@mikode13/harness 1.0.1 → 2.1.0

package/dist/llm/infrastructure/claudeLLMClient.js

@@ -0,0 +1,189 @@
+import Anthropic, { AnthropicError, APIError } from '@anthropic-ai/sdk';
+import { InvalidAgentConfigError, UnrecoverableError, withSpentTokens, } from '#src/agent/domain/errors';
+import { classifyHostFailure, classifyLocalFailure, classifyProviderFailure, describeFailure, treatErrors, } from '#src/agent/domain/providerFailure';
+import { isOneOf } from '#src/shared/domain/isOneOf';
+import { MaxContextError } from "../domain/errors.js";
+// The Messages API takes full model IDs, not the Agent SDK's aliases, so this list is its own.
+export const claudeLLMModels = ['claude-opus-5-5', 'claude-fable-5-1', 'claude-sonnet-5'];
+// Adaptive thinking spends from this budget too. Above roughly 21,000 the SDK refuses a request
+// that is not streamed, because it could outlast its ten-minute timeout.
+const maxOutputTokens = 16_000;
+// The request itself is wrong or not allowed, so sending it again cannot succeed.
+const unrecoverableStatuses = [400, 401, 403, 404, 413, 422];
+// The account is out of credit, or its key cannot make the request. Resending cannot fix
+// either, whatever status the error arrives with.
+const unrecoverableErrorTypes = ['billing_error', 'authentication_error', 'permission_error'];
+/**
+ * Only the text crosses back. Anthropic needs a thinking block again only to continue a turn
+ * that called a tool, and none is sent yet, so a reasoning part stays in the conversation for
+ * narration and is left out of the request. A message with no text is left out whole; the API
+ * joins the consecutive user turns that leaves.
+ */
+function toClaudeInput(message) {
+    const text = message.content
+        .filter(part => part.type === 'text')
+        .map(part => part.text)
+        .join('\n');
+    return text ? [{ role: message.role, content: text }] : [];
+}
+function describeBlock(block, logger) {
+    switch (block.type) {
+        case 'text':
+            return [{ type: 'text', text: block.text }];
+        case 'thinking':
+            // Summarized thinking; the signature only matters for replay, which is not done.
+            return block.thinking ? [{ type: 'reasoning', text: block.thinking }] : [];
+        case 'redacted_thinking':
+            // Encrypted by design: there is nothing to narrate.
+            return [];
+        default:
+            // No tools are sent yet, so any other block is output this client does not expect.
+            treatErrors(() => {
+                logger.warn(block, 'Claude returned a content block the client does not map');
+            }, classifyHostFailure, 'Claude client logger failed while mapping a response');
+            return [];
+    }
+}
+function toStopReason(response) {
+    switch (response.stop_reason) {
+        case 'end_turn':
+            return 'completed';
+        case 'max_tokens':
+            return 'truncated';
+        case 'refusal':
+            return 'refused';
+        case 'model_context_window_exceeded':
+            throw new MaxContextError('The context no longer fits the Claude model', {
+                cause: 'The response stopped at the model context window.',
+            });
+        default:
+            // `tool_use`, `pause_turn` and `stop_sequence` need tools, server tools or stop
+            // sequences, and this client sends none of them.
+            throw new UnrecoverableError('The Claude response stopped unexpectedly', {
+                cause: `The response stopped with "${response.stop_reason ?? 'unknown'}".`,
+            });
+    }
+}
+/** Unlike OpenAI, Anthropic counts cache reads and writes apart from `input_tokens`. */
+function toTokens(usage, logger) {
+    if (!usage) {
+        // Zeros would report a call that may have been billed as free.
+        treatErrors(() => {
+            logger.warn('Claude returned a response without usage; it cannot be accounted for');
+        }, classifyHostFailure, 'Claude client logger failed while reading usage');
+        return undefined;
+    }
+    return {
+        inputTokens: usage.input_tokens,
+        readCacheTokens: usage.cache_read_input_tokens ?? 0,
+        writtenCacheTokens: usage.cache_creation_input_tokens ?? 0,
+        outputTokens: usage.output_tokens,
+    };
+}
+function toLLMResponse(response, logger) {
+    let usage;
+    try {
+        // Inside the try: it warns when the usage is missing, and that warning can throw too.
+        usage = toTokens(response.usage, logger);
+        return {
+            message: {
+                role: 'assistant',
+                content: response.content.flatMap(block => describeBlock(block, logger)),
+            },
+            usage,
+            stopReason: toStopReason(response),
+        };
+    }
+    catch (error) {
+        // The call was billed whether the response failed or a logger did while mapping it, and
+        // an answer without usage leaves the run's count unknown.
+        throw withSpentTokens(error, usage, !response.usage);
+    }
+}
+/**
+ * The SDK resolves credentials lazily, so a missing key surfaces on the first call, as a plain
+ * `Error` the SDK gives no class of its own. An `AnthropicError` that is not an `APIError` never
+ * carries an answer from the API: on this client's path it is the SDK refusing to make the
+ * call, such as an unresolved profile or a request too large to send without streaming.
+ */
+function isLocalSDKFailure(error) {
+    return ((error instanceof AnthropicError && !(error instanceof APIError)) ||
+        (error instanceof Error && error.message.startsWith('Could not resolve authentication method')));
+}
+function classifyClaudeFailure(error) {
+    if (isLocalSDKFailure(error)) {
+        // The API never answered, and calling again would fail the same way.
