apex-code 0.0.1-alpha.0 → 0.0.1-alpha.2

package/dist/server/create-harness.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"create-harness.d.ts","sourceRoot":"","sources":["../../src/server/create-harness.ts"],"names":[],"mappings":"AAAA,OAAO,EACN,YAAY,EACZ,KAAK,mBAAmB,EAMxB,KAAK,YAAY,EAEjB,KAAK,WAAW,EAChB,MAAM,sBAAsB,CAAC;AAE9B,OAAO,EAAE,KAAK,wBAAwB,EAAqB,MAAM,0BAA0B,CAAC;AAM5F,MAAM,WAAW,sBAAuB,SAAQ,WAAW;IAC1D,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,gBAAgB,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;CACrC;AAeD,MAAM,WAAW,+BAAgC,SAAQ,IAAI,CAAC,mBAAmB,EAAE,aAAa,GAAG,OAAO,CAAC;IAC1G,GAAG,EAAE,YAAY,CAAC;IAClB,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,0FAA0F;IAC1F,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,KAAK,CAAC,EAAE,sBAAsB,EAAE,CAAC;IACjC,mBAAmB,CAAC,EAAE,IAAI,CAAC,wBAAwB,EAAE,KAAK,GAAG,kBAAkB,GAAG,eAAe,GAAG,cAAc,CAAC,CAAC;CACpH;AAED,MAAM,WAAW,0CAA0C;IAC1D,GAAG,EAAE,MAAM,CAAC;IACZ,KAAK,EAAE,SAAS,sBAAsB,EAAE,CAAC;IACzC,eAAe,EAAE,SAAS,MAAM,EAAE,CAAC;IACnC,mBAAmB,CAAC,EAAE,+BAA+B,CAAC,qBAAqB,CAAC,CAAC;CAC7E;AAED,wBAAgB,mCAAmC,CAAC,OAAO,EAAE,0CAA0C,GAAG,MAAM,CAsB/G;AAED,wBAAsB,wBAAwB,CAAC,OAAO,EAAE,+BAA+B;;;GAiFtF","sourcesContent":["import {\n\tAgentHarness,\n\ttype AgentHarnessOptions,\n\ttype AgentHarnessTool,\n\tcreateBashTool,\n\tcreateEditTool,\n\tcreateReadTool,\n\tcreateWriteTool,\n\ttype ExecutionEnv,\n\ttype ExecutionToolContext,\n\ttype HarnessTool,\n} from \"apex-code-agent-core\";\nimport type { Static, TSchema } from \"typebox\";\nimport { type BuildSystemPromptOptions, buildSystemPrompt } from \"../core/system-prompt.ts\";\nimport { bashToolSystemPromptContribution } from \"../core/tools/bash.ts\";\nimport { editToolSystemPromptContribution } from \"../core/tools/edit.ts\";\nimport { readToolSystemPromptContribution } from \"../core/tools/read.ts\";\nimport { writeToolSystemPromptContribution } from \"../core/tools/write.ts\";\n\nexport interface CodingAgentHarnessTool extends HarnessTool {\n\tpromptSnippet?: string;\n\tpromptGuidelines?: readonly string[];\n}\n\nfunction createCodingAgentHarnessTool<TParameters extends TSchema, TDetails>(\n\ttool: AgentHarnessTool<ExecutionToolContext, TParameters, TDetails>,\n\tcontext: ExecutionToolContext,\n\tprompt: Required<Pick<CodingAgentHarnessTool, \"promptSnippet\" | \"promptGuidelines\">>,\n): CodingAgentHarnessTool {\n\treturn {\n\t\t...tool,\n\t\t...prompt,\n\t\texecute: (toolCallId, params, signal, onUpdate) =>\n\t\t\ttool.execute(toolCallId, params as Static<TParameters>, signal, onUpdate, context),\n\t};\n}\n\nexport interface CreateCodingAgentHarnessOptions extends Omit<AgentHarnessOptions, \"toolContext\" | \"tools\"> {\n\tenv: ExecutionEnv;\n\tbashCommandPrefix?: string;\n\t/** Path to the JSONL session file exposed to default bash commands as PI_SESSION_FILE. */\n\tsessionFile?: string;\n\ttools?: CodingAgentHarnessTool[];\n\tsystemPromptOptions?: Omit<BuildSystemPromptOptions, \"cwd\" | \"promptGuidelines\" | \"selectedTools\" | \"toolSnippets\">;\n}\n\nexport interface BuildCodingAgentHarnessSystemPromptOptions {\n\tcwd: string;\n\ttools: readonly CodingAgentHarnessTool[];\n\tactiveToolNames: readonly string[];\n\tsystemPromptOptions?: CreateCodingAgentHarnessOptions[\"systemPromptOptions\"];\n}\n\nexport function buildCodingAgentHarnessSystemPrompt(options: BuildCodingAgentHarnessSystemPromptOptions): string {\n\tconst activeTools = options.activeToolNames.flatMap((name) => {\n\t\tconst tool = options.tools.find((candidate) => candidate.name === name);\n\t\treturn tool ? [tool] : [];\n\t});\n\tconst toolSnippets = Object.fromEntries(\n\t\tactiveTools.flatMap((tool) => {\n\t\t\tconst promptSnippet = tool.promptSnippet\n\t\t\t\t?.replace(/[\\r\\n]+/g, \" \")\n\t\t\t\t.replace(/\\s+/g, \" \")\n\t\t\t\t.trim();\n\t\t\treturn promptSnippet ? [[tool.name, promptSnippet]] : [];\n\t\t}),\n\t);\n\tconst promptGuidelines = activeTools.flatMap((tool) => tool.promptGuidelines ?? []);\n\treturn buildSystemPrompt({\n\t\t...options.systemPromptOptions,\n\t\tcwd: options.cwd,\n\t\tselectedTools: activeTools.map((tool) => tool.name),\n\t\ttoolSnippets,\n\t\tpromptGuidelines,\n\t});\n}\n\nexport async function createCodingAgentHarness(options: CreateCodingAgentHarnessOptions) {\n\tconst {\n\t\tenv,\n\t\tbashCommandPrefix,\n\t\tsessionFile,\n\t\tsystemPromptOptions,\n\t\ttools: providedTools,\n\t\tactiveToolNames: providedActiveToolNames,\n\t\tsystemPrompt: providedSystemPrompt,\n\t\t...harnessOptions\n\t} = options;\n\tlet harness: AgentHarness | undefined;\n\tconst getHarness = (): AgentHarness => {\n\t\tif (!harness) throw new Error(\"Coding-agent Harness callback ran before Harness initialization\");\n\t\treturn harness;\n\t};\n\tlet tools = providedTools;\n\tif (tools === undefined) {\n\t\tconst metadata = await options.session.getMetadata();\n\t\tconst toolContext = { env } satisfies ExecutionToolContext;\n\t\ttools = [\n\t\t\tcreateCodingAgentHarnessTool(createReadTool<ExecutionToolContext>(), toolContext, {\n\t\t\t\tpromptSnippet: readToolSystemPromptContribution.snippet,\n\t\t\t\tpromptGuidelines: readToolSystemPromptContribution.guidelines,\n\t\t\t}),\n\t\t\tcreateCodingAgentHarnessTool(\n\t\t\t\tcreateBashTool<ExecutionToolContext>({\n\t\t\t\t\tcommandPrefix: bashCommandPrefix,\n\t\t\t\t\tprepare: async (execution) => {\n\t\t\t\t\t\tconst currentHarness = getHarness();\n\t\t\t\t\t\tconst [model, thinkingLevel] = await Promise.all([\n\t\t\t\t\t\t\tcurrentHarness.getModel(),\n\t\t\t\t\t\t\tcurrentHarness.getThinkingLevel(),\n\t\t\t\t\t\t]);\n\t\t\t\t\t\texecution.env.PI_SESSION_ID = metadata.id;\n\t\t\t\t\t\texecution.env.PI_SESSION_FILE = sessionFile ?? \"\";\n\t\t\t\t\t\texecution.env.PI_PROVIDER = model.provider;\n\t\t\t\t\t\texecution.env.PI_MODEL = model.id;\n\t\t\t\t\t\texecution.env.PI_REASONING_LEVEL = thinkingLevel;\n\t\t\t\t\t},\n\t\t\t\t}),\n\t\t\t\ttoolContext,\n\t\t\t\t{\n\t\t\t\t\tpromptSnippet: bashToolSystemPromptContribution.snippet,\n\t\t\t\t\tpromptGuidelines: bashToolSystemPromptContribution.guidelines,\n\t\t\t\t},\n\t\t\t),\n\t\t\tcreateCodingAgentHarnessTool(createEditTool<ExecutionToolContext>(), toolContext, {\n\t\t\t\tpromptSnippet: editToolSystemPromptContribution.snippet,\n\t\t\t\tpromptGuidelines: editToolSystemPromptContribution.guidelines,\n\t\t\t}),\n\t\t\tcreateCodingAgentHarnessTool(createWriteTool<ExecutionToolContext>(), toolContext, {\n\t\t\t\tpromptSnippet: writeToolSystemPromptContribution.snippet,\n\t\t\t\tpromptGuidelines: writeToolSystemPromptContribution.guidelines,\n\t\t\t}),\n\t\t];\n\t}\n\tconst activeToolNames = [...(providedActiveToolNames ?? tools.map((tool) => tool.name))];\n\tconst systemPrompt =\n\t\tprovidedSystemPrompt ??\n\t\t(async () => {\n\t\t\tconst currentHarness = getHarness();\n\t\t\tconst [currentTools, currentActiveToolNames] = await Promise.all([\n\t\t\t\tcurrentHarness.getTools(),\n\t\t\t\tcurrentHarness.getActiveTools(),\n\t\t\t]);\n\t\t\treturn buildCodingAgentHarnessSystemPrompt({\n\t\t\t\tcwd: env.cwd,\n\t\t\t\ttools: currentTools,\n\t\t\t\tactiveToolNames: currentActiveToolNames,\n\t\t\t\tsystemPromptOptions,\n\t\t\t});\n\t\t});\n\tconst created = await AgentHarness.create({\n\t\t...harnessOptions,\n\t\ttools,\n\t\tactiveToolNames,\n\t\tsystemPrompt,\n\t});\n\tharness = created.harness;\n\treturn created;\n}\n"]}
+{"version":3,"file":"create-harness.d.ts","sourceRoot":"","sources":["../../src/server/create-harness.ts"],"names":[],"mappings":"AAAA,OAAO,EACN,YAAY,EACZ,KAAK,mBAAmB,EAMxB,KAAK,YAAY,EAEjB,KAAK,WAAW,EAChB,MAAM,sBAAsB,CAAC;AAG9B,OAAO,EAAE,KAAK,wBAAwB,EAAqB,MAAM,0BAA0B,CAAC;AAM5F,MAAM,WAAW,sBAAuB,SAAQ,WAAW;IAC1D,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,gBAAgB,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;CACrC;AAeD,MAAM,WAAW,+BAAgC,SAAQ,IAAI,CAAC,mBAAmB,EAAE,aAAa,GAAG,OAAO,CAAC;IAC1G,GAAG,EAAE,YAAY,CAAC;IAClB,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,iGAAiG;IACjG,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,KAAK,CAAC,EAAE,sBAAsB,EAAE,CAAC;IACjC,mBAAmB,CAAC,EAAE,IAAI,CAAC,wBAAwB,EAAE,KAAK,GAAG,kBAAkB,GAAG,eAAe,GAAG,cAAc,CAAC,CAAC;CACpH;AAED,MAAM,WAAW,0CAA0C;IAC1D,GAAG,EAAE,MAAM,CAAC;IACZ,KAAK,EAAE,SAAS,sBAAsB,EAAE,CAAC;IACzC,eAAe,EAAE,SAAS,MAAM,EAAE,CAAC;IACnC,mBAAmB,CAAC,EAAE,+BAA+B,CAAC,qBAAqB,CAAC,CAAC;CAC7E;AAED,wBAAgB,mCAAmC,CAAC,OAAO,EAAE,0CAA0C,GAAG,MAAM,CAsB/G;AAED,wBAAsB,wBAAwB,CAAC,OAAO,EAAE,+BAA+B;;;GAiFtF","sourcesContent":["import {\n\tAgentHarness,\n\ttype AgentHarnessOptions,\n\ttype AgentHarnessTool,\n\tcreateBashTool,\n\tcreateEditTool,\n\tcreateReadTool,\n\tcreateWriteTool,\n\ttype ExecutionEnv,\n\ttype ExecutionToolContext,\n\ttype HarnessTool,\n} from \"apex-code-agent-core\";\nimport type { Static, TSchema } from \"typebox\";\nimport { setApexEnvironment } from \"../core/environment.ts\";\nimport { type BuildSystemPromptOptions, buildSystemPrompt } from \"../core/system-prompt.ts\";\nimport { bashToolSystemPromptContribution } from \"../core/tools/bash.ts\";\nimport { editToolSystemPromptContribution } from \"../core/tools/edit.ts\";\nimport { readToolSystemPromptContribution } from \"../core/tools/read.ts\";\nimport { writeToolSystemPromptContribution } from \"../core/tools/write.ts\";\n\nexport interface CodingAgentHarnessTool extends HarnessTool {\n\tpromptSnippet?: string;\n\tpromptGuidelines?: readonly string[];\n}\n\nfunction createCodingAgentHarnessTool<TParameters extends TSchema, TDetails>(\n\ttool: AgentHarnessTool<ExecutionToolContext, TParameters, TDetails>,\n\tcontext: ExecutionToolContext,\n\tprompt: Required<Pick<CodingAgentHarnessTool, \"promptSnippet\" | \"promptGuidelines\">>,\n): CodingAgentHarnessTool {\n\treturn {\n\t\t...tool,\n\t\t...prompt,\n\t\texecute: (toolCallId, params, signal, onUpdate) =>\n\t\t\ttool.execute(toolCallId, params as Static<TParameters>, signal, onUpdate, context),\n\t};\n}\n\nexport interface CreateCodingAgentHarnessOptions extends Omit<AgentHarnessOptions, \"toolContext\" | \"tools\"> {\n\tenv: ExecutionEnv;\n\tbashCommandPrefix?: string;\n\t/** Path to the JSONL session file exposed to default bash commands as APEX_CODE_SESSION_FILE. */\n\tsessionFile?: string;\n\ttools?: CodingAgentHarnessTool[];\n\tsystemPromptOptions?: Omit<BuildSystemPromptOptions, \"cwd\" | \"promptGuidelines\" | \"selectedTools\" | \"toolSnippets\">;\n}\n\nexport interface BuildCodingAgentHarnessSystemPromptOptions {\n\tcwd: string;\n\ttools: readonly CodingAgentHarnessTool[];\n\tactiveToolNames: readonly string[];\n\tsystemPromptOptions?: CreateCodingAgentHarnessOptions[\"systemPromptOptions\"];\n}\n\nexport function buildCodingAgentHarnessSystemPrompt(options: BuildCodingAgentHarnessSystemPromptOptions): string {\n\tconst activeTools = options.activeToolNames.flatMap((name) => {\n\t\tconst tool = options.tools.find((candidate) => candidate.name === name);\n\t\treturn tool ? [tool] : [];\n\t});\n\tconst toolSnippets = Object.fromEntries(\n\t\tactiveTools.flatMap((tool) => {\n\t\t\tconst promptSnippet = tool.promptSnippet\n\t\t\t\t?.replace(/[\\r\\n]+/g, \" \")\n\t\t\t\t.replace(/\\s+/g, \" \")\n\t\t\t\t.trim();\n\t\t\treturn promptSnippet ? [[tool.name, promptSnippet]] : [];\n\t\t}),\n\t);\n\tconst promptGuidelines = activeTools.flatMap((tool) => tool.promptGuidelines ?? []);\n\treturn buildSystemPrompt({\n\t\t...options.systemPromptOptions,\n\t\tcwd: options.cwd,\n\t\tselectedTools: activeTools.map((tool) => tool.name),\n\t\ttoolSnippets,\n\t\tpromptGuidelines,\n\t});\n}\n\nexport async function createCodingAgentHarness(options: CreateCodingAgentHarnessOptions) {\n\tconst {\n\t\tenv,\n\t\tbashCommandPrefix,\n\t\tsessionFile,\n\t\tsystemPromptOptions,\n\t\ttools: providedTools,\n\t\tactiveToolNames: providedActiveToolNames,\n\t\tsystemPrompt: providedSystemPrompt,\n\t\t...harnessOptions\n\t} = options;\n\tlet harness: AgentHarness | undefined;\n\tconst getHarness = (): AgentHarness => {\n\t\tif (!harness) throw new Error(\"Coding-agent Harness callback ran before Harness initialization\");\n\t\treturn harness;\n\t};\n\tlet tools = providedTools;\n\tif (tools === undefined) {\n\t\tconst metadata = await options.session.getMetadata();\n\t\tconst toolContext = { env } satisfies ExecutionToolContext;\n\t\ttools = [\n\t\t\tcreateCodingAgentHarnessTool(createReadTool<ExecutionToolContext>(), toolContext, {\n\t\t\t\tpromptSnippet: readToolSystemPromptContribution.snippet,\n\t\t\t\tpromptGuidelines: readToolSystemPromptContribution.guidelines,\n\t\t\t}),\n\t\t\tcreateCodingAgentHarnessTool(\n\t\t\t\tcreateBashTool<ExecutionToolContext>({\n\t\t\t\t\tcommandPrefix: bashCommandPrefix,\n\t\t\t\t\tprepare: async (execution) => {\n\t\t\t\t\t\tconst currentHarness = getHarness();\n\t\t\t\t\t\tconst [model, thinkingLevel] = await Promise.all([\n\t\t\t\t\t\t\tcurrentHarness.getModel(),\n\t\t\t\t\t\t\tcurrentHarness.getThinkingLevel(),\n\t\t\t\t\t\t]);\n\t\t\t\t\t\tsetApexEnvironment(\"APEX_CODE_SESSION_ID\", metadata.id, execution.env);\n\t\t\t\t\t\tsetApexEnvironment(\"APEX_CODE_SESSION_FILE\", sessionFile ?? \"\", execution.env);\n\t\t\t\t\t\tsetApexEnvironment(\"APEX_CODE_PROVIDER\", model.provider, execution.env);\n\t\t\t\t\t\tsetApexEnvironment(\"APEX_CODE_MODEL\", model.id, execution.env);\n\t\t\t\t\t\tsetApexEnvironment(\"APEX_CODE_REASONING_LEVEL\", thinkingLevel, execution.env);\n\t\t\t\t\t},\n\t\t\t\t}),\n\t\t\t\ttoolContext,\n\t\t\t\t{\n\t\t\t\t\tpromptSnippet: bashToolSystemPromptContribution.snippet,\n\t\t\t\t\tpromptGuidelines: bashToolSystemPromptContribution.guidelines,\n\t\t\t\t},\n\t\t\t),\n\t\t\tcreateCodingAgentHarnessTool(createEditTool<ExecutionToolContext>(), toolContext, {\n\t\t\t\tpromptSnippet: editToolSystemPromptContribution.snippet,\n\t\t\t\tpromptGuidelines: editToolSystemPromptContribution.guidelines,\n\t\t\t}),\n\t\t\tcreateCodingAgentHarnessTool(createWriteTool<ExecutionToolContext>(), toolContext, {\n\t\t\t\tpromptSnippet: writeToolSystemPromptContribution.snippet,\n\t\t\t\tpromptGuidelines: writeToolSystemPromptContribution.guidelines,\n\t\t\t}),\n\t\t];\n\t}\n\tconst activeToolNames = [...(providedActiveToolNames ?? tools.map((tool) => tool.name))];\n\tconst systemPrompt =\n\t\tprovidedSystemPrompt ??\n\t\t(async () => {\n\t\t\tconst currentHarness = getHarness();\n\t\t\tconst [currentTools, currentActiveToolNames] = await Promise.all([\n\t\t\t\tcurrentHarness.getTools(),\n\t\t\t\tcurrentHarness.getActiveTools(),\n\t\t\t]);\n\t\t\treturn buildCodingAgentHarnessSystemPrompt({\n\t\t\t\tcwd: env.cwd,\n\t\t\t\ttools: currentTools,\n\t\t\t\tactiveToolNames: currentActiveToolNames,\n\t\t\t\tsystemPromptOptions,\n\t\t\t});\n\t\t});\n\tconst created = await AgentHarness.create({\n\t\t...harnessOptions,\n\t\ttools,\n\t\tactiveToolNames,\n\t\tsystemPrompt,\n\t});\n\tharness = created.harness;\n\treturn created;\n}\n"]}

