@mikode13/harness 3.0.0 → 4.0.0

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

@@ -1,4 +1,4 @@
-import type { Agent, AgentResponse, Callback } from '#src/agent/domain/agent';
+import { type Agent, type AgentResponse, type RunOptions } 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/tools/domain/tool';
@@ -15,20 +15,30 @@ export declare class LLMAgent implements Agent {
     private readonly tools;
     private readonly toolDefinitions;
     private readonly maxSteps;
-    constructor({ llmClient, messages, tools, maxSteps, }: {
+    private readonly autoApprove;
+    constructor({ llmClient, messages, tools, maxSteps, autoApprove, }: {
         llmClient: LLMClient;
         messages?: Message[];
         tools?: Tool[];
         /** How many calls to the model one run may make before it fails. */
         maxSteps?: number;
+        /** Runs every tool call without asking, `destructive` ones included. Meant for CI. */
+        autoApprove?: boolean;
     });
     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.
+     * Undefined when the call may run; otherwise why it may not, which the model receives in its
+     * place. Only a `destructive` call is asked about: git can undo a `mutating` one. A throwing
+     * approver ends the run, as any host callback does.
+     */
+    private denial;
+    /**
+     * Always answers the call: a missing tool, a denied call or a failing one becomes an error
+     * result, so the model can correct itself. Only a cancellation, or a failing approver,
+     * 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>;
+    run(prompt: string, options: RunOptions): Promise<AgentResponse>;
 }

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

@@ -1,5 +1,6 @@
-import { InvalidAgentConfigError, RecoverableError, UnrecoverableError, withSpentTokens, } from '#src/agent/domain/errors';
-import { classifyHostFailure, classifyProviderFailure, describeFailure, treatErrors, } from '#src/agent/domain/providerFailure';
+import { ignoreProgress, } from '#src/agent/domain/agent';
+import { InvalidAgentConfigError, RecoverableError, UnrecoverableError, withSpentTokens, } from '#src/shared/domain/errors';
+import { classifyHostFailure, classifyProviderFailure, describeFailure, treatErrors, } from '#src/shared/domain/providerFailure';
 import { Conversation } from '#src/llm/domain/conversation';
 import { addTokens } from '#src/shared/domain/tokens';
 function describePart(part) {
@@ -22,11 +23,25 @@ function describePart(part) {
             return undefined;
     }
 }
+/** Only an answer whose `approved` is a boolean is one; anything else is the approver failing. */
+function readDecision(answer) {
+    if (typeof answer !== 'object' || answer === null) {
+        throw new TypeError('The approver answered without a decision');
+    }
+    const { approved, reason } = answer;
+    if (typeof approved !== 'boolean') {
+        throw new TypeError('The approver answered without a boolean `approved`');
+    }
+    return approved
+        ? { approved }
+        : { approved, reason: typeof reason === 'string' ? reason : undefined };
+}
 function narrate(part, callback) {
     const event = describePart(part);
-    if (!event) {
-        return;
-    }
+    if (event)
+        emit(event, callback);
+}
+function emit(event, callback) {
     treatErrors(() => {
         callback(event);
     }, classifyHostFailure, 'LLM agent progress callback failed');
@@ -44,13 +59,15 @@ export class LLMAgent {
     tools;
     toolDefinitions;
     maxSteps;
-    constructor({ llmClient, messages, tools = [], maxSteps = 25, }) {
+    autoApprove;
+    constructor({ llmClient, messages, tools = [], maxSteps = 25, autoApprove = false, }) {
         if (!Number.isInteger(maxSteps) || maxSteps < 1) {
             throw new InvalidAgentConfigError(`maxSteps must be a positive integer; got ${String(maxSteps)}`);
         }
         this.llmClient = llmClient;
         this.conversation = new Conversation(messages);
         this.maxSteps = maxSteps;
+        this.autoApprove = autoApprove;
         this.tools = new Map(tools.map(tool => [tool.name, tool]));
         if (this.tools.size !== tools.length) {
             // The model calls tools by name, so a second one with the same name could never run.
@@ -86,51 +103,120 @@ export class LLMAgent {
         return response;
     }
     /**
-     * Always answers the call: a missing tool or a failing one becomes an error result, so the
-     * model can correct itself. Only a cancellation escapes, whatever error the tool turned it into.
+     * Undefined when the call may run; otherwise why it may not, which the model receives in its
+     * place. Only a `destructive` call is asked about: git can undo a `mutating` one. A throwing
+     * approver ends the run, as any host callback does.
      */
-    async runTool(call, signal) {
+    async denial(call, risk, signal, approve) {
+        if (this.autoApprove || risk !== 'destructive')
+            return undefined;
+        const retry = 'Do not retry the same call; choose another way, or ask the user.';
+        if (!approve) {
+            return `The call to "${call.name}" was denied: it is destructive, and no one was asked to allow it. ${retry}`;
+        }
+        let decision;
+        try {
+            // A copy, so what runs and what the conversation keeps is exactly what was approved.
+            const input = structuredClone(call.input);
+            // Read inside the boundary: an untyped approver can answer anything, and a bad answer
+            // is its failure, which must carry the run's tokens like any other.
+            decision = readDecision(await approve({ tool: call.name, input, risk }, signal));
+        }
+        catch (error) {
+            // An approver that gives up on a cancellation, whatever it throws, did not fail.
+            signal.throwIfAborted();
+            throw classifyHostFailure(error, 'The tool call approver failed');
+        }
+        // The user may cancel the run while they are being asked.
+        signal.throwIfAborted();
+        if (decision.approved)
+            return undefined;
+        const reason = decision.reason ? ` Their reason: ${decision.reason}` : '';
+        return `The user denied the call to "${call.name}".${reason} ${retry}`;
+    }
+    /**
+     * Always answers the call: a missing tool, a denied call or a failing one becomes an error
+     * result, so the model can correct itself. Only a cancellation, or a failing approver,
+     * escapes, whatever error the tool turned it into.
+     */
+    async runTool(call, { signal, onProgress: callback = ignoreProgress, approve }, effects) {
         const result = { type: 'toolResult', callId: call.id, name: call.name };
         const tool = this.tools.get(call.name);
         if (!tool) {
             return {
-                ...result,
-                output: `There is no tool named "${call.name}". Available tools: ${[...this.tools.keys()].join(', ') || 'none'}.`,
-                isError: true,
+                result: {
+                    ...result,
+                    output: `There is no tool named "${call.name}". Available tools: ${[...this.tools.keys()].join(', ') || 'none'}.`,
+                    isError: true,
+                },
+                denied: false,
+            };
+        }
+        let risk;
+        try {
+            risk = tool.risk(call.input);
+        }
+        catch (error) {
+            // A tool that cannot judge its own call is a failing tool, and the call does not run.
+            return {
+                result: { ...result, output: describeFailure(error), isError: true },
+                denied: false,
             };
         }
+        const denial = await this.denial(call, risk, signal, approve);
+        if (denial !== undefined) {
+            return { result: { ...result, output: denial, isError: true }, denied: true };
+        }
+        // Deciding took at least one await, and the run may have been cancelled meanwhile.
+        signal.throwIfAborted();
+        // Announced only once it will run: a missing, failing-to-judge or denied call never is.
+        narrate(call, callback);
         // Checked again at the last moment: the consumer may cancel on the announcement itself.
         signal.throwIfAborted();
+        // Set before `execute`, so a tool that fails halfway still counts: it may have had effects.
+        effects.toolRan = true;
         try {
-            return { ...result, output: await tool.execute(call.input, signal), isError: false };
+            return {
+                result: { ...result, output: await tool.execute(call.input, signal), isError: false },
+                denied: false,
+            };
         }
         catch (error) {
             signal.throwIfAborted();
-            return { ...result, output: describeFailure(error), isError: true };
+            return {
+                result: { ...result, output: describeFailure(error), isError: true },
+                denied: false,
+            };
         }
     }
     /** One at a time and in order, so their effects and their narration never interleave. */
-    async runTools(calls, signal, callback) {
+    async runTools(calls, options, effects) {
+        const { signal, onProgress: callback = ignoreProgress } = options;
         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);
+            // `runTool` announces the call only once it is allowed to run, so a consumer never
+            // shows a call as running early, nor while the user is still being asked.
+            const { result, denied } = await this.runTool(call, options, effects);
+            if (denied)
+                emit({ type: 'tool', id: call.id, name: call.name, status: 'denied' }, callback);
+            else
+                narrate(result, callback);
             results.push(result);
         }
         return { role: 'tool', content: results };
     }
-    async run(prompt, signal, callback) {
+    async run(prompt, options) {
+        const { signal, onProgress: callback = ignoreProgress } = options;
         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;
+        // Per run, not per instance: what one run executed says nothing about another.
+        const effects = { toolRan: false };
         try {
             for (let step = 1; step <= this.maxSteps; step++) {
                 const response = await this.llmCall(runMessages, signal);
@@ -158,10 +244,7 @@ export class LLMAgent {
                 }
                 // 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));
+                    runMessages.push(await this.runTools(calls, options, effects));
                 }
             }
             throw new UnrecoverableError('The LLM did not finish within its step limit', {
@@ -170,7 +253,7 @@ export class LLMAgent {
         }
         catch (error) {
             let failure = error;
-            if (toolRan && error instanceof RecoverableError) {
+            if (effects.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, {