+        return new UnrecoverableError('The Claude client cannot send the request', {
+            cause: describeFailure(error),
+        });
+    }
+    if (error instanceof APIError) {
+        // Anthropic gives an overlong prompt no code of its own, only a 400 with this message.
+        if (error.status === 400 && error.message.includes('prompt is too long')) {
+            return new MaxContextError('The context no longer fits the Claude model', {
+                cause: describeFailure(error),
+            });
+        }
+        if (unrecoverableErrorTypes.some(type => type === error.type) ||
+            unrecoverableStatuses.some(status => status === error.status)) {
+            return new UnrecoverableError('Claude rejected the request', {
+                cause: describeFailure(error),
+            });
+        }
+    }
+    // A rate limit, an overload (529), a 5xx or a network failure: a later attempt can succeed.
+    return classifyProviderFailure(error, 'The Claude request failed');
+}
+/**
+ * A stateless `LLMClient` on the Anthropic Messages API, which keeps nothing between calls:
+ * every call carries the whole context, so the conversation is MiKode's, not the provider's.
+ */
+export class ClaudeLLMClient {
+    client;
+    model;
+    systemPrompt;
+    logger;
+    // `model` is untyped on purpose: this adapter is the one that knows what Claude supports.
+    constructor({ model, systemPrompt, logger, }) {
+        if (!isOneOf(claudeLLMModels, model)) {
+            throw new InvalidAgentConfigError(`"${model}" is not a Claude API model; expected one of: ${claudeLLMModels.join(', ')}`);
+        }
+        try {
+            // The SDK's default retries stay on, as OpenAI's do. See "The provider SDKs keep their
+            // transport retries under RetryingAgent" in decisions.md.
+            this.client = new Anthropic();
+        }
+        catch (error) {
+            // Conflicting credential options, for example: building it again cannot fix them. A
+            // missing key does not fail here but on the first call; see `isLocalSDKFailure`.
+            throw new UnrecoverableError('The Claude client could not be set up', {
+                cause: describeFailure(error),
+            });
+        }
+        this.model = model;
+        this.systemPrompt = systemPrompt;
+        this.logger = logger;
+    }
+    async send(context, signal) {
+        let response;
+        try {
+            response = await this.client.messages.create({
+                model: this.model,
+                max_tokens: maxOutputTokens,
+                system: this.systemPrompt,
+                messages: context.flatMap(toClaudeInput),
+                thinking: { type: 'adaptive', display: 'summarized' },
+                // Anthropic caches only on request. This marks the last block, so the next call
+                // reads everything before it from the cache instead of paying for it again.
+                cache_control: { type: 'ephemeral' },
+            }, { signal });
+        }
+        catch (error) {
+            // The SDK's abort error is named 'Error', so it would pass for a provider failure.
+            signal.throwIfAborted();
+            throw classifyClaudeFailure(error);
+        }
+        return treatErrors(() => toLLMResponse(response, this.logger), classifyLocalFailure, 'Claude response mapping failed');
+    }
+}
+//# sourceMappingURL=claudeLLMClient.js.map

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

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

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

@@ -0,0 +1,22 @@
+import type { ILogger } from '#src/shared/domain/logger';
+import type { LLMClient, LLMResponse } from '../domain/llm.ts';
+import type { Message } from '../domain/message.ts';
+export declare const openAIModels: readonly ["gpt-6-astra", "gpt-5.6-sol", "gpt-5.6-luna", "gpt-5.6-terra"];
+export type OpenAIModel = (typeof openAIModels)[number];
+/**
+ * A stateless `LLMClient` on the OpenAI Responses API. With `store: false` OpenAI keeps
+ * nothing between calls, so every call carries the whole context: the conversation is
+ * MiKode's, not the provider's.
+ */
+export declare class OpenAILLMClient implements LLMClient {
+    private readonly client;
+    private readonly model;
+    private readonly systemPrompt;
+    private readonly logger;
+    constructor({ model, systemPrompt, logger, }: {
+        model: string;
+        systemPrompt: string;
+        logger: ILogger;
+    });
+    send(context: Message[], signal: AbortSignal): Promise<LLMResponse>;
+}

package/dist/llm/infrastructure/openAILLMClient.js

@@ -0,0 +1,172 @@
+import OpenAI, { APIError } from 'openai';
+import { InvalidAgentConfigError, RecoverableError, UnrecoverableError, withSpentTokens, } from '#src/agent/domain/errors';
+import { classifyHostFailure, classifyLocalFailure, classifyProviderFailure, describeFailure, treatErrors, } from '#src/agent/domain/providerFailure';
+import { isOneOf } from '#src/shared/domain/isOneOf';
+import { MaxContextError } from "../domain/errors.js";
+// The SDK types `model` as a plain string, so this list is maintained by hand.
+export const openAIModels = [
+    'gpt-6-astra',
+    'gpt-5.6-sol',
+    'gpt-5.6-luna',
+    'gpt-5.6-terra',
+];
+// The request itself is wrong or not allowed, so sending it again cannot succeed.
+const unrecoverableStatuses = [400, 401, 403, 404, 422];
+// Out of credit or over a spend limit. These arrive as a 429 like a rate limit, but the account
+// stays blocked however often the request is sent.
+const exhaustedQuotaCodes = [
+    'insufficient_quota',
+    'credit_balance_exhausted',
+    'project_spend_limit_exceeded',
+];
+// The codes OpenAI gives a failed response that a new attempt can succeed past.
