@mikode13/harness 4.0.1 → 5.0.0

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

@@ -1 +1 @@
-{"version":3,"file":"claudeLLMClient.js","sourceRoot":"","sources":["../../../src/llm/infrastructure/claudeLLMClient.ts"],"names":[],"mappings":"AAAA,OAAO,SAAS,EAAE,EAAE,cAAc,EAAE,QAAQ,EAAE,MAAM,mBAAmB,CAAC;AAYxE,OAAO,EACN,uBAAuB,EACvB,kBAAkB,EAClB,eAAe,GACf,MAAM,2BAA2B,CAAC;AACnC,OAAO,EACN,mBAAmB,EACnB,oBAAoB,EACpB,uBAAuB,EACvB,eAAe,EACf,WAAW,GACX,MAAM,oCAAoC,CAAC;AAC5C,OAAO,EAAE,OAAO,EAAE,MAAM,4BAA4B,CAAC;AAGrD,OAAO,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AAKtD,gGAAgG;AAChG,8FAA8F;AAC9F,yFAAyF;AACzF,MAAM,CAAC,MAAM,eAAe,GAAG,CAAC,MAAM,EAAE,OAAO,EAAE,QAAQ,CAAU,CAAC;AAGpE,MAAM,iBAAiB,GAAG;IACzB,IAAI,EAAE,iBAAiB;IACvB,KAAK,EAAE,kBAAkB;IACzB,MAAM,EAAE,iBAAiB;CACyB,CAAC;AAEpD,MAAM,CAAC,MAAM,yBAAyB,GAAG;IACxC,KAAK;IACL,QAAQ;IACR,MAAM;IACN,OAAO;IACP,KAAK;CAC6D,CAAC;AAGpE,gGAAgG;AAChG,yEAAyE;AACzE,MAAM,eAAe,GAAG,MAAM,CAAC;AAE/B,kFAAkF;AAClF,MAAM,qBAAqB,GAAG,CAAC,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,CAAC,CAAC;AAE7D,yFAAyF;AACzF,kDAAkD;AAClD,MAAM,uBAAuB,GAAG,CAAC,eAAe,EAAE,sBAAsB,EAAE,kBAAkB,CAAC,CAAC;AAE9F,0FAA0F;AAC1F,MAAM,YAAY,GAAG,QAAQ,CAAC;AAE9B;;;;GAIG;AACH,SAAS,mBAAmB,CAAC,KAAc;IAC1C,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC;AAC1F,CAAC;AAED,0FAA0F;AAC1F,SAAS,cAAc,CAAC,IAAiB;IACxC,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;QACnB,KAAK,MAAM;YACV,uCAAuC;YACvC,OAAO,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;QAC7D,KAAK,UAAU;YACd,OAAO;gBACN,EAAE,IAAI,EAAE,UAAU,EAAE,EAAE,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,KAAK,EAAE,mBAAmB,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE;aAC1F,CAAC;QACH,KAAK,YAAY;YAChB,OAAO;gBACN;oBACC,IAAI,EAAE,aAAa;oBACnB,WAAW,EAAE,IAAI,CAAC,MAAM;oBACxB,OAAO,EAAE,IAAI,CAAC,MAAM;oBACpB,QAAQ,EAAE,IAAI,CAAC,OAAO;iBACtB;aACD,CAAC;QACH,KAAK,cAAc;YAClB,8FAA8F;YAC9F,OAAO,IAAI,CAAC,MAAM,KAAK,YAAY;gBAClC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAuD,CAAC;gBAChE,CAAC,CAAC,EAAE,CAAC;QACP,KAAK,WAAW;YACf,mFAAmF;YACnF,OAAO,EAAE,CAAC;IACZ,CAAC;AACF,CAAC;AAED;;;GAGG;AACH,SAAS,aAAa,CAAC,OAAgB;IACtC,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC,cAAc,CAAC,CAAC;IACxD,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QAC1B,OAAO,EAAE,CAAC;IACX,CAAC;IAED,OAAO,CAAC,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC;AAC7E,CAAC;AAED;;;GAGG;AACH,SAAS,YAAY,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,WAAW,EAAkB;IACvE,OAAO,EAAE,IAAI,EAAE,WAAW,EAAE,YAAY,EAAE,WAAW,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC;AACvE,CAAC;AAED,SAAS,aAAa,CAAC,KAAmB,EAAE,MAAe;IAC1D,QAAQ,KAAK,CAAC,IAAI,EAAE,CAAC;QACpB,KAAK,MAAM;YACV,OAAO,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC;QAC7C,KAAK,UAAU,CAAC,CAAC,CAAC;YACjB,qEAAqE;YACrE,MAAM,MAAM,GAAuB;gBAClC,IAAI,EAAE,UAAU;gBAChB,QAAQ,EAAE,KAAK,CAAC,QAAQ;gBACxB,SAAS,EAAE,KAAK,CAAC,SAAS;aAC1B,CAAC;YACF,MAAM,MAAM,GAAgB,EAAE,IAAI,EAAE,cAAc,EAAE,MAAM,EAAE,YAAY,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC;YACzF,OAAO,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,KAAK,CAAC,QAAQ,EAAE,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;QAC1F,CAAC;QACD,KAAK,mBAAmB,CAAC,CAAC,CAAC;YAC1B,mEAAmE;YACnE,MAAM,QAAQ,GAA+B,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,CAAC;YAC7F,OAAO,CAAC,EAAE,IAAI,EAAE,cAAc,EAAE,MAAM,EAAE,YAAY,EAAE,IAAI,EAAE,QAAQ,EAAE,CAAC,CAAC;QACzE,CAAC;QACD,KAAK,UAAU;YACd,OAAO,CAAC,EAAE,IAAI,EAAE,UAAU,EAAE,EAAE,EAAE,KAAK,CAAC,EAAE,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,CAAC,CAAC;QACnF;YACC,4FAA4F;YAC5F,WAAW,CACV,GAAG,EAAE;gBACJ,MAAM,CAAC,IAAI,CAAC,KAAK,EAAE,yDAAyD,CAAC,CAAC;YAC/E,CAAC,EACD,mBAAmB,EACnB,sDAAsD,CACtD,CAAC;YACF,OAAO,EAAE,CAAC;IACZ,CAAC;AACF,CAAC;AAED,SAAS,YAAY,CAAC,QAA0B;IAC/C,QAAQ,QAAQ,CAAC,WAAW,EAAE,CAAC;QAC9B,yFAAyF;QACzF,KAAK,UAAU,CAAC;QAChB,KAAK,UAAU;YACd,OAAO,WAAW,CAAC;QACpB,KAAK,YAAY;YAChB,OAAO,WAAW,CAAC;QACpB,KAAK,SAAS;YACb,OAAO,SAAS,CAAC;QAClB,KAAK,+BAA+B;YACnC,MAAM,IAAI,eAAe,CAAC,6CAA6C,EAAE;gBACxE,KAAK,EAAE,mDAAmD;aAC1D,CAAC,CAAC;QACJ;YACC,iFAAiF;YACjF,wBAAwB;YACxB,MAAM,IAAI,kBAAkB,CAAC,0CAA0C,EAAE;gBACxE,KAAK,EAAE,8BAA8B,QAAQ,CAAC,WAAW,IAAI,SAAS,IAAI;aAC1E,CAAC,CAAC;IACL,CAAC;AACF,CAAC;AAED,wFAAwF;AACxF,SAAS,QAAQ,CAAC,KAAwB,EAAE,MAAe;IAC1D,IAAI,CAAC,KAAK,EAAE,CAAC;QACZ,+DAA+D;QAC/D,WAAW,CACV,GAAG,EAAE;YACJ,MAAM,CAAC,IAAI,CAAC,sEAAsE,CAAC,CAAC;QACrF,CAAC,EACD,mBAAmB,EACnB,iDAAiD,CACjD,CAAC;QACF,OAAO,SAAS,CAAC;IAClB,CAAC;IAED,OAAO;QACN,WAAW,EAAE,KAAK,CAAC,YAAY;QAC/B,eAAe,EAAE,KAAK,CAAC,uBAAuB,IAAI,CAAC;QACnD,kBAAkB,EAAE,KAAK,CAAC,2BAA2B,IAAI,CAAC;QAC1D,YAAY,EAAE,KAAK,CAAC,aAAa;KACjC,CAAC;AACH,CAAC;AAED,SAAS,aAAa,CAAC,QAA0B,EAAE,MAAe;IACjE,IAAI,KAAyB,CAAC;IAE9B,IAAI,CAAC;QACJ,sFAAsF;QACtF,KAAK,GAAG,QAAQ,CAAC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;QACzC,OAAO;YACN,OAAO,EAAE;gBACR,IAAI,EAAE,WAAW;gBACjB,OAAO,EAAE,QAAQ,CAAC,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC,aAAa,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;aACxE;YACD,KAAK;YACL,UAAU,EAAE,YAAY,CAAC,QAAQ,CAAC;SAClC,CAAC;IACH,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,wFAAwF;QACxF,0DAA0D;QAC1D,MAAM,eAAe,CAAC,KAAK,EAAE,KAAK,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;IACtD,CAAC;AACF,CAAC;AAED;;;;;GAKG;AACH,SAAS,iBAAiB,CAAC,KAAc;IACxC,OAAO,CACN,CAAC,KAAK,YAAY,cAAc,IAAI,CAAC,CAAC,KAAK,YAAY,QAAQ,CAAC,CAAC;QACjE,CAAC,KAAK,YAAY,KAAK,IAAI,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,yCAAyC,CAAC,CAAC,CAC/F,CAAC;AACH,CAAC;AAED,SAAS,qBAAqB,CAAC,KAAc;IAC5C,IAAI,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;QAC9B,qEAAqE;QACrE,OAAO,IAAI,kBAAkB,CAAC,2CAA2C,EAAE;YAC1E,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;SAC7B,CAAC,CAAC;IACJ,CAAC;IACD,IAAI,KAAK,YAAY,QAAQ,EAAE,CAAC;QAC/B,uFAAuF;QACvF,IAAI,KAAK,CAAC,MAAM,KAAK,GAAG,IAAI,KAAK,CAAC,OAAO,CAAC,QAAQ,CAAC,oBAAoB,CAAC,EAAE,CAAC;YAC1E,OAAO,IAAI,eAAe,CAAC,6CAA6C,EAAE;gBACzE,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;QACD,IACC,uBAAuB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,KAAK,KAAK,CAAC,IAAI,CAAC;YACzD,qBAAqB,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,CAAC,EAC5D,CAAC;YACF,OAAO,IAAI,kBAAkB,CAAC,6BAA6B,EAAE;gBAC5D,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;IACF,CAAC;IAED,4FAA4F;IAC5F,OAAO,uBAAuB,CAAC,KAAK,EAAE,2BAA2B,CAAC,CAAC;AACpE,CAAC;AAED;;;GAGG;AACH,MAAM,OAAO,eAAe;IACV,MAAM,CAAY;IAClB,KAAK,CAA6C;IAClD,eAAe,CAA2B;IAC1C,YAAY,CAAS;IACrB,MAAM,CAAU;IAEjC,2FAA2F;IAC3F,wBAAwB;IACxB,YAAY,EACX,KAAK,EACL,eAAe,GAAG,MAAM,EACxB,YAAY,EACZ,MAAM,GAMN;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;QACD,IAAI,CAAC,OAAO,CAAC,yBAAyB,EAAE,eAAe,CAAC,EAAE,CAAC;YAC1D,MAAM,IAAI,uBAAuB,CAChC,IAAI,eAAe,4DAA4D,yBAAyB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CACrH,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,iBAAiB,CAAC,KAAK,CAAC,CAAC;QACtC,IAAI,CAAC,eAAe,GAAG,eAAe,CAAC;QACvC,IAAI,CAAC,YAAY,GAAG,YAAY,CAAC;QACjC,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IACtB,CAAC;IAED,KAAK,CAAC,IAAI,CACT,EAAE,OAAO,EAAE,KAAK,EAAmD,EACnE,MAAmB;QAEnB,IAAI,QAA0B,CAAC;QAE/B,IAAI,CAAC;YACJ,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAC3C;gBACC,KAAK,EAAE,IAAI,CAAC,KAAK;gBACjB,UAAU,EAAE,eAAe;gBAC3B,MAAM,EAAE,IAAI,CAAC,YAAY;gBACzB,QAAQ,EAAE,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC;gBACxC,0EAA0E;gBAC1E,GAAG,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,EAAE,CAAC;gBAC3D,QAAQ,EAAE,EAAE,IAAI,EAAE,UAAU,EAAE,OAAO,EAAE,YAAY,EAAE;gBACrD,aAAa,EAAE,EAAE,MAAM,EAAE,IAAI,CAAC,eAAe,EAAE;gBAC/C,gFAAgF;gBAChF,4EAA4E;gBAC5E,aAAa,EAAE,EAAE,IAAI,EAAE,WAAW,EAAE;aACpC,EACD,EAAE,MAAM,EAAE,CACV,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,mFAAmF;YACnF,MAAM,CAAC,cAAc,EAAE,CAAC;YACxB,MAAM,qBAAqB,CAAC,KAAK,CAAC,CAAC;QACpC,CAAC;QAED,OAAO,WAAW,CACjB,GAAG,EAAE,CAAC,aAAa,CAAC,QAAQ,EAAE,IAAI,CAAC,MAAM,CAAC,EAC1C,oBAAoB,EACpB,gCAAgC,CAChC,CAAC;IACH,CAAC;CACD","sourcesContent":["import Anthropic, { AnthropicError, APIError } from '@anthropic-ai/sdk';\nimport type {\n\tContentBlock,\n\tContentBlockParam,\n\tMessage as AnthropicMessage,\n\tMessageParam,\n\tOutputConfig,\n\tRedactedThinkingBlockParam,\n\tThinkingBlockParam,\n\tTool as ClaudeTool,\n\tUsage,\n} from '@anthropic-ai/sdk/resources/messages';\nimport {\n\tInvalidAgentConfigError,\n\tUnrecoverableError,\n\twithSpentTokens,\n} from '#src/shared/domain/errors';\nimport {\n\tclassifyHostFailure,\n\tclassifyLocalFailure,\n\tclassifyProviderFailure,\n\tdescribeFailure,\n\ttreatErrors,\n} from '#src/shared/domain/providerFailure';\nimport { isOneOf } from '#src/shared/domain/isOneOf';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport type { Tokens } from '#src/shared/domain/tokens';\nimport { MaxContextError } from '../domain/errors.ts';\nimport type { LLMClient, LLMResponse, StopReason } from '../domain/llm.ts';\nimport type { Message, MessagePart } from '../domain/message.ts';\nimport type { ToolDefinition } from '../domain/tool.ts';\n\n// The Messages API takes full model IDs. Callers name a model by the Agent SDK's alias, so both\n// paths share one vocabulary; this table pins the ID each alias means here, and moves by hand\n// when Anthropic releases a newer model. Haiku is left out: it has no adaptive thinking.\nexport const claudeLLMModels = ['opus', 'fable', 'sonnet'] as const;\nexport type ClaudeLLMModel = (typeof claudeLLMModels)[number];\n\nconst claudeLLMModelIds = {\n\topus: 'claude-opus-5-5',\n\tfable: 'claude-fable-5-1',\n\tsonnet: 'claude-sonnet-5',\n} as const satisfies Record<ClaudeLLMModel, string>;\n\nexport const claudeLLMReasoningEfforts = [\n\t'low',\n\t'medium',\n\t'high',\n\t'xhigh',\n\t'max',\n] as const satisfies readonly NonNullable<OutputConfig['effort']>[];\nexport type ClaudeLLMReasoningEffort = (typeof claudeLLMReasoningEfforts)[number];\n\n// Adaptive thinking spends from this budget too. Above roughly 21,000 the SDK refuses a request\n// that is not streamed, because it could outlast its ten-minute timeout.\nconst maxOutputTokens = 16_000;\n\n// The request itself is wrong or not allowed, so sending it again cannot succeed.\nconst unrecoverableStatuses = [400, 401, 403, 404, 413, 422];\n\n// The account is out of credit, or its key cannot make the request. Resending cannot fix\n// either, whatever status the error arrives with.\nconst unrecoverableErrorTypes = ['billing_error', 'authentication_error', 'permission_error'];\n\n// This client's label on the thinking blocks it keeps for replay; see `ProviderDataPart`.\nconst claudeSource = 'claude';\n\n/**\n * Anthropic rejects a `tool_use` whose input is not an object, with a 400 on every later call.\n * Another provider's call can hold anything else, such as OpenAI arguments that did not parse;\n * its result already told the model the call failed, so sending it as empty loses nothing.\n */\nfunction toClaudeInputObject(input: unknown): unknown {\n\treturn typeof input === 'object' && input !== null && !Array.isArray(input) ? input : {};\n}\n\n/** Thinking crosses back only as the signed block this client kept, never as its text. */\nfunction toClaudeBlocks(part: MessagePart): ContentBlockParam[] {\n\tswitch (part.type) {\n\t\tcase 'text':\n\t\t\t// The API rejects an empty text block.\n\t\t\treturn part.text ? [{ type: 'text', text: part.text }] : [];\n\t\tcase 'toolCall':\n\t\t\treturn [\n\t\t\t\t{ type: 'tool_use', id: part.id, name: part.name, input: toClaudeInputObject(part.input) },\n\t\t\t];\n\t\tcase 'toolResult':\n\t\t\treturn [\n\t\t\t\t{\n\t\t\t\t\ttype: 'tool_result',\n\t\t\t\t\ttool_use_id: part.callId,\n\t\t\t\t\tcontent: part.output,\n\t\t\t\t\tis_error: part.isError,\n\t\t\t\t},\n\t\t\t];\n\t\tcase 'providerData':\n\t\t\t// Another provider's block would be rejected; losing it only costs that provider's reasoning.\n\t\t\treturn part.source === claudeSource\n\t\t\t\t? [part.data as ThinkingBlockParam | RedactedThinkingBlockParam]\n\t\t\t\t: [];\n\t\tcase 'reasoning':\n\t\t\t// Unsigned text cannot be replayed as thinking. The API accepts a turn without it.\n\t\t\treturn [];\n\t}\n}\n\n/**\n * Tool results travel in a user turn, as the API requires. A message left with no blocks is\n * left out whole; the API joins the consecutive user turns that leaves.\n */\nfunction toClaudeInput(message: Message): MessageParam[] {\n\tconst content = message.content.flatMap(toClaudeBlocks);\n\tif (content.length === 0) {\n\t\treturn [];\n\t}\n\n\treturn [{ role: message.role === 'tool' ? 