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,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;AAK5D,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 { ToolDefinition } from '#src/llm/domain/tool';\nimport type { Tool } from '#src/tools/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"]}
+{"version":3,"file":"llmAgent.js","sourceRoot":"","sources":["../../../../src/engines/domain/model/llmAgent.ts"],"names":[],"mappings":"AACA,OAAO,EAMN,cAAc,GACd,MAAM,yBAAyB,CAAC;AACjC,OAAO,EACN,uBAAuB,EACvB,gBAAgB,EAChB,kBAAkB,EAClB,eAAe,GACf,MAAM,2BAA2B,CAAC;AACnC,OAAO,EACN,mBAAmB,EACnB,uBAAuB,EACvB,eAAe,EACf,WAAW,GACX,MAAM,oCAAoC,CAAC;AAC5C,OAAO,EAAE,YAAY,EAAE,MAAM,8BAA8B,CAAC;AAK5D,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,kGAAkG;AAClG,SAAS,YAAY,CAAC,MAAe;IACpC,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,MAAM,KAAK,IAAI,EAAE,CAAC;QACnD,MAAM,IAAI,SAAS,CAAC,0CAA0C,CAAC,CAAC;IACjE,CAAC;IACD,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,GAAG,MAAkD,CAAC;IAChF,IAAI,OAAO,QAAQ,KAAK,SAAS,EAAE,CAAC;QACnC,MAAM,IAAI,SAAS,CAAC,oDAAoD,CAAC,CAAC;IAC3E,CAAC;IAED,OAAO,QAAQ;QACd,CAAC,CAAC,EAAE,QAAQ,EAAE;QACd,CAAC,CAAC,EAAE,QAAQ,EAAE,MAAM,EAAE,OAAO,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,EAAE,CAAC;AAC1E,CAAC;AAUD,SAAS,OAAO,CAAC,IAAiB,EAAE,QAAkB;IACrD,MAAM,KAAK,GAAG,YAAY,CAAC,IAAI,CAAC,CAAC;IACjC,IAAI,KAAK;QAAE,IAAI,CAAC,KAAK,EAAE,QAAQ,CAAC,CAAC;AAClC,CAAC;AAED,SAAS,IAAI,CAAC,KAAoB,EAAE,QAAkB;IACrD,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;IACjB,WAAW,CAAU;IAEtC,YAAY,EACX,SAAS,EACT,QAAQ,EACR,KAAK,GAAG,EAAE,EACV,QAAQ,GAAG,EAAE,EACb,WAAW,GAAG,KAAK,GASnB;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,WAAW,GAAG,WAAW,CAAC;QAC/B,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;;;;OAIG;IACK,KAAK,CAAC,MAAM,CACnB,IAAkB,EAClB,IAAc,EACd,MAAmB,EACnB,OAA6B;QAE7B,IAAI,IAAI,CAAC,WAAW,IAAI,IAAI,KAAK,aAAa;YAAE,OAAO,SAAS,CAAC;QAEjE,MAAM,KAAK,GAAG,kEAAkE,CAAC;QACjF,IAAI,CAAC,OAAO,EAAE,CAAC;YACd,OAAO,gBAAgB,IAAI,CAAC,IAAI,sEAAsE,KAAK,EAAE,CAAC;QAC/G,CAAC;QAED,IAAI,QAA0B,CAAC;QAC/B,IAAI,CAAC;YACJ,qFAAqF;YACrF,MAAM,KAAK,GAAY,eAAe,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;YACnD,sFAAsF;YACtF,oEAAoE;YACpE,QAAQ,GAAG,YAAY,CAAC,MAAM,OAAO,CAAC,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE,EAAE,MAAM,CAAC,CAAC,CAAC;QAClF,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,iFAAiF;YACjF,MAAM,CAAC,cAAc,EAAE,CAAC;YACxB,MAAM,mBAAmB,CAAC,KAAK,EAAE,+BAA+B,CAAC,CAAC;QACnE,CAAC;QACD,0DAA0D;QAC1D,MAAM,CAAC,cAAc,EAAE,CAAC;QACxB,IAAI,QAAQ,CAAC,QAAQ;YAAE,OAAO,SAAS,CAAC;QAExC,MAAM,MAAM,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,kBAAkB,QAAQ,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAC1E,OAAO,gCAAgC,IAAI,CAAC,IAAI,KAAK,MAAM,IAAI,KAAK,EAAE,CAAC;IACxE,CAAC;IAED;;;;OAIG;IACK,KAAK,CAAC,OAAO,CACpB,IAAkB,EAClB,EAAE,MAAM,EAAE,UAAU,EAAE,QAAQ,GAAG,cAAc,EAAE,OAAO,EAAc,EACtE,OAAmB;QAEnB,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,MAAM,EAAE;oBACP,GAAG,MAAM;oBACT,MAAM,EAAE,2BAA2B,IAAI,CAAC,IAAI,uBAAuB,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,MAAM,GAAG;oBACjH,OAAO,EAAE,IAAI;iBACb;gBACD,MAAM,EAAE,KAAK;aACb,CAAC;QACH,CAAC;QAED,IAAI,IAAc,CAAC;QACnB,IAAI,CAAC;YACJ,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QAC9B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,sFAAsF;YACtF,OAAO;gBACN,MAAM,EAAE,EAAE,GAAG,MAAM,EAAE,MAAM,EAAE,eAAe,CAAC,KAAK,CAAC,EAAE,OAAO,EAAE,IAAI,EAAE;gBACpE,MAAM,EAAE,KAAK;aACb,CAAC;QACH,CAAC;QAED,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,CAAC,CAAC;QAC9D,IAAI,MAAM,KAAK,SAAS,EAAE,CAAC;YAC1B,OAAO,EAAE,MAAM,EAAE,EAAE,GAAG,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,IAAI,EAAE,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC;QAC/E,CAAC;QAED,mFAAmF;QACnF,MAAM,CAAC,cAAc,EAAE,CAAC;QACxB,wFAAwF;QACxF,OAAO,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC;QACxB,wFAAwF;QACxF,MAAM,CAAC,cAAc,EAAE,CAAC;QACxB,4FAA4F;QAC5F,OAAO,CAAC,OAAO,GAAG,IAAI,CAAC;QACvB,IAAI,CAAC;YACJ,OAAO;gBACN,MAAM,EAAE,EAAE,GAAG,MAAM,EAAE,MAAM,EAAE,MAAM,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,KAAK,EAAE,MAAM,CAAC,EAAE,OAAO,EAAE,KAAK,EAAE;gBACrF,MAAM,EAAE,KAAK;aACb,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,MAAM,CAAC,cAAc,EAAE,CAAC;YACxB,OAAO;gBACN,MAAM,EAAE,EAAE,GAAG,MAAM,EAAE,MAAM,EAAE,eAAe,CAAC,KAAK,CAAC,EAAE,OAAO,EAAE,IAAI,EAAE;gBACpE,MAAM,EAAE,KAAK;aACb,CAAC;QACH,CAAC;IACF,CAAC;IAED,yFAAyF;IACjF,KAAK,CAAC,QAAQ,CACrB,KAAqB,EACrB,OAAmB,EACnB,OAAmB;QAEnB,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,QAAQ,GAAG,cAAc,EAAE,GAAG,OAAO,CAAC;QAClE,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,mFAAmF;YACnF,0EAA0E;YAC1E,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,GAAG,MAAM,IAAI,CAAC,OAAO,CAAC,IAAI,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;YACtE,IAAI,MAAM;gBAAE,IAAI,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,EAAE,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,MAAM,EAAE,QAAQ,EAAE,EAAE,QAAQ,CAAC,CAAC;;gBACxF,OAAO,CAAC,MAAM,EAAE,QAAQ,CAAC,CAAC;YAC/B,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,OAAmB;QAC5C,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,QAAQ,GAAG,cAAc,EAAE,GAAG,OAAO,CAAC;QAClE,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,+EAA+E;QAC/E,MAAM,OAAO,GAAe,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QAE/C,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,WAAW,CAAC,IAAI,CAAC,MAAM,IAAI,CAAC,QAAQ,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,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,CAAC,OAAO,IAAI,KAAK,YAAY,gBAAgB,EAAE,CAAC;gBAC1D,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 { ApprovalDecision, Approver, ToolRisk } from '#src/agent/domain/approval';\nimport {\n\ttype Agent,\n\ttype AgentResponse,\n\ttype Callback,\n\ttype ProgressEvent,\n\ttype RunOptions,\n\tignoreProgress,\n} from '#src/agent/domain/agent';\nimport {\n\tInvalidAgentConfigError,\n\tRecoverableError,\n\tUnrecoverableError,\n\twithSpentTokens,\n} from '#src/shared/domain/errors';\nimport {\n\tclassifyHostFailure,\n\tclassifyProviderFailure,\n\tdescribeFailure,\n\ttreatErrors,\n} from '#src/shared/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 { ToolDefinition } from '#src/llm/domain/tool';\nimport type { Tool } from '#src/tools/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\n/** Only an answer whose `approved` is a boolean is one; anything else is the approver failing. */\nfunction readDecision(answer: unknown): ApprovalDecision {\n\tif (typeof answer !== 'object' || answer === null) {\n\t\tthrow new TypeError('The approver answered without a decision');\n\t}\n\tconst { approved, reason } = answer as { approved?: unknown; reason?: unknown };\n\tif (typeof approved !== 'boolean') {\n\t\tthrow new TypeError('The approver answered without a boolean `approved`');\n\t}\n\n\treturn approved\n\t\t? { approved }\n\t\t: { approved, reason: typeof reason === 'string' ? reason : undefined };\n}\n\n/**\n * Whether a run reached a tool's `execute`. A missing tool or a denied call never does, so it\n * has no effect a retry could repeat.\n */\ninterface RunEffects {\n\ttoolRan: boolean;\n}\n\nfunction narrate(part: MessagePart, callback: Callback): void {\n\tconst event = describePart(part);\n\tif (event) emit(event, callback);\n}\n\nfunction emit(event: ProgressEvent, callback: Callback): void {\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\tprivate readonly autoApprove: boolean;\n\n\tconstructor({\n\t\tllmClient,\n\t\tmessages,\n\t\ttools = [],\n\t\tmaxSteps = 25,\n\t\tautoApprove = false,\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\t/** Runs every tool call without asking, `destructive` ones included. Meant for CI. */\n\t\tautoApprove?: boolean;\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 ${String(maxSteps)}`,\n\t\t\t);\n\t\t}\n\n\t\tthis.llmClient = llmClient;\n\t\tthis.conversation = new Conversation(messages);\n\t\tthis.maxSteps = maxSteps;\n\t\tthis.autoApprove = autoApprove;\n\t\tthis.tools = new Map(tools.map(tool => [tool.name, tool]));\n\t\tif (this.tools.size !== tools.length) {\n\t\t\t// The model calls tools by name, so a second one with the same name could never run.\n\t\t\tthrow new InvalidAgentConfigError('Two tools share a name; each name must be unique');\n\t\t}\n\t\t// The client only describes the tools to the model; running them stays with the agent.\n\t\tthis.toolDefinitions = tools.map(({ name, description, inputSchema }) => ({\n\t\t\tname,\n\t\t\tdescription,\n\t\t\tinputSchema,\n\t\t}));\n\t}\n\n\tprivate async llmCall(runMessages: Message[], signal: AbortSignal): Promise<LLMResponse> {\n\t\tlet response: LLMResponse;\n\t\ttry {\n\t\t\t// Sent as a copy, so what the run records is what the agent built, whatever the client does.\n\t\t\tresponse = await this.llmClient.send(\n\t\t\t\t{\n\t\t\t\t\tcontext: [...this.conversation.getContext(), ...structuredClone(runMessages)],\n\t\t\t\t\ttools: this.toolDefinitions,\n\t\t\t\t},\n\t\t\t\tsignal,\n\t\t\t);\n\t\t} catch (error) {\n\t\t\tthrow classifyProviderFailure(error, 'The LLM call failed');\n\t\t}\n\n\t\tif (response.stopReason !== 'completed') {\n\t\t\tthrow new UnrecoverableError('The LLM stopped before completing its answer', {\n\t\t\t\tcause: `The model stopped with \"${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 * Undefined when the call may run; otherwise why it may not, which the model receives in its\n\t * place. Only a `destructive` call is asked about: git can undo a `mutating` one. A throwing\n\t * approver ends the run, as any host callback does.\n\t */\n\tprivate async denial(\n\t\tcall: ToolCallPart,\n\t\trisk: ToolRisk,\n\t\tsignal: AbortSignal,\n\t\tapprove: Approver | undefined,\n\t): Promise<string | undefined> {\n\t\tif (this.autoApprove || risk !== 'destructive') return undefined;\n\n\t\tconst retry = 'Do not retry the same call; choose another way, or ask the user.';\n\t\tif (!approve) {\n\t\t\treturn `The call to \"${call.name}\" was denied: it is destructive, and no one was asked to allow it. ${retry}`;\n\t\t}\n\n\t\tlet decision: ApprovalDecision;\n\t\ttry {\n\t\t\t// A copy, so what runs and what the conversation keeps is exactly what was approved.\n\t\t\tconst input: unknown = structuredClone(call.input);\n\t\t\t// Read inside the boundary: an untyped approver can answer anything, and a bad answer\n\t\t\t// is its failure, which must carry the run's tokens like any other.\n\t\t\tdecision = readDecision(await approve({ tool: call.name, input, risk }, signal));\n\t\t} catch (error) {\n\t\t\t// An approver that gives up on a cancellation, whatever it throws, did not fail.\n\t\t\tsignal.throwIfAborted();\n\t\t\tthrow classifyHostFailure(error, 'The tool call approver failed');\n\t\t}\n\t\t// The user may cancel the run while they are being asked.\n\t\tsignal.throwIfAborted();\n\t\tif (decision.approved) return undefined;\n\n\t\tconst reason = decision.reason ? ` Their reason: ${decision.reason}` : '';\n\t\treturn `The user denied the call to \"${call.name}\".${reason} ${retry}`;\n\t}\n\n\t/**\n\t * Always answers the call: a missing tool, a denied call or a failing one becomes an error\n\t * result, so the model can correct itself. Only a cancellation, or a failing approver,\n\t * escapes, whatever error the tool turned it into.\n\t */\n\tprivate async runTool(\n\t\tcall: ToolCallPart,\n\t\t{ signal, onProgress: callback = ignoreProgress, approve }: RunOptions,\n\t\teffects: RunEffects,\n\t): Promise<{ result: ToolResultPart; denied: boolean }> {\n\t\tconst result = { type: 'toolResult' as const, callId: call.id, name: call.name };\n\t\tconst tool = this.tools.get(call.name);\n\t\tif (!tool) {\n\t\t\treturn {\n\t\t\t\tresult: {\n\t\t\t\t\t...result,\n\t\t\t\t\toutput: `There is no tool named \"${call.name}\". Available tools: ${[...this.tools.keys()].join(', ') || 'none'}.`,\n\t\t\t\t\tisError: true,\n\t\t\t\t},\n\t\t\t\tdenied: false,\n\t\t\t};\n\t\t}\n\n\t\tlet risk: ToolRisk;\n\t\ttry {\n\t\t\trisk = tool.risk(call.input);\n\t\t} catch (error) {\n\t\t\t// A tool that cannot judge its own call is a failing tool, and the call does not run.\n\t\t\treturn {\n\t\t\t\tresult: { ...result, output: describeFailure(error), isError: true },\n\t\t\t\tdenied: false,\n\t\t\t};\n\t\t}\n\n\t\tconst denial = await this.denial(call, risk, signal, approve);\n\t\tif (denial !== undefined) {\n\t\t\treturn { result: { ...result, output: denial, isError: true }, denied: true };\n\t\t}\n\n\t\t// Deciding took at least one await, and the run may have been cancelled meanwhile.\n\t\tsignal.throwIfAborted();\n\t\t// Announced only once it will run: a missing, failing-to-judge or denied call never is.\n\t\tnarrate(call, callback);\n\t\t// Checked again at the last moment: the consumer may cancel on the announcement itself.\n\t\tsignal.throwIfAborted();\n\t\t// Set before `execute`, so a tool that fails halfway still counts: it may have had effects.\n\t\teffects.toolRan = true;\n\t\ttry {\n\t\t\treturn {\n\t\t\t\tresult: { ...result, output: await tool.execute(call.input, signal), isError: false },\n\t\t\t\tdenied: false,\n\t\t\t};\n\t\t} catch (error) {\n\t\t\tsignal.throwIfAborted();\n\t\t\treturn {\n\t\t\t\tresult: { ...result, output: describeFailure(error), isError: true },\n\t\t\t\tdenied: false,\n\t\t\t};\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\toptions: RunOptions,\n\t\teffects: RunEffects,\n\t): Promise<Message> {\n\t\tconst { signal, onProgress: callback = ignoreProgress } = options;\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// `runTool` announces the call only once it is allowed to run, so a consumer never\n\t\t\t// shows a call as running early, nor while the user is still being asked.\n\t\t\tconst { result, denied } = await this.runTool(call, options, effects);\n\t\t\tif (denied) emit({ type: 'tool', id: call.id, name: call.name, status: 'denied' }, callback);\n\t\t\telse narrate(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, options: RunOptions): Promise<AgentResponse> {\n\t\tconst { signal, onProgress: callback = ignoreProgress } = options;\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\t// Per run, not per instance: what one run executed says nothing about another.\n\t\tconst effects: RunEffects = { 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\trunMessages.push(await this.runTools(calls, options, effects));\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 (effects.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/factory/infrastructure/agentFactory.js