+const transientResponseErrorCodes = ['server_error', 'rate_limit_exceeded'];
+/**
+ * Only the text crosses back. OpenAI replays reasoning only from the encrypted item it
+ * returned, which a `Message` does not keep, so a reasoning part stays in the conversation
+ * for narration and is left out of the request. A message with no text is left out whole.
+ */
+function toOpenAIInput(message) {
+    const text = message.content
+        .filter(part => part.type === 'text')
+        .map(part => part.text)
+        .join('\n');
+    return text ? [{ role: message.role, content: text }] : [];
+}
+function describeItem(item, logger) {
+    switch (item.type) {
+        case 'message':
+            // A refusal is not an answer: it only sets the stop reason.
+            return item.content.flatMap(content => content.type === 'output_text' ? [{ type: 'text', text: content.text }] : []);
+        case 'reasoning': {
+            // The summary is the readable part; the encrypted content is opaque by design.
+            const text = item.summary.map(summary => summary.text).join('\n');
+            return text ? [{ type: 'reasoning', text }] : [];
+        }
+        default:
+            // No tools are sent yet, so any other item is output this client does not expect.
+            treatErrors(() => {
+                logger.warn(item, 'OpenAI returned an output item the client does not map');
+            }, classifyHostFailure, 'OpenAI client logger failed while mapping a response');
+            return [];
+    }
+}
+function toStopReason(response) {
+    if (response.status === 'incomplete') {
+        return response.incomplete_details?.reason === 'content_filter' ? 'refused' : 'truncated';
+    }
+    if (response.status !== 'completed') {
+        const cause = response.error?.message ??
+            `The response ended with status "${response.status ?? 'unknown'}".`;
+        // The same failure is retried when it arrives as an SDK error, so it must be here too.
+        if (response.error && transientResponseErrorCodes.includes(response.error.code)) {
+            throw new RecoverableError('The OpenAI response failed', { cause });
+        }
+        throw new UnrecoverableError('The OpenAI response did not complete', { cause });
+    }
+    const refused = response.output.some(item => item.type === 'message' && item.content.some(content => content.type === 'refusal'));
+    return refused ? 'refused' : 'completed';
+}
+/** OpenAI counts both cache reads and cache writes inside `input_tokens`. */
+function toTokens(usage, logger) {
+    if (!usage) {
+        // Zeros would report a call that may have been billed as free.
+        treatErrors(() => {
+            logger.warn('OpenAI returned a response without usage; it cannot be accounted for');
+        }, classifyHostFailure, 'OpenAI client logger failed while reading usage');
+        return undefined;
+    }
+    const { cached_tokens, cache_write_tokens } = usage.input_tokens_details;
+    return {
+        inputTokens: usage.input_tokens - cached_tokens - cache_write_tokens,
+        readCacheTokens: cached_tokens,
+        writtenCacheTokens: cache_write_tokens,
+        outputTokens: usage.output_tokens,
+    };
+}
+function toLLMResponse(response, logger) {
+    let usage;
+    try {
+        // Inside the try: it warns when the usage is missing, and that warning can throw too.
+        usage = toTokens(response.usage, logger);
+        return {
+            message: {
+                role: 'assistant',
+                content: response.output.flatMap(item => describeItem(item, logger)),
+            },
+            usage,
+            stopReason: toStopReason(response),
+        };
+    }
+    catch (error) {
+        // The call was billed whether the response failed or a logger did while mapping it, and
+        // an answer without usage leaves the run's count unknown.
+        throw withSpentTokens(error, usage, !response.usage);
+    }
+}
+function classifyOpenAIFailure(error) {
+    if (error instanceof APIError) {
+        if (error.code === 'context_length_exceeded') {
+            return new MaxContextError('The context no longer fits the OpenAI model', {
+                cause: describeFailure(error),
+            });
+        }
+        if (exhaustedQuotaCodes.some(code => code === error.code) ||
+            unrecoverableStatuses.some(status => status === error.status)) {
+            return new UnrecoverableError('OpenAI rejected the request', {
+                cause: describeFailure(error),
+            });
+        }
+    }
+    return classifyProviderFailure(error, 'The OpenAI request failed');
+}
+/**
+ * A stateless `LLMClient` on the OpenAI Responses API. With `store: false` OpenAI keeps
+ * nothing between calls, so every call carries the whole context: the conversation is
+ * MiKode's, not the provider's.
+ */
+export class OpenAILLMClient {
+    client;
+    model;
+    systemPrompt;
+    logger;
+    // `model` is untyped on purpose: this adapter is the one that knows what OpenAI supports.
+    constructor({ model, systemPrompt, logger, }) {
+        if (!isOneOf(openAIModels, model)) {
+            throw new InvalidAgentConfigError(`"${model}" is not an OpenAI model; expected one of: ${openAIModels.join(', ')}`);
+        }
+        try {
+            // The SDK's default retries stay on: they back off and honour `Retry-After`, which
+            // `RetryingAgent` does not. See "The provider SDKs keep their transport retries" in decisions.md.
+            this.client = new OpenAI();
+        }
+        catch (error) {
+            // A missing API key, for example: building the client again cannot fix it.
