@sayknow-cli/agent-core 0.3.6 → 0.3.8

package/dist/types/types.d.ts

@@ -0,0 +1,430 @@
+import type { AssistantMessage, AssistantMessageEvent, AssistantMessageEventStream, Effort, ImageContent, Message, Model, SimpleStreamOptions, Static, streamSimple, TextContent, Tool, ToolChoice, ToolResultMessage, TSchema } from "@sayknow-cli/ai";
+import type { AppendOnlyContextManager } from "./append-only-context";
+import type { HarmonyAuditEvent } from "./harmony-leak";
+import type { AgentRunCoverage, AgentRunSummary } from "./run-collector";
+import type { AgentTelemetryConfig } from "./telemetry";
+/** Stream function - can return sync or Promise for async config lookup */
+export type StreamFn = (...args: Parameters<typeof streamSimple>) => AssistantMessageEventStream | Promise<AssistantMessageEventStream>;
+/**
+ * Configuration for the agent loop.
+ */
+export interface AgentLoopConfig extends SimpleStreamOptions {
+    model: Model;
+    /**
+     * When to interrupt tool execution for steering messages.
+     * - "immediate" = check after each tool call (default)
+     * - "wait" = defer steering until the current turn completes
+     */
+    interruptMode?: "immediate" | "wait";
+    /**
+     * Optional session identifier forwarded to LLM providers.
+     * Used by providers that support session-based caching (e.g., OpenAI code provider).
+     */
+    sessionId?: string;
+    /**
+     * Optional provider-facing cache/session affinity identifier. When set, this
+     * is forwarded to providers as StreamOptions.sessionId while `sessionId`
+     * remains the logical agent conversation id for telemetry/metadata.
+     */
+    providerSessionId?: string;
+    /**
+     * Optional resolver called per LLM request to produce request metadata.
+     * When set, the agent loop evaluates it **after** `getApiKey` resolves the
+     * session-sticky credential, ensuring the metadata's `account_uuid` reflects
+     * the credential actually used for the request (not the credential that was
+     * current when `AgentLoopConfig` was first constructed). Overrides the static
+     * `metadata` field when present.
+     */
+    metadataResolver?: (provider: string) => Record<string, unknown> | undefined;
+    /**
+     * Converts AgentMessage[] to LLM-compatible Message[] before each LLM call.
+     *
+     * Each AgentMessage must be converted to a UserMessage, AssistantMessage, or ToolResultMessage
+     * that the LLM can understand. AgentMessages that cannot be converted (e.g., UI-only notifications,
+     * status messages) should be filtered out.
+     *
+     * @example
+     * ```typescript
+     * convertToLlm: (messages) => messages.flatMap(m => {
+     *   if (m.role === "custom") {
+     *     // Convert custom message to user message
+     *     return [{ role: "user", content: m.content, timestamp: m.timestamp }];
+     *   }
+     *   if (m.role === "notification") {
+     *     // Filter out UI-only messages
+     *     return [];
+     *   }
+     *   // Pass through standard LLM messages
+     *   return [m];
+     * })
+     * ```
+     */
+    convertToLlm: (messages: AgentMessage[]) => Message[] | Promise<Message[]>;
+    /**
+     * Optional transform applied to the context before `convertToLlm`.
+     *
+     * Use this for operations that work at the AgentMessage level:
+     * - Context window management (pruning old messages)
+     * - Injecting context from external sources
+     *
+     * @example
+     * ```typescript
+     * transformContext: async (messages) => {
+     *   if (estimateTokens(messages) > MAX_TOKENS) {
+     *     return pruneOldMessages(messages);
+     *   }
+     *   return messages;
+     * }
+     * ```
+     */
+    transformContext?: (messages: AgentMessage[], signal?: AbortSignal) => Promise<AgentMessage[]>;
+    /**
+     * Resolves an API key dynamically for each LLM call.
+     *
+     * Useful for short-lived OAuth tokens (e.g., GitHub Copilot) that may expire
+     * during long-running tool execution phases.