@@ -1,4 +1,4 @@
-import { InvalidAgentConfigError } from '#src/agent/domain/errors';
+import { InvalidAgentConfigError } from '#src/shared/domain/errors';
 import { ClaudeAgent } from '#src/engines/claude/infrastructure/model/claudeAgent';
 import { CodexAgent } from '#src/engines/codex/infrastructure/model/codexAgent';
 import { InstructedAgent } from '#src/orchestration/domain/model/instructedAgent';

package/dist/factory/infrastructure/agentFactory.js.map

@@ -1 +1 @@
-{"version":3,"file":"agentFactory.js","sourceRoot":"","sources":["../../../src/factory/infrastructure/agentFactory.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,uBAAuB,EAAE,MAAM,0BAA0B,CAAC;AACnE,OAAO,EAAE,WAAW,EAAE,MAAM,sDAAsD,CAAC;AACnF,OAAO,EAAE,UAAU,EAAE,MAAM,oDAAoD,CAAC;AAChF,OAAO,EAAE,eAAe,EAAE,MAAM,iDAAiD,CAAC;AAClF,OAAO,EACN,oBAAoB,EACpB,iBAAiB,EACjB,mBAAmB,EACnB,oBAAoB,GACpB,MAAM,mDAAmD,CAAC;AAC3D,OAAO,EAAE,yBAAyB,EAAE,MAAM,mEAAmE,CAAC;AAC9G,OAAO,EAAE,aAAa,EAAE,MAAM,uCAAuC,CAAC;AACtE,OAAO,EAAE,OAAO,EAAE,MAAM,4BAA4B,CAAC;AAErD,OAAO,EAAE,MAAM,EAAE,MAAM,mCAAmC,CAAC;AAC3D,OAAO,EACN,cAAc,GAKd,MAAM,YAAY,CAAC;AAEpB,MAAM,UAAU,eAAe,CAAC,KAAa;IAC5C,OAAO,OAAO,CAAC,cAAc,EAAE,KAAK,CAAC,CAAC;AACvC,CAAC;AAaD,MAAM,aAAa,GAAG;IACrB,SAAS,EAAE,MAAM;IACjB,MAAM,EAAE,aAAa;CACgC,CAAC;AAEvD;;;;;;;;GAQG;AACH,MAAM,UAAU,WAAW,CAC1B,QAAuB,EACvB,EAAE,KAAK,EAAE,eAAe,EAAE,WAAW,EAAE,MAAM,GAAG,IAAI,MAAM,EAAE,KAAyB,EAAE;IAEvF,IAAI,MAAa,CAAC;IAElB,QAAQ,QAAQ,EAAE,CAAC;QAClB,KAAK,WAAW;YACf,MAAM,GAAG,IAAI,WAAW,CAAC;gBACxB,KAAK,EAAE,KAAK,IAAI,aAAa,CAAC,SAAS;gBACvC,eAAe;gBACf,WAAW;gBACX,MAAM;aACN,CAAC,CAAC;YACH,MAAM;QACP,KAAK,QAAQ;YACZ,MAAM,GAAG,IAAI,UAAU,CAAC;gBACvB,KAAK,EAAE,KAAK,IAAI,aAAa,CAAC,MAAM;gBACpC,eAAe;gBACf,WAAW;gBACX,MAAM;aACN,CAAC,CAAC;YACH,MAAM;QACP,OAAO,CAAC,CAAC,CAAC;YACT,+DAA+D;YAC/D,MAAM,eAAe,GAAU,QAAQ,CAAC;YACxC,MAAM,IAAI,uBAAuB,CAChC,IAAI,MAAM,CAAC,eAAe,CAAC,gDAAgD,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CACtG,CAAC;QACH,CAAC;IACF,CAAC;IAED,OAAO,IAAI,aAAa,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,CAAC,CAAC;AACrD,CAAC;AAQD,8FAA8F;AAC9F,4FAA4F;AAC5F,MAAM,iBAAiB,GAAG;IACzB,OAAO,EAAE;QACR,OAAO,EAAE,EAAE,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,aAAa,EAAE,eAAe,EAAE,MAAM,EAAE;QAC9E,QAAQ,EAAE,EAAE,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,cAAc,EAAE,eAAe,EAAE,OAAO,EAAE;QACjF,QAAQ,EAAE,EAAE,QAAQ,EAAE,WAAW,EAAE,KAAK,EAAE,MAAM,EAAE,eAAe,EAAE,MAAM,EAAE;KAC3E;IACD,SAAS,EAAE;QACV,OAAO,EAAE,EAAE,QAAQ,EAAE,WAAW,EAAE,KAAK,EAAE,MAAM,EAAE,eAAe,EAAE,MAAM,EAAE;QAC1E,QAAQ,EAAE,EAAE,QAAQ,EAAE,WAAW,EAAE,KAAK,EAAE,QAAQ,EAAE,eAAe,EAAE,OAAO,EAAE;QAC9E,QAAQ,EAAE,EAAE,QAAQ,EAAE,WAAW,EAAE,KAAK,EAAE,MAAM,EAAE,eAAe,EAAE,MAAM,EAAE;KAC3E;IACD,MAAM,EAAE;QACP,OAAO,EAAE,EAAE,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,aAAa,EAAE,eAAe,EAAE,MAAM,EAAE;QAC9E,QAAQ,EAAE,EAAE,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,cAAc,EAAE,eAAe,EAAE,OAAO,EAAE;QACjF,QAAQ,EAAE,EAAE,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,aAAa,EAAE,eAAe,EAAE,MAAM,EAAE;KAC/E;CAID,CAAC;AAEF;;;;GAIG;AACH,MAAM,UAAU,oBAAoB,CACnC,QAAmC;IAEnC,IAAI,QAAQ,KAAK,SAAS,IAAI,CAAC,eAAe,CAAC,QAAQ,CAAC,EAAE,CAAC;QAC1D,MAAM,IAAI,uBAAuB,CAChC,IAAI,MAAM,CAAC,QAAQ,CAAC,gDAAgD,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAC/F,CAAC;IACH,CAAC;IAED,OAAO,iBAAiB,CAAC,QAAQ,IAAI,SAAS,CAAC,CAAC;AACjD,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,kBAAkB,CAAC,EAClC,QAAQ,EACR,WAAW,EACX,aAAa,GAAG,EAAE,EAClB,MAAM,GAAG,IAAI,MAAM,EAAE,MACS,EAAE;IAChC,MAAM,KAAK,GAAG,oBAAoB,CAAC,QAAQ,CAAC,CAAC;IAC7C,MAAM,QAAQ,GAAG,CAAC,EAAE,QAAQ,EAAE,KAAK,EAAE,eAAe,EAAQ,EAAE,YAAoB,EAAE,EAAE,CACrF,IAAI,eAAe,CAAC;QACnB,KAAK,EAAE,WAAW,CAAC,QAAQ,EAAE,EAAE,KAAK,EAAE,eAAe,EAAE,WAAW,EAAE,MAAM,EAAE,CAAC;QAC7E,YAAY;KACZ,CAAC,CAAC;IAEJ,OAAO,IAAI,iBAAiB,CAAC;QAC5B,YAAY,EAAE,QAAQ,CAAC,KAAK,CAAC,OAAO,EAAE,aAAa,CAAC,OAAO,IAAI,mBAAmB,CAAC;QACnF,aAAa,EAAE,QAAQ,CAAC,KAAK,CAAC,QAAQ,EAAE,aAAa,CAAC,QAAQ,IAAI,oBAAoB,CAAC;QACvF,aAAa,EAAE,QAAQ,CAAC,KAAK,CAAC,QAAQ,EAAE,aAAa,CAAC,QAAQ,IAAI,oBAAoB,CAAC;QACvF,yBAAyB,EAAE,IAAI,yBAAyB,EAAE;QAC1D,MAAM;KACN,CAAC,CAAC;AACJ,CAAC","sourcesContent":["import type { Agent } from '#src/agent/domain/agent';\nimport { InvalidAgentConfigError } from '#src/agent/domain/errors';\nimport { ClaudeAgent } from '#src/engines/claude/infrastructure/model/claudeAgent';\nimport { CodexAgent } from '#src/engines/codex/infrastructure/model/codexAgent';\nimport { InstructedAgent } from '#src/orchestration/domain/model/instructedAgent';\nimport {\n\texecutorInstructions,\n\tOrchestratorAgent,\n\tplannerInstructions,\n\treviewerInstructions,\n} from '#src/orchestration/domain/model/orchestratorAgent';\nimport { ReviewerDecisionValidator } from '#src/orchestration/infrastructure/model/reviewerDecisionValidator';\nimport { RetryingAgent } from '#src/retry/domain/model/retryingAgent';\nimport { isOneOf } from '#src/shared/domain/isOneOf';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport { Logger } from '#src/shared/infrastructure/logger';\nimport {\n\tagentProviders,\n\ttype AgentModel,\n\ttype AgentProvider,\n\ttype CreateOrchestratorOptions,\n\ttype ReasoningEffort,\n} from './types.ts';\n\nexport function isAgentProvider(value: string): value is AgentProvider {\n\treturn isOneOf(agentProviders, value);\n}\n\nexport interface CreateAgentOptions {\n\t/** Defaults to `'opus'` on Anthropic and `'gpt-5.6-sol'` on OpenAI. */\n\tmodel?: AgentModel;\n\t/** Defaults to `'high'`. */\n\treasoningEffort?: ReasoningEffort;\n\t/** Maps to the provider's permission-bypass mode. Defaults to `false`. */\n\tautoApprove?: boolean;\n\t/** Defaults to warnings on stderr. */\n\tlogger?: ILogger;\n}\n\nconst defaultModels = {\n\tanthropic: 'opus',\n\topenai: 'gpt-5.6-sol',\n} as const satisfies Record<AgentProvider, AgentModel>;\n\n/**\n * Builds the agent for a provider on its Agent SDK — Claude Code for Anthropic, Codex for\n * OpenAI — already wrapped in the harness's retry policy.\n *\n * @throws {InvalidAgentConfigError} for an unknown provider, or a model or reasoning effort\n * the provider does not support.\n * @throws {UnrecoverableError} when Codex cannot be set up, for example because its CLI binary\n * is missing.\n */\nexport function createAgent(\n\tprovider: AgentProvider,\n\t{ model, reasoningEffort, autoApprove, logger = new Logger() }: CreateAgentOptions = {},\n): Agent {\n\tlet engine: Agent;\n\n\tswitch (provider) {\n\t\tcase 'anthropic':\n\t\t\tengine = new ClaudeAgent({\n\t\t\t\tmodel: model ?? defaultModels.anthropic,\n\t\t\t\treasoningEffort,\n\t\t\t\tautoApprove,\n\t\t\t\tlogger,\n\t\t\t});\n\t\t\tbreak;\n\t\tcase 'openai':\n\t\t\tengine = new CodexAgent({\n\t\t\t\tmodel: model ?? defaultModels.openai,\n\t\t\t\treasoningEffort,\n\t\t\t\tautoApprove,\n\t\t\t\tlogger,\n\t\t\t});\n\t\t\tbreak;\n\t\tdefault: {\n\t\t\t// Reachable from untyped input that skipped `isAgentProvider`.\n\t\t\tconst unknownProvider: never = provider;\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`\"${String(unknownProvider)}\" is not an agent provider; expected one of: ${agentProviders.join(', ')}`,\n\t\t\t);\n\t\t}\n\t}\n\n\treturn new RetryingAgent({ inner: engine, logger });\n}\n\nexport interface Role {\n\tprovider: AgentProvider;\n\tmodel: AgentModel;\n\treasoningEffort: ReasoningEffort;\n}\n\n// A stronger model plans and reviews; a cheaper one executes the plan. Every model and effort\n// here is one both the Agent SDKs and the model APIs accept, so either path can run a role.\nconst orchestratorRoles = {\n\tdefault: {\n\t\tplanner: { provider: 'openai', model: 'gpt-5.6-sol', reasoningEffort: 'high' },\n\t\texecutor: { provider: 'openai', model: 'gpt-5.6-luna', reasoningEffort: 'xhigh' },\n\t\treviewer: { provider: 'anthropic', model: 'opus', reasoningEffort: 'high' },\n\t},\n\tanthropic: {\n\t\tplanner: { provider: 'anthropic', model: 'opus', reasoningEffort: 'high' },\n\t\texecutor: { provider: 'anthropic', model: 'sonnet', reasoningEffort: 'xhigh' },\n\t\treviewer: { provider: 'anthropic', model: 'opus', reasoningEffort: 'high' },\n\t},\n\topenai: {\n\t\tplanner: { provider: 'openai', model: 'gpt-5.6-sol', reasoningEffort: 'high' },\n\t\texecutor: { provider: 'openai', model: 'gpt-5.6-luna', reasoningEffort: 'xhigh' },\n\t\treviewer: { provider: 'openai', model: 'gpt-5.6-sol', reasoningEffort: 'high' },\n\t},\n} as const satisfies Record<\n\tAgentProvider | 'default',\n\tRecord<'planner' | 'executor' | 'reviewer', Role>\n>;\n\n/**\n * The roles for `provider`, or the default mix without one.\n *\n * @throws {InvalidAgentConfigError} for an unknown provider.\n */\nexport function orchestratorRolesFor(\n\tprovider: AgentProvider | undefined,\n): Record<'planner' | 'executor' | 'reviewer', Role> {\n\tif (provider !== undefined && !isAgentProvider(provider)) {\n\t\tthrow new InvalidAgentConfigError(\n\t\t\t`\"${String(provider)}\" is not an agent provider; expected one of: ${agentProviders.join(', ')}`,\n\t\t);\n\t}\n\n\treturn orchestratorRoles[provider ?? 'default'];\n}\n\n/**\n * Builds the planner → executor → reviewer workflow, each role an agent from `createAgent`.\n *\n * @throws {InvalidAgentConfigError} for an unknown provider.\n * @throws {UnrecoverableError} when an OpenAI role cannot be set up, for example because the\n * Codex CLI binary is missing.\n */\nexport function createOrchestrator({\n\tprovider,\n\tautoApprove,\n\tsystemPrompts = {},\n\tlogger = new Logger(),\n}: CreateOrchestratorOptions = {}): Agent {\n\tconst roles = orchestratorRolesFor(provider);\n\tconst agentFor = ({ provider, model, reasoningEffort }: Role, instructions: string) =>\n\t\tnew InstructedAgent({\n\t\t\tinner: createAgent(provider, { model, reasoningEffort, autoApprove, logger }),\n\t\t\tinstructions,\n\t\t});\n\n\treturn new OrchestratorAgent({\n\t\tplannerAgent: agentFor(roles.planner, systemPrompts.planner ?? plannerInstructions),\n\t\texecutorAgent: agentFor(roles.executor, systemPrompts.executor ?? executorInstructions),\n\t\treviewerAgent: agentFor(roles.reviewer, systemPrompts.reviewer ?? reviewerInstructions),\n\t\treviewerDecisionValidator: new ReviewerDecisionValidator(),\n\t\tlogger,\n\t});\n}\n"]}