+            throw new UnrecoverableError('The OpenAI client could not be set up', {
+                cause: describeFailure(error),
+            });
+        }
+        this.model = model;
+        this.systemPrompt = systemPrompt;
+        this.logger = logger;
+    }
+    async send(context, signal) {
+        let response;
+        try {
+            response = await this.client.responses.create({
+                model: this.model,
+                instructions: this.systemPrompt,
+                input: context.flatMap(toOpenAIInput),
+                store: false,
+                reasoning: { summary: 'auto' },
+            }, { signal });
+        }
+        catch (error) {
+            // The SDK's abort error is named 'Error', so it would pass for a provider failure.
+            signal.throwIfAborted();
+            throw classifyOpenAIFailure(error);
+        }
+        return treatErrors(() => toLLMResponse(response, this.logger), classifyLocalFailure, 'OpenAI response mapping failed');
+    }
+}
+//# sourceMappingURL=openAILLMClient.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"openAILLMClient.js","sourceRoot":"","sources":["../../../src/llm/infrastructure/openAILLMClient.ts"],"names":[],"mappings":"AAAA,OAAO,MAAM,EAAE,EAAE,QAAQ,EAAE,MAAM,QAAQ,CAAC;AAO1C,OAAO,EACN,uBAAuB,EACvB,gBAAgB,EAChB,kBAAkB,EAClB,eAAe,GACf,MAAM,0BAA0B,CAAC;AAClC,OAAO,EACN,mBAAmB,EACnB,oBAAoB,EACpB,uBAAuB,EACvB,eAAe,EACf,WAAW,GACX,MAAM,mCAAmC,CAAC;AAC3C,OAAO,EAAE,OAAO,EAAE,MAAM,4BAA4B,CAAC;AAGrD,OAAO,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AAItD,+EAA+E;AAC/E,MAAM,CAAC,MAAM,YAAY,GAAG;IAC3B,aAAa;IACb,aAAa;IACb,cAAc;IACd,eAAe;CACN,CAAC;AAGX,kFAAkF;AAClF,MAAM,qBAAqB,GAAG,CAAC,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,CAAC,CAAC;AAExD,gGAAgG;AAChG,mDAAmD;AACnD,MAAM,mBAAmB,GAAG;IAC3B,oBAAoB;IACpB,0BAA0B;IAC1B,8BAA8B;CAC9B,CAAC;AAEF,gFAAgF;AAChF,MAAM,2BAA2B,GAAG,CAAC,cAAc,EAAE,qBAAqB,CAAC,CAAC;AAE5E;;;;GAIG;AACH,SAAS,aAAa,CAAC,OAAgB;IACtC,MAAM,IAAI,GAAG,OAAO,CAAC,OAAO;SAC1B,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,KAAK,MAAM,CAAC;SACpC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC;SACtB,IAAI,CAAC,IAAI,CAAC,CAAC;IAEb,OAAO,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;AAC5D,CAAC;AAED,SAAS,YAAY,CAAC,IAAwB,EAAE,MAAe;IAC9D,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;QACnB,KAAK,SAAS;YACb,4DAA4D;YAC5D,OAAO,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,CACrC,OAAO,CAAC,IAAI,KAAK,aAAa,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,MAAe,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CACrF,CAAC;QACH,KAAK,WAAW,CAAC,CAAC,CAAC;YAClB,+EAA+E;YAC/E,MAAM,IAAI,GAAG,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YAClE,OAAO,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;QAClD,CAAC;QACD;YACC,kFAAkF;YAClF,WAAW,CACV,GAAG,EAAE;gBACJ,MAAM,CAAC,IAAI,CAAC,IAAI,EAAE,wDAAwD,CAAC,CAAC;YAC7E,CAAC,EACD,mBAAmB,EACnB,sDAAsD,CACtD,CAAC;YACF,OAAO,EAAE,CAAC;IACZ,CAAC;AACF,CAAC;AAED,SAAS,YAAY,CAAC,QAAkB;IACvC,IAAI,QAAQ,CAAC,MAAM,KAAK,YAAY,EAAE,CAAC;QACtC,OAAO,QAAQ,CAAC,kBAAkB,EAAE,MAAM,KAAK,gBAAgB,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,WAAW,CAAC;IAC3F,CAAC;IACD,IAAI,QAAQ,CAAC,MAAM,KAAK,WAAW,EAAE,CAAC;QACrC,MAAM,KAAK,GACV,QAAQ,CAAC,KAAK,EAAE,OAAO;YACvB,mCAAmC,QAAQ,CAAC,MAAM,IAAI,SAAS,IAAI,CAAC;QACrE,uFAAuF;QACvF,IAAI,QAAQ,CAAC,KAAK,IAAI,2BAA2B,CAAC,QAAQ,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC;YACjF,MAAM,IAAI,gBAAgB,CAAC,4BAA4B,EAAE,EAAE,KAAK,EAAE,CAAC,CAAC;QACrE,CAAC;QACD,MAAM,IAAI,kBAAkB,CAAC,sCAAsC,EAAE,EAAE,KAAK,EAAE,CAAC,CAAC;IACjF,CAAC;IAED,MAAM,OAAO,GAAG,QAAQ,CAAC,MAAM,CAAC,IAAI,CACnC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,KAAK,SAAS,IAAI,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC,OAAO,CAAC,IAAI,KAAK,SAAS,CAAC,CAC3F,CAAC;IACF,OAAO,OAAO,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,WAAW,CAAC;AAC1C,CAAC;AAED,6EAA6E;AAC7E,SAAS,QAAQ,CAAC,KAAgC,EAAE,MAAe;IAClE,IAAI,CAAC,KAAK,EAAE,CAAC;QACZ,+DAA+D;QAC/D,WAAW,CACV,GAAG,EAAE;YACJ,MAAM,CAAC,IAAI,CAAC,sEAAsE,CAAC,CAAC;QACrF,CAAC,EACD,mBAAmB,EACnB,iDAAiD,CACjD,CAAC;QACF,OAAO,SAAS,CAAC;IAClB,CAAC;IAED,MAAM,EAAE,aAAa,EAAE,kBAAkB,EAAE,GAAG,KAAK,CAAC,oBAAoB,CAAC;IACzE,OAAO;QACN,WAAW,EAAE,KAAK,CAAC,YAAY,GAAG,aAAa,GAAG,kBAAkB;QACpE,eAAe,EAAE,aAAa;QAC9B,kBAAkB,EAAE,kBAAkB;QACtC,YAAY,EAAE,KAAK,CAAC,aAAa;KACjC,CAAC;AACH,CAAC;AAED