@illuma-ai/agents 1.5.0 → 1.5.1

package/dist/cjs/tools/memory/memoryGetTool.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"memoryGetTool.cjs","sources":["../../../../src/tools/memory/memoryGetTool.ts"],"sourcesContent":["/**\n * `memory_get` — safe snippet read.\n *\n * Port of upstream `createMemoryGetTool` at\n * `upstream reference`.\n */\nimport { tool } from '@langchain/core/tools';\nimport {\n  MEMORY_GET_DESCRIPTION,\n  MEMORY_GET_TOOL_NAME,\n} from '@/memory/constants';\nimport {\n  MemoryGetSchema,\n  type MemoryGetToolResult,\n  type MemoryToolBinding,\n} from './shared';\n\n// eslint-disable-next-line @typescript-eslint/explicit-function-return-type\nexport function createMemoryGetTool(binding: MemoryToolBinding) {\n  return tool(\n    async (args): Promise<string> => {\n      try {\n        const result = await binding.backend.get(binding.scope, {\n          path: args.path,\n          from: args.from,\n          lines: args.lines,\n        });\n        if (!result) {\n          const payload: MemoryGetToolResult = { path: args.path, text: '' };\n          return JSON.stringify(payload);\n        }\n        const payload: MemoryGetToolResult = {\n          path: result.path,\n          text: result.text,\n        };\n        return JSON.stringify(payload);\n      } catch (err) {\n        const payload: MemoryGetToolResult = {\n          path: args.path,\n          text: '',\n          disabled: true,\n          error: err instanceof Error ? err.message : String(err),\n        };\n        return JSON.stringify(payload);\n      }\n    },\n    {\n      name: MEMORY_GET_TOOL_NAME,\n      description: MEMORY_GET_DESCRIPTION,\n      schema: MemoryGetSchema,\n    }\n  );\n}\n"],"names":["tool","MEMORY_GET_TOOL_NAME","MEMORY_GET_DESCRIPTION","MemoryGetSchema"],"mappings":";;;;;;AAAA;;;;;AAKG;AAYH;AACM,SAAU,mBAAmB,CAAC,OAA0B,EAAA;AAC5D,IAAA,OAAOA,UAAI,CACT,OAAO,IAAI,KAAqB;AAC9B,QAAA,IAAI;AACF,YAAA,MAAM,MAAM,GAAG,MAAM,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,KAAK,EAAE;gBACtD,IAAI,EAAE,IAAI,CAAC,IAAI;gBACf,IAAI,EAAE,IAAI,CAAC,IAAI;gBACf,KAAK,EAAE,IAAI,CAAC,KAAK;AAClB,aAAA,CAAC;YACF,IAAI,CAAC,MAAM,EAAE;AACX,gBAAA,MAAM,OAAO,GAAwB,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,IAAI,EAAE,EAAE,EAAE;AAClE,gBAAA,OAAO,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC;YAChC;AACA,YAAA,MAAM,OAAO,GAAwB;gBACnC,IAAI,EAAE,MAAM,CAAC,IAAI;gBACjB,IAAI,EAAE,MAAM,CAAC,IAAI;aAClB;AACD,YAAA,OAAO,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC;QAChC;QAAE,OAAO,GAAG,EAAE;AACZ,YAAA,MAAM,OAAO,GAAwB;gBACnC,IAAI,EAAE,IAAI,CAAC,IAAI;AACf,gBAAA,IAAI,EAAE,EAAE;AACR,gBAAA,QAAQ,EAAE,IAAI;AACd,gBAAA,KAAK,EAAE,GAAG,YAAY,KAAK,GAAG,GAAG,CAAC,OAAO,GAAG,MAAM,CAAC,GAAG,CAAC;aACxD;AACD,YAAA,OAAO,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC;QAChC;AACF,IAAA,CAAC,EACD;AACE,QAAA,IAAI,EAAEC,8BAAoB;AAC1B,QAAA,WAAW,EAAEC,gCAAsB;AACnC,QAAA,MAAM,EAAEC,sBAAe;AACxB,KAAA,CACF;AACH;;;;"}
+{"version":3,"file":"memoryGetTool.cjs","sources":["../../../../src/tools/memory/memoryGetTool.ts"],"sourcesContent":["/**\n * `memory_get` — safe snippet read.\n *\n * Port of `createMemoryGetTool` at\n * `reference`.\n */\nimport { tool } from '@langchain/core/tools';\nimport {\n  MEMORY_GET_DESCRIPTION,\n  MEMORY_GET_TOOL_NAME,\n} from '@/memory/constants';\nimport {\n  MemoryGetSchema,\n  type MemoryGetToolResult,\n  type MemoryToolBinding,\n} from './shared';\n\n// eslint-disable-next-line @typescript-eslint/explicit-function-return-type\nexport function createMemoryGetTool(binding: MemoryToolBinding) {\n  return tool(\n    async (args): Promise<string> => {\n      try {\n        const result = await binding.backend.get(binding.scope, {\n          path: args.path,\n          from: args.from,\n          lines: args.lines,\n        });\n        if (!result) {\n          const payload: MemoryGetToolResult = { path: args.path, text: '' };\n          return JSON.stringify(payload);\n        }\n        const payload: MemoryGetToolResult = {\n          path: result.path,\n          text: result.text,\n        };\n        return JSON.stringify(payload);\n      } catch (err) {\n        const payload: MemoryGetToolResult = {\n          path: args.path,\n          text: '',\n          disabled: true,\n          error: err instanceof Error ? err.message : String(err),\n        };\n        return JSON.stringify(payload);\n      }\n    },\n    {\n      name: MEMORY_GET_TOOL_NAME,\n      description: MEMORY_GET_DESCRIPTION,\n      schema: MemoryGetSchema,\n    }\n  );\n}\n"],"names":["tool","MEMORY_GET_TOOL_NAME","MEMORY_GET_DESCRIPTION","MemoryGetSchema"],"mappings":";;;;;;AAAA;;;;;AAKG;AAYH;AACM,SAAU,mBAAmB,CAAC,OAA0B,EAAA;AAC5D,IAAA,OAAOA,UAAI,CACT,OAAO,IAAI,KAAqB;AAC9B,QAAA,IAAI;AACF,YAAA,MAAM,MAAM,GAAG,MAAM,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,KAAK,EAAE;gBACtD,IAAI,EAAE,IAAI,CAAC,IAAI;gBACf,IAAI,EAAE,IAAI,CAAC,IAAI;gBACf,KAAK,EAAE,IAAI,CAAC,KAAK;AAClB,aAAA,CAAC;YACF,IAAI,CAAC,MAAM,EAAE;AACX,gBAAA,MAAM,OAAO,GAAwB,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,IAAI,EAAE,EAAE,EAAE;AAClE,gBAAA,OAAO,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC;YAChC;AACA,YAAA,MAAM,OAAO,GAAwB;gBACnC,IAAI,EAAE,MAAM,CAAC,IAAI;gBACjB,IAAI,EAAE,MAAM,CAAC,IAAI;aAClB;AACD,YAAA,OAAO,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC;QAChC;QAAE,OAAO,GAAG,EAAE;AACZ,YAAA,MAAM,OAAO,GAAwB;gBACnC,IAAI,EAAE,IAAI,CAAC,IAAI;AACf,gBAAA,IAAI,EAAE,EAAE;AACR,gBAAA,QAAQ,EAAE,IAAI;AACd,gBAAA,KAAK,EAAE,GAAG,YAAY,KAAK,GAAG,GAAG,CAAC,OAAO,GAAG,MAAM,CAAC,GAAG,CAAC;aACxD;AACD,YAAA,OAAO,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC;QAChC;AACF,IAAA,CAAC,EACD;AACE,QAAA,IAAI,EAAEC,8BAAoB;AAC1B,QAAA,WAAW,EAAEC,gCAAsB;AACnC,QAAA,MAAM,EAAEC,sBAAe;AACxB,KAAA,CACF;AACH;;;;"}

package/dist/cjs/tools/memory/memorySearchTool.cjs

@@ -7,9 +7,9 @@ var shared = require('./shared.cjs');
 /**
  * `memory_search` — the mandatory-recall tool.
  *
- * Port of upstream `createMemorySearchTool` at
- * `upstream reference`. The tool description
- * is verbatim "Mandatory recall step..." from upstream — this is the load-
+ * Port of `createMemorySearchTool` at
+ * `reference`. The tool description
+ * is verbatim "Mandatory recall step..." from a reference implementation — this is the load-
  * bearing line that turns "the agent may recall" into "the agent will recall".
  */
 // eslint-disable-next-line @typescript-eslint/explicit-function-return-type

package/dist/cjs/tools/memory/memorySearchTool.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"memorySearchTool.cjs","sources":["../../../../src/tools/memory/memorySearchTool.ts"],"sourcesContent":["/**\n * `memory_search` — the mandatory-recall tool.\n *\n * Port of upstream `createMemorySearchTool` at\n * `upstream reference`. The tool description\n * is verbatim \"Mandatory recall step...\" from upstream — this is the load-\n * bearing line that turns \"the agent may recall\" into \"the agent will recall\".\n */\nimport { tool } from '@langchain/core/tools';\nimport {\n  MEMORY_SEARCH_DESCRIPTION,\n  MEMORY_SEARCH_TOOL_NAME,\n} from '@/memory/constants';\nimport {\n  buildMemorySearchUnavailableResult,\n  clampResultsByInjectedChars,\n  MemorySearchSchema,\n  type MemorySearchToolResult,\n  type MemoryToolBinding,\n} from './shared';\n\n// eslint-disable-next-line @typescript-eslint/explicit-function-return-type\nexport function createMemorySearchTool(binding: MemoryToolBinding) {\n  return tool(\n    async (args): Promise<string> => {\n      try {\n        const entries = await binding.backend.search(\n          binding.scope,\n          args.query,\n          {\n            maxResults: args.maxResults,\n            minScore: args.minScore,\n            mmr: binding.searchOptions?.mmr,\n            temporalDecay: binding.searchOptions?.temporalDecay,\n            citations: binding.searchOptions?.citations,\n          }\n        );\n        const clamped = clampResultsByInjectedChars(\n          entries,\n          binding.maxInjectedChars\n        );\n\n        // [phase2-recall-tracking] debug: fire-and-forget best-effort record\n        if (binding.recallTracker && clamped.length > 0) {\n          void binding.recallTracker\n            .record({\n              agentId: binding.scope.agentId,\n              query: args.query,\n              hits: clamped.map((e) => ({\n                id: e.id,\n                path: e.path,\n                score: e.score,\n              })),\n            })\n            .catch(() => undefined);\n        }\n\n        const payload: MemorySearchToolResult = {\n          results: clamped.map((entry) => ({\n            id: entry.id,\n            path: entry.path,\n            content: entry.content,\n            score: entry.score,\n            createdAt: entry.createdAt,\n            citation: entry.citation,\n          })),\n        };\n        return JSON.stringify(payload);\n      } catch (err) {\n        const message = err instanceof Error ? err.message : String(err);\n        return JSON.stringify(buildMemorySearchUnavailableResult(message));\n      }\n    },\n    {\n      name: MEMORY_SEARCH_TOOL_NAME,\n      description: MEMORY_SEARCH_DESCRIPTION,\n      schema: MemorySearchSchema,\n    }\n  );\n}\n"],"names":["tool","clampResultsByInjectedChars","buildMemorySearchUnavailableResult","MEMORY_SEARCH_TOOL_NAME","MEMORY_SEARCH_DESCRIPTION","MemorySearchSchema"],"mappings":";;;;;;AAAA;;;;;;;AAOG;AAcH;AACM,SAAU,sBAAsB,CAAC,OAA0B,EAAA;AAC/D,IAAA,OAAOA,UAAI,CACT,OAAO,IAAI,KAAqB;AAC9B,QAAA,IAAI;AACF,YAAA,MAAM,OAAO,GAAG,MAAM,OAAO,CAAC,OAAO,CAAC,MAAM,CAC1C,OAAO,CAAC,KAAK,EACb,IAAI,CAAC,KAAK,EACV;gBACE,UAAU,EAAE,IAAI,CAAC,UAAU;gBAC3B,QAAQ,EAAE,IAAI,CAAC,QAAQ;AACvB,gBAAA,GAAG,EAAE,OAAO,CAAC,aAAa,EAAE,GAAG;AAC/B,gBAAA,aAAa,EAAE,OAAO,CAAC,aAAa,EAAE,aAAa;AACnD,gBAAA,SAAS,EAAE,OAAO,CAAC,aAAa,EAAE,SAAS;AAC5C,aAAA,CACF;YACD,MAAM,OAAO,GAAGC,kCAA2B,CACzC,OAAO,EACP,OAAO,CAAC,gBAAgB,CACzB;;YAGD,IAAI,OAAO,CAAC,aAAa,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE;gBAC/C,KAAK,OAAO,CAAC;AACV,qBAAA,MAAM,CAAC;AACN,oBAAA,OAAO,EAAE,OAAO,CAAC,KAAK,CAAC,OAAO;oBAC9B,KAAK,EAAE,IAAI,CAAC,KAAK;oBACjB,IAAI,EAAE,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM;wBACxB,EAAE,EAAE,CAAC,CAAC,EAAE;wBACR,IAAI,EAAE,CAAC,CAAC,IAAI;wBACZ,KAAK,EAAE,CAAC,CAAC,KAAK;AACf,qBAAA,CAAC,CAAC;iBACJ;AACA,qBAAA,KAAK,CAAC,MAAM,SAAS,CAAC;YAC3B;AAEA,YAAA,MAAM,OAAO,GAA2B;gBACtC,OAAO,EAAE,OAAO,CAAC,GAAG,CAAC,CAAC,KAAK,MAAM;oBAC/B,EAAE,EAAE,KAAK,CAAC,EAAE;oBACZ,IAAI,EAAE,KAAK,CAAC,IAAI;oBAChB,OAAO,EAAE,KAAK,CAAC,OAAO;oBACtB,KAAK,EAAE,KAAK,CAAC,KAAK;oBAClB,SAAS,EAAE,KAAK,CAAC,SAAS;oBAC1B,QAAQ,EAAE,KAAK,CAAC,QAAQ;AACzB,iBAAA,CAAC,CAAC;aACJ;AACD,YAAA,OAAO,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC;QAChC;QAAE,OAAO,GAAG,EAAE;AACZ,YAAA,MAAM,OAAO,GAAG,GAAG,YAAY,KAAK,GAAG,GAAG,CAAC,OAAO,GAAG,MAAM,CAAC,GAAG,CAAC;YAChE,OAAO,IAAI,CAAC,SAAS,CAACC,yCAAkC,CAAC,OAAO,CAAC,CAAC;QACpE;AACF,IAAA,CAAC,EACD;AACE,QAAA,IAAI,EAAEC,iCAAuB;AAC7B,QAAA,WAAW,EAAEC,mCAAyB;AACtC,QAAA,MAAM,EAAEC,yBAAkB;AAC3B,KAAA,CACF;AACH;;;;"}