+{"version":3,"file":"agentFactory.js","sourceRoot":"","sources":["../../../src/factory/infrastructure/agentFactory.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,uBAAuB,EAAE,MAAM,2BAA2B,CAAC;AACpE,OAAO,EAAE,WAAW,EAAE,MAAM,sDAAsD,CAAC;AACnF,OAAO,EAAE,UAAU,EAAE,MAAM,oDAAoD,CAAC;AAChF,OAAO,EAAE,eAAe,EAAE,MAAM,iDAAiD,CAAC;AAClF,OAAO,EACN,oBAAoB,EACpB,iBAAiB,EACjB,mBAAmB,EACnB,oBAAoB,GACpB,MAAM,mDAAmD,CAAC;AAC3D,OAAO,EAAE,yBAAyB,EAAE,MAAM,mEAAmE,CAAC;AAC9G,OAAO,EAAE,aAAa,EAAE,MAAM,uCAAuC,CAAC;AACtE,OAAO,EAAE,OAAO,EAAE,MAAM,4BAA4B,CAAC;AAErD,OAAO,EAAE,MAAM,EAAE,MAAM,mCAAmC,CAAC;AAC3D,OAAO,EACN,cAAc,GAKd,MAAM,YAAY,CAAC;AAEpB,MAAM,UAAU,eAAe,CAAC,KAAa;IAC5C,OAAO,OAAO,CAAC,cAAc,EAAE,KAAK,CAAC,CAAC;AACvC,CAAC;AAaD,MAAM,aAAa,GAAG;IACrB,SAAS,EAAE,MAAM;IACjB,MAAM,EAAE,aAAa;CACgC,CAAC;AAEvD;;;;;;;;GAQG;AACH,MAAM,UAAU,WAAW,CAC1B,QAAuB,EACvB,EAAE,KAAK,EAAE,eAAe,EAAE,WAAW,EAAE,MAAM,GAAG,IAAI,MAAM,EAAE,KAAyB,EAAE;IAEvF,IAAI,MAAa,CAAC;IAElB,QAAQ,QAAQ,EAAE,CAAC;QAClB,KAAK,WAAW;YACf,MAAM,GAAG,IAAI,WAAW,CAAC;gBACxB,KAAK,EAAE,KAAK,IAAI,aAAa,CAAC,SAAS;gBACvC,eAAe;gBACf,WAAW;gBACX,MAAM;aACN,CAAC,CAAC;YACH,MAAM;QACP,KAAK,QAAQ;YACZ,MAAM,GAAG,IAAI,UAAU,CAAC;gBACvB,KAAK,EAAE,KAAK,IAAI,aAAa,CAAC,MAAM;gBACpC,eAAe;gBACf,WAAW;gBACX,MAAM;aACN,CAAC,CAAC;YACH,MAAM;QACP,OAAO,CAAC,CAAC,CAAC;YACT,+DAA+D;YAC/D,MAAM,eAAe,GAAU,QAAQ,CAAC;YACxC,MAAM,IAAI,uBAAuB,CAChC,IAAI,MAAM,CAAC,eAAe,CAAC,gDAAgD,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CACtG,CAAC;QACH,CAAC;IACF,CAAC;IAED,OAAO,IAAI,aAAa,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,CAAC,CAAC;AACrD,CAAC;AAQD,8FAA8F;AAC9F,4FAA4F;AAC5F,MAAM,iBAAiB,GAAG;IACzB,OAAO,EAAE;QACR,OAAO,EAAE,EAAE,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,aAAa,EAAE,eAAe,EAAE,MAAM,EAAE;QAC9E,QAAQ,EAAE,EAAE,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,cAAc,EAAE,eAAe,EAAE,OAAO,EAAE;QACjF,QAAQ,EAAE,EAAE,QAAQ,EAAE,WAAW,EAAE,KAAK,EAAE,MAAM,EAAE,eAAe,EAAE,MAAM,EAAE;KAC3E;IACD,SAAS,EAAE;QACV,OAAO,EAAE,EAAE,QAAQ,EAAE,WAAW,EAAE,KAAK,EAAE,MAAM,EAAE,eAAe,EAAE,MAAM,EAAE;QAC1E,QAAQ,EAAE,EAAE,QAAQ,EAAE,WAAW,EAAE,KAAK,EAAE,QAAQ,EAAE,eAAe,EAAE,OAAO,EAAE;QAC9E,QAAQ,EAAE,EAAE,QAAQ,EAAE,WAAW,EAAE,KAAK,EAAE,MAAM,EAAE,eAAe,EAAE,MAAM,EAAE;KAC3E;IACD,MAAM,EAAE;QACP,OAAO,EAAE,EAAE,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,aAAa,EAAE,eAAe,EAAE,MAAM,EAAE;QAC9E,QAAQ,EAAE,EAAE,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,cAAc,EAAE,eAAe,EAAE,OAAO,EAAE;QACjF,QAAQ,EAAE,EAAE,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,aAAa,EAAE,eAAe,EAAE,MAAM,EAAE;KAC/E;CAID,CAAC;AAEF;;;;GAIG;AACH,MAAM,UAAU,oBAAoB,CACnC,QAAmC;IAEnC,IAAI,QAAQ,KAAK,SAAS,IAAI,CAAC,eAAe,CAAC,QAAQ,CAAC,EAAE,CAAC;QAC1D,MAAM,IAAI,uBAAuB,CAChC,IAAI,MAAM,CAAC,QAAQ,CAAC,gDAAgD,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAC/F,CAAC;IACH,CAAC;IAED,OAAO,iBAAiB,CAAC,QAAQ,IAAI,SAAS,CAAC,CAAC;AACjD,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,kBAAkB,CAAC,EAClC,QAAQ,EACR,WAAW,EACX,aAAa,GAAG,EAAE,EAClB,MAAM,GAAG,IAAI,MAAM,EAAE,MACS,EAAE;IAChC,MAAM,KAAK,GAAG,oBAAoB,CAAC,QAAQ,CAAC,CAAC;IAC7C,MAAM,QAAQ,GAAG,CAAC,EAAE,QAAQ,EAAE,KAAK,EAAE,eAAe,EAAQ,EAAE,YAAoB,EAAE,EAAE,CACrF,IAAI,eAAe,CAAC;QACnB,KAAK,EAAE,WAAW,CAAC,QAAQ,EAAE,EAAE,KAAK,EAAE,eAAe,EAAE,WAAW,EAAE,MAAM,EAAE,CAAC;QAC7E,YAAY;KACZ,CAAC,CAAC;IAEJ,OAAO,IAAI,iBAAiB,CAAC;QAC5B,YAAY,EAAE,QAAQ,CAAC,KAAK,CAAC,OAAO,EAAE,aAAa,CAAC,OAAO,IAAI,mBAAmB,CAAC;QACnF,aAAa,EAAE,QAAQ,CAAC,KAAK,CAAC,QAAQ,EAAE,aAAa,CAAC,QAAQ,IAAI,oBAAoB,CAAC;QACvF,aAAa,EAAE,QAAQ,CAAC,KAAK,CAAC,QAAQ,EAAE,aAAa,CAAC,QAAQ,IAAI,oBAAoB,CAAC;QACvF,yBAAyB,EAAE,IAAI,yBAAyB,EAAE;QAC1D,MAAM;KACN,CAAC,CAAC;AACJ,CAAC","sourcesContent":["import type { Agent } from '#src/agent/domain/agent';\nimport { InvalidAgentConfigError } from '#src/shared/domain/errors';\nimport { ClaudeAgent } from '#src/engines/claude/infrastructure/model/claudeAgent';\nimport { CodexAgent } from '#src/engines/codex/infrastructure/model/codexAgent';\nimport { InstructedAgent } from '#src/orchestration/domain/model/instructedAgent';\nimport {\n\texecutorInstructions,\n\tOrchestratorAgent,\n\tplannerInstructions,\n\treviewerInstructions,\n} from '#src/orchestration/domain/model/orchestratorAgent';\nimport { ReviewerDecisionValidator } from '#src/orchestration/infrastructure/model/reviewerDecisionValidator';\nimport { RetryingAgent } from '#src/retry/domain/model/retryingAgent';\nimport { isOneOf } from '#src/shared/domain/isOneOf';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport { Logger } from '#src/shared/infrastructure/logger';\nimport {\n\tagentProviders,\n\ttype AgentModel,\n\ttype AgentProvider,\n\ttype CreateOrchestratorOptions,\n\ttype ReasoningEffort,\n} from './types.ts';\n\nexport function isAgentProvider(value: string): value is AgentProvider {\n\treturn isOneOf(agentProviders, value);\n}\n\nexport interface CreateAgentOptions {\n\t/** Defaults to `'opus'` on Anthropic and `'gpt-5.6-sol'` on OpenAI. */\n\tmodel?: AgentModel;\n\t/** Defaults to `'high'`. */\n\treasoningEffort?: ReasoningEffort;\n\t/** Maps to the provider's permission-bypass mode. Defaults to `false`. */\n\tautoApprove?: boolean;\n\t/** Defaults to warnings on stderr. */\n\tlogger?: ILogger;\n}\n\nconst defaultModels = {\n\tanthropic: 'opus',\n\topenai: 'gpt-5.6-sol',\n} as const satisfies Record<AgentProvider, AgentModel>;\n\n/**\n * Builds the agent for a provider on its Agent SDK — Claude Code for Anthropic, Codex for\n * OpenAI — already wrapped in the harness's retry policy.\n *\n * @throws {InvalidAgentConfigError} for an unknown provider, or a model or reasoning effort\n * the provider does not support.\n * @throws {UnrecoverableError} when Codex cannot be set up, for example because its CLI binary\n * is missing.\n */\nexport function createAgent(\n\tprovider: AgentProvider,\n\t{ model, reasoningEffort, autoApprove, logger = new Logger() }: CreateAgentOptions = {},\n): Agent {\n\tlet engine: Agent;\n\n\tswitch (provider) {\n\t\tcase 'anthropic':\n\t\t\tengine = new ClaudeAgent({\n\t\t\t\tmodel: model ?? defaultModels.anthropic,\n\t\t\t\treasoningEffort,\n\t\t\t\tautoApprove,\n\t\t\t\tlogger,\n\t\t\t});\n\t\t\tbreak;\n\t\tcase 'openai':\n\t\t\tengine = new CodexAgent({\n\t\t\t\tmodel: model ?? defaultModels.openai,\n\t\t\t\treasoningEffort,\n\t\t\t\tautoApprove,\n\t\t\t\tlogger,\n\t\t\t});\n\t\t\tbreak;\n\t\tdefault: {\n\t\t\t// Reachable from untyped input that skipped `isAgentProvider`.\n\t\t\tconst unknownProvider: never = provider;\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`\"${String(unknownProvider)}\" is not an agent provider; expected one of: ${agentProviders.join(', ')}`,\n\t\t\t);\n\t\t}\n\t}\n\n\treturn new RetryingAgent({ inner: engine, logger });\n}\n\nexport interface Role {\n\tprovider: AgentProvider;\n\tmodel: AgentModel;\n\treasoningEffort: ReasoningEffort;\n}\n\n// A stronger model plans and reviews; a cheaper one executes the plan. Every model and effort\n// here is one both the Agent SDKs and the model APIs accept, so either path can run a role.\nconst orchestratorRoles = {\n\tdefault: {\n\t\tplanner: { provider: 'openai', model: 'gpt-5.6-sol', reasoningEffort: 'high' },\n\t\texecutor: { provider: 'openai', model: 'gpt-5.6-luna', reasoningEffort: 'xhigh' },\n\t\treviewer: { provider: 'anthropic', model: 'opus', reasoningEffort: 'high' },\n\t},\n\tanthropic: {\n\t\tplanner: { provider: 'anthropic', model: 'opus', reasoningEffort: 'high' },\n\t\texecutor: { provider: 'anthropic', model: 'sonnet', reasoningEffort: 'xhigh' },\n\t\treviewer: { provider: 'anthropic', model: 'opus', reasoningEffort: 'high' },\n\t},\n\topenai: {\n\t\tplanner: { provider: 'openai', model: 'gpt-5.6-sol', reasoningEffort: 'high' },\n\t\texecutor: { provider: 'openai', model: 'gpt-5.6-luna', reasoningEffort: 'xhigh' },\n\t\treviewer: { provider: 'openai', model: 'gpt-5.6-sol', reasoningEffort: 'high' },\n\t},\n} as const satisfies Record<\n\tAgentProvider | 'default',\n\tRecord<'planner' | 'executor' | 'reviewer', Role>\n>;\n\n/**\n * The roles for `provider`, or the default mix without one.\n *\n * @throws {InvalidAgentConfigError} for an unknown provider.\n */\nexport function orchestratorRolesFor(\n\tprovider: AgentProvider | undefined,\n): Record<'planner' | 'executor' | 'reviewer', Role> {\n\tif (provider !== undefined && !isAgentProvider(provider)) {\n\t\tthrow new InvalidAgentConfigError(\n\t\t\t`\"${String(provider)}\" is not an agent provider; expected one of: ${agentProviders.join(', ')}`,\n\t\t);\n\t}\n\n\treturn orchestratorRoles[provider ?? 'default'];\n}\n\n/**\n * Builds the planner → executor → reviewer workflow, each role an agent from `createAgent`.\n *\n * @throws {InvalidAgentConfigError} for an unknown provider.\n * @throws {UnrecoverableError} when an OpenAI role cannot be set up, for example because the\n * Codex CLI binary is missing.\n */\nexport function createOrchestrator({\n\tprovider,\n\tautoApprove,\n\tsystemPrompts = {},\n\tlogger = new Logger(),\n}: CreateOrchestratorOptions = {}): Agent {\n\tconst roles = orchestratorRolesFor(provider);\n\tconst agentFor = ({ provider, model, reasoningEffort }: Role, instructions: string) =>\n\t\tnew InstructedAgent({\n\t\t\tinner: createAgent(provider, { model, reasoningEffort, autoApprove, logger }),\n\t\t\tinstructions,\n\t\t});\n\n\treturn new OrchestratorAgent({\n\t\tplannerAgent: agentFor(roles.planner, systemPrompts.planner ?? plannerInstructions),\n\t\texecutorAgent: agentFor(roles.executor, systemPrompts.executor ?? executorInstructions),\n\t\treviewerAgent: agentFor(roles.reviewer, systemPrompts.reviewer ?? reviewerInstructions),\n\t\treviewerDecisionValidator: new ReviewerDecisionValidator(),\n\t\tlogger,\n\t});\n}\n"]}

