/domain/llm.ts"],"names":[],"mappings":"","sourcesContent":["import type { Tokens } from '#src/shared/domain/tokens';\nimport type { Message } from './message.ts';\n\nexport type StopReason = 'completed' | 'truncated' | 'refused';\n\nexport interface LLMResponse {\n\tmessage: Message & { role: 'assistant' };\n\t/**\n\t * Missing when the provider did not report it. The call was made and may have been billed,\n\t * so the agent keeps the answer and only its tokens become unknown; a failure raised from\n\t * such a response is marked `usageUnreported`.\n\t */\n\tusage: Tokens | undefined;\n\tstopReason: StopReason;\n}\n\nexport interface LLMClient {\n\tsend(context: Message[], signal: AbortSignal): Promise<LLMResponse>;\n}\n"]}
+{"version":3,"file":"llm.js","sourceRoot":"","sources":["../../../src/llm/domain/llm.ts"],"names":[],"mappings":"","sourcesContent":["import type { Tokens } from '#src/shared/domain/tokens';\nimport type { Message } from './message.ts';\nimport type { ToolDefinition } from './tool.ts';\n\nexport type StopReason = 'completed' | 'truncated' | 'refused';\n\nexport interface LLMResponse {\n\tmessage: Message & { role: 'assistant' };\n\t/**\n\t * Missing when the provider did not report it. The call was made and may have been billed,\n\t * so the agent keeps the answer and only its tokens become unknown; a failure raised from\n\t * such a response is marked `usageUnreported`.\n\t */\n\tusage: Tokens | undefined;\n\tstopReason: StopReason;\n}\n\nexport interface LLMClient {\n\tsend(\n\t\t{ context, tools }: { context: Message[]; tools: ToolDefinition[] },\n\t\tsignal: AbortSignal,\n\t): Promise<LLMResponse>;\n}\n"]}

package/dist/llm/domain/message.d.ts

@@ -1,11 +1,37 @@
-export type MessagePart = {
+export type MessagePart = TextPart | ReasoningPart | ToolCallPart | ToolResultPart | ProviderDataPart;
+export interface TextPart {
     type: 'text';
     text: string;
-} | {
+}
+export interface ReasoningPart {
     type: 'reasoning';
     text: string;
-};
+}
+export interface ToolCallPart {
+    id: string;
+    type: 'toolCall';
+    name: string;
+    input: unknown;
+}
+export interface ToolResultPart {
+    callId: string;
+    type: 'toolResult';
+    name: string;
+    output: string;
+    isError: boolean;
+}
+/**
+ * A block only the client that produced it can read back, such as signed or encrypted
+ * reasoning, kept whole so that client can send it again on the next call. `source` is the
+ * client's own label and the domain gives it no meaning: every other client leaves the part out.
+ * `data` must stay plain JSON, since the conversation is copied and will be persisted.
+ */
+export interface ProviderDataPart {
+    type: 'providerData';
+    source: string;
+    data: unknown;
+}
 export interface Message {
-    role: 'user' | 'assistant';
+    role: 'user' | 'assistant' | 'tool';
     content: MessagePart[];
 }

package/dist/llm/domain/message.js.map

@@ -1 +1 @@
-{"version":3,"file":"message.js","sourceRoot":"","sources":["../../../src/llm/domain/message.ts"],"names":[],"mappings":"","sourcesContent":["export type MessagePart = { type: 'text'; text: string } | { type: 'reasoning'; text: string };\n\nexport interface Message {\n\trole: 'user' | 'assistant';\n\tcontent: MessagePart[];\n}\n"]}
+{"version":3,"file":"message.js","sourceRoot":"","sources":["../../../src/llm/domain/message.ts"],"names":[],"mappings":"","sourcesContent":["export type MessagePart =\n\tTextPart | ReasoningPart | ToolCallPart | ToolResultPart | ProviderDataPart;\n\nexport interface TextPart {\n\ttype: 'text';\n\ttext: string;\n}\n\nexport interface ReasoningPart {\n\ttype: 'reasoning';\n\ttext: string;\n}\n\nexport interface ToolCallPart {\n\tid: string;\n\ttype: 'toolCall';\n\tname: string;\n\tinput: unknown;\n}\n\nexport interface ToolResultPart {\n\tcallId: string;\n\ttype: 'toolResult';\n\tname: string;\n\toutput: string;\n\tisError: boolean;\n}\n\n/**\n * A block only the client that produced it can read back, such as signed or encrypted\n * reasoning, kept whole so that client can send it again on the next call. `source` is the\n * client's own label and the domain gives it no meaning: every other client leaves the part out.\n * `data` must stay plain JSON, since the conversation is copied and will be persisted.\n */\nexport interface ProviderDataPart {\n\ttype: 'providerData';\n\tsource: string;\n\tdata: unknown;\n}\n\nexport interface Message {\n\trole: 'user' | 'assistant' | 'tool';\n\tcontent: MessagePart[];\n}\n"]}

package/dist/llm/domain/tool.d.ts

@@ -0,0 +1,17 @@
+export interface JSONSchema {
+    type: 'object';
+    [keyword: string]: unknown;
+}
+export interface ToolDefinition {
+    name: string;
+    description: string;
+    inputSchema: JSONSchema;
+}
+export interface Tool extends ToolDefinition {
+    /**
+     * Runs the tool on the input the model sent. The input is unknown because the model can
+     * send anything: validating it is the tool's job, and a rejection returns to the model as
+     * an error result.
+     */
+    execute(input: unknown, signal: AbortSignal): Promise<string>;
+}

package/dist/llm/domain/tool.js

@@ -0,0 +1,2 @@
+export {};
+//# sourceMappingURL=tool.js.map

package/dist/llm/domain/tool.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"tool.js","sourceRoot":"","sources":["../../../src/llm/domain/tool.ts"],"names":[],"mappings":"","sourcesContent":["export interface JSONSchema {\n\ttype: 'object';\n\t[keyword: string]: unknown;\n}\n\nexport interface ToolDefinition {\n\tname: string;\n\tdescription: string;\n\tinputSchema: JSONSchema;\n}\n\nexport interface Tool extends ToolDefinition {\n\t/**\n\t * Runs the tool on the input the model sent. The input is unknown because the model can\n\t * send anything: validating it is the tool's job, and a rejection returns to the model as\n\t * an error result.\n\t */\n\texecute(input: unknown, signal: AbortSignal): Promise<string>;\n}\n"]}

package/dist/llm/infrastructure/claudeLLMClient.d.ts

@@ -1,6 +1,7 @@
 import type { ILogger } from '#src/shared/domain/logger';
 import type { LLMClient, LLMResponse } from '../domain/llm.ts';
 import type { Message } from '../domain/message.ts';
+import type { ToolDefinition } from '../domain/tool.ts';
 export declare const claudeLLMModels: readonly ["claude-opus-5-5", "claude-fable-5-1", "claude-sonnet-5"];
 export type ClaudeLLMModel = (typeof claudeLLMModels)[number];
 /**
@@ -17,5 +18,8 @@ export declare class ClaudeLLMClient implements LLMClient {
         systemPrompt: string;
         logger: ILogger;
     });
-    send(context: Message[], signal: AbortSignal): Promise<LLMResponse>;
+    send({ context, tools }: {
+        context: Message[];
+        tools: ToolDefinition[];
+    }, signal: AbortSignal): Promise<LLMResponse>;
 }

package/dist/llm/infrastructure/claudeLLMClient.js

@@ -13,31 +13,86 @@ const unrecoverableStatuses = [400, 401, 403, 404, 413, 422];
 // The account is out of credit, or its key cannot make the request. Resending cannot fix
 // either, whatever status the error arrives with.
 const unrecoverableErrorTypes = ['billing_error', 'authentication_error', 'permission_error'];
+// This client's label on the thinking blocks it keeps for replay; see `ProviderDataPart`.
+const claudeSource = 'claude';
 /**
- * Only the text crosses back. Anthropic needs a thinking block again only to continue a turn
- * that called a tool, and none is sent yet, so a reasoning part stays in the conversation for
- * narration and is left out of the request. A message with no text is left out whole; the API
- * joins the consecutive user turns that leaves.
+ * Anthropic rejects a `tool_use` whose input is not an object, with a 400 on every later call.
+ * Another provider's call can hold anything else, such as OpenAI arguments that did not parse;
+ * its result already told the model the call failed, so sending it as empty loses nothing.
+ */
+function toClaudeInputObject(input) {
+    return typeof input === 'object' && input !== null && !Array.isArray(input) ? input : {};
+}
+/** Thinking crosses back only as the signed block this client kept, never as its text. */
+function toClaudeBlocks(part) {
+    switch (part.type) {
+        case 'text':
+            // The API rejects an empty text block.
+            return part.text ? [{ type: 'text', text: part.text }] : [];
+        case 'toolCall':
+            return [
+                { type: 'tool_use', id: part.id, name: part.name, input: toClaudeInputObject(part.input) },
+            ];
+        case 'toolResult':
+            return [
+                {
+                    type: 'tool_result',
+                    tool_use_id: part.callId,
+                    content: part.output,
+                    is_error: part.isError,
+                },
+            ];
+        case 'providerData':
+            // Another provider's block would be rejected; losing it only costs that provider's reasoning.
+            return part.source === claudeSource
+                ? [part.data]
+                : [];
+        case 'reasoning':
+            // Unsigned text cannot be replayed as thinking. The API accepts a turn without it.
+            return [];
+    }
+}
+/**
+ * Tool results travel in a user turn, as the API requires. A message left with no blocks is
+ * left out whole; the API joins the consecutive user turns that leaves.
  */
 function toClaudeInput(message) {
-    const text = message.content
-        .filter(part => part.type === 'text')
-        .map(part => part.text)
-        .join('\n');
-    return text ? [{ role: message.role, content: text }] : [];
+    const content = message.content.flatMap(toClaudeBlocks);
+    if (content.length === 0) {
+        return [];
+    }
+    return [{ role: message.role === 'tool' ? 'user' : message.role, content }];
+}
+/**
+ * Strict: the API constrains the model to the schema, so its input always parses and matches.
+ * A schema outside what strict mode supports fails the request with a 400.
+ */
+function toClaudeTool({ name, description, inputSchema }) {
+    return { name, description, input_schema: inputSchema, strict: true };
 }
 function describeBlock(block, logger) {
     switch (block.type) {
         case 'text':
             return [{ type: 'text', text: block.text }];
-        case 'thinking':
-            // Summarized thinking; the signature only matters for replay, which is not done.
-            return block.thinking ? [{ type: 'reasoning', text: block.thinking }] : [];
-        case 'redacted_thinking':
-            // Encrypted by design: there is nothing to narrate.
-            return [];
+        case 'thinking': {
+            // Kept signed for replay; the summarized text is only for narration.
+            const signed = {
+                type: 'thinking',
+                thinking: block.thinking,
+                signature: block.signature,
+            };
+            const replay = { type: 'providerData', source: claudeSource, data: signed };
+            return block.thinking ? [{ type: 'reasoning', text: block.thinking }, replay] : [replay];
+        }
+        case 'redacted_thinking': {
+            // Encrypted by design: nothing to narrate, but it still goes back.
+            const redacted = { type: 'redacted_thinking', data: block.data };
+            return [{ type: 'providerData', source: claudeSource, data: redacted }];
+        }
+        case 'tool_use':
+            return [{ type: 'toolCall', id: block.id, name: block.name, input: block.input }];
         default:
-            // No tools are sent yet, so any other block is output this client does not expect.
+            // Server tools are never offered, so any other block is output this client does not expect.
             treatErrors(() => {
                 logger.warn(block, 'Claude returned a content block the client does not map');
             }, classifyHostFailure, 'Claude client logger failed while mapping a response');
@@ -46,7 +101,9 @@ function describeBlock(block, logger) {
 }
 function toStopReason(response) {
     switch (response.stop_reason) {
+        // `tool_use` is a complete answer that asks for tools: running them is the agent's call.
         case 'end_turn':
+        case 'tool_use':
             return 'completed';
         case 'max_tokens':
             return 'truncated';
@@ -57,8 +114,8 @@ function toStopReason(response) {
                 cause: 'The response stopped at the model context window.',
             });
         default:
-            // `tool_use`, `pause_turn` and `stop_sequence` need tools, server tools or stop
-            // sequences, and this client sends none of them.
+            // `pause_turn` and `stop_sequence` need server tools or stop sequences, and this
+            // client sends neither.
             throw new UnrecoverableError('The Claude response stopped unexpectedly', {
                 cause: `The response stopped with "${response.stop_reason ?? 'unknown'}".`,
             });
@@ -164,7 +221,7 @@ export class ClaudeLLMClient {
         this.systemPrompt = systemPrompt;
         this.logger = logger;
     }
-    async send(context, signal) {
+    async send({ context, tools }, signal) {
         let response;
         try {
             response = await this.client.messages.create({
@@ -172,6 +229,8 @@ export class ClaudeLLMClient {
                 max_tokens: maxOutputTokens,
                 system: this.systemPrompt,
                 messages: context.flatMap(toClaudeInput),
+                // Left out when empty, so a request without tools stays as it always was.
+                ...(tools.length > 0 && { tools: tools.map(toClaudeTool) }),
                 thinking: { type: 'adaptive', display: 'summarized' },
                 // Anthropic caches only on request. This marks the last block, so the next call
                 // reads everything before it from the cache instead of paying for it again.