package/dist/server/create-harness.js

@@ -1,4 +1,5 @@
 import { AgentHarness, createBashTool, createEditTool, createReadTool, createWriteTool, } from "apex-code-agent-core";
+import { setApexEnvironment } from "../core/environment.js";
 import { buildSystemPrompt } from "../core/system-prompt.js";
 import { bashToolSystemPromptContribution } from "../core/tools/bash.js";
 import { editToolSystemPromptContribution } from "../core/tools/edit.js";
@@ -57,11 +58,11 @@ export async function createCodingAgentHarness(options) {
                         currentHarness.getModel(),
                         currentHarness.getThinkingLevel(),
                     ]);
-                    execution.env.PI_SESSION_ID = metadata.id;
-                    execution.env.PI_SESSION_FILE = sessionFile ?? "";
-                    execution.env.PI_PROVIDER = model.provider;
-                    execution.env.PI_MODEL = model.id;
-                    execution.env.PI_REASONING_LEVEL = thinkingLevel;
+                    setApexEnvironment("APEX_CODE_SESSION_ID", metadata.id, execution.env);
+                    setApexEnvironment("APEX_CODE_SESSION_FILE", sessionFile ?? "", execution.env);
+                    setApexEnvironment("APEX_CODE_PROVIDER", model.provider, execution.env);
+                    setApexEnvironment("APEX_CODE_MODEL", model.id, execution.env);
+                    setApexEnvironment("APEX_CODE_REASONING_LEVEL", thinkingLevel, execution.env);
                 },
             }), toolContext, {
                 promptSnippet: bashToolSystemPromptContribution.snippet,