package/dist/factory/infrastructure/agentLLMFactory.d.ts

@@ -14,27 +14,34 @@ export interface CreateLLMAgentOptions {
     systemPrompt: string;
     /** The tools the model may call. Defaults to none. */
     tools?: Tool[];
+    /**
+     * Runs every tool call without asking, `destructive` ones included, for CI where no one can
+     * answer. Without it, a `destructive` call goes to the run's `approve`, or is denied.
+     */
+    autoApprove?: boolean;
     /** Defaults to warnings on stderr. */
     logger?: ILogger;
 }
 /**
  * Builds an agent whose conversation MiKode owns, on a provider's model API, already wrapped in
- * the harness's retry policy. Internal while #23 reaches parity with the Agent SDK engines: it
- * is not exported from `src/index.ts`. Its tools only read, so it has no `autoApprove`.
+ * the harness's retry policy. It needs the provider's API key and bills per token, where
+ * `createAgent` uses the Agent SDK's login. Each tool judges the risk of each call; a
+ * `destructive` one runs only when the run's `approve` allows it, or with `autoApprove`.
  *
- * @throws {InvalidAgentConfigError} for an unknown provider, or a model or reasoning effort its
- * client does not support.
+ * @throws {InvalidAgentConfigError} for an unknown provider, a model or reasoning effort its
+ * client does not support, or two tools that share a name.
  * @throws {UnrecoverableError} when the client cannot be set up, for example because
  * `OPENAI_API_KEY` is missing. The Anthropic SDK resolves credentials lazily, so a missing
  * `ANTHROPIC_API_KEY` fails the first run instead, with an `UnrecoverableError`.
  */
-export declare function createLLMAgent(provider: AgentProvider, { model, reasoningEffort, systemPrompt, tools, logger }: CreateLLMAgentOptions): Agent;
+export declare function createLLMAgent(provider: AgentProvider, { model, reasoningEffort, systemPrompt, tools, autoApprove, logger, }: CreateLLMAgentOptions): Agent;
 /**
  * Builds the planner → executor → reviewer workflow with the planner and reviewer on the model
  * APIs, from `createLLMAgent`, each holding its role's instructions, or `systemPrompts`, as its
  * system prompt, with the read-only repository tools for the current working directory. The
- * executor stays on its Agent SDK, from `createAgent`, until an agent of ours can change files; `autoApprove` reaches
- * only it. Asynchronous because finding the program that reads the repository is.
+ * executor stays on its Agent SDK, from `createAgent`, until an agent of ours can change files;
+ * `autoApprove` reaches only it. Asynchronous because finding the program that reads the
+ * repository is.
  *
  * @throws {InvalidAgentConfigError} for an unknown provider.
  * @throws {UnrecoverableError} when a role cannot be set up, for example because

package/dist/factory/infrastructure/agentLLMFactory.js

@@ -1,4 +1,4 @@
-import { InvalidAgentConfigError } from '#src/agent/domain/errors';
+import { InvalidAgentConfigError } from '#src/shared/domain/errors';
 import { LLMAgent } from '#src/engines/domain/model/llmAgent';
 import { ClaudeLLMClient } from '#src/llm/infrastructure/claudeLLMClient';
 import { OpenAILLMClient } from '#src/llm/infrastructure/openAILLMClient';
@@ -17,16 +17,17 @@ const defaultLLMAgentModels = {
 };
 /**
  * Builds an agent whose conversation MiKode owns, on a provider's model API, already wrapped in
- * the harness's retry policy. Internal while #23 reaches parity with the Agent SDK engines: it
- * is not exported from `src/index.ts`. Its tools only read, so it has no `autoApprove`.
+ * the harness's retry policy. It needs the provider's API key and bills per token, where
+ * `createAgent` uses the Agent SDK's login. Each tool judges the risk of each call; a
+ * `destructive` one runs only when the run's `approve` allows it, or with `autoApprove`.
  *
- * @throws {InvalidAgentConfigError} for an unknown provider, or a model or reasoning effort its
- * client does not support.
+ * @throws {InvalidAgentConfigError} for an unknown provider, a model or reasoning effort its
+ * client does not support, or two tools that share a name.
  * @throws {UnrecoverableError} when the client cannot be set up, for example because
  * `OPENAI_API_KEY` is missing. The Anthropic SDK resolves credentials lazily, so a missing
  * `ANTHROPIC_API_KEY` fails the first run instead, with an `UnrecoverableError`.
  */
-export function createLLMAgent(provider, { model, reasoningEffort, systemPrompt, tools, logger = new Logger() }) {
+export function createLLMAgent(provider, { model, reasoningEffort, systemPrompt, tools, autoApprove, logger = new Logger(), }) {
     let llmClient;
     switch (provider) {
         case 'anthropic':
@@ -53,7 +54,7 @@ export function createLLMAgent(provider, { model, reasoningEffort, systemPrompt,
     }
     // LLMAgent records nothing from a failed call, so the original prompt is the whole story.
     return new RetryingAgent({
-        inner: new LLMAgent({ llmClient, tools }),
+        inner: new LLMAgent({ llmClient, tools, autoApprove }),
         logger,
         noteFailures: false,
     });
@@ -65,8 +66,9 @@ const repositoryGuidance = 'Paths are relative to the repository root. Start wit
  * Builds the planner → executor → reviewer workflow with the planner and reviewer on the model
  * APIs, from `createLLMAgent`, each holding its role's instructions, or `systemPrompts`, as its
  * system prompt, with the read-only repository tools for the current working directory. The
- * executor stays on its Agent SDK, from `createAgent`, until an agent of ours can change files; `autoApprove` reaches
- * only it. Asynchronous because finding the program that reads the repository is.
+ * executor stays on its Agent SDK, from `createAgent`, until an agent of ours can change files;
+ * `autoApprove` reaches only it. Asynchronous because finding the program that reads the
+ * repository is.
  *
  * @throws {InvalidAgentConfigError} for an unknown provider.
  * @throws {UnrecoverableError} when a role cannot be set up, for example because