'user' : message.role, content }];\n}\n\n/**\n * Strict: the API constrains the model to the schema, so its input always parses and matches.\n * A schema outside what strict mode supports fails the request with a 400.\n */\nfunction toClaudeTool({ name, description, inputSchema }: ToolDefinition): ClaudeTool {\n\treturn { name, description, input_schema: inputSchema, strict: true };\n}\n\nfunction describeBlock(block: ContentBlock, logger: ILogger): MessagePart[] {\n\tswitch (block.type) {\n\t\tcase 'text':\n\t\t\treturn [{ type: 'text', text: block.text }];\n\t\tcase 'thinking': {\n\t\t\t// Kept signed for replay; the summarized text is only for narration.\n\t\t\tconst signed: ThinkingBlockParam = {\n\t\t\t\ttype: 'thinking',\n\t\t\t\tthinking: block.thinking,\n\t\t\t\tsignature: block.signature,\n\t\t\t};\n\t\t\tconst replay: MessagePart = { type: 'providerData', source: claudeSource, data: signed };\n\t\t\treturn block.thinking ? [{ type: 'reasoning', text: block.thinking }, replay] : [replay];\n\t\t}\n\t\tcase 'redacted_thinking': {\n\t\t\t// Encrypted by design: nothing to narrate, but it still goes back.\n\t\t\tconst redacted: RedactedThinkingBlockParam = { type: 'redacted_thinking', data: block.data };\n\t\t\treturn [{ type: 'providerData', source: claudeSource, data: redacted }];\n\t\t}\n\t\tcase 'tool_use':\n\t\t\treturn [{ type: 'toolCall', id: block.id, name: block.name, input: block.input }];\n\t\tdefault:\n\t\t\t// Server tools are never offered, so any other block is output this client does not expect.\n\t\t\ttreatErrors(\n\t\t\t\t() => {\n\t\t\t\t\tlogger.warn(block, 'Claude returned a content block the client does not map');\n\t\t\t\t},\n\t\t\t\tclassifyHostFailure,\n\t\t\t\t'Claude client logger failed while mapping a response',\n\t\t\t);\n\t\t\treturn [];\n\t}\n}\n\nfunction toStopReason(response: AnthropicMessage): StopReason {\n\tswitch (response.stop_reason) {\n\t\t// `tool_use` is a complete answer that asks for tools: running them is the agent's call.\n\t\tcase 'end_turn':\n\t\tcase 'tool_use':\n\t\t\treturn 'completed';\n\t\tcase 'max_tokens':\n\t\t\treturn 'truncated';\n\t\tcase 'refusal':\n\t\t\treturn 'refused';\n\t\tcase 'model_context_window_exceeded':\n\t\t\tthrow new MaxContextError('The context no longer fits the Claude model', {\n\t\t\t\tcause: 'The response stopped at the model context window.',\n\t\t\t});\n\t\tdefault:\n\t\t\t// `pause_turn` and `stop_sequence` need server tools or stop sequences, and this\n\t\t\t// client sends neither.\n\t\t\tthrow new UnrecoverableError('The Claude response stopped unexpectedly', {\n\t\t\t\tcause: `The response stopped with \"${response.stop_reason ?? 'unknown'}\".`,\n\t\t\t});\n\t}\n}\n\n/** Unlike OpenAI, Anthropic counts cache reads and writes apart from `input_tokens`. */\nfunction toTokens(usage: Usage | undefined, logger: ILogger): Tokens | undefined {\n\tif (!usage) {\n\t\t// Zeros would report a call that may have been billed as free.\n\t\ttreatErrors(\n\t\t\t() => {\n\t\t\t\tlogger.warn('Claude returned a response without usage; it cannot be accounted for');\n\t\t\t},\n\t\t\tclassifyHostFailure,\n\t\t\t'Claude client logger failed while reading usage',\n\t\t);\n\t\treturn undefined;\n\t}\n\n\treturn {\n\t\tinputTokens: usage.input_tokens,\n\t\treadCacheTokens: usage.cache_read_input_tokens ?? 0,\n\t\twrittenCacheTokens: usage.cache_creation_input_tokens ?? 0,\n\t\toutputTokens: usage.output_tokens,\n\t};\n}\n\nfunction toLLMResponse(response: AnthropicMessage, logger: ILogger): LLMResponse {\n\tlet usage: Tokens | undefined;\n\n\ttry {\n\t\t// Inside the try: it warns when the usage is missing, and that warning can throw too.\n\t\tusage = toTokens(response.usage, logger);\n\t\treturn {\n\t\t\tmessage: {\n\t\t\t\trole: 'assistant',\n\t\t\t\tcontent: response.content.flatMap(block => describeBlock(block, logger)),\n\t\t\t},\n\t\t\tusage,\n\t\t\tstopReason: toStopReason(response),\n\t\t};\n\t} catch (error) {\n\t\t// The call was billed whether the response failed or a logger did while mapping it, and\n\t\t// an answer without usage leaves the run's count unknown.\n\t\tthrow withSpentTokens(error, usage, !response.usage);\n\t}\n}\n\n/**\n * The SDK resolves credentials lazily, so a missing key surfaces on the first call, as a plain\n * `Error` the SDK gives no class of its own. An `AnthropicError` that is not an `APIError` never\n * carries an answer from the API: on this client's path it is the SDK refusing to make the\n * call, such as an unresolved profile or a request too large to send without streaming.\n */\nfunction isLocalSDKFailure(error: unknown): boolean {\n\treturn (\n\t\t(error instanceof AnthropicError && !(error instanceof APIError)) ||\n\t\t(error instanceof Error && error.message.startsWith('Could not resolve authentication method'))\n\t);\n}\n\nfunction classifyClaudeFailure(error: unknown): Error {\n\tif (isLocalSDKFailure(error)) {\n\t\t// The API never answered, and calling again would fail the same way.\n\t\treturn new UnrecoverableError('The Claude client cannot send the request', {\n\t\t\tcause: describeFailure(error),\n\t\t});\n\t}\n\tif (error instanceof APIError) {\n\t\t// Anthropic gives an overlong prompt no code of its own, only a 400 with this message.\n\t\tif (error.status === 400 && error.message.includes('prompt is too long')) {\n\t\t\treturn new MaxContextError('The context no longer fits the Claude model', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t\tif (\n\t\t\tunrecoverableErrorTypes.some(type => type === error.type) ||\n\t\t\tunrecoverableStatuses.some(status => status === error.status)\n\t\t) {\n\t\t\treturn new UnrecoverableError('Claude rejected the request', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t}\n\n\t// A rate limit, an overload (529), a 5xx or a network failure: a later attempt can succeed.\n\treturn classifyProviderFailure(error, 'The Claude request failed');\n}\n\n/**\n * A stateless `LLMClient` on the Anthropic Messages API, which keeps nothing between calls:\n * every call carries the whole context, so the conversation is MiKode's, not the provider's.\n */\nexport class ClaudeLLMClient implements LLMClient {\n\tprivate readonly client: Anthropic;\n\tprivate readonly model: (typeof claudeLLMModelIds)[ClaudeLLMModel];\n\tprivate readonly reasoningEffort: ClaudeLLMReasoningEffort;\n\tprivate readonly systemPrompt: string;\n\tprivate readonly logger: ILogger;\n\n\t// `model` and `reasoningEffort` are untyped on purpose: this adapter is the one that knows\n\t// what Claude supports.\n\tconstructor({\n\t\tmodel,\n\t\treasoningEffort = 'high',\n\t\tsystemPrompt,\n\t\tlogger,\n\t}: {\n\t\tmodel: string;\n\t\treasoningEffort?: 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: ${claudeLLMModels.join(', ')}`,\n\t\t\t);\n\t\t}\n\t\tif (!isOneOf(claudeLLMReasoningEfforts, reasoningEffort)) {\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`\"${reasoningEffort}\" is not a Claude API reasoning effort; expected one of: ${claudeLLMReasoningEfforts.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 = claudeLLMModelIds[model];\n\t\tthis.reasoningEffort = reasoningEffort;\n\t\tthis.systemPrompt = systemPrompt;\n\t\tthis.logger = logger;\n\t}\n\n\tasync send(\n\t\t{ context, tools }: { context: Message[]; tools: ToolDefinition[] },\n\t\tsignal: AbortSignal,\n\t): Promise<LLMResponse> {\n\t\tlet response: AnthropicMessage;\n\n\t\ttry {\n\t\t\tresponse = await this.client.messages.create(\n\t\t\t\t{\n\t\t\t\t\tmodel: this.model,\n\t\t\t\t\tmax_tokens: maxOutputTokens,\n\t\t\t\t\tsystem: this.systemPrompt,\n\t\t\t\t\tmessages: context.flatMap(toClaudeInput),\n\t\t\t\t\t// Left out when empty, so a request without tools stays as it always was.\n\t\t\t\t\t...(tools.length > 0 && { tools: tools.map(toClaudeTool) }),\n\t\t\t\t\tthinking: { type: 'adaptive', display: 'summarized' },\n\t\t\t\t\toutput_config: { effort: this.reasoningEffort },\n\t\t\t\t\t// Anthropic caches only on request. This marks the last block, so the next call\n\t\t\t\t\t// reads everything before it from the cache instead of paying for it again.\n\t\t\t\t\tcache_control: { type: 'ephemeral' },\n\t\t\t\t},\n\t\t\t\t{ signal },\n\t\t\t);\n\t\t} catch (error) {\n\t\t\t// The SDK's abort error is named 'Error', so it would pass for a provider failure.\n\t\t\tsignal.throwIfAborted();\n\t\t\tthrow classifyClaudeFailure(error);\n\t\t}\n\n\t\treturn treatErrors(\n\t\t\t() => toLLMResponse(response, this.logger),\n\t\t\tclassifyLocalFailure,\n\t\t\t'Claude response mapping failed',\n\t\t);\n\t}\n}\n"]}
+{"version":3,"file":"claudeLLMClient.js","sourceRoot":"","sources":["../../../src/llm/infrastructure/claudeLLMClient.ts"],"names":[],"mappings":"AAAA,OAAO,SAAS,EAAE,EAAE,cAAc,EAAE,QAAQ,EAAE,MAAM,mBAAmB,CAAC;AAaxE,OAAO,EACN,uBAAuB,EACvB,kBAAkB,EAClB,eAAe,GACf,MAAM,2BAA2B,CAAC;AACnC,OAAO,EACN,mBAAmB,EACnB,oBAAoB,EACpB,uBAAuB,EACvB,eAAe,EACf,WAAW,GACX,MAAM,oCAAoC,CAAC;AAC5C,OAAO,EAAE,OAAO,EAAE,MAAM,4BAA4B,CAAC;AAGrD,OAAO,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AAGtD,OAAO,EAAmB,QAAQ,EAAuB,MAAM,mBAAmB,CAAC;AAEnF,gGAAgG;AAChG,8FAA8F;AAC9F,kGAAkG;AAClG,2FAA2F;AAC3F,MAAM,CAAC,MAAM,eAAe,GAAG,CAAC,MAAM,EAAE,OAAO,EAAE,QAAQ,EAAE,OAAO,CAAU,CAAC;AAG7E,MAAM,iBAAiB,GAAG;IACzB,IAAI,EAAE,iBAAiB;IACvB,KAAK,EAAE,kBAAkB;IACzB,MAAM,EAAE,iBAAiB;IACzB,KAAK,EAAE,kBAAkB;CACyB,CAAC;AAEpD,MAAM,CAAC,MAAM,yBAAyB,GAAG;IACxC,KAAK;IACL,QAAQ;IACR,MAAM;IACN,OAAO;IACP,KAAK;CAC6D,CAAC;AAGpE,gGAAgG;AAChG,yEAAyE;AACzE,MAAM,eAAe,GAAG,MAAM,CAAC;AAE/B,kFAAkF;AAClF,MAAM,qBAAqB,GAAG,CAAC,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,CAAC,CAAC;AAE7D,yFAAyF;AACzF,kDAAkD;AAClD,MAAM,uBAAuB,GAAG,CAAC,eAAe,EAAE,sBAAsB,EAAE,kBAAkB,CAAC,CAAC;AAE9F,0FAA0F;AAC1F,MAAM,YAAY,GAAG,QAAQ,CAAC;AAE9B;;;;GAIG;AACH,SAAS,mBAAmB,CAAC,KAAc;IAC1C,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC;AAC1F,CAAC;AAED,0FAA0F;AAC1F,SAAS,cAAc,CAAC,IAAiB;IACxC,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;QACnB,KAAK,MAAM;YACV,uCAAuC;YACvC,OAAO,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;QAC7D,KAAK,UAAU;YACd,OAAO;gBACN,EAAE,IAAI,EAAE,UAAU,EAAE,EAAE,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,KAAK,EAAE,mBAAmB,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE;aAC1F,CAAC;QACH,KAAK,YAAY;YAChB,OAAO;gBACN;oBACC,IAAI,EAAE,aAAa;oBACnB,WAAW,EAAE,IAAI,CAAC,MAAM;oBACxB,OAAO,EAAE,IAAI,CAAC,MAAM;oBACpB,QAAQ,EAAE,IAAI,CAAC,OAAO;iBACtB;aACD,CAAC;QACH,KAAK,cAAc;YAClB,8FAA8F;YAC9F,OAAO,IAAI,CAAC,MAAM,KAAK,YAAY;gBAClC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAuD,CAAC;gBAChE,CAAC,CAAC,EAAE,CAAC;QACP,KAAK,WAAW;YACf,mFAAmF;YACnF,OAAO,EAAE,CAAC;IACZ,CAAC;AACF,CAAC;AAED;;;GAGG;AACH,SAAS,aAAa,CAAC,OAAgB;IACtC,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC,cAAc,CAAC,CAAC;IACxD,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QAC1B,OAAO,EAAE,CAAC;IACX,CAAC;IAED,OAAO,CAAC,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC;AAC7E,CAAC;AAED,qEAAqE;AACrE,MAAM,cAAc,GAAG,6BAA6B,CAAC;AAErD;;;;GAIG;AACH,SAAS,YAAY,CAAC,UAA0B;IAC/C,MAAM,MAAM,GAAG,QAAQ,CAAC,UAAU,CAAC,CAAC;IACpC,IAAI,MAAM,KAAK,YAAY,EAAE,CAAC;QAC7B,IAAI,UAAU,CAAC,IAAI,KAAK,cAAc,EAAE,CAAC;YACxC,MAAM,IAAI,uBAAuB,CAChC,uCAAuC,cAAc,WAAW,UAAU,CAAC,IAAI,GAAG,CAClF,CAAC;QACH,CAAC;QACD,OAAO,EAAE,IAAI,EAAE,sBAAsB,EAAE,IAAI,EAAE,cAAc,EAAE,CAAC;IAC/D,CAAC;IACD,IAAI,MAAM,KAAK,SAAS,EAAE,CAAC;QAC1B,MAAM,IAAI,uBAAuB,CAAC,8BAA8B,MAAM,GAAG,CAAC,CAAC;IAC5E,CAAC;IACD,MAAM,EAAE,IAAI,EAAE,WAAW,EAAE,WAAW,EAAE,GAAG,UAAU,CAAC;IACtD,OAAO,EAAE,IAAI,EAAE,WAAW,EAAE,YAAY,EAAE,WAAW,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC;AACvE,CAAC;AAED,SAAS,aAAa,CAAC,KAAmB,EAAE,MAAe;IAC1D,QAAQ,KAAK,CAAC,IAAI,EAAE,CAAC;QACpB,KAAK,MAAM;YACV,OAAO,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC;QAC7C,KAAK,UAAU,CAAC,CAAC,CAAC;YACjB,qEAAqE;YACrE,MAAM,MAAM,GAAuB;gBAClC,IAAI,EAAE,UAAU;gBAChB,QAAQ,EAAE,KAAK,CAAC,QAAQ;gBACxB,SAAS,EAAE,KAAK,CAAC,SAAS;aAC1B,CAAC;YACF,MAAM,MAAM,GAAgB,EAAE,IAAI,EAAE,cAAc,EAAE,MAAM,EAAE,YAAY,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC;YACzF,OAAO,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,KAAK,CAAC,QAAQ,EAAE,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;QAC1F,CAAC;QACD,KAAK,mBAAmB,CAAC,CAAC,CAAC;YAC1B,mEAAmE;YACnE,MAAM,QAAQ,GAA+B,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,CAAC;YAC7F,OAAO,CAAC,EAAE,IAAI,EAAE,cAAc,EAAE,MAAM,EAAE,YAAY,EAAE,IAAI,EAAE,QAAQ,EAAE,CAAC,CAAC;QACzE,CAAC;QACD,KAAK,UAAU;YACd,OAAO,CAAC,EAAE,IAAI,EAAE,UAAU,EAAE,EAAE,EAAE,KAAK,CAAC,EAAE,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,CAAC,CAAC;QACnF;YACC,4FAA4F;YAC5F,WAAW,CACV,GAAG,EAAE;gBACJ,MAAM,CAAC,IAAI,CAAC,KAAK,EAAE,yDAAyD,CAAC,CAAC;YAC/E,CAAC,EACD,mBAAmB,EACnB,sDAAsD,CACtD,CAAC;YACF,OAAO,EAAE,CAAC;IACZ,CAAC;AACF,CAAC;AAED,SAAS,YAAY,CAAC,QAA0B;IAC/C,QAAQ,QAAQ,CAAC,WAAW,EAAE,CAAC;QAC9B,yFAAyF;QACzF,KAAK,UAAU,CAAC;QAChB,KAAK,UAAU;YACd,OAAO,WAAW,CAAC;QACpB,KAAK,YAAY;YAChB,OAAO,WAAW,CAAC;QACpB,KAAK,SAAS;YACb,OAAO,SAAS,CAAC;QAClB,KAAK,+BAA+B;YACnC,MAAM,IAAI,eAAe,CAAC,6CAA6C,EAAE;gBACxE,KAAK,EAAE,mDAAmD;aAC1D,CAAC,CAAC;QACJ;YACC,iFAAiF;YACjF,wBAAwB;YACxB,MAAM,IAAI,kBAAkB,CAAC,0CAA0C,EAAE;gBACxE,KAAK,EAAE,8BAA8B,QAAQ,CAAC,WAAW,IAAI,SAAS,IAAI;aAC1E,CAAC,CAAC;IACL,CAAC;AACF,CAAC;AAED,wFAAwF;AACxF,SAAS,QAAQ,CAAC,KAAwB,EAAE,MAAe;IAC1D,IAAI,CAAC,KAAK,EAAE,CAAC;QACZ,+DAA+D;QAC/D,WAAW,CACV,GAAG,EAAE;YACJ,MAAM,CAAC,IAAI,CAAC,sEAAsE,CAAC,CAAC;QACrF,CAAC,EACD,mBAAmB,EACnB,iDAAiD,CACjD,CAAC;QACF,OAAO,SAAS,CAAC;IAClB,CAAC;IAED,OAAO;QACN,WAAW,EAAE,KAAK,CAAC,YAAY;QAC/B,eAAe,EAAE,KAAK,CAAC,uBAAuB,IAAI,CAAC;QACnD,kBAAkB,EAAE,KAAK,CAAC,2BAA2B,IAAI,CAAC;QAC1D,YAAY,EAAE,KAAK,CAAC,aAAa;KACjC,CAAC;AACH,CAAC;AAED,SAAS,aAAa,CAAC,QAA0B,EAAE,MAAe;IACjE,IAAI,KAAyB,CAAC;IAE9B,IAAI,CAAC;QACJ,sFAAsF;QACtF,KAAK,GAAG,QAAQ,CAAC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;QACzC,OAAO;YACN,OAAO,EAAE;gBACR,IAAI,EAAE,WAAW;gBACjB,OAAO,EAAE,QAAQ,CAAC,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC,aAAa,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;aACxE;YACD,KAAK;YACL,UAAU,EAAE,YAAY,CAAC,QAAQ,CAAC;SAClC,CAAC;IACH,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,wFAAwF;QACxF,0DAA0D;QAC1D,MAAM,eAAe,CAAC,KAAK,EAAE,KAAK,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;IACtD,CAAC;AACF,CAAC;AAED;;;;;GAKG;AACH,SAAS,iBAAiB,CAAC,KAAc;IACxC,OAAO,CACN,CAAC,KAAK,YAAY,cAAc,IAAI,CAAC,CAAC,KAAK,YAAY,QAAQ,CAAC,CAAC;QACjE,CAAC,KAAK,YAAY,KAAK,IAAI,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,yCAAyC,CAAC,CAAC,CAC/F,CAAC;AACH,CAAC;AAED,SAAS,qBAAqB,CAAC,KAAc;IAC5C,IAAI,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;QAC9B,qEAAqE;QACrE,OAAO,IAAI,kBAAkB,CAAC,2CAA2C,EAAE;YAC1E,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;SAC7B,CAAC,CAAC;IACJ,CAAC;IACD,IAAI,KAAK,YAAY,QAAQ,EAAE,CAAC;QAC/B,uFAAuF;QACvF,IAAI,KAAK,CAAC,MAAM,KAAK,GAAG,IAAI,KAAK,CAAC,OAAO,CAAC,QAAQ,CAAC,oBAAoB,CAAC,EAAE,CAAC;YAC1E,OAAO,IAAI,eAAe,CAAC,6CAA6C,EAAE;gBACzE,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;QACD,IACC,uBAAuB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,KAAK,KAAK,CAAC,IAAI,CAAC;YACzD,qBAAqB,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,CAAC,EAC5D,CAAC;YACF,OAAO,IAAI,kBAAkB,CAAC,6BAA6B,EAAE;gBAC5D,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;IACF,CAAC;IAED,4FAA4F;IAC5F,OAAO,uBAAuB,CAAC,KAAK,EAAE,2BAA2B,CAAC,CAAC;AACpE,CAAC;AAED;;;GAGG;AACH,MAAM,OAAO,eAAe;IACV,MAAM,CAAY;IAClB,KAAK,CAA6C;IACnE,kDAAkD;IACjC,eAAe,CAAuC;IACtD,YAAY,CAAS;IACrB,MAAM,CAAU;IAEjC,2FAA2F;IAC3F,wBAAwB;IACxB,YAAY,EACX,KAAK,EACL,eAAe,EACf,YAAY,EACZ,MAAM,GAMN;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;QACD,IAAI,KAAK,KAAK,OAAO,IAAI,eAAe,KAAK,SAAS,EAAE,CAAC;YACxD,MAAM,IAAI,uBAAuB,CAChC,kEAAkE,CAClE,CAAC;QACH,CAAC;QACD,MAAM,MAAM,GAAG,KAAK,KAAK,OAAO,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,eAAe,IAAI,MAAM,CAAC,CAAC;QAC3E,IAAI,MAAM,KAAK,SAAS,IAAI,CAAC,OAAO,CAAC,yBAAyB,EAAE,MAAM,CAAC,EAAE,CAAC;YACzE,MAAM,IAAI,uBAAuB,CAChC,IAAI,MAAM,4DAA4D,yBAAyB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAC5G,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,iBAAiB,CAAC,KAAK,CAAC,CAAC;QACtC,IAAI,CAAC,eAAe,GAAG,MAAM,CAAC;QAC9B,IAAI,CAAC,YAAY,GAAG,YAAY,CAAC;QACjC,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IACtB,CAAC;IAED,cAAc,CAAC,IAAgB;QAC9B,OAAO,IAAI,KAAK,YAAY,CAAC;IAC9B,CAAC;IAED,KAAK,CAAC,IAAI,CACT,EAAE,OAAO,EAAE,KAAK,EAAmD,EACnE,MAAmB;QAEnB,IAAI,QAA0B,CAAC;QAE/B,IAAI,CAAC;YACJ,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAC3C;gBACC,KAAK,EAAE,IAAI,CAAC,KAAK;gBACjB,UAAU,EAAE,eAAe;gBAC3B,MAAM,EAAE,IAAI,CAAC,YAAY;gBACzB,QAAQ,EAAE,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC;gBACxC,0EAA0E;gBAC1E,GAAG,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,EAAE,CAAC;gBAC3D,uEAAuE;gBACvE,GAAG,CAAC,IAAI,CAAC,eAAe,KAAK,SAAS,IAAI;oBACzC,QAAQ,EAAE,EAAE,IAAI,EAAE,UAAU,EAAE,OAAO,EAAE,YAAY,EAAE;oBACrD,aAAa,EAAE,EAAE,MAAM,EAAE,IAAI,CAAC,eAAe,EAAE;iBAC/C,CAAC;gBACF,gFAAgF;gBAChF,4EAA4E;gBAC5E,aAAa,EAAE,EAAE,IAAI,EAAE,WAAW,EAAE;aACpC,EACD,EAAE,MAAM,EAAE,CACV,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,mFAAmF;YACnF,MAAM,CAAC,cAAc,EAAE,CAAC;YACxB,MAAM,qBAAqB,CAAC,KAAK,CAAC,CAAC;QACpC,CAAC;QAED,OAAO,WAAW,CACjB,GAAG,EAAE,CAAC,aAAa,CAAC,QAAQ,EAAE,IAAI,CAAC,MAAM,CAAC,EAC1C,oBAAoB,EACpB,gCAAgC,CAChC,CAAC;IACH,CAAC;CACD","sourcesContent":["import Anthropic, { AnthropicError, APIError } from '@anthropic-ai/sdk';\nimport type {\n\tContentBlock,\n\tContentBlockParam,\n\tMessage as AnthropicMessage,\n\tMessageParam,\n\tOutputConfig,\n\tRedactedThinkingBlockParam,\n\tThinkingBlockParam,\n\tTool as ClaudeTool,\n\tToolTextEditor20250728,\n\tUsage,\n} from '@anthropic-ai/sdk/resources/messages';\nimport {\n\tInvalidAgentConfigError,\n\tUnrecoverableError,\n\twithSpentTokens,\n} from '#src/shared/domain/errors';\nimport {\n\tclassifyHostFailure,\n\tclassifyLocalFailure,\n\tclassifyProviderFailure,\n\tdescribeFailure,\n\ttreatErrors,\n} from '#src/shared/domain/providerFailure';\nimport { isOneOf } from '#src/shared/domain/isOneOf';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport type { Tokens } from '#src/shared/domain/tokens';\nimport { MaxContextError } from '../domain/errors.ts';\nimport type { LLMClient, LLMResponse, StopReason } from '../domain/llm.ts';\nimport type { Message, MessagePart } from '../domain/message.ts';\nimport { type NativeTool, nativeOf, type ToolDefinition } from '../domain/tool.ts';\n\n// The Messages API takes full model IDs. Callers name a model by the Agent SDK's alias, so both\n// paths share one vocabulary; this table pins the ID each alias means here, and moves by hand\n// when Anthropic releases a newer model. Haiku runs without thinking: it has no adaptive thinking\n// and takes no reasoning effort, and it is here for short, simple calls such as a summary.\nexport const claudeLLMModels = ['opus', 'fable', 'sonnet', 'haiku'] as const;\nexport type ClaudeLLMModel = (typeof claudeLLMModels)[number];\n\nconst claudeLLMModelIds = {\n\topus: 'claude-opus-5-5',\n\tfable: 'claude-fable-5-1',\n\tsonnet: 'claude-sonnet-5',\n\thaiku: 'claude-haiku-4-5',\n} as const satisfies Record<ClaudeLLMModel, string>;\n\nexport const claudeLLMReasoningEfforts = [\n\t'low',\n\t'medium',\n\t'high',\n\t'xhigh',\n\t'max',\n] as const satisfies readonly NonNullable<OutputConfig['effort']>[];\nexport type ClaudeLLMReasoningEffort = (typeof claudeLLMReasoningEfforts)[number];\n\n// Adaptive thinking spends from this budget too. Above roughly 21,000 the SDK refuses a request\n// that is not streamed, because it could outlast its ten-minute timeout.\nconst maxOutputTokens = 16_000;\n\n// The request itself is wrong or not allowed, so sending it again cannot succeed.\nconst unrecoverableStatuses = [400, 401, 403, 404, 413, 422];\n\n// The account is out of credit, or its key cannot make the request. Resending cannot fix\n// either, whatever status the error arrives with.\nconst unrecoverableErrorTypes = ['billing_error', 'authentication_error', 'permission_error'];\n\n// This client's label on the thinking blocks it keeps for replay; see `ProviderDataPart`.\nconst claudeSource = 'claude';\n\n/**\n * Anthropic rejects a `tool_use` whose input is not an object, with a 400 on every later call.\n * Another provider's call can hold anything else, such as OpenAI arguments that did not parse;\n * its result already told the model the call failed, so sending it as empty loses nothing.\n */\nfunction toClaudeInputObject(input: unknown): unknown {\n\treturn typeof input === 'object' && input !== null && !Array.isArray(input) ? input : {};\n}\n\n/** Thinking crosses back only as the signed block this client kept, never as its text. */\nfunction toClaudeBlocks(part: MessagePart): ContentBlockParam[] {\n\tswitch (part.type) {\n\t\tcase 'text':\n\t\t\t// The API rejects an empty text block.\n\t\t\treturn part.text ? [{ type: 'text', text: part.text }] : [];\n\t\tcase 'toolCall':\n\t\t\treturn [\n\t\t\t\t{ type: 'tool_use', id: part.id, name: part.name, input: toClaudeInputObject(part.input) },\n\t\t\t];\n\t\tcase 'toolResult':\n\t\t\treturn [\n\t\t\t\t{\n\t\t\t\t\ttype: 'tool_result',\n\t\t\t\t\ttool_use_id: part.callId,\n\t\t\t\t\tcontent: part.output,\n\t\t\t\t\tis_error: part.isError,\n\t\t\t\t},\n\t\t\t];\n\t\tcase 'providerData':\n\t\t\t// Another provider's block would be rejected; losing it only costs that provider's reasoning.\n\t\t\treturn part.source === claudeSource\n\t\t\t\t? [part.data as ThinkingBlockParam | RedactedThinkingBlockParam]\n\t\t\t\t: [];\n\t\tcase 'reasoning':\n\t\t\t// Unsigned text cannot be replayed as thinking. The API accepts a turn without it.\n\t\t\treturn [];\n\t}\n}\n\n/**\n * Tool results travel in a user turn, as the API requires. A message left with no blocks is\n * left out whole; the API joins the consecutive user turns that leaves.\n */\nfunction toClaudeInput(message: Message): MessageParam[] {\n\tconst content = message.content.flatMap(toClaudeBlocks);\n\tif (content.length === 0) {\n\t\treturn [];\n\t}\n\n\treturn [{ role: message.role === 'tool' ? 'user' : message.role, content }];\n}\n\n// Anthropic fixes the text editor's name: its calls arrive under it.\nconst textEditorName = 'str_replace_based_edit_tool';\n\n/**\n * Strict: the API constrains the model to the schema, so its input always parses and matches.\n * A schema outside what strict mode supports fails the request with a 400. The native text\n * editor has Anthropic's own schema, and its calls and results are ordinary tool blocks.\n */\nfunction toClaudeTool(definition: ToolDefinition): ClaudeTool | ToolTextEditor20250728 {\n\tconst native = nativeOf(definition);\n\tif (native === 'textEditor') {\n\t\tif (definition.name !== textEditorName) {\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`Claude's text editor must be named \"${textEditorName}\", not \"${definition.name}\"`,\n\t\t\t);\n\t\t}\n\t\treturn { type: 'text_editor_20250728', name: textEditorName };\n\t}\n\tif (native !== undefined) {\n\t\tthrow new InvalidAgentConfigError(`Claude has no native tool \"${native}\"`);\n\t}\n\tconst { name, description, inputSchema } = definition;\n\treturn { name, description, input_schema: inputSchema, strict: true };\n}\n\nfunction describeBlock(block: ContentBlock, logger: ILogger): MessagePart[] {\n\tswitch (block.type) {\n\t\tcase 'text':\n\t\t\treturn [{ type: 'text', text: block.text }];\n\t\tcase 'thinking': {\n\t\t\t// Kept signed for replay; the summarized text is only for narration.\n\t\t\tconst signed: ThinkingBlockParam = {\n\t\t\t\ttype: 'thinking',\n\t\t\t\tthinking: block.thinking,\n\t\t\t\tsignature: block.signature,\n\t\t\t};\n\t\t\tconst replay: MessagePart = { type: 'providerData', source: claudeSource, data: signed };\n\t\t\treturn block.thinking ? [{ type: 'reasoning', text: block.thinking }, replay] : [replay];\n\t\t}\n\t\tcase 'redacted_thinking': {\n\t\t\t// Encrypted by design: nothing to narrate, but it still goes back.\n\t\t\tconst redacted: RedactedThinkingBlockParam = { type: 'redacted_thinking', data: block.data };\n\t\t\treturn [{ type: 'providerData', source: claudeSource, data: redacted }];\n\t\t}\n\t\tcase 'tool_use':\n\t\t\treturn [{ type: 'toolCall', id: block.id, name: block.name, input: block.input }];\n\t\tdefault:\n\t\t\t// Server tools are never offered, so any other block is output this client does not expect.\n\t\t\ttreatErrors(\n\t\t\t\t() => {\n\t\t\t\t\tlogger.warn(block, 'Claude returned a content block the client does not map');\n\t\t\t\t},\n\t\t\t\tclassifyHostFailure,\n\t\t\t\t'Claude client logger failed while mapping a response',\n\t\t\t);\n\t\t\treturn [];\n\t}\n}\n\nfunction toStopReason(response: AnthropicMessage): StopReason {\n\tswitch (response.stop_reason) {\n\t\t// `tool_use` is a complete answer that asks for tools: running them is the agent's call.