package/dist/server/create-harness.js.map

@@ -1 +1 @@
-{"version":3,"file":"create-harness.js","sourceRoot":"","sources":["../../src/server/create-harness.ts"],"names":[],"mappings":"AAAA,OAAO,EACN,YAAY,EAGZ,cAAc,EACd,cAAc,EACd,cAAc,EACd,eAAe,GAIf,MAAM,sBAAsB,CAAC;AAE9B,OAAO,EAAiC,iBAAiB,EAAE,MAAM,0BAA0B,CAAC;AAC5F,OAAO,EAAE,gCAAgC,EAAE,MAAM,uBAAuB,CAAC;AACzE,OAAO,EAAE,gCAAgC,EAAE,MAAM,uBAAuB,CAAC;AACzE,OAAO,EAAE,gCAAgC,EAAE,MAAM,uBAAuB,CAAC;AACzE,OAAO,EAAE,iCAAiC,EAAE,MAAM,wBAAwB,CAAC;AAO3E,SAAS,4BAA4B,CACpC,IAAmE,EACnE,OAA6B,EAC7B,MAAoF,EAC3D;IACzB,OAAO;QACN,GAAG,IAAI;QACP,GAAG,MAAM;QACT,OAAO,EAAE,CAAC,UAAU,EAAE,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,EAAE,CACjD,IAAI,CAAC,OAAO,CAAC,UAAU,EAAE,MAA6B,EAAE,MAAM,EAAE,QAAQ,EAAE,OAAO,CAAC;KACnF,CAAC;AAAA,CACF;AAkBD,MAAM,UAAU,mCAAmC,CAAC,OAAmD,EAAU;IAChH,MAAM,WAAW,GAAG,OAAO,CAAC,eAAe,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC;QAC7D,MAAM,IAAI,GAAG,OAAO,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,SAAS,EAAE,EAAE,CAAC,SAAS,CAAC,IAAI,KAAK,IAAI,CAAC,CAAC;QACxE,OAAO,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IAAA,CAC1B,CAAC,CAAC;IACH,MAAM,YAAY,GAAG,MAAM,CAAC,WAAW,CACtC,WAAW,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC;QAC7B,MAAM,aAAa,GAAG,IAAI,CAAC,aAAa;YACvC,EAAE,OAAO,CAAC,UAAU,EAAE,GAAG,CAAC;aACzB,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC;aACpB,IAAI,EAAE,CAAC;QACT,OAAO,aAAa,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,EAAE,aAAa,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IAAA,CACzD,CAAC,CACF,CAAC;IACF,MAAM,gBAAgB,GAAG,WAAW,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,gBAAgB,IAAI,EAAE,CAAC,CAAC;IACpF,OAAO,iBAAiB,CAAC;QACxB,GAAG,OAAO,CAAC,mBAAmB;QAC9B,GAAG,EAAE,OAAO,CAAC,GAAG;QAChB,aAAa,EAAE,WAAW,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC;QACnD,YAAY;QACZ,gBAAgB;KAChB,CAAC,CAAC;AAAA,CACH;AAED,MAAM,CAAC,KAAK,UAAU,wBAAwB,CAAC,OAAwC,EAAE;IACxF,MAAM,EACL,GAAG,EACH,iBAAiB,EACjB,WAAW,EACX,mBAAmB,EACnB,KAAK,EAAE,aAAa,EACpB,eAAe,EAAE,uBAAuB,EACxC,YAAY,EAAE,oBAAoB,EAClC,GAAG,cAAc,EACjB,GAAG,OAAO,CAAC;IACZ,IAAI,OAAiC,CAAC;IACtC,MAAM,UAAU,GAAG,GAAiB,EAAE,CAAC;QACtC,IAAI,CAAC,OAAO;YAAE,MAAM,IAAI,KAAK,CAAC,iEAAiE,CAAC,CAAC;QACjG,OAAO,OAAO,CAAC;IAAA,CACf,CAAC;IACF,IAAI,KAAK,GAAG,aAAa,CAAC;IAC1B,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;QACzB,MAAM,QAAQ,GAAG,MAAM,OAAO,CAAC,OAAO,CAAC,WAAW,EAAE,CAAC;QACrD,MAAM,WAAW,GAAG,EAAE,GAAG,EAAiC,CAAC;QAC3D,KAAK,GAAG;YACP,4BAA4B,CAAC,cAAc,EAAwB,EAAE,WAAW,EAAE;gBACjF,aAAa,EAAE,gCAAgC,CAAC,OAAO;gBACvD,gBAAgB,EAAE,gCAAgC,CAAC,UAAU;aAC7D,CAAC;YACF,4BAA4B,CAC3B,cAAc,CAAuB;gBACpC,aAAa,EAAE,iBAAiB;gBAChC,OAAO,EAAE,KAAK,EAAE,SAAS,EAAE,EAAE,CAAC;oBAC7B,MAAM,cAAc,GAAG,UAAU,EAAE,CAAC;oBACpC,MAAM,CAAC,KAAK,EAAE,aAAa,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC;wBAChD,cAAc,CAAC,QAAQ,EAAE;wBACzB,cAAc,CAAC,gBAAgB,EAAE;qBACjC,CAAC,CAAC;oBACH,SAAS,CAAC,GAAG,CAAC,aAAa,GAAG,QAAQ,CAAC,EAAE,CAAC;oBAC1C,SAAS,CAAC,GAAG,CAAC,eAAe,GAAG,WAAW,IAAI,EAAE,CAAC;oBAClD,SAAS,CAAC,GAAG,CAAC,WAAW,GAAG,KAAK,CAAC,QAAQ,CAAC;oBAC3C,SAAS,CAAC,GAAG,CAAC,QAAQ,GAAG,KAAK,CAAC,EAAE,CAAC;oBAClC,SAAS,CAAC,GAAG,CAAC,kBAAkB,GAAG,aAAa,CAAC;gBAAA,CACjD;aACD,CAAC,EACF,WAAW,EACX;gBACC,aAAa,EAAE,gCAAgC,CAAC,OAAO;gBACvD,gBAAgB,EAAE,gCAAgC,CAAC,UAAU;aAC7D,CACD;YACD,4BAA4B,CAAC,cAAc,EAAwB,EAAE,WAAW,EAAE;gBACjF,aAAa,EAAE,gCAAgC,CAAC,OAAO;gBACvD,gBAAgB,EAAE,gCAAgC,CAAC,UAAU;aAC7D,CAAC;YACF,4BAA4B,CAAC,eAAe,EAAwB,EAAE,WAAW,EAAE;gBAClF,aAAa,EAAE,iCAAiC,CAAC,OAAO;gBACxD,gBAAgB,EAAE,iCAAiC,CAAC,UAAU;aAC9D,CAAC;SACF,CAAC;IACH,CAAC;IACD,MAAM,eAAe,GAAG,CAAC,GAAG,CAAC,uBAAuB,IAAI,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IACzF,MAAM,YAAY,GACjB,oBAAoB;QACpB,CAAC,KAAK,IAAI,EAAE,CAAC;YACZ,MAAM,cAAc,GAAG,UAAU,EAAE,CAAC;YACpC,MAAM,CAAC,YAAY,EAAE,sBAAsB,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC;gBAChE,cAAc,CAAC,QAAQ,EAAE;gBACzB,cAAc,CAAC,cAAc,EAAE;aAC/B,CAAC,CAAC;YACH,OAAO,mCAAmC,CAAC;gBAC1C,GAAG,EAAE,GAAG,CAAC,GAAG;gBACZ,KAAK,EAAE,YAAY;gBACnB,eAAe,EAAE,sBAAsB;gBACvC,mBAAmB;aACnB,CAAC,CAAC;QAAA,CACH,CAAC,CAAC;IACJ,MAAM,OAAO,GAAG,MAAM,YAAY,CAAC,MAAM,CAAC;QACzC,GAAG,cAAc;QACjB,KAAK;QACL,eAAe;QACf,YAAY;KACZ,CAAC,CAAC;IACH,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC;IAC1B,OAAO,OAAO,CAAC;AAAA,CACf","sourcesContent":["import {\n\tAgentHarness,\n\ttype AgentHarnessOptions,\n\ttype AgentHarnessTool,\n\tcreateBashTool,\n\tcreateEditTool,\n\tcreateReadTool,\n\tcreateWriteTool,\n\ttype ExecutionEnv,\n\ttype ExecutionToolContext,\n\ttype HarnessTool,\n} from \"apex-code-agent-core\";\nimport type { Static, TSchema } from \"typebox\";\nimport { type BuildSystemPromptOptions, buildSystemPrompt } from \"../core/system-prompt.ts\";\nimport { bashToolSystemPromptContribution } from \"../core/tools/bash.ts\";\nimport { editToolSystemPromptContribution } from \"../core/tools/edit.ts\";\nimport { readToolSystemPromptContribution } from \"../core/tools/read.ts\";\nimport { writeToolSystemPromptContribution } from \"../core/tools/write.ts\";\n\nexport interface CodingAgentHarnessTool extends HarnessTool {\n\tpromptSnippet?: string;\n\tpromptGuidelines?: readonly string[];\n}\n\nfunction createCodingAgentHarnessTool<TParameters extends TSchema, TDetails>(\n\ttool: AgentHarnessTool<ExecutionToolContext, TParameters, TDetails>,\n\tcontext: ExecutionToolContext,\n\tprompt: Required<Pick<CodingAgentHarnessTool, \"promptSnippet\" | \"promptGuidelines\">>,\n): CodingAgentHarnessTool {\n\treturn {\n\t\t...tool,\n\t\t...prompt,\n\t\texecute: (toolCallId, params, signal, onUpdate) =>\n\t\t\ttool.execute(toolCallId, params as Static<TParameters>, signal, onUpdate, context),\n\t};\n}\n\nexport interface CreateCodingAgentHarnessOptions extends Omit<AgentHarnessOptions, \"toolContext\" | \"tools\"> {\n\tenv: ExecutionEnv;\n\tbashCommandPrefix?: string;\n\t/** Path to the JSONL session file exposed to default bash commands as PI_SESSION_FILE. */\n\tsessionFile?: string;\n\ttools?: CodingAgentHarnessTool[];\n\tsystemPromptOptions?: Omit<BuildSystemPromptOptions, \"cwd\" | \"promptGuidelines\" | \"selectedTools\" | \"toolSnippets\">;\n}\n\nexport interface BuildCodingAgentHarnessSystemPromptOptions {\n\tcwd: string;\n\ttools: readonly CodingAgentHarnessTool[];\n\tactiveToolNames: readonly string[];\n\tsystemPromptOptions?: CreateCodingAgentHarnessOptions[\"systemPromptOptions\"];\n}\n\nexport function buildCodingAgentHarnessSystemPrompt(options: BuildCodingAgentHarnessSystemPromptOptions): string {\n\tconst activeTools = options.activeToolNames.flatMap((name) => {\n\t\tconst tool = options.tools.find((candidate) => candidate.name === name);\n\t\treturn tool ? [tool] : [];\n\t});\n\tconst toolSnippets = Object.fromEntries(\n\t\tactiveTools.flatMap((tool) => {\n\t\t\tconst promptSnippet = tool.promptSnippet\n\t\t\t\t?.replace(/[\\r\\n]+/g, \" \")\n\t\t\t\t.replace(/\\s+/g, \" \")\n\t\t\t\t.trim();\n\t\t\treturn promptSnippet ? [[tool.name, promptSnippet]] : [];\n\t\t}),\n\t);\n\tconst promptGuidelines = activeTools.flatMap((tool) => tool.promptGuidelines ?? []);\n\treturn buildSystemPrompt({\n\t\t...options.systemPromptOptions,\n\t\tcwd: options.cwd,\n\t\tselectedTools: activeTools.map((tool) => tool.name),\n\t\ttoolSnippets,\n\t\tpromptGuidelines,\n\t});\n}\n\nexport async function createCodingAgentHarness(options: CreateCodingAgentHarnessOptions) {\n\tconst {\n\t\tenv,\n\t\tbashCommandPrefix,\n\t\tsessionFile,\n\t\tsystemPromptOptions,\n\t\ttools: providedTools,\n\t\tactiveToolNames: providedActiveToolNames,\n\t\tsystemPrompt: providedSystemPrompt,\n\t\t...harnessOptions\n\t} = options;\n\tlet harness: AgentHarness | undefined;\n\tconst getHarness = (): AgentHarness => {\n\t\tif (!harness) throw new Error(\"Coding-agent Harness callback ran before Harness initialization\");\n\t\treturn harness;\n\t};\n\tlet tools = providedTools;\n\tif (tools === undefined) {\n\t\tconst metadata = await options.session.getMetadata();\n\t\tconst toolContext = { env } satisfies ExecutionToolContext;\n\t\ttools = [\n\t\t\tcreateCodingAgentHarnessTool(createReadTool<ExecutionToolContext>(), toolContext, {\n\t\t\t\tpromptSnippet: readToolSystemPromptContribution.snippet,\n\t\t\t\tpromptGuidelines: readToolSystemPromptContribution.guidelines,\n\t\t\t}),\n\t\t\tcreateCodingAgentHarnessTool(\n\t\t\t\tcreateBashTool<ExecutionToolContext>({\n\t\t\t\t\tcommandPrefix: bashCommandPrefix,\n\t\t\t\t\tprepare: async (execution) => {\n\t\t\t\t\t\tconst currentHarness = getHarness();\n\t\t\t\t\t\tconst [model, thinkingLevel] = await Promise.all([\n\t\t\t\t\t\t\tcurrentHarness.getModel(),\n\t\t\t\t\t\t\tcurrentHarness.getThinkingLevel(),\n\t\t\t\t\t\t]);\n\t\t\t\t\t\texecution.env.PI_SESSION_ID = metadata.id;\n\t\t\t\t\t\texecution.env.PI_SESSION_FILE = sessionFile ?? \"\";\n\t\t\t\t\t\texecution.env.PI_PROVIDER = model.provider;\n\t\t\t\t\t\texecution.env.PI_MODEL = model.id;\n\t\t\t\t\t\texecution.env.PI_REASONING_LEVEL = thinkingLevel;\n\t\t\t\t\t},\n\t\t\t\t}),\n\t\t\t\ttoolContext,\n\t\t\t\t{\n\t\t\t\t\tpromptSnippet: bashToolSystemPromptContribution.snippet,\n\t\t\t\t\tpromptGuidelines: bashToolSystemPromptContribution.guidelines,\n\t\t\t\t},\n\t\t\t),\n\t\t\tcreateCodingAgentHarnessTool(createEditTool<ExecutionToolContext>(), toolContext, {\n\t\t\t\tpromptSnippet: editToolSystemPromptContribution.snippet,\n\t\t\t\tpromptGuidelines: editToolSystemPromptContribution.guidelines,\n\t\t\t}),\n\t\t\tcreateCodingAgentHarnessTool(createWriteTool<ExecutionToolContext>(), toolContext, {\n\t\t\t\tpromptSnippet: writeToolSystemPromptContribution.snippet,\n\t\t\t\tpromptGuidelines: writeToolSystemPromptContribution.guidelines,\n\t\t\t}),\n\t\t];\n\t}\n\tconst activeToolNames = [...(providedActiveToolNames ?? tools.map((tool) => tool.name))];\n\tconst systemPrompt =\n\t\tprovidedSystemPrompt ??\n\t\t(async () => {\n\t\t\tconst currentHarness = getHarness();\n\t\t\tconst [currentTools, currentActiveToolNames] = await Promise.all([\n\t\t\t\tcurrentHarness.getTools(),\n\t\t\t\tcurrentHarness.getActiveTools(),\n\t\t\t]);\n\t\t\treturn buildCodingAgentHarnessSystemPrompt({\n\t\t\t\tcwd: env.cwd,\n\t\t\t\ttools: currentTools,\n\t\t\t\tactiveToolNames: currentActiveToolNames,\n\t\t\t\tsystemPromptOptions,\n\t\t\t});\n\t\t});\n\tconst created = await AgentHarness.create({\n\t\t...harnessOptions,\n\t\ttools,\n\t\tactiveToolNames,\n\t\tsystemPrompt,\n\t});\n\tharness = created.harness;\n\treturn created;\n}\n"]}
+{"version":3,"file":"create-harness.js","sourceRoot":"","sources":["../../src/server/create-harness.ts"],"names":[],"mappings":"AAAA,OAAO,EACN,YAAY,EAGZ,cAAc,EACd,cAAc,EACd,cAAc,EACd,eAAe,GAIf,MAAM,sBAAsB,CAAC;AAE9B,OAAO,EAAE,kBAAkB,EAAE,MAAM,wBAAwB,CAAC;AAC5D,OAAO,EAAiC,iBAAiB,EAAE,MAAM,0BAA0B,CAAC;AAC5F,OAAO,EAAE,gCAAgC,EAAE,MAAM,uBAAuB,CAAC;AACzE,OAAO,EAAE,gCAAgC,EAAE,MAAM,uBAAuB,CAAC;AACzE,OAAO,EAAE,gCAAgC,EAAE,MAAM,uBAAuB,CAAC;AACzE,OAAO,EAAE,iCAAiC,EAAE,MAAM,wBAAwB,CAAC;AAO3E,SAAS,4BAA4B,CACpC,IAAmE,EACnE,OAA6B,EAC7B,MAAoF,EAC3D;IACzB,OAAO;QACN,GAAG,IAAI;QACP,GAAG,MAAM;QACT,OAAO,EAAE,CAAC,UAAU,EAAE,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,EAAE,CACjD,IAAI,CAAC,OAAO,CAAC,UAAU,EAAE,MAA6B,EAAE,MAAM,EAAE,QAAQ,EAAE,OAAO,CAAC;KACnF,CAAC;AAAA,CACF;AAkBD,MAAM,UAAU,mCAAmC,CAAC,OAAmD,EAAU;IAChH,MAAM,WAAW,GAAG,OAAO,CAAC,eAAe,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC;QAC7D,MAAM,IAAI,GAAG,OAAO,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,SAAS,EAAE,EAAE,CAAC,SAAS,CAAC,IAAI,KAAK,IAAI,CAAC,CAAC;QACxE,OAAO,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IAAA,CAC1B,CAAC,CAAC;IACH,MAAM,YAAY,GAAG,MAAM,CAAC,WAAW,CACtC,WAAW,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC;QAC7B,MAAM,aAAa,GAAG,IAAI,CAAC,aAAa;YACvC,EAAE,OAAO,CAAC,UAAU,EAAE,GAAG,CAAC;aACzB,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC;aACpB,IAAI,EAAE,CAAC;QACT,OAAO,aAAa,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,EAAE,aAAa,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IAAA,CACzD,CAAC,CACF,CAAC;IACF,MAAM,gBAAgB,GAAG,WAAW,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,gBAAgB,IAAI,EAAE,CAAC,CAAC;IACpF,OAAO,iBAAiB,CAAC;QACxB,GAAG,OAAO,CAAC,mBAAmB;QAC9B,GAAG,EAAE,OAAO,CAAC,GAAG;QAChB,aAAa,EAAE,WAAW,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC;QACnD,YAAY;QACZ,gBAAgB;KAChB,CAAC,CAAC;AAAA,CACH;AAED,MAAM,CAAC,KAAK,UAAU,wBAAwB,CAAC,OAAwC,EAAE;IACxF,MAAM,EACL,GAAG,EACH,iBAAiB,EACjB,WAAW,EACX,mBAAmB,EACnB,KAAK,EAAE,aAAa,EACpB,eAAe,EAAE,uBAAuB,EACxC,YAAY,EAAE,oBAAoB,EAClC,GAAG,cAAc,EACjB,GAAG,OAAO,CAAC;IACZ,IAAI,OAAiC,CAAC;IACtC,MAAM,UAAU,GAAG,GAAiB,EAAE,CAAC;QACtC,IAAI,CAAC,OAAO;YAAE,MAAM,IAAI,KAAK,CAAC,iEAAiE,CAAC,CAAC;QACjG,OAAO,OAAO,CAAC;IAAA,CACf,CAAC;IACF,IAAI,KAAK,GAAG,aAAa,CAAC;IAC1B,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;QACzB,MAAM,QAAQ,GAAG,MAAM,OAAO,CAAC,OAAO,CAAC,WAAW,EAAE,CAAC;QACrD,MAAM,WAAW,GAAG,EAAE,GAAG,EAAiC,CAAC;QAC3D,KAAK,GAAG;YACP,4BAA4B,CAAC,cAAc,EAAwB,EAAE,WAAW,EAAE;gBACjF,aAAa,EAAE,gCAAgC,CAAC,OAAO;gBACvD,gBAAgB,EAAE,gCAAgC,CAAC,UAAU;aAC7D,CAAC;YACF,4BAA4B,CAC3B,cAAc,CAAuB;gBACpC,aAAa,EAAE,iBAAiB;gBAChC,OAAO,EAAE,KAAK,EAAE,SAAS,EAAE,EAAE,CAAC;oBAC7B,MAAM,cAAc,GAAG,UAAU,EAAE,CAAC;oBACpC,MAAM,CAAC,KAAK,EAAE,aAAa,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC;wBAChD,cAAc,CAAC,QAAQ,EAAE;wBACzB,cAAc,CAAC,gBAAgB,EAAE;qBACjC,CAAC,CAAC;oBACH,kBAAkB,CAAC,sBAAsB,EAAE,QAAQ,CAAC,EAAE,EAAE,SAAS,CAAC,GAAG,CAAC,CAAC;oBACvE,kBAAkB,CAAC,wBAAwB,EAAE,WAAW,IAAI,EAAE,EAAE,SAAS,CAAC,GAAG,CAAC,CAAC;oBAC/E,kBAAkB,CAAC,oBAAoB,EAAE,KAAK,CAAC,QAAQ,EAAE,SAAS,CAAC,GAAG,CAAC,CAAC;oBACxE,kBAAkB,CAAC,iBAAiB,EAAE,KAAK,CAAC,EAAE,EAAE,SAAS,CAAC,GAAG,CAAC,CAAC;oBAC/D,kBAAkB,CAAC,2BAA2B,EAAE,aAAa,EAAE,SAAS,CAAC,GAAG,CAAC,CAAC;gBAAA,CAC9E;aACD,CAAC,EACF,WAAW,EACX;gBACC,aAAa,EAAE,gCAAgC,CAAC,OAAO;gBACvD,gBAAgB,EAAE,gCAAgC,CAAC,UAAU;aAC7D,CACD;YACD,4BAA4B,CAAC,cAAc,EAAwB,EAAE,WAAW,EAAE;gBACjF,aAAa,EAAE,gCAAgC,CAAC,OAAO;gBACvD,gBAAgB,EAAE,gCAAgC,CAAC,UAAU;aAC7D,CAAC;YACF,4BAA4B,CAAC,eAAe,EAAwB,EAAE,WAAW,EAAE;gBAClF,aAAa,EAAE,iCAAiC,CAAC,OAAO;gBACxD,gBAAgB,EAAE,iCAAiC,CAAC,UAAU;aAC9D,CAAC;SACF,CAAC;IACH,CAAC;IACD,MAAM,eAAe,GAAG,CAAC,GAAG,CAAC,uBAAuB,IAAI,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IACzF,MAAM,YAAY,GACjB,oBAAoB;QACpB,CAAC,KAAK,IAAI,EAAE,CAAC;YACZ,MAAM,cAAc,GAAG,UAAU,EAAE,CAAC;YACpC,MAAM,CAAC,YAAY,EAAE,sBAAsB,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC;gBAChE,cAAc,CAAC,QAAQ,EAAE;gBACzB,cAAc,CAAC,cAAc,EAAE;aAC/B,CAAC,CAAC;YACH,OAAO,mCAAmC,CAAC;gBAC1C,GAAG,EAAE,GAAG,CAAC,GAAG;gBACZ,KAAK,EAAE,YAAY;gBACnB,eAAe,EAAE,sBAAsB;gBACvC,mBAAmB;aACnB,CAAC,CAAC;QAAA,CACH,CAAC,CAAC;IACJ,MAAM,OAAO,GAAG,MAAM,YAAY,CAAC,MAAM,CAAC;QACzC,GAAG,cAAc;QACjB,KAAK;QACL,eAAe;QACf,YAAY;KACZ,CAAC,CAAC;IACH,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC;IAC1B,OAAO,OAAO,CAAC;AAAA,CACf","sourcesContent":["import {\n\tAgentHarness,\n\ttype AgentHarnessOptions,\n\ttype AgentHarnessTool,\n\tcreateBashTool,\n\tcreateEditTool,\n\tcreateReadTool,\n\tcreateWriteTool,\n\ttype ExecutionEnv,\n\ttype ExecutionToolContext,\n\ttype HarnessTool,\n} from \"apex-code-agent-core\";\nimport type { Static, TSchema } from \"typebox\";\nimport { setApexEnvironment } from \"../core/environment.ts\";\nimport { type BuildSystemPromptOptions, buildSystemPrompt } from \"../core/system-prompt.ts\";\nimport { bashToolSystemPromptContribution } from \"../core/tools/bash.ts\";\nimport { editToolSystemPromptContribution } from \"../core/tools/edit.ts\";\nimport { readToolSystemPromptContribution } from \"../core/tools/read.ts\";\nimport { writeToolSystemPromptContribution } from \"../core/tools/write.ts\";\n\nexport interface CodingAgentHarnessTool extends HarnessTool {\n\tpromptSnippet?: string;\n\tpromptGuidelines?: readonly string[];\n}\n\nfunction createCodingAgentHarnessTool<TParameters extends TSchema, TDetails>(\n\ttool: AgentHarnessTool<ExecutionToolContext, TParameters, TDetails>,\n\tcontext: ExecutionToolContext,\n\tprompt: Required<Pick<CodingAgentHarnessTool, \"promptSnippet\" | \"promptGuidelines\">>,\n): CodingAgentHarnessTool {\n\treturn {\n\t\t...tool,\n\t\t...prompt,\n\t\texecute: (toolCallId, params, signal, onUpdate) =>\n\t\t\ttool.execute(toolCallId, params as Static<TParameters>, signal, onUpdate, context),\n\t};\n}\n\nexport interface CreateCodingAgentHarnessOptions extends Omit<AgentHarnessOptions, \"toolContext\" | \"tools\"> {\n\tenv: ExecutionEnv;\n\tbashCommandPrefix?: string;\n\t/** Path to the JSONL session file exposed to default bash commands as APEX_CODE_SESSION_FILE. */\n\tsessionFile?: string;\n\ttools?: CodingAgentHarnessTool[];\n\tsystemPromptOptions?: Omit<BuildSystemPromptOptions, \"cwd\" | \"promptGuidelines\" | \"selectedTools\" | \"toolSnippets\">;\n}\n\nexport interface BuildCodingAgentHarnessSystemPromptOptions {\n\tcwd: string;\n\ttools: readonly CodingAgentHarnessTool[];\n\tactiveToolNames: readonly string[];\n\tsystemPromptOptions?: CreateCodingAgentHarnessOptions[\"systemPromptOptions\"];\n}\n\nexport function buildCodingAgentHarnessSystemPrompt(options: BuildCodingAgentHarnessSystemPromptOptions): string {\n\tconst activeTools = options.activeToolNames.flatMap((name) => {\n\t\tconst tool = options.tools.find((candidate) => candidate.name === name);\n\t\treturn tool ? [tool] : [];\n\t});\n\tconst toolSnippets = Object.fromEntries(\n\t\tactiveTools.flatMap((tool) => {\n\t\t\tconst promptSnippet = tool.promptSnippet\n\t\t\t\t?.replace(/[\\r\\n]+/g, \" \")\n\t\t\t\t.replace(/\\s+/g, \" \")\n\t\t\t\t.trim();\n\t\t\treturn promptSnippet ? [[tool.name, promptSnippet]] : [];\n\t\t}),\n\t);\n\tconst promptGuidelines = activeTools.flatMap((tool) => tool.promptGuidelines ?? []);\n\treturn buildSystemPrompt({\n\t\t...options.systemPromptOptions,\n\t\tcwd: options.cwd,\n\t\tselectedTools: activeTools.map((tool) => tool.name),\n\t\ttoolSnippets,\n\t\tpromptGuidelines,\n\t});\n}\n\nexport async function createCodingAgentHarness(options: CreateCodingAgentHarnessOptions) {\n\tconst {\n\t\tenv,\n\t\tbashCommandPrefix,\n\t\tsessionFile,\n\t\tsystemPromptOptions,\n\t\ttools: providedTools,\n\t\tactiveToolNames: providedActiveToolNames,\n\t\tsystemPrompt: providedSystemPrompt,\n\t\t...harnessOptions\n\t} = options;\n\tlet harness: AgentHarness | undefined;\n\tconst getHarness = (): AgentHarness => {\n\t\tif (!harness) throw new Error(\"Coding-agent Harness callback ran before Harness initialization\");\n\t\treturn harness;\n\t};\n\tlet tools = providedTools;\n\tif (tools === undefined) {\n\t\tconst metadata = await options.session.getMetadata();\n\t\tconst toolContext = { env } satisfies ExecutionToolContext;\n\t\ttools = [\n\t\t\tcreateCodingAgentHarnessTool(createReadTool<ExecutionToolContext>(), toolContext, {\n\t\t\t\tpromptSnippet: readToolSystemPromptContribution.snippet,\n\t\t\t\tpromptGuidelines: readToolSystemPromptContribution.guidelines,\n\t\t\t}),\n\t\t\tcreateCodingAgentHarnessTool(\n\t\t\t\tcreateBashTool<ExecutionToolContext>({\n\t\t\t\t\tcommandPrefix: bashCommandPrefix,\n\t\t\t\t\tprepare: async (execution) => {\n\t\t\t\t\t\tconst currentHarness = getHarness();\n\t\t\t\t\t\tconst [model, thinkingLevel] = await Promise.all([\n\t\t\t\t\t\t\tcurrentHarness.getModel(),\n\t\t\t\t\t\t\tcurrentHarness.getThinkingLevel(),\n\t\t\t\t\t\t]);\n\t\t\t\t\t\tsetApexEnvironment(\"APEX_CODE_SESSION_ID\", metadata.id, execution.env);\n\t\t\t\t\t\tsetApexEnvironment(\"APEX_CODE_SESSION_FILE\", sessionFile ?? \"\", execution.env);\n\t\t\t\t\t\tsetApexEnvironment(\"APEX_CODE_PROVIDER\", model.provider, execution.env);\n\t\t\t\t\t\tsetApexEnvironment(\"APEX_CODE_MODEL\", model.id, execution.env);\n\t\t\t\t\t\tsetApexEnvironment(\"APEX_CODE_REASONING_LEVEL\", thinkingLevel, execution.env);\n\t\t\t\t\t},\n\t\t\t\t}),\n\t\t\t\ttoolContext,\n\t\t\t\t{\n\t\t\t\t\tpromptSnippet: bashToolSystemPromptContribution.snippet,\n\t\t\t\t\tpromptGuidelines: bashToolSystemPromptContribution.guidelines,\n\t\t\t\t},\n\t\t\t),\n\t\t\tcreateCodingAgentHarnessTool(createEditTool<ExecutionToolContext>(), toolContext, {\n\t\t\t\tpromptSnippet: editToolSystemPromptContribution.snippet,\n\t\t\t\tpromptGuidelines: editToolSystemPromptContribution.guidelines,\n\t\t\t}),\n\t\t\tcreateCodingAgentHarnessTool(createWriteTool<ExecutionToolContext>(), toolContext, {\n\t\t\t\tpromptSnippet: writeToolSystemPromptContribution.snippet,\n\t\t\t\tpromptGuidelines: writeToolSystemPromptContribution.guidelines,\n\t\t\t}),\n\t\t];\n\t}\n\tconst activeToolNames = [...(providedActiveToolNames ?? tools.map((tool) => tool.name))];\n\tconst systemPrompt =\n\t\tprovidedSystemPrompt ??\n\t\t(async () => {\n\t\t\tconst currentHarness = getHarness();\n\t\t\tconst [currentTools, currentActiveToolNames] = await Promise.all([\n\t\t\t\tcurrentHarness.getTools(),\n\t\t\t\tcurrentHarness.getActiveTools(),\n\t\t\t]);\n\t\t\treturn buildCodingAgentHarnessSystemPrompt({\n\t\t\t\tcwd: env.cwd,\n\t\t\t\ttools: currentTools,\n\t\t\t\tactiveToolNames: currentActiveToolNames,\n\t\t\t\tsystemPromptOptions,\n\t\t\t});\n\t\t});\n\tconst created = await AgentHarness.create({\n\t\t...harnessOptions,\n\t\ttools,\n\t\tactiveToolNames,\n\t\tsystemPrompt,\n\t});\n\tharness = created.harness;\n\treturn created;\n}\n"]}

