@mikode13/harness 5.0.0 → 5.2.0

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;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"]}
+{"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;IAC/C,YAAY,CAAS;IACrB,kBAAkB,GAAG,YAAY,CAAC;IAC1B,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,SAAS,EAAE;gBAC/D,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\treadonly systemPrompt: string;\n\treadonly providerDataSource = openAISource;\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: 'concise' },\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

@@ -2,11 +2,11 @@ import type { Agent, AgentResponse, RunOptions } from '#src/agent/domain/agent';
 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. 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 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. Communicate tersely and professionally. In reasoning updates, state only the next concrete action or a verified finding, in one short sentence. Do not narrate routine intentions, uncertainty, self-reflection, encouragement, or a desire to be thorough. Do not restate tool results. Once you decide what to do, act immediately. Use only the detail required by your role in the final response.";
+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. Communicate tersely and professionally. In reasoning updates, state only the next concrete action or a verified finding, in one short sentence. Do not narrate routine intentions, uncertainty, self-reflection, encouragement, or a desire to be thorough. Do not restate tool results. Once you decide what to do, act immediately. Use only the detail required by your role in the final response. 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. Communicate tersely and professionally. In reasoning updates, state only the next concrete action or a verified finding, in one short sentence. Do not narrate routine intentions, uncertainty, self-reflection, encouragement, or a desire to be thorough. Do not restate tool results. Once you decide what to do, act immediately. Use only the detail required by your role in the final response. 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. Communicate tersely and professionally. In reasoning updates, state only the next concrete action or a verified finding, in one short sentence. Do not narrate routine intentions, uncertainty, self-reflection, encouragement, or a desire to be thorough. Do not restate tool results. Once you decide what to do, act immediately. Use only the detail required by your role in the final response.";
+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. Communicate tersely and professionally. In reasoning updates, state only the next concrete action or a verified finding, in one short sentence. Do not narrate routine intentions, uncertainty, self-reflection, encouragement, or a desire to be thorough. Do not restate tool results. Once you decide what to do, act immediately. Use only the detail required by your role in the final response.";
 export declare class OrchestratorAgent implements Agent {
     private plannerAgent;
     private executorAgent;
@@ -33,4 +33,9 @@ export declare class OrchestratorAgent implements Agent {
     private runRounds;
     private getReviewerDecision;
     private warn;
+    /**
+     * Keep the options object extensible and the private run context intact while making the
+     * orchestrator's attribution authoritative over anything a nested agent supplied.
+     */
+    private optionsForRole;
 }

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

@@ -6,8 +6,9 @@ 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.';
-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.";
+const conciseCommunication = 'Communicate tersely and professionally. In reasoning updates, state only the next concrete action or a verified finding, in one short sentence. Do not narrate routine intentions, uncertainty, self-reflection, encouragement, or a desire to be thorough. Do not restate tool results. Once you decide what to do, act immediately. Use only the detail required by your role in the final response.';
+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. ${conciseCommunication}`;
+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. ${conciseCommunication}`;
 // For an executor that can run the project's tests and checks: an Agent SDK one, which has a
 // shell, or a model-API one given commands. The harness cannot tell which command runs the
 // tests, so any command counts.
@@ -15,7 +16,7 @@ export const executorInstructions = `${executorTask} Before you finish, run the
 // 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.';
+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. ${conciseCommunication}`;
 // 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}`;
@@ -160,7 +161,7 @@ export class OrchestratorAgent {
         let lastFailureReason;
         for (let attempt = 1; attempt <= this.maxAttempts; attempt++) {
             const isLastAttempt = attempt === this.maxAttempts;
-            const plannerResponse = await this.plannerAgent.run(getPlannerPrompt(prompt, lastFailureReason), options);
+            const plannerResponse = await this.plannerAgent.run(getPlannerPrompt(prompt, lastFailureReason), this.optionsForRole(options, 'planner'));
             addToTotals(totals, plannerResponse);
             if (!plannerResponse.response) {
                 lastFailureReason = 'The planner produced no response.';
@@ -170,7 +171,7 @@ export class OrchestratorAgent {
                 this.warn(`Attempt ${String(attempt)}/${String(this.maxAttempts)}: the planner produced no response; starting another round`);
                 continue;
             }
-            const executorResponse = await this.executorAgent.run(getExecutorPrompt(prompt, plannerResponse.response), options);
+            const executorResponse = await this.executorAgent.run(getExecutorPrompt(prompt, plannerResponse.response), this.optionsForRole(options, 'executor'));
             addToTotals(totals, executorResponse);
             if (!executorResponse.response) {
                 lastFailureReason = 'The executor produced no response.';
@@ -199,7 +200,7 @@ export class OrchestratorAgent {
     async getReviewerDecision(userPrompt, plannerPrompt, executorResult, options, totals) {
         let parseFailureReason;
         for (let attempt = 1; attempt <= this.maxAttempts; attempt++) {
-            const reviewerResponse = await this.reviewerAgent.run(getReviewerPrompt(userPrompt, plannerPrompt, executorResult, parseFailureReason), options);
+            const reviewerResponse = await this.reviewerAgent.run(getReviewerPrompt(userPrompt, plannerPrompt, executorResult, parseFailureReason), this.optionsForRole(options, 'reviewer'));
             addToTotals(totals, reviewerResponse);
             try {
                 return parseReviewerDecision(reviewerResponse, this.reviewerDecisionValidator);
@@ -223,5 +224,22 @@ export class OrchestratorAgent {
             this.logger.warn(...args);
         }, classifyHostFailure, 'Orchestrator logger failed while reporting a retry');
     }
+    /**
+     * Keep the options object extensible and the private run context intact while making the
+     * orchestrator's attribution authoritative over anything a nested agent supplied.
+     */
+    optionsForRole(options, agent) {
+        const onProgress = options.onProgress;
+        return {
+            ...options,
+            onProgress: onProgress
+                ? (event) => {
+                    treatErrors(() => {
+                        onProgress({ ...event, agent });
+                    }, classifyHostFailure, 'Orchestrator progress callback failed');
+                }
+                : undefined,
+        };
+    }
 }
 //# sourceMappingURL=orchestratorAgent.js.map