package/dist/llm/infrastructure/claudeLLMClient.js.map

@@ -1 +1 @@
-{"version":3,"file":"claudeLLMClient.js","sourceRoot":"","sources":["../../../src/llm/infrastructure/claudeLLMClient.ts"],"names":[],"mappings":"AAAA,OAAO,SAAS,EAAE,EAAE,cAAc,EAAE,QAAQ,EAAE,MAAM,mBAAmB,CAAC;AAOxE,OAAO,EACN,uBAAuB,EACvB,kBAAkB,EAClB,eAAe,GACf,MAAM,0BAA0B,CAAC;AAClC,OAAO,EACN,mBAAmB,EACnB,oBAAoB,EACpB,uBAAuB,EACvB,eAAe,EACf,WAAW,GACX,MAAM,mCAAmC,CAAC;AAC3C,OAAO,EAAE,OAAO,EAAE,MAAM,4BAA4B,CAAC;AAGrD,OAAO,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AAItD,+FAA+F;AAC/F,MAAM,CAAC,MAAM,eAAe,GAAG,CAAC,iBAAiB,EAAE,kBAAkB,EAAE,iBAAiB,CAAU,CAAC;AAGnG,gGAAgG;AAChG,yEAAyE;AACzE,MAAM,eAAe,GAAG,MAAM,CAAC;AAE/B,kFAAkF;AAClF,MAAM,qBAAqB,GAAG,CAAC,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,CAAC,CAAC;AAE7D,yFAAyF;AACzF,kDAAkD;AAClD,MAAM,uBAAuB,GAAG,CAAC,eAAe,EAAE,sBAAsB,EAAE,kBAAkB,CAAC,CAAC;AAE9F;;;;;GAKG;AACH,SAAS,aAAa,CAAC,OAAgB;IACtC,MAAM,IAAI,GAAG,OAAO,CAAC,OAAO;SAC1B,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,KAAK,MAAM,CAAC;SACpC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC;SACtB,IAAI,CAAC,IAAI,CAAC,CAAC;IAEb,OAAO,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;AAC5D,CAAC;AAED,SAAS,aAAa,CAAC,KAAmB,EAAE,MAAe;IAC1D,QAAQ,KAAK,CAAC,IAAI,EAAE,CAAC;QACpB,KAAK,MAAM;YACV,OAAO,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC;QAC7C,KAAK,UAAU;YACd,iFAAiF;YACjF,OAAO,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,KAAK,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;QAC5E,KAAK,mBAAmB;YACvB,oDAAoD;YACpD,OAAO,EAAE,CAAC;QACX;YACC,mFAAmF;YACnF,WAAW,CACV,GAAG,EAAE;gBACJ,MAAM,CAAC,IAAI,CAAC,KAAK,EAAE,yDAAyD,CAAC,CAAC;YAC/E,CAAC,EACD,mBAAmB,EACnB,sDAAsD,CACtD,CAAC;YACF,OAAO,EAAE,CAAC;IACZ,CAAC;AACF,CAAC;AAED,SAAS,YAAY,CAAC,QAA0B;IAC/C,QAAQ,QAAQ,CAAC,WAAW,EAAE,CAAC;QAC9B,KAAK,UAAU;YACd,OAAO,WAAW,CAAC;QACpB,KAAK,YAAY;YAChB,OAAO,WAAW,CAAC;QACpB,KAAK,SAAS;YACb,OAAO,SAAS,CAAC;QAClB,KAAK,+BAA+B;YACnC,MAAM,IAAI,eAAe,CAAC,6CAA6C,EAAE;gBACxE,KAAK,EAAE,mDAAmD;aAC1D,CAAC,CAAC;QACJ;YACC,gFAAgF;YAChF,iDAAiD;YACjD,MAAM,IAAI,kBAAkB,CAAC,0CAA0C,EAAE;gBACxE,KAAK,EAAE,8BAA8B,QAAQ,CAAC,WAAW,IAAI,SAAS,IAAI;aAC1E,CAAC,CAAC;IACL,CAAC;AACF,CAAC;AAED,wFAAwF;AACxF,SAAS,QAAQ,CAAC,KAAwB,EAAE,MAAe;IAC1D,IAAI,CAAC,KAAK,EAAE,CAAC;QACZ,+DAA+D;QAC/D,WAAW,CACV,GAAG,EAAE;YACJ,MAAM,CAAC,IAAI,CAAC,sEAAsE,CAAC,CAAC;QACrF,CAAC,EACD,mBAAmB,EACnB,iDAAiD,CACjD,CAAC;QACF,OAAO,SAAS,CAAC;IAClB,CAAC;IAED,OAAO;QACN,WAAW,EAAE,KAAK,CAAC,YAAY;QAC/B,eAAe,EAAE,KAAK,CAAC,uBAAuB,IAAI,CAAC;QACnD,kBAAkB,EAAE,KAAK,CAAC,2BAA2B,IAAI,CAAC;QAC1D,YAAY,EAAE,KAAK,CAAC,aAAa;KACjC,CAAC;AACH,CAAC;AAED,SAAS,aAAa,CAAC,QAA0B,EAAE,MAAe;IACjE,IAAI,KAAyB,CAAC;IAE9B,IAAI,CAAC;QACJ,sFAAsF;QACtF,KAAK,GAAG,QAAQ,CAAC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;QACzC,OAAO;YACN,OAAO,EAAE;gBACR,IAAI,EAAE,WAAW;gBACjB,OAAO,EAAE,QAAQ,CAAC,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC,aAAa,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;aACxE;YACD,KAAK;YACL,UAAU,EAAE,YAAY,CAAC,QAAQ,CAAC;SAClC,CAAC;IACH,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,wFAAwF;QACxF,0DAA0D;QAC1D,MAAM,eAAe,CAAC,KAAK,EAAE,KAAK,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;IACtD,CAAC;AACF,CAAC;AAED;;;;;GAKG;AACH,SAAS,iBAAiB,CAAC,KAAc;IACxC,OAAO,CACN,CAAC,KAAK,YAAY,cAAc,IAAI,CAAC,CAAC,KAAK,YAAY,QAAQ,CAAC,CAAC;QACjE,CAAC,KAAK,YAAY,KAAK,IAAI,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,yCAAyC,CAAC,CAAC,CAC/F,CAAC;AACH,CAAC;AAED,SAAS,qBAAqB,CAAC,KAAc;IAC5C,IAAI,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;QAC9B,qEAAqE;QACrE,OAAO,IAAI,kBAAkB,CAAC,2CAA2C,EAAE;YAC1E,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;SAC7B,CAAC,CAAC;IACJ,CAAC;IACD,IAAI,KAAK,YAAY,QAAQ,EAAE,CAAC;QAC/B,uFAAuF;QACvF,IAAI,KAAK,CAAC,MAAM,KAAK,GAAG,IAAI,KAAK,CAAC,OAAO,CAAC,QAAQ,CAAC,oBAAoB,CAAC,EAAE,CAAC;YAC1E,OAAO,IAAI,eAAe,CAAC,6CAA6C,EAAE;gBACzE,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;QACD,IACC,uBAAuB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,KAAK,KAAK,CAAC,IAAI,CAAC;YACzD,qBAAqB,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,CAAC,EAC5D,CAAC;YACF,OAAO,IAAI,kBAAkB,CAAC,6BAA6B,EAAE;gBAC5D,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;IACF,CAAC;IAED,4FAA4F;IAC5F,OAAO,uBAAuB,CAAC,KAAK,EAAE,2BAA2B,CAAC,CAAC;AACpE,CAAC;AAED;;;GAGG;AACH,MAAM,OAAO,eAAe;IACV,MAAM,CAAY;IAClB,KAAK,CAAiB;IACtB,YAAY,CAAS;IACrB,MAAM,CAAU;IAEjC,0FAA0F;IAC1F,YAAY,EACX,KAAK,EACL,YAAY,EACZ,MAAM,GAKN;QACA,IAAI,CAAC,OAAO,CAAC,eAAe,EAAE,KAAK,CAAC,EAAE,CAAC;YACtC,MAAM,IAAI,uBAAuB,CAChC,IAAI,KAAK,iDAAiD,eAAe,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CACtF,CAAC;QACH,CAAC;QAED,IAAI,CAAC;YACJ,uFAAuF;YACvF,0DAA0D;YAC1D,IAAI,CAAC,MAAM,GAAG,IAAI,SAAS,EAAE,CAAC;QAC/B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,oFAAoF;YACpF,iFAAiF;YACjF,MAAM,IAAI,kBAAkB,CAAC,uCAAuC,EAAE;gBACrE,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;QAED,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,YAAY,GAAG,YAAY,CAAC;QACjC,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IACtB,CAAC;IAED,KAAK,CAAC,IAAI,CAAC,OAAkB,EAAE,MAAmB;QACjD,IAAI,QAA0B,CAAC;QAE/B,IAAI,CAAC;YACJ,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAC3C;gBACC,KAAK,EAAE,IAAI,CAAC,KAAK;gBACjB,UAAU,EAAE,eAAe;gBAC3B,MAAM,EAAE,IAAI,CAAC,YAAY;gBACzB,QAAQ,EAAE,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC;gBACxC,QAAQ,EAAE,EAAE,IAAI,EAAE,UAAU,EAAE,OAAO,EAAE,YAAY,EAAE;gBACrD,gFAAgF;gBAChF,4EAA4E;gBAC5E,aAAa,EAAE,EAAE,IAAI,EAAE,WAAW,EAAE;aACpC,EACD,EAAE,MAAM,EAAE,CACV,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,mFAAmF;YACnF,MAAM,CAAC,cAAc,EAAE,CAAC;YACxB,MAAM,qBAAqB,CAAC,KAAK,CAAC,CAAC;QACpC,CAAC;QAED,OAAO,WAAW,CACjB,GAAG,EAAE,CAAC,aAAa,CAAC,QAAQ,EAAE,IAAI,CAAC,MAAM,CAAC,EAC1C,oBAAoB,EACpB,gCAAgC,CAChC,CAAC;IACH,CAAC;CACD","sourcesContent":["import Anthropic, { AnthropicError, APIError } from '@anthropic-ai/sdk';\nimport type {\n\tContentBlock,\n\tMessage as AnthropicMessage,\n\tMessageParam,\n\tUsage,\n} from '@anthropic-ai/sdk/resources/messages';\nimport {\n\tInvalidAgentConfigError,\n\tUnrecoverableError,\n\twithSpentTokens,\n} from '#src/agent/domain/errors';\nimport {\n\tclassifyHostFailure,\n\tclassifyLocalFailure,\n\tclassifyProviderFailure,\n\tdescribeFailure,\n\ttreatErrors,\n} from '#src/agent/domain/providerFailure';\nimport { isOneOf } from '#src/shared/domain/isOneOf';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport type { Tokens } from '#src/shared/domain/tokens';\nimport { MaxContextError } from '../domain/errors.ts';\nimport type { LLMClient, LLMResponse, StopReason } from '../domain/llm.ts';\nimport type { Message, MessagePart } from '../domain/message.ts';\n\n// The Messages API takes full model IDs, not the Agent SDK's aliases, so this list is its own.\nexport const claudeLLMModels = ['claude-opus-5-5', 'claude-fable-5-1', 'claude-sonnet-5'] as const;\nexport type ClaudeLLMModel = (typeof claudeLLMModels)[number];\n\n// Adaptive thinking spends from this budget too. Above roughly 21,000 the SDK refuses a request\n// that is not streamed, because it could outlast its ten-minute timeout.\nconst maxOutputTokens = 16_000;\n\n// The request itself is wrong or not allowed, so sending it again cannot succeed.\nconst unrecoverableStatuses = [400, 401, 403, 404, 413, 422];\n\n// The account is out of credit, or its key cannot make the request. Resending cannot fix\n// either, whatever status the error arrives with.\nconst unrecoverableErrorTypes = ['billing_error', 'authentication_error', 'permission_error'];\n\n/**\n * Only the text crosses back. Anthropic needs a thinking block again only to continue a turn\n * that called a tool, and none is sent yet, so a reasoning part stays in the conversation for\n * narration and is left out of the request. A message with no text is left out whole; the API\n * joins the consecutive user turns that leaves.\n */\nfunction toClaudeInput(message: Message): MessageParam[] {\n\tconst text = message.content\n\t\t.filter(part => part.type === 'text')\n\t\t.map(part => part.text)\n\t\t.join('\\n');\n\n\treturn text ? [{ role: message.role, content: text }] : [];\n}\n\nfunction describeBlock(block: ContentBlock, logger: ILogger): MessagePart[] {\n\tswitch (block.type) {\n\t\tcase 'text':\n\t\t\treturn [{ type: 'text', text: block.text }];\n\t\tcase 'thinking':\n\t\t\t// Summarized thinking; the signature only matters for replay, which is not done.\n\t\t\treturn block.thinking ? [{ type: 'reasoning', text: block.thinking }] : [];\n\t\tcase 'redacted_thinking':\n\t\t\t// Encrypted by design: there is nothing to narrate.\n\t\t\treturn [];\n\t\tdefault:\n\t\t\t// No tools are sent yet, so any other block is output this client does not expect.\n\t\t\ttreatErrors(\n\t\t\t\t() => {\n\t\t\t\t\tlogger.warn(block, 'Claude returned a content block the client does not map');\n\t\t\t\t},\n\t\t\t\tclassifyHostFailure,\n\t\t\t\t'Claude client logger failed while mapping a response',\n\t\t\t);\n\t\t\treturn [];\n\t}\n}\n\nfunction toStopReason(response: AnthropicMessage): StopReason {\n\tswitch (response.stop_reason) {\n\t\tcase 'end_turn':\n\t\t\treturn 'completed';\n\t\tcase 'max_tokens':\n\t\t\treturn 'truncated';\n\t\tcase 'refusal':\n\t\t\treturn 'refused';\n\t\tcase 'model_context_window_exceeded':\n\t\t\tthrow new MaxContextError('The context no longer fits the Claude model', {\n\t\t\t\tcause: 'The response stopped at the model context window.',\n\t\t\t});\n\t\tdefault:\n\t\t\t// `tool_use`, `pause_turn` and `stop_sequence` need tools, server tools or stop\n\t\t\t// sequences, and this client sends none of them.\n\t\t\tthrow new UnrecoverableError('The Claude response stopped unexpectedly', {\n\t\t\t\tcause: `The response stopped with \"${response.stop_reason ?? 'unknown'}\".`,\n\t\t\t});\n\t}\n}\n\n/** Unlike OpenAI, Anthropic counts cache reads and writes apart from `input_tokens`. */\nfunction toTokens(usage: Usage | undefined, logger: ILogger): Tokens | undefined {\n\tif (!usage) {\n\t\t// Zeros would report a call that may have been billed as free.\n\t\ttreatErrors(\n\t\t\t() => {\n\t\t\t\tlogger.warn('Claude returned a response without usage; it cannot be accounted for');\n\t\t\t},\n\t\t\tclassifyHostFailure,\n\t\t\t'Claude client logger failed while reading usage',\n\t\t);\n\t\treturn undefined;\n\t}\n\n\treturn {\n\t\tinputTokens: usage.input_tokens,\n\t\treadCacheTokens: usage.cache_read_input_tokens ?? 0,\n\t\twrittenCacheTokens: usage.cache_creation_input_tokens ?? 0,\n\t\toutputTokens: usage.output_tokens,\n\t};\n}\n\nfunction toLLMResponse(response: AnthropicMessage, logger: ILogger): LLMResponse {\n\tlet usage: Tokens | undefined;\n\n\ttry {\n\t\t// Inside the try: it warns when the usage is missing, and that warning can throw too.\n\t\tusage = toTokens(response.usage, logger);\n\t\treturn {\n\t\t\tmessage: {\n\t\t\t\trole: 'assistant',\n\t\t\t\tcontent: response.content.flatMap(block => describeBlock(block, logger)),\n\t\t\t},\n\t\t\tusage,\n\t\t\tstopReason: toStopReason(response),\n\t\t};\n\t} catch (error) {\n\t\t// The call was billed whether the response failed or a logger did while mapping it, and\n\t\t// an answer without usage leaves the run's count unknown.\n\t\tthrow withSpentTokens(error, usage, !response.usage);\n\t}\n}\n\n/**\n * The SDK resolves credentials lazily, so a missing key surfaces on the first call, as a plain\n * `Error` the SDK gives no class of its own. An `AnthropicError` that is not an `APIError` never\n * carries an answer from the API: on this client's path it is the SDK refusing to make the\n * call, such as an unresolved profile or a request too large to send without streaming.\n */\nfunction isLocalSDKFailure(error: unknown): boolean {\n\treturn (\n\t\t(error instanceof AnthropicError && !(error instanceof APIError)) ||\n\t\t(error instanceof Error && error.message.startsWith('Could not resolve authentication method'))\n\t);\n}\n\nfunction classifyClaudeFailure(error: unknown): Error {\n\tif (isLocalSDKFailure(error)) {\n\t\t// The API never answered, and calling again would fail the same way.