package/dist/testing/replay/metrics.d.ts

@@ -0,0 +1,24 @@
+import type { AssistantMessage, Tool } from "@earendil-works/pi-ai";
+/** The stable, provider-independent measurements emitted by an offline replay. */
+export interface ReplayMetrics {
+    contextTokensByTurn: number[];
+    /** Estimated static provider input: the production system prompt plus tool schemas. */
+    systemPromptTokens: number;
+    cacheHitRate: number;
+    toolCallsByName: Record<string, number>;
+    /** Deterministic placeholder until recordings carry request durations. */
+    wallTimeMs: number;
+    costUsd: number;
+    turnsCompleted: number;
+}
+export interface ReplayMetricsInput {
+    systemPrompt: string;
+    tools: readonly Tool[];
+    recordedResponses: readonly AssistantMessage[];
+    replayedResponses: readonly AssistantMessage[];
+    contextTokensByTurn: readonly number[];
+    turnsCompleted: number;
+}
+/** Build metrics from stable provider-boundary and replay output data. */
+export declare function buildReplayMetrics(input: ReplayMetricsInput): ReplayMetrics;
+//# sourceMappingURL=metrics.d.ts.map

package/dist/testing/replay/metrics.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"metrics.d.ts","sourceRoot":"","sources":["../../../src/testing/replay/metrics.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,MAAM,uBAAuB,CAAC;AAEpE,kFAAkF;AAClF,MAAM,WAAW,aAAa;IAC7B,mBAAmB,EAAE,MAAM,EAAE,CAAC;IAC9B,uFAAuF;IACvF,kBAAkB,EAAE,MAAM,CAAC;IAC3B,YAAY,EAAE,MAAM,CAAC;IACrB,eAAe,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACxC,0EAA0E;IAC1E,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC;IAChB,cAAc,EAAE,MAAM,CAAC;CACvB;AAED,MAAM,WAAW,kBAAkB;IAClC,YAAY,EAAE,MAAM,CAAC;IACrB,KAAK,EAAE,SAAS,IAAI,EAAE,CAAC;IACvB,iBAAiB,EAAE,SAAS,gBAAgB,EAAE,CAAC;IAC/C,iBAAiB,EAAE,SAAS,gBAAgB,EAAE,CAAC;IAC/C,mBAAmB,EAAE,SAAS,MAAM,EAAE,CAAC;IACvC,cAAc,EAAE,MAAM,CAAC;CACvB;AAED,0EAA0E;AAC1E,wBAAgB,kBAAkB,CAAC,KAAK,EAAE,kBAAkB,GAAG,aAAa,CA+B3E","sourcesContent":["import type { AssistantMessage, Tool } from \"@earendil-works/pi-ai\";\n\n/** The stable, provider-independent measurements emitted by an offline replay. */\nexport interface ReplayMetrics {\n\tcontextTokensByTurn: number[];\n\t/** Estimated static provider input: the production system prompt plus tool schemas. */\n\tsystemPromptTokens: number;\n\tcacheHitRate: number;\n\ttoolCallsByName: Record<string, number>;\n\t/** Deterministic placeholder until recordings carry request durations. */\n\twallTimeMs: number;\n\tcostUsd: number;\n\tturnsCompleted: number;\n}\n\nexport interface ReplayMetricsInput {\n\tsystemPrompt: string;\n\ttools: readonly Tool[];\n\trecordedResponses: readonly AssistantMessage[];\n\treplayedResponses: readonly AssistantMessage[];\n\tcontextTokensByTurn: readonly number[];\n\tturnsCompleted: number;\n}\n\n/** Build metrics from stable provider-boundary and replay output data. */\nexport function buildReplayMetrics(input: ReplayMetricsInput): ReplayMetrics {\n\tlet inputTokens = 0;\n\tlet cacheRead = 0;\n\tlet cacheWrite = 0;\n\tlet costUsd = 0;\n\tfor (const response of input.recordedResponses) {\n\t\tinputTokens += response.usage.input;\n\t\tcacheRead += response.usage.cacheRead;\n\t\tcacheWrite += response.usage.cacheWrite;\n\t\tcostUsd += response.usage.cost.total;\n\t}\n\tconst toolCallCounts = new Map<string, number>();\n\tfor (const response of input.replayedResponses) {\n\t\tfor (const part of response.content) {\n\t\t\tif (part.type === \"toolCall\") toolCallCounts.set(part.name, (toolCallCounts.get(part.name) ?? 0) + 1);\n\t\t}\n\t}\n\tconst toolCallsByName = Object.fromEntries([...toolCallCounts.entries()].sort(([a], [b]) => a.localeCompare(b)));\n\tconst staticProviderInput = `${input.systemPrompt}\n${JSON.stringify(input.tools)}`;\n\tconst promptTokens = inputTokens + cacheRead + cacheWrite;\n\treturn {\n\t\tcontextTokensByTurn: [...input.contextTokensByTurn],\n\t\tsystemPromptTokens: Math.ceil(staticProviderInput.length / 4),\n\t\tcacheHitRate: promptTokens === 0 ? 0 : cacheRead / promptTokens,\n\t\ttoolCallsByName,\n\t\t// A live timer would make the equality gate flaky; recordings do not yet carry durations.\n\t\twallTimeMs: 0,\n\t\tcostUsd,\n\t\tturnsCompleted: input.turnsCompleted,\n\t};\n}\n"]}