+     */
+    getApiKey?: (provider: string) => Promise<string | undefined> | string | undefined;
+    /** Returns the credential type selected by the most recent getApiKey call for this session/provider. */
+    getAuthCredentialType?: (provider: string) => "api_key" | "oauth" | undefined;
+    /**
+     * Returns steering messages to inject into the conversation mid-run.
+     *
+     * Called after each tool execution to check for user interruptions unless interruptMode is "wait".
+     * If messages are returned, remaining tool calls are skipped and
+     * these messages are added to the context before the next LLM call.
+     */
+    getSteeringMessages?: () => Promise<AgentMessage[]>;
+    /**
+     * Returns follow-up messages to process after the agent would otherwise stop.
+     *
+     * Called when the agent has no more tool calls and no steering messages.
+     * If messages are returned, they're added to the context and the agent
+     * continues with another turn.
+     */
+    getFollowUpMessages?: () => Promise<AgentMessage[]>;
+    /**
+     * Cooperative pause checkpoint evaluated at safe loop boundaries.
+     *
+     * Called after completed tool execution has been emitted and before the loop
+     * polls steering/follow-up queues or schedules another assistant response.
+     * Returning true ends the current loop with `agent_end.stopReason === "paused"`
+     * without aborting any in-flight model or tool work.
+     */
+    shouldPause?: () => boolean;
+    /**
+     * Hook fired right before the loop would exit.
+     *
+     * Called when the agent has no more tool calls and no steering messages,
+     * immediately before polling follow-up messages.
+     */
+    onBeforeYield?: () => Promise<void> | void;
+    /**
+     * Provides tool execution context, resolved per tool call.
+     * Use for late-bound UI or session state access.
+     */
+    getToolContext?: (toolCall?: ToolCallContext) => AgentToolContext | undefined;
+    /**
+     * Refreshes prompt/tool context from live session state before each model call.
+     * Use this when tool availability or the system prompt can change mid-turn.
+     */
+    syncContextBeforeModelCall?: (context: AgentContext) => void | Promise<void>;
+    /**
+     * Optional transform applied to tool call arguments before execution.
+     * Use for deobfuscating secrets or rewriting arguments.
+     */
+    transformToolCallArguments?: (args: Record<string, unknown>, toolName: string) => Record<string, unknown>;
+    /**
+     * Enable intent tracing for tool calls.
+     * When enabled, the harness injects a `string` field into tool schemas sent to the model,
+     * then strips from arguments before executing tools.
+     */
+    intentTracing?: boolean;
+    /**
+     * Append-only context mode — stabilizes system prompt + tool spec bytes
+     * across turns so provider prefix caches hit at maximum rate.
+     *
+     * When set, the loop reads messages from the append-only log (stable
+     * byte prefix) and caches system prompt + tools. Tools exclude per-turn
+     * `_i` intent fields.
+     */
+    appendOnlyContext?: AppendOnlyContextManager;
+    /**
+     * Inspect assistant streaming events before they are published to the outer agent event stream.
+     * Callers may abort synchronously to stop consuming buffered provider events.
+     */
+    onAssistantMessageEvent?: (message: AssistantMessage, event: AssistantMessageEvent) => void;
+    /** Called for non-content tool-choice incapability stream events. */
+    onToolChoiceIncapability?: (event: Extract<AssistantMessageEvent, {
+        type: "toolChoiceIncapability";
+    }>) => void;
+    /**
+     * Called when GPT-5 Harmony protocol leakage is detected and mitigated.
+     */
+    onHarmonyLeak?: (event: HarmonyAuditEvent) => void | Promise<void>;
+    /**
+     * Dynamic tool choice override, resolved per LLM call.
+     * When set and returns a value, overrides the static `toolChoice`.
+     */
+    getToolChoice?: () => ToolChoice | undefined;
+    /**
+     * Dynamic reasoning effort override, resolved per LLM call.