\n\t\treturn new UnrecoverableError('The Claude client cannot send the request', {\n\t\t\tcause: describeFailure(error),\n\t\t});\n\t}\n\tif (error instanceof APIError) {\n\t\t// Anthropic gives an overlong prompt no code of its own, only a 400 with this message.\n\t\tif (error.status === 400 && error.message.includes('prompt is too long')) {\n\t\t\treturn new MaxContextError('The context no longer fits the Claude model', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t\tif (\n\t\t\tunrecoverableErrorTypes.some(type => type === error.type) ||\n\t\t\tunrecoverableStatuses.some(status => status === error.status)\n\t\t) {\n\t\t\treturn new UnrecoverableError('Claude rejected the request', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t}\n\n\t// A rate limit, an overload (529), a 5xx or a network failure: a later attempt can succeed.\n\treturn classifyProviderFailure(error, 'The Claude request failed');\n}\n\n/**\n * A stateless `LLMClient` on the Anthropic Messages API, which keeps nothing between calls:\n * every call carries the whole context, so the conversation is MiKode's, not the provider's.\n */\nexport class ClaudeLLMClient implements LLMClient {\n\tprivate readonly client: Anthropic;\n\tprivate readonly model: ClaudeLLMModel;\n\tprivate readonly systemPrompt: string;\n\tprivate readonly logger: ILogger;\n\n\t// `model` is untyped on purpose: this adapter is the one that knows what Claude supports.\n\tconstructor({\n\t\tmodel,\n\t\tsystemPrompt,\n\t\tlogger,\n\t}: {\n\t\tmodel: string;\n\t\tsystemPrompt: string;\n\t\tlogger: ILogger;\n\t}) {\n\t\tif (!isOneOf(claudeLLMModels, model)) {\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`\"${model}\" is not a Claude API model; expected one of: ${claudeLLMModels.join(', ')}`,\n\t\t\t);\n\t\t}\n\n\t\ttry {\n\t\t\t// The SDK's default retries stay on, as OpenAI's do. See \"The provider SDKs keep their\n\t\t\t// transport retries under RetryingAgent\" in decisions.md.\n\t\t\tthis.client = new Anthropic();\n\t\t} catch (error) {\n\t\t\t// Conflicting credential options, for example: building it again cannot fix them. A\n\t\t\t// missing key does not fail here but on the first call; see `isLocalSDKFailure`.\n\t\t\tthrow new UnrecoverableError('The Claude client could not be set up', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\n\t\tthis.model = model;\n\t\tthis.systemPrompt = systemPrompt;\n\t\tthis.logger = logger;\n\t}\n\n\tasync send(context: Message[], signal: AbortSignal): Promise<LLMResponse> {\n\t\tlet response: AnthropicMessage;\n\n\t\ttry {\n\t\t\tresponse = await this.client.messages.create(\n\t\t\t\t{\n\t\t\t\t\tmodel: this.model,\n\t\t\t\t\tmax_tokens: maxOutputTokens,\n\t\t\t\t\tsystem: this.systemPrompt,\n\t\t\t\t\tmessages: context.flatMap(toClaudeInput),\n\t\t\t\t\tthinking: { type: 'adaptive', display: 'summarized' },\n\t\t\t\t\t// Anthropic caches only on request. This marks the last block, so the next call\n\t\t\t\t\t// reads everything before it from the cache instead of paying for it again.\n\t\t\t\t\tcache_control: { type: 'ephemeral' },\n\t\t\t\t},\n\t\t\t\t{ signal },\n\t\t\t);\n\t\t} catch (error) {\n\t\t\t// The SDK's abort error is named 'Error', so it would pass for a provider failure.\n\t\t\tsignal.throwIfAborted();\n\t\t\tthrow classifyClaudeFailure(error);\n\t\t}\n\n\t\treturn treatErrors(\n\t\t\t() => toLLMResponse(response, this.logger),\n\t\t\tclassifyLocalFailure,\n\t\t\t'Claude response mapping failed',\n\t\t);\n\t}\n}\n"]}
+{"version":3,"file":"claudeLLMClient.js","sourceRoot":"","sources":["../../../src/llm/infrastructure/claudeLLMClient.ts"],"names":[],"mappings":"AAAA,OAAO,SAAS,EAAE,EAAE,cAAc,EAAE,QAAQ,EAAE,MAAM,mBAAmB,CAAC;AAWxE,OAAO,EACN,uBAAuB,EACvB,kBAAkB,EAClB,eAAe,GACf,MAAM,0BAA0B,CAAC;AAClC,OAAO,EACN,mBAAmB,EACnB,oBAAoB,EACpB,uBAAuB,EACvB,eAAe,EACf,WAAW,GACX,MAAM,mCAAmC,CAAC;AAC3C,OAAO,EAAE,OAAO,EAAE,MAAM,4BAA4B,CAAC;AAGrD,OAAO,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AAKtD,+FAA+F;AAC/F,MAAM,CAAC,MAAM,eAAe,GAAG,CAAC,iBAAiB,EAAE,kBAAkB,EAAE,iBAAiB,CAAU,CAAC;AAGnG,gGAAgG;AAChG,yEAAyE;AACzE,MAAM,eAAe,GAAG,MAAM,CAAC;AAE/B,kFAAkF;AAClF,MAAM,qBAAqB,GAAG,CAAC,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,CAAC,CAAC;AAE7D,yFAAyF;AACzF,kDAAkD;AAClD,MAAM,uBAAuB,GAAG,CAAC,eAAe,EAAE,sBAAsB,EAAE,kBAAkB,CAAC,CAAC;AAE9F,0FAA0F;AAC1F,MAAM,YAAY,GAAG,QAAQ,CAAC;AAE9B;;;;GAIG;AACH,SAAS,mBAAmB,CAAC,KAAc;IAC1C,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC;AAC1F,CAAC;AAED,0FAA0F;AAC1F,SAAS,cAAc,CAAC,IAAiB;IACxC,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;QACnB,KAAK,MAAM;YACV,uCAAuC;YACvC,OAAO,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;QAC7D,KAAK,UAAU;YACd,OAAO;gBACN,EAAE,IAAI,EAAE,UAAU,EAAE,EAAE,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,KAAK,EAAE,mBAAmB,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE;aAC1F,CAAC;QACH,KAAK,YAAY;YAChB,OAAO;gBACN;oBACC,IAAI,EAAE,aAAa;oBACnB,WAAW,EAAE,IAAI,CAAC,MAAM;oBACxB,OAAO,EAAE,IAAI,CAAC,MAAM;oBACpB,QAAQ,EAAE,IAAI,CAAC,OAAO;iBACtB;aACD,CAAC;QACH,KAAK,cAAc;YAClB,8FAA8F;YAC9F,OAAO,IAAI,CAAC,MAAM,KAAK,YAAY;gBAClC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAuD,CAAC;gBAChE,CAAC,CAAC,EAAE,CAAC;QACP,KAAK,WAAW;YACf,mFAAmF;YACnF,OAAO,EAAE,CAAC;IACZ,CAAC;AACF,CAAC;AAED;;;GAGG;AACH,SAAS,aAAa,CAAC,OAAgB;IACtC,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC,cAAc,CAAC,CAAC;IACxD,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QAC1B,OAAO,EAAE,CAAC;IACX,CAAC;IAED,OAAO,CAAC,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC;AAC7E,CAAC;AAED;;;GAGG;AACH,SAAS,YAAY,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,WAAW,EAAkB;IACvE,OAAO,EAAE,IAAI,EAAE,WAAW,EAAE,YAAY,EAAE,WAAW,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC;AACvE,CAAC;AAED,SAAS,aAAa,CAAC,KAAmB,EAAE,MAAe;IAC1D,QAAQ,KAAK,CAAC,IAAI,EAAE,CAAC;QACpB,KAAK,MAAM;YACV,OAAO,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC;QAC7C,KAAK,UAAU,CAAC,CAAC,CAAC;YACjB,qEAAqE;YACrE,MAAM,MAAM,GAAuB;gBAClC,IAAI,EAAE,UAAU;gBAChB,QAAQ,EAAE,KAAK,CAAC,QAAQ;gBACxB,SAAS,EAAE,KAAK,CAAC,SAAS;aAC1B,CAAC;YACF,MAAM,MAAM,GAAgB,EAAE,IAAI,EAAE,cAAc,EAAE,MAAM,EAAE,YAAY,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC;YACzF,OAAO,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,KAAK,CAAC,QAAQ,EAAE,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;QAC1F,CAAC;QACD,KAAK,mBAAmB,CAAC,CAAC,CAAC;YAC1B,mEAAmE;YACnE,MAAM,QAAQ,GAA+B,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,CAAC;YAC7F,OAAO,CAAC,EAAE,IAAI,EAAE,cAAc,EAAE,MAAM,EAAE,YAAY,EAAE,IAAI,EAAE,QAAQ,EAAE,CAAC,CAAC;QACzE,CAAC;QACD,KAAK,UAAU;YACd,OAAO,CAAC,EAAE,IAAI,EAAE,UAAU,EAAE,EAAE,EAAE,KAAK,CAAC,EAAE,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,CAAC,CAAC;QACnF;YACC,4FAA4F;YAC5F,WAAW,CACV,GAAG,EAAE;gBACJ,MAAM,CAAC,IAAI,CAAC,KAAK,EAAE,yDAAyD,CAAC,CAAC;YAC/E,CAAC,EACD,mBAAmB,EACnB,sDAAsD,CACtD,CAAC;YACF,OAAO,EAAE,CAAC;IACZ,CAAC;AACF,CAAC;AAED,SAAS,YAAY,CAAC,QAA0B;IAC/C,QAAQ,QAAQ,CAAC,WAAW,EAAE,CAAC;QAC9B,yFAAyF;QACzF,KAAK,UAAU,CAAC;QAChB,KAAK,UAAU;YACd,OAAO,WAAW,CAAC;QACpB,KAAK,YAAY;YAChB,OAAO,WAAW,CAAC;QACpB,KAAK,SAAS;YACb,OAAO,SAAS,CAAC;QAClB,KAAK,+BAA+B;YACnC,MAAM,IAAI,eAAe,CAAC,6CAA6C,EAAE;gBACxE,KAAK,EAAE,mDAAmD;aAC1D,CAAC,CAAC;QACJ;YACC,iFAAiF;YACjF,wBAAwB;YACxB,MAAM,IAAI,kBAAkB,CAAC,0CAA0C,EAAE;gBACxE,KAAK,EAAE,8BAA8B,QAAQ,CAAC,WAAW,IAAI,SAAS,IAAI;aAC1E,CAAC,CAAC;IACL,CAAC;AACF,CAAC;AAED,wFAAwF;AACxF,SAAS,QAAQ,CAAC,KAAwB,EAAE,MAAe;IAC1D,IAAI,CAAC,KAAK,EAAE,CAAC;QACZ,+DAA+D;QAC/D,WAAW,CACV,GAAG,EAAE;YACJ,MAAM,CAAC,IAAI,CAAC,sEAAsE,CAAC,CAAC;QACrF,CAAC,EACD,mBAAmB,EACnB,iDAAiD,CACjD,CAAC;QACF,OAAO,SAAS,CAAC;IAClB,CAAC;IAED,OAAO;QACN,WAAW,EAAE,KAAK,CAAC,YAAY;QAC/B,eAAe,EAAE,KAAK,CAAC,uBAAuB,IAAI,CAAC;QACnD,kBAAkB,EAAE,KAAK,CAAC,2BAA2B,IAAI,CAAC;QAC1D,YAAY,EAAE,KAAK,CAAC,aAAa;KACjC,CAAC;AACH,CAAC;AAED,SAAS,aAAa,CAAC,QAA0B,EAAE,MAAe;IACjE,IAAI,KAAyB,CAAC;IAE9B,IAAI,CAAC;QACJ,sFAAsF;QACtF,KAAK,GAAG,QAAQ,CAAC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;QACzC,OAAO;YACN,OAAO,EAAE;gBACR,IAAI,EAAE,WAAW;gBACjB,OAAO,EAAE,QAAQ,CAAC,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC,aAAa,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;aACxE;YACD,KAAK;YACL,UAAU,EAAE,YAAY,CAAC,QAAQ,CAAC;SAClC,CAAC;IACH,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,wFAAwF;QACxF,0DAA0D;QAC1D,MAAM,eAAe,CAAC,KAAK,EAAE,KAAK,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;IACtD,CAAC;AACF,CAAC;AAED;;;;;GAKG;AACH,SAAS,iBAAiB,CAAC,KAAc;IACxC,OAAO,CACN,CAAC,KAAK,YAAY,cAAc,IAAI,CAAC,CAAC,KAAK,YAAY,QAAQ,CAAC,CAAC;QACjE,CAAC,KAAK,YAAY,KAAK,IAAI,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,yCAAyC,CAAC,CAAC,CAC/F,CAAC;AACH,CAAC;AAED,SAAS,qBAAqB,CAAC,KAAc;IAC5C,IAAI,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;QAC9B,qEAAqE;QACrE,OAAO,IAAI,kBAAkB,CAAC,2CAA2C,EAAE;YAC1E,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;SAC7B,CAAC,CAAC;IACJ,CAAC;IACD,IAAI,KAAK,YAAY,QAAQ,EAAE,CAAC;QAC/B,uFAAuF;QACvF,IAAI,KAAK,CAAC,MAAM,KAAK,GAAG,IAAI,KAAK,CAAC,OAAO,CAAC,QAAQ,CAAC,oBAAoB,CAAC,EAAE,CAAC;YAC1E,OAAO,IAAI,eAAe,CAAC,6CAA6C,EAAE;gBACzE,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;QACD,IACC,uBAAuB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,KAAK,KAAK,CAAC,IAAI,CAAC;YACzD,qBAAqB,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,CAAC,EAC5D,CAAC;YACF,OAAO,IAAI,kBAAkB,CAAC,6BAA6B,EAAE;gBAC5D,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;IACF,CAAC;IAED,4FAA4F;IAC5F,OAAO,uBAAuB,CAAC,KAAK,EAAE,2BAA2B,CAAC,CAAC;AACpE,CAAC;AAED;;;GAGG;AACH,MAAM,OAAO,eAAe;IACV,MAAM,CAAY;IAClB,KAAK,CAAiB;IACtB,YAAY,CAAS;IACrB,MAAM,CAAU;IAEjC,0FAA0F;IAC1F,YAAY,EACX,KAAK,EACL,YAAY,EACZ,MAAM,GAKN;QACA,IAAI,CAAC,OAAO,CAAC,eAAe,EAAE,KAAK,CAAC,EAAE,CAAC;YACtC,MAAM,IAAI,uBAAuB,CAChC,IAAI,KAAK,iDAAiD,eAAe,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CACtF,CAAC;QACH,CAAC;QAED,IAAI,CAAC;YACJ,uFAAuF;YACvF,0DAA0D;YAC1D,IAAI,CAAC,MAAM,GAAG,IAAI,SAAS,EAAE,CAAC;QAC/B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,oFAAoF;YACpF,iFAAiF;YACjF,MAAM,IAAI,kBAAkB,CAAC,uCAAuC,EAAE;gBACrE,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;QAED,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,YAAY,GAAG,YAAY,CAAC;QACjC,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IACtB,CAAC;IAED,KAAK,CAAC,IAAI,CACT,EAAE,OAAO,EAAE,KAAK,EAAmD,EACnE,MAAmB;QAEnB,IAAI,QAA0B,CAAC;QAE/B,IAAI,CAAC;YACJ,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAC3C;gBACC,KAAK,EAAE,IAAI,CAAC,KAAK;gBACjB,UAAU,EAAE,eAAe;gBAC3B,MAAM,EAAE,IAAI,CAAC,YAAY;gBACzB,QAAQ,EAAE,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC;gBACxC,0EAA0E;gBAC1E,GAAG,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,EAAE,CAAC;gBAC3D,QAAQ,EAAE,EAAE,IAAI,EAAE,UAAU,EAAE,OAAO,EAAE,YAAY,EAAE;gBACrD,gFAAgF;gBAChF,4EAA4E;gBAC5E,aAAa,EAAE,EAAE,IAAI,EAAE,WAAW,EAAE;aACpC,EACD,EAAE,MAAM,EAAE,CACV,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,mFAAmF;YACnF,MAAM,CAAC,cAAc,EAAE,CAAC;YACxB,MAAM,qBAAqB,CAAC,KAAK,CAAC,CAAC;QACpC,CAAC;QAED,OAAO,WAAW,CACjB,GAAG,EAAE,CAAC,aAAa,CAAC,QAAQ,EAAE,IAAI,CAAC,MAAM,CAAC,EAC1C,oBAAoB,EACpB,gCAAgC,CAChC,CAAC;IACH,CAAC;CACD","sourcesContent":["import Anthropic, { AnthropicError, APIError } from '@anthropic-ai/sdk';\nimport type {\n\tContentBlock,\n\tContentBlockParam,\n\tMessage as AnthropicMessage,\n\tMessageParam,\n\tRedactedThinkingBlockParam,\n\tThinkingBlockParam,\n\tTool as ClaudeTool,\n\tUsage,\n} from '@anthropic-ai/sdk/resources/messages';\nimport {\n\tInvalidAgentConfigError,\n\tUnrecoverableError,\n\twithSpentTokens,\n} from '#src/agent/domain/errors';\nimport {\n\tclassifyHostFailure,\n\tclassifyLocalFailure,\n\tclassifyProviderFailure,\n\tdescribeFailure,\n\ttreatErrors,\n} from '#src/agent/domain/providerFailure';\nimport { isOneOf } from '#src/shared/domain/isOneOf';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport type { Tokens } from '#src/shared/domain/tokens';\nimport { MaxContextError } from '../domain/errors.ts';\nimport type { LLMClient, LLMResponse, StopReason } from '../domain/llm.ts';\nimport type { Message, MessagePart } from '../domain/message.ts';\nimport type { ToolDefinition } from '../domain/tool.ts';\n\n// The Messages API takes full model IDs, not the Agent SDK's aliases, so this list is its own.