package/dist/testing/replay/metrics.js

@@ -0,0 +1,35 @@
+/** Build metrics from stable provider-boundary and replay output data. */
+export function buildReplayMetrics(input) {
+    let inputTokens = 0;
+    let cacheRead = 0;
+    let cacheWrite = 0;
+    let costUsd = 0;
+    for (const response of input.recordedResponses) {
+        inputTokens += response.usage.input;
+        cacheRead += response.usage.cacheRead;
+        cacheWrite += response.usage.cacheWrite;
+        costUsd += response.usage.cost.total;
+    }
+    const toolCallCounts = new Map();
+    for (const response of input.replayedResponses) {
+        for (const part of response.content) {
+            if (part.type === "toolCall")
+                toolCallCounts.set(part.name, (toolCallCounts.get(part.name) ?? 0) + 1);
+        }
+    }
+    const toolCallsByName = Object.fromEntries([...toolCallCounts.entries()].sort(([a], [b]) => a.localeCompare(b)));
+    const staticProviderInput = `${input.systemPrompt}
+${JSON.stringify(input.tools)}`;
+    const promptTokens = inputTokens + cacheRead + cacheWrite;
+    return {
+        contextTokensByTurn: [...input.contextTokensByTurn],
+        systemPromptTokens: Math.ceil(staticProviderInput.length / 4),
+        cacheHitRate: promptTokens === 0 ? 0 : cacheRead / promptTokens,
+        toolCallsByName,
+        // A live timer would make the equality gate flaky; recordings do not yet carry durations.
+        wallTimeMs: 0,
+        costUsd,
+        turnsCompleted: input.turnsCompleted,
+    };
+}
+//# sourceMappingURL=metrics.js.map

package/dist/testing/replay/metrics.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"metrics.js","sourceRoot":"","sources":["../../../src/testing/replay/metrics.ts"],"names":[],"mappings":"AAwBA,0EAA0E;AAC1E,MAAM,UAAU,kBAAkB,CAAC,KAAyB,EAAiB;IAC5E,IAAI,WAAW,GAAG,CAAC,CAAC;IACpB,IAAI,SAAS,GAAG,CAAC,CAAC;IAClB,IAAI,UAAU,GAAG,CAAC,CAAC;IACnB,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,KAAK,MAAM,QAAQ,IAAI,KAAK,CAAC,iBAAiB,EAAE,CAAC;QAChD,WAAW,IAAI,QAAQ,CAAC,KAAK,CAAC,KAAK,CAAC;QACpC,SAAS,IAAI,QAAQ,CAAC,KAAK,CAAC,SAAS,CAAC;QACtC,UAAU,IAAI,QAAQ,CAAC,KAAK,CAAC,UAAU,CAAC;QACxC,OAAO,IAAI,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC;IACtC,CAAC;IACD,MAAM,cAAc,GAAG,IAAI,GAAG,EAAkB,CAAC;IACjD,KAAK,MAAM,QAAQ,IAAI,KAAK,CAAC,iBAAiB,EAAE,CAAC;QAChD,KAAK,MAAM,IAAI,IAAI,QAAQ,CAAC,OAAO,EAAE,CAAC;YACrC,IAAI,IAAI,CAAC,IAAI,KAAK,UAAU;gBAAE,cAAc,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,cAAc,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QACvG,CAAC;IACF,CAAC;IACD,MAAM,eAAe,GAAG,MAAM,CAAC,WAAW,CAAC,CAAC,GAAG,cAAc,CAAC,OAAO,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,aAAa,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACjH,MAAM,mBAAmB,GAAG,GAAG,KAAK,CAAC,YAAY;EAChD,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC;IAC/B,MAAM,YAAY,GAAG,WAAW,GAAG,SAAS,GAAG,UAAU,CAAC;IAC1D,OAAO;QACN,mBAAmB,EAAE,CAAC,GAAG,KAAK,CAAC,mBAAmB,CAAC;QACnD,kBAAkB,EAAE,IAAI,CAAC,IAAI,CAAC,mBAAmB,CAAC,MAAM,GAAG,CAAC,CAAC;QAC7D,YAAY,EAAE,YAAY,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,GAAG,YAAY;QAC/D,eAAe;QACf,0FAA0F;QAC1F,UAAU,EAAE,CAAC;QACb,OAAO;QACP,cAAc,EAAE,KAAK,CAAC,cAAc;KACpC,CAAC;AAAA,CACF","sourcesContent":["import type { AssistantMessage, Tool } from \"@earendil-works/pi-ai\";\n\n/** The stable, provider-independent measurements emitted by an offline replay. */\nexport interface ReplayMetrics {\n\tcontextTokensByTurn: number[];\n\t/** Estimated static provider input: the production system prompt plus tool schemas. */\n\tsystemPromptTokens: number;\n\tcacheHitRate: number;\n\ttoolCallsByName: Record<string, number>;\n\t/** Deterministic placeholder until recordings carry request durations. */\n\twallTimeMs: number;\n\tcostUsd: number;\n\tturnsCompleted: number;\n}\n\nexport interface ReplayMetricsInput {\n\tsystemPrompt: string;\n\ttools: readonly Tool[];\n\trecordedResponses: readonly AssistantMessage[];\n\treplayedResponses: readonly AssistantMessage[];\n\tcontextTokensByTurn: readonly number[];\n\tturnsCompleted: number;\n}\n\n/** Build metrics from stable provider-boundary and replay output data. */\nexport function buildReplayMetrics(input: ReplayMetricsInput): ReplayMetrics {\n\tlet inputTokens = 0;\n\tlet cacheRead = 0;\n\tlet cacheWrite = 0;\n\tlet costUsd = 0;\n\tfor (const response of input.recordedResponses) {\n\t\tinputTokens += response.usage.input;\n\t\tcacheRead += response.usage.cacheRead;\n\t\tcacheWrite += response.usage.cacheWrite;\n\t\tcostUsd += response.usage.cost.total;\n\t}\n\tconst toolCallCounts = new Map<string, number>();\n\tfor (const response of input.replayedResponses) {\n\t\tfor (const part of response.content) {\n\t\t\tif (part.type === \"toolCall\") toolCallCounts.set(part.name, (toolCallCounts.get(part.name) ?? 0) + 1);\n\t\t}\n\t}\n\tconst toolCallsByName = Object.fromEntries([...toolCallCounts.entries()].sort(([a], [b]) => a.localeCompare(b)));\n\tconst staticProviderInput = `${input.systemPrompt}\n${JSON.stringify(input.tools)}`;\n\tconst promptTokens = inputTokens + cacheRead + cacheWrite;\n\treturn {\n\t\tcontextTokensByTurn: [...input.contextTokensByTurn],\n\t\tsystemPromptTokens: Math.ceil(staticProviderInput.length / 4),\n\t\tcacheHitRate: promptTokens === 0 ? 0 : cacheRead / promptTokens,\n\t\ttoolCallsByName,\n\t\t// A live timer would make the equality gate flaky; recordings do not yet carry durations.\n\t\twallTimeMs: 0,\n\t\tcostUsd,\n\t\tturnsCompleted: input.turnsCompleted,\n\t};\n}\n"]}

package/dist/testing/replay/recorded-provider.d.ts

@@ -0,0 +1,10 @@
+import { type AssistantMessage, type Model } from "@earendil-works/pi-ai";
+import type { ModelRuntime } from "../../core/model-runtime.ts";
+export interface RecordedProvider {
+    readonly requestCount: number;
+    readonly contexts: readonly import("@earendil-works/pi-ai").Context[];
+    getModel(modelId: string): Model<string>;
+    assertExhausted(): void;
+}
+export declare function registerRecordedProvider(runtime: ModelRuntime, responses: readonly AssistantMessage[]): RecordedProvider;
+//# sourceMappingURL=recorded-provider.d.ts.map

package/dist/testing/replay/recorded-provider.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"recorded-provider.d.ts","sourceRoot":"","sources":["../../../src/testing/replay/recorded-provider.ts"],"names":[],"mappings":"AAAA,OAAO,EACN,KAAK,gBAAgB,EAGrB,KAAK,KAAK,EACV,MAAM,uBAAuB,CAAC;AAC/B,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,6BAA6B,CAAC;AAMhE,MAAM,WAAW,gBAAgB;IAChC,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,QAAQ,EAAE,SAAS,OAAO,uBAAuB,EAAE,OAAO,EAAE,CAAC;IACtE,QAAQ,CAAC,OAAO,EAAE,MAAM,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC;IACzC,eAAe,IAAI,IAAI,CAAC;CACxB;AA8DD,wBAAgB,wBAAwB,CACvC,OAAO,EAAE,YAAY,EACrB,SAAS,EAAE,SAAS,gBAAgB,EAAE,GACpC,gBAAgB,CA2ClB","sourcesContent":["import {\n\ttype AssistantMessage,\n\ttype AssistantMessageEventStream,\n\tcreateAssistantMessageEventStream,\n\ttype Model,\n} from \"@earendil-works/pi-ai\";\nimport type { ModelRuntime } from \"../../core/model-runtime.ts\";\n\nconst RECORDED_PROVIDER_ID = \"recorded\";\nconst RECORDED_API = \"apex-recorded\";\nconst ZERO_COST = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 };\n\nexport interface RecordedProvider {\n\treadonly requestCount: number;\n\treadonly contexts: readonly import(\"@earendil-works/pi-ai\").Context[];\n\tgetModel(modelId: string): Model<string>;\n\tassertExhausted(): void;\n}\n\nfunction completedStream(message: AssistantMessage): AssistantMessageEventStream {\n\tconst stream = createAssistantMessageEventStream();\n\tqueueMicrotask(() => {\n\t\tconst cloned = structuredClone(message);\n\t\tstream.push({ type: \"start\", partial: { ...cloned, content: [] } });\n\t\tif (cloned.stopReason === \"error\" || cloned.stopReason === \"aborted\") {\n\t\t\tstream.push({ type: \"error\", reason: cloned.stopReason, error: cloned });\n\t\t} else if (cloned.stopReason === \"pending\") {\n\t\t\tconst error = errorMessage(\"Recorded response cannot have a pending stop reason\", cloned.model);\n\t\t\tstream.push({ type: \"error\", reason: \"error\", error });\n\t\t\tstream.end(error);\n\t\t\treturn;\n\t\t} else {\n\t\t\tstream.push({ type: \"done\", reason: cloned.stopReason, message: cloned });\n\t\t}\n\t\tstream.end(cloned);\n\t});\n\treturn stream;\n}\n\nfunction errorMessage(message: string, modelId: string): AssistantMessage {\n\treturn {\n\t\trole: \"assistant\",\n\t\tcontent: [],\n\t\tapi: RECORDED_API,\n\t\tprovider: RECORDED_PROVIDER_ID,\n\t\tmodel: modelId,\n\t\tusage: {\n\t\t\tinput: 0,\n\t\t\toutput: 0,\n\t\t\tcacheRead: 0,\n\t\t\tcacheWrite: 0,\n\t\t\ttotalTokens: 0,\n\t\t\tcost: { ...ZERO_COST, total: 0 },\n\t\t},\n\t\tstopReason: \"error\",\n\t\terrorMessage: message,\n\t\ttimestamp: 0,\n\t};\n}\n\nfunction failedStream(message: string, modelId: string): AssistantMessageEventStream {\n\treturn completedStream(errorMessage(message, modelId));\n}\n\nfunction recordedModel(modelId: string): Model<string> {\n\treturn {\n\t\tid: modelId,\n\t\tname: modelId,\n\t\tapi: RECORDED_API,\n\t\tprovider: RECORDED_PROVIDER_ID,\n\t\tbaseUrl: \"replay://recorded\",\n\t\treasoning: false,\n\t\tinput: [\"text\", \"image\"],\n\t\tcost: ZERO_COST,\n\t\tcontextWindow: 1_000_000,\n\t\tmaxTokens: 100_000,\n\t};\n}\n\nexport function registerRecordedProvider(\n\truntime: ModelRuntime,\n\tresponses: readonly AssistantMessage[],\n): RecordedProvider {\n\tconst pending = responses.map((message) => structuredClone(message));\n\tconst modelIds = [...new Set(pending.map((message) => message.model))];\n\tconst models = new Map(modelIds.map((modelId) => [modelId, recordedModel(modelId)]));\n\tlet requestCount = 0;\n\tconst contexts: import(\"@earendil-works/pi-ai\").Context[] = [];\n\n\truntime.registerProvider(RECORDED_PROVIDER_ID, {\n\t\tname: \"Recorded replay provider\",\n\t\tapi: RECORDED_API,\n\t\tapiKey: \"offline-replay\",\n\t\tmodels: [...models.values()],\n\t\tstreamSimple: (model, context) => {\n\t\t\trequestCount++;\n\t\t\tcontexts.push({\n\t\t\t\tsystemPrompt: context.systemPrompt,\n\t\t\t\tmessages: structuredClone(context.messages),\n\t\t\t});\n\t\t\tconst response = pending.shift();\n\t\t\tif (!response) return failedStream(`Recorded replay exhausted before request ${requestCount}`, model.id);\n\t\t\tif (response.model !== model.id) {\n\t\t\t\treturn failedStream(`Recorded model mismatch: expected ${response.model}, received ${model.id}`, model.id);\n\t\t\t}\n\t\t\treturn completedStream(response);\n\t\t},\n\t});\n\n\treturn {\n\t\tget requestCount() {\n\t\t\treturn requestCount;\n\t\t},\n\t\tget contexts() {\n\t\t\treturn contexts;\n\t\t},\n\t\tgetModel(modelId) {\n\t\t\tconst model = models.get(modelId);\n\t\t\tif (!model) throw new Error(`Unknown recorded model: ${modelId}`);\n\t\t\treturn model;\n\t\t},\n\t\tassertExhausted() {\n\t\t\tif (pending.length > 0) throw new Error(`Recorded replay left ${pending.length} response(s) unused`);\n\t\t},\n\t};\n}\n"]}