+     * When set and returns a value, overrides the static `reasoning` captured
+     * at run-loop start. Use this so mid-run thinking-level changes apply on
+     * the next model call instead of waiting for the next prompt.
+     */
+    getReasoning?: () => Effort | undefined;
+    /**
+     * Called after a tool call has been validated and is about to execute.
+     *
+     * Return `{ block: true }` to prevent execution. The loop emits an error tool
+     * result instead (using `reason` as the error text, or a default if omitted).
+     *
+     * Mutating `context.args` in place changes the arguments passed to `tool.execute`
+     * — the loop does **not** re-validate after this hook runs.
+     *
+     * The hook receives the tool abort signal (`signal`) and is responsible for
+     * honoring it. Throwing surfaces as a tool-error result and does not abort the
+     * rest of the batch.
+     */
+    beforeToolCall?: (context: BeforeToolCallContext, signal?: AbortSignal) => Promise<BeforeToolCallResult | undefined> | BeforeToolCallResult | undefined;
+    /**
+     * Called after a tool finishes executing, before `tool_execution_end` and the
+     * tool-result message are emitted.
+     *
+     * Return an `AfterToolCallResult` to override individual fields of the executed
+     * tool result. Omitted fields keep their original values; there is no deep merge.
+     *
+     * Throwing surfaces as a tool-error result and does not abort the rest of the batch.
+     */
+    afterToolCall?: (context: AfterToolCallContext, signal?: AbortSignal) => Promise<AfterToolCallResult | undefined> | AfterToolCallResult | undefined;
+    /**
+     * Opt-in OpenTelemetry instrumentation. Passing `{}` enables the loop's
+     * GenAI-semantic-convention spans (`invoke_agent`, `chat`, `execute_tool`)
+     * using the global tracer provider. Leaving this field undefined disables
+     * the instrumentation entirely — the loop performs zero tracer lookups.
+     *
+     * See {@link AgentTelemetryConfig} for the full surface (hooks, content
+     * capture, cost estimator, agent identity).
+     */
+    telemetry?: AgentTelemetryConfig;
+}
+/**
+ * Batch/sequencing metadata for the tool call currently being processed.
+ */
+export interface ToolCallContext {
+    batchId: string;
+    index: number;
+    total: number;
+    toolCalls: Array<{
+        id: string;
+        name: string;
+    }>;
+}
+/** A single tool-call content block emitted by an assistant message. */
+export type AgentToolCall = Extract<AssistantMessage["content"][number], {
+    type: "toolCall";
+}>;
+/**
+ * Result returned from `beforeToolCall`.
+ *
+ * Set `block: true` to prevent the tool from executing. The loop emits an error tool
+ * result instead, using `reason` as the error text (or a default if omitted).
+ *
+ * Mutating the `args` reference passed in `BeforeToolCallContext` is supported and
+ * survives into execution — the loop does **not** re-validate after this hook runs.
+ */
+export interface BeforeToolCallResult {
+    block?: boolean;
+    reason?: string;
+}
+/**
+ * Partial override returned from `afterToolCall`.
+ *
+ * Merge semantics are field-by-field; omitted fields keep the executed values.
+ * No deep merge is performed.
+ */
+export interface AfterToolCallResult {
+    /** If provided, replaces the tool result content array in full. */
+    content?: (TextContent | ImageContent)[];
+    /** If provided, replaces the tool result details payload in full. */
+    details?: unknown;
+    /** If provided, replaces the error flag carried with the tool result. */
+    isError?: boolean;
+}
+/** Context passed to `beforeToolCall`. */
+export interface BeforeToolCallContext {
+    /** The assistant message that requested the tool call. */
+    assistantMessage: AssistantMessage;
+    /** The raw tool call block from `assistantMessage.content`. */
+    toolCall: AgentToolCall;
+    /**
+     * Validated tool arguments. The same reference is forwarded to `tool.execute`
+     * (after any `transformToolCallArguments` pass), so in-place mutations stick.