,SAAS,aAAa,CAAC,QAAkB,EAAE,MAAe;IACzD,IAAI,KAAyB,CAAC;IAE9B,IAAI,CAAC;QACJ,sFAAsF;QACtF,KAAK,GAAG,QAAQ,CAAC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;QACzC,OAAO;YACN,OAAO,EAAE;gBACR,IAAI,EAAE,WAAW;gBACjB,OAAO,EAAE,QAAQ,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC,YAAY,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;aACpE;YACD,KAAK;YACL,UAAU,EAAE,YAAY,CAAC,QAAQ,CAAC;SAClC,CAAC;IACH,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,wFAAwF;QACxF,0DAA0D;QAC1D,MAAM,eAAe,CAAC,KAAK,EAAE,KAAK,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;IACtD,CAAC;AACF,CAAC;AAED,SAAS,qBAAqB,CAAC,KAAc;IAC5C,IAAI,KAAK,YAAY,QAAQ,EAAE,CAAC;QAC/B,IAAI,KAAK,CAAC,IAAI,KAAK,yBAAyB,EAAE,CAAC;YAC9C,OAAO,IAAI,eAAe,CAAC,6CAA6C,EAAE;gBACzE,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;QACD,IACC,mBAAmB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,KAAK,KAAK,CAAC,IAAI,CAAC;YACrD,qBAAqB,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,CAAC,EAC5D,CAAC;YACF,OAAO,IAAI,kBAAkB,CAAC,6BAA6B,EAAE;gBAC5D,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;IACF,CAAC;IAED,OAAO,uBAAuB,CAAC,KAAK,EAAE,2BAA2B,CAAC,CAAC;AACpE,CAAC;AAED;;;;GAIG;AACH,MAAM,OAAO,eAAe;IACV,MAAM,CAAS;IACf,KAAK,CAAc;IACnB,YAAY,CAAS;IACrB,MAAM,CAAU;IAEjC,0FAA0F;IAC1F,YAAY,EACX,KAAK,EACL,YAAY,EACZ,MAAM,GAKN;QACA,IAAI,CAAC,OAAO,CAAC,YAAY,EAAE,KAAK,CAAC,EAAE,CAAC;YACnC,MAAM,IAAI,uBAAuB,CAChC,IAAI,KAAK,8CAA8C,YAAY,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAChF,CAAC;QACH,CAAC;QAED,IAAI,CAAC;YACJ,mFAAmF;YACnF,kGAAkG;YAClG,IAAI,CAAC,MAAM,GAAG,IAAI,MAAM,EAAE,CAAC;QAC5B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,2EAA2E;YAC3E,MAAM,IAAI,kBAAkB,CAAC,uCAAuC,EAAE;gBACrE,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;QAED,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,YAAY,GAAG,YAAY,CAAC;QACjC,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IACtB,CAAC;IAED,KAAK,CAAC,IAAI,CAAC,OAAkB,EAAE,MAAmB;QACjD,IAAI,QAAkB,CAAC;QAEvB,IAAI,CAAC;YACJ,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,MAAM,CAC5C;gBACC,KAAK,EAAE,IAAI,CAAC,KAAK;gBACjB,YAAY,EAAE,IAAI,CAAC,YAAY;gBAC/B,KAAK,EAAE,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC;gBACrC,KAAK,EAAE,KAAK;gBACZ,SAAS,EAAE,EAAE,OAAO,EAAE,MAAM,EAAE;aAC9B,EACD,EAAE,MAAM,EAAE,CACV,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,mFAAmF;YACnF,MAAM,CAAC,cAAc,EAAE,CAAC;YACxB,MAAM,qBAAqB,CAAC,KAAK,CAAC,CAAC;QACpC,CAAC;QAED,OAAO,WAAW,CACjB,GAAG,EAAE,CAAC,aAAa,CAAC,QAAQ,EAAE,IAAI,CAAC,MAAM,CAAC,EAC1C,oBAAoB,EACpB,gCAAgC,CAChC,CAAC;IACH,CAAC;CACD","sourcesContent":["import OpenAI, { APIError } from 'openai';\nimport type {\n\tResponse,\n\tResponseInputItem,\n\tResponseOutputItem,\n\tResponseUsage,\n} from 'openai/resources/responses/responses';\nimport {\n\tInvalidAgentConfigError,\n\tRecoverableError,\n\tUnrecoverableError,\n\twithSpentTokens,\n} from '#src/agent/domain/errors';\nimport {\n\tclassifyHostFailure,\n\tclassifyLocalFailure,\n\tclassifyProviderFailure,\n\tdescribeFailure,\n\ttreatErrors,\n} from '#src/agent/domain/providerFailure';\nimport { isOneOf } from '#src/shared/domain/isOneOf';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport type { Tokens } from '#src/shared/domain/tokens';\nimport { MaxContextError } from '../domain/errors.ts';\nimport type { LLMClient, LLMResponse, StopReason } from '../domain/llm.ts';\nimport type { Message, MessagePart } from '../domain/message.ts';\n\n// The SDK types `model` as a plain string, so this list is maintained by hand.\nexport const openAIModels = [\n\t'gpt-6-astra',\n\t'gpt-5.6-sol',\n\t'gpt-5.6-luna',\n\t'gpt-5.6-terra',\n] as const;\nexport type OpenAIModel = (typeof openAIModels)[number];\n\n// The request itself is wrong or not allowed, so sending it again cannot succeed.\nconst unrecoverableStatuses = [400, 401, 403, 404, 422];\n\n// Out of credit or over a spend limit. These arrive as a 429 like a rate limit, but the account\n// stays blocked however often the request is sent.\nconst exhaustedQuotaCodes = [\n\t'insufficient_quota',\n\t'credit_balance_exhausted',\n\t'project_spend_limit_exceeded',\n];\n\n// The codes OpenAI gives a failed response that a new attempt can succeed past.