package/dist/testing/replay/recorded-provider.js

@@ -0,0 +1,108 @@
+import { createAssistantMessageEventStream, } from "@earendil-works/pi-ai";
+const RECORDED_PROVIDER_ID = "recorded";
+const RECORDED_API = "apex-recorded";
+const ZERO_COST = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 };
+function completedStream(message) {
+    const stream = createAssistantMessageEventStream();
+    queueMicrotask(() => {
+        const cloned = structuredClone(message);
+        stream.push({ type: "start", partial: { ...cloned, content: [] } });
+        if (cloned.stopReason === "error" || cloned.stopReason === "aborted") {
+            stream.push({ type: "error", reason: cloned.stopReason, error: cloned });
+        }
+        else if (cloned.stopReason === "pending") {
+            const error = errorMessage("Recorded response cannot have a pending stop reason", cloned.model);
+            stream.push({ type: "error", reason: "error", error });
+            stream.end(error);
+            return;
+        }
+        else {
+            stream.push({ type: "done", reason: cloned.stopReason, message: cloned });
+        }
+        stream.end(cloned);
+    });
+    return stream;
+}
+function errorMessage(message, modelId) {
+    return {
+        role: "assistant",
+        content: [],
+        api: RECORDED_API,
+        provider: RECORDED_PROVIDER_ID,
+        model: modelId,
+        usage: {
+            input: 0,
+            output: 0,
+            cacheRead: 0,
+            cacheWrite: 0,
+            totalTokens: 0,
+            cost: { ...ZERO_COST, total: 0 },
+        },
+        stopReason: "error",
+        errorMessage: message,
+        timestamp: 0,
+    };
+}
+function failedStream(message, modelId) {
+    return completedStream(errorMessage(message, modelId));
+}
+function recordedModel(modelId) {
+    return {
+        id: modelId,
+        name: modelId,
+        api: RECORDED_API,
+        provider: RECORDED_PROVIDER_ID,
+        baseUrl: "replay://recorded",
+        reasoning: false,
+        input: ["text", "image"],
+        cost: ZERO_COST,
+        contextWindow: 1_000_000,
+        maxTokens: 100_000,
+    };
+}
+export function registerRecordedProvider(runtime, responses) {
+    const pending = responses.map((message) => structuredClone(message));
+    const modelIds = [...new Set(pending.map((message) => message.model))];
+    const models = new Map(modelIds.map((modelId) => [modelId, recordedModel(modelId)]));
+    let requestCount = 0;
+    const contexts = [];
+    runtime.registerProvider(RECORDED_PROVIDER_ID, {
+        name: "Recorded replay provider",
+        api: RECORDED_API,
+        apiKey: "offline-replay",
+        models: [...models.values()],
+        streamSimple: (model, context) => {
+            requestCount++;
+            contexts.push({
+                systemPrompt: context.systemPrompt,
+                messages: structuredClone(context.messages),
+            });
+            const response = pending.shift();
+            if (!response)
+                return failedStream(`Recorded replay exhausted before request ${requestCount}`, model.id);
+            if (response.model !== model.id) {
+                return failedStream(`Recorded model mismatch: expected ${response.model}, received ${model.id}`, model.id);
+            }
+            return completedStream(response);
+        },
+    });
+    return {
+        get requestCount() {
+            return requestCount;
+        },
+        get contexts() {
+            return contexts;
+        },
+        getModel(modelId) {
+            const model = models.get(modelId);
+            if (!model)
+                throw new Error(`Unknown recorded model: ${modelId}`);
+            return model;
+        },
+        assertExhausted() {
+            if (pending.length > 0)
+                throw new Error(`Recorded replay left ${pending.length} response(s) unused`);
+        },
+    };
+}
+//# sourceMappingURL=recorded-provider.js.map

package/dist/testing/replay/recorded-provider.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"recorded-provider.js","sourceRoot":"","sources":["../../../src/testing/replay/recorded-provider.ts"],"names":[],"mappings":"AAAA,OAAO,EAGN,iCAAiC,GAEjC,MAAM,uBAAuB,CAAC;AAG/B,MAAM,oBAAoB,GAAG,UAAU,CAAC;AACxC,MAAM,YAAY,GAAG,eAAe,CAAC;AACrC,MAAM,SAAS,GAAG,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,CAAC;AASvE,SAAS,eAAe,CAAC,OAAyB,EAA+B;IAChF,MAAM,MAAM,GAAG,iCAAiC,EAAE,CAAC;IACnD,cAAc,CAAC,GAAG,EAAE,CAAC;QACpB,MAAM,MAAM,GAAG,eAAe,CAAC,OAAO,CAAC,CAAC;QACxC,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,EAAE,GAAG,MAAM,EAAE,OAAO,EAAE,EAAE,EAAE,EAAE,CAAC,CAAC;QACpE,IAAI,MAAM,CAAC,UAAU,KAAK,OAAO,IAAI,MAAM,CAAC,UAAU,KAAK,SAAS,EAAE,CAAC;YACtE,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,MAAM,EAAE,MAAM,CAAC,UAAU,EAAE,KAAK,EAAE,MAAM,EAAE,CAAC,CAAC;QAC1E,CAAC;aAAM,IAAI,MAAM,CAAC,UAAU,KAAK,SAAS,EAAE,CAAC;YAC5C,MAAM,KAAK,GAAG,YAAY,CAAC,qDAAqD,EAAE,MAAM,CAAC,KAAK,CAAC,CAAC;YAChG,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC,CAAC;YACvD,MAAM,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC;YAClB,OAAO;QACR,CAAC;aAAM,CAAC;YACP,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC,UAAU,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;QAC3E,CAAC;QACD,MAAM,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC;IAAA,CACnB,CAAC,CAAC;IACH,OAAO,MAAM,CAAC;AAAA,CACd;AAED,SAAS,YAAY,CAAC,OAAe,EAAE,OAAe,EAAoB;IACzE,OAAO;QACN,IAAI,EAAE,WAAW;QACjB,OAAO,EAAE,EAAE;QACX,GAAG,EAAE,YAAY;QACjB,QAAQ,EAAE,oBAAoB;QAC9B,KAAK,EAAE,OAAO;QACd,KAAK,EAAE;YACN,KAAK,EAAE,CAAC;YACR,MAAM,EAAE,CAAC;YACT,SAAS,EAAE,CAAC;YACZ,UAAU,EAAE,CAAC;YACb,WAAW,EAAE,CAAC;YACd,IAAI,EAAE,EAAE,GAAG,SAAS,EAAE,KAAK,EAAE,CAAC,EAAE;SAChC;QACD,UAAU,EAAE,OAAO;QACnB,YAAY,EAAE,OAAO;QACrB,SAAS,EAAE,CAAC;KACZ,CAAC;AAAA,CACF;AAED,SAAS,YAAY,CAAC,OAAe,EAAE,OAAe,EAA+B;IACpF,OAAO,eAAe,CAAC,YAAY,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC,CAAC;AAAA,CACvD;AAED,SAAS,aAAa,CAAC,OAAe,EAAiB;IACtD,OAAO;QACN,EAAE,EAAE,OAAO;QACX,IAAI,EAAE,OAAO;QACb,GAAG,EAAE,YAAY;QACjB,QAAQ,EAAE,oBAAoB;QAC9B,OAAO,EAAE,mBAAmB;QAC5B,SAAS,EAAE,KAAK;QAChB,KAAK,EAAE,CAAC,MAAM,EAAE,OAAO,CAAC;QACxB,IAAI,EAAE,SAAS;QACf,aAAa,EAAE,SAAS;QACxB,SAAS,EAAE,OAAO;KAClB,CAAC;AAAA,CACF;AAED,MAAM,UAAU,wBAAwB,CACvC,OAAqB,EACrB,SAAsC,EACnB;IACnB,MAAM,OAAO,GAAG,SAAS,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,eAAe,CAAC,OAAO,CAAC,CAAC,CAAC;IACrE,MAAM,QAAQ,GAAG,CAAC,GAAG,IAAI,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IACvE,MAAM,MAAM,GAAG,IAAI,GAAG,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,OAAO,EAAE,aAAa,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC;IACrF,IAAI,YAAY,GAAG,CAAC,CAAC;IACrB,MAAM,QAAQ,GAA8C,EAAE,CAAC;IAE/D,OAAO,CAAC,gBAAgB,CAAC,oBAAoB,EAAE;QAC9C,IAAI,EAAE,0BAA0B;QAChC,GAAG,EAAE,YAAY;QACjB,MAAM,EAAE,gBAAgB;QACxB,MAAM,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC;QAC5B,YAAY,EAAE,CAAC,KAAK,EAAE,OAAO,EAAE,EAAE,CAAC;YACjC,YAAY,EAAE,CAAC;YACf,QAAQ,CAAC,IAAI,CAAC;gBACb,YAAY,EAAE,OAAO,CAAC,YAAY;gBAClC,QAAQ,EAAE,eAAe,CAAC,OAAO,CAAC,QAAQ,CAAC;aAC3C,CAAC,CAAC;YACH,MAAM,QAAQ,GAAG,OAAO,CAAC,KAAK,EAAE,CAAC;YACjC,IAAI,CAAC,QAAQ;gBAAE,OAAO,YAAY,CAAC,4CAA4C,YAAY,EAAE,EAAE,KAAK,CAAC,EAAE,CAAC,CAAC;YACzG,IAAI,QAAQ,CAAC,KAAK,KAAK,KAAK,CAAC,EAAE,EAAE,CAAC;gBACjC,OAAO,YAAY,CAAC,qCAAqC,QAAQ,CAAC,KAAK,cAAc,KAAK,CAAC,EAAE,EAAE,EAAE,KAAK,CAAC,EAAE,CAAC,CAAC;YAC5G,CAAC;YACD,OAAO,eAAe,CAAC,QAAQ,CAAC,CAAC;QAAA,CACjC;KACD,CAAC,CAAC;IAEH,OAAO;QACN,IAAI,YAAY,GAAG;YAClB,OAAO,YAAY,CAAC;QAAA,CACpB;QACD,IAAI,QAAQ,GAAG;YACd,OAAO,QAAQ,CAAC;QAAA,CAChB;QACD,QAAQ,CAAC,OAAO,EAAE;YACjB,MAAM,KAAK,GAAG,MAAM,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC;YAClC,IAAI,CAAC,KAAK;gBAAE,MAAM,IAAI,KAAK,CAAC,2BAA2B,OAAO,EAAE,CAAC,CAAC;YAClE,OAAO,KAAK,CAAC;QAAA,CACb;QACD,eAAe,GAAG;YACjB,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC;gBAAE,MAAM,IAAI,KAAK,CAAC,wBAAwB,OAAO,CAAC,MAAM,qBAAqB,CAAC,CAAC;QAAA,CACrG;KACD,CAAC;AAAA,CACF","sourcesContent":["import {\n\ttype AssistantMessage,\n\ttype AssistantMessageEventStream,\n\tcreateAssistantMessageEventStream,\n\ttype Model,\n} from \"@earendil-works/pi-ai\";\nimport type { ModelRuntime } from \"../../core/model-runtime.ts\";\n\nconst RECORDED_PROVIDER_ID = \"recorded\";\nconst RECORDED_API = \"apex-recorded\";\nconst ZERO_COST = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 };\n\nexport interface RecordedProvider {\n\treadonly requestCount: number;\n\treadonly contexts: readonly import(\"@earendil-works/pi-ai\").Context[];\n\tgetModel(modelId: string): Model<string>;\n\tassertExhausted(): void;\n}\n\nfunction completedStream(message: AssistantMessage): AssistantMessageEventStream {\n\tconst stream = createAssistantMessageEventStream();\n\tqueueMicrotask(() => {\n\t\tconst cloned = structuredClone(message);\n\t\tstream.push({ type: \"start\", partial: { ...cloned, content: [] } });\n\t\tif (cloned.stopReason === \"error\" || cloned.stopReason === \"aborted\") {\n\t\t\tstream.push({ type: \"error\", reason: cloned.stopReason, error: cloned });\n\t\t} else if (cloned.stopReason === \"pending\") {\n\t\t\tconst error = errorMessage(\"Recorded response cannot have a pending stop reason\", cloned.model);\n\t\t\tstream.push({ type: \"error\", reason: \"error\", error });\n\t\t\tstream.end(error);\n\t\t\treturn;\n\t\t} else {\n\t\t\tstream.push({ type: \"done\", reason: cloned.stopReason, message: cloned });\n\t\t}\n\t\tstream.end(cloned);\n\t});\n\treturn stream;\n}\n\nfunction errorMessage(message: string, modelId: string): AssistantMessage {\n\treturn {\n\t\trole: \"assistant\",\n\t\tcontent: [],\n\t\tapi: RECORDED_API,\n\t\tprovider: RECORDED_PROVIDER_ID,\n\t\tmodel: modelId,\n\t\tusage: {\n\t\t\tinput: 0,\n\t\t\toutput: 0,\n\t\t\tcacheRead: 0,\n\t\t\tcacheWrite: 0,\n\t\t\ttotalTokens: 0,\n\t\t\tcost: { ...ZERO_COST, total: 0 },\n\t\t},\n\t\tstopReason: \"error\",\n\t\terrorMessage: message,\n\t\ttimestamp: 0,\n\t};\n}\n\nfunction failedStream(message: string, modelId: string): AssistantMessageEventStream {\n\treturn completedStream(errorMessage(message, modelId));\n}\n\nfunction recordedModel(modelId: string): Model<string> {\n\treturn {\n\t\tid: modelId,\n\t\tname: modelId,\n\t\tapi: RECORDED_API,\n\t\tprovider: RECORDED_PROVIDER_ID,\n\t\tbaseUrl: \"replay://recorded\",\n\t\treasoning: false,\n\t\tinput: [\"text\", \"image\"],\n\t\tcost: ZERO_COST,\n\t\tcontextWindow: 1_000_000,\n\t\tmaxTokens: 100_000,\n\t};\n}\n\nexport function registerRecordedProvider(\n\truntime: ModelRuntime,\n\tresponses: readonly AssistantMessage[],\n): RecordedProvider {\n\tconst pending = responses.map((message) => structuredClone(message));\n\tconst modelIds = [...new Set(pending.map((message) => message.model))];\n\tconst models = new Map(modelIds.map((modelId) => [modelId, recordedModel(modelId)]));\n\tlet requestCount = 0;\n\tconst contexts: import(\"@earendil-works/pi-ai\").Context[] = [];\n\n\truntime.registerProvider(RECORDED_PROVIDER_ID, {\n\t\tname: \"Recorded replay provider\",\n\t\tapi: RECORDED_API,\n\t\tapiKey: \"offline-replay\",\n\t\tmodels: [...models.values()],\n\t\tstreamSimple: (model, context) => {\n\t\t\trequestCount++;\n\t\t\tcontexts.push({\n\t\t\t\tsystemPrompt: context.systemPrompt,\n\t\t\t\tmessages: structuredClone(context.messages),\n\t\t\t});\n\t\t\tconst response = pending.shift();\n\t\t\tif (!response) return failedStream(`Recorded replay exhausted before request ${requestCount}`, model.id);\n\t\t\tif (response.model !== model.id) {\n\t\t\t\treturn failedStream(`Recorded model mismatch: expected ${response.model}, received ${model.id}`, model.id);\n\t\t\t}\n\t\t\treturn completedStream(response);\n\t\t},\n\t});\n\n\treturn {\n\t\tget requestCount() {\n\t\t\treturn requestCount;\n\t\t},\n\t\tget contexts() {\n\t\t\treturn contexts;\n\t\t},\n\t\tgetModel(modelId) {\n\t\t\tconst model = models.get(modelId);\n\t\t\tif (!model) throw new Error(`Unknown recorded model: ${modelId}`);\n\t\t\treturn model;\n\t\t},\n\t\tassertExhausted() {\n\t\t\tif (pending.length > 0) throw new Error(`Recorded replay left ${pending.length} response(s) unused`);\n\t\t},\n\t};\n}\n"]}