+     */
+    args: Record<string, unknown>;
+    /** Current agent context at the time the tool call is prepared. */
+    context: AgentContext;
+}
+/** Context passed to `afterToolCall`. */
+export interface AfterToolCallContext {
+    /** The assistant message that requested the tool call. */
+    assistantMessage: AssistantMessage;
+    /** The raw tool call block from `assistantMessage.content`. */
+    toolCall: AgentToolCall;
+    /** Validated tool arguments used for execution (post `beforeToolCall` mutations). */
+    args: Record<string, unknown>;
+    /** The executed tool result before any `afterToolCall` overrides are applied. */
+    result: AgentToolResult<any>;
+    /** Whether the executed tool result is currently treated as an error. */
+    isError: boolean;
+    /** Current agent context at the time the tool call is finalized. */
+    context: AgentContext;
+}
+/**
+ * Extensible interface for custom app messages.
+ * Apps can extend via declaration merging:
+ *
+ * @example
+ * ```typescript
+ * declare module "@sayknow-cli/agent" {
+ *   interface CustomAgentMessages {
+ *     artifact: ArtifactMessage;
+ *     notification: NotificationMessage;
+ *   }
+ * }
+ * ```
+ */
+export interface CustomAgentMessages {
+}
+/**
+ * AgentMessage: Union of LLM messages + custom messages.
+ * This abstraction allows apps to add custom message types while maintaining
+ * type safety and compatibility with the base LLM messages.
+ */
+export type AgentMessage = Message | CustomAgentMessages[keyof CustomAgentMessages];
+/**
+ * Agent state containing all configuration and conversation data.
+ */
+export interface AgentState {
+    systemPrompt: string[];
+    model: Model;
+    thinkingLevel?: Effort;
+    tools: AgentTool<any>[];
+    messages: AgentMessage[];
+    isStreaming: boolean;
+    streamMessage: AgentMessage | null;
+    pendingToolCalls: Set<string>;
+    error?: string;
+}
+export interface AgentToolResult<T = any, _TInput = unknown> {
+    content: (TextContent | ImageContent)[];
+    details?: T;
+    isError?: boolean;
+}
+export type AgentToolUpdateCallback<T = any, TInput = unknown> = (partialResult: AgentToolResult<T, TInput>) => void;
+/** Options passed to renderResult */
+export interface RenderResultOptions {
+    /** Whether the result view is expanded */
+    expanded: boolean;
+    /** Whether this is a partial/streaming result */
+    isPartial: boolean;
+    /** Current spinner frame index for animated elements (optional) */
+    spinnerFrame?: number;
+}
+/**
+ * Context passed to tool execution.
+ * Apps can extend via declaration merging.
+ */
+export interface AgentToolContext {
+}
+export type AgentToolExecFn<TParameters extends TSchema = TSchema, TDetails = any, TTheme = unknown> = (this: AgentTool<TParameters, TDetails, TTheme>, toolCallId: string, params: Static<TParameters>, signal?: AbortSignal, onUpdate?: AgentToolUpdateCallback<TDetails, TParameters>, context?: AgentToolContext) => Promise<AgentToolResult<TDetails, TParameters>>;
+export interface AgentTool<TParameters extends TSchema = TSchema, TDetails = any, TTheme = unknown> extends Tool<TParameters> {
+    label: string;
+    /** If true, tool is excluded unless explicitly listed in --tools or agent's tools field */
+    hidden?: boolean;
+    /** If true, tool can stage a pending action that requires explicit resolution via the resolve tool. */
+    deferrable?: boolean;
+    /** Built-in tool loading behavior. "essential" loads initially; "discoverable" can be activated by tool search. */
+    loadMode?: "essential" | "discoverable";
+    /** Short one-line summary used for tool discovery indexes. */
+    summary?: string;
+    /** If true, tool execution ignores abort signals (runs to completion) */
+    nonAbortable?: boolean;
+    /**
+     * Concurrency mode for tool scheduling when multiple calls are in one turn.