\nexport const claudeLLMModels = ['claude-opus-5-5', 'claude-fable-5-1', 'claude-sonnet-5'] as const;\nexport type ClaudeLLMModel = (typeof claudeLLMModels)[number];\n\n// Adaptive thinking spends from this budget too. Above roughly 21,000 the SDK refuses a request\n// that is not streamed, because it could outlast its ten-minute timeout.\nconst maxOutputTokens = 16_000;\n\n// The request itself is wrong or not allowed, so sending it again cannot succeed.\nconst unrecoverableStatuses = [400, 401, 403, 404, 413, 422];\n\n// The account is out of credit, or its key cannot make the request. Resending cannot fix\n// either, whatever status the error arrives with.\nconst unrecoverableErrorTypes = ['billing_error', 'authentication_error', 'permission_error'];\n\n// This client's label on the thinking blocks it keeps for replay; see `ProviderDataPart`.\nconst claudeSource = 'claude';\n\n/**\n * Anthropic rejects a `tool_use` whose input is not an object, with a 400 on every later call.\n * Another provider's call can hold anything else, such as OpenAI arguments that did not parse;\n * its result already told the model the call failed, so sending it as empty loses nothing.\n */\nfunction toClaudeInputObject(input: unknown): unknown {\n\treturn typeof input === 'object' && input !== null && !Array.isArray(input) ? input : {};\n}\n\n/** Thinking crosses back only as the signed block this client kept, never as its text. */\nfunction toClaudeBlocks(part: MessagePart): ContentBlockParam[] {\n\tswitch (part.type) {\n\t\tcase 'text':\n\t\t\t// The API rejects an empty text block.\n\t\t\treturn part.text ? [{ type: 'text', text: part.text }] : [];\n\t\tcase 'toolCall':\n\t\t\treturn [\n\t\t\t\t{ type: 'tool_use', id: part.id, name: part.name, input: toClaudeInputObject(part.input) },\n\t\t\t];\n\t\tcase 'toolResult':\n\t\t\treturn [\n\t\t\t\t{\n\t\t\t\t\ttype: 'tool_result',\n\t\t\t\t\ttool_use_id: part.callId,\n\t\t\t\t\tcontent: part.output,\n\t\t\t\t\tis_error: part.isError,\n\t\t\t\t},\n\t\t\t];\n\t\tcase 'providerData':\n\t\t\t// Another provider's block would be rejected; losing it only costs that provider's reasoning.\n\t\t\treturn part.source === claudeSource\n\t\t\t\t? [part.data as ThinkingBlockParam | RedactedThinkingBlockParam]\n\t\t\t\t: [];\n\t\tcase 'reasoning':\n\t\t\t// Unsigned text cannot be replayed as thinking. The API accepts a turn without it.\n\t\t\treturn [];\n\t}\n}\n\n/**\n * Tool results travel in a user turn, as the API requires. A message left with no blocks is\n * left out whole; the API joins the consecutive user turns that leaves.\n */\nfunction toClaudeInput(message: Message): MessageParam[] {\n\tconst content = message.content.flatMap(toClaudeBlocks);\n\tif (content.length === 0) {\n\t\treturn [];\n\t}\n\n\treturn [{ role: message.role === 'tool' ? 'user' : message.role, content }];\n}\n\n/**\n * Strict: the API constrains the model to the schema, so its input always parses and matches.\n * A schema outside what strict mode supports fails the request with a 400.\n */\nfunction toClaudeTool({ name, description, inputSchema }: ToolDefinition): ClaudeTool {\n\treturn { name, description, input_schema: inputSchema, strict: true };\n}\n\nfunction describeBlock(block: ContentBlock, logger: ILogger): MessagePart[] {\n\tswitch (block.type) {\n\t\tcase 'text':\n\t\t\treturn [{ type: 'text', text: block.text }];\n\t\tcase 'thinking': {\n\t\t\t// Kept signed for replay; the summarized text is only for narration.\n\t\t\tconst signed: ThinkingBlockParam = {\n\t\t\t\ttype: 'thinking',\n\t\t\t\tthinking: block.thinking,\n\t\t\t\tsignature: block.signature,\n\t\t\t};\n\t\t\tconst replay: MessagePart = { type: 'providerData', source: claudeSource, data: signed };\n\t\t\treturn block.thinking ? [{ type: 'reasoning', text: block.thinking }, replay] : [replay];\n\t\t}\n\t\tcase 'redacted_thinking': {\n\t\t\t// Encrypted by design: nothing to narrate, but it still goes back.\n\t\t\tconst redacted: RedactedThinkingBlockParam = { type: 'redacted_thinking', data: block.data };\n\t\t\treturn [{ type: 'providerData', source: claudeSource, data: redacted }];\n\t\t}\n\t\tcase 'tool_use':\n\t\t\treturn [{ type: 'toolCall', id: block.id, name: block.name, input: block.input }];\n\t\tdefault:\n\t\t\t// Server tools are never offered, so any other block is output this client does not expect.\n\t\t\ttreatErrors(\n\t\t\t\t() => {\n\t\t\t\t\tlogger.warn(block, 'Claude returned a content block the client does not map');\n\t\t\t\t},\n\t\t\t\tclassifyHostFailure,\n\t\t\t\t'Claude client logger failed while mapping a response',\n\t\t\t);\n\t\t\treturn [];\n\t}\n}\n\nfunction toStopReason(response: AnthropicMessage): StopReason {\n\tswitch (response.stop_reason) {\n\t\t// `tool_use` is a complete answer that asks for tools: running them is the agent's call.\n\t\tcase 'end_turn':\n\t\tcase 'tool_use':\n\t\t\treturn 'completed';\n\t\tcase 'max_tokens':\n\t\t\treturn 'truncated';\n\t\tcase 'refusal':\n\t\t\treturn 'refused';\n\t\tcase 'model_context_window_exceeded':\n\t\t\tthrow new MaxContextError('The context no longer fits the Claude model', {\n\t\t\t\tcause: 'The response stopped at the model context window.',\n\t\t\t});\n\t\tdefault:\n\t\t\t// `pause_turn` and `stop_sequence` need server tools or stop sequences, and this\n\t\t\t// client sends neither.\n\t\t\tthrow new UnrecoverableError('The Claude response stopped unexpectedly', {\n\t\t\t\tcause: `The response stopped with \"${response.stop_reason ?? 'unknown'}\".`,\n\t\t\t});\n\t}\n}\n\n/** Unlike OpenAI, Anthropic counts cache reads and writes apart from `input_tokens`. */\nfunction toTokens(usage: Usage | undefined, logger: ILogger): Tokens | undefined {\n\tif (!usage) {\n\t\t// Zeros would report a call that may have been billed as free.\n\t\ttreatErrors(\n\t\t\t() => {\n\t\t\t\tlogger.warn('Claude returned a response without usage; it cannot be accounted for');\n\t\t\t},\n\t\t\tclassifyHostFailure,\n\t\t\t'Claude client logger failed while reading usage',\n\t\t);\n\t\treturn undefined;\n\t}\n\n\treturn {\n\t\tinputTokens: usage.input_tokens,\n\t\treadCacheTokens: usage.cache_read_input_tokens ?? 0,\n\t\twrittenCacheTokens: usage.cache_creation_input_tokens ?? 0,\n\t\toutputTokens: usage.output_tokens,\n\t};\n}\n\nfunction toLLMResponse(response: AnthropicMessage, logger: ILogger): LLMResponse {\n\tlet usage: Tokens | undefined;\n\n\ttry {\n\t\t// Inside the try: it warns when the usage is missing, and that warning can throw too.\n\t\tusage = toTokens(response.usage, logger);\n\t\treturn {\n\t\t\tmessage: {\n\t\t\t\trole: 'assistant',\n\t\t\t\tcontent: response.content.flatMap(block => describeBlock(block, logger)),\n\t\t\t},\n\t\t\tusage,\n\t\t\tstopReason: toStopReason(response),\n\t\t};\n\t} catch (error) {\n\t\t// The call was billed whether the response failed or a logger did while mapping it, and\n\t\t// an answer without usage leaves the run's count unknown.\n\t\tthrow withSpentTokens(error, usage, !response.usage);\n\t}\n}\n\n/**\n * The SDK resolves credentials lazily, so a missing key surfaces on the first call, as a plain\n * `Error` the SDK gives no class of its own. An `AnthropicError` that is not an `APIError` never\n * carries an answer from the API: on this client's path it is the SDK refusing to make the\n * call, such as an unresolved profile or a request too large to send without streaming.\n */\nfunction isLocalSDKFailure(error: unknown): boolean {\n\treturn (\n\t\t(error instanceof AnthropicError && !(error instanceof APIError)) ||\n\t\t(error instanceof Error && error.message.startsWith('Could not resolve authentication method'))\n\t);\n}\n\nfunction classifyClaudeFailure(error: unknown): Error {\n\tif (isLocalSDKFailure(error)) {\n\t\t// The API never answered, and calling again would fail the same way.\n\t\treturn new UnrecoverableError('The Claude client cannot send the request', {\n\t\t\tcause: describeFailure(error),\n\t\t});\n\t}\n\tif (error instanceof APIError) {\n\t\t// Anthropic gives an overlong prompt no code of its own, only a 400 with this message.\n\t\tif (error.status === 400 && error.message.includes('prompt is too long')) {\n\t\t\treturn new MaxContextError('The context no longer fits the Claude model', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t\tif (\n\t\t\tunrecoverableErrorTypes.some(type => type === error.type) ||\n\t\t\tunrecoverableStatuses.some(status => status === error.status)\n\t\t) {\n\t\t\treturn new UnrecoverableError('Claude rejected the request', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t}\n\n\t// A rate limit, an overload (529), a 5xx or a network failure: a later attempt can succeed.\n\treturn classifyProviderFailure(error, 'The Claude request failed');\n}\n\n/**\n * A stateless `LLMClient` on the Anthropic Messages API, which keeps nothing between calls:\n * every call carries the whole context, so the conversation is MiKode's, not the provider's.\n */\nexport class ClaudeLLMClient implements LLMClient {\n\tprivate readonly client: Anthropic;\n\tprivate readonly model: ClaudeLLMModel;\n\tprivate readonly systemPrompt: string;\n\tprivate readonly logger: ILogger;\n\n\t// `model` is untyped on purpose: this adapter is the one that knows what Claude supports.\n\tconstructor({\n\t\tmodel,\n\t\tsystemPrompt,\n\t\tlogger,\n\t}: {\n\t\tmodel: string;\n\t\tsystemPrompt: string;\n\t\tlogger: ILogger;\n\t}) {\n\t\tif (!isOneOf(claudeLLMModels, model)) {\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`\"${model}\" is not a Claude API model; expected one of: ${claudeLLMModels.join(', ')}`,\n\t\t\t);\n\t\t}\n\n\t\ttry {\n\t\t\t// The SDK's default retries stay on, as OpenAI's do. See \"The provider SDKs keep their\n\t\t\t// transport retries under RetryingAgent\" in decisions.md.\n\t\t\tthis.client = new Anthropic();\n\t\t} catch (error) {\n\t\t\t// Conflicting credential options, for example: building it again cannot fix them. A\n\t\t\t// missing key does not fail here but on the first call; see `isLocalSDKFailure`.\n\t\t\tthrow new UnrecoverableError('The Claude client could not be set up', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\n\t\tthis.model = model;\n\t\tthis.systemPrompt = systemPrompt;\n\t\tthis.logger = logger;\n\t}\n\n\tasync send(\n\t\t{ context, tools }: { context: Message[]; tools: ToolDefinition[] },\n\t\tsignal: AbortSignal,\n\t): Promise<LLMResponse> {\n\t\tlet response: AnthropicMessage;\n\n\t\ttry {\n\t\t\tresponse = await this.client.messages.create(\n\t\t\t\t{\n\t\t\t\t\tmodel: this.model,\n\t\t\t\t\tmax_tokens: maxOutputTokens,\n\t\t\t\t\tsystem: this.systemPrompt,\n\t\t\t\t\tmessages: context.flatMap(toClaudeInput),\n\t\t\t\t\t// Left out when empty, so a request without tools stays as it always was.\n\t\t\t\t\t...(tools.length > 0 && { tools: tools.map(toClaudeTool) }),\n\t\t\t\t\tthinking: { type: 'adaptive', display: 'summarized' },\n\t\t\t\t\t// Anthropic caches only on request. This marks the last block, so the next call\n\t\t\t\t\t// reads everything before it from the cache instead of paying for it again.\n\t\t\t\t\tcache_control: { type: 'ephemeral' },\n\t\t\t\t},\n\t\t\t\t{ signal },\n\t\t\t);\n\t\t} catch (error) {\n\t\t\t// The SDK's abort error is named 'Error', so it would pass for a provider failure.\n\t\t\tsignal.throwIfAborted();\n\t\t\tthrow classifyClaudeFailure(error);\n\t\t}\n\n\t\treturn treatErrors(\n\t\t\t() => toLLMResponse(response, this.logger),\n\t\t\tclassifyLocalFailure,\n\t\t\t'Claude response mapping failed',\n\t\t);\n\t}\n}\n"]}