package/dist/testing/replay/runner.d.ts

@@ -0,0 +1,56 @@
+import { type AssistantMessage, type Message, type ToolResultMessage } from "@earendil-works/pi-ai";
+import { type ReplayMetrics } from "./metrics.ts";
+export interface ReplayResult {
+    turns: number;
+    requests: number;
+    networkCalls: number;
+    responses: AssistantMessage[];
+    toolResults: ToolResultMessage[];
+    metrics: ReplayMetrics;
+    /**
+     * The final outbound provider request's messages for each user turn — the same
+     * array `contextTokensByTurn` is measured from, exposed so tests can assert
+     * directly on context-pipeline effects (eviction markers, deferred schemas)
+     * without reaching into `RecordedProvider` internals.
+     */
+    contextsByTurn: Message[][];
+}
+/**
+ * Dedicated eviction budget for the replay harness — deliberately **not** derived
+ * from the fixture model's `contextWindow` (`recorded-provider.ts` hardcodes
+ * `1_000_000` for unrelated reasons predating this pipeline, shared by every
+ * fixture in the corpus, and is not to be changed here). Feeding that real
+ * `contextWindow` through production's `evictionBudget` formula
+ * (`../../core/context/pipeline.ts`) would yield a budget of roughly 490,000 tokens
+ * — orders of magnitude above anything in this corpus, so eviction would never
+ * observably fire and the gate it is meant to measure would stay blind, which is
+ * exactly the gap this task closes.
+ *
+ * `1_000` was chosen empirically — measured directly against `replayCorpus()`, not
+ * estimated — against three constraints that all have to hold simultaneously:
+ *
+ *  - `tool-error-recovery.jsonl`'s single turn carries its entire ~759 tokens
+ *    (mostly the static system prompt/tool-schema prefix) under budget, so its
+ *    recorded `contextTokensByTurn: [759]` (`test/replay-runner.test.ts`) is
+ *    unaffected — confirmed unchanged at this budget.
+ *  - `heavy-tool-output.jsonl` carries a single ~8,400-token recoverable `read`
+ *    result — comfortably over budget, so it evicts, which is the fixture's
+ *    documented purpose ("Large deterministic result for eviction measurements",
+ *    `fixtures/corpus/README.md`). Measured post-eviction turn: 748 tokens.
+ *  - `long-tool-heavy.jsonl` accumulates roughly 14,300 evictable tokens by turn 20
+ *    across ten interleaved `read`/`grep` results. `1_000` was verified against
+ *    `REPLAY_EVICTION_BUDGET: 0` (the theoretical floor — evict every eligible
+ *    result unconditionally) and produces the *same* turn-20 result, 1,769 tokens
+ *    — an 88.4% drop from the pre-pipeline baseline of 15,272 — confirming `1_000`
+ *    already reaches that floor rather than under-evicting. A first pass at
+ *    `3_000` measured 3,195 (79.1%, just short of the phase's 80% per-fixture
+ *    goal); `1_000` was chosen after confirming the floor is actually lower.
+ *
+ * This constant governs replay measurement only; it has no effect on production,
+ * which always calls `evictionBudget` with the real model's `contextWindow`.
+ */
+export declare const REPLAY_EVICTION_BUDGET = 1000;
+export declare function replay(filePath: string): Promise<ReplayResult>;
+/** Replay every JSONL fixture in lexical order for byte-stable corpus output. */
+export declare function replayCorpus(directory: string): Promise<Record<string, ReplayMetrics>>;
+//# sourceMappingURL=runner.d.ts.map