\n\t\tcase 'end_turn':\n\t\tcase 'tool_use':\n\t\t\treturn 'completed';\n\t\tcase 'max_tokens':\n\t\t\treturn 'truncated';\n\t\tcase 'refusal':\n\t\t\treturn 'refused';\n\t\tcase 'model_context_window_exceeded':\n\t\t\tthrow new MaxContextError('The context no longer fits the Claude model', {\n\t\t\t\tcause: 'The response stopped at the model context window.',\n\t\t\t});\n\t\tdefault:\n\t\t\t// `pause_turn` and `stop_sequence` need server tools or stop sequences, and this\n\t\t\t// client sends neither.\n\t\t\tthrow new UnrecoverableError('The Claude response stopped unexpectedly', {\n\t\t\t\tcause: `The response stopped with \"${response.stop_reason ?? 'unknown'}\".`,\n\t\t\t});\n\t}\n}\n\n/** Unlike OpenAI, Anthropic counts cache reads and writes apart from `input_tokens`. */\nfunction toTokens(usage: Usage | undefined, logger: ILogger): Tokens | undefined {\n\tif (!usage) {\n\t\t// Zeros would report a call that may have been billed as free.\n\t\ttreatErrors(\n\t\t\t() => {\n\t\t\t\tlogger.warn('Claude returned a response without usage; it cannot be accounted for');\n\t\t\t},\n\t\t\tclassifyHostFailure,\n\t\t\t'Claude client logger failed while reading usage',\n\t\t);\n\t\treturn undefined;\n\t}\n\n\treturn {\n\t\tinputTokens: usage.input_tokens,\n\t\treadCacheTokens: usage.cache_read_input_tokens ?? 0,\n\t\twrittenCacheTokens: usage.cache_creation_input_tokens ?? 0,\n\t\toutputTokens: usage.output_tokens,\n\t};\n}\n\nfunction toLLMResponse(response: AnthropicMessage, logger: ILogger): LLMResponse {\n\tlet usage: Tokens | undefined;\n\n\ttry {\n\t\t// Inside the try: it warns when the usage is missing, and that warning can throw too.\n\t\tusage = toTokens(response.usage, logger);\n\t\treturn {\n\t\t\tmessage: {\n\t\t\t\trole: 'assistant',\n\t\t\t\tcontent: response.content.flatMap(block => describeBlock(block, logger)),\n\t\t\t},\n\t\t\tusage,\n\t\t\tstopReason: toStopReason(response),\n\t\t};\n\t} catch (error) {\n\t\t// The call was billed whether the response failed or a logger did while mapping it, and\n\t\t// an answer without usage leaves the run's count unknown.\n\t\tthrow withSpentTokens(error, usage, !response.usage);\n\t}\n}\n\n/**\n * The SDK resolves credentials lazily, so a missing key surfaces on the first call, as a plain\n * `Error` the SDK gives no class of its own. An `AnthropicError` that is not an `APIError` never\n * carries an answer from the API: on this client's path it is the SDK refusing to make the\n * call, such as an unresolved profile or a request too large to send without streaming.\n */\nfunction isLocalSDKFailure(error: unknown): boolean {\n\treturn (\n\t\t(error instanceof AnthropicError && !(error instanceof APIError)) ||\n\t\t(error instanceof Error && error.message.startsWith('Could not resolve authentication method'))\n\t);\n}\n\nfunction classifyClaudeFailure(error: unknown): Error {\n\tif (isLocalSDKFailure(error)) {\n\t\t// The API never answered, and calling again would fail the same way.\n\t\treturn new UnrecoverableError('The Claude client cannot send the request', {\n\t\t\tcause: describeFailure(error),\n\t\t});\n\t}\n\tif (error instanceof APIError) {\n\t\t// Anthropic gives an overlong prompt no code of its own, only a 400 with this message.\n\t\tif (error.status === 400 && error.message.includes('prompt is too long')) {\n\t\t\treturn new MaxContextError('The context no longer fits the Claude model', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t\tif (\n\t\t\tunrecoverableErrorTypes.some(type => type === error.type) ||\n\t\t\tunrecoverableStatuses.some(status => status === error.status)\n\t\t) {\n\t\t\treturn new UnrecoverableError('Claude rejected the request', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t}\n\n\t// A rate limit, an overload (529), a 5xx or a network failure: a later attempt can succeed.\n\treturn classifyProviderFailure(error, 'The Claude request failed');\n}\n\n/**\n * A stateless `LLMClient` on the Anthropic Messages API, which keeps nothing between calls:\n * every call carries the whole context, so the conversation is MiKode's, not the provider's.\n */\nexport class ClaudeLLMClient implements LLMClient {\n\tprivate readonly client: Anthropic;\n\tprivate readonly model: (typeof claudeLLMModelIds)[ClaudeLLMModel];\n\t// Missing for Haiku, which runs without thinking.\n\tprivate readonly reasoningEffort: ClaudeLLMReasoningEffort | undefined;\n\tprivate readonly systemPrompt: string;\n\tprivate readonly logger: ILogger;\n\n\t// `model` and `reasoningEffort` are untyped on purpose: this adapter is the one that knows\n\t// what Claude supports.\n\tconstructor({\n\t\tmodel,\n\t\treasoningEffort,\n\t\tsystemPrompt,\n\t\tlogger,\n\t}: {\n\t\tmodel: string;\n\t\treasoningEffort?: 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: ${claudeLLMModels.join(', ')}`,\n\t\t\t);\n\t\t}\n\t\tif (model === 'haiku' && reasoningEffort !== undefined) {\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t'Claude Haiku takes no reasoning effort: it runs without thinking',\n\t\t\t);\n\t\t}\n\t\tconst effort = model === 'haiku' ? undefined : (reasoningEffort ?? 'high');\n\t\tif (effort !== undefined && !isOneOf(claudeLLMReasoningEfforts, effort)) {\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`\"${effort}\" is not a Claude API reasoning effort; expected one of: ${claudeLLMReasoningEfforts.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 = claudeLLMModelIds[model];\n\t\tthis.reasoningEffort = effort;\n\t\tthis.systemPrompt = systemPrompt;\n\t\tthis.logger = logger;\n\t}\n\n\tsupportsNative(tool: NativeTool): boolean {\n\t\treturn tool === 'textEditor';\n\t}\n\n\tasync send(\n\t\t{ context, tools }: { context: Message[]; tools: ToolDefinition[] },\n\t\tsignal: AbortSignal,\n\t): Promise<LLMResponse> {\n\t\tlet response: AnthropicMessage;\n\n\t\ttry {\n\t\t\tresponse = await this.client.messages.create(\n\t\t\t\t{\n\t\t\t\t\tmodel: this.model,\n\t\t\t\t\tmax_tokens: maxOutputTokens,\n\t\t\t\t\tsystem: this.systemPrompt,\n\t\t\t\t\tmessages: context.flatMap(toClaudeInput),\n\t\t\t\t\t// Left out when empty, so a request without tools stays as it always was.\n\t\t\t\t\t...(tools.length > 0 && { tools: tools.map(toClaudeTool) }),\n\t\t\t\t\t// Haiku has no adaptive thinking, and an effort is refused with a 400.\n\t\t\t\t\t...(this.reasoningEffort !== undefined && {\n\t\t\t\t\t\tthinking: { type: 'adaptive', display: 'summarized' },\n\t\t\t\t\t\toutput_config: { effort: this.reasoningEffort },\n\t\t\t\t\t}),\n\t\t\t\t\t// Anthropic caches only on request. This marks the last block, so the next call\n\t\t\t\t\t// reads everything before it from the cache instead of paying for it again.\n\t\t\t\t\tcache_control: { type: 'ephemeral' },\n\t\t\t\t},\n\t\t\t\t{ signal },\n\t\t\t);\n\t\t} catch (error) {\n\t\t\t// The SDK's abort error is named 'Error', so it would pass for a provider failure.\n\t\t\tsignal.throwIfAborted();\n\t\t\tthrow classifyClaudeFailure(error);\n\t\t}\n\n\t\treturn treatErrors(\n\t\t\t() => toLLMResponse(response, this.logger),\n\t\t\tclassifyLocalFailure,\n\t\t\t'Claude response mapping failed',\n\t\t);\n\t}\n}\n"]}

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

@@ -1,7 +1,7 @@
 import type { ILogger } from '#src/shared/domain/logger';
 import type { LLMClient, LLMResponse } from '../domain/llm.ts';
 import type { Message } from '../domain/message.ts';
-import type { ToolDefinition } from '../domain/tool.ts';
+import { type NativeTool, type ToolDefinition } from '../domain/tool.ts';
 export declare const openAIModels: readonly ["gpt-6-astra", "gpt-5.6-sol", "gpt-5.6-luna", "gpt-5.6-terra"];
 export type OpenAIModel = (typeof openAIModels)[number];
 export declare const openAIReasoningEfforts: readonly ["low", "medium", "high", "xhigh", "max"];
@@ -23,6 +23,7 @@ export declare class OpenAILLMClient implements LLMClient {
         systemPrompt: string;
         logger: ILogger;
     });
+    supportsNative(tool: NativeTool): boolean;
     send({ context, tools }: {
         context: Message[];
         tools: ToolDefinition[];

package/dist/llm/infrastructure/openAILLMClient.js

@@ -3,6 +3,7 @@ import { InvalidAgentConfigError, RecoverableError, UnrecoverableError, withSpen
 import { classifyHostFailure, classifyLocalFailure, classifyProviderFailure, describeFailure, treatErrors, } from '#src/shared/domain/providerFailure';
 import { isOneOf } from '#src/shared/domain/isOneOf';
 import { MaxContextError } from "../domain/errors.js";
+import { nativeOf } from "../domain/tool.js";
 // The SDK types `model` as a plain string, so this list is maintained by hand.
 export const openAIModels = [
     'gpt-6-astra',
@@ -36,13 +37,26 @@ const failedToolPrefix = 'The tool call failed: ';
 /**
  * OpenAI takes a flat list of items, so every part becomes its own, in order. Reasoning crosses
  * back only as the encrypted item this client kept, which OpenAI requires beside the tool calls
- * that followed it; a reasoning part is its summary, kept for narration.
+ * that followed it; a reasoning part is its summary, kept for narration. A call to the tool
+ * offered as `apply_patch`, named `applyPatch` here, and its result go back as that tool's own
+ * items.
  */
-function toOpenAIItems(part, role) {
+function toOpenAIItems(part, role, applyPatch) {
     switch (part.type) {
         case 'text':
             return part.text ? [{ role, content: part.text }] : [];
         case 'toolCall':
+            if (part.name === applyPatch) {
+                return [
+                    {
+                        type: 'apply_patch_call',
+                        call_id: part.id,
+                        // Kept as OpenAI sent it: the conversation holds the operation unchanged.
+                        operation: part.input,
+                        status: 'completed',
+                    },
+                ];
+            }
             return [
                 {
                     type: 'function_call',
@@ -53,6 +67,16 @@ function toOpenAIItems(part, role) {
                 },
             ];
         case 'toolResult':
+            if (part.name === applyPatch) {
+                return [
+                    {
+                        type: 'apply_patch_call_output',
+                        call_id: part.callId,
+                        status: part.isError ? 'failed' : 'completed',
+                        output: part.output,
+                    },
+                ];
+            }
             return [
                 {
                     type: 'function_call_output',
@@ -68,18 +92,30 @@ function toOpenAIItems(part, role) {
             return [];
     }
 }
-function toOpenAIInput(message) {
+function toOpenAIInput(message, applyPatch) {
     // A tool message holds only results, which carry no role.
     const role = message.role === 'assistant' ? 'assistant' : 'user';
-    return message.content.flatMap(part => toOpenAIItems(part, role));
+    return message.content.flatMap(part => toOpenAIItems(part, role, applyPatch));
 }
 /**
  * Strict: the API constrains the model to the schema, so its arguments always parse and match.
- * A schema outside what strict mode supports fails the request with a 400.
+ * A schema outside what strict mode supports fails the request with a 400. The native
+ * `apply_patch` has OpenAI's own schema.
  */
-function toOpenAITool({ name, description, inputSchema }) {
+function toOpenAITool(definition) {
+    const native = nativeOf(definition);
+    if (native === 'applyPatch')
+        return { type: 'apply_patch' };
+    if (native !== undefined) {
+        throw new InvalidAgentConfigError(`OpenAI has no native tool "${native}"`);
+    }
+    const { name, description, inputSchema } = definition;
     return { type: 'function', name, description, parameters: inputSchema, strict: true };
 }
+/** The name the tool offered as `apply_patch` has in the conversation, if one is offered. */
+function applyPatchName(tools) {
+    return tools.find(tool => nativeOf(tool) === 'applyPatch')?.name;
+}
 /**
  * Strict mode makes arguments that do not parse rare: a response cut short mid-call ends as
  * `truncated` before any tool runs. Should one arrive, the tool gets the raw string and its
@@ -93,7 +129,7 @@ function parseArguments(serialized) {
         return serialized;
     }
 }
-function describeItem(item, logger) {
+function describeItem(item, logger, applyPatch) {
     switch (item.type) {
         case 'message':
             // A refusal is not an answer: it only sets the stop reason.