+{"version":3,"file":"memorySearchTool.cjs","sources":["../../../../src/tools/memory/memorySearchTool.ts"],"sourcesContent":["/**\n * `memory_search` — the mandatory-recall tool.\n *\n * Port of `createMemorySearchTool` at\n * `reference`. The tool description\n * is verbatim \"Mandatory recall step...\" from a reference implementation — this is the load-\n * bearing line that turns \"the agent may recall\" into \"the agent will recall\".\n */\nimport { tool } from '@langchain/core/tools';\nimport {\n  MEMORY_SEARCH_DESCRIPTION,\n  MEMORY_SEARCH_TOOL_NAME,\n} from '@/memory/constants';\nimport {\n  buildMemorySearchUnavailableResult,\n  clampResultsByInjectedChars,\n  MemorySearchSchema,\n  type MemorySearchToolResult,\n  type MemoryToolBinding,\n} from './shared';\n\n// eslint-disable-next-line @typescript-eslint/explicit-function-return-type\nexport function createMemorySearchTool(binding: MemoryToolBinding) {\n  return tool(\n    async (args): Promise<string> => {\n      try {\n        const entries = await binding.backend.search(\n          binding.scope,\n          args.query,\n          {\n            maxResults: args.maxResults,\n            minScore: args.minScore,\n            mmr: binding.searchOptions?.mmr,\n            temporalDecay: binding.searchOptions?.temporalDecay,\n            citations: binding.searchOptions?.citations,\n          }\n        );\n        const clamped = clampResultsByInjectedChars(\n          entries,\n          binding.maxInjectedChars\n        );\n\n        // [phase2-recall-tracking] debug: fire-and-forget best-effort record\n        if (binding.recallTracker && clamped.length > 0) {\n          void binding.recallTracker\n            .record({\n              agentId: binding.scope.agentId,\n              query: args.query,\n              hits: clamped.map((e) => ({\n                id: e.id,\n                path: e.path,\n                score: e.score,\n              })),\n            })\n            .catch(() => undefined);\n        }\n\n        const payload: MemorySearchToolResult = {\n          results: clamped.map((entry) => ({\n            id: entry.id,\n            path: entry.path,\n            content: entry.content,\n            score: entry.score,\n            createdAt: entry.createdAt,\n            citation: entry.citation,\n          })),\n        };\n        return JSON.stringify(payload);\n      } catch (err) {\n        const message = err instanceof Error ? err.message : String(err);\n        return JSON.stringify(buildMemorySearchUnavailableResult(message));\n      }\n    },\n    {\n      name: MEMORY_SEARCH_TOOL_NAME,\n      description: MEMORY_SEARCH_DESCRIPTION,\n      schema: MemorySearchSchema,\n    }\n  );\n}\n"],"names":["tool","clampResultsByInjectedChars","buildMemorySearchUnavailableResult","MEMORY_SEARCH_TOOL_NAME","MEMORY_SEARCH_DESCRIPTION","MemorySearchSchema"],"mappings":";;;;;;AAAA;;;;;;;AAOG;AAcH;AACM,SAAU,sBAAsB,CAAC,OAA0B,EAAA;AAC/D,IAAA,OAAOA,UAAI,CACT,OAAO,IAAI,KAAqB;AAC9B,QAAA,IAAI;AACF,YAAA,MAAM,OAAO,GAAG,MAAM,OAAO,CAAC,OAAO,CAAC,MAAM,CAC1C,OAAO,CAAC,KAAK,EACb,IAAI,CAAC,KAAK,EACV;gBACE,UAAU,EAAE,IAAI,CAAC,UAAU;gBAC3B,QAAQ,EAAE,IAAI,CAAC,QAAQ;AACvB,gBAAA,GAAG,EAAE,OAAO,CAAC,aAAa,EAAE,GAAG;AAC/B,gBAAA,aAAa,EAAE,OAAO,CAAC,aAAa,EAAE,aAAa;AACnD,gBAAA,SAAS,EAAE,OAAO,CAAC,aAAa,EAAE,SAAS;AAC5C,aAAA,CACF;YACD,MAAM,OAAO,GAAGC,kCAA2B,CACzC,OAAO,EACP,OAAO,CAAC,gBAAgB,CACzB;;YAGD,IAAI,OAAO,CAAC,aAAa,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE;gBAC/C,KAAK,OAAO,CAAC;AACV,qBAAA,MAAM,CAAC;AACN,oBAAA,OAAO,EAAE,OAAO,CAAC,KAAK,CAAC,OAAO;oBAC9B,KAAK,EAAE,IAAI,CAAC,KAAK;oBACjB,IAAI,EAAE,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM;wBACxB,EAAE,EAAE,CAAC,CAAC,EAAE;wBACR,IAAI,EAAE,CAAC,CAAC,IAAI;wBACZ,KAAK,EAAE,CAAC,CAAC,KAAK;AACf,qBAAA,CAAC,CAAC;iBACJ;AACA,qBAAA,KAAK,CAAC,MAAM,SAAS,CAAC;YAC3B;AAEA,YAAA,MAAM,OAAO,GAA2B;gBACtC,OAAO,EAAE,OAAO,CAAC,GAAG,CAAC,CAAC,KAAK,MAAM;oBAC/B,EAAE,EAAE,KAAK,CAAC,EAAE;oBACZ,IAAI,EAAE,KAAK,CAAC,IAAI;oBAChB,OAAO,EAAE,KAAK,CAAC,OAAO;oBACtB,KAAK,EAAE,KAAK,CAAC,KAAK;oBAClB,SAAS,EAAE,KAAK,CAAC,SAAS;oBAC1B,QAAQ,EAAE,KAAK,CAAC,QAAQ;AACzB,iBAAA,CAAC,CAAC;aACJ;AACD,YAAA,OAAO,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC;QAChC;QAAE,OAAO,GAAG,EAAE;AACZ,YAAA,MAAM,OAAO,GAAG,GAAG,YAAY,KAAK,GAAG,GAAG,CAAC,OAAO,GAAG,MAAM,CAAC,GAAG,CAAC;YAChE,OAAO,IAAI,CAAC,SAAS,CAACC,yCAAkC,CAAC,OAAO,CAAC,CAAC;QACpE;AACF,IAAA,CAAC,EACD;AACE,QAAA,IAAI,EAAEC,iCAAuB;AAC7B,QAAA,WAAW,EAAEC,mCAAyB;AACtC,QAAA,MAAM,EAAEC,yBAAkB;AAC3B,KAAA,CACF;AACH;;;;"}

package/dist/cjs/tools/memory/shared.cjs

@@ -71,7 +71,7 @@ function buildMemorySearchUnavailableResult(error) {
 }
 /**
  * Clamp a ranked result list to a total character budget.
- * Ported from upstream `tools.citations.ts::clampResultsByInjectedChars`.
+ * Ported from a reference implementation `tools.citations.ts::clampResultsByInjectedChars`.
  */
 function clampResultsByInjectedChars(results, maxInjectedChars = constants.DEFAULT_MAX_INJECTED_CHARS) {
     const out = [];

package/dist/cjs/tools/memory/shared.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"shared.cjs","sources":["../../../../src/tools/memory/shared.ts"],"sourcesContent":["/**\n * Shared Zod schemas + helpers for the memory tool family.\n *\n * The tool schemas deliberately do NOT expose `agent_id` or `user_id`. Scope\n * is captured in a closure at tool construction time by {@link buildMemoryTools},\n * so even a hallucinated tool call with extra fields cannot influence it.\n */\nimport { z } from 'zod';\nimport {\n  DEFAULT_MAX_INJECTED_CHARS,\n  MEMORY_PATH_PREFIX,\n} from '@/memory/constants';\nimport type {\n  MemoryAppendInput,\n  MemoryBackend,\n  MemoryEntry,\n  MemoryScope,\n} from '@/memory/types';\n\nexport const MemorySearchSchema = z.object({\n  query: z.string().describe('Natural-language query to search memory for'),\n  maxResults: z\n    .number()\n    .int()\n    .min(1)\n    .max(50)\n    .optional()\n    .describe('Maximum number of results to return (default 10)'),\n  minScore: z\n    .number()\n    .min(0)\n    .max(1)\n    .optional()\n    .describe('Minimum hybrid score (0..1) below which results are dropped'),\n});\n\nexport const MemoryGetSchema = z.object({\n  path: z\n    .string()\n    .describe(\n      'Memory path, e.g. \"memory/decisions.md\". Must start with \"memory/\"'\n    ),\n  from: z\n    .number()\n    .int()\n    .min(1)\n    .optional()\n    .describe('Starting line number (1-indexed)'),\n  lines: z\n    .number()\n    .int()\n    .min(1)\n    .optional()\n    .describe('Number of lines to read starting at `from`'),\n});\n\nexport const MemoryAppendSchema = z.object({\n  path: z\n    .string()\n    .describe(\n      'Memory path to append to (e.g. \"memory/decisions.md\"). Must start with \"memory/\"'\n    ),\n  content: z\n    .string()\n    .min(1)\n    .describe(\n      \"Note content in the agent's own voice. Markdown allowed. Each call appends an immutable entry.\"\n    ),\n});\n\nexport type MemorySearchArgs = z.infer<typeof MemorySearchSchema>;\nexport type MemoryGetArgs = z.infer<typeof MemoryGetSchema>;\nexport type MemoryAppendArgs = z.infer<typeof MemoryAppendSchema>;\n\n/** Shape returned by `memory_search` — serialised as the tool result string. */\nexport interface MemorySearchToolResult {\n  results: Array<\n    Pick<MemoryEntry, 'id' | 'path' | 'content' | 'score' | 'createdAt'> & {\n      citation?: string;\n    }\n  >;\n  disabled?: boolean;\n  unavailable?: boolean;\n  error?: string;\n  warning?: string;\n  action?: string;\n}\n\n/** Shape returned by `memory_get`. */\nexport interface MemoryGetToolResult {\n  path: string;\n  text: string;\n  disabled?: boolean;\n  error?: string;\n}\n\nexport function buildMemorySearchUnavailableResult(\n  error: string | undefined\n): MemorySearchToolResult {\n  const reason =\n    (error ?? 'memory search unavailable').trim() ||\n    'memory search unavailable';\n  const isQuota = /insufficient_quota|quota|429/i.test(reason);\n  return {\n    results: [],\n    disabled: true,\n    unavailable: true,\n    error: reason,\n    warning: isQuota\n      ? 'Memory search is unavailable because the embedding provider quota is exhausted.'\n      : 'Memory search is unavailable due to an embedding/provider error.',\n    action: isQuota\n      ? 'Top up or switch embedding provider, then retry memory_search.'\n      : 'Check embedding provider configuration and retry memory_search.',\n  };\n}\n\n/**\n * Clamp a ranked result list to a total character budget.\n * Ported from upstream `tools.citations.ts::clampResultsByInjectedChars`.\n */\nexport function clampResultsByInjectedChars<T extends { content: string }>(\n  results: T[],\n  maxInjectedChars: number = DEFAULT_MAX_INJECTED_CHARS\n): T[] {\n  const out: T[] = [];\n  let total = 0;\n  for (const result of results) {\n    const size = result.content.length ?? 