package/dist/llm/infrastructure/openAILLMClient.d.ts

@@ -1,6 +1,7 @@
 import type { ILogger } from '#src/shared/domain/logger';
 import type { LLMClient, LLMResponse } from '../domain/llm.ts';
 import type { Message } from '../domain/message.ts';
+import type { ToolDefinition } from '../domain/tool.ts';
 export declare const openAIModels: readonly ["gpt-6-astra", "gpt-5.6-sol", "gpt-5.6-luna", "gpt-5.6-terra"];
 export type OpenAIModel = (typeof openAIModels)[number];
 /**
@@ -18,5 +19,8 @@ export declare class OpenAILLMClient implements LLMClient {
         systemPrompt: string;
         logger: ILogger;
     });
-    send(context: Message[], signal: AbortSignal): Promise<LLMResponse>;
+    send({ context, tools }: {
+        context: Message[];
+        tools: ToolDefinition[];
+    }, signal: AbortSignal): Promise<LLMResponse>;
 }

package/dist/llm/infrastructure/openAILLMClient.js

@@ -21,17 +21,69 @@ const exhaustedQuotaCodes = [
 ];
 // The codes OpenAI gives a failed response that a new attempt can succeed past.
 const transientResponseErrorCodes = ['server_error', 'rate_limit_exceeded'];
+// This client's label on the reasoning items it keeps for replay; see `ProviderDataPart`.
+const openAISource = 'openai';
+// Marks a failed call's output, which would otherwise read the same as a successful one.
+const failedToolPrefix = 'The tool call failed: ';
 /**
- * Only the text crosses back. OpenAI replays reasoning only from the encrypted item it
- * returned, which a `Message` does not keep, so a reasoning part stays in the conversation
- * for narration and is left out of the request. A message with no text is left out whole.
+ * OpenAI takes a flat list of items, so every part becomes its own, in order. Reasoning crosses
+ * back only as the encrypted item this client kept, which OpenAI requires beside the tool calls
+ * that followed it; a reasoning part is its summary, kept for narration.
  */
+function toOpenAIItems(part, role) {
+    switch (part.type) {
+        case 'text':
+            return part.text ? [{ role, content: part.text }] : [];
+        case 'toolCall':
+            return [
+                {
+                    type: 'function_call',
+                    call_id: part.id,
+                    name: part.name,
+                    // A string is arguments that did not parse, sent back as the model wrote them.
+                    arguments: typeof part.input === 'string' ? part.input : JSON.stringify(part.input),
+                },
+            ];
+        case 'toolResult':
+            return [
+                {
+                    type: 'function_call_output',
+                    call_id: part.callId,
+                    // OpenAI has no error flag, so the text says it; the conversation keeps the output as it was.
+                    output: part.isError ? `${failedToolPrefix}${part.output}` : part.output,
+                },
+            ];
+        case 'providerData':
+            // Another provider's item would be rejected; losing it only costs that provider's reasoning.
+            return part.source === openAISource ? [part.data] : [];
+        case 'reasoning':
+            return [];
+    }
+}
 function toOpenAIInput(message) {
-    const text = message.content
-        .filter(part => part.type === 'text')
-        .map(part => part.text)
-        .join('\n');
-    return text ? [{ role: message.role, content: text }] : [];
+    // A tool message holds only results, which carry no role.
+    const role = message.role === 'assistant' ? 'assistant' : 'user';
+    return message.content.flatMap(part => toOpenAIItems(part, role));
+}
+/**
+ * Strict: the API constrains the model to the schema, so its arguments always parse and match.
+ * A schema outside what strict mode supports fails the request with a 400.
+ */
+function toOpenAITool({ name, description, inputSchema }) {
+    return { type: 'function', name, description, parameters: inputSchema, strict: true };
+}
+/**
+ * Strict mode makes arguments that do not parse rare: a response cut short mid-call ends as
+ * `truncated` before any tool runs. Should one arrive, the tool gets the raw string and its
+ * rejection lets the model correct itself, instead of the run failing.
+ */
+function parseArguments(serialized) {
+    try {
+        return JSON.parse(serialized);
+    }
+    catch {
+        return serialized;
+    }
 }
 function describeItem(item, logger) {
     switch (item.type) {
@@ -39,12 +91,32 @@ function describeItem(item, logger) {
             // A refusal is not an answer: it only sets the stop reason.
             return item.content.flatMap(content => content.type === 'output_text' ? [{ type: 'text', text: content.text }] : []);
         case 'reasoning': {
-            // The summary is the readable part; the encrypted content is opaque by design.
+            // The summary is the readable part; the encrypted item is what OpenAI reads back.
             const text = item.summary.map(summary => summary.text).join('\n');
-            return text ? [{ type: 'reasoning', text }] : [];
+            const parts = text ? [{ type: 'reasoning', text }] : [];
+            // Without its encrypted content the item cannot be replayed: OpenAI stores nothing.
+            if (item.encrypted_content) {
+                const encrypted = {
+                    type: 'reasoning',
+                    id: item.id,
+                    summary: item.summary,
+                    encrypted_content: item.encrypted_content,
+                };
+                parts.push({ type: 'providerData', source: openAISource, data: encrypted });
+            }
+            return parts;
         }
+        case 'function_call':
+            return [
+                {
+                    type: 'toolCall',
+                    id: item.call_id,
+                    name: item.name,
+                    input: parseArguments(item.arguments),
+                },
+            ];
         default:
-            // No tools are sent yet, so any other item is output this client does not expect.
+            // No built-in tools are offered, so any other item is output this client does not expect.
             treatErrors(() => {
                 logger.warn(item, 'OpenAI returned an output item the client does not map');
             }, classifyHostFailure, 'OpenAI client logger failed while mapping a response');
@@ -150,15 +222,19 @@ export class OpenAILLMClient {
         this.systemPrompt = systemPrompt;
         this.logger = logger;
     }
-    async send(context, signal) {
+    async send({ context, tools }, signal) {
         let response;
         try {
             response = await this.client.responses.create({
                 model: this.model,
                 instructions: this.systemPrompt,
                 input: context.flatMap(toOpenAIInput),
+                // Left out when empty, so a request without tools stays as it always was.
+                ...(tools.length > 0 && { tools: tools.map(toOpenAITool) }),
                 store: false,
                 reasoning: { summary: 'auto' },
+                // With `store: false` this is the only way to send the reasoning back.
+                include: ['reasoning.encrypted_content'],
             }, { signal });
         }
         catch (error) {