@@ -123,14 +159,24 @@ function describeItem(item, logger) {
                     input: parseArguments(item.arguments),
                 },
             ];
+        case 'apply_patch_call':
+            // Only when `apply_patch` was offered; otherwise it is unexpected, as below.
+            if (applyPatch !== undefined) {
+                return [{ type: 'toolCall', id: item.call_id, name: applyPatch, input: item.operation }];
+            }
+            return unmapped(item, logger);
         default:
-            // No built-in tools are offered, so any other item is output this client does not expect.
-            treatErrors(() => {
-                logger.warn(item, 'OpenAI returned an output item the client does not map');
-            }, classifyHostFailure, 'OpenAI client logger failed while mapping a response');
-            return [];
+            return unmapped(item, logger);
     }
 }
+/** An item this client does not map, which it reports and leaves out. */
+function unmapped(item, logger) {
+    // No other built-in tool is offered, so any other item is output this client does not expect.
+    treatErrors(() => {
+        logger.warn(item, 'OpenAI returned an output item the client does not map');
+    }, classifyHostFailure, 'OpenAI client logger failed while mapping a response');
+    return [];
+}
 function toStopReason(response) {
     if (response.status === 'incomplete') {
         return response.incomplete_details?.reason === 'content_filter' ? 'refused' : 'truncated';
@@ -164,7 +210,7 @@ function toTokens(usage, logger) {
         outputTokens: usage.output_tokens,
     };
 }
-function toLLMResponse(response, logger) {
+function toLLMResponse(response, logger, applyPatch) {
     let usage;
     try {
         // Inside the try: it warns when the usage is missing, and that warning can throw too.
@@ -172,7 +218,7 @@ function toLLMResponse(response, logger) {
         return {
             message: {
                 role: 'assistant',
-                content: response.output.flatMap(item => describeItem(item, logger)),
+                content: response.output.flatMap(item => describeItem(item, logger, applyPatch)),
             },
             usage,
             stopReason: toStopReason(response),
@@ -236,13 +282,17 @@ export class OpenAILLMClient {
         this.systemPrompt = systemPrompt;
         this.logger = logger;
     }
+    supportsNative(tool) {
+        return tool === 'applyPatch';
+    }
     async send({ context, tools }, signal) {
         let response;
+        const applyPatch = applyPatchName(tools);
         try {
             response = await this.client.responses.create({
                 model: this.model,
                 instructions: this.systemPrompt,
-                input: context.flatMap(toOpenAIInput),
+                input: context.flatMap(message => toOpenAIInput(message, applyPatch)),
                 // Left out when empty, so a request without tools stays as it always was.
                 ...(tools.length > 0 && { tools: tools.map(toOpenAITool) }),
                 store: false,
@@ -256,7 +306,7 @@ export class OpenAILLMClient {
             signal.throwIfAborted();
             throw classifyOpenAIFailure(error);
         }
-        return treatErrors(() => toLLMResponse(response, this.logger), classifyLocalFailure, 'OpenAI response mapping failed');
+        return treatErrors(() => toLLMResponse(response, this.logger, applyPatch), classifyLocalFailure, 'OpenAI response mapping failed');
     }
 }
 //# sourceMappingURL=openAILLMClient.js.map

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

@@ -1 +1 @@
-{"version":3,"file":"openAILLMClient.js","sourceRoot":"","sources":["../../../src/llm/infrastructure/openAILLMClient.ts"],"names":[],"mappings":"AAAA,OAAO,MAAM,EAAE,EAAE,QAAQ,EAAE,MAAM,QAAQ,CAAC;AAU1C,OAAO,EACN,uBAAuB,EACvB,gBAAgB,EAChB,kBAAkB,EAClB,eAAe,GACf,MAAM,2BAA2B,CAAC;AACnC,OAAO,EACN,mBAAmB,EACnB,oBAAoB,EACpB,uBAAuB,EACvB,eAAe,EACf,WAAW,GACX,MAAM,oCAAoC,CAAC;AAC5C,OAAO,EAAE,OAAO,EAAE,MAAM,4BAA4B,CAAC;AAGrD,OAAO,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AAKtD,+EAA+E;AAC/E,MAAM,CAAC,MAAM,YAAY,GAAG;IAC3B,aAAa;IACb,aAAa;IACb,cAAc;IACd,eAAe;CACN,CAAC;AAGX,0FAA0F;AAC1F,MAAM,CAAC,MAAM,sBAAsB,GAAG;IACrC,KAAK;IACL,QAAQ;IACR,MAAM;IACN,OAAO;IACP,KAAK;CACsD,CAAC;AAG7D,kFAAkF;AAClF,MAAM,qBAAqB,GAAG,CAAC,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,CAAC,CAAC;AAExD,gGAAgG;AAChG,mDAAmD;AACnD,MAAM,mBAAmB,GAAG;IAC3B,oBAAoB;IACpB,0BAA0B;IAC1B,8BAA8B;CAC9B,CAAC;AAEF,gFAAgF;AAChF,MAAM,2BAA2B,GAAG,CAAC,cAAc,EAAE,qBAAqB,CAAC,CAAC;AAE5E,0FAA0F;AAC1F,MAAM,YAAY,GAAG,QAAQ,CAAC;AAE9B,yFAAyF;AACzF,MAAM,gBAAgB,GAAG,wBAAwB,CAAC;AAElD;;;;GAIG;AACH,SAAS,aAAa,CAAC,IAAiB,EAAE,IAA0B;IACnE,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;QACnB,KAAK,MAAM;YACV,OAAO,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;QACxD,KAAK,UAAU;YACd,OAAO;gBACN;oBACC,IAAI,EAAE,eAAe;oBACrB,OAAO,EAAE,IAAI,CAAC,EAAE;oBAChB,IAAI,EAAE,IAAI,CAAC,IAAI;oBACf,+EAA+E;oBAC/E,SAAS,EAAE,OAAO,IAAI,CAAC,KAAK,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC;iBACnF;aACD,CAAC;QACH,KAAK,YAAY;YAChB,OAAO;gBACN;oBACC,IAAI,EAAE,sBAAsB;oBAC5B,OAAO,EAAE,IAAI,CAAC,MAAM;oBACpB,8FAA8F;oBAC9F,MAAM,EAAE,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,gBAAgB,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM;iBACxE;aACD,CAAC;QACH,KAAK,cAAc;YAClB,6FAA6F;YAC7F,OAAO,IAAI,CAAC,MAAM,KAAK,YAAY,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAA6B,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;QACjF,KAAK,WAAW;YACf,OAAO,EAAE,CAAC;IACZ,CAAC;AACF,CAAC;AAED,SAAS,aAAa,CAAC,OAAgB;IACtC,0DAA0D;IAC1D,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,KAAK,WAAW,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,MAAM,CAAC;IACjE,OAAO,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC,aAAa,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC;AACnE,CAAC;AAED;;;GAGG;AACH,SAAS,YAAY,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,WAAW,EAAkB;IACvE,OAAO,EAAE,IAAI,EAAE,UAAU,EAAE,IAAI,EAAE,WAAW,EAAE,UAAU,EAAE,WAAW,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC;AACvF,CAAC;AAED;;;;GAIG;AACH,SAAS,cAAc,CAAC,UAAkB;IACzC,IAAI,CAAC;QACJ,OAAO,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC;IAC/B,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,UAAU,CAAC;IACnB,CAAC;AACF,CAAC;AAED,SAAS,YAAY,CAAC,IAAwB,EAAE,MAAe;IAC9D,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;QACnB,KAAK,SAAS;YACb,4DAA4D;YAC5D,OAAO,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,CACrC,OAAO,CAAC,IAAI,KAAK,aAAa,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,MAAe,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CACrF,CAAC;QACH,KAAK,WAAW,CAAC,CAAC,CAAC;YAClB,kFAAkF;YAClF,MAAM,IAAI,GAAG,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YAClE,MAAM,KAAK,GAAkB,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;YACvE,oFAAoF;YACpF,IAAI,IAAI,CAAC,iBAAiB,EAAE,CAAC;gBAC5B,MAAM,SAAS,GAA0B;oBACxC,IAAI,EAAE,WAAW;oBACjB,EAAE,EAAE,IAAI,CAAC,EAAE;oBACX,OAAO,EAAE,IAAI,CAAC,OAAO;oBACrB,iBAAiB,EAAE,IAAI,CAAC,iBAAiB;iBACzC,CAAC;gBACF,KAAK,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,cAAc,EAAE,MAAM,EAAE,YAAY,EAAE,IAAI,EAAE,SAAS,EAAE,CAAC,CAAC;YAC7E,CAAC;YACD,OAAO,KAAK,CAAC;QACd,CAAC;QACD,KAAK,eAAe;YACnB,OAAO;gBACN;oBACC,IAAI,EAAE,UAAU;oBAChB,EAAE,EAAE,IAAI,CAAC,OAAO;oBAChB,IAAI,EAAE,IAAI,CAAC,IAAI;oBACf,KAAK,EAAE,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC;iBACrC;aACD,CAAC;QACH;YACC,0FAA0F;YAC1F,WAAW,CACV,GAAG,EAAE;gBACJ,MAAM,CAAC,IAAI,CAAC,IAAI,EAAE,wDAAwD,CAAC,CAAC;YAC7E,CAAC,EACD,mBAAmB,EACnB,sDAAsD,CACtD,CAAC;YACF,OAAO,EAAE,CAAC;IACZ,CAAC;AACF,CAAC;AAED,SAAS,YAAY,CAAC,QAAkB;IACvC,IAAI,QAAQ,CAAC,MAAM,KAAK,YAAY,EAAE,CAAC;QACtC,OAAO,QAAQ,CAAC,kBAAkB,EAAE,MAAM,KAAK,gBAAgB,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,WAAW,CAAC;IAC3F,CAAC;IACD,IAAI,QAAQ,CAAC,MAAM,KAAK,WAAW,EAAE,CAAC;QACrC,MAAM,KAAK,GACV,QAAQ,CAAC,KAAK,EAAE,OAAO;YACvB,mCAAmC,QAAQ,CAAC,MAAM,IAAI,SAAS,IAAI,CAAC;QACrE,uFAAuF;QACvF,IAAI,QAAQ,CAAC,KAAK,IAAI,2BAA2B,CAAC,QAAQ,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC;YACjF,MAAM,IAAI,gBAAgB,CAAC,4BAA4B,EAAE,EAAE,KAAK,EAAE,CAAC,CAAC;QACrE,CAAC;QACD,MAAM,IAAI,kBAAkB,CAAC,sCAAsC,EAAE,EAAE,KAAK,EAAE,CAAC,CAAC;IACjF,CAAC;IAED,MAAM,OAAO,GAAG,QAAQ,CAAC,MAAM,CAAC,IAAI,CACnC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,KAAK,SAAS,IAAI,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC,OAAO,CAAC,IAAI,KAAK,SAAS,CAAC,CAC3F,CAAC;IACF,OAAO,OAAO,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,WAAW,CAAC;AAC1C,CAAC;AAED,6EAA6E;AAC7E,SAAS,QAAQ,CAAC,KAAgC,EAAE,MAAe;IAClE,IAAI,CAAC,KAAK,EAAE,CAAC;QACZ,+DAA+D;QAC/D,WAAW,CACV,GAAG,EAAE;YACJ,MAAM,CAAC,IAAI,CAAC,sEAAsE,CAAC,CAAC;QACrF,CAAC,EACD,mBAAmB,EACnB,iDAAiD,CACjD,CAAC;QACF,OAAO,SAAS,CAAC;IAClB,CAAC;IAED,MAAM,EAAE,aAAa,EAAE,kBAAkB,EAAE,GAAG,KAAK,CAAC,oBAAoB,CAAC;IACzE,OAAO;QACN,WAAW,EAAE,KAAK,CAAC,YAAY,GAAG,aAAa,GAAG,kBAAkB;QACpE,eAAe,EAAE,aAAa;QAC9B,kBAAkB,EAAE,kBAAkB;QACtC,YAAY,EAAE,KAAK,CAAC,aAAa;KACjC,CAAC;AACH,CAAC;AAED,SAAS,aAAa,CAAC,QAAkB,EAAE,MAAe;IACzD,IAAI,KAAyB,CAAC;IAE9B,IAAI,CAAC;QACJ,sFAAsF;QACtF,KAAK,GAAG,QAAQ,CAAC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;QACzC,OAAO;YACN,OAAO,EAAE;gBACR,IAAI,EAAE,WAAW;gBACjB,OAAO,EAAE,QAAQ,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC,YAAY,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;aACpE;YACD,KAAK;YACL,UAAU,EAAE,YAAY,CAAC,QAAQ,CAAC;SAClC,CAAC;IACH,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,wFAAwF;QACxF,0DAA0D;QAC1D,MAAM,eAAe,CAAC,KAAK,EAAE,KAAK,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;IACtD,CAAC;AACF,CAAC;AAED,SAAS,qBAAqB,CAAC,KAAc;IAC5C,IAAI,KAAK,YAAY,QAAQ,EAAE,CAAC;QAC/B,IAAI,KAAK,CAAC,IAAI,KAAK,yBAAyB,EAAE,CAAC;YAC9C,OAAO,IAAI,eAAe,CAAC,6CAA6C,EAAE;gBACzE,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;QACD,IACC,mBAAmB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,KAAK,KAAK,CAAC,IAAI,CAAC;YACrD,qBAAqB,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,CAAC,EAC5D,CAAC;YACF,OAAO,IAAI,kBAAkB,CAAC,6BAA6B,EAAE;gBAC5D,KAAK,EAAE,eAAe,CAAC,KAAK,CAAC;aAC7B,CAAC,CAAC;QACJ,CAAC;IACF,CAAC;IAED,OAAO,uBAAuB,CAAC,KAAK,EAAE,2BAA2B,CAAC,CAAC;AACpE,CAAC;AAED;;;;GAIG;AACH,MAAM,OAAO,eAAe;IACV,MAAM,CAAS;IACf,KAAK,CAAc;IACnB,eAAe,CAAwB;IACvC,YAAY,CAAS;IACrB,MAAM,CAAU;IAEjC,2FAA2F;IAC3F,wBAAwB;IACxB,YAAY,EACX,KAAK,EACL,eAAe,GAAG,MAAM,EACxB,YAAY,EACZ,MAAM,GAMN;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;QACD,IAAI,CAAC,OAAO,CAAC,sBAAsB,EAAE,eAAe,CAAC,EAAE,CAAC;YACvD,MAAM,IAAI,uBAAuB,CAChC,IAAI,eAAe,yDAAyD,sBAAsB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAC/G,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,eAAe,GAAG,eAAe,CAAC;QACvC,IAAI,CAAC,YAAY,GAAG,YAAY,CAAC;QACjC,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IACtB,CAAC;IAED,KAAK,CAAC,IAAI,CACT,EAAE,OAAO,EAAE,KAAK,EAAmD,EACnE,MAAmB;QAEnB,IAAI,QAAkB,CAAC;QAEvB,IAAI,CAAC;YACJ,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,MAAM,CAC5C;gBACC,KAAK,EAAE,IAAI,CAAC,KAAK;gBACjB,YAAY,EAAE,IAAI,CAAC,YAAY;gBAC/B,KAAK,EAAE,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC;gBACrC,0EAA0E;gBAC1E,GAAG,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,EAAE,CAAC;gBAC3D,KAAK,EAAE,KAAK;gBACZ,SAAS,EAAE,EAAE,MAAM,EAAE,IAAI,CAAC,eAAe,EAAE,OAAO,EAAE,MAAM,EAAE;gBAC5D,uEAAuE;gBACvE,OAAO,EAAE,CAAC,6BAA6B,CAAC;aACxC,EACD,EAAE,MAAM,EAAE,CACV,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,mFAAmF;YACnF,MAAM,CAAC,cAAc,EAAE,CAAC;YACxB,MAAM,qBAAqB,CAAC,KAAK,CAAC,CAAC;QACpC,CAAC;QAED,OAAO,WAAW,CACjB,GAAG,EAAE,CAAC,aAAa,CAAC,QAAQ,EAAE,IAAI,CAAC,MAAM,CAAC,EAC1C,oBAAoB,EACpB,gCAAgC,CAChC,CAAC;IACH,CAAC;CACD","sourcesContent":["import OpenAI, { APIError } from 'openai';\nimport type { ReasoningEffort } from 'openai/resources/shared';\nimport type {\n\tFunctionTool,\n\tResponse,\n\tResponseInputItem,\n\tResponseOutputItem,\n\tResponseReasoningItem,\n\tResponseUsage,\n} from 'openai/resources/responses/responses';\nimport {\n\tInvalidAgentConfigError,\n\tRecoverableError,\n\tUnrecoverableError,\n\twithSpentTokens,\n} from '#src/shared/domain/errors';\nimport {\n\tclassifyHostFailure,\n\tclassifyLocalFailure,\n\tclassifyProviderFailure,\n\tdescribeFailure,\n\ttreatErrors,\n} from '#src/shared/domain/providerFailure';\nimport { isOneOf } from '#src/shared/domain/isOneOf';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport type { Tokens } from '#src/shared/domain/tokens';\nimport { MaxContextError } from '../domain/errors.ts';\nimport type { LLMClient, LLMResponse, StopReason } from '../domain/llm.ts';\nimport type { Message, MessagePart } from '../domain/message.ts';\nimport type { ToolDefinition } from '../domain/tool.ts';\n\n// The SDK types `model` as a plain string, so this list is maintained by hand.\nexport const openAIModels = [\n\t'gpt-6-astra',\n\t'gpt-5.6-sol',\n\t'gpt-5.6-luna',\n\t'gpt-5.6-terra',\n] as const;\nexport type OpenAIModel = (typeof openAIModels)[number];\n\n// The efforts both the Responses API and Codex's model catalog list for the models above.\nexport const openAIReasoningEfforts = [\n\t'low',\n\t'medium',\n\t'high',\n\t'xhigh',\n\t'max',\n] as const satisfies readonly NonNullable<ReasoningEffort>[];\nexport type OpenAIReasoningEffort = (typeof openAIReasoningEfforts)[number];\n\n// The request itself is wrong or not allowed, so sending it again cannot succeed.\nconst unrecoverableStatuses = [400, 401, 403, 404, 422];\n\n// Out of credit or over a spend limit. These arrive as a 429 like a rate limit, but the account\n// stays blocked however often the request is sent.\nconst exhaustedQuotaCodes = [\n\t'insufficient_quota',\n\t'credit_balance_exhausted',\n\t'project_spend_limit_exceeded',\n];\n\n// The codes OpenAI gives a failed response that a new attempt can succeed past.\nconst transientResponseErrorCodes = ['server_error', 'rate_limit_exceeded'];\n\n// This client's label on the reasoning items it keeps for replay; see `ProviderDataPart`.