0;\n    if (total + size > maxInjectedChars && out.length > 0) break;\n    out.push(result);\n    total += size;\n  }\n  return out;\n}\n\nexport interface MemoryToolBinding {\n  backend: MemoryBackend;\n  scope: MemoryScope;\n  maxInjectedChars?: number;\n  /** Phase 2 — per-call search options propagated into backend.search. */\n  searchOptions?: {\n    mmr?: { enabled?: boolean; lambda?: number };\n    temporalDecay?: { enabled?: boolean; halfLifeDays?: number };\n    citations?: 'on' | 'off' | 'auto';\n  };\n  /** Phase 2 — optional best-effort recall recorder. */\n  recallTracker?: {\n    record(params: {\n      agentId: string;\n      query: string;\n      hits: Array<{ id: string; path: string; score: number }>;\n    }): Promise<void>;\n  };\n}\n\nexport function assertAppendAllowed(path: string): void {\n  if (!path || !path.startsWith(MEMORY_PATH_PREFIX)) {\n    throw new Error(\n      `memory_append path must start with \"${MEMORY_PATH_PREFIX}\"`\n    );\n  }\n}\n\n/** Normalise an append call into the backend input shape. */\nexport function toAppendInput(args: MemoryAppendArgs): MemoryAppendInput {\n  assertAppendAllowed(args.path);\n  return { path: args.path, content: args.content };\n}\n"],"names":["z","DEFAULT_MAX_INJECTED_CHARS","MEMORY_PATH_PREFIX"],"mappings":";;;;;AAAA;;;;;;AAMG;AAaI,MAAM,kBAAkB,GAAGA,KAAC,CAAC,MAAM,CAAC;IACzC,KAAK,EAAEA,KAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,6CAA6C,CAAC;AACzE,IAAA,UAAU,EAAEA;AACT,SAAA,MAAM;AACN,SAAA,GAAG;SACH,GAAG,CAAC,CAAC;SACL,GAAG,CAAC,EAAE;AACN,SAAA,QAAQ;SACR,QAAQ,CAAC,kDAAkD,CAAC;AAC/D,IAAA,QAAQ,EAAEA;AACP,SAAA,MAAM;SACN,GAAG,CAAC,CAAC;SACL,GAAG,CAAC,CAAC;AACL,SAAA,QAAQ;SACR,QAAQ,CAAC,6DAA6D,CAAC;AAC3E,CAAA;AAEM,MAAM,eAAe,GAAGA,KAAC,CAAC,MAAM,CAAC;AACtC,IAAA,IAAI,EAAEA;AACH,SAAA,MAAM;SACN,QAAQ,CACP,oEAAoE,CACrE;AACH,IAAA,IAAI,EAAEA;AACH,SAAA,MAAM;AACN,SAAA,GAAG;SACH,GAAG,CAAC,CAAC;AACL,SAAA,QAAQ;SACR,QAAQ,CAAC,kCAAkC,CAAC;AAC/C,IAAA,KAAK,EAAEA;AACJ,SAAA,MAAM;AACN,SAAA,GAAG;SACH,GAAG,CAAC,CAAC;AACL,SAAA,QAAQ;SACR,QAAQ,CAAC,4CAA4C,CAAC;AAC1D,CAAA;AAEM,MAAM,kBAAkB,GAAGA,KAAC,CAAC,MAAM,CAAC;AACzC,IAAA,IAAI,EAAEA;AACH,SAAA,MAAM;SACN,QAAQ,CACP,kFAAkF,CACnF;AACH,IAAA,OAAO,EAAEA;AACN,SAAA,MAAM;SACN,GAAG,CAAC,CAAC;SACL,QAAQ,CACP,gGAAgG,CACjG;AACJ,CAAA;AA4BK,SAAU,kCAAkC,CAChD,KAAyB,EAAA;IAEzB,MAAM,MAAM,GACV,CAAC,KAAK,IAAI,2BAA2B,EAAE,IAAI,EAAE;AAC7C,QAAA,2BAA2B;IAC7B,MAAM,OAAO,GAAG,+BAA+B,CAAC,IAAI,CAAC,MAAM,CAAC;IAC5D,OAAO;AACL,QAAA,OAAO,EAAE,EAAE;AACX,QAAA,QAAQ,EAAE,IAAI;AACd,QAAA,WAAW,EAAE,IAAI;AACjB,QAAA,KAAK,EAAE,MAAM;AACb,QAAA,OAAO,EAAE;AACP,cAAE;AACF,cAAE,kEAAkE;AACtE,QAAA,MAAM,EAAE;AACN,cAAE;AACF,cAAE,iEAAiE;KACtE;AACH;AAEA;;;AAGG;SACa,2BAA2B,CACzC,OAAY,EACZ,mBAA2BC,oCAA0B,EAAA;IAErD,MAAM,GAAG,GAAQ,EAAE;IACnB,IAAI,KAAK,GAAG,CAAC;AACb,IAAA,KAAK,MAAM,MAAM,IAAI,OAAO,EAAE;QAC5B,MAAM,IAAI,GAAG,MAAM,CAAC,OAAO,CAAC,MAAM,IAAI,CAAC;QACvC,IAAI,KAAK,GAAG,IAAI,GAAG,gBAAgB,IAAI,GAAG,CAAC,MAAM,GAAG,CAAC;YAAE;AACvD,QAAA,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC;QAChB,KAAK,IAAI,IAAI;IACf;AACA,IAAA,OAAO,GAAG;AACZ;AAsBM,SAAU,mBAAmB,CAAC,IAAY,EAAA;IAC9C,IAAI,CAAC,IAAI,IAAI,CAAC,IAAI,CAAC,UAAU,CAACC,4BAAkB,CAAC,EAAE;AACjD,QAAA,MAAM,IAAI,KAAK,CACb,uCAAuCA,4BAAkB,CAAA,CAAA,CAAG,CAC7D;IACH;AACF;AAEA;AACM,SAAU,aAAa,CAAC,IAAsB,EAAA;AAClD,IAAA,mBAAmB,CAAC,IAAI,CAAC,IAAI,CAAC;AAC9B,IAAA,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE;AACnD;;;;;;;;;;"}
+{"version":3,"file":"shared.cjs","sources":["../../../../src/tools/memory/shared.ts"],"sourcesContent":["/**\n * Shared Zod schemas + helpers for the memory tool family.\n *\n * The tool schemas deliberately do NOT expose `agent_id` or `user_id`. Scope\n * is captured in a closure at tool construction time by {@link buildMemoryTools},\n * so even a hallucinated tool call with extra fields cannot influence it.\n */\nimport { z } from 'zod';\nimport {\n  DEFAULT_MAX_INJECTED_CHARS,\n  MEMORY_PATH_PREFIX,\n} from '@/memory/constants';\nimport type {\n  MemoryAppendInput,\n  MemoryBackend,\n  MemoryEntry,\n  MemoryScope,\n} from '@/memory/types';\n\nexport const MemorySearchSchema = z.object({\n  query: z.string().describe('Natural-language query to search memory for'),\n  maxResults: z\n    .number()\n    .int()\n    .min(1)\n    .max(50)\n    .optional()\n    .describe('Maximum number of results to return (default 10)'),\n  minScore: z\n    .number()\n    .min(0)\n    .max(1)\n    .optional()\n    .describe('Minimum hybrid score (0..1) below which results are dropped'),\n});\n\nexport const MemoryGetSchema = z.object({\n  path: z\n    .string()\n    .describe(\n      'Memory path, e.g. \"memory/decisions.md\". Must start with \"memory/\"'\n    ),\n  from: z\n    .number()\n    .int()\n    .min(1)\n    .optional()\n    .describe('Starting line number (1-indexed)'),\n  lines: z\n    .number()\n    .int()\n    .min(1)\n    .optional()\n    .describe('Number of lines to read starting at `from`'),\n});\n\nexport const MemoryAppendSchema = z.object({\n  path: z\n    .string()\n    .describe(\n      'Memory path to append to (e.g. \"memory/decisions.md\"). Must start with \"memory/\"'\n    ),\n  content: z\n    .string()\n    .min(1)\n    .describe(\n      \"Note content in the agent's own voice. Markdown allowed. Each call appends an immutable entry.\"\n    ),\n});\n\nexport type MemorySearchArgs = z.infer<typeof MemorySearchSchema>;\nexport type MemoryGetArgs = z.infer<typeof MemoryGetSchema>;\nexport type MemoryAppendArgs = z.infer<typeof MemoryAppendSchema>;\n\n/** Shape returned by `memory_search` — serialised as the tool result string. */\nexport interface MemorySearchToolResult {\n  results: Array<\n    Pick<MemoryEntry, 'id' | 'path' | 'content' | 'score' | 'createdAt'> & {\n      citation?: string;\n    }\n  >;\n  disabled?: boolean;\n  unavailable?: boolean;\n  error?: string;\n  warning?: string;\n  action?: string;\n}\n\n/** Shape returned by `memory_get`. */\nexport interface MemoryGetToolResult {\n  path: string;\n  text: string;\n  disabled?: boolean;\n  error?: string;\n}\n\nexport function buildMemorySearchUnavailableResult(\n  error: string | undefined\n): MemorySearchToolResult {\n  const reason =\n    (error ?? 'memory search unavailable').trim() ||\n    'memory search unavailable';\n  const isQuota = /insufficient_quota|quota|429/i.test(reason);\n  return {\n    results: [],\n    disabled: true,\n    unavailable: true,\n    error: reason,\n    warning: isQuota\n      ? 'Memory search is unavailable because the embedding provider quota is exhausted.'\n      : 'Memory search is unavailable due to an embedding/provider error.',\n    action: isQuota\n      ? 'Top up or switch embedding provider, then retry memory_search.'\n      : 'Check embedding provider configuration and retry memory_search.',\n  };\n}\n\n/**\n * Clamp a ranked result list to a total character budget.\n * Ported from a reference implementation `tools.citations.ts::clampResultsByInjectedChars`.\n */\nexport function clampResultsByInjectedChars<T extends { content: string }>(\n  results: T[],\n  maxInjectedChars: number = DEFAULT_MAX_INJECTED_CHARS\n): T[] {\n  const out: T[] = [];\n  let total = 0;\n  for (const result of results) {\n    const size = result.content.length ?? 0;\n    if (total + size > maxInjectedChars && out.length > 0) break;\n    out.push(result);\n    total += size;\n  }\n  return out;\n}\n\nexport interface MemoryToolBinding {\n  backend: MemoryBackend;\n  scope: MemoryScope;\n  maxInjectedChars?: number;\n  /** Phase 2 — per-call search options propagated into backend.search. */\n  searchOptions?: {\n    mmr?: { enabled?: boolean; lambda?: number };\n    temporalDecay?: { enabled?: boolean; halfLifeDays?: number };\n    citations?: 'on' | 'off' | 'auto';\n  };\n  /** Phase 2 — optional best-effort recall recorder. */\n  recallTracker?: {\n    record(params: {\n      agentId: string;\n      query: string;\n      hits: Array<{ id: string; path: string; score: number }>;\n    }): Promise<void>;\n  };\n}\n\nexport function assertAppendAllowed(path: string): void {\n  if (!path || !path.startsWith(MEMORY_PATH_PREFIX)) {\n    throw new Error(\n      `memory_append path must start with \"${MEMORY_PATH_PREFIX}\"`\n    );\n  }\n}\n\n/** Normalise an append call into the backend input shape. */\nexport function toAppendInput(args: MemoryAppendArgs): MemoryAppendInput {\n  assertAppendAllowed(args.path);\n  return { path: args.path, content: args.content };\n}\n"],"names":["z","DEFAULT_MAX_INJECTED_CHARS","MEMORY_PATH_PREFIX"],"mappings":";;;;;AAAA;;;;;;AAMG;AAaI,MAAM,kBAAkB,GAAGA,KAAC,CAAC,MAAM,CAAC;IACzC,KAAK,EAAEA,KAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,6CAA6C,CAAC;AACzE,IAAA,UAAU,EAAEA;AACT,SAAA,MAAM;AACN,SAAA,GAAG;SACH,GAAG,CAAC,CAAC;SACL,GAAG,CAAC,EAAE;AACN,SAAA,QAAQ;SACR,QAAQ,CAAC,kDAAkD,CAAC;AAC/D,IAAA,QAAQ,EAAEA;AACP,SAAA,MAAM;SACN,GAAG,CAAC,CAAC;SACL,GAAG,CAAC,CAAC;AACL,SAAA,QAAQ;SACR,QAAQ,CAAC,6DAA6D,CAAC;AAC3E,CAAA;AAEM,MAAM,eAAe,GAAGA,KAAC,CAAC,MAAM,CAAC;AACtC,IAAA,IAAI,EAAEA;AACH,SAAA,MAAM;SACN,QAAQ,CACP,oEAAoE,CACrE;AACH,IAAA,IAAI,EAAEA;AACH,SAAA,MAAM;AACN,SAAA,GAAG;SACH,GAAG,CAAC,CAAC;AACL,SAAA,QAAQ;SACR,QAAQ,CAAC,kCAAkC,CAAC;AAC/C,IAAA,KAAK,EAAEA;AACJ,SAAA,MAAM;AACN,SAAA,GAAG;SACH,GAAG,CAAC,CAAC;AACL,SAAA,QAAQ;SACR,QAAQ,CAAC,4CAA4C,CAAC;AAC1D,CAAA;AAEM,MAAM,kBAAkB,GAAGA,KAAC,CAAC,MAAM,CAAC;AACzC,IAAA,IAAI,EAAEA;AACH,SAAA,MAAM;SACN,QAAQ,CACP,kFAAkF,CACnF;AACH,IAAA,OAAO,EAAEA;AACN,SAAA,MAAM;SACN,GAAG,CAAC,CAAC;SACL,QAAQ,CACP,gGAAgG,CACjG;AACJ,CAAA;AA4BK,SAAU,kCAAkC,CAChD,KAAyB,EAAA;IAEzB,MAAM,MAAM,GACV,CAAC,KAAK,IAAI,2BAA2B,EAAE,IAAI,EAAE;AAC7C,QAAA,2BAA2B;IAC7B,MAAM,OAAO,GAAG,+BAA+B,CAAC,IAAI,CAAC,MAAM,CAAC;IAC5D,OAAO;AACL,QAAA,OAAO,EAAE,EAAE;AACX,QAAA,QAAQ,EAAE,IAAI;AACd,QAAA,WAAW,EAAE,IAAI;AACjB,QAAA,KAAK,EAAE,MAAM;AACb,QAAA,OAAO,EAAE;AACP,cAAE;AACF,cAAE,kEAAkE;AACtE,QAAA,MAAM,EAAE;AACN,cAAE;AACF,cAAE,iEAAiE;KACtE;AACH;AAEA;;;AAGG;SACa,2BAA2B,CACzC,OAAY,EACZ,mBAA2BC,oCAA0B,EAAA;IAErD,MAAM,GAAG,GAAQ,EAAE;IACnB,IAAI,KAAK,GAAG,CAAC;AACb,IAAA,KAAK,MAAM,MAAM,IAAI,OAAO,EAAE;QAC5B,MAAM,IAAI,GAAG,MAAM,CAAC,OAAO,CAAC,MAAM,IAAI,CAAC;QACvC,IAAI,KAAK,GAAG,IAAI,GAAG,gBAAgB,IAAI,GAAG,CAAC,MAAM,GAAG,CAAC;YAAE;AACvD,QAAA,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC;QAChB,KAAK,IAAI,IAAI;IACf;AACA,IAAA,OAAO,GAAG;AACZ;AAsBM,SAAU,mBAAmB,CAAC,IAAY,EAAA;IAC9C,IAAI,CAAC,IAAI,IAAI,CAAC,IAAI,CAAC,UAAU,CAACC,4BAAkB,CAAC,EAAE;AACjD,QAAA,MAAM,IAAI,KAAK,CACb,uCAAuCA,4BAAkB,CAAA,CAAA,CAAG,CAC7D;IACH;AACF;AAEA;AACM,SAAU,aAAa,CAAC,IAAsB,EAAA;AAClD,IAAA,mBAAmB,CAAC,IAAI,CAAC,IAAI,CAAC;AAC9B,IAAA,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE;AACnD;;;;;;;;;;"}

package/dist/cjs/types/agent-cache.cjs

@@ -11,13 +11,14 @@
  * for the active provider; consumers decide what content to put in each
  * block.
  *
- * Why this is generic rather than ranger-specific
+ * Why this is generic rather than tier-named
  * ------------------------------------------------
  * Earlier iterations exposed `platform_instructions` and `cache_ttl:
- * { platform, agent }` — vocabulary borrowed from a multi-tenant chat
- * platform's mental model. That leaked a specific consumer's domain into
- * the library and made it confusing for other consumers (CLI tools,
- * desktop apps, gateways) which have no notion of "platforms" or "agents".
