effect-agent 0.1.0-beta.140 → 0.1.0-beta.142

package/dist/{AgentRuntime-RdmnK_aD.mjs → AgentRuntime-xeCUXutN.mjs}

@@ -24,7 +24,7 @@ import { CompactionDecision, CompactionError, ContextCompactor } from "./engine/
 import { RunEventSink, RunEventSinkClosedError } from "./engine/RunEventSink.mjs";
 import { AgentUpdateAcceptance, CurrentToolFailureObserver, ModelUsageAccounting, RunContextPreparation, RunResumeUsageSchema, RunToolAuthorization, RunToolScheduling, RunTurnResumeSettledCallSchema } from "./engine/RunOptions.mjs";
 import { ToolBroker, ToolBrokerConfigurationError, ToolBrokerUnavailableError } from "./engine/ToolBroker.mjs";
-import { AiError, LanguageModel, Model, Prompt, Response, Tool, Toolkit } from "effect/unstable/ai";
+import { AiError, LanguageModel, Model, Prompt, Response, ResponseIdTracker, Tool, Toolkit } from "effect/unstable/ai";
 import { Cause, Channel, Clock, Context, DateTime, Duration, Effect, Equal, ErrorReporter, Exit, Fiber, Layer, Metric, Option, PubSub, Queue, Redacted, Result, Schema, SchemaAST, SchemaGetter, Scope, Semaphore, Stream, Tracer } from "effect";
 //#region src/engine/internal/bounded-value.ts
 const DEFAULT_MAX_DEPTH = 128;
@@ -1860,11 +1860,13 @@ const formatRunStatus = (view) => {
 /**
 * The run-status message is derived per request and appended only to the
 * OUTGOING prompt: official history and durable commits never carry it, so it
-* can never accumulate or replay.
+* can never accumulate or replay. OpenAI receives it as trailing system guidance,
+* leaving the last user/tool result as the implicit cache-write boundary.
 */
 const outgoingModelPrompt = (policy, context, prepared, turn, declaredToolCalls) => Effect.gen(function* () {
 	if (policy.runStatus !== "appended") return prepared;
 	const now = yield* Clock.currentTimeMillis;
+	const provider = yield* Model.ProviderName;
 	const status = formatRunStatus({
 		turn,
 		maxTurns: policy.maxTurns,
@@ -1877,7 +1879,7 @@ const outgoingModelPrompt = (policy, context, prepared, turn, declaredToolCalls)
 		elapsedSeconds: Math.max(0, Math.floor((now - context.startedAtMillis) / 1e3)),
 		maxDurationSeconds: Math.floor(Duration.toMillis(policy.maxDuration) / 1e3)
 	});
-	return Prompt.fromMessages([...prepared.content, Prompt.makeMessage("user", { content: [Prompt.makePart("text", { text: status })] })]);
+	return Prompt.fromMessages([...prepared.content, provider === "openai" ? Prompt.systemMessage({ content: status }) : Prompt.userMessage({ content: [Prompt.textPart({ text: status })] })]);
 });
 const ProviderUsage = Schema.Struct({
 	inputTokens: Schema.Struct({
@@ -2900,7 +2902,7 @@ const makeTurn = (agent, context, prompt, turn, priorToolCalls, options) => Stre
 	}));
 	const admissionRequired = contextTokenLimit !== void 0 || !context.finalizing && !finalAnswerOnly && policy.tokenBudget !== void 0;
 	const outputContractTokens = !admissionRequired || outputContractMessage === void 0 ? 0 : yield* estimateCallTokens([Prompt.makeMessage("system", { content: outputContractMessage })]);
-	const canonicalDecoration = !admissionRequired ? Prompt.empty : yield* outgoingModelPrompt(policy, context, Prompt.empty, turn, priorToolCalls);
+	const canonicalDecoration = !admissionRequired ? Prompt.empty : yield* outgoingModelPrompt(policy, context, Prompt.empty, turn, priorToolCalls).pipe(withCallModel);
 	const estimateToolSchemaTokens = Effect.suspend(() => {
 		const choice = modelToolChoice();
 		const tools = typeof choice === "object" && "oneOf" in choice ? catalog.filter((entry) => entry.kind === "native" && choice.oneOf.includes(entry.nativeToolName)).map((entry) => entry.tool) : Object.values(modelToolkit.tools);
@@ -3006,7 +3008,7 @@ const makeTurn = (agent, context, prompt, turn, priorToolCalls, options) => Stre
 		});
 	}
 	const transientContext = options.transientContext === void 0 ? Prompt.empty : yield* options.transientContext.load(contextRequest).pipe(Effect.flatMap(transientInputToPrompt));
-	const derivedPrompt = options.transientContext === void 0 ? canonicalDecoration : yield* outgoingModelPrompt(policy, context, transientContext, turn, priorToolCalls);