package/dist/testing/replay/runner.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"runner.d.ts","sourceRoot":"","sources":["../../../src/testing/replay/runner.ts"],"names":[],"mappings":"AAEA,OAAO,EACN,KAAK,gBAAgB,EAErB,KAAK,OAAO,EAGZ,KAAK,iBAAiB,EAEtB,MAAM,uBAAuB,CAAC;AAiB/B,OAAO,EAAsB,KAAK,aAAa,EAAE,MAAM,cAAc,CAAC;AAGtE,MAAM,WAAW,YAAY;IAC5B,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;IACjB,YAAY,EAAE,MAAM,CAAC;IACrB,SAAS,EAAE,gBAAgB,EAAE,CAAC;IAC9B,WAAW,EAAE,iBAAiB,EAAE,CAAC;IACjC,OAAO,EAAE,aAAa,CAAC;IACvB;;;;;OAKG;IACH,cAAc,EAAE,OAAO,EAAE,EAAE,CAAC;CAC5B;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAiCG;AACH,eAAO,MAAM,sBAAsB,OAAQ,CAAC;AA8L5C,wBAAsB,MAAM,CAAC,QAAQ,EAAE,MAAM,GAAG,OAAO,CAAC,YAAY,CAAC,CAmLpE;AAED,iFAAiF;AACjF,wBAAsB,YAAY,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,MAAM,EAAE,aAAa,CAAC,CAAC,CAQ5F","sourcesContent":["import { readdirSync, readFileSync } from \"node:fs\";\nimport { join } from \"node:path\";\nimport {\n\ttype AssistantMessage,\n\tInMemoryModelsStore,\n\ttype Message,\n\ttype Tool,\n\ttype ToolCall,\n\ttype ToolResultMessage,\n\tType,\n} from \"@earendil-works/pi-ai\";\nimport { Agent, type AgentTool, type AgentToolResult } from \"apex-code-agent-core\";\nimport { AuthStorage } from \"../../core/auth-storage.ts\";\nimport { estimateTokens } from \"../../core/compaction/compaction.ts\";\nimport type { ContractLookup } from \"../../core/context/eviction.ts\";\nimport { installContextPipeline } from \"../../core/context/pipeline.ts\";\nimport { ModelRuntime } from \"../../core/model-runtime.ts\";\nimport {\n\tbuildSessionContext,\n\tloadEntriesFromFile,\n\ttype SessionEntry,\n\ttype SessionHeader,\n\tSessionManager,\n\ttype SessionMessageEntry,\n} from \"../../core/session-manager.ts\";\nimport { buildSystemPrompt } from \"../../core/system-prompt.ts\";\nimport { createAllToolDefinitions, type ToolDef } from \"../../core/tools/index.ts\";\nimport { buildReplayMetrics, type ReplayMetrics } from \"./metrics.ts\";\nimport { registerRecordedProvider } from \"./recorded-provider.ts\";\n\nexport interface ReplayResult {\n\tturns: number;\n\trequests: number;\n\tnetworkCalls: number;\n\tresponses: AssistantMessage[];\n\ttoolResults: ToolResultMessage[];\n\tmetrics: ReplayMetrics;\n\t/**\n\t * The final outbound provider request's messages for each user turn — the same\n\t * array `contextTokensByTurn` is measured from, exposed so tests can assert\n\t * directly on context-pipeline effects (eviction markers, deferred schemas)\n\t * without reaching into `RecordedProvider` internals.\n\t */\n\tcontextsByTurn: Message[][];\n}\n\n/**\n * Dedicated eviction budget for the replay harness — deliberately **not** derived\n * from the fixture model's `contextWindow` (`recorded-provider.ts` hardcodes\n * `1_000_000` for unrelated reasons predating this pipeline, shared by every\n * fixture in the corpus, and is not to be changed here). Feeding that real\n * `contextWindow` through production's `evictionBudget` formula\n * (`../../core/context/pipeline.ts`) would yield a budget of roughly 490,000 tokens\n * — orders of magnitude above anything in this corpus, so eviction would never\n * observably fire and the gate it is meant to measure would stay blind, which is\n * exactly the gap this task closes.\n *\n * `1_000` was chosen empirically — measured directly against `replayCorpus()`, not\n * estimated — against three constraints that all have to hold simultaneously:\n *\n *  - `tool-error-recovery.jsonl`'s single turn carries its entire ~759 tokens\n *    (mostly the static system prompt/tool-schema prefix) under budget, so its\n *    recorded `contextTokensByTurn: [759]` (`test/replay-runner.test.ts`) is\n *    unaffected — confirmed unchanged at this budget.\n *  - `heavy-tool-output.jsonl` carries a single ~8,400-token recoverable `read`\n *    result — comfortably over budget, so it evicts, which is the fixture's\n *    documented purpose (\"Large deterministic result for eviction measurements\",\n *    `fixtures/corpus/README.md`). Measured post-eviction turn: 748 tokens.\n *  - `long-tool-heavy.jsonl` accumulates roughly 14,300 evictable tokens by turn 20\n *    across ten interleaved `read`/`grep` results. `1_000` was verified against\n *    `REPLAY_EVICTION_BUDGET: 0` (the theoretical floor — evict every eligible\n *    result unconditionally) and produces the *same* turn-20 result, 1,769 tokens\n *    — an 88.4% drop from the pre-pipeline baseline of 15,272 — confirming `1_000`\n *    already reaches that floor rather than under-evicting. A first pass at\n *    `3_000` measured 3,195 (79.1%, just short of the phase's 80% per-fixture\n *    goal); `1_000` was chosen after confirming the floor is actually lower.\n *\n * This constant governs replay measurement only; it has no effect on production,\n * which always calls `evictionBudget` with the real model's `contextWindow`.\n */\nexport const REPLAY_EVICTION_BUDGET = 1_000;\n\n/**\n * Builds a `ContractLookup` from the real production tool definitions — a second,\n * independent call to `createAllToolDefinitions` alongside the one\n * `productionPromptAndSchemas` already makes for the static-prompt baseline. Kept\n * separate rather than threading that function's `definitions` out: the two\n * concerns (building the static prompt/schema baseline vs. resolving a tool's\n * contract for the pipeline) are unrelated, and `createAllToolDefinitions` is cheap\n * and pure, so a second call costs nothing and keeps both call sites simple.\n */\nfunction replayContractLookup(cwd: string): ContractLookup {\n\tconst definitions = createAllToolDefinitions(cwd);\n\treturn (name) => (Reflect.get(definitions, name) as Partial<ToolDef> | undefined)?.contract;\n}\n\nfunction isVersionThreeHeader(entry: unknown): entry is SessionHeader & { version: 3 } {\n\tif (entry === null || typeof entry !== \"object\") return false;\n\treturn Reflect.get(entry, \"type\") === \"session\" && Reflect.get(entry, \"version\") === 3;\n}\n\nfunction isSessionEntry(entry: unknown): entry is SessionEntry {\n\tif (entry === null || typeof entry !== \"object\") return false;\n\tconst type = Reflect.get(entry, \"type\");\n\tconst id = Reflect.get(entry, \"id\");\n\tconst parentId = Reflect.get(entry, \"parentId\");\n\treturn (\n\t\ttype !== \"session\" &&\n\t\ttypeof type === \"string\" &&\n\t\ttypeof id === \"string\" &&\n\t\t(parentId === null || typeof parentId === \"string\")\n\t);\n}\n\nfunction validateSessionTree(entries: readonly SessionEntry[]): void {\n\tconst ids = new Set<string>();\n\tconst parentById = new Map<string, string | null>();\n\tlet roots = 0;\n\tfor (const entry of entries) {\n\t\tif (!entry || typeof entry.id !== \"string\" || ids.has(entry.id))\n\t\t\tthrow new Error(\"Replay session contains invalid or duplicate entry ids\");\n\t\tids.add(entry.id);\n\t\tparentById.set(entry.id, entry.parentId);\n\t\tif (entry.parentId === null) roots++;\n\t\telse if (typeof entry.parentId !== \"string\") throw new Error(\"Replay session contains an invalid parentId\");\n\t}\n\tif (roots !== 1) throw new Error(\"Replay session must contain exactly one tree root\");\n\tfor (const entry of entries) {\n\t\tif (entry.parentId !== null && !ids.has(entry.parentId))\n\t\t\tthrow new Error(`Replay session has missing parent ${entry.parentId}`);\n\t\tconst seen = new Set<string>();\n\t\tlet current: string | null = entry.id;\n\t\twhile (current !== null) {\n\t\t\tif (seen.has(current)) throw new Error(\"Replay session contains a parent cycle\");\n\t\t\tseen.add(current);\n\t\t\tcurrent = parentById.get(current) ?? null;\n\t\t}\n\t}\n}\n\nfunction isMessageEntry(entry: SessionEntry): entry is SessionMessageEntry {\n\treturn entry.type === \"message\";\n}\n\nfunction selectedResponses(entries: readonly SessionEntry[]): AssistantMessage[] {\n\treturn entries\n\t\t.filter(isMessageEntry)\n\t\t.map((entry) => entry.message)\n\t\t.filter((message): message is AssistantMessage => message.role === \"assistant\");\n}\n\nfunction selectedToolResults(entries: readonly SessionEntry[]): ToolResultMessage[] {\n\treturn entries\n\t\t.filter(isMessageEntry)\n\t\t.map((entry) => entry.message)\n\t\t.filter((message): message is ToolResultMessage => message.role === \"toolResult\");\n}\n\nfunction selectedUserEntries(entries: readonly SessionEntry[]): SessionMessageEntry[] {\n\treturn entries.filter(isMessageEntry).filter((entry) => entry.message.role === \"user\");\n}\n\nfunction toolCalls(responses: readonly AssistantMessage[]): ToolCall[] {\n\treturn responses.flatMap((message) => message.content.filter((part): part is ToolCall => part.type === \"toolCall\"));\n}\n\nfunction createRecordedTools(\n\tcalls: readonly ToolCall[],\n\tresults: readonly ToolResultMessage[],\n): { tools: AgentTool[]; assertExhausted(): void } {\n\tconst resultByCallId = new Map(results.map((result) => [result.toolCallId, structuredClone(result)]));\n\tconst unconsumed = new Set(resultByCallId.keys());\n\tconst tools = [...new Set(calls.map((call) => call.name))].map(\n\t\t(name): AgentTool => ({\n\t\t\tname,\n\t\t\tlabel: `Recorded ${name}`,\n\t\t\tdescription: \"Replays a recorded tool result without external side effects.\",\n\t\t\tparameters: Type.Any(),\n\t\t\texecute: async (toolCallId): Promise<AgentToolResult<unknown>> => {\n\t\t\t\tconst result = resultByCallId.get(toolCallId);\n\t\t\t\tif (!result) throw new Error(`Missing recorded tool result for ${toolCallId}`);\n\t\t\t\tunconsumed.delete(toolCallId);\n\t\t\t\tif (result.isError) {\n\t\t\t\t\tconst text = result.content.find((part) => part.type === \"text\")?.text ?? \"Recorded tool error\";\n\t\t\t\t\tthrow new Error(text);\n\t\t\t\t}\n\t\t\t\treturn {\n\t\t\t\t\tcontent: structuredClone(result.content),\n\t\t\t\t\tdetails: structuredClone(result.details),\n\t\t\t\t\tusage: structuredClone(result.usage),\n\t\t\t\t\taddedToolNames: structuredClone(result.addedToolNames),\n\t\t\t\t};\n\t\t\t},\n\t\t}),\n\t);\n\treturn {\n\t\ttools,\n\t\tassertExhausted() {\n\t\t\tif (unconsumed.size > 0) throw new Error(`Replay left ${unconsumed.size} tool result(s) unused`);\n\t\t},\n\t};\n}\n\nfunction assertResponsesMatch(actual: readonly AssistantMessage[], expected: readonly AssistantMessage[]): void {\n\tif (actual.length !== expected.length) {\n\t\tthrow new Error(`Replayed ${actual.length} assistant response(s), expected ${expected.length}`);\n\t}\n\tfor (const [index, response] of actual.entries()) {\n\t\tif (JSON.stringify(response) !== JSON.stringify(expected[index])) {\n\t\t\tthrow new Error(`Replayed assistant response ${index + 1} does not match the recording`);\n\t\t}\n\t}\n}\n\nfunction assertToolResultsMatch(actual: readonly ToolResultMessage[], expected: readonly ToolResultMessage[]): void {\n\tif (actual.length !== expected.length) {\n\t\tthrow new Error(`Replayed ${actual.length} tool result(s), expected ${expected.length}`);\n\t}\n\tfor (const [index, result] of actual.entries()) {\n\t\tconst recorded = expected[index];\n\t\tif (!recorded) throw new Error(`Missing recorded tool result ${index + 1}`);\n\t\tconst { timestamp: _actualTimestamp, ...actualStable } = result;\n\t\tconst { timestamp: _recordedTimestamp, ...recordedStable } = recorded;\n\t\tconst actualComparable = { ...actualStable };\n\t\tconst recordedComparable = { ...recordedStable };\n\t\tif (\n\t\t\tactualComparable.details === undefined ||\n\t\t\t(typeof actualComparable.details === \"object\" && Object.keys(actualComparable.details ?? {}).length === 0)\n\t\t)\n\t\t\tdelete actualComparable.details;\n\t\tif (\n\t\t\trecordedComparable.details === undefined ||\n\t\t\t(typeof recordedComparable.details === \"object\" && Object.keys(recordedComparable.details ?? {}).length === 0)\n\t\t)\n\t\t\tdelete recordedComparable.details;\n\t\tif (JSON.stringify(actualComparable) !== JSON.stringify(recordedComparable)) {\n\t\t\tthrow new Error(`Replayed tool result ${index + 1} does not match the recording`);\n\t\t}\n\t}\n}\n\nfunction productionPromptAndSchemas(toolNames: readonly string[]): { systemPrompt: string; tools: Tool[] } {\n\tconst definitions = createAllToolDefinitions(\"$HOME/replay-fixtures\");\n\tconst defaultToolNames = [\"read\", \"bash\", \"edit\", \"write\"];\n\tconst selected = [...new Set([...defaultToolNames, ...toolNames])].sort().flatMap((name) => {\n\t\tconst definition = Reflect.get(definitions, name);\n\t\treturn definition && typeof definition === \"object\" ? [definition] : [];\n\t});\n\tconst toolSnippets = Object.fromEntries(\n\t\tselected.flatMap((definition) =>\n\t\t\ttypeof definition.promptSnippet === \"string\" ? [[definition.name, definition.promptSnippet]] : [],\n\t\t),\n\t);\n\tconst promptGuidelines = selected.flatMap((definition) => definition.promptGuidelines ?? []);\n\treturn {\n\t\tsystemPrompt: buildSystemPrompt({\n\t\t\tcwd: \"$HOME/replay-fixtures\",\n\t\t\tcustomPrompt: \"You are an expert coding assistant operating inside Apex Code.\",\n\t\t\tselectedTools: selected.map((definition) => definition.name),\n\t\t\ttoolSnippets,\n\t\t\tpromptGuidelines,\n\t\t}),\n\t\ttools: selected.map((definition) => ({\n\t\t\tname: definition.name,\n\t\t\tdescription: definition.description,\n\t\t\tparameters: definition.parameters,\n\t\t})),\n\t};\n}\n\nexport async function replay(filePath: string): Promise<ReplayResult> {\n\tlet physicalLines: unknown[];\n\ttry {\n\t\tphysicalLines = readFileSync(filePath, \"utf8\")\n\t\t\t.split(\"\\n\")\n\t\t\t.filter((line) => line.trim().length > 0)\n\t\t\t.map((line) => JSON.parse(line));\n\t} catch (error) {\n\t\tthrow new Error(`Replay requires valid JSONL: ${filePath}`, { cause: error });\n\t}\n\tconst fileEntries = loadEntriesFromFile(filePath);\n\tif (physicalLines.length !== fileEntries.length) throw new Error(\"Replay session contains an invalid physical line\");\n\tif (!isVersionThreeHeader(fileEntries[0])) {\n\t\tthrow new Error(`Replay requires a native version 3 session: ${filePath}`);\n\t}\n\tconst rawEntries = fileEntries.slice(1);\n\tconst rawSessionEntries = rawEntries.filter(isSessionEntry);\n\tif (rawSessionEntries.length !== rawEntries.length) throw new Error(\"Replay session contains an invalid entry\");\n\tvalidateSessionTree(rawSessionEntries);\n\tconst sessionManager = SessionManager.open(filePath);\n\tconst entries = sessionManager.getBranch();\n\tvalidateSessionTree(entries);\n\tconst responses = selectedResponses(entries);\n\tconst recordedToolResults = selectedToolResults(entries);\n\tconst users = selectedUserEntries(entries);\n\tif (users.length === 0 || responses.length === 0)\n\t\tthrow new Error(\"Replay session must contain user and assistant messages\");\n\n\tconst runtime = await ModelRuntime.create({\n\t\tcredentials: AuthStorage.inMemory(),\n\t\tmodelsPath: null,\n\t\tmodelsStore: new InMemoryModelsStore(),\n\t\tallowModelNetwork: false,\n\t\trefreshOnCreate: false,\n\t});\n\tconst recorded = registerRecordedProvider(runtime, responses);\n\tconst recordedTools = createRecordedTools(toolCalls(responses), recordedToolResults);\n\tconst replayedResponses: AssistantMessage[] = [];\n\tconst replayedToolResults: ToolResultMessage[] = [];\n\tconst contextTokensByTurn: number[] = [];\n\tlet turnsCompleted = 0;\n\tconst promptBaseline = productionPromptAndSchemas(recordedTools.tools.map((tool) => tool.name));\n\tlet networkCalls = 0;\n\tconst rejectNetwork: typeof globalThis.fetch = async () => {\n\t\tnetworkCalls++;\n\t\tthrow new Error(\"Network access is disabled during replay\");\n\t};\n\tconst firstContext = buildSessionContext(entries, users[0]?.id);\n\tconst firstResponse = responses[0];\n\tif (!firstResponse) throw new Error(\"Replay session has no recorded response\");\n\tconst agent = new Agent({\n\t\tinitialState: {\n\t\t\tmodel: recorded.getModel(firstContext.model?.modelId ?? firstResponse.model),\n\t\t\tsystemPrompt: promptBaseline.systemPrompt,\n\t\t\ttools: recordedTools.tools,\n\t\t},\n\t\tstreamFn: (model, context, options) => runtime.streamSimple(model, context, { ...options, fetch: rejectNetwork }),\n\t\tgetApiKey: () => \"offline-replay\",\n\t\tconvertToLlm: (messages) => {\n\t\t\tconst converted: Message[] = [];\n\t\t\tfor (const message of messages) {\n\t\t\t\tif (message.role === \"user\") {\n\t\t\t\t\tconverted.push({ role: \"user\", content: message.content, timestamp: message.timestamp });\n\t\t\t\t\tcontinue;\n\t\t\t\t}\n\t\t\t\tif (message.role === \"assistant\") {\n\t\t\t\t\tconverted.push(message);\n\t\t\t\t\tcontinue;\n\t\t\t\t}\n\t\t\t\tif (message.role !== \"toolResult\") continue;\n\t\t\t\t// Validate the call id is one this replay actually recorded, but send the\n\t\t\t\t// *incoming* message — not a fresh lookup by id — so that whatever the\n\t\t\t\t// context pipeline's eviction stage did to `message.content` upstream\n\t\t\t\t// (in `transformContext`) is what reaches the provider request. Re-fetching\n\t\t\t\t// the original recorded content here would silently discard eviction,\n\t\t\t\t// since eviction never touches `recordedToolResults` itself (see\n\t\t\t\t// `assertToolResultsMatch`, which intentionally still compares against the\n\t\t\t\t// untouched recording).\n\t\t\t\tconst recordedResult = recordedToolResults.find((result) => result.toolCallId === message.toolCallId);\n\t\t\t\tif (!recordedResult) throw new Error(`Missing recorded tool result for ${message.toolCallId}`);\n\t\t\t\tconverted.push(structuredClone(message));\n\t\t\t}\n\t\t\treturn converted;\n\t\t},\n\t\tafterToolCall: async ({ toolCall }) => {\n\t\t\tconst result = recordedToolResults.find((candidate) => candidate.toolCallId === toolCall.id);\n\t\t\tif (!result) throw new Error(`Missing recorded tool result for ${toolCall.id}`);\n\t\t\treturn {\n\t\t\t\tcontent: structuredClone(result.content),\n\t\t\t\tdetails: result.details,\n\t\t\t\tusage: result.usage,\n\t\t\t\tisError: result.isError,\n\t\t\t};\n\t\t},\n\t\ttoolExecution: \"sequential\",\n\t});\n\n\t// Same pipeline `AgentSession` installs in production (`_installContextPipeline`\n\t// in `core/agent-session.ts`), via the shared `installContextPipeline` — deferred\n\t// -schema resolution ahead of every request, then tool-result eviction — so the\n\t// replay gate can observe both. The contract lookup is built from the real\n\t// production tool definitions; the budget is `REPLAY_EVICTION_BUDGET`, a\n\t// dedicated constant documented above (not derived from the fixture model's\n\t// `contextWindow`; see that constant's comment for why).\n\tinstallContextPipeline(agent, {\n\t\tcontractLookup: replayContractLookup(\"$HOME/replay-fixtures\"),\n\t\tevictionBudget: () => REPLAY_EVICTION_BUDGET,\n\t});\n\n\tconst contextsByTurn: Message[][] = [];\n\tfor (const user of users) {\n\t\tconst historical = buildSessionContext(entries, user.id);\n\t\tconst nextResponse = responses[recorded.requestCount];\n\t\tif (!nextResponse) throw new Error(`Missing recorded response for user turn ${contextTokensByTurn.length + 1}`);\n\t\tagent.state.model = recorded.getModel(historical.model?.modelId ?? nextResponse.model);\n\t\tagent.state.messages = historical.messages.slice(0, -1);\n\t\tconst requestStart = recorded.requestCount;\n\t\tawait agent.prompt(user.message);\n\t\tconst turnContexts = recorded.contexts.slice(requestStart);\n\t\tconst lastContext = turnContexts.at(-1);\n\t\tif (!lastContext) throw new Error(`User turn ${contextTokensByTurn.length + 1} made no provider request`);\n\t\tconst staticInputTokens = Math.ceil(\n\t\t\t`${promptBaseline.systemPrompt}\\n${JSON.stringify(promptBaseline.tools)}`.length / 4,\n\t\t);\n\t\tcontextTokensByTurn.push(\n\t\t\tstaticInputTokens + lastContext.messages.reduce((total, message) => total + estimateTokens(message), 0),\n\t\t);\n\t\t// Normalize toolResult timestamps to the recorded value before exposing this\n\t\t// turn's context, mirroring the same normalization the `toolResults` field\n\t\t// below already does and for the same reason: a toolResult's timestamp is\n\t\t// stamped live (via `agent.state.messages`) rather than replayed from the\n\t\t// fixture, so leaving it as-is would make `contextsByTurn` — and anything\n\t\t// that serializes it, like the determinism test below — flaky across runs.\n\t\t// Content (what eviction may have rewritten) is untouched.\n\t\tcontextsByTurn.push(\n\t\t\tlastContext.messages.map((message) =>\n\t\t\t\tmessage.role === \"toolResult\"\n\t\t\t\t\t? {\n\t\t\t\t\t\t\t...message,\n\t\t\t\t\t\t\ttimestamp:\n\t\t\t\t\t\t\t\trecordedToolResults.find((result) => result.toolCallId === message.toolCallId)?.timestamp ??\n\t\t\t\t\t\t\t\tmessage.timestamp,\n\t\t\t\t\t\t}\n\t\t\t\t\t: message,\n\t\t\t),\n\t\t);\n\t\tfor (const message of agent.state.messages.slice(historical.messages.length - 1)) {\n\t\t\tif (message.role === \"assistant\") replayedResponses.push(message);\n\t\t\tif (message.role === \"toolResult\") replayedToolResults.push(message);\n\t\t}\n\t\tconst terminalResponse = replayedResponses.at(-1);\n\t\tif (terminalResponse && terminalResponse.stopReason !== \"error\" && terminalResponse.stopReason !== \"aborted\") {\n\t\t\tturnsCompleted++;\n\t\t}\n\t}\n\trecorded.assertExhausted();\n\trecordedTools.assertExhausted();\n\tassertResponsesMatch(replayedResponses, responses);\n\tassertToolResultsMatch(replayedToolResults, recordedToolResults);\n\n\treturn {\n\t\tturns: users.length,\n\t\trequests: recorded.requestCount,\n\t\tnetworkCalls,\n\t\tresponses: replayedResponses,\n\t\ttoolResults: replayedToolResults.map((result, index) => ({\n\t\t\t...structuredClone(result),\n\t\t\ttimestamp: recordedToolResults[index]?.timestamp ?? result.timestamp,\n\t\t})),\n\t\tcontextsByTurn,\n\t\tmetrics: buildReplayMetrics({\n\t\t\tsystemPrompt: promptBaseline.systemPrompt,\n\t\t\ttools: promptBaseline.tools,\n\t\t\trecordedResponses: responses,\n\t\t\treplayedResponses,\n\t\t\tcontextTokensByTurn,\n\t\t\tturnsCompleted,\n\t\t}),\n\t};\n}\n\n/** Replay every JSONL fixture in lexical order for byte-stable corpus output. */\nexport async function replayCorpus(directory: string): Promise<Record<string, ReplayMetrics>> {\n\tconst metrics: Record<string, ReplayMetrics> = {};\n\tfor (const name of readdirSync(directory)\n\t\t.filter((entry) => entry.endsWith(\".jsonl\"))\n\t\t.sort()) {\n\t\tmetrics[name] = (await replay(join(directory, name))).metrics;\n\t}\n\treturn metrics;\n}\n"]}