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

@@ -1 +1 @@
-{"version":3,"file":"orchestratorAgent.js","sourceRoot":"","sources":["../../../../src/orchestration/domain/model/orchestratorAgent.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,SAAS,EAAe,MAAM,2BAA2B,CAAC;AACnE,OAAO,EAAE,gBAAgB,EAAE,kBAAkB,EAAE,eAAe,EAAE,MAAM,2BAA2B,CAAC;AAIlG,OAAO,EACN,mBAAmB,EACnB,eAAe,EACf,WAAW,GACX,MAAM,oCAAoC,CAAC;AAC5C,OAAO,EACN,UAAU,EAEV,YAAY,EACZ,cAAc,GACd,MAAM,8BAA8B,CAAC;AACtC,OAAO,EAAE,YAAY,EAAE,MAAM,iCAAiC,CAAC;AAE/D,4FAA4F;AAC5F,8FAA8F;AAC9F,2EAA2E;AAC3E,MAAM,CAAC,MAAM,mBAAmB,GAC/B,qXAAqX,CAAC;AAEvX,MAAM,YAAY,GACjB,yVAAyV,CAAC;AAE3V,6FAA6F;AAC7F,2FAA2F;AAC3F,gCAAgC;AAChC,MAAM,CAAC,MAAM,oBAAoB,GAAG,GAAG,YAAY,6SAA6S,CAAC;AAEjW,2FAA2F;AAC3F,oFAAoF;AACpF,MAAM,CAAC,MAAM,mCAAmC,GAAG,GAAG,YAAY,iRAAiR,CAAC;AAEpV,MAAM,YAAY,GACjB,uxBAAuxB,CAAC;AAEzxB,6DAA6D;AAC7D,MAAM,CAAC,MAAM,oBAAoB,GAAG,6IAA6I,YAAY,EAAE,CAAC;AAEhM,MAAM,CAAC,MAAM,mCAAmC,GAAG,kOAAkO,YAAY,EAAE,CAAC;AAEpS,MAAM,gBAAgB,GAAG,CAAC,UAAkB,EAAE,qBAA8B,EAAE,EAAE;IAC/E,MAAM,QAAQ,GAAG,qBAAqB;QACrC,CAAC,CAAC,iDAAiD,qBAAqB,OAAO;QAC/E,CAAC,CAAC,EAAE,CAAC;IAEN,OAAO;;EAEN,UAAU;KACP,QAAQ,EAAE,CAAC;AAChB,CAAC,CAAC;AAEF,MAAM,iBAAiB,GAAG,CAAC,UAAkB,EAAE,aAAqB,EAAE,EAAE,CACvE;;EAEC,UAAU;;;;;EAKV,aAAa;IACX,CAAC;AAEL,MAAM,iBAAiB,GAAG,CACzB,UAAkB,EAClB,aAAqB,EACrB,cAAsB,EACtB,kBAA2B,EAC1B,EAAE;IACH,MAAM,WAAW,GAAG,kBAAkB;QACrC,CAAC,CAAC,iDAAiD,kBAAkB,6GAA6G;QAClL,CAAC,CAAC,EAAE,CAAC;IAEN,OAAO;;EAEN,UAAU;;;;;EAKV,aAAa;;;;;EAKb,cAAc;KACX,WAAW,EAAE,CAAC;AACnB,CAAC,CAAC;AAQF,4FAA4F;AAC5F,SAAS,WAAW,CAAC,MAAiB,EAAE,QAAuB;IAC9D,IAAI,CAAC,QAAQ,CAAC,MAAM;QAAE,MAAM,CAAC,UAAU,GAAG,IAAI,CAAC;IAC/C,MAAM,CAAC,MAAM,GAAG,SAAS,CAAC,MAAM,CAAC,MAAM,EAAE,QAAQ,CAAC,MAAM,CAAC,CAAC;AAC3D,CAAC;AAED,uFAAuF;AACvF,SAAS,SAAS,CAAC,MAAiB;IACnC,OAAO,MAAM,CAAC,UAAU,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC;AACtD,CAAC;AAED,SAAS,cAAc,CAAC,IAAY;IACnC,MAAM,KAAK,GAAG,mCAAmC,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC;IACpE,OAAO,KAAK,EAAE,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC;AAC3B,CAAC;AAED,SAAS,qBAAqB,CAC7B,QAAuB,EACvB,yBAAsD;IAEtD,IAAI,CAAC,QAAQ,CAAC,QAAQ,EAAE,CAAC;QACxB,MAAM,IAAI,gBAAgB,CAAC,6DAA6D,EAAE;YACzF,KAAK,EAAE,+BAA+B;SACtC,CAAC,CAAC;IACJ,CAAC;IAED,IAAI,cAAuB,CAAC;IAC5B,IAAI,CAAC;QACJ,cAAc,GAAG,IAAI,CAAC,KAAK,CAAC,cAAc,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC,CAAC;IAChE,CAAC;IAAC,MAAM,CAAC;QACR,MAAM,IAAI,gBAAgB,CAAC,uCAAuC,EAAE;YACnE,KAAK,EAAE,uCAAuC;SAC9C,CAAC,CAAC;IACJ,CAAC;IAED,MAAM,QAAQ,GAAG,yBAAyB,CAAC,QAAQ,CAAC,cAAc,CAAC,CAAC;IACpE,IAAI,CAAC,QAAQ,EAAE,CAAC;QACf,MAAM,IAAI,gBAAgB,CAAC,uCAAuC,EAAE;YACnE,KAAK,EAAE,sDAAsD;SAC7D,CAAC,CAAC;IACJ,CAAC;IAED,OAAO,QAAQ,CAAC;AACjB,CAAC;AAED,MAAM,OAAO,iBAAiB;IACrB,YAAY,CAAQ;IACpB,aAAa,CAAQ;IACrB,aAAa,CAAQ;IACrB,yBAAyB,CAA8B;IACvD,WAAW,CAAS;IACpB,MAAM,CAAU;IAExB,YAAY,EACX,YAAY,EACZ,aAAa,EACb,aAAa,EACb,yBAAyB,EACzB,WAAW,GAAG,CAAC,EACf,MAAM,GAQN;QACA,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,WAAW,CAAC,IAAI,WAAW,GAAG,CAAC,EAAE,CAAC;YACvD,MAAM,IAAI,UAAU,CAAC,wCAAwC,CAAC,CAAC;QAChE,CAAC;QAED,IAAI,CAAC,YAAY,GAAG,YAAY,CAAC;QACjC,IAAI,CAAC,aAAa,GAAG,aAAa,CAAC;QACnC,IAAI,CAAC,aAAa,GAAG,aAAa,CAAC;QACnC,IAAI,CAAC,yBAAyB,GAAG,yBAAyB,CAAC;QAC3D,IAAI,CAAC,WAAW,GAAG,WAAW,CAAC;QAC/B,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IACtB,CAAC;IAED;;;;OAIG;IACH,KAAK,CAAC,GAAG,CAAC,MAAc,EAAE,OAAmB;QAC5C,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QACzB,wFAAwF;QACxF,MAAM,MAAM,GAAc,EAAE,MAAM,EAAE,SAAS,EAAE,UAAU,EAAE,KAAK,EAAE,CAAC;QACnE,MAAM,SAAS,GAAG,YAAY,CAAC,OAAO,CAAC,CAAC;QACxC,MAAM,OAAO,GAAG,SAAS,IAAI,IAAI,UAAU,EAAE,CAAC;QAC9C,IAAI,GAAG,GAAW,QAAQ,CAAC;QAE3B,IAAI,CAAC;YACJ,MAAM,IAAI,CAAC,SAAS,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,cAAc,CAAC,OAAO,EAAE,OAAO,CAAC,EAAE,MAAM,CAAC,CAAC;YAC7F,GAAG,GAAG,WAAW,CAAC;QACnB,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,IAAI,YAAY,CAAC,KAAK,CAAC;gBAAE,GAAG,GAAG,WAAW,CAAC;YAC3C,gFAAgF;YAChF,MAAM,eAAe,CAAC,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,UAAU,CAAC,CAAC;QAChE,CAAC;gBAAS,CAAC;YACV,IAAI,CAAC,SAAS;gBAAE,MAAM,IAAI,CAAC,aAAa,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC;QACxD,CAAC;QAED,6EAA6E;QAC7E,MAAM,EAAE,KAAK,EAAE,GAAG,OAAO,CAAC,UAAU,CAAC;QACrC,OAAO;YACN,QAAQ,EAAE,sBAAsB;YAChC,MAAM,EAAE,SAAS,CAAC,MAAM,CAAC;YACzB,sEAAsE;YACtE,QAAQ,EAAE,CAAC,IAAI,CAAC,GAAG,EAAE,GAAG,KAAK,CAAC,GAAG,IAAI;YACrC,GAAG,CAAC,CAAC,SAAS,IAAI,KAAK,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;SACzC,CAAC;IACH,CAAC;IAED,6FAA6F;IACrF,KAAK,CAAC,aAAa,CAAC,OAAmB,EAAE,GAAW;QAC3D,IAAI,CAAC;YACJ,MAAM,OAAO,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;QAC3B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,IAAI,CAAC,IAAI,CAAC,wCAAwC,eAAe,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;QAC7E,CAAC;IACF,CAAC;IAEO,KAAK,CAAC,SAAS,CAAC,MAAc,EAAE,OAAmB,EAAE,MAAiB;QAC7E,IAAI,iBAAqC,CAAC;QAE1C,KAAK,IAAI,OAAO,GAAG,CAAC,EAAE,OAAO,IAAI,IAAI,CAAC,WAAW,EAAE,OAAO,EAAE,EAAE,CAAC;YAC9D,MAAM,aAAa,GAAG,OAAO,KAAK,IAAI,CAAC,WAAW,CAAC;YAEnD,MAAM,eAAe,GAAG,MAAM,IAAI,CAAC,YAAY,CAAC,GAAG,CAClD,gBAAgB,CAAC,MAAM,EAAE,iBAAiB,CAAC,EAC3C,OAAO,CACP,CAAC;YACF,WAAW,CAAC,MAAM,EAAE,eAAe,CAAC,CAAC;YAErC,IAAI,CAAC,eAAe,CAAC,QAAQ,EAAE,CAAC;gBAC/B,iBAAiB,GAAG,mCAAmC,CAAC;gBACxD,IAAI,aAAa,EAAE,CAAC;oBACnB,MAAM,IAAI,kBAAkB,CAAC,wBAAwB,EAAE,EAAE,KAAK,EAAE,iBAAiB,EAAE,CAAC,CAAC;gBACtF,CAAC;gBACD,IAAI,CAAC,IAAI,CACR,WAAW,MAAM,CAAC,OAAO,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,4DAA4D,CAClH,CAAC;gBACF,SAAS;YACV,CAAC;YAED,MAAM,gBAAgB,GAAG,MAAM,IAAI,CAAC,aAAa,CAAC,GAAG,CACpD,iBAAiB,CAAC,MAAM,EAAE,eAAe,CAAC,QAAQ,CAAC,EACnD,OAAO,CACP,CAAC;YACF,WAAW,CAAC,MAAM,EAAE,gBAAgB,CAAC,CAAC;YAEtC,IAAI,CAAC,gBAAgB,CAAC,QAAQ,EAAE,CAAC;gBAChC,iBAAiB,GAAG,oCAAoC,CAAC;gBACzD,IAAI,aAAa,EAAE,CAAC;oBACnB,MAAM,IAAI,kBAAkB,CAAC,wBAAwB,EAAE,EAAE,KAAK,EAAE,iBAAiB,EAAE,CAAC,CAAC;gBACtF,CAAC;gBAED,IAAI,CAAC,IAAI,CACR,WAAW,MAAM,CAAC,OAAO,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,6DAA6D,CACnH,CAAC;gBACF,SAAS;YACV,CAAC;YAED,MAAM,gBAAgB,GAAG,MAAM,IAAI,CAAC,mBAAmB,CACtD,MAAM,EACN,eAAe,CAAC,QAAQ,EACxB,gBAAgB,CAAC,QAAQ,EACzB,OAAO,EACP,MAAM,CACN,CAAC;YAEF,IAAI,gBAAgB,CAAC,QAAQ,KAAK,UAAU;gBAAE,OAAO;YAErD,iBAAiB,GAAG,gBAAgB,CAAC,QAAQ,CAAC;YAC9C,IAAI,aAAa,EAAE,CAAC;gBACnB,MAAM,IAAI,kBAAkB,CAAC,wBAAwB,EAAE,EAAE,KAAK,EAAE,iBAAiB,EAAE,CAAC,CAAC;YACtF,CAAC;YAED,IAAI,CAAC,IAAI,CACR,WAAW,MAAM,CAAC,OAAO,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,uEAAuE,EAC7H,iBAAiB,CACjB,CAAC;QACH,CAAC;QAED,MAAM,IAAI,kBAAkB,CAAC,wBAAwB,EAAE;YACtD,KAAK,EAAE,iBAAiB,IAAI,kBAAkB;SAC9C,CAAC,CAAC;IACJ,CAAC;IAED,6EAA6E;IAC7E,6EAA6E;IAC7E,6CAA6C;IACrC,KAAK,CAAC,mBAAmB,CAChC,UAAkB,EAClB,aAAqB,EACrB,cAAsB,EACtB,OAAmB,EACnB,MAAiB;QAEjB,IAAI,kBAAsC,CAAC;QAE3C,KAAK,IAAI,OAAO,GAAG,CAAC,EAAE,OAAO,IAAI,IAAI,CAAC,WAAW,EAAE,OAAO,EAAE,EAAE,CAAC;YAC9D,MAAM,gBAAgB,GAAG,MAAM,IAAI,CAAC,aAAa,CAAC,GAAG,CACpD,iBAAiB,CAAC,UAAU,EAAE,aAAa,EAAE,cAAc,EAAE,kBAAkB,CAAC,EAChF,OAAO,CACP,CAAC;YAEF,WAAW,CAAC,MAAM,EAAE,gBAAgB,CAAC,CAAC;YAEtC,IAAI,CAAC;gBACJ,OAAO,qBAAqB,CAAC,gBAAgB,EAAE,IAAI,CAAC,yBAAyB,CAAC,CAAC;YAChF,CAAC;YAAC,OAAO,KAAK,EAAE,CAAC;gBAChB,IAAI,CAAC,CAAC,KAAK,YAAY,gBAAgB,CAAC;oBAAE,MAAM,KAAK,CAAC;gBAEtD,kBAAkB,GAAG,KAAK,CAAC,KAAK,CAAC;gBACjC,IAAI,OAAO,KAAK,IAAI,CAAC,WAAW,EAAE,CAAC;oBAClC,MAAM,IAAI,kBAAkB,CAAC,wBAAwB,EAAE,EAAE,KAAK,EAAE,kBAAkB,EAAE,CAAC,CAAC;gBACvF,CAAC;gBAED,IAAI,CAAC,IAAI,CACR,qBAAqB,MAAM,CAAC,OAAO,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,0CAA0C,EAC1G,kBAAkB,CAClB,CAAC;YACH,CAAC;QACF,CAAC;QAED,MAAM,IAAI,kBAAkB,CAAC,wBAAwB,EAAE;YACtD,KAAK,EAAE,kBAAkB,IAAI,kBAAkB;SAC/C,CAAC,CAAC;IACJ,CAAC;IAEO,IAAI,CAAC,GAAG,IAAe;QAC9B,WAAW,CACV,GAAG,EAAE;YACJ,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,IAAI,CAAC,CAAC;QAC3B,CAAC,EACD,mBAAmB,EACnB,oDAAoD,CACpD,CAAC;IACH,CAAC;CACD","sourcesContent":["import type { Agent, AgentResponse, RunOptions } from '#src/agent/domain/agent';\nimport { addTokens, type Tokens } from '#src/shared/domain/tokens';\nimport { RecoverableError, UnrecoverableError, withSpentTokens } from '#src/shared/domain/errors';\nimport type { ReviewerDecision } from './reviewerDecision.ts';\nimport type { Validator } from '../interface/validator.ts';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport {\n\tclassifyHostFailure,\n\tdescribeFailure,\n\ttreatErrors,\n} from '#src/shared/domain/providerFailure';\nimport {\n\tRunContext,\n\ttype RunEnd,\n\trunContextOf,\n\twithRunContext,\n} from '#src/agent/domain/runContext';\nimport { isAbortError } from '#src/shared/domain/isAbortError';\n\n// Each role's standing instructions, kept apart from the data of a round so an agent with a\n// system prompt can hold them there instead of receiving them again in every prompt. An agent\n// without one gets them at the head of each prompt; see `InstructedAgent`.