+	const derivedPrompt = options.transientContext === void 0 ? canonicalDecoration : yield* outgoingModelPrompt(policy, context, transientContext, turn, priorToolCalls).pipe(withCallModel);
 	const derivedPromptContentTokens = !admissionRequired ? 0 : options.transientContext === void 0 ? canonicalDecorationPromptTokens : outputContractTokens + (yield* estimateCallTokens(derivedPrompt.content));
 	const derivedPromptTokens = () => toolSchemaTokens + derivedPromptContentTokens;
 	if (!context.finalizing && admissionRequired && options.transientContext !== void 0) {
@@ -3078,7 +3080,7 @@ const makeTurn = (agent, context, prompt, turn, priorToolCalls, options) => Stre
 		if (trace.usage === void 0) return yield* noteIncompleteUsage(context, turn);
 		yield* consumeTurnUsage(0).pipe(Effect.exit);
 	});
-	const attempt = (basis) => Stream.unwrap(outgoingModelPrompt(policy, context, Prompt.fromMessages([...basis.content, ...transientContext.content]), turn, priorToolCalls).pipe(Effect.flatMap((outgoing) => Effect.gen(function* () {
+	const attempt = (basis) => Stream.unwrap(outgoingModelPrompt(policy, context, prepareModelPrompt(Prompt.fromMessages([...basis.content, ...transientContext.content]), outputContract._tag === "rendered" ? outputContract.part : void 0), turn, priorToolCalls).pipe(withCallModel, Effect.flatMap((providerPrompt) => Effect.gen(function* () {
 		const toolChoice = modelToolChoice();
 		let requestToolkit = modelToolkit;
 		snapshot = selectionSnapshot;
@@ -3092,7 +3094,6 @@ const makeTurn = (agent, context, prompt, turn, priorToolCalls, options) => Stre
 			}));
 		}
 		context.toolExposure = snapshot;
-		const providerPrompt = prepareModelPrompt(outgoing, outputContract._tag === "rendered" ? outputContract.part : void 0);
 		return yield* estimateCallTokens(providerPrompt.content).pipe(Effect.map((tokens) => tokens + toolSchemaTokens), Effect.tap((estimatedTokens) => contextTokenLimit !== void 0 && (options.context !== void 0 || options.transientContext !== void 0) && estimatedTokens > contextTokenLimit ? ContextBudgetError.make({
 			message: `Prepared context could not fit the next model prompt inside the ${contextTokenLimit} token context target`,
 			estimatedTokens,
@@ -3105,7 +3106,7 @@ const makeTurn = (agent, context, prompt, turn, priorToolCalls, options) => Stre
 			toolkit: deferredToolParameterToolkit(requestToolkit),
 			disableToolCallResolution: true,
 			...toolChoice === void 0 ? {} : { toolChoice }
-		}), options.budget).pipe(Stream.provideServiceEffect(Tracer.Tracer, modelTelemetryTracer(context, turnId)), Stream.onStart(Effect.sync(() => {
+		}), options.budget).pipe((stream) => policy.runStatus === "appended" || options.context !== void 0 || options.transientContext !== void 0 ? stream.pipe(Stream.provideServiceEffect(ResponseIdTracker.ResponseIdTracker, ResponseIdTracker.make)) : stream, Stream.provideServiceEffect(Tracer.Tracer, modelTelemetryTracer(context, turnId)), Stream.onStart(Effect.sync(() => {
 			trace.usageConsumed = false;
 		})), Stream.mapEffect((part) => ownModelResponsePart(part, agent.definition.toolkit, trace, context.bufferLimits).pipe(Effect.flatMap((owned) => eventsForPart(context, turnId, turn, agent.definition.toolkit.tools, trace, owned.ownedPart, owned.retainedBytes)))), Stream.flatMap(Stream.fromIterable), Stream.tapCause(() => retainFailedUsage()))));
 	}))));
@@ -4736,4 +4737,4 @@ var AgentRuntime_exports = /* @__PURE__ */ __exportAll({
 //#endregion
 export { stream as _, SubagentDurability as a, withTerminalDefectEvent as b, decodeFinalOutput as c, projectCompletionOutput as d, renderInputPrompt as f, startUnknown as g, start as h, AgentSpawner as i, encodeRunDisposition as l, runUnknown as m, AgentChildPending as n, SubagentDurabilityError as o, run as p, AgentResultSchema as r, ToolCallWaiting as s, AgentRuntime_exports as t, projectCompletionFromToolOutput as u, streamUnknown as v, streamWithUsageAccountingUnknown as y };
 
-//# sourceMappingURL=AgentRuntime-RdmnK_aD.mjs.map
+//# sourceMappingURL=AgentRuntime-xeCUXutN.mjs.map