package/dist/llm/infrastructure/openAILLMClient.js.map

@@ -1 +1 @@
-{"version":3,"file":"openAILLMClient.js","sourceRoot":"","sources":["../../../src/llm/infrastructure/openAILLMClient.ts"],"names":[],"mappings":"AAAA,OAAO,MAAM,EAAE,EAAE,QAAQ,EAAE,MAAM,QAAQ,CAAC;AAO1C,OAAO,EACN,uBAAuB,EACvB,gBAAgB,EAChB,kBAAkB,EAClB,eAAe,GACf,MAAM,0BAA0B,CAAC;AAClC,OAAO,EACN,mBAAmB,EACnB,oBAAoB,EACpB,uBAAuB,EACvB,eAAe,EACf,WAAW,GACX,MAAM,mCAAmC,CAAC;AAC3C,OAAO,EAAE,OAAO,EAAE,MAAM,4BAA4B,CAAC;AAGrD,OAAO,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AAItD,+EAA+E;AAC/E,MAAM,CAAC,MAAM,YAAY,GAAG;IAC3B,aAAa;IACb,aAAa;IACb,cAAc;IACd,eAAe;CACN,CAAC;AAGX,kFAAkF;AAClF,MAAM,qBAAqB,GAAG,CAAC,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,CAAC,CAAC;AAExD,gGAAgG;AAChG,mDAAmD;AACnD,MAAM,mBAAmB,GAAG;IAC3B,oBAAoB;IACpB,0BAA0B;IAC1B,8BAA8B;CAC9B,CAAC;AAEF,gFAAgF;AAChF,MAAM,2BAA2B,GAAG,CAAC,cAAc,EAAE,qBAAqB,CAAC,CAAC;AAE5E;;;;GAIG;AACH,SAAS,aAAa,CAAC,OAAgB;IACtC,MAAM,IAAI,GAAG,OAAO,CAAC,OAAO;SAC1B,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,KAAK,MAAM,CAAC;SACpC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC;SACtB,IAAI,CAAC,IAAI,CAAC,CAAC;IAEb,OAAO,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;AAC5D,CAAC;AAED,SAAS,YAAY,CAAC,IAAwB,EAAE,MAAe;IAC9D,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;QACnB,KAAK,SAAS;YACb,4DAA4D;YAC5D,OAAO,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,CACrC,OAAO,CAAC,IAAI,KAAK,aAAa,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,MAAe,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CACrF,CAAC;QACH,KAAK,WAAW,CAAC,CAAC,CAAC;YAClB,+EAA+E;YAC/E,MAAM,IAAI,GAAG,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YAClE,OAAO,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;QAClD,CAAC;QACD;YACC,kFAAkF;YAClF,WAAW,CACV,GAAG,EAAE;gBACJ,MAAM,CAAC,IAAI,CAAC,IAAI,EAAE,wDAAwD,CAAC,CAAC;YAC7E,CAAC,EACD,mBAAmB,EACnB,sDAAsD,CACtD,CAAC;YACF,OAAO,EAAE,CAAC;IACZ,CAAC;AACF,CAAC;AAED,SAAS,YAAY,CAAC,QAAkB;IACvC,IAAI,QAAQ,CAAC,MAAM,KAAK,YAAY,EAAE,CAAC;QACtC,OAAO,QAAQ,CAAC,kBAAkB,EAAE,MAAM,KAAK,gBAAgB,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,WAAW,CAAC;IAC3F,CAAC;IACD,IAAI,QAAQ,CAAC,MAAM,KAAK,WAAW,EAAE,CAAC;QACrC,MAAM,KAAK,GACV,QAAQ,CAAC,KAAK,EAAE,OAAO;YACvB,mCAAmC,QAAQ,CAAC,MAAM,IAAI,SAAS,IAAI,CAAC;QACrE,uFAAuF;QACvF,IAAI,QAAQ,CAAC,KAAK,IAAI,2BAA2B,CAAC,QAAQ,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC;YACjF,MAAM,IAAI,gBAAgB,CAAC,4BAA4B,EAAE,EAAE,KAAK,EAAE,CAAC,CAAC;QACrE,CAAC;QACD,MAAM,IAAI,kBAAkB,CAAC,sCAAsC,EAAE,EAAE,KAAK,EAAE,CAAC,CAAC;IACjF,CAAC;IAED,MAAM,OAAO,GAAG,QAAQ,CAAC,MAAM,CAAC,IAAI,CACnC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,KAAK,SAAS,IAAI,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC,OAAO,CAAC,IAAI,KAAK,SAAS,CAAC,CAC3F,CAAC;IACF,OAAO,OAAO,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,WAAW,CAAC;AAC1C,CAAC;AAED,6EAA6E;AAC7E,SAAS,QAAQ,CAAC,KAAgC,EAAE,MAAe;IAClE,IAAI,CAAC,KAAK,EAAE,CAAC;QACZ,+DAA+D;QAC/D,WAAW,CACV,GAAG,EAAE;YACJ,MAAM,CAAC,IAAI,CAAC,sEAAsE,CAAC,CAAC;QACrF,CAAC,EACD,mBAAmB,EACnB,iDAAiD,CACjD,CAAC;QACF,OAAO,SAAS,CAAC;IAClB,CAAC;IAED,MAAM,EAAE,aAAa,EAAE,kBAAkB,EAAE,GAAG,KAAK,CAAC,oBAAoB,CAAC;IACzE,OAAO;QACN,WAAW,EAAE,KAAK,CAAC,YAAY,GAAG,aAAa,GAAG,kBAAkB;QACpE,eAAe,EAAE,aAAa;QAC9B,kBAAkB,EAAE,kBAAkB;QACtC,YAAY,EAAE,KAAK,CAAC,aAAa;KACjC,CAAC;AACH,CAAC;AAED,SAAS,aAAa,CAAC,QAAkB,EAAE,MAAe;IACzD,IAAI,KAAyB,CAAC;IAE9B,IAAI,CAAC;QACJ,sFAAsF;QACtF,KAAK,GAAG,QAAQ,CAAC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;QACzC,OAAO;YACN,OAAO,EAAE;gBACR,IAAI,EAAE,WAAW;gBACjB,OAAO,EAAE,QAAQ,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC,YAAY,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;aACpE;YACD,KAAK;YACL,UAAU,EAAE,YAAY,CAAC,QAAQ,CAAC;SAClC,CAAC;IACH,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,wFAAwF;QACxF,0DAA0D;QAC1D,MAAM,eAAe,CAAC,KAAK,EAAE,KAAK,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;IACtD,CAAC;AACF,CAAC;AAED,SAAS,qBAAqB,CAAC,KAAc;IAC5C,IAAI,KAAK,YAAY,QAAQ,EAAE,CAAC;QAC/B,IAAI,KAAK,CAAC,IAAI,KAAK,yBAAyB,EAAE,CAAC;YAC9C,OAAO,IAAI,eAAe,CAAC,6CAA6C,EAAE;gBACzE,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;QACD,IACC,mBAAmB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,KAAK,KAAK,CAAC,IAAI,CAAC;YACrD,qBAAqB,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,CAAC,EAC5D,CAAC;YACF,OAAO,IAAI,kBAAkB,CAAC,6BAA6B,EAAE;gBAC5D,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;IACF,CAAC;IAED,OAAO,uBAAuB,CAAC,KAAK,EAAE,2BAA2B,CAAC,CAAC;AACpE,CAAC;AAED;;;;GAIG;AACH,MAAM,OAAO,eAAe;IACV,MAAM,CAAS;IACf,KAAK,CAAc;IACnB,YAAY,CAAS;IACrB,MAAM,CAAU;IAEjC,0FAA0F;IAC1F,YAAY,EACX,KAAK,EACL,YAAY,EACZ,MAAM,GAKN;QACA,IAAI,CAAC,OAAO,CAAC,YAAY,EAAE,KAAK,CAAC,EAAE,CAAC;YACnC,MAAM,IAAI,uBAAuB,CAChC,IAAI,KAAK,8CAA8C,YAAY,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAChF,CAAC;QACH,CAAC;QAED,IAAI,CAAC;YACJ,mFAAmF;YACnF,kGAAkG;YAClG,IAAI,CAAC,MAAM,GAAG,IAAI,MAAM,EAAE,CAAC;QAC5B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,2EAA2E;YAC3E,MAAM,IAAI,kBAAkB,CAAC,uCAAuC,EAAE;gBACrE,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;QAED,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,YAAY,GAAG,YAAY,CAAC;QACjC,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IACtB,CAAC;IAED,KAAK,CAAC,IAAI,CAAC,OAAkB,EAAE,MAAmB;QACjD,IAAI,QAAkB,CAAC;QAEvB,IAAI,CAAC;YACJ,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,MAAM,CAC5C;gBACC,KAAK,EAAE,IAAI,CAAC,KAAK;gBACjB,YAAY,EAAE,IAAI,CAAC,YAAY;gBAC/B,KAAK,EAAE,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC;gBACrC,KAAK,EAAE,KAAK;gBACZ,SAAS,EAAE,EAAE,OAAO,EAAE,MAAM,EAAE;aAC9B,EACD,EAAE,MAAM,EAAE,CACV,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,mFAAmF;YACnF,MAAM,CAAC,cAAc,EAAE,CAAC;YACxB,MAAM,qBAAqB,CAAC,KAAK,CAAC,CAAC;QACpC,CAAC;QAED,OAAO,WAAW,CACjB,GAAG,EAAE,CAAC,aAAa,CAAC,QAAQ,EAAE,IAAI,CAAC,MAAM,CAAC,EAC1C,oBAAoB,EACpB,gCAAgC,CAChC,CAAC;IACH,CAAC;CACD","sourcesContent":["import OpenAI, { APIError } from 'openai';\nimport type {\n\tResponse,\n\tResponseInputItem,\n\tResponseOutputItem,\n\tResponseUsage,\n} from 'openai/resources/responses/responses';\nimport {\n\tInvalidAgentConfigError,\n\tRecoverableError,\n\tUnrecoverableError,\n\twithSpentTokens,\n} from '#src/agent/domain/errors';\nimport {\n\tclassifyHostFailure,\n\tclassifyLocalFailure,\n\tclassifyProviderFailure,\n\tdescribeFailure,\n\ttreatErrors,\n} from '#src/agent/domain/providerFailure';\nimport { isOneOf } from '#src/shared/domain/isOneOf';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport type { Tokens } from '#src/shared/domain/tokens';\nimport { MaxContextError } from '../domain/errors.ts';\nimport type { LLMClient, LLMResponse, StopReason } from '../domain/llm.ts';\nimport type { Message, MessagePart } from '../domain/message.ts';\n\n// The SDK types `model` as a plain string, so this list is maintained by hand.\nexport const openAIModels = [\n\t'gpt-6-astra',\n\t'gpt-5.6-sol',\n\t'gpt-5.6-luna',\n\t'gpt-5.6-terra',\n] as const;\nexport type OpenAIModel = (typeof openAIModels)[number];\n\n// The request itself is wrong or not allowed, so sending it again cannot succeed.\nconst unrecoverableStatuses = [400, 401, 403, 404, 422];\n\n// Out of credit or over a spend limit. These arrive as a 429 like a rate limit, but the account\n// stays blocked however often the request is sent.\nconst exhaustedQuotaCodes = [\n\t'insufficient_quota',\n\t'credit_balance_exhausted',\n\t'project_spend_limit_exceeded',\n];\n\n// The codes OpenAI gives a failed response that a new attempt can succeed past.\nconst transientResponseErrorCodes = ['server_error', 'rate_limit_exceeded'];\n\n/**\n * Only the text crosses back. OpenAI replays reasoning only from the encrypted item it\n * returned, which a `Message` does not keep, so a reasoning part stays in the conversation\n * for narration and is left out of the request. A message with no text is left out whole.\n */\nfunction toOpenAIInput(message: Message): ResponseInputItem[] {\n\tconst text = message.content\n\t\t.filter(part => part.type === 'text')\n\t\t.map(part => part.text)\n\t\t.join('\\n');\n\n\treturn text ? [{ role: message.role, content: text }] : [];\n}\n\nfunction describeItem(item: ResponseOutputItem, logger: ILogger): MessagePart[] {\n\tswitch (item.type) {\n\t\tcase 'message':\n\t\t\t// A refusal is not an answer: it only sets the stop reason.\n\t\t\treturn item.content.flatMap(content =>\n\t\t\t\tcontent.type === 'output_text' ? [{ type: 'text' as const, text: content.text }] : [],\n\t\t\t);\n\t\tcase 'reasoning': {\n\t\t\t// The summary is the readable part; the encrypted content is opaque by design.\n\t\t\tconst text = item.summary.map(summary => summary.text).join('\\n');\n\t\t\treturn text ? [{ type: 'reasoning', text }] : [];\n\t\t}\n\t\tdefault:\n\t\t\t// No tools are sent yet, so any other item is output this client does not expect.\n\t\t\ttreatErrors(\n\t\t\t\t() => {\n\t\t\t\t\tlogger.warn(item, 'OpenAI returned an output item the client does not map');\n\t\t\t\t},\n\t\t\t\tclassifyHostFailure,\n\t\t\t\t'OpenAI client logger failed while mapping a response',\n\t\t\t);\n\t\t\treturn [];\n\t}\n}\n\nfunction toStopReason(response: Response): StopReason {\n\tif (response.status === 'incomplete') {\n\t\treturn response.incomplete_details?.reason === 'content_filter' ? 'refused' : 'truncated';\n\t}\n\tif (response.status !== 'completed') {\n\t\tconst cause =\n\t\t\tresponse.error?.message ??\n\t\t\t`The response ended with status \"${response.status ?? 'unknown'}\".`;\n\t\t// The same failure is retried when it arrives as an SDK error, so it must be here too.\n\t\tif (response.error && transientResponseErrorCodes.includes(response.error.code)) {\n\t\t\tthrow new RecoverableError('The OpenAI response failed', { cause });\n\t\t}\n\t\tthrow new UnrecoverableError('The OpenAI response did not complete', { cause });\n\t}\n\n\tconst refused = response.output.some(\n\t\titem => item.type === 'message' && item.content.some(content => content.type === 'refusal'),\n\t);\n\treturn refused ? 'refused' : 'completed';\n}\n\n/** OpenAI counts both cache reads and cache writes inside `input_tokens`. */\nfunction toTokens(usage: ResponseUsage | undefined, logger: ILogger): Tokens | undefined {\n\tif (!usage) {\n\t\t// Zeros would report a call that may have been billed as free.\n\t\ttreatErrors(\n\t\t\t() => {\n\t\t\t\tlogger.warn('OpenAI returned a response without usage; it cannot be accounted for');\n\t\t\t},\n\t\t\tclassifyHostFailure,\n\t\t\t'OpenAI client logger failed while reading usage',\n\t\t);\n\t\treturn undefined;\n\t}\n\n\tconst { cached_tokens, cache_write_tokens } = usage.input_tokens_details;\n\treturn {\n\t\tinputTokens: usage.input_tokens - cached_tokens - cache_write_tokens,\n\t\treadCacheTokens: cached_tokens,\n\t\twrittenCacheTokens: cache_write_tokens,\n\t\toutputTokens: usage.output_tokens,\n\t};\n}\n\nfunction toLLMResponse(response: Response, logger: ILogger): LLMResponse {\n\tlet usage: Tokens | undefined;\n\n\ttry {\n\t\t// Inside the try: it warns when the usage is missing, and that warning can throw too.\n\t\tusage = toTokens(response.usage, logger);\n\t\treturn {\n\t\t\tmessage: {\n\t\t\t\trole: 'assistant',\n\t\t\t\tcontent: response.output.flatMap(item => describeItem(item, logger)),\n\t\t\t},\n\t\t\tusage,\n\t\t\tstopReason: toStopReason(response),\n\t\t};\n\t} catch (error) {\n\t\t// The call was billed whether the response failed or a logger did while mapping it, and\n\t\t// an answer without usage leaves the run's count unknown.