\nexport const plannerInstructions =\n\t'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.';\n\nconst executorTask =\n\t\"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.\";\n\n// For an executor that can run the project's tests and checks: an Agent SDK one, which has a\n// shell, or a model-API one given commands. The harness cannot tell which command runs the\n// tests, so any command counts.\nexport const executorInstructions = `${executorTask} Before you finish, run the project's tests and checks, and finish only when they pass. Once they pass and the plan is carried out, finish at once: do not add work beyond the plan, because the reviewer asks for what is missing. If you cannot make them pass, finish by saying which ones fail and why.`;\n\n// For a model-API executor with no command, which cannot run anything: told so, it neither\n// spends its steps looking for a way to test nor claims a result it could not have.\nexport const executorInstructionsWithoutCommands = `${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.`;\n\nconst reviewerTask =\n\t'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.';\n\n// Each goes with the executor instructions of the same name.\nexport 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}`;\n\nexport 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. ${reviewerTask}`;\n\nconst getPlannerPrompt = (userPrompt: string, previousFailureReason?: string) => {\n\tconst feedback = previousFailureReason\n\t\t? `\\n\\nFeedback from the previous attempt:\\n---\\n${previousFailureReason}\\n---`\n\t\t: '';\n\n\treturn `Original user request:\n---\n${userPrompt}\n---${feedback}`;\n};\n\nconst getExecutorPrompt = (userPrompt: string, plannerPrompt: string) =>\n\t`Original user request:\n---\n${userPrompt}\n---\n\nCurrent implementation plan:\n---\n${plannerPrompt}\n---`;\n\nconst getReviewerPrompt = (\n\tuserPrompt: string,\n\tplannerPrompt: string,\n\texecutorResult: string,\n\tparseFailureReason?: string,\n) => {\n\tconst retryNotice = parseFailureReason\n\t\t? `\\n\\nYour previous response could not be used: ${parseFailureReason} Respond with JSON only, matching the schema exactly, with no surrounding text and no markdown code fences.`\n\t\t: '';\n\n\treturn `Original user request:\n---\n${userPrompt}\n---\n\nPlanner's current plan:\n---\n${plannerPrompt}\n---\n\nExecutor response (context only; verify it independently):\n---\n${executorResult}\n---${retryNotice}`;\n};\n\ninterface RunTotals {\n\ttokens: Tokens | undefined;\n\t/** A role answered without usage, so the run's total cannot be known. */\n\tunreported: boolean;\n}\n\n// Every role's response counts, even an empty one that starts another round: it was billed.\nfunction addToTotals(totals: RunTotals, response: AgentResponse): void {\n\tif (!response.tokens) totals.unreported = true;\n\ttotals.tokens = addTokens(totals.tokens, response.tokens);\n}\n\n/** Unknown rather than partial: a sum missing a billed call must not look complete. */\nfunction runTokens(totals: RunTotals): Tokens | undefined {\n\treturn totals.unreported ? undefined : totals.tokens;\n}\n\nfunction stripCodeFence(text: string): string {\n\tconst match = /^```(?:json)?\\s*([\\s\\S]*?)\\s*```$/.exec(text.trim());\n\treturn match?.[1] ?? text;\n}\n\nfunction parseReviewerDecision(\n\tresponse: AgentResponse,\n\treviewerDecisionValidator: Validator<ReviewerDecision>,\n): ReviewerDecision {\n\tif (!response.response) {\n\t\tthrow new RecoverableError(\"There's no decision from the reviewer, something went wrong\", {\n\t\t\tcause: 'Reviewer decision is missing.',\n\t\t});\n\t}\n\n\tlet parsedResponse: unknown;\n\ttry {\n\t\tparsedResponse = JSON.parse(stripCodeFence(response.response));\n\t} catch {\n\t\tthrow new RecoverableError('Reviewer returned an invalid decision', {\n\t\t\tcause: 'Reviewer response must be valid JSON.',\n\t\t});\n\t}\n\n\tconst decision = reviewerDecisionValidator.validate(parsedResponse);\n\tif (!decision) {\n\t\tthrow new RecoverableError('Reviewer returned an invalid decision', {\n\t\t\tcause: 'Reviewer response did not match the decision schema.',\n\t\t});\n\t}\n\n\treturn decision;\n}\n\nexport class OrchestratorAgent implements Agent {\n\tprivate plannerAgent: Agent;\n\tprivate executorAgent: Agent;\n\tprivate reviewerAgent: Agent;\n\tprivate reviewerDecisionValidator: Validator<ReviewerDecision>;\n\tprivate maxAttempts: number;\n\tprivate logger: ILogger;\n\n\tconstructor({\n\t\tplannerAgent,\n\t\texecutorAgent,\n\t\treviewerAgent,\n\t\treviewerDecisionValidator,\n\t\tmaxAttempts = 3,\n\t\tlogger,\n\t}: {\n\t\tplannerAgent: Agent;\n\t\texecutorAgent: Agent;\n\t\treviewerAgent: Agent;\n\t\treviewerDecisionValidator: Validator<ReviewerDecision>;\n\t\tmaxAttempts?: number;\n\t\tlogger: ILogger;\n\t}) {\n\t\tif (!Number.isInteger(maxAttempts) || maxAttempts < 1) {\n\t\t\tthrow new RangeError('maxAttempts must be a positive integer');\n\t\t}\n\n\t\tthis.plannerAgent = plannerAgent;\n\t\tthis.executorAgent = executorAgent;\n\t\tthis.reviewerAgent = reviewerAgent;\n\t\tthis.reviewerDecisionValidator = reviewerDecisionValidator;\n\t\tthis.maxAttempts = maxAttempts;\n\t\tthis.logger = logger;\n\t}\n\n\t/**\n\t * Every role, round and retry of one run shares one run context, so what the executor writes\n\t * is one run of the workspace's history, and the planner's and reviewer's turns are tied to\n\t * it. The orchestrator ends that context with the run, unless an outer agent owns it.\n\t */\n\tasync run(prompt: string, options: RunOptions): Promise<AgentResponse> {\n\t\tconst start = Date.now();\n\t\t// Per invocation, not per instance: the CLI keeps one orchestrator for a whole session.\n\t\tconst totals: RunTotals = { tokens: undefined, unreported: false };\n\t\tconst inherited = runContextOf(options);\n\t\tconst context = inherited ?? new RunContext();\n\t\tlet end: RunEnd = 'failed';\n\n\t\ttry {\n\t\t\tawait this.runRounds(prompt, inherited ? options : withRunContext(options, context), totals);\n\t\t\tend = 'completed';\n\t\t} catch (error) {\n\t\t\tif (isAbortError(error)) end = 'cancelled';\n\t\t\t// A failing role carries its own tokens; the roles before it are in the totals.\n\t\t\tthrow withSpentTokens(error, totals.tokens, totals.unreported);\n\t\t} finally {\n\t\t\tif (!inherited) await this.finishContext(context, end);\n\t\t}\n\n\t\t// Read once the run ended: a run that changed nothing left the history then.\n\t\tconst { runId } = context.historyRun;\n\t\treturn {\n\t\t\tresponse: 'All job has finished',\n\t\t\ttokens: runTokens(totals),\n\t\t\t// The whole run's wall clock, so it matches what the consumer waited.\n\t\t\tduration: (Date.now() - start) / 1000,\n\t\t\t...