+ * { platform, agent }` — vocabulary borrowed from a specific consumer's
+ * multi-tenant chat-platform mental model. That leaked one consumer's
+ * domain into the library and made it confusing for other consumers
+ * (CLI tools, desktop apps, gateways) which have no notion of
+ * "platforms" or "agents".
  *
  * The current shape lets every consumer define their own cache-tier
  * organization without the library knowing or caring.

package/dist/cjs/types/agent-cache.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"agent-cache.cjs","sources":["../../../src/types/agent-cache.ts"],"sourcesContent":["/**\n * Agent system-message prompt-cache types.\n * ==========================================\n *\n * `@illuma-ai/agents` exposes a generic primitive — `system_cache_blocks` —\n * for composing a system message that gets explicit cache markers\n * (cachePoint on Bedrock, cache_control on Anthropic) at consumer-defined\n * boundaries. The library encodes each entry into the right wire format\n * for the active provider; consumers decide what content to put in each\n * block.\n *\n * Why this is generic rather than ranger-specific\n * ------------------------------------------------\n * Earlier iterations exposed `platform_instructions` and `cache_ttl:\n * { platform, agent }` — vocabulary borrowed from a multi-tenant chat\n * platform's mental model. That leaked a specific consumer's domain into\n * the library and made it confusing for other consumers (CLI tools,\n * desktop apps, gateways) which have no notion of \"platforms\" or \"agents\".\n *\n * The current shape lets every consumer define their own cache-tier\n * organization without the library knowing or caring.\n *\n * Cache key composition\n * ---------------------\n * Both Anthropic and Bedrock hash the message bytes leading up to each\n * cache marker. So entry N's cache key is the concatenation of bytes\n * 0..N. Stable content (universal rules, framework docs) goes earlier;\n * volatile content (per-agent identity, per-tool docs) goes later. The\n * earlier entries get cross-tenant cache hits; the later entries get\n * narrower hits.\n *\n * Provider budget\n * ---------------\n * Bedrock's Converse API supports up to 4 cachePoint blocks per request.\n * The default tools-array cache strategy (see `IllumaBedrockConverse`)\n * uses up to 2 of those. So at most 2 entries should appear in\n * `system_cache_blocks` to leave budget for the trailing\n * `instructions` block. The library throws on more than 2.\n *\n * Anthropic supports up to 4 cache breakpoints per request workspace-wide;\n * the same constraint applies to keep tools + system within budget.\n */\n\n/** Provider TTL hint applied to a cachePoint / cache_control marker. */\nexport type AgentCacheTTL = '5m' | '1h';\n\n/**\n * One cacheable system content block. Each entry produces a text block\n * followed by a provider-specific cache marker in the assembled system\n * message. Entries are emitted in array order — earlier = more stable\n * = wider cache key.\n */\nexport interface SystemCacheBlock {\n  /**\n   * Text payload. Caller is responsible for keeping this byte-stable\n   * across requests that should share a cache entry. ANY change to\n   * these bytes invalidates every cache key that includes this entry.\n   */\n  text: string;\n  /**\n   * Optional TTL hint for the cache marker. Defaults to '5m' when\n   * omitted. '1h' costs more to write (2× base vs 1.25×) but survives\n   * longer idle, which is the right tradeoff for low-traffic blocks.\n   */\n  ttl?: AgentCacheTTL;\n}\n\n/**\n * Maximum number of `system_cache_blocks` entries the library accepts.\n * Higher than this would exceed Bedrock's 4-cachePoint budget once the\n * tools array and trailing `instructions` block are counted.\n */\nexport const MAX_SYSTEM_CACHE_BLOCKS = 2;\n"],"names":[],"mappings":";;AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyCG;AA0BH;;;;AAIG;AACI,MAAM,uBAAuB,GAAG;;;;"}
+{"version":3,"file":"agent-cache.cjs","sources":["../../../src/types/agent-cache.ts"],"sourcesContent":["/**\n * Agent system-message prompt-cache types.\n * ==========================================\n *\n * `@illuma-ai/agents` exposes a generic primitive — `system_cache_blocks` —\n * for composing a system message that gets explicit cache markers\n * (cachePoint on Bedrock, cache_control on Anthropic) at consumer-defined\n * boundaries. The library encodes each entry into the right wire format\n * for the active provider; consumers decide what content to put in each\n * block.\n *\n * Why this is generic rather than tier-named\n * ------------------------------------------------\n * Earlier iterations exposed `platform_instructions` and `cache_ttl:\n * { platform, agent }` — vocabulary borrowed from a specific consumer's\n * multi-tenant chat-platform mental model. That leaked one consumer's\n * domain into the library and made it confusing for other consumers\n * (CLI tools, desktop apps, gateways) which have no notion of\n * \"platforms\" or \"agents\".\n *\n * The current shape lets every consumer define their own cache-tier\n * organization without the library knowing or caring.\n *\n * Cache key composition\n * ---------------------\n * Both Anthropic and Bedrock hash the message bytes leading up to each\n * cache marker. So entry N's cache key is the concatenation of bytes\n * 0..N. Stable content (universal rules, framework docs) goes earlier;\n * volatile content (per-agent identity, per-tool docs) goes later. The\n * earlier entries get cross-tenant cache hits; the later entries get\n * narrower hits.\n *\n * Provider budget\n * ---------------\n * Bedrock's Converse API supports up to 4 cachePoint blocks per request.\n * The default tools-array cache strategy (see `IllumaBedrockConverse`)\n * uses up to 2 of those. So at most 2 entries should appear in\n * `system_cache_blocks` to leave budget for the trailing\n * `instructions` block. The library throws on more than 2.\n *\n * Anthropic supports up to 4 cache breakpoints per request workspace-wide;\n * the same constraint applies to keep tools + system within budget.\n */\n\n/** Provider TTL hint applied to a cachePoint / cache_control marker. */\nexport type AgentCacheTTL = '5m' | '1h';\n\n/**\n * One cacheable system content block. Each entry produces a text block\n * followed by a provider-specific cache marker in the assembled system\n * message. Entries are emitted in array order — earlier = more stable\n * = wider cache key.\n */\nexport interface SystemCacheBlock {\n  /**\n   * Text payload. Caller is responsible for keeping this byte-stable\n   * across requests that should share a cache entry. ANY change to\n   * these bytes invalidates every cache key that includes this entry.\n   */\n  text: string;\n  /**\n   * Optional TTL hint for the cache marker. Defaults to '5m' when\n   * omitted. '1h' costs more to write (2× base vs 1.25×) but survives\n   * longer idle, which is the right tradeoff for low-traffic blocks.\n   */\n  ttl?: AgentCacheTTL;\n}\n\n/**\n * Maximum number of `system_cache_blocks` entries the library accepts.\n * Higher than this would exceed Bedrock's 4-cachePoint budget once the\n * tools array and trailing `instructions` block are counted.\n */\nexport const MAX_SYSTEM_CACHE_BLOCKS = 2;\n"],"names":[],"mappings":";;AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0CG;AA0BH;;;;AAIG;AACI,MAAM,uBAAuB,GAAG;;;;"}

package/dist/cjs/types/graph.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"graph.cjs","sources":["../../../src/types/graph.ts"],"sourcesContent":["// src/types/graph.ts\nimport type {\n  START,\n  StateType,\n  UpdateType,\n  StateGraph,\n  StateGraphArgs,\n  StateDefinition,\n  CompiledStateGraph,\n  BinaryOperatorAggregate,\n} from '@langchain/langgraph';\nimport type { BindToolsInput } from '@langchain/core/language_models/chat_models';\nimport type {\n  BaseMessage,\n  AIMessageChunk,\n  SystemMessage,\n} from '@langchain/core/messages';\nimport type { RunnableConfig, Runnable } from '@langchain/core/runnables';\nimport type { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { GoogleAIToolType } from '@langchain/google-common';\nimport type {\n  ToolMap,\n  ToolEndEvent,\n  GenericTool,\n  LCTool,\n  ToolApprovalConfig,\n  ToolExecuteBatchRequest,\n} from '@/types/tools';\nimport type { Providers, Callback, GraphNodeKeys } from '@/common';\nimport type { StandardGraph, MultiAgentGraph } from '@/graphs';\nimport type { ClientOptions } from '@/types/llm';\nimport type {\n  RunStep,\n  RunStepDeltaEvent,\n  MessageDeltaEvent,\n  ReasoningDeltaEvent,\n} from '@/types/stream';\nimport type { TokenCounter } from '@/types/run';\nimport type { SystemCacheBlock, AgentCacheTTL } from '@/types/agent-cache';\n\n/** Interface for bound model with stream and invoke methods */\nexport interface ChatModel {\n  stream?: (\n    messages: BaseMessage[],\n    config?: RunnableConfig\n  ) => Promise<AsyncIterable<AIMessageChunk>>;\n  invoke: (\n    messages: BaseMessage[],\n    config?: RunnableConfig\n  ) => Promise<AIMessageChunk>;\n}\n\n/** Payload for ON_AGENT_TRANSITION events */\nexport type AgentTransitionEvent = {\n  sourceAgentId?: string;\n  sourceAgentName?: string;\n  destinationAgentId: string;\n  destinationAgentName: string;\n  edgeType: string; // 'handoff' | 'transfer' | 'sequence'\n  timestamp: number;\n  /** When true, this event signals handoff completion (child → parent return) */\n  isCompletion?: boolean;\n  /** Duration of child agent execution in milliseconds (only on completion events) */\n  durationMs?: number;\n  /** Length of child agent result text in characters (only on completion events) */\n  resultLength?: number;\n};\n\nexport type GraphNode = GraphNodeKeys | typeof START;\nexport type ClientCallback<T extends unknown[]> = (\n  graph: StandardGraph,\n  ...args: T\n) => void;\n\nexport type ClientCallbacks = {\n  [Callback.TOOL_ERROR]?: ClientCallback<[Error, string]>;\n  [Callback.TOOL_START]?: ClientCallback<unknown[]>;\n  [Callback.TOOL_END]?: ClientCallback<unknown[]>;\n};\n\nexport type SystemCallbacks = {\n  [K in keyof ClientCallbacks]: ClientCallbacks[K] extends ClientCallback<\n    infer Args\n  >\n    ? (...args: Args) => void\n    : never;\n};\n\nexport type BaseGraphState = {\n  messages: BaseMessage[];\n  /**\n   * Structured response when using structured output mode.\n   * Contains the validated JSON response conforming to the configured schema.\n   */\n  structuredResponse?: Record<string, unknown>;\n};\n\nexport type MultiAgentGraphState = BaseGraphState & {\n  agentMessages?: BaseMessage[];\n};\n\nexport type IState = BaseGraphState;\n\nexport interface EventHandler {\n  handle(\n    event: string,\n    data:\n      | StreamEventData\n      | ModelEndData\n      | RunStep\n      | RunStepDeltaEvent\n      | MessageDeltaEvent\n      | ReasoningDeltaEvent\n      | SubagentUpdateEvent\n      | ToolExecuteBatchRequest\n      | { result: ToolEndEvent },\n    metadata?: Record<string, unknown>,\n    graph?: StandardGraph | MultiAgentGraph\n  ): void | Promise<void>;\n}\n\nexport type GraphStateChannels<T extends BaseGraphState> =\n  StateGraphArgs<T>['channels'];\n\nexport type Workflow<\n  T extends BaseGraphState = BaseGraphState,\n  U extends Partial<T> = Partial<T>,\n  N extends string = string,\n> = StateGraph<T, U, N>;\n\nexport type CompiledWorkflow<\n  T extends BaseGraphState = BaseGraphState,\n  U extends Partial<T> = Partial<T>,\n  N extends string = string,\n> = CompiledStateGraph<T, U, N>;\n\nexport type CompiledStateWorkflow = CompiledStateGraph<\n  StateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  UpdateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  string,\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  StateDefinition\n>;\n\nexport type CompiledMultiAgentWorkflow = CompiledStateGraph<\n  StateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  UpdateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  string,\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  StateDefinition\n>;\n\nexport type CompiledAgentWorfklow = CompiledStateGraph<\n  {\n    messages: BaseMessage[];\n  },\n  {\n    messages?: BaseMessage[] | undefined;\n  },\n  '__start__' | `agent=${string}` | `tools=${string}`,\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  StateDefinition,\n  {\n    [x: `agent=${string}`]: Partial<BaseGraphState>;\n    // eslint-disable-next-line @typescript-eslint/no-explicit-any\n    [x: `tools=${string}`]: any;\n  }\n>;\n\nexport type SystemRunnable =\n  | Runnable<\n      BaseMessage[],\n      (BaseMessage | SystemMessage)[],\n      RunnableConfig<Record<string, unknown>>\n    >\n  | undefined;\n\n/**\n * Optional compile options passed to workflow.compile().\n * These are intentionally untyped to avoid coupling to library internals.\n */\nexport type CompileOptions = {\n  // A checkpointer instance (e.g., MemorySaver, SQL saver)\n  // eslint-disable-next-line @typescript-eslint/no-explicit-any\n  checkpointer?: any;\n  interruptBefore?: string[];\n  interruptAfter?: string[];\n  /**\n   * Human-in-the-loop tool approval configuration.\n   * When set, tools matching the policy will trigger an interrupt()\n   * before execution, pausing the graph for human approval.\n   * Requires a checkpointer to be set for interrupt/resume to work.\n   */\n  toolApprovalConfig?: ToolApprovalConfig;\n};\n\nexport type EventStreamCallbackHandlerInput =\n  Parameters<CompiledWorkflow['streamEvents']>[2] extends Omit<\n    infer T,\n    'autoClose'\n  >\n    ? T\n    : never;\n\nexport type StreamChunk =\n  | (ChatGenerationChunk & {\n      message: AIMessageChunk;\n    })\n  | AIMessageChunk;\n\n/**\n * Data associated with a StreamEvent.\n */\nexport type StreamEventData = {\n  /**\n   * The input passed to the runnable that generated the event.\n   * Inputs will sometimes be available at the *START* of the runnable, and\n   * sometimes at the *END* of the runnable.\n   * If a runnable is able to stream its inputs, then its input by definition\n   * won't be known until the *END* of the runnable when it has finished streaming\n   * its inputs.\n   */\n  input?: unknown;\n  /**\n   * The output of the runnable that generated the event.