+     * - "shared": can run alongside other shared tools (default)
+     * - "exclusive": runs alone; other tools wait until it finishes
+     */
+    concurrency?: "shared" | "exclusive";
+    /** If true, argument validation errors are non-fatal: raw args are passed to execute() instead of returning an error to the LLM. */
+    lenientArgValidation?: boolean;
+    /**
+     * Controls how the INTENT_FIELD (`_i`) is handled for this tool.
+     * - `"require"` (default): `_i` is injected and required in the parameter schema.
+     * - `"optional"`: `_i` is injected as an optional/nullable field.
+     * - `"omit"`: `_i` is NOT injected. Use for tools where intent is obvious (yield, resolve, todo_write, …).
+     * - function: `_i` is NOT injected; intent is derived dynamically from (potentially partial / streaming) args.
+     */
+    intent?: "omit" | "optional" | "require" | ((args: Partial<Static<TParameters>>) => string | undefined);
+    /** The main execution callback for this tool. */
+    execute: AgentToolExecFn<TParameters, TDetails, TTheme>;
+    /** Optional custom rendering for tool call display (returns UI component) */
+    renderCall?: (args: Static<TParameters>, options: RenderResultOptions, theme: TTheme) => unknown;
+    /** Optional custom rendering for tool result display (returns UI component) */
+    renderResult?: (result: AgentToolResult<TDetails, TParameters>, options: RenderResultOptions, theme: TTheme) => unknown;
+}
+export interface AgentContext {
+    systemPrompt: string[];
+    messages: AgentMessage[];
+    tools?: AgentTool<any>[];
+}
+/**
+ * Events emitted by the Agent for UI updates.
+ * These events provide fine-grained lifecycle information for messages, turns, and tool executions.
+ */
+export type AgentEvent = {
+    type: "agent_start";
+} | {
+    type: "agent_end";
+    messages: AgentMessage[];
+    /** Indicates whether the loop ended normally or suspended at a pause checkpoint. */
+    stopReason?: "completed" | "paused";
+    /** Present iff `AgentTelemetryConfig` was supplied on this run. */
+    telemetry?: AgentRunSummary;
+    coverage?: AgentRunCoverage;
+} | {
+    type: "turn_start";
+} | {
+    type: "turn_end";
+    message: AgentMessage;
+    toolResults: ToolResultMessage[];
+} | {
+    type: "message_start";
+    message: AgentMessage;
+} | {
+    type: "message_update";
+    message: AgentMessage;
+    assistantMessageEvent: AssistantMessageEvent;
+} | {
+    type: "message_end";
+    message: AgentMessage;
+} | {
+    type: "tool_execution_start";
+    toolCallId: string;
+    toolName: string;
+    args: any;
+    intent?: string;
+} | {
+    type: "tool_execution_update";
+    toolCallId: string;
+    toolName: string;
+    args: any;
+    partialResult: any;
+} | {
+    type: "tool_execution_end";
+    toolCallId: string;
+    toolName: string;
+    result: any;
+    isError?: boolean;
+};

package/package.json

@@ -1,9 +1,9 @@
 {
 	"type": "module",
 	"name": "@sayknow-cli/agent-core",
-	"version": "0.3.6",
+	"version": "0.3.8",
 	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
-	"homepage": "https://github.com/jaybeyond/Sayknow_CLI",
+	"homepage": "https://sayknow-cli.com",
 	"author": "jaybeyond",
 	"contributors": [
 		"Mario Zechner"
@@ -25,7 +25,7 @@
 		"state-management"
 	],
 	"main": "./src/index.ts",
-	"types": "./src/index.ts",
+	"types": "./dist/types/index.d.ts",
 	"scripts": {
 		"check": "biome check . && bun run check:types",
 		"check:types": "tsgo -p tsconfig.json --noEmit",
@@ -35,9 +35,9 @@
 		"fmt": "biome format --write ."