(!inherited && runId ? { runId } : {}),\n\t\t};\n\t}\n\n\t/** The run's records end here, but its result stands: what it changed is already on disk. */\n\tprivate async finishContext(context: RunContext, end: RunEnd): Promise<void> {\n\t\ttry {\n\t\t\tawait context.finish(end);\n\t\t} catch (error) {\n\t\t\tthis.warn(`The run could not close its records: ${describeFailure(error)}`);\n\t\t}\n\t}\n\n\tprivate async runRounds(prompt: string, options: RunOptions, totals: RunTotals): Promise<void> {\n\t\tlet lastFailureReason: string | undefined;\n\n\t\tfor (let attempt = 1; attempt <= this.maxAttempts; attempt++) {\n\t\t\tconst isLastAttempt = attempt === this.maxAttempts;\n\n\t\t\tconst plannerResponse = await this.plannerAgent.run(\n\t\t\t\tgetPlannerPrompt(prompt, lastFailureReason),\n\t\t\t\toptions,\n\t\t\t);\n\t\t\taddToTotals(totals, plannerResponse);\n\n\t\t\tif (!plannerResponse.response) {\n\t\t\t\tlastFailureReason = 'The planner produced no response.';\n\t\t\t\tif (isLastAttempt) {\n\t\t\t\t\tthrow new UnrecoverableError('Max attempts exhausted', { cause: lastFailureReason });\n\t\t\t\t}\n\t\t\t\tthis.warn(\n\t\t\t\t\t`Attempt ${String(attempt)}/${String(this.maxAttempts)}: the planner produced no response; starting another round`,\n\t\t\t\t);\n\t\t\t\tcontinue;\n\t\t\t}\n\n\t\t\tconst executorResponse = await this.executorAgent.run(\n\t\t\t\tgetExecutorPrompt(prompt, plannerResponse.response),\n\t\t\t\toptions,\n\t\t\t);\n\t\t\taddToTotals(totals, executorResponse);\n\n\t\t\tif (!executorResponse.response) {\n\t\t\t\tlastFailureReason = 'The executor produced no response.';\n\t\t\t\tif (isLastAttempt) {\n\t\t\t\t\tthrow new UnrecoverableError('Max attempts exhausted', { cause: lastFailureReason });\n\t\t\t\t}\n\n\t\t\t\tthis.warn(\n\t\t\t\t\t`Attempt ${String(attempt)}/${String(this.maxAttempts)}: the executor produced no response; starting another round`,\n\t\t\t\t);\n\t\t\t\tcontinue;\n\t\t\t}\n\n\t\t\tconst reviewerDecision = await this.getReviewerDecision(\n\t\t\t\tprompt,\n\t\t\t\tplannerResponse.response,\n\t\t\t\texecutorResponse.response,\n\t\t\t\toptions,\n\t\t\t\ttotals,\n\t\t\t);\n\n\t\t\tif (reviewerDecision.decision === 'approved') return;\n\n\t\t\tlastFailureReason = reviewerDecision.feedback;\n\t\t\tif (isLastAttempt) {\n\t\t\t\tthrow new UnrecoverableError('Max attempts exhausted', { cause: lastFailureReason });\n\t\t\t}\n\n\t\t\tthis.warn(\n\t\t\t\t`Attempt ${String(attempt)}/${String(this.maxAttempts)}: the reviewer rejected the round; starting another with its feedback`,\n\t\t\t\tlastFailureReason,\n\t\t\t);\n\t\t}\n\n\t\tthrow new UnrecoverableError('Max attempts exhausted', {\n\t\t\tcause: lastFailureReason ?? 'Unknown failure.',\n\t\t});\n\t}\n\n\t// Retries only the reviewer call on a malformed decision, instead of redoing\n\t// planning/execution — a bad or unparsable reviewer response is not evidence\n\t// the plan or the implementation were wrong.\n\tprivate async getReviewerDecision(\n\t\tuserPrompt: string,\n\t\tplannerPrompt: string,\n\t\texecutorResult: string,\n\t\toptions: RunOptions,\n\t\ttotals: RunTotals,\n\t): Promise<ReviewerDecision> {\n\t\tlet parseFailureReason: string | undefined;\n\n\t\tfor (let attempt = 1; attempt <= this.maxAttempts; attempt++) {\n\t\t\tconst reviewerResponse = await this.reviewerAgent.run(\n\t\t\t\tgetReviewerPrompt(userPrompt, plannerPrompt, executorResult, parseFailureReason),\n\t\t\t\toptions,\n\t\t\t);\n\n\t\t\taddToTotals(totals, reviewerResponse);\n\n\t\t\ttry {\n\t\t\t\treturn parseReviewerDecision(reviewerResponse, this.reviewerDecisionValidator);\n\t\t\t} catch (error) {\n\t\t\t\tif (!(error instanceof RecoverableError)) throw error;\n\n\t\t\t\tparseFailureReason = error.cause;\n\t\t\t\tif (attempt === this.maxAttempts) {\n\t\t\t\t\tthrow new UnrecoverableError('Max attempts exhausted', { cause: parseFailureReason });\n\t\t\t\t}\n\n\t\t\t\tthis.warn(\n\t\t\t\t\t`Reviewer decision ${String(attempt)}/${String(this.maxAttempts)} was unusable; asking the reviewer again`,\n\t\t\t\t\tparseFailureReason,\n\t\t\t\t);\n\t\t\t}\n\t\t}\n\n\t\tthrow new UnrecoverableError('Max attempts exhausted', {\n\t\t\tcause: parseFailureReason ?? 'Unknown failure.',\n\t\t});\n\t}\n\n\tprivate warn(...args: unknown[]): void {\n\t\ttreatErrors(\n\t\t\t() => {\n\t\t\t\tthis.logger.warn(...args);\n\t\t\t},\n\t\t\tclassifyHostFailure,\n\t\t\t'Orchestrator logger failed while reporting a retry',\n\t\t);\n\t}\n}\n"]}
+{"version":3,"file":"orchestratorAgent.js","sourceRoot":"","sources":["../../../../src/orchestration/domain/model/orchestratorAgent.ts"],"names":[],"mappings":"AAOA,OAAO,EAAE,SAAS,EAAe,MAAM,2BAA2B,CAAC;AACnE,OAAO,EAAE,gBAAgB,EAAE,kBAAkB,EAAE,eAAe,EAAE,MAAM,2BAA2B,CAAC;AAIlG,OAAO,EACN,mBAAmB,EACnB,eAAe,EACf,WAAW,GACX,MAAM,oCAAoC,CAAC;AAC5C,OAAO,EACN,UAAU,EAEV,YAAY,EACZ,cAAc,GACd,MAAM,8BAA8B,CAAC;AACtC,OAAO,EAAE,YAAY,EAAE,MAAM,iCAAiC,CAAC;AAE/D,4FAA4F;AAC5F,8FAA8F;AAC9F,2EAA2E;AAC3E,MAAM,oBAAoB,GACzB,wYAAwY,CAAC;AAE1Y,MAAM,CAAC,MAAM,mBAAmB,GAAG,uXAAuX,oBAAoB,EAAE,CAAC;AAEjb,MAAM,YAAY,GAAG,2VAA2V,oBAAoB,EAAE,CAAC;AAEvY,6FAA6F;AAC7F,2FAA2F;AAC3F,gCAAgC;AAChC,MAAM,CAAC,MAAM,oBAAoB,GAAG,GAAG,YAAY,6SAA6S,CAAC;AAEjW,2FAA2F;AAC3F,oFAAoF;AACpF,MAAM,CAAC,MAAM,mCAAmC,GAAG,GAAG,YAAY,iRAAiR,CAAC;AAEpV,MAAM,YAAY,GAAG,wxBAAwxB,oBAAoB,EAAE,CAAC;AAEp0B,6DAA6D;AAC7D,MAAM,CAAC,MAAM,oBAAoB,GAAG,6IAA6I,YAAY,EAAE,CAAC;AAEhM,MAAM,CAAC,MAAM,mCAAmC,GAAG,kOAAkO,YAAY,EAAE,CAAC;AAEpS,MAAM,gBAAgB,GAAG,CAAC,UAAkB,EAAE,qBAA8B,EAAE,EAAE;IAC/E,MAAM,QAAQ,GAAG,qBAAqB;QACrC,CAAC,CAAC,iDAAiD,qBAAqB,OAAO;QAC/E,CAAC,CAAC,EAAE,CAAC;IAEN,OAAO;;EAEN,UAAU;KACP,QAAQ,EAAE,CAAC;AAChB,CAAC,CAAC;AAEF,MAAM,iBAAiB,GAAG,CAAC,UAAkB,EAAE,aAAqB,EAAE,EAAE,CACvE;;EAEC,UAAU;;;;;EAKV,aAAa;IACX,CAAC;AAEL,MAAM,iBAAiB,GAAG,CACzB,UAAkB,EAClB,aAAqB,EACrB,cAAsB,EACtB,kBAA2B,EAC1B,EAAE;IACH,MAAM,WAAW,GAAG,kBAAkB;QACrC,CAAC,CAAC,iDAAiD,kBAAkB,6GAA6G;QAClL,CAAC,CAAC,EAAE,CAAC;IAEN,OAAO;;EAEN,UAAU;;;;;EAKV,aAAa;;;;;EAKb,cAAc;KACX,WAAW,EAAE,CAAC;AACnB,CAAC,CAAC;AAQF,4FAA4F;AAC5F,SAAS,WAAW,CAAC,MAAiB,EAAE,QAAuB;IAC9D,IAAI,CAAC,QAAQ,CAAC,MAAM;QAAE,MAAM,CAAC,UAAU,GAAG,IAAI,CAAC;IAC/C,MAAM,CAAC,MAAM,GAAG,SAAS,CAAC,MAAM,CAAC,MAAM,EAAE,QAAQ,CAAC,MAAM,CAAC,CAAC;AAC3D,CAAC;AAED,uFAAuF;AACvF,SAAS,SAAS,CAAC,MAAiB;IACnC,OAAO,MAAM,CAAC,UAAU,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC;AACtD,CAAC;AAED,SAAS,cAAc,CAAC,IAAY;IACnC,MAAM,KAAK,GAAG,mCAAmC,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC;IACpE,OAAO,KAAK,EAAE,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC;AAC3B,CAAC;AAED,SAAS,qBAAqB,CAC7B,QAAuB,EACvB,yBAAsD;IAEtD,IAAI,CAAC,QAAQ,CAAC,QAAQ,EAAE,CAAC;QACxB,MAAM,IAAI,gBAAgB,CAAC,6DAA6D,EAAE;YACzF,KAAK,EAAE,+BAA+B;SACtC,CAAC,CAAC;IACJ,CAAC;IAED,IAAI,cAAuB,CAAC;IAC5B,IAAI,CAAC;QACJ,cAAc,GAAG,IAAI,CAAC,KAAK,CAAC,cAAc,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC,CAAC;IAChE,CAAC;IAAC,MAAM,CAAC;QACR,MAAM,IAAI,gBAAgB,CAAC,uCAAuC,EAAE;YACnE,KAAK,EAAE,uCAAuC;SAC9C,CAAC,CAAC;IACJ,CAAC;IAED,MAAM,QAAQ,GAAG,yBAAyB,CAAC,QAAQ,CAAC,cAAc,CAAC,CAAC;IACpE,IAAI,CAAC,QAAQ,EAAE,CAAC;QACf,MAAM,IAAI,gBAAgB,CAAC,uCAAuC,EAAE;YACnE,KAAK,EAAE,sDAAsD;SAC7D,CAAC,CAAC;IACJ,CAAC;IAED,OAAO,QAAQ,CAAC;AACjB,CAAC;AAED,MAAM,OAAO,iBAAiB;IACrB,YAAY,CAAQ;IACpB,aAAa,CAAQ;IACrB,aAAa,CAAQ;IACrB,yBAAyB,CAA8B;IACvD,WAAW,CAAS;IACpB,MAAM,CAAU;IAExB,YAAY,EACX,YAAY,EACZ,aAAa,EACb,aAAa,EACb,yBAAyB,EACzB,WAAW,GAAG,CAAC,EACf,MAAM,GAQN;QACA,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,WAAW,CAAC,IAAI,WAAW,GAAG,CAAC,EAAE,CAAC;YACvD,MAAM,IAAI,UAAU,CAAC,wCAAwC,CAAC,CAAC;QAChE,CAAC;QAED,IAAI,CAAC,YAAY,GAAG,YAAY,CAAC;QACjC,IAAI,CAAC,aAAa,GAAG,aAAa,CAAC;QACnC,IAAI,CAAC,aAAa,GAAG,aAAa,CAAC;QACnC,IAAI,CAAC,yBAAyB,GAAG,yBAAyB,CAAC;QAC3D,IAAI,CAAC,WAAW,GAAG,WAAW,CAAC;QAC/B,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IACtB,CAAC;IAED;;;;OAIG;IACH,KAAK,CAAC,GAAG,CAAC,MAAc,EAAE,OAAmB;QAC5C,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QACzB,wFAAwF;QACxF,MAAM,MAAM,GAAc,EAAE,MAAM,EAAE,SAAS,EAAE,UAAU,EAAE,KAAK,EAAE,CAAC;QACnE,MAAM,SAAS,GAAG,YAAY,CAAC,OAAO,CAAC,CAAC;QACxC,MAAM,OAAO,GAAG,SAAS,IAAI,IAAI,UAAU,EAAE,CAAC;QAC9C,IAAI,GAAG,GAAW,QAAQ,CAAC;QAE3B,IAAI,CAAC;YACJ,MAAM,IAAI,CAAC,SAAS,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,cAAc,CAAC,OAAO,EAAE,OAAO,CAAC,EAAE,MAAM,CAAC,CAAC;YAC7F,GAAG,GAAG,WAAW,CAAC;QACnB,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,IAAI,YAAY,CAAC,KAAK,CAAC;gBAAE,GAAG,GAAG,WAAW,CAAC;YAC3C,gFAAgF;YAChF,MAAM,eAAe,CAAC,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,UAAU,CAAC,CAAC;QAChE,CAAC;gBAAS,CAAC;YACV,IAAI,CAAC,SAAS;gBAAE,MAAM,IAAI,CAAC,aAAa,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC;QACxD,CAAC;QAED,6EAA6E;QAC7E,MAAM,EAAE,KAAK,EAAE,GAAG,OAAO,CAAC,UAAU,CAAC;QACrC,OAAO;YACN,QAAQ,EAAE,sBAAsB;YAChC,MAAM,EAAE,SAAS,CAAC,MAAM,CAAC;YACzB,sEAAsE;YACtE,QAAQ,EAAE,CAAC,IAAI,CAAC,GAAG,EAAE,GAAG,KAAK,CAAC,GAAG,IAAI;YACrC,GAAG,CAAC,CAAC,SAAS,IAAI,KAAK,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;SACzC,CAAC;IACH,CAAC;IAED,6FAA6F;IACrF,KAAK,CAAC,aAAa,CAAC,OAAmB,EAAE,GAAW;QAC3D,IAAI,CAAC;YACJ,MAAM,OAAO,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;QAC3B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,IAAI,CAAC,IAAI,CAAC,wCAAwC,eAAe,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;QAC7E,CAAC;IACF,CAAC;IAEO,KAAK,CAAC,SAAS,CAAC,MAAc,EAAE,OAAmB,EAAE,MAAiB;QAC7E,IAAI,iBAAqC,CAAC