\nconst openAISource = 'openai';\n\n// Marks a failed call's output, which would otherwise read the same as a successful one.\nconst failedToolPrefix = 'The tool call failed: ';\n\n/**\n * OpenAI takes a flat list of items, so every part becomes its own, in order. Reasoning crosses\n * back only as the encrypted item this client kept, which OpenAI requires beside the tool calls\n * that followed it; a reasoning part is its summary, kept for narration.\n */\nfunction toOpenAIItems(part: MessagePart, role: 'user' | 'assistant'): ResponseInputItem[] {\n\tswitch (part.type) {\n\t\tcase 'text':\n\t\t\treturn part.text ? [{ role, content: part.text }] : [];\n\t\tcase 'toolCall':\n\t\t\treturn [\n\t\t\t\t{\n\t\t\t\t\ttype: 'function_call',\n\t\t\t\t\tcall_id: part.id,\n\t\t\t\t\tname: part.name,\n\t\t\t\t\t// A string is arguments that did not parse, sent back as the model wrote them.\n\t\t\t\t\targuments: typeof part.input === 'string' ? part.input : JSON.stringify(part.input),\n\t\t\t\t},\n\t\t\t];\n\t\tcase 'toolResult':\n\t\t\treturn [\n\t\t\t\t{\n\t\t\t\t\ttype: 'function_call_output',\n\t\t\t\t\tcall_id: part.callId,\n\t\t\t\t\t// OpenAI has no error flag, so the text says it; the conversation keeps the output as it was.\n\t\t\t\t\toutput: part.isError ? `${failedToolPrefix}${part.output}` : part.output,\n\t\t\t\t},\n\t\t\t];\n\t\tcase 'providerData':\n\t\t\t// Another provider's item would be rejected; losing it only costs that provider's reasoning.\n\t\t\treturn part.source === openAISource ? [part.data as ResponseReasoningItem] : [];\n\t\tcase 'reasoning':\n\t\t\treturn [];\n\t}\n}\n\nfunction toOpenAIInput(message: Message): ResponseInputItem[] {\n\t// A tool message holds only results, which carry no role.\n\tconst role = message.role === 'assistant' ? 'assistant' : 'user';\n\treturn message.content.flatMap(part => toOpenAIItems(part, role));\n}\n\n/**\n * Strict: the API constrains the model to the schema, so its arguments always parse and match.\n * A schema outside what strict mode supports fails the request with a 400.\n */\nfunction toOpenAITool({ name, description, inputSchema }: ToolDefinition): FunctionTool {\n\treturn { type: 'function', name, description, parameters: inputSchema, strict: true };\n}\n\n/**\n * Strict mode makes arguments that do not parse rare: a response cut short mid-call ends as\n * `truncated` before any tool runs. Should one arrive, the tool gets the raw string and its\n * rejection lets the model correct itself, instead of the run failing.\n */\nfunction parseArguments(serialized: string): unknown {\n\ttry {\n\t\treturn JSON.parse(serialized);\n\t} catch {\n\t\treturn serialized;\n\t}\n}\n\nfunction describeItem(item: ResponseOutputItem, logger: ILogger): MessagePart[] {\n\tswitch (item.type) {\n\t\tcase 'message':\n\t\t\t// A refusal is not an answer: it only sets the stop reason.\n\t\t\treturn item.content.flatMap(content =>\n\t\t\t\tcontent.type === 'output_text' ? [{ type: 'text' as const, text: content.text }] : [],\n\t\t\t);\n\t\tcase 'reasoning': {\n\t\t\t// The summary is the readable part; the encrypted item is what OpenAI reads back.\n\t\t\tconst text = item.summary.map(summary => summary.text).join('\\n');\n\t\t\tconst parts: MessagePart[] = text ? [{ type: 'reasoning', text }] : [];\n\t\t\t// Without its encrypted content the item cannot be replayed: OpenAI stores nothing.\n\t\t\tif (item.encrypted_content) {\n\t\t\t\tconst encrypted: ResponseReasoningItem = {\n\t\t\t\t\ttype: 'reasoning',\n\t\t\t\t\tid: item.id,\n\t\t\t\t\tsummary: item.summary,\n\t\t\t\t\tencrypted_content: item.encrypted_content,\n\t\t\t\t};\n\t\t\t\tparts.push({ type: 'providerData', source: openAISource, data: encrypted });\n\t\t\t}\n\t\t\treturn parts;\n\t\t}\n\t\tcase 'function_call':\n\t\t\treturn [\n\t\t\t\t{\n\t\t\t\t\ttype: 'toolCall',\n\t\t\t\t\tid: item.call_id,\n\t\t\t\t\tname: item.name,\n\t\t\t\t\tinput: parseArguments(item.arguments),\n\t\t\t\t},\n\t\t\t];\n\t\tdefault:\n\t\t\t// No built-in tools are offered, so any other item is output this client does not expect.\n\t\t\ttreatErrors(\n\t\t\t\t() => {\n\t\t\t\t\tlogger.warn(item, 'OpenAI returned an output item the client does not map');\n\t\t\t\t},\n\t\t\t\tclassifyHostFailure,\n\t\t\t\t'OpenAI client logger failed while mapping a response',\n\t\t\t);\n\t\t\treturn [];\n\t}\n}\n\nfunction toStopReason(response: Response): StopReason {\n\tif (response.status === 'incomplete') {\n\t\treturn response.incomplete_details?.reason === 'content_filter' ? 'refused' : 'truncated';\n\t}\n\tif (response.status !== 'completed') {\n\t\tconst cause =\n\t\t\tresponse.error?.message ??\n\t\t\t`The response ended with status \"${response.status ?? 'unknown'}\".`;\n\t\t// The same failure is retried when it arrives as an SDK error, so it must be here too.\n\t\tif (response.error && transientResponseErrorCodes.includes(response.error.code)) {\n\t\t\tthrow new RecoverableError('The OpenAI response failed', { cause });\n\t\t}\n\t\tthrow new UnrecoverableError('The OpenAI response did not complete', { cause });\n\t}\n\n\tconst refused = response.output.some(\n\t\titem => item.type === 'message' && item.content.some(content => content.type === 'refusal'),\n\t);\n\treturn refused ? 'refused' : 'completed';\n}\n\n/** OpenAI counts both cache reads and cache writes inside `input_tokens`. */\nfunction toTokens(usage: ResponseUsage | undefined, logger: ILogger): Tokens | undefined {\n\tif (!usage) {\n\t\t// Zeros would report a call that may have been billed as free.\n\t\ttreatErrors(\n\t\t\t() => {\n\t\t\t\tlogger.warn('OpenAI returned a response without usage; it cannot be accounted for');\n\t\t\t},\n\t\t\tclassifyHostFailure,\n\t\t\t'OpenAI client logger failed while reading usage',\n\t\t);\n\t\treturn undefined;\n\t}\n\n\tconst { cached_tokens, cache_write_tokens } = usage.input_tokens_details;\n\treturn {\n\t\tinputTokens: usage.input_tokens - cached_tokens - cache_write_tokens,\n\t\treadCacheTokens: cached_tokens,\n\t\twrittenCacheTokens: cache_write_tokens,\n\t\toutputTokens: usage.output_tokens,\n\t};\n}\n\nfunction toLLMResponse(response: Response, logger: ILogger): LLMResponse {\n\tlet usage: Tokens | undefined;\n\n\ttry {\n\t\t// Inside the try: it warns when the usage is missing, and that warning can throw too.\n\t\tusage = toTokens(response.usage, logger);\n\t\treturn {\n\t\t\tmessage: {\n\t\t\t\trole: 'assistant',\n\t\t\t\tcontent: response.output.flatMap(item => describeItem(item, logger)),\n\t\t\t},\n\t\t\tusage,\n\t\t\tstopReason: toStopReason(response),\n\t\t};\n\t} catch (error) {\n\t\t// The call was billed whether the response failed or a logger did while mapping it, and\n\t\t// an answer without usage leaves the run's count unknown.\n\t\tthrow withSpentTokens(error, usage, !response.usage);\n\t}\n}\n\nfunction classifyOpenAIFailure(error: unknown): Error {\n\tif (error instanceof APIError) {\n\t\tif (error.code === 'context_length_exceeded') {\n\t\t\treturn new MaxContextError('The context no longer fits the OpenAI model', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t\tif (\n\t\t\texhaustedQuotaCodes.some(code => code === error.code) ||\n\t\t\tunrecoverableStatuses.some(status => status === error.status)\n\t\t) {\n\t\t\treturn new UnrecoverableError('OpenAI rejected the request', {\n\t\t\t\tcause: describeFailure(error),\n\t\t\t});\n\t\t}\n\t}\n\n\treturn classifyProviderFailure(error, 'The OpenAI request failed');\n}\n\n/**\n * A stateless `LLMClient` on the OpenAI Responses API. With `store: false` OpenAI keeps\n * nothing between calls, so every call carries the whole context: the conversation is\n * MiKode's, not the provider's.\n */\nexport class OpenAILLMClient implements LLMClient {\n\tprivate readonly client: OpenAI;\n\tprivate readonly model: OpenAIModel;\n\tprivate readonly reasoningEffort: OpenAIReasoningEffort;\n\tprivate readonly systemPrompt: string;\n\tprivate readonly logger: ILogger;\n\n\t// `model` and `reasoningEffort` are untyped on purpose: this adapter is the one that knows\n\t// what OpenAI supports.\n\tconstructor({\n\t\tmodel,\n\t\treasoningEffort = 'high',\n\t\tsystemPrompt,\n\t\tlogger,\n\t}: {\n\t\tmodel: string;\n\t\treasoningEffort?: 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: ${openAIModels.join(', ')}`,\n\t\t\t);\n\t\t}\n\t\tif (!isOneOf(openAIReasoningEfforts, reasoningEffort)) {\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`\"${reasoningEffort}\" is not an OpenAI reasoning effort; expected one of: ${openAIReasoningEfforts.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.reasoningEffort = reasoningEffort;\n\t\tthis.systemPrompt = systemPrompt;\n\t\tthis.logger = logger;\n\t}\n\n\tasync send(\n\t\t{ context, tools }: { context: Message[]; tools: ToolDefinition[] },\n\t\tsignal: AbortSignal,\n\t): Promise<LLMResponse> {\n\t\tlet response: Response;\n\n\t\ttry {\n\t\t\tresponse = await this.client.responses.create(\n\t\t\t\t{\n\t\t\t\t\tmodel: this.model,\n\t\t\t\t\tinstructions: this.systemPrompt,\n\t\t\t\t\tinput: context.flatMap(toOpenAIInput),\n\t\t\t\t\t// Left out when empty, so a request without tools stays as it always was.\n\t\t\t\t\t...(tools.length > 0 && { tools: tools.map(toOpenAITool) }),\n\t\t\t\t\tstore: false,\n\t\t\t\t\treasoning: { effort: this.reasoningEffort, summary: 'auto' },\n\t\t\t\t\t// With `store: false` this is the only way to send the reasoning back.\n\t\t\t\t\tinclude: ['reasoning.encrypted_content'],\n\t\t\t\t},\n\t\t\t\t{ signal },\n\t\t\t);\n\t\t} catch (error) {\n\t\t\t// The SDK's abort error is named 'Error', so it would pass for a provider failure.\n\t\t\tsignal.throwIfAborted();\n\t\t\tthrow classifyOpenAIFailure(error);\n\t\t}\n\n\t\treturn treatErrors(\n\t\t\t() => toLLMResponse(response, this.logger),\n\t\t\tclassifyLocalFailure,\n\t\t\t'OpenAI response mapping failed',\n\t\t);\n\t}\n}\n"]}
+{"version":3,"file":"openAILLMClient.js","sourceRoot":"","sources":["../../../src/llm/infrastructure/openAILLMClient.ts"],"names":[],"mappings":"AAAA,OAAO,MAAM,EAAE,EAAE,QAAQ,EAAE,MAAM,QAAQ,CAAC;AAY1C,OAAO,EACN,uBAAuB,EACvB,gBAAgB,EAChB,kBAAkB,EAClB,eAAe,GACf,MAAM,2BAA2B,CAAC;AACnC,OAAO,EACN,mBAAmB,EACnB,oBAAoB,EACpB,uBAAuB,EACvB,eAAe,EACf,WAAW,GACX,MAAM,oCAAoC,CAAC;AAC5C,OAAO,EAAE,OAAO,EAAE,MAAM,4BAA4B,CAAC;AAGrD,OAAO,EAAE,eAAe,EAAE,MAAM,qBAAqB,CAAC;AAGtD,OAAO,EAAmB,QAAQ,EAAuB,MAAM,mBAAmB,CAAC;AAEnF,+EAA+E;AAC/E,MAAM,CAAC,MAAM,YAAY,GAAG;IAC3B,aAAa;IACb,aAAa;IACb,cAAc;IACd,eAAe;CACN,CAAC;AAGX,0FAA0F;AAC1F,MAAM,CAAC,MAAM,sBAAsB,GAAG;IACrC,KAAK;IACL,QAAQ;IACR,MAAM;IACN,OAAO;IACP,KAAK;CACsD,CAAC;AAG7D,kFAAkF;AAClF,MAAM,qBAAqB,GAAG,CAAC,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,CAAC,CAAC;AAExD,gGAAgG;AAChG,mDAAmD;AACnD,MAAM,mBAAmB,GAAG;IAC3B,oBAAoB;IACpB,0BAA0B;IAC1B,8BAA8B;CAC9B,CAAC;AAEF,gFAAgF;AAChF,MAAM,2BAA2B,GAAG,CAAC,cAAc,EAAE,qBAAqB,CAAC,CAAC;AAE5E,0FAA0F;AAC1F,MAAM,YAAY,GAAG,QAAQ,CAAC;AAE9B,yFAAyF;AACzF,MAAM,gBAAgB,GAAG,wBAAwB,CAAC;AAElD;;;;;;GAMG;AACH,SAAS,aAAa,CACrB,IAAiB,EACjB,IAA0B,EAC1B,UAA8B;IAE9B,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;QACnB,KAAK,MAAM;YACV,OAAO,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;QACxD,KAAK,UAAU;YACd,IAAI,IAAI,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;gBAC9B,OAAO;oBACN;wBACC,IAAI,EAAE,kBAAkB;wBACxB,OAAO,EAAE,IAAI,CAAC,EAAE;wBAChB,0EAA0E;wBAC1E,SAAS,EAAE,IAAI,CAAC,KAAgD;wBAChE,MAAM,EAAE,WAAW;qBACnB;iBACD,CAAC;YACH,CAAC;YACD,OAAO;gBACN;oBACC,IAAI,EAAE,eAAe;oBACrB,OAAO,EAAE,IAAI,CAAC,EAAE;oBAChB,IAAI,EAAE,IAAI,CAAC,IAAI;oBACf,+EAA+E;oBAC/E,SAAS,EAAE,OAAO,IAAI,CAAC,KAAK,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC;iBACnF;aACD,CAAC;QACH,KAAK,YAAY;YAChB,IAAI,IAAI,CAAC,IAAI,KAAK,UAAU,EAAE,CAAC;gBAC9B,OAAO;oBACN;wBACC,IAAI,EAAE,yBAAyB;wBAC/B,OAAO,EAAE,IAAI,CAAC,MAAM;wBACpB,MAAM,EAAE,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,WAAW;wBAC7C,MAAM,EAAE,IAAI,CAAC,MAAM;qBACnB;iBACD,CAAC;YACH,CAAC;YACD,OAAO;gBACN;oBACC,IAAI,EAAE,sBAAsB;oBAC5B,OAAO,EAAE,IAAI,CAAC,MAAM;oBACpB,8FAA8F;oBAC9F,MAAM,EAAE,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,gBAAgB,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM;iBACxE;aACD,CAAC;QACH,KAAK,cAAc;YAClB,6FAA6F;YAC7F,OAAO,IAAI,CAAC,MAAM,KAAK,YAAY,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAA6B,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;QACjF,KAAK,WAAW;YACf,OAAO,EAAE,CAAC;IACZ,CAAC;AACF,CAAC;AAED,SAAS,aAAa,CAAC,OAAgB,EAAE,UAA8B;IACtE,0DAA0D;IAC1D,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,KAAK,WAAW,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,MAAM,CAAC;IACjE,OAAO,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC,aAAa,CAAC,IAAI,EAAE,IAAI,EAAE,UAAU,CAAC,CAAC,CAAC;AAC/E,CAAC;AAED;;;;GAIG;AACH,SAAS,YAAY,CAAC,UAA0B;IAC/C,MAAM,MAAM,GAAG,QAAQ,CAAC,UAAU,CAAC,CAAC;IACpC,IAAI,MAAM,KAAK,YAAY;QAAE,OAAO,EAAE,IAAI,EAAE,aAAa,EAAE,CAAC;IAC5D,IAAI,MAAM,KAAK,SAAS,EAAE,CAAC;QAC1B,MAAM,IAAI,uBAAuB,CAAC,8BAA8B,MAAM,GAAG,CAAC,CAAC;IAC5E,CAAC;IACD,MAAM,EAAE,IAAI,EAAE,WAAW,EAAE,WAAW,EAAE,GAAG,UAAU,CAAC;IACtD,OAAO,EAAE,IAAI,EAAE,UAAU,EAAE,IAAI,EAAE,WAAW,EAAE,UAAU,EAAE,WAAW,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC;AACvF,CAAC;AAED,6FAA6F;AAC7F,SAAS,cAAc,CAAC,KAAuB;IAC9C,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,YAAY,CAAC,EAAE,IAAI,CAAC;AAClE,CAAC;AAED;;;;GAIG;AACH,SAAS,cAAc,CAAC,UAAkB;IACzC,IAAI,CAAC;QACJ,OAAO,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC;IAC/B,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,UAAU,CAAC;IACnB,CAAC;AACF,CAAC;AAED,SAAS,YAAY,CACpB,IAAwB,EACxB,MAAe,EACf,UAA8B;IAE9B,QAAQ,IAAI,CAAC,IAAI,EAAE,CAAC;QACnB,KAAK,SAAS;YACb,4DAA4D;YAC5D,OAAO,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,CACrC,OAAO,CAAC,IAAI,KAAK,aAAa,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,MAAe,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CACrF,CAAC;QACH,KAAK,WAAW,