\n   * Outputs will only be available at the *END* of the runnable.\n   * For most runnables, this field can be inferred from the `chunk` field,\n   * though there might be some exceptions for special cased runnables (e.g., like\n   * chat models), which may return more information.\n   */\n  output?: unknown;\n  /**\n   * A streaming chunk from the output that generated the event.\n   * chunks support addition in general, and adding them up should result\n   * in the output of the runnable that generated the event.\n   */\n  chunk?: StreamChunk;\n  /**\n   * Runnable config for invoking other runnables within handlers.\n   */\n  config?: RunnableConfig;\n  /**\n   * Custom result from the runnable that generated the event.\n   */\n  result?: unknown;\n  /**\n   * Custom field to indicate the event was manually emitted, and may have been handled already\n   */\n  emitted?: boolean;\n};\n\n/**\n * A streaming event.\n *\n * Schema of a streaming event which is produced from the streamEvents method.\n */\nexport type StreamEvent = {\n  /**\n   * Event names are of the format: on_[runnable_type]_(start|stream|end).\n   *\n   * Runnable types are one of:\n   * - llm - used by non chat models\n   * - chat_model - used by chat models\n   * - prompt --  e.g., ChatPromptTemplate\n   * - tool -- LangChain tools\n   * - chain - most Runnables are of this type\n   *\n   * Further, the events are categorized as one of:\n   * - start - when the runnable starts\n   * - stream - when the runnable is streaming\n   * - end - when the runnable ends\n   *\n   * start, stream and end are associated with slightly different `data` payload.\n   *\n   * Please see the documentation for `EventData` for more details.\n   */\n  event: string;\n  /** The name of the runnable that generated the event. */\n  name: string;\n  /**\n   * An randomly generated ID to keep track of the execution of the given runnable.\n   *\n   * Each child runnable that gets invoked as part of the execution of a parent runnable\n   * is assigned its own unique ID.\n   */\n  run_id: string;\n  /**\n   * Tags associated with the runnable that generated this event.\n   * Tags are always inherited from parent runnables.\n   */\n  tags?: string[];\n  /** Metadata associated with the runnable that generated this event. */\n  metadata: Record<string, unknown>;\n  /**\n   * Event data.\n   *\n   * The contents of the event data depend on the event type.\n   */\n  data: StreamEventData;\n};\n\nexport type GraphConfig = {\n  provider: string;\n  thread_id?: string;\n  run_id?: string;\n};\n\nexport type PartMetadata = {\n  progress?: number;\n  asset_pointer?: string;\n  status?: string;\n  action?: boolean;\n  output?: string;\n};\n\nexport type ModelEndData =\n  | (StreamEventData & { output: AIMessageChunk | undefined })\n  | undefined;\nexport type GraphTools = GenericTool[] | BindToolsInput[] | GoogleAIToolType[];\nexport type StandardGraphInput = {\n  runId?: string;\n  signal?: AbortSignal;\n  agents: AgentInputs[];\n  tokenCounter?: TokenCounter;\n  indexTokenCountMap?: Record<string, number>;\n};\n\n/**\n * Configuration for an approval gate placed on a sequence edge.\n * When present, the graph inserts an approval gate node between the source\n * and destination agents. The gate ALWAYS fires (regardless of ExecutionContext)\n * and calls interrupt() to pause the graph for human approval.\n */\nexport type ApprovalGateConfig = {\n  /** Unique identifier for this gate (used as node ID suffix) */\n  gateId: string;\n  /**\n   * Approval channel — where the approval UI is rendered.\n   * - 'chat': SSE-based chat UI (default)\n   * - 'outlook': MS Graph Actionable Messages\n   * - 'telegram': Telegram Bot inline keyboard\n   */\n  channel?: 'chat' | 'outlook' | 'telegram';\n  /** Optional human-readable prompt shown to the approver */\n  prompt?: string;\n  /** Optional approver identifier (e.g., email, user ID) */\n  approver?: string;\n  /** Timeout in ms before the gate auto-expires (default: 5 minutes) */\n  timeoutMs?: number;\n  /** What to do on denial: 'stop' ends the graph, 'skip' skips the destination agent */\n  onDeny?: 'stop' | 'skip';\n};\n\nexport type GraphEdge = {\n  /** Agent ID, use a list for multiple sources */\n  from: string | string[];\n  /** Agent ID, use a list for multiple destinations */\n  to: string | string[];\n  description?: string;\n  /** Can return boolean or specific destination(s) */\n  condition?: (state: BaseGraphState) => boolean | string | string[];\n  /**\n   * EdgeType.HANDOFF — one-way routing: parent emits an `lc_transfer_to_*`\n   *   tool call and exits, child takes over and responds directly to the\n   *   user. Aligns with upstream's handoff semantics.\n   * EdgeType.DIRECT — fixed graph edges for automatic sequential / parallel\n   *   transitions. Ranger preserves its enriched wiring (fan-in with prompt,\n   *   parallel groups, ApprovalGateNode, excludeResults / agentMessages).\n   */\n  edgeType?: import('@/common').EdgeType;\n  /**\n   * For sequence edges: Optional prompt to add when transitioning through this edge.\n   * String prompts can include variables like {results} which will be replaced with\n   * messages from startIndex onwards. When {results} is used, excludeResults defaults to true.\n   *\n   * For transfer edges: Description for the input parameter that the transfer tool accepts,\n   * allowing the supervisor to pass specific instructions/context to the transferred agent.\n   */\n  prompt?:\n    | string\n    | ((\n        messages: BaseMessage[],\n        runStartIndex: number\n      ) => string | Promise<string> | undefined);\n  /**\n   * When true, excludes messages from startIndex when adding prompt.\n   * Automatically set to true when {results} variable is used in prompt.\n   */\n  excludeResults?: boolean;\n  /**\n   * For transfer edges: Customizes the parameter name for the transfer input.\n   * Defaults to \"instructions\" if not specified.\n   * Only applies when prompt is provided for transfer edges.\n   *\n   * For handoff edges: Customizes the parameter name for the handoff instruction input.\n   */\n  promptKey?: string;\n  /**\n   * Approval gate configuration for sequence edges.\n   * When set, inserts an approval gate node between source and destination.\n   * The gate ALWAYS fires regardless of ExecutionContext (unlike tool approval).\n   */\n  approvalGate?: ApprovalGateConfig;\n};\n\nexport type MultiAgentGraphInput = StandardGraphInput & {\n  edges: GraphEdge[];\n  /**\n   * When set, the graph routes START to this agent instead of the default\n   * starting nodes. Used for multi-turn resumption: the caller reads the\n   * last active agent id from the previous turn (e.g. via\n   * `Run.getLastActiveAgentId()`, which derives it from the run's content\n   * data trail) and passes it here so follow-up messages route to the\n   * agent that last handled the conversation.\n   *\n   * If the agent ID is invalid (not in the graph), falls back to default\n   * starting nodes.\n   */\n  resumeFromAgentId?: string;\n};\n\n/**\n * Structured output mode determines how the agent returns structured data.\n * - 'tool': Uses tool calling to return structured output (works with all tool-calling models)\n * - 'provider': Uses provider-native structured output via LangChain's jsonMode (OpenAI, Anthropic, etc.)\n * - 'native': Uses provider's constrained decoding API directly for guaranteed schema compliance\n *   (Anthropic output_config.format, OpenAI response_format.json_schema). Falls back to 'tool' for unsupported providers.\n * - 'auto': Automatically selects the best strategy — 'native' for supported providers, 'tool' for others\n */\nexport type StructuredOutputMode = 'tool' | 'provider' | 'native' | 'auto';\n\n/**\n * Resolved method used internally after mode resolution.\n * Maps to LangChain's withStructuredOutput method parameter plus our native path.\n */\nexport type ResolvedStructuredOutputMethod =\n  | 'functionCalling'\n  | 'jsonMode'\n  | 'jsonSchema'\n  | 'native'\n  | undefined;\n\n/**\n * Error thrown when the model refuses to produce structured output due to safety policies.\n */\nexport class StructuredOutputRefusalError extends Error {\n  constructor(public refusalText: string) {\n    super(`Model refused to produce structured output: ${refusalText}`);\n    this.name = 'StructuredOutputRefusalError';\n  }\n}\n\n/**\n * Error thrown when the structured output response was truncated due to max_tokens.\n */\nexport class StructuredOutputTruncatedError extends Error {\n  constructor(public stopReason: string) {\n    super(\n      `Structured output was truncated (stop_reason: ${stopReason}). ` +\n        'Increase max_tokens to allow the full JSON response to be generated.'\n    );\n    this.name = 'StructuredOutputTruncatedError';\n  }\n}\n\n/**\n * Configuration for structured JSON output from agents.\n * When configured, the agent will return a validated JSON response\n * instead of streaming text.\n */\nexport interface StructuredOutputConfig {\n  /**\n   * JSON Schema defining the output structure.\n   * The model will be forced to return data conforming to this schema.\n   */\n  schema: Record<string, unknown>;\n  /**\n   * Name for the structured output format (used in tool mode).\n   * @default 'StructuredResponse'\n   */\n  name?: string;\n  /**\n   * Description of what the structured output represents.\n   * Helps the model understand the expected format.\n   */\n  description?: string;\n  /**\n   * Output mode strategy.\n   * @default 'auto'\n   */\n  mode?: StructuredOutputMode;\n  /**\n   * Enable strict schema validation.\n   * When true, the response must exactly match the schema.\n   * @default true\n   */\n  strict?: boolean;\n  /**\n   * Error handling configuration.\n   * - true: Auto-retry on validation errors (default)\n   * - false: Throw error on validation failure\n   * - string: Custom error message for retry\n   */\n  handleErrors?: boolean | string;\n  /**\n   * Maximum number of retry attempts on validation failure.\n   * @default 2\n   */\n  maxRetries?: number;\n  /**\n   * Include the raw AI message along with structured response.\n   * Useful for debugging.\n   * @default false\n   */\n  includeRaw?: boolean;\n}\n\n/**\n * Database/API structured output format (snake_case with enabled flag).\n * This matches the format stored in MongoDB and sent from frontends.\n */\nexport interface StructuredOutputInput {\n  /** Whether structured output is enabled */\n  enabled?: boolean;\n  /** JSON Schema defining the expected response structure */\n  schema?: Record<string, unknown>;\n  /** Name identifier for the structured output */\n  name?: string;\n  /** Description of what the structured output represents */\n  description?: string;\n  /** Mode for structured output: 'tool' | 'provider' | 'native' | 'auto' */\n  mode?: StructuredOutputMode;\n  /** Whether to enforce strict schema validation */\n  strict?: boolean;\n}\n\n/**\n * Trigger strategy for when summarization should activate.\n * - 'contextPercentage': Trigger when context utilization exceeds a threshold percentage\n * - 'messageCount': Trigger when pruned message count exceeds a threshold\n * - 'tokenThreshold': Trigger when total token count exceeds a raw threshold\n */\nexport type SummarizationTriggerType =\n  | 'contextPercentage'\n  | 'messageCount'\n  | 'tokenThreshold';\n\n/**\n * Configuration for summarization behavior within the agent pipeline.\n * All fields are optional — sensible defaults are provided via constants.\n *\n * @see SUMMARIZATION_CONTEXT_THRESHOLD, SUMMARIZATION_RESERVE_RATIO, PRUNING_EMA_ALPHA\n */\nexport interface SummarizationConfig {\n  /**\n   * Strategy for when summarization triggers.\n   * @default 'contextPercentage'\n   */\n  triggerType?: SummarizationTriggerType;\n\n  /**\n   * Threshold value interpreted based on triggerType:\n   * - contextPercentage: 0-100 (percentage of context window)\n   * - messageCount: absolute count of messages pruned\n   * - tokenThreshold: absolute token count\n   * @default 80 (for contextPercentage)\n   */\n  triggerThreshold?: number;\n\n  /**\n   * Fraction of context window (0-1) reserved for recent messages.\n   * Prevents over-pruning by ensuring at least this fraction of the\n   * context budget is preserved as recent conversation history.\n   * @default 0.3\n   */\n  reserveRatio?: number;\n\n  /**\n   * Whether context pruning is enabled (can be disabled for debugging).\n   * @default true\n   */\n  contextPruning?: boolean;\n\n  /**\n   * Initial summary text to seed across runs.\n   * Different from persistedSummary: this is provided by the caller as a\n   * cross-conversation seed (e.g., agent personality or recurring context),\n   * while persistedSummary is loaded from the conversation's own history.\n   */\n  initialSummary?: string;\n\n  /**\n   * Upstream-aligned optional fields. When the host wants the summarization\n   * pass to run on a different LLM than the agent's own (e.g. a cheaper\n   * model for compaction), these provider/model overrides flow through to\n   * the summarize callback.\n   */\n  provider?: Providers;\n  model?: string;\n  parameters?: Record<string, unknown>;\n  prompt?: string;\n  updatePrompt?: string;\n  trigger?: { type: string; value: number };\n  maxSummaryTokens?: number;\n}\n\n/**\n * Runtime state for EMA-based pruning calibration.\n * Maintained across iterations within a single run to smooth pruning decisions.\n */\nexport interface PruneCalibrationState {\n  /** Current EMA calibration ratio */\n  ratio: number;\n  /** Number of calibration updates applied */\n  iterations: number;\n}\n\n/**\n * Lightweight file metadata entry for conversation-level file awareness.\n * Contains only IDs and names — NOT full content — so the agent always knows\n * what files exist in the conversation even after compaction pushes old messages\n * behind the summary window. The agent can retrieve full content on-demand\n * via file_search (RAG) or content_tool read (by contentId).