;QAE1C,KAAK,IAAI,OAAO,GAAG,CAAC,EAAE,OAAO,IAAI,IAAI,CAAC,WAAW,EAAE,OAAO,EAAE,EAAE,CAAC;YAC9D,MAAM,aAAa,GAAG,OAAO,KAAK,IAAI,CAAC,WAAW,CAAC;YAEnD,MAAM,eAAe,GAAG,MAAM,IAAI,CAAC,YAAY,CAAC,GAAG,CAClD,gBAAgB,CAAC,MAAM,EAAE,iBAAiB,CAAC,EAC3C,IAAI,CAAC,cAAc,CAAC,OAAO,EAAE,SAAS,CAAC,CACvC,CAAC;YACF,WAAW,CAAC,MAAM,EAAE,eAAe,CAAC,CAAC;YAErC,IAAI,CAAC,eAAe,CAAC,QAAQ,EAAE,CAAC;gBAC/B,iBAAiB,GAAG,mCAAmC,CAAC;gBACxD,IAAI,aAAa,EAAE,CAAC;oBACnB,MAAM,IAAI,kBAAkB,CAAC,wBAAwB,EAAE,EAAE,KAAK,EAAE,iBAAiB,EAAE,CAAC,CAAC;gBACtF,CAAC;gBACD,IAAI,CAAC,IAAI,CACR,WAAW,MAAM,CAAC,OAAO,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,4DAA4D,CAClH,CAAC;gBACF,SAAS;YACV,CAAC;YAED,MAAM,gBAAgB,GAAG,MAAM,IAAI,CAAC,aAAa,CAAC,GAAG,CACpD,iBAAiB,CAAC,MAAM,EAAE,eAAe,CAAC,QAAQ,CAAC,EACnD,IAAI,CAAC,cAAc,CAAC,OAAO,EAAE,UAAU,CAAC,CACxC,CAAC;YACF,WAAW,CAAC,MAAM,EAAE,gBAAgB,CAAC,CAAC;YAEtC,IAAI,CAAC,gBAAgB,CAAC,QAAQ,EAAE,CAAC;gBAChC,iBAAiB,GAAG,oCAAoC,CAAC;gBACzD,IAAI,aAAa,EAAE,CAAC;oBACnB,MAAM,IAAI,kBAAkB,CAAC,wBAAwB,EAAE,EAAE,KAAK,EAAE,iBAAiB,EAAE,CAAC,CAAC;gBACtF,CAAC;gBAED,IAAI,CAAC,IAAI,CACR,WAAW,MAAM,CAAC,OAAO,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,6DAA6D,CACnH,CAAC;gBACF,SAAS;YACV,CAAC;YAED,MAAM,gBAAgB,GAAG,MAAM,IAAI,CAAC,mBAAmB,CACtD,MAAM,EACN,eAAe,CAAC,QAAQ,EACxB,gBAAgB,CAAC,QAAQ,EACzB,OAAO,EACP,MAAM,CACN,CAAC;YAEF,IAAI,gBAAgB,CAAC,QAAQ,KAAK,UAAU;gBAAE,OAAO;YAErD,iBAAiB,GAAG,gBAAgB,CAAC,QAAQ,CAAC;YAC9C,IAAI,aAAa,EAAE,CAAC;gBACnB,MAAM,IAAI,kBAAkB,CAAC,wBAAwB,EAAE,EAAE,KAAK,EAAE,iBAAiB,EAAE,CAAC,CAAC;YACtF,CAAC;YAED,IAAI,CAAC,IAAI,CACR,WAAW,MAAM,CAAC,OAAO,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,uEAAuE,EAC7H,iBAAiB,CACjB,CAAC;QACH,CAAC;QAED,MAAM,IAAI,kBAAkB,CAAC,wBAAwB,EAAE;YACtD,KAAK,EAAE,iBAAiB,IAAI,kBAAkB;SAC9C,CAAC,CAAC;IACJ,CAAC;IAED,6EAA6E;IAC7E,6EAA6E;IAC7E,6CAA6C;IACrC,KAAK,CAAC,mBAAmB,CAChC,UAAkB,EAClB,aAAqB,EACrB,cAAsB,EACtB,OAAmB,EACnB,MAAiB;QAEjB,IAAI,kBAAsC,CAAC;QAE3C,KAAK,IAAI,OAAO,GAAG,CAAC,EAAE,OAAO,IAAI,IAAI,CAAC,WAAW,EAAE,OAAO,EAAE,EAAE,CAAC;YAC9D,MAAM,gBAAgB,GAAG,MAAM,IAAI,CAAC,aAAa,CAAC,GAAG,CACpD,iBAAiB,CAAC,UAAU,EAAE,aAAa,EAAE,cAAc,EAAE,kBAAkB,CAAC,EAChF,IAAI,CAAC,cAAc,CAAC,OAAO,EAAE,UAAU,CAAC,CACxC,CAAC;YAEF,WAAW,CAAC,MAAM,EAAE,gBAAgB,CAAC,CAAC;YAEtC,IAAI,CAAC;gBACJ,OAAO,qBAAqB,CAAC,gBAAgB,EAAE,IAAI,CAAC,yBAAyB,CAAC,CAAC;YAChF,CAAC;YAAC,OAAO,KAAK,EAAE,CAAC;gBAChB,IAAI,CAAC,CAAC,KAAK,YAAY,gBAAgB,CAAC;oBAAE,MAAM,KAAK,CAAC;gBAEtD,kBAAkB,GAAG,KAAK,CAAC,KAAK,CAAC;gBACjC,IAAI,OAAO,KAAK,IAAI,CAAC,WAAW,EAAE,CAAC;oBAClC,MAAM,IAAI,kBAAkB,CAAC,wBAAwB,EAAE,EAAE,KAAK,EAAE,kBAAkB,EAAE,CAAC,CAAC;gBACvF,CAAC;gBAED,IAAI,CAAC,IAAI,CACR,qBAAqB,MAAM,CAAC,OAAO,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,0CAA0C,EAC1G,kBAAkB,CAClB,CAAC;YACH,CAAC;QACF,CAAC;QAED,MAAM,IAAI,kBAAkB,CAAC,wBAAwB,EAAE;YACtD,KAAK,EAAE,kBAAkB,IAAI,kBAAkB;SAC/C,CAAC,CAAC;IACJ,CAAC;IAEO,IAAI,CAAC,GAAG,IAAe;QAC9B,WAAW,CACV,GAAG,EAAE;YACJ,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,IAAI,CAAC,CAAC;QAC3B,CAAC,EACD,mBAAmB,EACnB,oDAAoD,CACpD,CAAC;IACH,CAAC;IAED;;;OAGG;IACK,cAAc,CAAC,OAAmB,EAAE,KAAoB;QAC/D,MAAM,UAAU,GAAG,OAAO,CAAC,UAAU,CAAC;QACtC,OAAO;YACN,GAAG,OAAO;YACV,UAAU,EAAE,UAAU;gBACrB,CAAC,CAAC,CAAC,KAAoB,EAAE,EAAE;oBACzB,WAAW,CACV,GAAG,EAAE;wBACJ,UAAU,CAAC,EAAE,GAAG,KAAK,EAAE,KAAK,EAAE,CAAC,CAAC;oBACjC,CAAC,EACD,mBAAmB,EACnB,uCAAuC,CACvC,CAAC;gBACH,CAAC;gBACF,CAAC,CAAC,SAAS;SACZ,CAAC;IACH,CAAC;CACD","sourcesContent":["import type {\n\tAgent,\n\tAgentResponse,\n\tProgressAgent,\n\tProgressEvent,\n\tRunOptions,\n} from '#src/agent/domain/agent';\nimport { addTokens, type Tokens } from '#src/shared/domain/tokens';\nimport { RecoverableError, UnrecoverableError, withSpentTokens } from '#src/shared/domain/errors';\nimport type { ReviewerDecision } from './reviewerDecision.ts';\nimport type { Validator } from '../interface/validator.ts';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport {\n\tclassifyHostFailure,\n\tdescribeFailure,\n\ttreatErrors,\n} from '#src/shared/domain/providerFailure';\nimport {\n\tRunContext,\n\ttype RunEnd,\n\trunContextOf,\n\twithRunContext,\n} from '#src/agent/domain/runContext';\nimport { isAbortError } from '#src/shared/domain/isAbortError';\n\n// Each role's standing instructions, kept apart from the data of a round so an agent with a\n// system prompt can hold them there instead of receiving them again in every prompt. An agent\n// without one gets them at the head of each prompt; see `InstructedAgent`.\nconst conciseCommunication =\n\t'Communicate tersely and professionally. In reasoning updates, state only the next concrete action or a verified finding, in one short sentence. Do not narrate routine intentions, uncertainty, self-reflection, encouragement, or a desire to be thorough. Do not restate tool results. Once you decide what to do, act immediately. Use only the detail required by your role in the final response.';\n\nexport 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. ${conciseCommunication}`;\n\nconst 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. ${conciseCommunication}`;\n\n// For an executor that can run the project's tests and checks: an Agent SDK one, which has a\n// shell, or a model-API one given commands. The harness cannot tell which command runs the\n// tests, so any command counts.\nexport const executorInstructions = `${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.`;\n\n// For a model-API executor with no command, which cannot run anything: told so, it neither\n// spends its steps looking for a way to test nor claims a result it could not have.\nexport 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.`;\n\nconst 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. ${conciseCommunication}`;\n\n// Each goes with the executor instructions of the same name.\nexport 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}`;\n\nexport 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. ${reviewerTask}`;\n\nconst getPlannerPrompt = (userPrompt: string, previousFailureReason?: string) => {\n\tconst feedback = previousFailureReason\n\t\t? `\\n\\nFeedback from the previous attempt:\\n---\\n${previousFailureReason}\\n---`\n\t\t: '';\n\n\treturn `Original user request:\n---\n${userPrompt}\n---${feedback}`;\n};\n\nconst getExecutorPrompt = (userPrompt: string, plannerPrompt: string) =>\n\t`Original user request:\n---\n${userPrompt}\n---\n\nCurrent implementation plan:\n---\n${plannerPrompt}\n---`;\n\nconst getReviewerPrompt = (\n\tuserPrompt: string,\n\tplannerPrompt: string,\n\texecutorResult: string,\n\tparseFailureReason?: string,\n) => {\n\tconst retryNotice = parseFailureReason\n\t\t? `\\n\\nYour previous response could not be used: ${parseFailureReason} Respond with JSON only, matching the schema exactly, with no surrounding text and no markdown code fences.`\n\t\t: '';\n\n\treturn `Original user request:\n---\n${userPrompt}\n---\n\nPlanner's current plan:\n---\n${plannerPrompt}\n---\n\nExecutor response (context only; verify it independently):\n---\n${executorResult}\n---${retryNotice}`;\n};\n\ninterface RunTotals {\n\ttokens: Tokens | undefined;\n\t/** A role answered without usage, so the run's total cannot be known. */\n\tunreported: boolean;\n}\n\n// Every role's response counts, even an empty one that starts another round: it was billed.\nfunction addToTotals(totals: RunTotals, response: AgentResponse): void {\n\tif (!response.tokens) totals.unreported = true;\n\ttotals.tokens = addTokens(totals.tokens, response.tokens);\n}\n\n/** Unknown rather than partial: a sum missing a billed call must not look complete. */\nfunction runTokens(totals: RunTotals): Tokens | undefined {\n\treturn totals.unreported ? undefined : totals.tokens;\n}\n\nfunction stripCodeFence(text: string): string {\n\tconst match = /^```(?:json)?\\s*([\\s\\S]*?)\\s*```$/.exec(text.trim());\n\treturn match?.