\nconst transientResponseErrorCodes = ['server_error', 'rate_limit_exceeded'];\n\n/**\n * Only the text crosses back. OpenAI replays reasoning only from the encrypted item it\n * returned, which a `Message` does not keep, so a reasoning part stays in the conversation\n * for narration and is left out of the request. A message with no text is left out whole.\n */\nfunction toOpenAIInput(message: Message): ResponseInputItem[] {\n\tconst text = message.content\n\t\t.filter(part => part.type === 'text')\n\t\t.map(part => part.text)\n\t\t.join('\\n');\n\n\treturn text ? [{ role: message.role, content: text }] : [];\n}\n\nfunction describeItem(item: ResponseOutputItem, logger: ILogger): MessagePart[] {\n\tswitch (item.type) {\n\t\tcase 'message':\n\t\t\t// A refusal is not an answer: it only sets the stop reason.\n\t\t\treturn item.content.flatMap(content =>\n\t\t\t\tcontent.type === 'output_text' ? [{ type: 'text' as const, text: content.text }] : [],\n\t\t\t);\n\t\tcase 'reasoning': {\n\t\t\t// The summary is the readable part; the encrypted content is opaque by design.\n\t\t\tconst text = item.summary.map(summary => summary.text).join('\\n');\n\t\t\treturn text ? [{ type: 'reasoning', text }] : [];\n\t\t}\n\t\tdefault:\n\t\t\t// No tools are sent yet, so any other item is output this client does not expect.\n\t\t\ttreatErrors(\n\t\t\t\t() => {\n\t\t\t\t\tlogger.warn(item, 'OpenAI returned an output item the client does not map');\n\t\t\t\t},\n\t\t\t\tclassifyHostFailure,\n\t\t\t\t'OpenAI client logger failed while mapping a response',\n\t\t\t);\n\t\t\treturn [];\n\t}\n}\n\nfunction toStopReason(response: Response): StopReason {\n\tif (response.status === 'incomplete') {\n\t\treturn response.incomplete_details?.reason === 'content_filter' ? 'refused' : 'truncated';\n\t}\n\tif (response.status !== 'completed') {\n\t\tconst cause =\n\t\t\tresponse.error?.message ??\n\t\t\t`The response ended with status \"${response.status ?? 'unknown'}\".`;\n\t\t// The same failure is retried when it arrives as an SDK error, so it must be here too.\n\t\tif (response.error && transientResponseErrorCodes.includes(response.error.code)) {\n\t\t\tthrow new RecoverableError('The OpenAI response failed', { cause });\n\t\t}\n\t\tthrow new UnrecoverableError('The OpenAI response did not complete', { cause });\n\t}\n\n\tconst refused = response.output.some(\n\t\titem => item.type === 'message' && item.content.some(content => content.type === 'refusal'),\n\t);\n\treturn refused ? 'refused' : 'completed';\n}\n\n/** OpenAI counts both cache reads and cache writes inside `input_tokens`. */\nfunction toTokens(usage: ResponseUsage | undefined, logger: ILogger): Tokens | undefined {\n\tif (!usage) {\n\t\t// Zeros would report a call that may have been billed as free.\n\t\ttreatErrors(\n\t\t\t() => {\n\t\t\t\tlogger.warn('OpenAI returned a response without usage; it cannot be accounted for');\n\t\t\t},\n\t\t\tclassifyHostFailure,\n\t\t\t'OpenAI client logger failed while reading usage',\n\t\t);\n\t\treturn undefined;\n\t}\n\n\tconst { cached_tokens, cache_write_tokens } = usage.input_tokens_details;\n\treturn {\n\t\tinputTokens: usage.input_tokens - cached_tokens - cache_write_tokens,\n\t\treadCacheTokens: cached_tokens,\n\t\twrittenCacheTokens: cache_write_tokens,\n\t\toutputTokens: usage.output_tokens,\n\t};\n}\n\nfunction toLLMResponse(response: Response, logger: ILogger): LLMResponse {\n\tlet usage: Tokens | undefined;\n\n\ttry {\n\t\t// Inside the try: it warns when the usage is missing, and that warning can throw too.\n\t\tusage = toTokens(response.usage, logger);\n\t\treturn {\n\t\t\tmessage: {\n\t\t\t\trole: 'assistant',\n\t\t\t\tcontent: response.output.flatMap(item => describeItem(item, logger)),\n\t\t\t},\n\t\t\tusage,\n\t\t\tstopReason: toStopReason(response),\n\t\t};\n\t} catch (error) {\n\t\t// The call was billed whether the response failed or a logger did while mapping it, and\n\t\t// an answer without usage leaves the run's count unknown.\n\t\tthrow withSpentTokens(error, usage, !response.usage);\n\t}\n}\n\nfunction classifyOpenAIFailure(error: unknown): Error {\n\tif (error instanceof APIError) {\n\t\tif (error.code === 'context_length_exceeded') {\n\t\t\treturn new MaxContextError('The context no longer fits the OpenAI model', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t\tif (\n\t\t\texhaustedQuotaCodes.some(code => code === error.code) ||\n\t\t\tunrecoverableStatuses.some(status => status === error.status)\n\t\t) {\n\t\t\treturn new UnrecoverableError('OpenAI rejected the request', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t}\n\n\treturn classifyProviderFailure(error, 'The OpenAI request failed');\n}\n\n/**\n * A stateless `LLMClient` on the OpenAI Responses API. With `store: false` OpenAI keeps\n * nothing between calls, so every call carries the whole context: the conversation is\n * MiKode's, not the provider's.