\n */\nexport interface FileManifestEntry {\n  /** Unique file identifier (e.g., MongoDB ObjectId or UUID) */\n  fileId: string;\n  /** Original filename (e.g., \"quarterly-report.pdf\") */\n  filename: string;\n  /** Content identifier for on-demand retrieval via content_tool read */\n  contentId?: string;\n  /** File source (e.g., \"local\", \"sharepoint\", \"onedrive\") */\n  source?: string;\n  /** Index of the message that introduced this file (0-based in the original message array) */\n  messageIndex?: number;\n}\n\n/** Configuration for a subagent type that can be spawned by a parent agent. */\nexport type SubagentConfig = {\n  /** Identifier used in the tool's `subagent_type` enum (e.g. 'researcher', 'coder'). */\n  type: string;\n  /** Human-readable display name. */\n  name: string;\n  /** What this subagent specializes in — shown to the LLM. */\n  description: string;\n  /** Full agent config for the child graph. Omit when `self` is true. */\n  agentInputs?: AgentInputs;\n  /** When true, reuse the parent's AgentInputs (context isolation without separate config). */\n  self?: boolean;\n  /** Max AGENT→TOOLS cycles before forced stop (default: 25). */\n  maxTurns?: number;\n  /** Allow this subagent to spawn its own subagents (default: false). */\n  allowNested?: boolean;\n};\n\n/** SubagentConfig with agentInputs guaranteed present (self-spawn resolved). */\nexport type ResolvedSubagentConfig = SubagentConfig & {\n  agentInputs: AgentInputs;\n};\n\n/** Lifecycle phase carried on {@link SubagentUpdateEvent}. */\nexport type SubagentUpdatePhase =\n  | 'start'\n  | 'run_step'\n  | 'run_step_delta'\n  | 'run_step_completed'\n  | 'message_delta'\n  | 'reasoning_delta'\n  | 'stop'\n  | 'error';\n\n/**\n * Wrapper event emitted when a subagent's child graph dispatches activity.\n * Lets hosts show subagent progress in a UI surface separate from the parent\n * conversation without having to untangle events by agent ID.\n */\nexport interface SubagentUpdateEvent {\n  /** Parent run ID. */\n  runId: string;\n  /** Child run ID (unique per subagent execution). */\n  subagentRunId: string;\n  /**\n   * Parent-side `tool_call_id` for the `subagent` tool invocation that\n   * triggered this run. Stable for the duration of the child; lets hosts\n   * correlate updates deterministically instead of inferring by ordering.\n   * Omitted when the executor was invoked outside of a tool-call context.\n   */\n  parentToolCallId?: string;\n  /** Subagent `type` identifier from the SubagentConfig. */\n  subagentType: string;\n  /** Child agent ID assigned to this subagent execution. */\n  subagentAgentId: string;\n  /** Parent agent ID that spawned this subagent. */\n  parentAgentId?: string;\n  /** Lifecycle phase carried by this update. */\n  phase: SubagentUpdatePhase;\n  /** Underlying event payload (shape depends on phase). */\n  data?: unknown;\n  /** Short human-readable description. Hosts can render this directly. */\n  label?: string;\n  /** ISO timestamp for ordering / display. */\n  timestamp: string;\n}\n\nexport interface AgentInputs {\n  agentId: string;\n  /** Human-readable name for the agent (used in handoff context). Defaults to agentId if not provided. */\n  name?: string;\n  /** Description of what this agent does (used to enrich handoff tool descriptions). */\n  description?: string;\n  toolEnd?: boolean;\n  toolMap?: ToolMap;\n  tools?: GraphTools;\n  provider: Providers;\n  /**\n   * Ordered list of cacheable system content blocks emitted BEFORE\n   * `instructions` in the system message. Each block gets its own cache\n   * marker (cachePoint on Bedrock, cache_control on Anthropic) so the\n   * cache key for each block includes only the bytes up to and including\n   * its own marker. Entries are emitted in array order — earlier =\n   * stabler = wider cache key (best for cross-tenant sharing).\n   *\n   * See `src/types/agent-cache.ts` for full semantics. Capped at\n   * `MAX_SYSTEM_CACHE_BLOCKS` (2) entries to stay within Bedrock's\n   * 4-cachePoint budget after the tools array consumes 2.\n   */\n  system_cache_blocks?: SystemCacheBlock[];\n  /**\n   * Stable/cacheable system instructions for this agent. Always emitted\n   * with its own trailing cache marker (when caching is supported by\n   * the provider/model). Use `instructions_cache_ttl` to override the\n   * default '5m' TTL.\n   */\n  instructions?: string;\n  /** TTL for the trailing `instructions` cache marker. Defaults to '5m'. */\n  instructions_cache_ttl?: AgentCacheTTL;\n  streamBuffer?: number;\n  maxContextTokens?: number;\n  clientOptions?: ClientOptions;\n  /**\n   * Dynamic system tail appended after the cacheable instructions\n   * without any cache marker. Per-user / per-message context belongs\n   * here so it does not invalidate the cacheable prefix.\n   */\n  additional_instructions?: string;\n  reasoningKey?: 'reasoning_content' | 'reasoning';\n  /**\n   * Subagent types this agent may spawn. When non-empty, Graph injects the\n   * `subagent` tool into the agent's toolset with the per-config enum and\n   * description populated.\n   */\n  subagentConfigs?: SubagentConfig[];\n  /**\n   * Maximum subagent depth allowed below this agent. Used by SubagentExecutor\n   * to abort runaway nesting. Defaults to a small constant.\n   */\n  maxSubagentDepth?: number;\n  /**\n   * Pre-existing summary injected into the system message via\n   * AgentContext.buildInstructionsString. Populated by SubagentExecutor when\n   * spawning a self-config subagent that should inherit the parent's prior\n   * compaction summary; otherwise host-supplied at run time.\n   */\n  initialSummary?: { text: string; tokenCount: number };\n  /**\n   * Upstream-aligned opt-in for summarization. When false, the agent never\n   * invokes the summarize callback regardless of context utilization.\n   * Defaults to undefined (treated as enabled in Ranger to preserve prior\n   * behaviour); hosts that want strict opt-in semantics should set it\n   * explicitly.\n   */\n  summarizationEnabled?: boolean;\n  /** Format content blocks as strings (for legacy compatibility i.e. Ollama/Azure Serverless) */\n  useLegacyContent?: boolean;\n  /**\n   * Tool definitions for all tools, including deferred and programmatic.\n   * Used for tool search and programmatic tool calling.\n   * Maps tool name to LCTool definition.\n   */\n  toolRegistry?: Map<string, LCTool>;\n  /**\n   * Dynamic context that changes per-request (e.g., current time, user info).\n   * This is injected as a user message rather than system prompt to preserve cache.\n   * Keeping this separate from instructions ensures the system message stays static\n   * and can be cached by Bedrock/Anthropic prompt caching.\n   */\n  dynamicContext?: string;\n  /**\n   * Structured output configuration (camelCase).\n   * When set, disables streaming and returns a validated JSON response\n   * conforming to the specified schema.\n   */\n  structuredOutput?: StructuredOutputConfig;\n  /**\n   * Structured output configuration (snake_case - database/API format).\n   * Alternative to structuredOutput for compatibility with MongoDB/frontend.\n   * Uses an `enabled` flag to control activation.\n   * @deprecated Use structuredOutput instead when possible\n   */\n  structured_output?: StructuredOutputInput;\n  /**\n   * Serializable tool definitions for event-driven execution.\n   * When provided, ToolNode operates in event-driven mode, dispatching\n   * ON_TOOL_EXECUTE events instead of invoking tools directly.\n   */\n  toolDefinitions?: LCTool[];\n  /**\n   * Tool names discovered from previous conversation history.\n   * These tools will be pre-marked as discovered so they're included\n   * in tool binding without requiring tool_search.\n   */\n  discoveredTools?: string[];\n  /**\n   * Optional callback for summarizing messages that were pruned from context.\n   * When provided, discarded messages are summarized by the host caller\n   * using a cheap LLM call, and the summary is prepended to the context.\n   */\n  summarizeCallback?: (\n    messagesToRefine: import('@langchain/core/messages').BaseMessage[]\n  ) => Promise<string | undefined>;\n  /**\n   * Pre-existing summary text loaded from persistent storage (MongoDB/Redis).\n   * When provided, this summary is injected into the initial message context\n   * so the agent has prior conversation history even on new turns.\n   * Set by the host's summary store when resuming a conversation.\n   */\n  persistedSummary?: string;\n  /**\n   * Summarization configuration controlling trigger strategy, reserve ratio,\n   * and EMA calibration for pruning. When omitted, sensible defaults apply.\n   * @see SummarizationConfig\n   */\n  summarizationConfig?: SummarizationConfig;\n  /**\n   * Lightweight file manifest for the conversation.\n   * Contains file IDs, names, and metadata — NOT full content.\n   *\n   * Used by the compaction engine to inject a [Conversation Files] block\n   * into the windowed view, ensuring the LLM always knows what files exist\n   * even when old messages (with full file content) are behind the summary.\n   *\n   * The agent can retrieve full content on-demand via:\n   *   - file_search (RAG semantic search over embedded files)\n   *   - content_tool read (by contentId for exact file retrieval)\n   *\n   * Built by the host orchestrator from message_file_map\n   * and metadata.context_files across all conversation messages.\n   */\n  fileManifest?: FileManifestEntry[];\n}\n\n/**\n * Tunable knobs for position-based content degradation.\n * See `messages/contextPruning.ts` and `messages/contextPruningSettings.ts`.\n */\nexport interface ContextPruningConfig {\n  enabled?: boolean;\n  keepLastAssistants?: number;\n  softTrimRatio?: number;\n  hardClearRatio?: number;\n  minPrunableToolChars?: number;\n  softTrim?: {\n    maxChars?: number;\n    headChars?: number;\n    tailChars?: number;\n  };\n  hardClear?: {\n    enabled?: boolean;\n    placeholder?: string;\n  };\n}\n"],"names":[],"mappings":";;AAqdA;;AAEG;AACG,MAAO,4BAA6B,SAAQ,KAAK,CAAA;AAClC,IAAA,WAAA;AAAnB,IAAA,WAAA,CAAmB,WAAmB,EAAA;AACpC,QAAA,KAAK,CAAC,CAAA,4CAAA,EAA+C,WAAW,CAAA,CAAE,CAAC;QADlD,IAAA,CAAA,WAAW,GAAX,WAAW;AAE5B,QAAA,IAAI,CAAC,IAAI,GAAG,8BAA8B;IAC5C;AACD;AAED;;AAEG;AACG,MAAO,8BAA+B,SAAQ,KAAK,CAAA;AACpC,IAAA,UAAA;AAAnB,IAAA,WAAA,CAAmB,UAAkB,EAAA;QACnC,KAAK,CACH,CAAA,8CAAA,EAAiD,UAAU,CAAA,GAAA,CAAK;AAC9D,YAAA,sEAAsE,CACzE;QAJgB,IAAA,CAAA,UAAU,GAAV,UAAU;AAK3B,QAAA,IAAI,CAAC,IAAI,GAAG,gCAAgC;IAC9C;AACD;;;;;"}
+{"version":3,"file":"graph.cjs","sources":["../../../src/types/graph.ts"],"sourcesContent":["// src/types/graph.ts\nimport type {\n  START,\n  StateType,\n  UpdateType,\n  StateGraph,\n  StateGraphArgs,\n  StateDefinition,\n  CompiledStateGraph,\n  BinaryOperatorAggregate,\n} from '@langchain/langgraph';\nimport type { BindToolsInput } from '@langchain/core/language_models/chat_models';\nimport type {\n  BaseMessage,\n  AIMessageChunk,\n  SystemMessage,\n} from '@langchain/core/messages';\nimport type { RunnableConfig, Runnable } from '@langchain/core/runnables';\nimport type { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { GoogleAIToolType } from '@langchain/google-common';\nimport type {\n  ToolMap,\n  ToolEndEvent,\n  GenericTool,\n  LCTool,\n  ToolApprovalConfig,\n  ToolExecuteBatchRequest,\n} from '@/types/tools';\nimport type { Providers, Callback, GraphNodeKeys } from '@/common';\nimport type { StandardGraph, MultiAgentGraph } from '@/graphs';\nimport type { ClientOptions } from '@/types/llm';\nimport type {\n  RunStep,\n  RunStepDeltaEvent,\n  MessageDeltaEvent,\n  ReasoningDeltaEvent,\n} from '@/types/stream';\nimport type { TokenCounter } from '@/types/run';\nimport type { SystemCacheBlock, AgentCacheTTL } from '@/types/agent-cache';\n\n/** Interface for bound model with stream and invoke methods */\nexport interface ChatModel {\n  stream?: (\n    messages: BaseMessage[],\n    config?: RunnableConfig\n  ) => Promise<AsyncIterable<AIMessageChunk>>;\n  invoke: (\n    messages: BaseMessage[],\n    config?: RunnableConfig\n  ) => Promise<AIMessageChunk>;\n}\n\n/** Payload for ON_AGENT_TRANSITION events */\nexport type AgentTransitionEvent = {\n  sourceAgentId?: string;\n  sourceAgentName?: string;\n  destinationAgentId: string;\n  destinationAgentName: string;\n  edgeType: string; // 'handoff' | 'transfer' | 'sequence'\n  timestamp: number;\n  /** When true, this event signals handoff completion (child → parent return) */\n  isCompletion?: boolean;\n  /** Duration of child agent execution in milliseconds (only on completion events) */\n  durationMs?: number;\n  /** Length of child agent result text in characters (only on completion events) */\n  resultLength?: number;\n};\n\nexport type GraphNode = GraphNodeKeys | typeof START;\nexport type ClientCallback<T extends unknown[]> = (\n  graph: StandardGraph,\n  ...args: T\n) => void;\n\nexport type ClientCallbacks = {\n  [Callback.TOOL_ERROR]?: ClientCallback<[Error, string]>;\n  [Callback.TOOL_START]?: ClientCallback<unknown[]>;\n  [Callback.TOOL_END]?: ClientCallback<unknown[]>;\n};\n\nexport type SystemCallbacks = {\n  [K in keyof ClientCallbacks]: ClientCallbacks[K] extends ClientCallback<\n    infer Args\n  >\n    ? (...args: Args) => void\n    : never;\n};\n\nexport type BaseGraphState = {\n  messages: BaseMessage[];\n  /**\n   * Structured response when using structured output mode.\n   * Contains the validated JSON response conforming to the configured schema.