[1] ?? text;\n}\n\nfunction parseReviewerDecision(\n\tresponse: AgentResponse,\n\treviewerDecisionValidator: Validator<ReviewerDecision>,\n): ReviewerDecision {\n\tif (!response.response) {\n\t\tthrow new RecoverableError(\"There's no decision from the reviewer, something went wrong\", {\n\t\t\tcause: 'Reviewer decision is missing.',\n\t\t});\n\t}\n\n\tlet parsedResponse: unknown;\n\ttry {\n\t\tparsedResponse = JSON.parse(stripCodeFence(response.response));\n\t} catch {\n\t\tthrow new RecoverableError('Reviewer returned an invalid decision', {\n\t\t\tcause: 'Reviewer response must be valid JSON.',\n\t\t});\n\t}\n\n\tconst decision = reviewerDecisionValidator.validate(parsedResponse);\n\tif (!decision) {\n\t\tthrow new RecoverableError('Reviewer returned an invalid decision', {\n\t\t\tcause: 'Reviewer response did not match the decision schema.',\n\t\t});\n\t}\n\n\treturn decision;\n}\n\nexport class OrchestratorAgent implements Agent {\n\tprivate plannerAgent: Agent;\n\tprivate executorAgent: Agent;\n\tprivate reviewerAgent: Agent;\n\tprivate reviewerDecisionValidator: Validator<ReviewerDecision>;\n\tprivate maxAttempts: number;\n\tprivate logger: ILogger;\n\n\tconstructor({\n\t\tplannerAgent,\n\t\texecutorAgent,\n\t\treviewerAgent,\n\t\treviewerDecisionValidator,\n\t\tmaxAttempts = 3,\n\t\tlogger,\n\t}: {\n\t\tplannerAgent: Agent;\n\t\texecutorAgent: Agent;\n\t\treviewerAgent: Agent;\n\t\treviewerDecisionValidator: Validator<ReviewerDecision>;\n\t\tmaxAttempts?: number;\n\t\tlogger: ILogger;\n\t}) {\n\t\tif (!Number.isInteger(maxAttempts) || maxAttempts < 1) {\n\t\t\tthrow new RangeError('maxAttempts must be a positive integer');\n\t\t}\n\n\t\tthis.plannerAgent = plannerAgent;\n\t\tthis.executorAgent = executorAgent;\n\t\tthis.reviewerAgent = reviewerAgent;\n\t\tthis.reviewerDecisionValidator = reviewerDecisionValidator;\n\t\tthis.maxAttempts = maxAttempts;\n\t\tthis.logger = logger;\n\t}\n\n\t/**\n\t * Every role, round and retry of one run shares one run context, so what the executor writes\n\t * is one run of the workspace's history, and the planner's and reviewer's turns are tied to\n\t * it. The orchestrator ends that context with the run, unless an outer agent owns it.\n\t */\n\tasync run(prompt: string, options: RunOptions): Promise<AgentResponse> {\n\t\tconst start = Date.now();\n\t\t// Per invocation, not per instance: the CLI keeps one orchestrator for a whole session.\n\t\tconst totals: RunTotals = { tokens: undefined, unreported: false };\n\t\tconst inherited = runContextOf(options);\n\t\tconst context = inherited ?? new RunContext();\n\t\tlet end: RunEnd = 'failed';\n\n\t\ttry {\n\t\t\tawait this.runRounds(prompt, inherited ? options : withRunContext(options, context), totals);\n\t\t\tend = 'completed';\n\t\t} catch (error) {\n\t\t\tif (isAbortError(error)) end = 'cancelled';\n\t\t\t// A failing role carries its own tokens; the roles before it are in the totals.\n\t\t\tthrow withSpentTokens(error, totals.tokens, totals.unreported);\n\t\t} finally {\n\t\t\tif (!inherited) await this.finishContext(context, end);\n\t\t}\n\n\t\t// Read once the run ended: a run that changed nothing left the history then.\n\t\tconst { runId } = context.historyRun;\n\t\treturn {\n\t\t\tresponse: 'All job has finished',\n\t\t\ttokens: runTokens(totals),\n\t\t\t// The whole run's wall clock, so it matches what the consumer waited.\n\t\t\tduration: (Date.now() - start) / 1000,\n\t\t\t...(!inherited && runId ? { runId } : {}),\n\t\t};\n\t}\n\n\t/** The run's records end here, but its result stands: what it changed is already on disk. */\n\tprivate async finishContext(context: RunContext, end: RunEnd): Promise<void> {\n\t\ttry {\n\t\t\tawait context.finish(end);\n\t\t} catch (error) {\n\t\t\tthis.warn(`The run could not close its records: ${describeFailure(error)}`);\n\t\t}\n\t}\n\n\tprivate async runRounds(prompt: string, options: RunOptions, totals: RunTotals): Promise<void> {\n\t\tlet lastFailureReason: string | undefined;\n\n\t\tfor (let attempt = 1; attempt <= this.maxAttempts; attempt++) {\n\t\t\tconst isLastAttempt = attempt === this.maxAttempts;\n\n\t\t\tconst plannerResponse = await this.plannerAgent.run(\n\t\t\t\tgetPlannerPrompt(prompt, lastFailureReason),\n\t\t\t\tthis.optionsForRole(options, 'planner'),\n\t\t\t);\n\t\t\taddToTotals(totals, plannerResponse);\n\n\t\t\tif (!plannerResponse.response) {\n\t\t\t\tlastFailureReason = 'The planner produced no response.';\n\t\t\t\tif (isLastAttempt) {\n\t\t\t\t\tthrow new UnrecoverableError('Max attempts exhausted', { cause: lastFailureReason });\n\t\t\t\t}\n\t\t\t\tthis.warn(\n\t\t\t\t\t`Attempt ${String(attempt)}/${String(this.maxAttempts)}: the planner produced no response; starting another round`,\n\t\t\t\t);\n\t\t\t\tcontinue;\n\t\t\t}\n\n\t\t\tconst executorResponse = await this.executorAgent.run(\n\t\t\t\tgetExecutorPrompt(prompt, plannerResponse.response),\n\t\t\t\tthis.optionsForRole(options, 'executor'),\n\t\t\t);\n\t\t\taddToTotals(totals, executorResponse);\n\n\t\t\tif (!executorResponse.response) {\n\t\t\t\tlastFailureReason = 'The executor produced no response.';\n\t\t\t\tif (isLastAttempt) {\n\t\t\t\t\tthrow new UnrecoverableError('Max attempts exhausted', { cause: lastFailureReason });\n\t\t\t\t}\n\n\t\t\t\tthis.warn(\n\t\t\t\t\t`Attempt ${String(attempt)}/${String(this.maxAttempts)}: the executor produced no response; starting another round`,\n\t\t\t\t);\n\t\t\t\tcontinue;\n\t\t\t}\n\n\t\t\tconst reviewerDecision = await this.getReviewerDecision(\n\t\t\t\tprompt,\n\t\t\t\tplannerResponse.response,\n\t\t\t\texecutorResponse.response,\n\t\t\t\toptions,\n\t\t\t\ttotals,\n\t\t\t);\n\n\t\t\tif (reviewerDecision.decision === 'approved') return;\n\n\t\t\tlastFailureReason = reviewerDecision.feedback;\n\t\t\tif (isLastAttempt) {\n\t\t\t\tthrow new UnrecoverableError('Max attempts exhausted', { cause: lastFailureReason });\n\t\t\t}\n\n\t\t\tthis.warn(\n\t\t\t\t`Attempt ${String(attempt)}/${String(this.maxAttempts)}: the reviewer rejected the round; starting another with its feedback`,\n\t\t\t\tlastFailureReason,\n\t\t\t);\n\t\t}\n\n\t\tthrow new UnrecoverableError('Max attempts exhausted', {\n\t\t\tcause: lastFailureReason ?? 'Unknown failure.',\n\t\t});\n\t}\n\n\t// Retries only the reviewer call on a malformed decision, instead of redoing\n\t// planning/execution — a bad or unparsable reviewer response is not evidence\n\t// the plan or the implementation were wrong.\n\tprivate async getReviewerDecision(\n\t\tuserPrompt: string,\n\t\tplannerPrompt: string,\n\t\texecutorResult: string,\n\t\toptions: RunOptions,\n\t\ttotals: RunTotals,\n\t): Promise<ReviewerDecision> {\n\t\tlet parseFailureReason: string | undefined;\n\n\t\tfor (let attempt = 1; attempt <= this.maxAttempts; attempt++) {\n\t\t\tconst reviewerResponse = await this.reviewerAgent.run(\n\t\t\t\tgetReviewerPrompt(userPrompt, plannerPrompt, executorResult, parseFailureReason),\n\t\t\t\tthis.optionsForRole(options, 'reviewer'),\n\t\t\t);\n\n\t\t\taddToTotals(totals, reviewerResponse);\n\n\t\t\ttry {\n\t\t\t\treturn parseReviewerDecision(reviewerResponse, this.reviewerDecisionValidator);\n\t\t\t} catch (error) {\n\t\t\t\tif (!(error instanceof RecoverableError)) throw error;\n\n\t\t\t\tparseFailureReason = error.cause;\n\t\t\t\tif (attempt === this.maxAttempts) {\n\t\t\t\t\tthrow new UnrecoverableError('Max attempts exhausted', { cause: parseFailureReason });\n\t\t\t\t}\n\n\t\t\t\tthis.warn(\n\t\t\t\t\t`Reviewer decision ${String(attempt)}/${String(this.maxAttempts)} was unusable; asking the reviewer again`,\n\t\t\t\t\tparseFailureReason,\n\t\t\t\t);\n\t\t\t}\n\t\t}\n\n\t\tthrow new UnrecoverableError('Max attempts exhausted', {\n\t\t\tcause: parseFailureReason ?? 'Unknown failure.',\n\t\t});\n\t}\n\n\tprivate warn(...args: unknown[]): void {\n\t\ttreatErrors(\n\t\t\t() => {\n\t\t\t\tthis.logger.warn(...args);\n\t\t\t},\n\t\t\tclassifyHostFailure,\n\t\t\t'Orchestrator logger failed while reporting a retry',\n\t\t);\n\t}\n\n\t/**\n\t * Keep the options object extensible and the private run context intact while making the\n\t * orchestrator's attribution authoritative over anything a nested agent supplied.\n\t */\n\tprivate optionsForRole(options: RunOptions, agent: ProgressAgent): RunOptions {\n\t\tconst onProgress = options.onProgress;\n\t\treturn {\n\t\t\t...options,\n\t\t\tonProgress: onProgress\n\t\t\t\t? (event: ProgressEvent) => {\n\t\t\t\t\t\ttreatErrors(\n\t\t\t\t\t\t\t() => {\n\t\t\t\t\t\t\t\tonProgress({ ...event, agent });\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\tclassifyHostFailure,\n\t\t\t\t\t\t\t'Orchestrator progress callback failed',\n\t\t\t\t\t\t);\n\t\t\t\t\t}\n\t\t\t\t: undefined,\n\t\t};\n\t}\n}\n"]}