\n\t\tthrow withSpentTokens(error, usage, !response.usage);\n\t}\n}\n\nfunction classifyOpenAIFailure(error: unknown): Error {\n\tif (error instanceof APIError) {\n\t\tif (error.code === 'context_length_exceeded') {\n\t\t\treturn new MaxContextError('The context no longer fits the OpenAI model', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t\tif (\n\t\t\texhaustedQuotaCodes.some(code => code === error.code) ||\n\t\t\tunrecoverableStatuses.some(status => status === error.status)\n\t\t) {\n\t\t\treturn new UnrecoverableError('OpenAI rejected the request', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t}\n\n\treturn classifyProviderFailure(error, 'The OpenAI request failed');\n}\n\n/**\n * A stateless `LLMClient` on the OpenAI Responses API. With `store: false` OpenAI keeps\n * nothing between calls, so every call carries the whole context: the conversation is\n * MiKode's, not the provider's.\n */\nexport class OpenAILLMClient implements LLMClient {\n\tprivate readonly client: OpenAI;\n\tprivate readonly model: OpenAIModel;\n\tprivate readonly systemPrompt: string;\n\tprivate readonly logger: ILogger;\n\n\t// `model` is untyped on purpose: this adapter is the one that knows what OpenAI supports.\n\tconstructor({\n\t\tmodel,\n\t\tsystemPrompt,\n\t\tlogger,\n\t}: {\n\t\tmodel: string;\n\t\tsystemPrompt: string;\n\t\tlogger: ILogger;\n\t}) {\n\t\tif (!isOneOf(openAIModels, model)) {\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`\"${model}\" is not an OpenAI model; expected one of: ${openAIModels.join(', ')}`,\n\t\t\t);\n\t\t}\n\n\t\ttry {\n\t\t\t// The SDK's default retries stay on: they back off and honour `Retry-After`, which\n\t\t\t// `RetryingAgent` does not. See \"The provider SDKs keep their transport retries\" in decisions.md.\n\t\t\tthis.client = new OpenAI();\n\t\t} catch (error) {\n\t\t\t// A missing API key, for example: building the client again cannot fix it.\n\t\t\tthrow new UnrecoverableError('The OpenAI client could not be set up', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\n\t\tthis.model = model;\n\t\tthis.systemPrompt = systemPrompt;\n\t\tthis.logger = logger;\n\t}\n\n\tasync send(context: Message[], signal: AbortSignal): Promise<LLMResponse> {\n\t\tlet response: Response;\n\n\t\ttry {\n\t\t\tresponse = await this.client.responses.create(\n\t\t\t\t{\n\t\t\t\t\tmodel: this.model,\n\t\t\t\t\tinstructions: this.systemPrompt,\n\t\t\t\t\tinput: context.flatMap(toOpenAIInput),\n\t\t\t\t\tstore: false,\n\t\t\t\t\treasoning: { summary: 'auto' },\n\t\t\t\t},\n\t\t\t\t{ signal },\n\t\t\t);\n\t\t} catch (error) {\n\t\t\t// The SDK's abort error is named 'Error', so it would pass for a provider failure.\n\t\t\tsignal.throwIfAborted();\n\t\t\tthrow classifyOpenAIFailure(error);\n\t\t}\n\n\t\treturn treatErrors(\n\t\t\t() => toLLMResponse(response, this.logger),\n\t\t\tclassifyLocalFailure,\n\t\t\t'OpenAI response mapping failed',\n\t\t);\n\t}\n}\n"]}
+{"version":3,"file":"openAILLMClient.js","sourceRoot":"","sources":["../../../src/llm/infrastructure/openAILLMClient.ts"],"names":[],"mappings":"AAAA,OAAO,MAAM,EAAE,EAAE,QAAQ,EAAE,MAAM,QAAQ,CAAC;AAS1C,OAAO,EACN,uBAAuB,EACvB,gBAAgB,EAChB,kBAAkB,EAClB,eAAe,GACf,MAAM,0BAA0B,CAAC;AAClC,OAAO,EACN,mBAAmB,EACnB,oBAAoB,EACpB,uBAAuB,EACvB,eAAe,EACf,WAAW,GACX,MAAM,mCAAmC,CAAC;AAC3C,OAAO,EAAE,OAAO,EAAE,MAAM,4BAA4B,CAAC;AAGrD,OAAO,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AAKtD,+EAA+E;AAC/E,MAAM,CAAC,MAAM,YAAY,GAAG;IAC3B,aAAa;IACb,aAAa;IACb,cAAc;IACd,eAAe;CACN,CAAC;AAGX,kFAAkF;AAClF,MAAM,qBAAqB,GAAG,CAAC,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,CAAC,CAAC;AAExD,gGAAgG;AAChG,mDAAmD;AACnD,MAAM,mBAAmB,GAAG;IAC3B,oBAAoB;IACpB,0BAA0B;IAC1B,8BAA8B;CAC9B,CAAC;AAEF,gFAAgF;AAChF,MAAM,2BAA2B,GAAG,CAAC,cAAc,EAAE,qBAAqB,CAAC,CAAC;AAE5E,0FAA0F;AAC1F,MAAM,YAAY,GAAG,QAAQ,CAAC;AAE9B,yFAAyF;AACzF,MAAM,gBAAgB,GAAG,wBAAwB,CAAC;AAElD;;;;GAIG;AACH,SAAS,aAAa,CAAC,IAAiB,EAAE,IAA0B;IACnE,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;QACnB,KAAK,MAAM;YACV,OAAO,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;QACxD,KAAK,UAAU;YACd,OAAO;gBACN;oBACC,IAAI,EAAE,eAAe;oBACrB,OAAO,EAAE,IAAI,CAAC,EAAE;oBAChB,IAAI,EAAE,IAAI,CAAC,IAAI;oBACf,+EAA+E;oBAC/E,SAAS,EAAE,OAAO,IAAI,CAAC,KAAK,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC;iBACnF;aACD,CAAC;QACH,KAAK,YAAY;YAChB,OAAO;gBACN;oBACC,IAAI,EAAE,sBAAsB;oBAC5B,OAAO,EAAE,IAAI,CAAC,MAAM;oBACpB,8FAA8F;oBAC9F,MAAM,EAAE,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,gBAAgB,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM;iBACxE;aACD,CAAC;QACH,KAAK,cAAc;YAClB,6FAA6F;YAC7F,OAAO,IAAI,CAAC,MAAM,KAAK,YAAY,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAA6B,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;QACjF,KAAK,WAAW;YACf,OAAO,EAAE,CAAC;IACZ,CAAC;AACF,CAAC;AAED,SAAS,aAAa,CAAC,OAAgB;IACtC,0DAA0D;IAC1D,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,KAAK,WAAW,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,MAAM,CAAC;IACjE,OAAO,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC,aAAa,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC;AACnE,CAAC;AAED;;;GAGG;AACH,SAAS,YAAY,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,WAAW,EAAkB;IACvE,OAAO,EAAE,IAAI,EAAE,UAAU,EAAE,IAAI,EAAE,WAAW,EAAE,UAAU,EAAE,WAAW,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC;AACvF,CAAC;AAED;;;;GAIG;AACH,SAAS,cAAc,CAAC,UAAkB;IACzC,IAAI,CAAC;QACJ,OAAO,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC;IAC/B,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,UAAU,CAAC;IACnB,CAAC;AACF,CAAC;AAED,SAAS,YAAY,CAAC,IAAwB,EAAE,MAAe;IAC9D,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;QACnB,KAAK,SAAS;YACb,4DAA4D;YAC5D,OAAO,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,CACrC,OAAO,CAAC,IAAI,KAAK,aAAa,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,MAAe,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CACrF,CAAC;QACH,KAAK,WAAW,CAAC,CAAC,CAAC;YAClB,kFAAkF;YAClF,MAAM,IAAI,GAAG,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YAClE,MAAM,KAAK,GAAkB,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;YACvE,oFAAoF;YACpF,IAAI,IAAI,CAAC,iBAAiB,EAAE,CAAC;gBAC5B,MAAM,SAAS,GAA0B;oBACxC,IAAI,EAAE,WAAW;oBACjB,EAAE,EAAE,IAAI,CAAC,EAAE;oBACX,OAAO,EAAE,IAAI,CAAC,OAAO;oBACrB,iBAAiB,EAAE,IAAI,CAAC,iBAAiB;iBACzC,CAAC;gBACF,KAAK,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,cAAc,EAAE,MAAM,EAAE,YAAY,EAAE,IAAI,EAAE,SAAS,EAAE,CAAC,CAAC;YAC7E,CAAC;YACD,OAAO,KAAK,CAAC;QACd,CAAC;QACD,KAAK,eAAe;YACnB,OAAO;gBACN;oBACC,IAAI,EAAE,UAAU;oBAChB,EAAE,EAAE,IAAI,CAAC,OAAO;oBAChB,IAAI,EAAE,IAAI,CAAC,IAAI;oBACf,KAAK,EAAE,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC;iBACrC;aACD,CAAC;QACH;YACC,0FAA0F;YAC1F,WAAW,CACV,GAAG,EAAE;gBACJ,MAAM,CAAC,IAAI,CAAC,IAAI,EAAE,wDAAwD,CAAC,CAAC;YAC7E,CAAC,EACD,mBAAmB,EACnB,sDAAsD,CACtD,CAAC;YACF,OAAO,EAAE,CAAC;IACZ,CAAC;AACF,CAAC;AAED,SAAS,YAAY,CAAC,QAAkB;IACvC,IAAI,QAAQ,CAAC,MAAM,KAAK,YAAY,EAAE,CAAC;QACtC,OAAO,QAAQ,CAAC,kBAAkB,EAAE,MAAM,KAAK,gBAAgB,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,WAAW,CAAC;IAC3F,CAAC;IACD,IAAI,QAAQ,CAAC,MAAM,KAAK,WAAW,EAAE,CAAC;QACrC,MAAM,KAAK,GACV,QAAQ,CAAC,KAAK,EAAE,OAAO;YACvB,mCAAmC,QAAQ,CAAC,MAAM,IAAI,SAAS,IAAI,CAAC;QACrE,uFAAuF;QACvF,IAAI,QAAQ,CAAC,KAAK,IAAI,2BAA2B,CAAC,QAAQ,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC;YACjF,MAAM,IAAI,gBAAgB,CAAC,4BAA4B,EAAE,EAAE,KAAK,EAAE,CAAC,CAAC;QACrE,CAAC;QACD,MAAM,IAAI,kBAAkB,CAAC,sCAAsC,EAAE,EAAE,KAAK,EAAE,CAAC,CAAC;IACjF,CAAC;IAED,MAAM,OAAO,GAAG,QAAQ,CAAC,MAAM,CAAC,IAAI,CACnC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,KAAK,SAAS,IAAI,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC,OAAO,CAAC,IAAI,KAAK,SAAS,CAAC,CAC3F,CAAC;IACF,OAAO,OAAO,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,WAAW,CAAC;AAC1C,CAAC;AAED,6EAA6E;AAC7E,SAAS,QAAQ,CAAC,KAAgC,EAAE,MAAe;IAClE,IAAI,CAAC,KAAK,EAAE,CAAC;QACZ,+DAA+D;QAC/D,WAAW,CACV,GAAG,EAAE;YACJ,MAAM,CAAC,IAAI,CAAC,sEAAsE,CAAC,CAAC;QACrF,CAAC,EACD,mBAAmB,EACnB,iDAAiD,CACjD,CAAC;QACF,OAAO,SAAS,CAAC;IAClB,CAAC;IAED,MAAM,EAAE,aAAa,EAAE,kBAAkB,EAAE,GAAG,KAAK,CAAC,oBAAoB,CAAC;IACzE,OAAO;QACN,WAAW,EAAE,KAAK,CAAC,YAAY,GAAG,aAAa,GAAG,kBAAkB;QACpE,eAAe,EAAE,aAAa;QAC9B,kBAAkB,EAAE,kBAAkB;QACtC,YAAY,EAAE,KAAK,CAAC,aAAa;KACjC,CAAC;AACH,CAAC;AAED,SAAS,aAAa,CAAC,QAAkB,EAAE,MAAe;IACzD,IAAI,KAAyB,CAAC;IAE9B,IAAI,CAAC;QACJ,sFAAsF;QACtF,KAAK,GAAG,QAAQ,CAAC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;QACzC,OAAO;YACN,OAAO,EAAE;gBACR,IAAI,EAAE,WAAW;gBACjB,OAAO,EAAE,QAAQ,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC,YAAY,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;aACpE;YACD,KAAK;YACL,UAAU,EAAE,YAAY,CAAC,QAAQ,CAAC;SAClC,CAAC;IACH,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,wFAAwF;QACxF,0DAA0D;QAC1D,MAAM,eAAe,CAAC,KAAK,EAAE,KAAK,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;IACtD,CAAC;AACF,CAAC;AAED,SAAS,qBAAqB,CAAC,KAAc;IAC5C,IAAI,KAAK,YAAY,QAAQ,EAAE,CAAC;QAC/B,IAAI,KAAK,CAAC,IAAI,KAAK,yBAAyB,EAAE,CAAC;YAC9C,OAAO,IAAI,eAAe,CAAC,6CAA6C,EAAE;gBACzE,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;QACD,IACC,mBAAmB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,KAAK,KAAK,CAAC,IAAI,CAAC;YACrD,qBAAqB,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,CAAC,EAC5D,CAAC;YACF,OAAO,IAAI,kBAAkB,CAAC,6BAA6B,EAAE;gBAC5D,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;IACF,CAAC;IAED,OAAO,uBAAuB,CAAC,KAAK,EAAE,2BAA2B,CAAC,CAAC;AACpE,CAAC;AAED;;;;GAIG;AACH,MAAM,OAAO,eAAe;IACV,MAAM,CAAS;IACf,KAAK,CAAc;IACnB,YAAY,CAAS;IACrB,MAAM,CAAU;IAEjC,0FAA0F;IAC1F,YAAY,EACX,KAAK,EACL,YAAY,EACZ,MAAM,GAKN;QACA,IAAI,CAAC,OAAO,CAAC,YAAY,EAAE,KAAK,CAAC,EAAE,CAAC;YACnC,MAAM,IAAI,uBAAuB,CAChC,IAAI,KAAK,8CAA8C,YAAY,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAChF,CAAC;QACH,CAAC;QAED,IAAI,CAAC;YACJ,mFAAmF;YACnF,kGAAkG;YAClG,IAAI,CAAC,MAAM,GAAG,IAAI,MAAM,EAAE,CAAC;QAC5B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,2EAA2E;YAC3E,MAAM,IAAI,kBAAkB,CAAC,uCAAuC,EAAE;gBACrE,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;QAED,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,YAAY,GAAG,YAAY,CAAC;QACjC,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IACtB,CAAC;IAED,KAAK,CAAC,IAAI,CACT,EAAE,OAAO,EAAE,KAAK,EAAmD,EACnE,MAAmB;QAEnB,IAAI,QAAkB,CAAC;QAEvB,IAAI,CAAC;YACJ,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,MAAM,CAC5C;gBACC,KAAK,EAAE,IAAI,CAAC,KAAK;gBACjB,YAAY,EAAE,IAAI,CAAC,YAAY;gBAC/B,KAAK,EAAE,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC;gBACrC,0EAA0E;gBAC1E,GAAG,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,EAAE,CAAC;gBAC3D,KAAK,EAAE,KAAK;gBACZ,SAAS,EAAE,EAAE,OAAO,EAAE,MAAM,EAAE;gBAC9B,uEAAuE;gBACvE,OAAO,EAAE,CAAC,6BAA6B,CAAC;aACxC,EACD,EAAE,MAAM,EAAE,CACV,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,mFAAmF;YACnF,MAAM,CAAC,cAAc,EAAE,CAAC;YACxB,MAAM,qBAAqB,CAAC,KAAK,CAAC,CAAC;QACpC,CAAC;QAED,OAAO,WAAW,CACjB,GAAG,EAAE,CAAC,aAAa,CAAC,QAAQ,EAAE,IAAI,CAAC,MAAM,CAAC,EAC1C,oBAAoB,EACpB,gCAAgC,CAChC,CAAC;IACH,CAAC;CACD","sourcesContent":["import OpenAI, { APIError } from 'openai';\nimport type {\n\tFunctionTool,\n\tResponse,\n\tResponseInputItem,\n\tResponseOutputItem,\n\tResponseReasoningItem,\n\tResponseUsage,\n} from 'openai/resources/responses/responses';\nimport {\n\tInvalidAgentConfigError,\n\tRecoverableError,\n\tUnrecoverableError,\n\twithSpentTokens,\n} from '#src/agent/domain/errors';\nimport {\n\tclassifyHostFailure,\n\tclassifyLocalFailure,\n\tclassifyProviderFailure,\n\tdescribeFailure,\n\ttreatErrors,\n} from '#src/agent/domain/providerFailure';\nimport { isOneOf } from '#src/shared/domain/isOneOf';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport type { Tokens } from '#src/shared/domain/tokens';\nimport { MaxContextError } from '../domain/errors.ts';\nimport type { LLMClient, LLMResponse, StopReason } from '../domain/llm.ts';\nimport type { Message, MessagePart } from '../domain/message.ts';\nimport type { ToolDefinition } from '../domain/tool.ts';\n\n// The SDK types `model` as a plain string, so this list is maintained by hand.