\n   */\n  structuredResponse?: Record<string, unknown>;\n};\n\nexport type MultiAgentGraphState = BaseGraphState & {\n  agentMessages?: BaseMessage[];\n};\n\nexport type IState = BaseGraphState;\n\nexport interface EventHandler {\n  handle(\n    event: string,\n    data:\n      | StreamEventData\n      | ModelEndData\n      | RunStep\n      | RunStepDeltaEvent\n      | MessageDeltaEvent\n      | ReasoningDeltaEvent\n      | SubagentUpdateEvent\n      | ToolExecuteBatchRequest\n      | { result: ToolEndEvent },\n    metadata?: Record<string, unknown>,\n    graph?: StandardGraph | MultiAgentGraph\n  ): void | Promise<void>;\n}\n\nexport type GraphStateChannels<T extends BaseGraphState> =\n  StateGraphArgs<T>['channels'];\n\nexport type Workflow<\n  T extends BaseGraphState = BaseGraphState,\n  U extends Partial<T> = Partial<T>,\n  N extends string = string,\n> = StateGraph<T, U, N>;\n\nexport type CompiledWorkflow<\n  T extends BaseGraphState = BaseGraphState,\n  U extends Partial<T> = Partial<T>,\n  N extends string = string,\n> = CompiledStateGraph<T, U, N>;\n\nexport type CompiledStateWorkflow = CompiledStateGraph<\n  StateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  UpdateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  string,\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  StateDefinition\n>;\n\nexport type CompiledMultiAgentWorkflow = CompiledStateGraph<\n  StateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  UpdateType<{\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  }>,\n  string,\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n    agentMessages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  StateDefinition\n>;\n\nexport type CompiledAgentWorfklow = CompiledStateGraph<\n  {\n    messages: BaseMessage[];\n  },\n  {\n    messages?: BaseMessage[] | undefined;\n  },\n  '__start__' | `agent=${string}` | `tools=${string}`,\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  {\n    messages: BinaryOperatorAggregate<BaseMessage[], BaseMessage[]>;\n  },\n  StateDefinition,\n  {\n    [x: `agent=${string}`]: Partial<BaseGraphState>;\n    // eslint-disable-next-line @typescript-eslint/no-explicit-any\n    [x: `tools=${string}`]: any;\n  }\n>;\n\nexport type SystemRunnable =\n  | Runnable<\n      BaseMessage[],\n      (BaseMessage | SystemMessage)[],\n      RunnableConfig<Record<string, unknown>>\n    >\n  | undefined;\n\n/**\n * Optional compile options passed to workflow.compile().\n * These are intentionally untyped to avoid coupling to library internals.\n */\nexport type CompileOptions = {\n  // A checkpointer instance (e.g., MemorySaver, SQL saver)\n  // eslint-disable-next-line @typescript-eslint/no-explicit-any\n  checkpointer?: any;\n  interruptBefore?: string[];\n  interruptAfter?: string[];\n  /**\n   * Human-in-the-loop tool approval configuration.\n   * When set, tools matching the policy will trigger an interrupt()\n   * before execution, pausing the graph for human approval.\n   * Requires a checkpointer to be set for interrupt/resume to work.\n   */\n  toolApprovalConfig?: ToolApprovalConfig;\n};\n\nexport type EventStreamCallbackHandlerInput =\n  Parameters<CompiledWorkflow['streamEvents']>[2] extends Omit<\n    infer T,\n    'autoClose'\n  >\n    ? T\n    : never;\n\nexport type StreamChunk =\n  | (ChatGenerationChunk & {\n      message: AIMessageChunk;\n    })\n  | AIMessageChunk;\n\n/**\n * Data associated with a StreamEvent.\n */\nexport type StreamEventData = {\n  /**\n   * The input passed to the runnable that generated the event.\n   * Inputs will sometimes be available at the *START* of the runnable, and\n   * sometimes at the *END* of the runnable.\n   * If a runnable is able to stream its inputs, then its input by definition\n   * won't be known until the *END* of the runnable when it has finished streaming\n   * its inputs.\n   */\n  input?: unknown;\n  /**\n   * The output of the runnable that generated the event.\n   * Outputs will only be available at the *END* of the runnable.\n   * For most runnables, this field can be inferred from the `chunk` field,\n   * though there might be some exceptions for special cased runnables (e.g., like\n   * chat models), which may return more information.\n   */\n  output?: unknown;\n  /**\n   * A streaming chunk from the output that generated the event.\n   * chunks support addition in general, and adding them up should result\n   * in the output of the runnable that generated the event.\n   */\n  chunk?: StreamChunk;\n  /**\n   * Runnable config for invoking other runnables within handlers.\n   */\n  config?: RunnableConfig;\n  /**\n   * Custom result from the runnable that generated the event.\n   */\n  result?: unknown;\n  /**\n   * Custom field to indicate the event was manually emitted, and may have been handled already\n   */\n  emitted?: boolean;\n};\n\n/**\n * A streaming event.\n *\n * Schema of a streaming event which is produced from the streamEvents method.\n */\nexport type StreamEvent = {\n  /**\n   * Event names are of the format: on_[runnable_type]_(start|stream|end).\n   *\n   * Runnable types are one of:\n   * - llm - used by non chat models\n   * - chat_model - used by chat models\n   * - prompt --  e.g., ChatPromptTemplate\n   * - tool -- LangChain tools\n   * - chain - most Runnables are of this type\n   *\n   * Further, the events are categorized as one of:\n   * - start - when the runnable starts\n   * - stream - when the runnable is streaming\n   * - end - when the runnable ends\n   *\n   * start, stream and end are associated with slightly different `data` payload.\n   *\n   * Please see the documentation for `EventData` for more details.\n   */\n  event: string;\n  /** The name of the runnable that generated the event. */\n  name: string;\n  /**\n   * An randomly generated ID to keep track of the execution of the given runnable.\n   *\n   * Each child runnable that gets invoked as part of the execution of a parent runnable\n   * is assigned its own unique ID.\n   */\n  run_id: string;\n  /**\n   * Tags associated with the runnable that generated this event.\n   * Tags are always inherited from parent runnables.\n   */\n  tags?: string[];\n  /** Metadata associated with the runnable that generated this event. */\n  metadata: Record<string, unknown>;\n  /**\n   * Event data.\n   *\n   * The contents of the event data depend on the event type.\n   */\n  data: StreamEventData;\n};\n\nexport type GraphConfig = {\n  provider: string;\n  thread_id?: string;\n  run_id?: string;\n};\n\nexport type PartMetadata = {\n  progress?: number;\n  asset_pointer?: string;\n  status?: string;\n  action?: boolean;\n  output?: string;\n};\n\nexport type ModelEndData =\n  | (StreamEventData & { output: AIMessageChunk | undefined })\n  | undefined;\nexport type GraphTools = GenericTool[] | BindToolsInput[] | GoogleAIToolType[];\nexport type StandardGraphInput = {\n  runId?: string;\n  signal?: AbortSignal;\n  agents: AgentInputs[];\n  tokenCounter?: TokenCounter;\n  indexTokenCountMap?: Record<string, number>;\n};\n\n/**\n * Configuration for an approval gate placed on a sequence edge.\n * When present, the graph inserts an approval gate node between the source\n * and destination agents. The gate ALWAYS fires (regardless of ExecutionContext)\n * and calls interrupt() to pause the graph for human approval.\n */\nexport type ApprovalGateConfig = {\n  /** Unique identifier for this gate (used as node ID suffix) */\n  gateId: string;\n  /**\n   * Approval channel — where the approval UI is rendered.\n   * - 'chat': SSE-based chat UI (default)\n   * - 'outlook': MS Graph Actionable Messages\n   * - 'telegram': Telegram Bot inline keyboard\n   */\n  channel?: 'chat' | 'outlook' | 'telegram';\n  /** Optional human-readable prompt shown to the approver */\n  prompt?: string;\n  /** Optional approver identifier (e.g., email, user ID) */\n  approver?: string;\n  /** Timeout in ms before the gate auto-expires (default: 5 minutes) */\n  timeoutMs?: number;\n  /** What to do on denial: 'stop' ends the graph, 'skip' skips the destination agent */\n  onDeny?: 'stop' | 'skip';\n};\n\nexport type GraphEdge = {\n  /** Agent ID, use a list for multiple sources */\n  from: string | string[];\n  /** Agent ID, use a list for multiple destinations */\n  to: string | string[];\n  description?: string;\n  /** Can return boolean or specific destination(s) */\n  condition?: (state: BaseGraphState) => boolean | string | string[];\n  /**\n   * EdgeType.HANDOFF — one-way routing: parent emits an `lc_transfer_to_*`\n   *   tool call and exits, child takes over and responds directly to the\n   *   user. Aligns with the standard handoff semantics.\n   * EdgeType.DIRECT — fixed graph edges for automatic sequential / parallel\n   *   transitions. Downstream consumers may layer additional wiring on top\n   *   (fan-in with prompt, parallel groups, approval-gate nodes,\n   *   excludeResults / agentMessages).\n   */\n  edgeType?: import('@/common').EdgeType;\n  /**\n   * For sequence edges: Optional prompt to add when transitioning through this edge.\n   * String prompts can include variables like {results} which will be replaced with\n   * messages from startIndex onwards. When {results} is used, excludeResults defaults to true.\n   *\n   * For transfer edges: Description for the input parameter that the transfer tool accepts,\n   * allowing the supervisor to pass specific instructions/context to the transferred agent.\n   */\n  prompt?:\n    | string\n    | ((\n        messages: BaseMessage[],\n        runStartIndex: number\n      ) => string | Promise<string> | undefined);\n  /**\n   * When true, excludes messages from startIndex when adding prompt.\n   * Automatically set to true when {results} variable is used in prompt.\n   */\n  excludeResults?: boolean;\n  /**\n   * For transfer edges: Customizes the parameter name for the transfer input.\n   * Defaults to \"instructions\" if not specified.\n   * Only applies when prompt is provided for transfer edges.\n   *\n   * For handoff edges: Customizes the parameter name for the handoff instruction input.\n   */\n  promptKey?: string;\n  /**\n   * Approval gate configuration for sequence edges.\n   * When set, inserts an approval gate node between source and destination.\n   * The gate ALWAYS fires regardless of ExecutionContext (unlike tool approval).\n   */\n  approvalGate?: ApprovalGateConfig;\n};\n\nexport type MultiAgentGraphInput = StandardGraphInput & {\n  edges: GraphEdge[];\n  /**\n   * When set, the graph routes START to this agent instead of the default\n   * starting nodes. Used for multi-turn resumption: the caller reads the\n   * last active agent id from the previous turn (e.g. via\n   * `Run.getLastActiveAgentId()`, which derives it from the run's content\n   * data trail) and passes it here so follow-up messages route to the\n   * agent that last handled the conversation.\n   *\n   * If the agent ID is invalid (not in the graph), falls back to default\n   * starting nodes.\n   */\n  resumeFromAgentId?: string;\n};\n\n/**\n * Structured output mode determines how the agent returns structured data.\n * - 'tool': Uses tool calling to return structured output (works with all tool-calling models)\n * - 'provider': Uses provider-native structured output via LangChain's jsonMode (OpenAI, Anthropic, etc.)\n * - 'native': Uses provider's constrained decoding API directly for guaranteed schema compliance\n *   (Anthropic output_config.format, OpenAI response_format.json_schema). Falls back to 'tool' for unsupported providers.\n * - 'auto': Automatically selects the best strategy — 'native' for supported providers, 'tool' for others\n */\nexport type StructuredOutputMode = 'tool' | 'provider' | 'native' | 'auto';\n\n/**\n * Resolved method used internally after mode resolution.\n * Maps to LangChain's withStructuredOutput method parameter plus our native path.\n */\nexport type ResolvedStructuredOutputMethod =\n  | 'functionCalling'\n  | 'jsonMode'\n  | 'jsonSchema'\n  | 'native'\n  | undefined;\n\n/**\n * Error thrown when the model refuses to produce structured output due to safety policies.\n */\nexport class StructuredOutputRefusalError extends Error {\n  constructor(public refusalText: string) {\n    super(`Model refused to produce structured output: ${refusalText}`);\n    this.name = 'StructuredOutputRefusalError';\n  }\n}\n\n/**\n * Error thrown when the structured output response was truncated due to max_tokens.\n */\nexport class StructuredOutputTruncatedError extends Error {\n  constructor(public stopReason: string) {\n    super(\n      `Structured output was truncated (stop_reason: ${stopReason}). ` +\n        'Increase max_tokens to allow the full JSON response to be generated.'\n    );\n    this.name = 'StructuredOutputTruncatedError';\n  }\n}\n\n/**\n * Configuration for structured JSON output from agents.\n * When configured, the agent will return a validated JSON response\n * instead of streaming text.\n */\nexport interface StructuredOutputConfig {\n  /**\n   * JSON Schema defining the output structure.\n   * The model will be forced to return data conforming to this schema.\n   */\n  schema: Record<string, unknown>;\n  /**\n   * Name for the structured output format (used in tool mode).\n   * @default 'StructuredResponse'\n   */\n  name?: string;\n  /**\n   * Description of what the structured output represents.\n   * Helps the model understand the expected format.\n   */\n  description?: string;\n  /**\n   * Output mode strategy.\n   * @default 'auto'\n   */\n  mode?: StructuredOutputMode;\n  /**\n   * Enable strict schema validation.\n   * When true, the response must exactly match the schema.\n   * @default true\n   */\n  strict?: boolean;\n  /**\n   * Error handling configuration.\n   * - true: Auto-retry on validation errors (default)\n   * - false: Throw error on validation failure\n   * - string: Custom error message for retry\n   */\n  handleErrors?: boolean | string;\n  /**\n   * Maximum number of retry attempts on validation failure.\n   * @default 2\n   */\n  maxRetries?: number;\n  /**\n   * Include the raw AI message along with structured response.\n   * Useful for debugging.\n   * @default false\n   */\n  includeRaw?: boolean;\n}\n\n/**\n * Database/API structured output format (snake_case with enabled flag).\n * This matches the format stored in MongoDB and sent from frontends.\n */\nexport interface StructuredOutputInput {\n  /** Whether structured output is enabled */\n  enabled?: boolean;\n  /** JSON Schema defining the expected response structure */\n  schema?: Record<string, unknown>;\n  /** Name identifier for the structured output */\n  name?: string;\n  /** Description of what the structured output represents */\n  description?: string;\n  /** Mode for structured output: 'tool' | 'provider' | 'native' | 'auto' */\n  mode?: StructuredOutputMode;\n  /** Whether to enforce strict schema validation */\n  strict?: boolean;\n}\n\n/**\n * Trigger strategy for when summarization should activate.