package/dist/tools/domain/preparedCall.d.ts

@@ -1,7 +1,7 @@
 import type { CommandLine, ToolRisk } from '#src/agent/domain/approval';
 import type { ToolDefinition } from '#src/llm/domain/tool';
 import type { ToolCallContext } from './readRegistry.ts';
-import type { Tool } from './tool.ts';
+import type { Tool, ToolSummarizer } from './tool.ts';
 /**
  * A tool call checked and fixed before anyone approves it: its risk is judged on what will
  * really happen, and `run` does exactly that, nothing worked out again.
@@ -14,11 +14,24 @@ export interface PreparedCall {
     readonly command?: CommandLine;
     run(signal: AbortSignal): Promise<string>;
 }
+/**
+ * A tool ran and produced a result the model should see as an error. Internal tools use this
+ * instead of a plain error when the result also carries narrative metadata, such as a command's
+ * non-zero exit code.
+ */
+export declare class ToolExecutionError extends Error {
+    readonly exitCode: number | undefined;
+    constructor(message: string, { exitCode }?: {
+        exitCode?: number;
+    });
+}
 /**
  * A tool built by the harness, such as a write tool, which prepares each call against the
  * state of the run before it is approved. Internal: consumers build a `Tool`.
  */
 export interface PreparingTool extends ToolDefinition {
+    /** Describes the model's input without exposing preparation details to the provider. */
+    summarize?: ToolSummarizer;
     /** Rejects with a message for the model when the call cannot run. */
     prepare(input: unknown, signal: AbortSignal, call: ToolCallContext): Promise<PreparedCall>;
 }

package/dist/tools/domain/preparedCall.js

@@ -1,3 +1,15 @@
+/**
+ * A tool ran and produced a result the model should see as an error. Internal tools use this
+ * instead of a plain error when the result also carries narrative metadata, such as a command's
+ * non-zero exit code.
+ */
+export class ToolExecutionError extends Error {
+    exitCode;
+    constructor(message, { exitCode } = {}) {
+        super(message);
+        this.exitCode = exitCode;
+    }
+}
 function isPreparing(tool) {
     return 'prepare' in tool;
 }

package/dist/tools/domain/preparedCall.js.map

@@ -1 +1 @@
-{"version":3,"file":"preparedCall.js","sourceRoot":"","sources":["../../../src/tools/domain/preparedCall.ts"],"names":[],"mappings":"AA8BA,SAAS,WAAW,CAAC,IAAe;IACnC,OAAO,SAAS,IAAI,IAAI,CAAC;AAC1B,CAAC;AAED;;;GAGG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW,CAChC,IAAe,EACf,KAAc,EACd,MAAmB,EACnB,IAAqB;IAErB,IAAI,WAAW,CAAC,IAAI,CAAC;QAAE,OAAO,IAAI,CAAC,OAAO,CAAC,KAAK,EAAE,MAAM,EAAE,IAAI,CAAC,CAAC;IAEhE,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;IAC9B,OAAO,EAAE,IAAI,EAAE,GAAG,EAAE,SAAS,CAAC,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,KAAK,EAAE,SAAS,CAAC,EAAE,CAAC;AACnE,CAAC","sourcesContent":["import type { CommandLine, ToolRisk } from '#src/agent/domain/approval';\nimport type { ToolDefinition } from '#src/llm/domain/tool';\nimport type { ToolCallContext } from './readRegistry.ts';\nimport type { Tool } from './tool.ts';\n\n/**\n * A tool call checked and fixed before anyone approves it: its risk is judged on what will\n * really happen, and `run` does exactly that, nothing worked out again.\n */\nexport interface PreparedCall {\n\treadonly risk: ToolRisk;\n\t/** For a change to a file, what it will change as a unified diff, fixed when prepared. */\n\treadonly diff?: string;\n\t/** For a call that runs a configured command, what it will run, fixed when prepared. */\n\treadonly command?: CommandLine;\n\trun(signal: AbortSignal): Promise<string>;\n}\n\n/**\n * A tool built by the harness, such as a write tool, which prepares each call against the\n * state of the run before it is approved. Internal: consumers build a `Tool`.\n */\nexport interface PreparingTool extends ToolDefinition {\n\t/** Rejects with a message for the model when the call cannot run. */\n\tprepare(input: unknown, signal: AbortSignal, call: ToolCallContext): Promise<PreparedCall>;\n}\n\n/** What an agent can run: a consumer's `Tool`, or one the harness built. */\nexport type AgentTool = Tool | PreparingTool;\n\nfunction isPreparing(tool: AgentTool): tool is PreparingTool {\n\treturn 'prepare' in tool;\n}\n\n/**\n * Brings any tool into the one lifecycle of prepare, approve, run. A consumer's `Tool` has its\n * risk judged now, and later runs on the input as the model sent it.\n */\nexport async function prepareCall(\n\ttool: AgentTool,\n\tinput: unknown,\n\tsignal: AbortSignal,\n\tcall: ToolCallContext,\n): Promise<PreparedCall> {\n\tif (isPreparing(tool)) return tool.prepare(input, signal, call);\n\n\tconst risk = tool.risk(input);\n\treturn { risk, run: runSignal => tool.execute(input, runSignal) };\n}\n"]}
+{"version":3,"file":"preparedCall.js","sourceRoot":"","sources":["../../../src/tools/domain/preparedCall.ts"],"names":[],"mappings":"AAkBA;;;;GAIG;AACH,MAAM,OAAO,kBAAmB,SAAQ,KAAK;IACnC,QAAQ,CAAqB;IAEtC,YAAY,OAAe,EAAE,EAAE,QAAQ,KAA4B,EAAE;QACpE,KAAK,CAAC,OAAO,CAAC,CAAC;QACf,IAAI,CAAC,QAAQ,GAAG,QAAQ,CAAC;IAC1B,CAAC;CACD;AAgBD,SAAS,WAAW,CAAC,IAAe;IACnC,OAAO,SAAS,IAAI,IAAI,CAAC;AAC1B,CAAC;AAED;;;GAGG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW,CAChC,IAAe,EACf,KAAc,EACd,MAAmB,EACnB,IAAqB;IAErB,IAAI,WAAW,CAAC,IAAI,CAAC;QAAE,OAAO,IAAI,CAAC,OAAO,CAAC,KAAK,EAAE,MAAM,EAAE,IAAI,CAAC,CAAC;IAEhE,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;IAC9B,OAAO,EAAE,IAAI,EAAE,GAAG,EAAE,SAAS,CAAC,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,KAAK,EAAE,SAAS,CAAC,EAAE,CAAC;AACnE,CAAC","sourcesContent":["import type { CommandLine, ToolRisk } from '#src/agent/domain/approval';\nimport type { ToolDefinition } from '#src/llm/domain/tool';\nimport type { ToolCallContext } from './readRegistry.ts';\nimport type { Tool, ToolSummarizer } from './tool.ts';\n\n/**\n * A tool call checked and fixed before anyone approves it: its risk is judged on what will\n * really happen, and `run` does exactly that, nothing worked out again.\n */\nexport interface PreparedCall {\n\treadonly risk: ToolRisk;\n\t/** For a change to a file, what it will change as a unified diff, fixed when prepared. */\n\treadonly diff?: string;\n\t/** For a call that runs a configured command, what it will run, fixed when prepared. */\n\treadonly command?: CommandLine;\n\trun(signal: AbortSignal): Promise<string>;\n}\n\n/**\n * A tool ran and produced a result the model should see as an error. Internal tools use this\n * instead of a plain error when the result also carries narrative metadata, such as a command's\n * non-zero exit code.\n */\nexport class ToolExecutionError extends Error {\n\treadonly exitCode: number | undefined;\n\n\tconstructor(message: string, { exitCode }: { exitCode?: number } = {}) {\n\t\tsuper(message);\n\t\tthis.exitCode = exitCode;\n\t}\n}\n\n/**\n * A tool built by the harness, such as a write tool, which prepares each call against the\n * state of the run before it is approved. Internal: consumers build a `Tool`.\n */\nexport interface PreparingTool extends ToolDefinition {\n\t/** Describes the model's input without exposing preparation details to the provider. */\n\tsummarize?: ToolSummarizer;\n\t/** Rejects with a message for the model when the call cannot run. */\n\tprepare(input: unknown, signal: AbortSignal, call: ToolCallContext): Promise<PreparedCall>;\n}\n\n/** What an agent can run: a consumer's `Tool`, or one the harness built. */\nexport type AgentTool = Tool | PreparingTool;\n\nfunction isPreparing(tool: AgentTool): tool is PreparingTool {\n\treturn 'prepare' in tool;\n}\n\n/**\n * Brings any tool into the one lifecycle of prepare, approve, run. A consumer's `Tool` has its\n * risk judged now, and later runs on the input as the model sent it.\n */\nexport async function prepareCall(\n\ttool: AgentTool,\n\tinput: unknown,\n\tsignal: AbortSignal,\n\tcall: ToolCallContext,\n): Promise<PreparedCall> {\n\tif (isPreparing(tool)) return tool.prepare(input, signal, call);\n\n\tconst risk = tool.risk(input);\n\treturn { risk, run: runSignal => tool.execute(input, runSignal) };\n}\n"]}