package/dist/factory/infrastructure/agentLLMFactory.js.map

@@ -1 +1 @@
-{"version":3,"file":"agentLLMFactory.js","sourceRoot":"","sources":["../../../src/factory/infrastructure/agentLLMFactory.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,uBAAuB,EAAE,MAAM,0BAA0B,CAAC;AACnE,OAAO,EAAE,QAAQ,EAAE,MAAM,oCAAoC,CAAC;AAE9D,OAAO,EAAE,eAAe,EAAE,MAAM,yCAAyC,CAAC;AAC1E,OAAO,EAAE,eAAe,EAAE,MAAM,yCAAyC,CAAC;AAC1E,OAAO,EAAE,eAAe,EAAE,MAAM,iDAAiD,CAAC;AAClF,OAAO,EACN,oBAAoB,EACpB,iBAAiB,EACjB,mBAAmB,EACnB,oBAAoB,GACpB,MAAM,mDAAmD,CAAC;AAC3D,OAAO,EAAE,yBAAyB,EAAE,MAAM,mEAAmE,CAAC;AAC9G,OAAO,EAAE,aAAa,EAAE,MAAM,uCAAuC,CAAC;AAEtE,OAAO,EAAE,MAAM,EAAE,MAAM,mCAAmC,CAAC;AAE3D,OAAO,EAAE,eAAe,EAAE,MAAM,2CAA2C,CAAC;AAC5E,OAAO,EAAE,oBAAoB,EAAE,MAAM,0CAA0C,CAAC;AAChF,OAAO,EAAE,WAAW,EAAE,oBAAoB,EAAa,MAAM,mBAAmB,CAAC;AACjF,OAAO,EACN,cAAc,GAKd,MAAM,YAAY,CAAC;AAkBpB,MAAM,qBAAqB,GAAG;IAC7B,SAAS,EAAE,QAAQ;IACnB,MAAM,EAAE,cAAc;CAC+B,CAAC;AAEvD;;;;;;;;;;GAUG;AACH,MAAM,UAAU,cAAc,CAC7B,QAAuB,EACvB,EAAE,KAAK,EAAE,eAAe,EAAE,YAAY,EAAE,KAAK,EAAE,MAAM,GAAG,IAAI,MAAM,EAAE,EAAyB;IAE7F,IAAI,SAAoB,CAAC;IAEzB,QAAQ,QAAQ,EAAE,CAAC;QAClB,KAAK,WAAW;YACf,SAAS,GAAG,IAAI,eAAe,CAAC;gBAC/B,KAAK,EAAE,KAAK,IAAI,qBAAqB,CAAC,SAAS;gBAC/C,eAAe;gBACf,YAAY;gBACZ,MAAM;aACN,CAAC,CAAC;YACH,MAAM;QACP,KAAK,QAAQ;YACZ,SAAS,GAAG,IAAI,eAAe,CAAC;gBAC/B,KAAK,EAAE,KAAK,IAAI,qBAAqB,CAAC,MAAM;gBAC5C,eAAe;gBACf,YAAY;gBACZ,MAAM;aACN,CAAC,CAAC;YACH,MAAM;QACP,OAAO,CAAC,CAAC,CAAC;YACT,gCAAgC;YAChC,MAAM,eAAe,GAAU,QAAQ,CAAC;YACxC,MAAM,IAAI,uBAAuB,CAChC,IAAI,MAAM,CAAC,eAAe,CAAC,gDAAgD,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CACtG,CAAC;QACH,CAAC;IACF,CAAC;IAED,0FAA0F;IAC1F,OAAO,IAAI,aAAa,CAAC;QACxB,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,EAAE,KAAK,EAAE,CAAC;QACzC,MAAM;QACN,YAAY,EAAE,KAAK;KACnB,CAAC,CAAC;AACJ,CAAC;AAED,gGAAgG;AAChG,6DAA6D;AAC7D,MAAM,kBAAkB,GACvB,8KAA8K,CAAC;AAEhL;;;;;;;;;;;GAWG;AACH,MAAM,CAAC,KAAK,UAAU,qBAAqB,CAAC,EAC3C,QAAQ,EACR,WAAW,EACX,aAAa,GAAG,EAAE,EAClB,MAAM,GAAG,IAAI,MAAM,EAAE,MACS,EAAE;IAChC,MAAM,KAAK,GAAG,oBAAoB,CAAC,QAAQ,CAAC,CAAC;IAC7C,MAAM,KAAK,GAAG,oBAAoB,CAAC,MAAM,eAAe,CAAC,EAAE,IAAI,EAAE,OAAO,CAAC,GAAG,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC;IAC3F,MAAM,WAAW,GAAG,CAAC,EAAE,QAAQ,EAAE,KAAK,EAAE,eAAe,EAAQ,EAAE,YAAoB,EAAE,EAAE,CACxF,cAAc,CAAC,QAAQ,EAAE,EAAE,KAAK,EAAE,eAAe,EAAE,YAAY,EAAE,KAAK,EAAE,MAAM,EAAE,CAAC,CAAC;IACnF,uEAAuE;IACvE,MAAM,YAAY,GAAG,CAAC,YAAoB,EAAE,EAAE,CAAC,GAAG,YAAY,OAAO,kBAAkB,EAAE,CAAC;IAE1F,OAAO,IAAI,iBAAiB,CAAC;QAC5B,YAAY,EAAE,WAAW,CACxB,KAAK,CAAC,OAAO,EACb,aAAa,CAAC,OAAO,IAAI,YAAY,CAAC,mBAAmB,CAAC,CAC1D;QACD,aAAa,EAAE,IAAI,eAAe,CAAC;YAClC,KAAK,EAAE,WAAW,CAAC,KAAK,CAAC,QAAQ,CAAC,QAAQ,EAAE;gBAC3C,KAAK,EAAE,KAAK,CAAC,QAAQ,CAAC,KAAK;gBAC3B,eAAe,EAAE,KAAK,CAAC,QAAQ,CAAC,eAAe;gBAC/C,WAAW;gBACX,MAAM;aACN,CAAC;YACF,YAAY,EAAE,aAAa,CAAC,QAAQ,IAAI,oBAAoB;SAC5D,CAAC;QACF,aAAa,EAAE,WAAW,CACzB,KAAK,CAAC,QAAQ,EACd,aAAa,CAAC,QAAQ,IAAI,YAAY,CAAC,oBAAoB,CAAC,CAC5D;QACD,yBAAyB,EAAE,IAAI,yBAAyB,EAAE;QAC1D,MAAM;KACN,CAAC,CAAC;AACJ,CAAC","sourcesContent":["import type { Agent } from '#src/agent/domain/agent';\nimport { InvalidAgentConfigError } from '#src/agent/domain/errors';\nimport { LLMAgent } from '#src/engines/domain/model/llmAgent';\nimport type { LLMClient } from '#src/llm/domain/llm';\nimport { ClaudeLLMClient } from '#src/llm/infrastructure/claudeLLMClient';\nimport { OpenAILLMClient } from '#src/llm/infrastructure/openAILLMClient';\nimport { InstructedAgent } from '#src/orchestration/domain/model/instructedAgent';\nimport {\n\texecutorInstructions,\n\tOrchestratorAgent,\n\tplannerInstructions,\n\treviewerInstructions,\n} from '#src/orchestration/domain/model/orchestratorAgent';\nimport { ReviewerDecisionValidator } from '#src/orchestration/infrastructure/model/reviewerDecisionValidator';\nimport { RetryingAgent } from '#src/retry/domain/model/retryingAgent';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport { Logger } from '#src/shared/infrastructure/logger';\nimport type { Tool } from '#src/tools/domain/tool';\nimport { createWorkspace } from '#src/tools/infrastructure/createWorkspace';\nimport { createWorkspaceTools } from '#src/tools/infrastructure/workspaceTools';\nimport { createAgent, orchestratorRolesFor, type Role } from './agentFactory.ts';\nimport {\n\tagentProviders,\n\ttype AgentModel,\n\ttype AgentProvider,\n\ttype CreateOrchestratorOptions,\n\ttype ReasoningEffort,\n} from './types.ts';\n\nexport interface CreateLLMAgentOptions {\n\t/**\n\t * Defaults to `'sonnet'` on Anthropic and `'gpt-5.6-luna'` on OpenAI. The provider's client\n\t * rejects a model it does not support.\n\t */\n\tmodel?: AgentModel;\n\t/** Defaults to `'high'`. The provider's client rejects an effort it does not support. */\n\treasoningEffort?: ReasoningEffort;\n\t/** Sent as the model's instructions on every call. */\n\tsystemPrompt: string;\n\t/** The tools the model may call. Defaults to none. */\n\ttools?: Tool[];\n\t/** Defaults to warnings on stderr. */\n\tlogger?: ILogger;\n}\n\nconst defaultLLMAgentModels = {\n\tanthropic: 'sonnet',\n\topenai: 'gpt-5.6-luna',\n} as const satisfies Record<AgentProvider, AgentModel>;\n\n/**\n * Builds an agent whose conversation MiKode owns, on a provider's model API, already wrapped in\n * the harness's retry policy. Internal while #23 reaches parity with the Agent SDK engines: it\n * is not exported from `src/index.ts`. Its tools only read, so it has no `autoApprove`.\n *\n * @throws {InvalidAgentConfigError} for an unknown provider, or a model or reasoning effort its\n * client does not support.\n * @throws {UnrecoverableError} when the client cannot be set up, for example because\n * `OPENAI_API_KEY` is missing. The Anthropic SDK resolves credentials lazily, so a missing\n * `ANTHROPIC_API_KEY` fails the first run instead, with an `UnrecoverableError`.\n */\nexport function createLLMAgent(\n\tprovider: AgentProvider,\n\t{ model, reasoningEffort, systemPrompt, tools, logger = new Logger() }: CreateLLMAgentOptions,\n): Agent {\n\tlet llmClient: LLMClient;\n\n\tswitch (provider) {\n\t\tcase 'anthropic':\n\t\t\tllmClient = new ClaudeLLMClient({\n\t\t\t\tmodel: model ?? defaultLLMAgentModels.anthropic,\n\t\t\t\treasoningEffort,\n\t\t\t\tsystemPrompt,\n\t\t\t\tlogger,\n\t\t\t});\n\t\t\tbreak;\n\t\tcase 'openai':\n\t\t\tllmClient = new OpenAILLMClient({\n\t\t\t\tmodel: model ?? defaultLLMAgentModels.openai,\n\t\t\t\treasoningEffort,\n\t\t\t\tsystemPrompt,\n\t\t\t\tlogger,\n\t\t\t});\n\t\t\tbreak;\n\t\tdefault: {\n\t\t\t// Reachable from untyped input.\n\t\t\tconst unknownProvider: never = provider;\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`\"${String(unknownProvider)}\" is not an agent provider; expected one of: ${agentProviders.join(', ')}`,\n\t\t\t);\n\t\t}\n\t}\n\n\t// LLMAgent records nothing from a failed call, so the original prompt is the whole story.\n\treturn new RetryingAgent({\n\t\tinner: new LLMAgent({ llmClient, tools }),\n\t\tlogger,\n\t\tnoteFailures: false,\n\t});\n}\n\n// Sent on every call after the role's instructions, so it stays short: it only says how to find\n// one's way around the repository without reading all of it.\nconst repositoryGuidance =\n\t'Paths are relative to the repository root. Start with AGENTS.md and the architecture document it links to learn where things live, then read only the files your task needs.';\n\n/**\n * Builds the planner → executor → reviewer workflow with the planner and reviewer on the model\n * APIs, from `createLLMAgent`, each holding its role's instructions, or `systemPrompts`, as its\n * system prompt, with the read-only repository tools for the current working directory. The\n * executor stays on its Agent SDK, from `createAgent`, until an agent of ours can change files; `autoApprove` reaches\n * only it. Asynchronous because finding the program that reads the repository is.\n *\n * @throws {InvalidAgentConfigError} for an unknown provider.\n * @throws {UnrecoverableError} when a role cannot be set up, for example because\n * `OPENAI_API_KEY` or the Codex CLI binary is missing, or when no program can read the\n * repository.\n */\nexport async function createLLMOrchestrator({\n\tprovider,\n\tautoApprove,\n\tsystemPrompts = {},\n\tlogger = new Logger(),\n}: CreateOrchestratorOptions = {}): Promise<Agent> {\n\tconst roles = orchestratorRolesFor(provider);\n\tconst tools = createWorkspaceTools(await createWorkspace({ root: process.cwd(), logger }));\n\tconst llmAgentFor = ({ provider, model, reasoningEffort }: Role, systemPrompt: string) =>\n\t\tcreateLLMAgent(provider, { model, reasoningEffort, systemPrompt, tools, logger });\n\t// A caller's own prompt replaces the whole of ours, guidance included.\n\tconst withGuidance = (instructions: string) => `${instructions}\\n\\n${repositoryGuidance}`;\n\n\treturn new OrchestratorAgent({\n\t\tplannerAgent: llmAgentFor(\n\t\t\troles.planner,\n\t\t\tsystemPrompts.planner ?? withGuidance(plannerInstructions),\n\t\t),\n\t\texecutorAgent: new InstructedAgent({\n\t\t\tinner: createAgent(roles.executor.provider, {\n\t\t\t\tmodel: roles.executor.model,\n\t\t\t\treasoningEffort: roles.executor.reasoningEffort,\n\t\t\t\tautoApprove,\n\t\t\t\tlogger,\n\t\t\t}),\n\t\t\tinstructions: systemPrompts.executor ?? executorInstructions,\n\t\t}),\n\t\treviewerAgent: llmAgentFor(\n\t\t\troles.reviewer,\n\t\t\tsystemPrompts.reviewer ?? withGuidance(reviewerInstructions),\n\t\t),\n\t\treviewerDecisionValidator: new ReviewerDecisionValidator(),\n\t\tlogger,\n\t});\n}\n"]}