\n * - 'contextPercentage': Trigger when context utilization exceeds a threshold percentage\n * - 'messageCount': Trigger when pruned message count exceeds a threshold\n * - 'tokenThreshold': Trigger when total token count exceeds a raw threshold\n */\nexport type SummarizationTriggerType =\n  | 'contextPercentage'\n  | 'messageCount'\n  | 'tokenThreshold';\n\n/**\n * Configuration for summarization behavior within the agent pipeline.\n * All fields are optional — sensible defaults are provided via constants.\n *\n * @see SUMMARIZATION_CONTEXT_THRESHOLD, SUMMARIZATION_RESERVE_RATIO, PRUNING_EMA_ALPHA\n */\nexport interface SummarizationConfig {\n  /**\n   * Strategy for when summarization triggers.\n   * @default 'contextPercentage'\n   */\n  triggerType?: SummarizationTriggerType;\n\n  /**\n   * Threshold value interpreted based on triggerType:\n   * - contextPercentage: 0-100 (percentage of context window)\n   * - messageCount: absolute count of messages pruned\n   * - tokenThreshold: absolute token count\n   * @default 80 (for contextPercentage)\n   */\n  triggerThreshold?: number;\n\n  /**\n   * Fraction of context window (0-1) reserved for recent messages.\n   * Prevents over-pruning by ensuring at least this fraction of the\n   * context budget is preserved as recent conversation history.\n   * @default 0.3\n   */\n  reserveRatio?: number;\n\n  /**\n   * Whether context pruning is enabled (can be disabled for debugging).\n   * @default true\n   */\n  contextPruning?: boolean;\n\n  /**\n   * Initial summary text to seed across runs.\n   * Different from persistedSummary: this is provided by the caller as a\n   * cross-conversation seed (e.g., agent personality or recurring context),\n   * while persistedSummary is loaded from the conversation's own history.\n   */\n  initialSummary?: string;\n\n  /**\n   * optional fields. When the host wants the summarization\n   * pass to run on a different LLM than the agent's own (e.g. a cheaper\n   * model for compaction), these provider/model overrides flow through to\n   * the summarize callback.\n   */\n  provider?: Providers;\n  model?: string;\n  parameters?: Record<string, unknown>;\n  prompt?: string;\n  updatePrompt?: string;\n  trigger?: { type: string; value: number };\n  maxSummaryTokens?: number;\n}\n\n/**\n * Runtime state for EMA-based pruning calibration.\n * Maintained across iterations within a single run to smooth pruning decisions.\n */\nexport interface PruneCalibrationState {\n  /** Current EMA calibration ratio */\n  ratio: number;\n  /** Number of calibration updates applied */\n  iterations: number;\n}\n\n/**\n * Lightweight file metadata entry for conversation-level file awareness.\n * Contains only IDs and names — NOT full content — so the agent always knows\n * what files exist in the conversation even after compaction pushes old messages\n * behind the summary window. The agent can retrieve full content on-demand\n * via file_search (RAG) or content_tool read (by contentId).\n */\nexport interface FileManifestEntry {\n  /** Unique file identifier (e.g., MongoDB ObjectId or UUID) */\n  fileId: string;\n  /** Original filename (e.g., \"quarterly-report.pdf\") */\n  filename: string;\n  /** Content identifier for on-demand retrieval via content_tool read */\n  contentId?: string;\n  /** File source (e.g., \"local\", \"sharepoint\", \"onedrive\") */\n  source?: string;\n  /** Index of the message that introduced this file (0-based in the original message array) */\n  messageIndex?: number;\n}\n\n/** Configuration for a subagent type that can be spawned by a parent agent. */\nexport type SubagentConfig = {\n  /** Identifier used in the tool's `subagent_type` enum (e.g. 'researcher', 'coder'). */\n  type: string;\n  /** Human-readable display name. */\n  name: string;\n  /** What this subagent specializes in — shown to the LLM. */\n  description: string;\n  /** Full agent config for the child graph. Omit when `self` is true. */\n  agentInputs?: AgentInputs;\n  /** When true, reuse the parent's AgentInputs (context isolation without separate config). */\n  self?: boolean;\n  /** Max AGENT→TOOLS cycles before forced stop (default: 25). */\n  maxTurns?: number;\n  /** Allow this subagent to spawn its own subagents (default: false). */\n  allowNested?: boolean;\n};\n\n/** SubagentConfig with agentInputs guaranteed present (self-spawn resolved). */\nexport type ResolvedSubagentConfig = SubagentConfig & {\n  agentInputs: AgentInputs;\n};\n\n/** Lifecycle phase carried on {@link SubagentUpdateEvent}. */\nexport type SubagentUpdatePhase =\n  | 'start'\n  | 'run_step'\n  | 'run_step_delta'\n  | 'run_step_completed'\n  | 'message_delta'\n  | 'reasoning_delta'\n  | 'stop'\n  | 'error';\n\n/**\n * Wrapper event emitted when a subagent's child graph dispatches activity.\n * Lets hosts show subagent progress in a UI surface separate from the parent\n * conversation without having to untangle events by agent ID.\n */\nexport interface SubagentUpdateEvent {\n  /** Parent run ID. */\n  runId: string;\n  /** Child run ID (unique per subagent execution). */\n  subagentRunId: string;\n  /**\n   * Parent-side `tool_call_id` for the `subagent` tool invocation that\n   * triggered this run. Stable for the duration of the child; lets hosts\n   * correlate updates deterministically instead of inferring by ordering.\n   * Omitted when the executor was invoked outside of a tool-call context.\n   */\n  parentToolCallId?: string;\n  /** Subagent `type` identifier from the SubagentConfig. */\n  subagentType: string;\n  /** Child agent ID assigned to this subagent execution. */\n  subagentAgentId: string;\n  /** Parent agent ID that spawned this subagent. */\n  parentAgentId?: string;\n  /** Lifecycle phase carried by this update. */\n  phase: SubagentUpdatePhase;\n  /** Underlying event payload (shape depends on phase). */\n  data?: unknown;\n  /** Short human-readable description. Hosts can render this directly. */\n  label?: string;\n  /** ISO timestamp for ordering / display. */\n  timestamp: string;\n}\n\nexport interface AgentInputs {\n  agentId: string;\n  /** Human-readable name for the agent (used in handoff context). Defaults to agentId if not provided. */\n  name?: string;\n  /** Description of what this agent does (used to enrich handoff tool descriptions). */\n  description?: string;\n  toolEnd?: boolean;\n  toolMap?: ToolMap;\n  tools?: GraphTools;\n  provider: Providers;\n  /**\n   * Ordered list of cacheable system content blocks emitted BEFORE\n   * `instructions` in the system message. Each block gets its own cache\n   * marker (cachePoint on Bedrock, cache_control on Anthropic) so the\n   * cache key for each block includes only the bytes up to and including\n   * its own marker. Entries are emitted in array order — earlier =\n   * stabler = wider cache key (best for cross-tenant sharing).\n   *\n   * See `src/types/agent-cache.ts` for full semantics. Capped at\n   * `MAX_SYSTEM_CACHE_BLOCKS` (2) entries to stay within Bedrock's\n   * 4-cachePoint budget after the tools array consumes 2.\n   */\n  system_cache_blocks?: SystemCacheBlock[];\n  /**\n   * Stable/cacheable system instructions for this agent. Always emitted\n   * with its own trailing cache marker (when caching is supported by\n   * the provider/model). Use `instructions_cache_ttl` to override the\n   * default '5m' TTL.\n   */\n  instructions?: string;\n  /** TTL for the trailing `instructions` cache marker. Defaults to '5m'. */\n  instructions_cache_ttl?: AgentCacheTTL;\n  streamBuffer?: number;\n  maxContextTokens?: number;\n  clientOptions?: ClientOptions;\n  /**\n   * Dynamic system tail appended after the cacheable instructions\n   * without any cache marker. Per-user / per-message context belongs\n   * here so it does not invalidate the cacheable prefix.\n   */\n  additional_instructions?: string;\n  reasoningKey?: 'reasoning_content' | 'reasoning';\n  /**\n   * Subagent types this agent may spawn. When non-empty, Graph injects the\n   * `subagent` tool into the agent's toolset with the per-config enum and\n   * description populated.\n   */\n  subagentConfigs?: SubagentConfig[];\n  /**\n   * Maximum subagent depth allowed below this agent. Used by SubagentExecutor\n   * to abort runaway nesting. Defaults to a small constant.\n   */\n  maxSubagentDepth?: number;\n  /**\n   * Pre-existing summary injected into the system message via\n   * AgentContext.buildInstructionsString. Populated by SubagentExecutor when\n   * spawning a self-config subagent that should inherit the parent's prior\n   * compaction summary; otherwise host-supplied at run time.\n   */\n  initialSummary?: { text: string; tokenCount: number };\n  /**\n   * opt-in for summarization. When false, the agent never\n   * invokes the summarize callback regardless of context utilization.\n   * Defaults to undefined (treated as enabled by default to preserve prior\n   * behaviour); hosts that want strict opt-in semantics should set it\n   * explicitly.\n   */\n  summarizationEnabled?: boolean;\n  /** Format content blocks as strings (for legacy compatibility i.e. Ollama/Azure Serverless) */\n  useLegacyContent?: boolean;\n  /**\n   * Tool definitions for all tools, including deferred and programmatic.\n   * Used for tool search and programmatic tool calling.\n   * Maps tool name to LCTool definition.\n   */\n  toolRegistry?: Map<string, LCTool>;\n  /**\n   * Dynamic context that changes per-request (e.g., current time, user info).\n   * This is injected as a user message rather than system prompt to preserve cache.\n   * Keeping this separate from instructions ensures the system message stays static\n   * and can be cached by Bedrock/Anthropic prompt caching.\n   */\n  dynamicContext?: string;\n  /**\n   * Structured output configuration (camelCase).\n   * When set, disables streaming and returns a validated JSON response\n   * conforming to the specified schema.\n   */\n  structuredOutput?: StructuredOutputConfig;\n  /**\n   * Structured output configuration (snake_case - database/API format).\n   * Alternative to structuredOutput for compatibility with MongoDB/frontend.\n   * Uses an `enabled` flag to control activation.\n   * @deprecated Use structuredOutput instead when possible\n   */\n  structured_output?: StructuredOutputInput;\n  /**\n   * Serializable tool definitions for event-driven execution.\n   * When provided, ToolNode operates in event-driven mode, dispatching\n   * ON_TOOL_EXECUTE events instead of invoking tools directly.\n   */\n  toolDefinitions?: LCTool[];\n  /**\n   * Tool names discovered from previous conversation history.\n   * These tools will be pre-marked as discovered so they're included\n   * in tool binding without requiring tool_search.\n   */\n  discoveredTools?: string[];\n  /**\n   * Optional callback for summarizing messages that were pruned from context.\n   * When provided, discarded messages are summarized by the host caller\n   * using a cheap LLM call, and the summary is prepended to the context.\n   */\n  summarizeCallback?: (\n    messagesToRefine: import('@langchain/core/messages').BaseMessage[]\n  ) => Promise<string | undefined>;\n  /**\n   * Pre-existing summary text loaded from persistent storage (MongoDB/Redis).\n   * When provided, this summary is injected into the initial message context\n   * so the agent has prior conversation history even on new turns.\n   * Set by the host's summary store when resuming a conversation.\n   */\n  persistedSummary?: string;\n  /**\n   * Summarization configuration controlling trigger strategy, reserve ratio,\n   * and EMA calibration for pruning. When omitted, sensible defaults apply.\n   * @see SummarizationConfig\n   */\n  summarizationConfig?: SummarizationConfig;\n  /**\n   * Lightweight file manifest for the conversation.\n   * Contains file IDs, names, and metadata — NOT full content.\n   *\n   * Used by the compaction engine to inject a [Conversation Files] block\n   * into the windowed view, ensuring the LLM always knows what files exist\n   * even when old messages (with full file content) are behind the summary.\n   *\n   * The agent can retrieve full content on-demand via:\n   *   - file_search (RAG semantic search over embedded files)\n   *   - content_tool read (by contentId for exact file retrieval)\n   *\n   * Built by the host orchestrator from message_file_map\n   * and metadata.context_files across all conversation messages.\n   */\n  fileManifest?: FileManifestEntry[];\n}\n\n/**\n * Tunable knobs for position-based content degradation.\n * See `messages/contextPruning.ts` and `messages/contextPruningSettings.ts`.\n */\nexport interface ContextPruningConfig {\n  enabled?: boolean;\n  keepLastAssistants?: number;\n  softTrimRatio?: number;\n  hardClearRatio?: number;\n  minPrunableToolChars?: number;\n  softTrim?: {\n    maxChars?: number;\n    headChars?: number;\n    tailChars?: number;\n  };\n  hardClear?: {\n    enabled?: boolean;\n    placeholder?: string;\n  };\n}\n"],"names":[],"mappings":";;AAsdA;;AAEG;AACG,MAAO,4BAA6B,SAAQ,KAAK,CAAA;AAClC,IAAA,WAAA;AAAnB,IAAA,WAAA,CAAmB,WAAmB,EAAA;AACpC,QAAA,KAAK,CAAC,CAAA,4CAAA,EAA+C,WAAW,CAAA,CAAE,CAAC;QADlD,IAAA,CAAA,WAAW,GAAX,WAAW;AAE5B,QAAA,IAAI,CAAC,IAAI,GAAG,8BAA8B;IAC5C;AACD;AAED;;AAEG;AACG,MAAO,8BAA+B,SAAQ,KAAK,CAAA;AACpC,IAAA,UAAA;AAAnB,IAAA,WAAA,CAAmB,UAAkB,EAAA;QACnC,KAAK,CACH,CAAA,8CAAA,EAAiD,UAAU,CAAA,GAAA,CAAK;AAC9D,YAAA,sEAAsE,CACzE;QAJgB,IAAA,CAAA,UAAU,GAAV,UAAU;AAK3B,QAAA,IAAI,CAAC,IAAI,GAAG,gCAAgC;IAC9C;AACD;;;;;"}

package/dist/esm/agents/AgentContext.mjs

@@ -65,7 +65,7 @@ class AgentContext {
             fileManifest,
         });
         /**
-         * Track upstream-aligned subagent inputs on the context. `_sourceInputs`
+         * Track subagent inputs on the context. `_sourceInputs`
          * preserves the original AgentInputs so SubagentExecutor can self-spawn
          * (`SubagentConfig.self === true`) without separate config; the other
          * two flow straight from agentConfig.
@@ -765,7 +765,7 @@ class AgentContext {
      * appended unfiltered; outside event-driven mode they pass through
      * `filterToolsForBinding`. Centralizing the decision here prevents the
      * accounting/binding paths from drifting apart, which was the root
-     * cause of the original miscount (Fixes upstream #121).
+     * cause of the original miscount (Fixes #121).
      */
     getEffectiveInstanceTools() {
         if (!this.tools) {