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

@@ -1,6 +1,20 @@
 import type { ToolRisk } from '#src/agent/domain/approval';
 import type { ToolDefinition } from '#src/llm/domain/tool';
+/** Narration for a call, kept out of the definition sent to an LLM provider. */
+export type ToolSummarizer = (input: unknown) => string;
+/**
+ * An input rejected by a harness-created schema is a tool error, not a failure of host code.
+ * It is kept internal to the tool lifecycle so `LLMAgent` can return it to the model without
+ * treating invalid model input as an unrecoverable consumer callback failure.
+ */
+export declare class InvalidToolInputError extends Error {
+}
 export interface Tool extends ToolDefinition {
+    /**
+     * Describes the model's input without reading resources. `defineTool` validates the input
+     * before calling this function.
+     */
+    summarize?: ToolSummarizer;
     /**
      * Judges this call from its input before it runs, so the agent can ask before a
      * `destructive` one. Input the tool would reject cannot do harm, so it may count as `safe`.

package/dist/tools/domain/tool.js

@@ -1,2 +1,8 @@
-export {};
+/**
+ * An input rejected by a harness-created schema is a tool error, not a failure of host code.
+ * It is kept internal to the tool lifecycle so `LLMAgent` can return it to the model without
+ * treating invalid model input as an unrecoverable consumer callback failure.
+ */
+export class InvalidToolInputError extends Error {
+}
 //# sourceMappingURL=tool.js.map

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

@@ -1 +1 @@
-{"version":3,"file":"tool.js","sourceRoot":"","sources":["../../../src/tools/domain/tool.ts"],"names":[],"mappings":"","sourcesContent":["import type { ToolRisk } from '#src/agent/domain/approval';\nimport type { ToolDefinition } from '#src/llm/domain/tool';\n\nexport interface Tool extends ToolDefinition {\n\t/**\n\t * Judges this call from its input before it runs, so the agent can ask before a\n\t * `destructive` one. Input the tool would reject cannot do harm, so it may count as `safe`.\n\t */\n\trisk(input: unknown): ToolRisk;\n\t/**\n\t * Runs the tool on the input the model sent. The input is unknown because the model can\n\t * send anything: validating it is the tool's job, and a rejection returns to the model as\n\t * an error result.\n\t */\n\texecute(input: unknown, signal: AbortSignal): Promise<string>;\n}\n"]}
+{"version":3,"file":"tool.js","sourceRoot":"","sources":["../../../src/tools/domain/tool.ts"],"names":[],"mappings":"AAMA;;;;GAIG;AACH,MAAM,OAAO,qBAAsB,SAAQ,KAAK;CAAG","sourcesContent":["import type { ToolRisk } from '#src/agent/domain/approval';\nimport type { ToolDefinition } from '#src/llm/domain/tool';\n\n/** Narration for a call, kept out of the definition sent to an LLM provider. */\nexport type ToolSummarizer = (input: unknown) => string;\n\n/**\n * An input rejected by a harness-created schema is a tool error, not a failure of host code.\n * It is kept internal to the tool lifecycle so `LLMAgent` can return it to the model without\n * treating invalid model input as an unrecoverable consumer callback failure.\n */\nexport class InvalidToolInputError extends Error {}\n\nexport interface Tool extends ToolDefinition {\n\t/**\n\t * Describes the model's input without reading resources. `defineTool` validates the input\n\t * before calling this function.\n\t */\n\tsummarize?: ToolSummarizer;\n\t/**\n\t * Judges this call from its input before it runs, so the agent can ask before a\n\t * `destructive` one. Input the tool would reject cannot do harm, so it may count as `safe`.\n\t */\n\trisk(input: unknown): ToolRisk;\n\t/**\n\t * Runs the tool on the input the model sent. The input is unknown because the model can\n\t * send anything: validating it is the tool's job, and a rejection returns to the model as\n\t * an error result.\n\t */\n\texecute(input: unknown, signal: AbortSignal): Promise<string>;\n}\n"]}

package/dist/tools/infrastructure/commandTool.d.ts

@@ -1,6 +1,6 @@
 import type { AccessPolicy } from '../domain/accessPolicy.ts';
 import { type CommandDefinition, type Commands } from '../domain/commands.ts';
-import type { PreparingTool } from '../domain/preparedCall.ts';
+import { type PreparingTool } from '../domain/preparedCall.ts';
 import { type CoveredFiles } from './commandCheckpoint.ts';
 import { runCommandProcess } from './commandProcess.ts';
 import type { WorkspaceWrites } from './workspaceWrites.ts';

package/dist/tools/infrastructure/commandTool.js

@@ -5,6 +5,7 @@ import { displayPath } from '#src/recovery/domain/runChanges';
 import { InvalidAgentConfigError } from '#src/shared/domain/errors';
 import { buildArguments, defaultCommandOutputBytes, defaultCommandTimeoutMs, parametersOf, validateCommands, } from "../domain/commands.js";
 import { EditRefusedError, EditUnconfirmedError } from "../domain/fileEdits.js";
+import { ToolExecutionError } from "../domain/preparedCall.js";
 import { CommandCheckpoint } from "./commandCheckpoint.js";
 import { runCommandProcess } from "./commandProcess.js";
 import { definePreparingTool } from "./defineTool.js";
@@ -43,6 +44,12 @@ export async function createCommandTool({ commands, policy, writes, covered, bas
         name: 'runCommand',
         description: describeTool(commands),
         input: inputSchema(commands),
+        summarize: ({ command: { name, ...values } }) => {
+            const definition = commands[name];
+            // The schema only permits configured names, but keep this defensive for callers that
+            // use a tool outside LLMAgent. This description never uses the resolved cwd.
+            return definition ? visibleCommand(definition, values) : name;
+        },
         prepare: async ({ command: { name, ...values } }, signal, { run }) => {
             const command = resolved.get(name);
             // The schema allows only the configured names.
@@ -123,7 +130,12 @@ export async function createCommandTool({ commands, policy, writes, covered, bas
         if (result.outcome === 'cancelled')
             signal.throwIfAborted();
         if (result.outcome !== 'exited')
-            throw new Error(report);
+            throw new ToolExecutionError(report);
+        if (result.exitCode !== 0) {
+            throw new ToolExecutionError(report, {
+                ...(result.exitCode === null ? {} : { exitCode: result.exitCode }),
+            });
+        }
         return report;
     }
     function describeRun(name, definition, result, { changes, unrecorded }) {
@@ -164,6 +176,25 @@ export async function createCommandTool({ commands, policy, writes, covered, bas
 function kilobytes(bytes) {
     return `${String(Math.ceil(bytes / 1024))} KB`;
 }
+/** The command line the model can see, before paths are resolved against the host workspace. */
+function visibleCommand(definition, values) {
+    const args = [];
+    for (const argument of definition.args ?? []) {
+        if (typeof argument === 'string') {
+            args.push(argument);
+            continue;
+        }
+        const value = values[argument.name] ?? null;
+        if (argument.type === 'paths') {
+            if (Array.isArray(value))
+                args.push(...value);
+        }
+        else if (value !== null && typeof value === 'string') {
+            args.push(value);
+        }
+    }
+    return [definition.executable, ...args].join(' ');
+}
 /** Its working directory, as a real path inside a `write` root. */
 async function resolveCwd(name, definition, policy, base) {
     const named = resolve(base, definition.cwd ?? '.');