\n */\nexport class OpenAILLMClient implements LLMClient {\n\tprivate readonly client: OpenAI;\n\tprivate readonly model: OpenAIModel;\n\tprivate readonly systemPrompt: string;\n\tprivate readonly logger: ILogger;\n\n\t// `model` is untyped on purpose: this adapter is the one that knows what OpenAI supports.\n\tconstructor({\n\t\tmodel,\n\t\tsystemPrompt,\n\t\tlogger,\n\t}: {\n\t\tmodel: string;\n\t\tsystemPrompt: string;\n\t\tlogger: ILogger;\n\t}) {\n\t\tif (!isOneOf(openAIModels, model)) {\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`\"${model}\" is not an OpenAI model; expected one of: ${openAIModels.join(', ')}`,\n\t\t\t);\n\t\t}\n\n\t\ttry {\n\t\t\t// The SDK's default retries stay on: they back off and honour `Retry-After`, which\n\t\t\t// `RetryingAgent` does not. See \"The provider SDKs keep their transport retries\" in decisions.md.\n\t\t\tthis.client = new OpenAI();\n\t\t} catch (error) {\n\t\t\t// A missing API key, for example: building the client again cannot fix it.\n\t\t\tthrow new UnrecoverableError('The OpenAI client could not be set up', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\n\t\tthis.model = model;\n\t\tthis.systemPrompt = systemPrompt;\n\t\tthis.logger = logger;\n\t}\n\n\tasync send(context: Message[], signal: AbortSignal): Promise<LLMResponse> {\n\t\tlet response: Response;\n\n\t\ttry {\n\t\t\tresponse = await this.client.responses.create(\n\t\t\t\t{\n\t\t\t\t\tmodel: this.model,\n\t\t\t\t\tinstructions: this.systemPrompt,\n\t\t\t\t\tinput: context.flatMap(toOpenAIInput),\n\t\t\t\t\tstore: false,\n\t\t\t\t\treasoning: { summary: 'auto' },\n\t\t\t\t},\n\t\t\t\t{ signal },\n\t\t\t);\n\t\t} catch (error) {\n\t\t\t// The SDK's abort error is named 'Error', so it would pass for a provider failure.\n\t\t\tsignal.throwIfAborted();\n\t\t\tthrow classifyOpenAIFailure(error);\n\t\t}\n\n\t\treturn treatErrors(\n\t\t\t() => toLLMResponse(response, this.logger),\n\t\t\tclassifyLocalFailure,\n\t\t\t'OpenAI response mapping failed',\n\t\t);\n\t}\n}\n"]}

package/dist/orchestration/domain/model/orchestratorAgent.d.ts

@@ -1,7 +1,7 @@
-import type { Agent, AgentResponse, Callback } from '../../../agent/domain/agent.ts';
+import type { Agent, AgentResponse, Callback } from '#src/agent/domain/agent';
 import type { ReviewerDecision } from './reviewerDecision.ts';
 import type { Validator } from '../interface/validator.ts';
-import type { ILogger } from '../../../shared/domain/logger.ts';
+import type { ILogger } from '#src/shared/domain/logger';
 export declare class OrchestratorAgent implements Agent {
     private plannerAgent;
     private executorAgent;
@@ -17,7 +17,8 @@ export declare class OrchestratorAgent implements Agent {
         maxAttempts?: number;
         logger: ILogger;
     });
-    run(prompt: string, signal: AbortSignal, callback: Callback): Promise<AgentResponse | undefined>;
+    run(prompt: string, signal: AbortSignal, callback: Callback): Promise<AgentResponse>;
+    private runRounds;
     private getReviewerDecision;
     private warn;
 }

package/dist/orchestration/domain/model/orchestratorAgent.js

@@ -1,5 +1,6 @@
-import { RecoverableError, UnrecoverableError } from "../../../agent/domain/errors.js";
-import { classifyHostFailure, treatErrors } from "../../../agent/domain/providerFailure.js";
+import { addTokens } from '#src/shared/domain/tokens';
+import { RecoverableError, UnrecoverableError, withSpentTokens } from '#src/agent/domain/errors';
+import { classifyHostFailure, treatErrors } from '#src/agent/domain/providerFailure';
 const getPlannerPrompt = (userPrompt, previousFailureReason) => {
     const feedback = previousFailureReason
         ? `\n\nFeedback from the previous attempt:\n---\n${previousFailureReason}\n---`
@@ -43,17 +44,22 @@ Executor response (context only; verify it independently):
 ${executorResult}
 ---${retryNotice}`;
 };
+// Every role's response counts, even an empty one that starts another round: it was billed.