+{"version":3,"file":"agentLLMFactory.js","sourceRoot":"","sources":["../../../src/factory/infrastructure/agentLLMFactory.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,uBAAuB,EAAE,MAAM,2BAA2B,CAAC;AACpE,OAAO,EAAE,QAAQ,EAAE,MAAM,oCAAoC,CAAC;AAE9D,OAAO,EAAE,eAAe,EAAE,MAAM,yCAAyC,CAAC;AAC1E,OAAO,EAAE,eAAe,EAAE,MAAM,yCAAyC,CAAC;AAC1E,OAAO,EAAE,eAAe,EAAE,MAAM,iDAAiD,CAAC;AAClF,OAAO,EACN,oBAAoB,EACpB,iBAAiB,EACjB,mBAAmB,EACnB,oBAAoB,GACpB,MAAM,mDAAmD,CAAC;AAC3D,OAAO,EAAE,yBAAyB,EAAE,MAAM,mEAAmE,CAAC;AAC9G,OAAO,EAAE,aAAa,EAAE,MAAM,uCAAuC,CAAC;AAEtE,OAAO,EAAE,MAAM,EAAE,MAAM,mCAAmC,CAAC;AAE3D,OAAO,EAAE,eAAe,EAAE,MAAM,2CAA2C,CAAC;AAC5E,OAAO,EAAE,oBAAoB,EAAE,MAAM,0CAA0C,CAAC;AAChF,OAAO,EAAE,WAAW,EAAE,oBAAoB,EAAa,MAAM,mBAAmB,CAAC;AACjF,OAAO,EACN,cAAc,GAKd,MAAM,YAAY,CAAC;AAuBpB,MAAM,qBAAqB,GAAG;IAC7B,SAAS,EAAE,QAAQ;IACnB,MAAM,EAAE,cAAc;CAC+B,CAAC;AAEvD;;;;;;;;;;;GAWG;AACH,MAAM,UAAU,cAAc,CAC7B,QAAuB,EACvB,EACC,KAAK,EACL,eAAe,EACf,YAAY,EACZ,KAAK,EACL,WAAW,EACX,MAAM,GAAG,IAAI,MAAM,EAAE,GACE;IAExB,IAAI,SAAoB,CAAC;IAEzB,QAAQ,QAAQ,EAAE,CAAC;QAClB,KAAK,WAAW;YACf,SAAS,GAAG,IAAI,eAAe,CAAC;gBAC/B,KAAK,EAAE,KAAK,IAAI,qBAAqB,CAAC,SAAS;gBAC/C,eAAe;gBACf,YAAY;gBACZ,MAAM;aACN,CAAC,CAAC;YACH,MAAM;QACP,KAAK,QAAQ;YACZ,SAAS,GAAG,IAAI,eAAe,CAAC;gBAC/B,KAAK,EAAE,KAAK,IAAI,qBAAqB,CAAC,MAAM;gBAC5C,eAAe;gBACf,YAAY;gBACZ,MAAM;aACN,CAAC,CAAC;YACH,MAAM;QACP,OAAO,CAAC,CAAC,CAAC;YACT,gCAAgC;YAChC,MAAM,eAAe,GAAU,QAAQ,CAAC;YACxC,MAAM,IAAI,uBAAuB,CAChC,IAAI,MAAM,CAAC,eAAe,CAAC,gDAAgD,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CACtG,CAAC;QACH,CAAC;IACF,CAAC;IAED,0FAA0F;IAC1F,OAAO,IAAI,aAAa,CAAC;QACxB,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,EAAE,KAAK,EAAE,WAAW,EAAE,CAAC;QACtD,MAAM;QACN,YAAY,EAAE,KAAK;KACnB,CAAC,CAAC;AACJ,CAAC;AAED,gGAAgG;AAChG,6DAA6D;AAC7D,MAAM,kBAAkB,GACvB,8KAA8K,CAAC;AAEhL;;;;;;;;;;;;GAYG;AACH,MAAM,CAAC,KAAK,UAAU,qBAAqB,CAAC,EAC3C,QAAQ,EACR,WAAW,EACX,aAAa,GAAG,EAAE,EAClB,MAAM,GAAG,IAAI,MAAM,EAAE,MACS,EAAE;IAChC,MAAM,KAAK,GAAG,oBAAoB,CAAC,QAAQ,CAAC,CAAC;IAC7C,MAAM,KAAK,GAAG,oBAAoB,CAAC,MAAM,eAAe,CAAC,EAAE,IAAI,EAAE,OAAO,CAAC,GAAG,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC;IAC3F,MAAM,WAAW,GAAG,CAAC,EAAE,QAAQ,EAAE,KAAK,EAAE,eAAe,EAAQ,EAAE,YAAoB,EAAE,EAAE,CACxF,cAAc,CAAC,QAAQ,EAAE,EAAE,KAAK,EAAE,eAAe,EAAE,YAAY,EAAE,KAAK,EAAE,MAAM,EAAE,CAAC,CAAC;IACnF,uEAAuE;IACvE,MAAM,YAAY,GAAG,CAAC,YAAoB,EAAE,EAAE,CAAC,GAAG,YAAY,OAAO,kBAAkB,EAAE,CAAC;IAE1F,OAAO,IAAI,iBAAiB,CAAC;QAC5B,YAAY,EAAE,WAAW,CACxB,KAAK,CAAC,OAAO,EACb,aAAa,CAAC,OAAO,IAAI,YAAY,CAAC,mBAAmB,CAAC,CAC1D;QACD,aAAa,EAAE,IAAI,eAAe,CAAC;YAClC,KAAK,EAAE,WAAW,CAAC,KAAK,CAAC,QAAQ,CAAC,QAAQ,EAAE;gBAC3C,KAAK,EAAE,KAAK,CAAC,QAAQ,CAAC,KAAK;gBAC3B,eAAe,EAAE,KAAK,CAAC,QAAQ,CAAC,eAAe;gBAC/C,WAAW;gBACX,MAAM;aACN,CAAC;YACF,YAAY,EAAE,aAAa,CAAC,QAAQ,IAAI,oBAAoB;SAC5D,CAAC;QACF,aAAa,EAAE,WAAW,CACzB,KAAK,CAAC,QAAQ,EACd,aAAa,CAAC,QAAQ,IAAI,YAAY,CAAC,oBAAoB,CAAC,CAC5D;QACD,yBAAyB,EAAE,IAAI,yBAAyB,EAAE;QAC1D,MAAM;KACN,CAAC,CAAC;AACJ,CAAC","sourcesContent":["import type { Agent } from '#src/agent/domain/agent';\nimport { InvalidAgentConfigError } from '#src/shared/domain/errors';\nimport { LLMAgent } from '#src/engines/domain/model/llmAgent';\nimport type { LLMClient } from '#src/llm/domain/llm';\nimport { ClaudeLLMClient } from '#src/llm/infrastructure/claudeLLMClient';\nimport { OpenAILLMClient } from '#src/llm/infrastructure/openAILLMClient';\nimport { InstructedAgent } from '#src/orchestration/domain/model/instructedAgent';\nimport {\n\texecutorInstructions,\n\tOrchestratorAgent,\n\tplannerInstructions,\n\treviewerInstructions,\n} from '#src/orchestration/domain/model/orchestratorAgent';\nimport { ReviewerDecisionValidator } from '#src/orchestration/infrastructure/model/reviewerDecisionValidator';\nimport { RetryingAgent } from '#src/retry/domain/model/retryingAgent';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport { Logger } from '#src/shared/infrastructure/logger';\nimport type { Tool } from '#src/tools/domain/tool';\nimport { createWorkspace } from '#src/tools/infrastructure/createWorkspace';\nimport { createWorkspaceTools } from '#src/tools/infrastructure/workspaceTools';\nimport { createAgent, orchestratorRolesFor, type Role } from './agentFactory.ts';\nimport {\n\tagentProviders,\n\ttype AgentModel,\n\ttype AgentProvider,\n\ttype CreateOrchestratorOptions,\n\ttype ReasoningEffort,\n} from './types.ts';\n\nexport interface CreateLLMAgentOptions {\n\t/**\n\t * Defaults to `'sonnet'` on Anthropic and `'gpt-5.6-luna'` on OpenAI. The provider's client\n\t * rejects a model it does not support.\n\t */\n\tmodel?: AgentModel;\n\t/** Defaults to `'high'`. The provider's client rejects an effort it does not support. */\n\treasoningEffort?: ReasoningEffort;\n\t/** Sent as the model's instructions on every call. */\n\tsystemPrompt: string;\n\t/** The tools the model may call. Defaults to none. */\n\ttools?: Tool[];\n\t/**\n\t * Runs every tool call without asking, `destructive` ones included, for CI where no one can\n\t * answer. Without it, a `destructive` call goes to the run's `approve`, or is denied.\n\t */\n\tautoApprove?: boolean;\n\t/** Defaults to warnings on stderr. */\n\tlogger?: ILogger;\n}\n\nconst defaultLLMAgentModels = {\n\tanthropic: 'sonnet',\n\topenai: 'gpt-5.6-luna',\n} as const satisfies Record<AgentProvider, AgentModel>;\n\n/**\n * Builds an agent whose conversation MiKode owns, on a provider's model API, already wrapped in\n * the harness's retry policy. It needs the provider's API key and bills per token, where\n * `createAgent` uses the Agent SDK's login. Each tool judges the risk of each call; a\n * `destructive` one runs only when the run's `approve` allows it, or with `autoApprove`.\n *\n * @throws {InvalidAgentConfigError} for an unknown provider, a model or reasoning effort its\n * client does not support, or two tools that share a name.\n * @throws {UnrecoverableError} when the client cannot be set up, for example because\n * `OPENAI_API_KEY` is missing. The Anthropic SDK resolves credentials lazily, so a missing\n * `ANTHROPIC_API_KEY` fails the first run instead, with an `UnrecoverableError`.\n */\nexport function createLLMAgent(\n\tprovider: AgentProvider,\n\t{\n\t\tmodel,\n\t\treasoningEffort,\n\t\tsystemPrompt,\n\t\ttools,\n\t\tautoApprove,\n\t\tlogger = new Logger(),\n\t}: CreateLLMAgentOptions,\n): Agent {\n\tlet llmClient: LLMClient;\n\n\tswitch (provider) {\n\t\tcase 'anthropic':\n\t\t\tllmClient = new ClaudeLLMClient({\n\t\t\t\tmodel: model ?? defaultLLMAgentModels.anthropic,\n\t\t\t\treasoningEffort,\n\t\t\t\tsystemPrompt,\n\t\t\t\tlogger,\n\t\t\t});\n\t\t\tbreak;\n\t\tcase 'openai':\n\t\t\tllmClient = new OpenAILLMClient({\n\t\t\t\tmodel: model ?? defaultLLMAgentModels.openai,\n\t\t\t\treasoningEffort,\n\t\t\t\tsystemPrompt,\n\t\t\t\tlogger,\n\t\t\t});\n\t\t\tbreak;\n\t\tdefault: {\n\t\t\t// Reachable from untyped input.\n\t\t\tconst unknownProvider: never = provider;\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`\"${String(unknownProvider)}\" is not an agent provider; expected one of: ${agentProviders.join(', ')}`,\n\t\t\t);\n\t\t}\n\t}\n\n\t// LLMAgent records nothing from a failed call, so the original prompt is the whole story.\n\treturn new RetryingAgent({\n\t\tinner: new LLMAgent({ llmClient, tools, autoApprove }),\n\t\tlogger,\n\t\tnoteFailures: false,\n\t});\n}\n\n// Sent on every call after the role's instructions, so it stays short: it only says how to find\n// one's way around the repository without reading all of it.\nconst repositoryGuidance =\n\t'Paths are relative to the repository root. Start with AGENTS.md and the architecture document it links to learn where things live, then read only the files your task needs.';\n\n/**\n * Builds the planner → executor → reviewer workflow with the planner and reviewer on the model\n * APIs, from `createLLMAgent`, each holding its role's instructions, or `systemPrompts`, as its\n * system prompt, with the read-only repository tools for the current working directory. The\n * executor stays on its Agent SDK, from `createAgent`, until an agent of ours can change files;\n * `autoApprove` reaches only it. Asynchronous because finding the program that reads the\n * repository is.\n *\n * @throws {InvalidAgentConfigError} for an unknown provider.\n * @throws {UnrecoverableError} when a role cannot be set up, for example because\n * `OPENAI_API_KEY` or the Codex CLI binary is missing, or when no program can read the\n * repository.\n */\nexport async function createLLMOrchestrator({\n\tprovider,\n\tautoApprove,\n\tsystemPrompts = {},\n\tlogger = new Logger(),\n}: CreateOrchestratorOptions = {}): Promise<Agent> {\n\tconst roles = orchestratorRolesFor(provider);\n\tconst tools = createWorkspaceTools(await createWorkspace({ root: process.cwd(), logger }));\n\tconst llmAgentFor = ({ provider, model, reasoningEffort }: Role, systemPrompt: string) =>\n\t\tcreateLLMAgent(provider, { model, reasoningEffort, systemPrompt, tools, logger });\n\t// A caller's own prompt replaces the whole of ours, guidance included.\n\tconst withGuidance = (instructions: string) => `${instructions}\\n\\n${repositoryGuidance}`;\n\n\treturn new OrchestratorAgent({\n\t\tplannerAgent: llmAgentFor(\n\t\t\troles.planner,\n\t\t\tsystemPrompts.planner ?? withGuidance(plannerInstructions),\n\t\t),\n\t\texecutorAgent: new InstructedAgent({\n\t\t\tinner: createAgent(roles.executor.provider, {\n\t\t\t\tmodel: roles.executor.model,\n\t\t\t\treasoningEffort: roles.executor.reasoningEffort,\n\t\t\t\tautoApprove,\n\t\t\t\tlogger,\n\t\t\t}),\n\t\t\tinstructions: systemPrompts.executor ?? executorInstructions,\n\t\t}),\n\t\treviewerAgent: llmAgentFor(\n\t\t\troles.reviewer,\n\t\t\tsystemPrompts.reviewer ?? withGuidance(reviewerInstructions),\n\t\t),\n\t\treviewerDecisionValidator: new ReviewerDecisionValidator(),\n\t\tlogger,\n\t});\n}\n"]}