CAAC,CAAC,CAAC;YAClB,kFAAkF;YAClF,MAAM,IAAI,GAAG,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YAClE,MAAM,KAAK,GAAkB,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;YACvE,oFAAoF;YACpF,IAAI,IAAI,CAAC,iBAAiB,EAAE,CAAC;gBAC5B,MAAM,SAAS,GAA0B;oBACxC,IAAI,EAAE,WAAW;oBACjB,EAAE,EAAE,IAAI,CAAC,EAAE;oBACX,OAAO,EAAE,IAAI,CAAC,OAAO;oBACrB,iBAAiB,EAAE,IAAI,CAAC,iBAAiB;iBACzC,CAAC;gBACF,KAAK,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,cAAc,EAAE,MAAM,EAAE,YAAY,EAAE,IAAI,EAAE,SAAS,EAAE,CAAC,CAAC;YAC7E,CAAC;YACD,OAAO,KAAK,CAAC;QACd,CAAC;QACD,KAAK,eAAe;YACnB,OAAO;gBACN;oBACC,IAAI,EAAE,UAAU;oBAChB,EAAE,EAAE,IAAI,CAAC,OAAO;oBAChB,IAAI,EAAE,IAAI,CAAC,IAAI;oBACf,KAAK,EAAE,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC;iBACrC;aACD,CAAC;QACH,KAAK,kBAAkB;YACtB,6EAA6E;YAC7E,IAAI,UAAU,KAAK,SAAS,EAAE,CAAC;gBAC9B,OAAO,CAAC,EAAE,IAAI,EAAE,UAAU,EAAE,EAAE,EAAE,IAAI,CAAC,OAAO,EAAE,IAAI,EAAE,UAAU,EAAE,KAAK,EAAE,IAAI,CAAC,SAAS,EAAE,CAAC,CAAC;YAC1F,CAAC;YACD,OAAO,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;QAC/B;YACC,OAAO,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;IAChC,CAAC;AACF,CAAC;AAED,yEAAyE;AACzE,SAAS,QAAQ,CAAC,IAAwB,EAAE,MAAe;IAC1D,8FAA8F;IAC9F,WAAW,CACV,GAAG,EAAE;QACJ,MAAM,CAAC,IAAI,CAAC,IAAI,EAAE,wDAAwD,CAAC,CAAC;IAC7E,CAAC,EACD,mBAAmB,EACnB,sDAAsD,CACtD,CAAC;IACF,OAAO,EAAE,CAAC;AACX,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,CACrB,QAAkB,EAClB,MAAe,EACf,UAA8B;IAE9B,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,EAAE,UAAU,CAAC,CAAC;aAChF;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,eAAe,CAAwB;IACvC,YAAY,CAAS;IACrB,MAAM,CAAU;IAEjC,2FAA2F;IAC3F,wBAAwB;IACxB,YAAY,EACX,KAAK,EACL,eAAe,GAAG,MAAM,EACxB,YAAY,EACZ,MAAM,GAMN;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;QACD,IAAI,CAAC,OAAO,CAAC,sBAAsB,EAAE,eAAe,CAAC,EAAE,CAAC;YACvD,MAAM,IAAI,uBAAuB,CAChC,IAAI,eAAe,yDAAyD,sBAAsB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAC/G,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,eAAe,GAAG,eAAe,CAAC;QACvC,IAAI,CAAC,YAAY,GAAG,YAAY,CAAC;QACjC,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IACtB,CAAC;IAED,cAAc,CAAC,IAAgB;QAC9B,OAAO,IAAI,KAAK,YAAY,CAAC;IAC9B,CAAC;IAED,KAAK,CAAC,IAAI,CACT,EAAE,OAAO,EAAE,KAAK,EAAmD,EACnE,MAAmB;QAEnB,IAAI,QAAkB,CAAC;QACvB,MAAM,UAAU,GAAG,cAAc,CAAC,KAAK,CAAC,CAAC;QAEzC,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,OAAO,CAAC,EAAE,CAAC,aAAa,CAAC,OAAO,EAAE,UAAU,CAAC,CAAC;gBACrE,0EAA0E;gBAC1E,GAAG,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,YAAY,CAAC,EAAE,CAAC;gBAC3D,KAAK,EAAE,KAAK;gBACZ,SAAS,EAAE,EAAE,MAAM,EAAE,IAAI,CAAC,eAAe,EAAE,OAAO,EAAE,MAAM,EAAE;gBAC5D,uEAAuE;gBACvE,OAAO,EAAE,CAAC,6BAA6B,CAAC;aACxC,EACD,EAAE,MAAM,EAAE,CACV,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,mFAAmF;YACnF,MAAM,CAAC,cAAc,EAAE,CAAC;YACxB,MAAM,qBAAqB,CAAC,KAAK,CAAC,CAAC;QACpC,CAAC;QAED,OAAO,WAAW,CACjB,GAAG,EAAE,CAAC,aAAa,CAAC,QAAQ,EAAE,IAAI,CAAC,MAAM,EAAE,UAAU,CAAC,EACtD,oBAAoB,EACpB,gCAAgC,CAChC,CAAC;IACH,CAAC;CACD","sourcesContent":["import OpenAI, { APIError } from 'openai';\nimport type { ReasoningEffort } from 'openai/resources/shared';\nimport type {\n\tApplyPatchTool,\n\tFunctionTool,\n\tResponse,\n\tResponseApplyPatchToolCall,\n\tResponseInputItem,\n\tResponseOutputItem,\n\tResponseReasoningItem,\n\tResponseUsage,\n} from 'openai/resources/responses/responses';\nimport {\n\tInvalidAgentConfigError,\n\tRecoverableError,\n\tUnrecoverableError,\n\twithSpentTokens,\n} from '#src/shared/domain/errors';\nimport {\n\tclassifyHostFailure,\n\tclassifyLocalFailure,\n\tclassifyProviderFailure,\n\tdescribeFailure,\n\ttreatErrors,\n} from '#src/shared/domain/providerFailure';\nimport { isOneOf } from '#src/shared/domain/isOneOf';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport type { Tokens } from '#src/shared/domain/tokens';\nimport { MaxContextError } from '../domain/errors.ts';\nimport type { LLMClient, LLMResponse, StopReason } from '../domain/llm.ts';\nimport type { Message, MessagePart } from '../domain/message.ts';\nimport { type NativeTool, nativeOf, type ToolDefinition } from '../domain/tool.ts';\n\n// The SDK types `model` as a plain string, so this list is maintained by hand.\nexport const openAIModels = [\n\t'gpt-6-astra',\n\t'gpt-5.6-sol',\n\t'gpt-5.6-luna',\n\t'gpt-5.6-terra',\n] as const;\nexport type OpenAIModel = (typeof openAIModels)[number];\n\n// The efforts both the Responses API and Codex's model catalog list for the models above.\nexport const openAIReasoningEfforts = [\n\t'low',\n\t'medium',\n\t'high',\n\t'xhigh',\n\t'max',\n] as const satisfies readonly NonNullable<ReasoningEffort>[];\nexport type OpenAIReasoningEffort = (typeof openAIReasoningEfforts)[number];\n\n// The request itself is wrong or not allowed, so sending it again cannot succeed.\nconst unrecoverableStatuses = [400, 401, 403, 404, 422];\n\n// Out of credit or over a spend limit. These arrive as a 429 like a rate limit, but the account\n// stays blocked however often the request is sent.\nconst exhaustedQuotaCodes = [\n\t'insufficient_quota',\n\t'credit_balance_exhausted',\n\t'project_spend_limit_exceeded',\n];\n\n// The codes OpenAI gives a failed response that a new attempt can succeed past.\nconst transientResponseErrorCodes = ['server_error', 'rate_limit_exceeded'];\n\n// This client's label on the reasoning items it keeps for replay; see `ProviderDataPart`.\nconst openAISource = 'openai';\n\n// Marks a failed call's output, which would otherwise read the same as a successful one.\nconst failedToolPrefix = 'The tool call failed: ';\n\n/**\n * OpenAI takes a flat list of items, so every part becomes its own, in order. Reasoning crosses\n * back only as the encrypted item this client kept, which OpenAI requires beside the tool calls\n * that followed it; a reasoning part is its summary, kept for narration. A call to the tool\n * offered as `apply_patch`, named `applyPatch` here, and its result go back as that tool's own\n * items.\n */\nfunction toOpenAIItems(\n\tpart: MessagePart,\n\trole: 'user' | 'assistant',\n\tapplyPatch: string | undefined,\n): ResponseInputItem[] {\n\tswitch (part.type) {\n\t\tcase 'text':\n\t\t\treturn part.text ? [{ role, content: part.text }] : [];\n\t\tcase 'toolCall':\n\t\t\tif (part.name === applyPatch) {\n\t\t\t\treturn [\n\t\t\t\t\t{\n\t\t\t\t\t\ttype: 'apply_patch_call',\n\t\t\t\t\t\tcall_id: part.id,\n\t\t\t\t\t\t// Kept as OpenAI sent it: the conversation holds the operation unchanged.\n\t\t\t\t\t\toperation: part.input as ResponseApplyPatchToolCall['operation'],\n\t\t\t\t\t\tstatus: 'completed',\n\t\t\t\t\t},\n\t\t\t\t];\n\t\t\t}\n\t\t\treturn [\n\t\t\t\t{\n\t\t\t\t\ttype: 'function_call',\n\t\t\t\t\tcall_id: part.id,\n\t\t\t\t\tname: part.name,\n\t\t\t\t\t// A string is arguments that did not parse, sent back as the model wrote them.\n\t\t\t\t\targuments: typeof part.input === 'string' ? part.input : JSON.stringify(part.input),\n\t\t\t\t},\n\t\t\t];\n\t\tcase 'toolResult':\n\t\t\tif (part.name === applyPatch) {\n\t\t\t\treturn [\n\t\t\t\t\t{\n\t\t\t\t\t\ttype: 'apply_patch_call_output',\n\t\t\t\t\t\tcall_id: part.callId,\n\t\t\t\t\t\tstatus: part.isError ? 'failed' : 'completed',\n\t\t\t\t\t\toutput: part.output,\n\t\t\t\t\t},\n\t\t\t\t];\n\t\t\t}\n\t\t\treturn [\n\t\t\t\t{\n\t\t\t\t\ttype: 'function_call_output',\n\t\t\t\t\tcall_id: part.callId,\n\t\t\t\t\t// OpenAI has no error flag, so the text says it; the conversation keeps the output as it was.\n\t\t\t\t\toutput: part.isError ? `${failedToolPrefix}${part.output}` : part.output,\n\t\t\t\t},\n\t\t\t];\n\t\tcase 'providerData':\n\t\t\t// Another provider's item would be rejected; losing it only costs that provider's reasoning.\n\t\t\treturn part.source === openAISource ? [part.data as ResponseReasoningItem] : [];\n\t\tcase 'reasoning':\n\t\t\treturn [];\n\t}\n}\n\nfunction toOpenAIInput(message: Message, applyPatch: string | undefined): ResponseInputItem[] {\n\t// A tool message holds only results, which carry no role.\n\tconst role = message.role === 'assistant' ? 'assistant' : 'user';\n\treturn message.content.flatMap(part => toOpenAIItems(part, role, applyPatch));\n}\n\n/**\n * Strict: the API constrains the model to the schema, so its arguments always parse and match.\n * A schema outside what strict mode supports fails the request with a 400. The native\n * `apply_patch` has OpenAI's own schema.\n */\nfunction toOpenAITool(definition: ToolDefinition): FunctionTool | ApplyPatchTool {\n\tconst native = nativeOf(definition);\n\tif (native === 'applyPatch') return { type: 'apply_patch' };\n\tif (native !== undefined) {\n\t\tthrow new InvalidAgentConfigError(`OpenAI has no native tool \"${native}\"`);\n\t}\n\tconst { name, description, inputSchema } = definition;\n\treturn { type: 'function', name, description, parameters: inputSchema, strict: true };\n}\n\n/** The name the tool offered as `apply_patch` has in the conversation, if one is offered. */\nfunction applyPatchName(tools: ToolDefinition[]): string | undefined {\n\treturn tools.find(tool => nativeOf(tool) === 'applyPatch')?.name;\n}\n\n/**\n * Strict mode makes arguments that do not parse rare: a response cut short mid-call ends as\n * `truncated` before any tool runs. Should one arrive, the tool gets the raw string and its\n * rejection lets the model correct itself, instead of the run failing.\n */\nfunction parseArguments(serialized: string): unknown {\n\ttry {\n\t\treturn JSON.parse(serialized);\n\t} catch {\n\t\treturn serialized;\n\t}\n}\n\nfunction describeItem(\n\titem: ResponseOutputItem,\n\tlogger: ILogger,\n\tapplyPatch: string | undefined,\n): MessagePart[] {\n\tswitch (item.type) {\n\t\tcase 'message':\n\t\t\t// A refusal is not an answer: it only sets the stop reason.\n\t\t\treturn item.content.flatMap(content =>\n\t\t\t\tcontent.type === 'output_text' ? [{ type: 'text' as const, text: content.text }] : [],\n\t\t\t);\n\t\tcase 'reasoning': {\n\t\t\t// The summary is the readable part; the encrypted item is what OpenAI reads back.\n\t\t\tconst text = item.summary.map(summary => summary.text).join('\\n');\n\t\t\tconst parts: MessagePart[] = text ? [{ type: 'reasoning', text }] : [];\n\t\t\t// Without its encrypted content the item cannot be replayed: OpenAI stores nothing.\n\t\t\tif (item.encrypted_content) {\n\t\t\t\tconst encrypted: ResponseReasoningItem = {\n\t\t\t\t\ttype: 'reasoning',\n\t\t\t\t\tid: item.id,\n\t\t\t\t\tsummary: item.summary,\n\t\t\t\t\tencrypted_content: item.encrypted_content,\n\t\t\t\t};\n\t\t\t\tparts.push({ type: 'providerData', source: openAISource, data: encrypted });\n\t\t\t}\n\t\t\treturn parts;\n\t\t}\n\t\tcase 'function_call':\n\t\t\treturn [\n\t\t\t\t{\n\t\t\t\t\ttype: 'toolCall',\n\t\t\t\t\tid: item.call_id,\n\t\t\t\t\tname: item.name,\n\t\t\t\t\tinput: parseArguments(item.arguments),\n\t\t\t\t},\n\t\t\t];\n\t\tcase 'apply_patch_call':\n\t\t\t// Only when `apply_patch` was offered; otherwise it is unexpected, as below.\n\t\t\tif (applyPatch !== undefined) {\n\t\t\t\treturn [{ type: 'toolCall', id: item.call_id, name: applyPatch, input: item.operation }];\n\t\t\t}\n\t\t\treturn unmapped(item, logger);\n\t\tdefault:\n\t\t\treturn unmapped(item, logger);\n\t}\n}\n\n/** An item this client does not map, which it reports and leaves out. */\nfunction unmapped(item: ResponseOutputItem, logger: ILogger): MessagePart[] {\n\t// No other built-in tool is offered, so any other item is output this client does not expect.\n\ttreatErrors(\n\t\t() => {\n\t\t\tlogger.warn(item, 'OpenAI returned an output item the client does not map');\n\t\t},\n\t\tclassifyHostFailure,\n\t\t'OpenAI client logger failed while mapping a response',\n\t);\n\treturn [];\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(\n\tresponse: Response,\n\tlogger: ILogger,\n\tapplyPatch: string | undefined,\n): 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, applyPatch)),\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 reasoningEffort: OpenAIReasoningEffort;\n\tprivate readonly systemPrompt: string;\n\tprivate readonly logger: ILogger;\n\n\t// `model` and `reasoningEffort` are untyped on purpose: this adapter is the one that knows\n\t// what OpenAI supports.\n\tconstructor({\n\t\tmodel,\n\t\treasoningEffort = 'high',\n\t\tsystemPrompt,\n\t\tlogger,\n\t}: {\n\t\tmodel: string;\n\t\treasoningEffort?: 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: ${openAIModels.join(', ')}`,\n\t\t\t);\n\t\t}\n\t\tif (!isOneOf(openAIReasoningEfforts, reasoningEffort)) {\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`\"${reasoningEffort}\" is not an OpenAI reasoning effort; expected one of: ${openAIReasoningEfforts.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.reasoningEffort = reasoningEffort;\n\t\tthis.systemPrompt = systemPrompt;\n\t\tthis.logger = logger;\n\t}\n\n\tsupportsNative(tool: NativeTool): boolean {\n\t\treturn tool === 'applyPatch';\n\t}\n\n\tasync send(\n\t\t{ context, tools }: { context: Message[]; tools: ToolDefinition[] },\n\t\tsignal: AbortSignal,\n\t): Promise<LLMResponse> {\n\t\tlet response: Response;\n\t\tconst applyPatch = applyPatchName(tools);\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(message => toOpenAIInput(message, applyPatch)),\n\t\t\t\t\t// Left out when empty, so a request without tools stays as it always was.\n\t\t\t\t\t...(tools.length > 0 && { tools: tools.map(toOpenAITool) }),\n\t\t\t\t\tstore: false,\n\t\t\t\t\treasoning: { effort: this.reasoningEffort, summary: 'auto' },\n\t\t\t\t\t// With `store: false` this is the only way to send the reasoning back.\n\t\t\t\t\tinclude: ['reasoning.encrypted_content'],\n\t\t\t\t},\n\t\t\t\t{ signal },\n\t\t\t);\n\t\t} catch (error) {\n\t\t\t// The SDK's abort error is named 'Error', so it would pass for a provider failure.\n\t\t\tsignal.throwIfAborted();\n\t\t\tthrow classifyOpenAIFailure(error);\n\t\t}\n\n\t\treturn treatErrors(\n\t\t\t() => toLLMResponse(response, this.logger, applyPatch),\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