 function addToTotals(totals, response) {
-    totals.duration += response.duration;
-    totals.inputTokens += response.inputTokens;
-    totals.outputTokens += response.outputTokens;
+    if (!response.tokens)
+        totals.unreported = true;
+    totals.tokens = addTokens(totals.tokens, response.tokens);
+}
+/** Unknown rather than partial: a sum missing a billed call must not look complete. */
+function runTokens(totals) {
+    return totals.unreported ? undefined : totals.tokens;
 }
 function stripCodeFence(text) {
     const match = /^```(?:json)?\s*([\s\S]*?)\s*```$/.exec(text.trim());
     return match?.[1] ?? text;
 }
 function parseReviewerDecision(response, reviewerDecisionValidator) {
-    if (!response?.response) {
+    if (!response.response) {
         throw new RecoverableError("There's no decision from the reviewer, something went wrong", {
             cause: 'Reviewer decision is missing.',
         });
@@ -94,13 +100,30 @@ export class OrchestratorAgent {
         this.logger = logger;
     }
     async run(prompt, signal, callback) {
+        const start = Date.now();
         // Per invocation, not per instance: the CLI keeps one orchestrator for a whole session.
-        const totals = { duration: 0, inputTokens: 0, outputTokens: 0 };
+        const totals = { tokens: undefined, unreported: false };
+        try {
+            await this.runRounds(prompt, signal, callback, totals);
+        }
+        catch (error) {
+            // A failing role carries its own tokens; the roles before it are in the totals.
+            throw withSpentTokens(error, totals.tokens, totals.unreported);
+        }
+        return {
+            response: 'All job has finished',
+            tokens: runTokens(totals),
+            // The whole run's wall clock, so it matches what the consumer waited.
+            duration: (Date.now() - start) / 1000,
+        };
+    }
+    async runRounds(prompt, signal, callback, totals) {
         let lastFailureReason;
         for (let attempt = 1; attempt <= this.maxAttempts; attempt++) {
             const isLastAttempt = attempt === this.maxAttempts;
             const plannerResponse = await this.plannerAgent.run(getPlannerPrompt(prompt, lastFailureReason), signal, callback);
-            if (!plannerResponse?.response) {
+            addToTotals(totals, plannerResponse);
+            if (!plannerResponse.response) {
                 lastFailureReason = 'The planner produced no response.';
                 if (isLastAttempt) {
                     throw new UnrecoverableError('Max attempts exhausted', { cause: lastFailureReason });
@@ -108,9 +131,9 @@ export class OrchestratorAgent {
                 this.warn(`Attempt ${String(attempt)}/${String(this.maxAttempts)}: the planner produced no response; starting another round`);
                 continue;
             }
-            addToTotals(totals, plannerResponse);
             const executorResponse = await this.executorAgent.run(getExecutorPrompt(prompt, plannerResponse.response), signal, callback);
-            if (!executorResponse?.response) {
+            addToTotals(totals, executorResponse);
+            if (!executorResponse.response) {
                 lastFailureReason = 'The executor produced no response.';
                 if (isLastAttempt) {
                     throw new UnrecoverableError('Max attempts exhausted', { cause: lastFailureReason });
@@ -118,11 +141,9 @@ export class OrchestratorAgent {
                 this.warn(`Attempt ${String(attempt)}/${String(this.maxAttempts)}: the executor produced no response; starting another round`);
                 continue;
             }
-            addToTotals(totals, executorResponse);
             const reviewerDecision = await this.getReviewerDecision(prompt, plannerResponse.response, executorResponse.response, signal, callback, totals);
-            if (reviewerDecision.decision === 'approved') {
-                return { response: 'All job has finished', ...totals };
-            }
+            if (reviewerDecision.decision === 'approved')
+                return;
             lastFailureReason = reviewerDecision.feedback;
             if (isLastAttempt) {
                 throw new UnrecoverableError('Max attempts exhausted', { cause: lastFailureReason });
@@ -140,9 +161,7 @@ export class OrchestratorAgent {
         let parseFailureReason;
         for (let attempt = 1; attempt <= this.maxAttempts; attempt++) {
             const reviewerResponse = await this.reviewerAgent.run(getReviewerPrompt(userPrompt, plannerPrompt, executorResult, parseFailureReason), signal, callback);
-            if (reviewerResponse) {
-                addToTotals(totals, reviewerResponse);
-            }
+            addToTotals(totals, reviewerResponse);
             try {
                 return parseReviewerDecision(reviewerResponse, this.reviewerDecisionValidator);
             }