package/dist/index.d.ts

@@ -1,7 +1,15 @@
 export type { ILogger } from './shared/domain/logger.ts';
 export { createAgent, createOrchestrator, isAgentProvider, type CreateAgentOptions, } from './factory/infrastructure/agentFactory.ts';
+export { createLLMAgent, createLLMOrchestrator, type CreateLLMAgentOptions, } from './factory/infrastructure/agentLLMFactory.ts';
 export { agentProviders, type AgentModel, type AgentProvider, type CreateOrchestratorOptions, type ReasoningEffort, } from './factory/infrastructure/types.ts';
-export { type Agent, type AgentResponse, type Callback, type ProgressEvent, } from './agent/domain/agent.ts';
-export { InvalidAgentConfigError, RecoverableError, UnrecoverableError, } from './agent/domain/errors.ts';
+export { type Agent, type AgentResponse, type Callback, type ProgressEvent, type RunOptions, } from './agent/domain/agent.ts';
+export { InvalidAgentConfigError, RecoverableError, UnrecoverableError, } from './shared/domain/errors.ts';
 export { isAbortError } from './shared/domain/isAbortError.ts';
 export type { Tokens } from './shared/domain/tokens.ts';
+export { rememberApprovals, type ApprovalDecision, type ApprovalRequest, type Approver, type RememberableDecision, type ToolRisk, } from './agent/domain/approval.ts';
+export type { JSONSchema, ToolDefinition } from './llm/domain/tool.ts';
+export type { Tool } from './tools/domain/tool.ts';
+export type { TextMatch, Workspace } from './tools/domain/workspace.ts';
+export { defineTool } from './tools/infrastructure/defineTool.ts';
+export { createWorkspace } from './tools/infrastructure/createWorkspace.ts';
+export { createWorkspaceTools } from './tools/infrastructure/workspaceTools.ts';

package/dist/index.js

@@ -1,6 +1,11 @@
 export { createAgent, createOrchestrator, isAgentProvider, } from "./factory/infrastructure/agentFactory.js";
+export { createLLMAgent, createLLMOrchestrator, } from "./factory/infrastructure/agentLLMFactory.js";
 export { agentProviders, } from "./factory/infrastructure/types.js";
 export {} from "./agent/domain/agent.js";
-export { InvalidAgentConfigError, RecoverableError, UnrecoverableError, } from "./agent/domain/errors.js";
+export { InvalidAgentConfigError, RecoverableError, UnrecoverableError, } from "./shared/domain/errors.js";
 export { isAbortError } from "./shared/domain/isAbortError.js";
+export { rememberApprovals, } from "./agent/domain/approval.js";
+export { defineTool } from "./tools/infrastructure/defineTool.js";
+export { createWorkspace } from "./tools/infrastructure/createWorkspace.js";
+export { createWorkspaceTools } from "./tools/infrastructure/workspaceTools.js";
 //# sourceMappingURL=index.js.map

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AACA,OAAO,EACN,WAAW,EACX,kBAAkB,EAClB,eAAe,GAEf,MAAM,0CAA0C,CAAC;AAClD,OAAO,EACN,cAAc,GAKd,MAAM,mCAAmC,CAAC;AAC3C,OAAO,EAKN,MAAM,yBAAyB,CAAC;AACjC,OAAO,EACN,uBAAuB,EACvB,gBAAgB,EAChB,kBAAkB,GAClB,MAAM,0BAA0B,CAAC;AAClC,OAAO,EAAE,YAAY,EAAE,MAAM,iCAAiC,CAAC","sourcesContent":["export type { ILogger } from './shared/domain/logger.ts';\nexport {\n\tcreateAgent,\n\tcreateOrchestrator,\n\tisAgentProvider,\n\ttype CreateAgentOptions,\n} from './factory/infrastructure/agentFactory.ts';\nexport {\n\tagentProviders,\n\ttype AgentModel,\n\ttype AgentProvider,\n\ttype CreateOrchestratorOptions,\n\ttype ReasoningEffort,\n} from './factory/infrastructure/types.ts';\nexport {\n\ttype Agent,\n\ttype AgentResponse,\n\ttype Callback,\n\ttype ProgressEvent,\n} from './agent/domain/agent.ts';\nexport {\n\tInvalidAgentConfigError,\n\tRecoverableError,\n\tUnrecoverableError,\n} from './agent/domain/errors.ts';\nexport { isAbortError } from './shared/domain/isAbortError.ts';\nexport type { Tokens } from './shared/domain/tokens.ts';\n"]}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AACA,OAAO,EACN,WAAW,EACX,kBAAkB,EAClB,eAAe,GAEf,MAAM,0CAA0C,CAAC;AAClD,OAAO,EACN,cAAc,EACd,qBAAqB,GAErB,MAAM,6CAA6C,CAAC;AACrD,OAAO,EACN,cAAc,GAKd,MAAM,mCAAmC,CAAC;AAC3C,OAAO,EAMN,MAAM,yBAAyB,CAAC;AACjC,OAAO,EACN,uBAAuB,EACvB,gBAAgB,EAChB,kBAAkB,GAClB,MAAM,2BAA2B,CAAC;AACnC,OAAO,EAAE,YAAY,EAAE,MAAM,iCAAiC,CAAC;AAE/D,OAAO,EACN,iBAAiB,GAMjB,MAAM,4BAA4B,CAAC;AAIpC,OAAO,EAAE,UAAU,EAAE,MAAM,sCAAsC,CAAC;AAClE,OAAO,EAAE,eAAe,EAAE,MAAM,2CAA2C,CAAC;AAC5E,OAAO,EAAE,oBAAoB,EAAE,MAAM,0CAA0C,CAAC","sourcesContent":["export type { ILogger } from './shared/domain/logger.ts';\nexport {\n\tcreateAgent,\n\tcreateOrchestrator,\n\tisAgentProvider,\n\ttype CreateAgentOptions,\n} from './factory/infrastructure/agentFactory.ts';\nexport {\n\tcreateLLMAgent,\n\tcreateLLMOrchestrator,\n\ttype CreateLLMAgentOptions,\n} from './factory/infrastructure/agentLLMFactory.ts';\nexport {\n\tagentProviders,\n\ttype AgentModel,\n\ttype AgentProvider,\n\ttype CreateOrchestratorOptions,\n\ttype ReasoningEffort,\n} from './factory/infrastructure/types.ts';\nexport {\n\ttype Agent,\n\ttype AgentResponse,\n\ttype Callback,\n\ttype ProgressEvent,\n\ttype RunOptions,\n} from './agent/domain/agent.ts';\nexport {\n\tInvalidAgentConfigError,\n\tRecoverableError,\n\tUnrecoverableError,\n} from './shared/domain/errors.ts';\nexport { isAbortError } from './shared/domain/isAbortError.ts';\nexport type { Tokens } from './shared/domain/tokens.ts';\nexport {\n\trememberApprovals,\n\ttype ApprovalDecision,\n\ttype ApprovalRequest,\n\ttype Approver,\n\ttype RememberableDecision,\n\ttype ToolRisk,\n} from './agent/domain/approval.ts';\nexport type { JSONSchema, ToolDefinition } from './llm/domain/tool.ts';\nexport type { Tool } from './tools/domain/tool.ts';\nexport type { TextMatch, Workspace } from './tools/domain/workspace.ts';\nexport { defineTool } from './tools/infrastructure/defineTool.ts';\nexport { createWorkspace } from './tools/infrastructure/createWorkspace.ts';\nexport { createWorkspaceTools } from './tools/infrastructure/workspaceTools.ts';\n"]}

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

@@ -1,4 +1,4 @@
-import { UnrecoverableError } from '#src/agent/domain/errors';
+import { UnrecoverableError } from '#src/shared/domain/errors';
 /** The context no longer fits the model's window; the agent must compact before sending again. */
 export declare class MaxContextError extends UnrecoverableError {
 }

package/dist/llm/domain/errors.js

@@ -1,4 +1,4 @@
-import { UnrecoverableError } from '#src/agent/domain/errors';
+import { UnrecoverableError } from '#src/shared/domain/errors';
 /** The context no longer fits the model's window; the agent must compact before sending again. */
 export class MaxContextError extends UnrecoverableError {
 }