@@ -3,8 +3,10 @@ import type { ReviewerDecision } from './reviewerDecision.ts';
 import type { Validator } from '../interface/validator.ts';
 import type { ILogger } from '#src/shared/domain/logger';
 export declare const plannerInstructions = "You are the planner agent. Read the relevant repository code and create a concise, engineering-grade plan for the executor. Cover the requested behaviour, structural issues, affected files or symbols, API contracts and boundaries, important failure paths, and meaningful tests. Keep the plan focused on the current request and do not require an architecture redesign yet.";
-export declare const executorInstructions = "You are the executor agent. Read the relevant repository code and implement the original user request according to the planner's current plan. Preserve contracts and boundaries, handle important failure paths, keep the change focused, and add or update meaningful tests. Inspect and change the code; do not merely describe what should be done.";
-export declare const reviewerInstructions = "You are the reviewer agent. Independently inspect the repository, the current diff, and relevant surrounding code; do not rely on the executor's narrative. Evaluate the original request first; plan compliance is secondary and provides supporting context. Check correctness and logic errors, important failure paths, unused or artificial abstractions, races or shared state, API contract breaks, boundary violations, regressions, scope, and test quality. If it is correct, respond with JSON only: {\"decision\":\"approved\"}. If it is not correct, respond with JSON only: {\"decision\":\"rejected\",\"feedback\":\"list concrete, prioritized findings and the required direction for each\"}. The feedback must contain actionable engineering findings, not a general summary.";
+export declare const executorInstructions = "You are the executor agent. Read the relevant repository code and implement the original user request according to the planner's current plan. Preserve contracts and boundaries, handle important failure paths, keep the change focused, and add or update meaningful tests. Inspect and change the code; do not merely describe what should be done. Before you finish, run the project's tests and checks, and finish only when they pass. Once they pass and the plan is carried out, finish at once: do not add work beyond the plan, because the reviewer asks for what is missing. If you cannot make them pass, finish by saying which ones fail and why.";
+export declare const executorInstructionsWithoutCommands = "You are the executor agent. Read the relevant repository code and implement the original user request according to the planner's current plan. Preserve contracts and boundaries, handle important failure paths, keep the change focused, and add or update meaningful tests. Inspect and change the code; do not merely describe what should be done. You have no command that runs the project's tests or checks, so you cannot run them. Once the plan is carried out, finish at once: do not add work beyond the plan, because the reviewer asks for what is missing. Say in your summary that the tests and checks were not run.";
+export declare const reviewerInstructions = "You are the reviewer agent. The executor has already run the tests and checks, and finished only when they passed: do not run them again. Review the code. Independently inspect the repository, the current diff, and relevant surrounding code; do not rely on the executor's narrative. Judge it against the original request, and any issue it names, first: check that the plan is right for it, and that the change fulfills it. Check correctness and logic errors, important failure paths, unused or artificial abstractions, races or shared state, API contract breaks, boundary violations, regressions, scope, and test quality. If it is correct, respond with JSON only: {\"decision\":\"approved\"}. If it is not correct, respond with JSON only: {\"decision\":\"rejected\",\"feedback\":\"list concrete, prioritized findings and the required direction for each\"}. The feedback must contain actionable engineering findings, not a general summary.";
+export declare const reviewerInstructionsWithoutCommands = "You are the reviewer agent. Nobody has run the tests or checks: the executor had no command for them. Judge from the code whether the tests it added or changed would pass, and do not approve a change whose tests could not. Review the code. Independently inspect the repository, the current diff, and relevant surrounding code; do not rely on the executor's narrative. Judge it against the original request, and any issue it names, first: check that the plan is right for it, and that the change fulfills it. Check correctness and logic errors, important failure paths, unused or artificial abstractions, races or shared state, API contract breaks, boundary violations, regressions, scope, and test quality. If it is correct, respond with JSON only: {\"decision\":\"approved\"}. If it is not correct, respond with JSON only: {\"decision\":\"rejected\",\"feedback\":\"list concrete, prioritized findings and the required direction for each\"}. The feedback must contain actionable engineering findings, not a general summary.";
 export declare class OrchestratorAgent implements Agent {
     private plannerAgent;
     private executorAgent;
@@ -20,7 +22,14 @@ export declare class OrchestratorAgent implements Agent {
         maxAttempts?: number;
         logger: ILogger;
     });
+    /**
+     * Every role, round and retry of one run shares one run context, so what the executor writes
+     * is one run of the workspace's history, and the planner's and reviewer's turns are tied to
+     * it. The orchestrator ends that context with the run, unless an outer agent owns it.
+     */
     run(prompt: string, options: RunOptions): Promise<AgentResponse>;
+    /** The run's records end here, but its result stands: what it changed is already on disk. */
+    private finishContext;
     private runRounds;
     private getReviewerDecision;
     private warn;

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

@@ -1,12 +1,24 @@
 import { addTokens } from '#src/shared/domain/tokens';
 import { RecoverableError, UnrecoverableError, withSpentTokens } from '#src/shared/domain/errors';
-import { classifyHostFailure, treatErrors } from '#src/shared/domain/providerFailure';
+import { classifyHostFailure, describeFailure, treatErrors, } from '#src/shared/domain/providerFailure';
+import { RunContext, runContextOf, withRunContext, } from '#src/agent/domain/runContext';
+import { isAbortError } from '#src/shared/domain/isAbortError';
 // Each role's standing instructions, kept apart from the data of a round so an agent with a
 // system prompt can hold them there instead of receiving them again in every prompt. An agent
 // without one gets them at the head of each prompt; see `InstructedAgent`.
 export const plannerInstructions = 'You are the planner agent. Read the relevant repository code and create a concise, engineering-grade plan for the executor. Cover the requested behaviour, structural issues, affected files or symbols, API contracts and boundaries, important failure paths, and meaningful tests. Keep the plan focused on the current request and do not require an architecture redesign yet.';
-export const executorInstructions = "You are the executor agent. Read the relevant repository code and implement the original user request according to the planner's current plan. Preserve contracts and boundaries, handle important failure paths, keep the change focused, and add or update meaningful tests. Inspect and change the code; do not merely describe what should be done.";
-export const reviewerInstructions = 'You are the reviewer agent. Independently inspect the repository, the current diff, and relevant surrounding code; do not rely on the executor\'s narrative. Evaluate the original request first; plan compliance is secondary and provides supporting context. Check correctness and logic errors, important failure paths, unused or artificial abstractions, races or shared state, API contract breaks, boundary violations, regressions, scope, and test quality. If it is correct, respond with JSON only: {"decision":"approved"}. If it is not correct, respond with JSON only: {"decision":"rejected","feedback":"list concrete, prioritized findings and the required direction for each"}. The feedback must contain actionable engineering findings, not a general summary.';
+const executorTask = "You are the executor agent. Read the relevant repository code and implement the original user request according to the planner's current plan. Preserve contracts and boundaries, handle important failure paths, keep the change focused, and add or update meaningful tests. Inspect and change the code; do not merely describe what should be done.";
+// For an executor that can run the project's tests and checks: an Agent SDK one, which has a
+// shell, or a model-API one given commands. The harness cannot tell which command runs the
+// tests, so any command counts.
+export const executorInstructions = `${executorTask} Before you finish, run the project's tests and checks, and finish only when they pass. Once they pass and the plan is carried out, finish at once: do not add work beyond the plan, because the reviewer asks for what is missing. If you cannot make them pass, finish by saying which ones fail and why.`;
+// For a model-API executor with no command, which cannot run anything: told so, it neither
+// spends its steps looking for a way to test nor claims a result it could not have.
+export const executorInstructionsWithoutCommands = `${executorTask} You have no command that runs the project's tests or checks, so you cannot run them. Once the plan is carried out, finish at once: do not add work beyond the plan, because the reviewer asks for what is missing. Say in your summary that the tests and checks were not run.`;
+const reviewerTask = 'Review the code. Independently inspect the repository, the current diff, and relevant surrounding code; do not rely on the executor\'s narrative. Judge it against the original request, and any issue it names, first: check that the plan is right for it, and that the change fulfills it. Check correctness and logic errors, important failure paths, unused or artificial abstractions, races or shared state, API contract breaks, boundary violations, regressions, scope, and test quality. If it is correct, respond with JSON only: {"decision":"approved"}. If it is not correct, respond with JSON only: {"decision":"rejected","feedback":"list concrete, prioritized findings and the required direction for each"}. The feedback must contain actionable engineering findings, not a general summary.';
+// Each goes with the executor instructions of the same name.
+export const reviewerInstructions = `You are the reviewer agent. The executor has already run the tests and checks, and finished only when they passed: do not run them again. ${reviewerTask}`;
+export const reviewerInstructionsWithoutCommands = `You are the reviewer agent. Nobody has run the tests or checks: the executor had no command for them. Judge from the code whether the tests it added or changed would pass, and do not approve a change whose tests could not. ${reviewerTask}`;
 const getPlannerPrompt = (userPrompt, previousFailureReason) => {
     const feedback = previousFailureReason
         ? `\n\nFeedback from the previous attempt:\n---\n${previousFailureReason}\n---`
@@ -99,24 +111,51 @@ export class OrchestratorAgent {
         this.maxAttempts = maxAttempts;
         this.logger = logger;
     }
+    /**
+     * Every role, round and retry of one run shares one run context, so what the executor writes
+     * is one run of the workspace's history, and the planner's and reviewer's turns are tied to
+     * it. The orchestrator ends that context with the run, unless an outer agent owns it.
+     */
     async run(prompt, options) {
         const start = Date.now();
         // Per invocation, not per instance: the CLI keeps one orchestrator for a whole session.
         const totals = { tokens: undefined, unreported: false };
+        const inherited = runContextOf(options);
+        const context = inherited ?? new RunContext();
+        let end = 'failed';
         try {
-            await this.runRounds(prompt, options, totals);
+            await this.runRounds(prompt, inherited ? options : withRunContext(options, context), totals);
+            end = 'completed';
         }
         catch (error) {
+            if (isAbortError(error))
+                end = 'cancelled';
             // A failing role carries its own tokens; the roles before it are in the totals.
             throw withSpentTokens(error, totals.tokens, totals.unreported);
         }
+        finally {
+            if (!inherited)
+                await this.finishContext(context, end);
+        }
+        // Read once the run ended: a run that changed nothing left the history then.
+        const { runId } = context.historyRun;
         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,
+            ...(!inherited && runId ? { runId } : {}),
         };
     }
+    /** The run's records end here, but its result stands: what it changed is already on disk. */
+    async finishContext(context, end) {
+        try {
+            await context.finish(end);
+        }
+        catch (error) {
+            this.warn(`The run could not close its records: ${describeFailure(error)}`);
+        }
+    }
     async runRounds(prompt, options, totals) {
         let lastFailureReason;
         for (let attempt = 1; attempt <= this.maxAttempts; attempt++) {