\nexport const openAIModels = [\n\t'gpt-6-astra',\n\t'gpt-5.6-sol',\n\t'gpt-5.6-luna',\n\t'gpt-5.6-terra',\n] as const;\nexport type OpenAIModel = (typeof openAIModels)[number];\n\n// The request itself is wrong or not allowed, so sending it again cannot succeed.\nconst unrecoverableStatuses = [400, 401, 403, 404, 422];\n\n// Out of credit or over a spend limit. These arrive as a 429 like a rate limit, but the account\n// stays blocked however often the request is sent.\nconst exhaustedQuotaCodes = [\n\t'insufficient_quota',\n\t'credit_balance_exhausted',\n\t'project_spend_limit_exceeded',\n];\n\n// The codes OpenAI gives a failed response that a new attempt can succeed past.\nconst transientResponseErrorCodes = ['server_error', 'rate_limit_exceeded'];\n\n// This client's label on the reasoning items it keeps for replay; see `ProviderDataPart`.\nconst openAISource = 'openai';\n\n// Marks a failed call's output, which would otherwise read the same as a successful one.\nconst failedToolPrefix = 'The tool call failed: ';\n\n/**\n * OpenAI takes a flat list of items, so every part becomes its own, in order. Reasoning crosses\n * back only as the encrypted item this client kept, which OpenAI requires beside the tool calls\n * that followed it; a reasoning part is its summary, kept for narration.\n */\nfunction toOpenAIItems(part: MessagePart, role: 'user' | 'assistant'): ResponseInputItem[] {\n\tswitch (part.type) {\n\t\tcase 'text':\n\t\t\treturn part.text ? [{ role, content: part.text }] : [];\n\t\tcase 'toolCall':\n\t\t\treturn [\n\t\t\t\t{\n\t\t\t\t\ttype: 'function_call',\n\t\t\t\t\tcall_id: part.id,\n\t\t\t\t\tname: part.name,\n\t\t\t\t\t// A string is arguments that did not parse, sent back as the model wrote them.\n\t\t\t\t\targuments: typeof part.input === 'string' ? part.input : JSON.stringify(part.input),\n\t\t\t\t},\n\t\t\t];\n\t\tcase 'toolResult':\n\t\t\treturn [\n\t\t\t\t{\n\t\t\t\t\ttype: 'function_call_output',\n\t\t\t\t\tcall_id: part.callId,\n\t\t\t\t\t// OpenAI has no error flag, so the text says it; the conversation keeps the output as it was.\n\t\t\t\t\toutput: part.isError ? `${failedToolPrefix}${part.output}` : part.output,\n\t\t\t\t},\n\t\t\t];\n\t\tcase 'providerData':\n\t\t\t// Another provider's item would be rejected; losing it only costs that provider's reasoning.\n\t\t\treturn part.source === openAISource ? [part.data as ResponseReasoningItem] : [];\n\t\tcase 'reasoning':\n\t\t\treturn [];\n\t}\n}\n\nfunction toOpenAIInput(message: Message): ResponseInputItem[] {\n\t// A tool message holds only results, which carry no role.\n\tconst role = message.role === 'assistant' ? 'assistant' : 'user';\n\treturn message.content.flatMap(part => toOpenAIItems(part, role));\n}\n\n/**\n * Strict: the API constrains the model to the schema, so its arguments always parse and match.\n * A schema outside what strict mode supports fails the request with a 400.\n */\nfunction toOpenAITool({ name, description, inputSchema }: ToolDefinition): FunctionTool {\n\treturn { type: 'function', name, description, parameters: inputSchema, strict: true };\n}\n\n/**\n * Strict mode makes arguments that do not parse rare: a response cut short mid-call ends as\n * `truncated` before any tool runs. Should one arrive, the tool gets the raw string and its\n * rejection lets the model correct itself, instead of the run failing.\n */\nfunction parseArguments(serialized: string): unknown {\n\ttry {\n\t\treturn JSON.parse(serialized);\n\t} catch {\n\t\treturn serialized;\n\t}\n}\n\nfunction describeItem(item: ResponseOutputItem, logger: ILogger): MessagePart[] {\n\tswitch (item.type) {\n\t\tcase 'message':\n\t\t\t// A refusal is not an answer: it only sets the stop reason.\n\t\t\treturn item.content.flatMap(content =>\n\t\t\t\tcontent.type === 'output_text' ? [{ type: 'text' as const, text: content.text }] : [],\n\t\t\t);\n\t\tcase 'reasoning': {\n\t\t\t// The summary is the readable part; the encrypted item is what OpenAI reads back.\n\t\t\tconst text = item.summary.map(summary => summary.text).join('\\n');\n\t\t\tconst parts: MessagePart[] = text ? [{ type: 'reasoning', text }] : [];\n\t\t\t// Without its encrypted content the item cannot be replayed: OpenAI stores nothing.\n\t\t\tif (item.encrypted_content) {\n\t\t\t\tconst encrypted: ResponseReasoningItem = {\n\t\t\t\t\ttype: 'reasoning',\n\t\t\t\t\tid: item.id,\n\t\t\t\t\tsummary: item.summary,\n\t\t\t\t\tencrypted_content: item.encrypted_content,\n\t\t\t\t};\n\t\t\t\tparts.push({ type: 'providerData', source: openAISource, data: encrypted });\n\t\t\t}\n\t\t\treturn parts;\n\t\t}\n\t\tcase 'function_call':\n\t\t\treturn [\n\t\t\t\t{\n\t\t\t\t\ttype: 'toolCall',\n\t\t\t\t\tid: item.call_id,\n\t\t\t\t\tname: item.name,\n\t\t\t\t\tinput: parseArguments(item.arguments),\n\t\t\t\t},\n\t\t\t];\n\t\tdefault:\n\t\t\t// No built-in tools are offered, so any other item is output this client does not expect.\n\t\t\ttreatErrors(\n\t\t\t\t() => {\n\t\t\t\t\tlogger.warn(item, 'OpenAI returned an output item the client does not map');\n\t\t\t\t},\n\t\t\t\tclassifyHostFailure,\n\t\t\t\t'OpenAI client logger failed while mapping a response',\n\t\t\t);\n\t\t\treturn [];\n\t}\n}\n\nfunction toStopReason(response: Response): StopReason {\n\tif (response.status === 'incomplete') {\n\t\treturn response.incomplete_details?.reason === 'content_filter' ? 'refused' : 'truncated';\n\t}\n\tif (response.status !== 'completed') {\n\t\tconst cause =\n\t\t\tresponse.error?.message ??\n\t\t\t`The response ended with status \"${response.status ?? 'unknown'}\".`;\n\t\t// The same failure is retried when it arrives as an SDK error, so it must be here too.\n\t\tif (response.error && transientResponseErrorCodes.includes(response.error.code)) {\n\t\t\tthrow new RecoverableError('The OpenAI response failed', { cause });\n\t\t}\n\t\tthrow new UnrecoverableError('The OpenAI response did not complete', { cause });\n\t}\n\n\tconst refused = response.output.some(\n\t\titem => item.type === 'message' && item.content.some(content => content.type === 'refusal'),\n\t);\n\treturn refused ? 'refused' : 'completed';\n}\n\n/** OpenAI counts both cache reads and cache writes inside `input_tokens`. */\nfunction toTokens(usage: ResponseUsage | undefined, logger: ILogger): Tokens | undefined {\n\tif (!usage) {\n\t\t// Zeros would report a call that may have been billed as free.\n\t\ttreatErrors(\n\t\t\t() => {\n\t\t\t\tlogger.warn('OpenAI returned a response without usage; it cannot be accounted for');\n\t\t\t},\n\t\t\tclassifyHostFailure,\n\t\t\t'OpenAI client logger failed while reading usage',\n\t\t);\n\t\treturn undefined;\n\t}\n\n\tconst { cached_tokens, cache_write_tokens } = usage.input_tokens_details;\n\treturn {\n\t\tinputTokens: usage.input_tokens - cached_tokens - cache_write_tokens,\n\t\treadCacheTokens: cached_tokens,\n\t\twrittenCacheTokens: cache_write_tokens,\n\t\toutputTokens: usage.output_tokens,\n\t};\n}\n\nfunction toLLMResponse(response: Response, logger: ILogger): LLMResponse {\n\tlet usage: Tokens | undefined;\n\n\ttry {\n\t\t// Inside the try: it warns when the usage is missing, and that warning can throw too.\n\t\tusage = toTokens(response.usage, logger);\n\t\treturn {\n\t\t\tmessage: {\n\t\t\t\trole: 'assistant',\n\t\t\t\tcontent: response.output.flatMap(item => describeItem(item, logger)),\n\t\t\t},\n\t\t\tusage,\n\t\t\tstopReason: toStopReason(response),\n\t\t};\n\t} catch (error) {\n\t\t// The call was billed whether the response failed or a logger did while mapping it, and\n\t\t// an answer without usage leaves the run's count unknown.\n\t\tthrow withSpentTokens(error, usage, !response.usage);\n\t}\n}\n\nfunction classifyOpenAIFailure(error: unknown): Error {\n\tif (error instanceof APIError) {\n\t\tif (error.code === 'context_length_exceeded') {\n\t\t\treturn new MaxContextError('The context no longer fits the OpenAI model', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t\tif (\n\t\t\texhaustedQuotaCodes.some(code => code === error.code) ||\n\t\t\tunrecoverableStatuses.some(status => status === error.status)\n\t\t) {\n\t\t\treturn new UnrecoverableError('OpenAI rejected the request', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t}\n\n\treturn classifyProviderFailure(error, 'The OpenAI request failed');\n}\n\n/**\n * A stateless `LLMClient` on the OpenAI Responses API. With `store: false` OpenAI keeps\n * nothing between calls, so every call carries the whole context: the conversation is\n * MiKode's, not the provider's.\n */\nexport class OpenAILLMClient implements LLMClient {\n\tprivate readonly client: OpenAI;\n\tprivate readonly model: OpenAIModel;\n\tprivate readonly systemPrompt: string;\n\tprivate readonly logger: ILogger;\n\n\t// `model` is untyped on purpose: this adapter is the one that knows what OpenAI supports.\n\tconstructor({\n\t\tmodel,\n\t\tsystemPrompt,\n\t\tlogger,\n\t}: {\n\t\tmodel: string;\n\t\tsystemPrompt: string;\n\t\tlogger: ILogger;\n\t}) {\n\t\tif (!isOneOf(openAIModels, model)) {\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`\"${model}\" is not an OpenAI model; expected one of: ${openAIModels.join(', ')}`,\n\t\t\t);\n\t\t}\n\n\t\ttry {\n\t\t\t// The SDK's default retries stay on: they back off and honour `Retry-After`, which\n\t\t\t// `RetryingAgent` does not. See \"The provider SDKs keep their transport retries\" in decisions.md.\n\t\t\tthis.client = new OpenAI();\n\t\t} catch (error) {\n\t\t\t// A missing API key, for example: building the client again cannot fix it.\n\t\t\tthrow new UnrecoverableError('The OpenAI client could not be set up', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\n\t\tthis.model = model;\n\t\tthis.systemPrompt = systemPrompt;\n\t\tthis.logger = logger;\n\t}\n\n\tasync send(\n\t\t{ context, tools }: { context: Message[]; tools: ToolDefinition[] },\n\t\tsignal: AbortSignal,\n\t): Promise<LLMResponse> {\n\t\tlet response: Response;\n\n\t\ttry {\n\t\t\tresponse = await this.client.responses.create(\n\t\t\t\t{\n\t\t\t\t\tmodel: this.model,\n\t\t\t\t\tinstructions: this.systemPrompt,\n\t\t\t\t\tinput: context.flatMap(toOpenAIInput),\n\t\t\t\t\t// Left out when empty, so a request without tools stays as it always was.\n\t\t\t\t\t...(tools.length > 0 && { tools: tools.map(toOpenAITool) }),\n\t\t\t\t\tstore: false,\n\t\t\t\t\treasoning: { summary: 'auto' },\n\t\t\t\t\t// With `store: false` this is the only way to send the reasoning back.\n\t\t\t\t\tinclude: ['reasoning.encrypted_content'],\n\t\t\t\t},\n\t\t\t\t{ signal },\n\t\t\t);\n\t\t} catch (error) {\n\t\t\t// The SDK's abort error is named 'Error', so it would pass for a provider failure.\n\t\t\tsignal.throwIfAborted();\n\t\t\tthrow classifyOpenAIFailure(error);\n\t\t}\n\n\t\treturn treatErrors(\n\t\t\t() => toLLMResponse(response, this.logger),\n\t\t\tclassifyLocalFailure,\n\t\t\t'OpenAI response mapping failed',\n\t\t);\n\t}\n}\n"]}

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@mikode13/harness",
-	"version": "2.1.0",
+	"version": "2.2.0",
 	"description": "Test/agent harness utility",
 	"type": "module",
 	"license": "SEE LICENSE IN LICENSE",