 	},
 	"dependencies": {
-		"@sayknow-cli/ai": "0.3.6",
-		"@sayknow-cli/natives": "0.3.6",
-		"@sayknow-cli/utils": "0.3.6",
+		"@sayknow-cli/ai": "0.3.8",
+		"@sayknow-cli/natives": "0.3.8",
+		"@sayknow-cli/utils": "0.3.8",
 		"@opentelemetry/api": "^1.9.0"
 	},
 	"devDependencies": {
@@ -51,23 +51,24 @@
 	"files": [
 		"src",
 		"README.md",
-		"CHANGELOG.md"
+		"CHANGELOG.md",
+		"dist/types"
 	],
 	"exports": {
 		".": {
-			"types": "./src/index.ts",
+			"types": "./dist/types/index.d.ts",
 			"import": "./src/index.ts"
 		},
 		"./compaction": {
-			"types": "./src/compaction.ts",
+			"types": "./dist/types/compaction.d.ts",
 			"import": "./src/compaction.ts"
 		},
 		"./compaction/*": {
-			"types": "./src/compaction/*.ts",
+			"types": "./dist/types/compaction/*.d.ts",
 			"import": "./src/compaction/*.ts"
 		},
 		"./*": {
-			"types": "./src/*.ts",
+			"types": "./dist/types/*.d.ts",
 			"import": "./src/*.ts"
 		}
 	}

package/src/agent-loop.ts

@@ -342,9 +342,136 @@ export function normalizeMessagesForProvider(
 	return changed ? normalized : messages;
 }
 
+interface ConvertedContextCacheEntry {
+	messageHashes: string[];
+	modelKey: string;
+	toolKey: string;
+	intentTracing: boolean;
+	convertToLlm: AgentLoopConfig["convertToLlm"];
+	transformContext: AgentLoopConfig["transformContext"];
+	llmMessages: Context["messages"];
+	normalizedMessages: Context["messages"];
+}
+
+const convertedContextCache = new WeakMap<AgentLoopConfig, ConvertedContextCacheEntry>();
+
+function stableCacheString(value: unknown): string | undefined {
+	try {
+		return JSON.stringify(value, (_key, item) =>
+			typeof item === "function" ? `[Function:${item.name || "anonymous"}]` : item,
+		);
+	} catch {
+		return undefined;
+	}
+}
+
+/**
+ * Hash a message by full content serialization.
+ *
+ * Deliberately NOT memoized by object identity: callers mutate messages in
+ * place (compaction rewrites, obfuscation, abort markers) and the cache's
+ * correctness contract requires detecting those mutations. The per-turn
+ * serialization cost is the price of that contract; the win is skipping
+ * convertToLlm + normalize on stable contexts, which dominates for
+ * image-heavy histories.
+ */
+function hashMessageContent(message: AgentMessage): string | undefined {
+	return stableCacheString(message);
+}
+
+function buildConvertedContextCacheKeys(
+	messages: AgentMessage[],
+	context: AgentContext,
+	config: AgentLoopConfig,
+): Pick<ConvertedContextCacheEntry, "messageHashes" | "modelKey" | "toolKey" | "intentTracing"> | undefined {
+	const intentTracing = !!config.intentTracing;
+	const messageHashes = messages.map(hashMessageContent);
+	const modelKey = stableCacheString(config.model);
+	const toolKey = stableCacheString(normalizeTools(context.tools, intentTracing) ?? []);
+	if (messageHashes.some(hash => hash === undefined) || modelKey === undefined || toolKey === undefined) {
+		return undefined;
+	}
+	return {
+		messageHashes: messageHashes as string[],
+		modelKey,
+		toolKey,
+		intentTracing,
+	};
+}
+
+function findStablePrefixLength(previous: string[], next: string[]): number {
+	const max = Math.min(previous.length, next.length);
+	let index = 0;
+	while (index < max && previous[index] === next[index]) index++;
+	return index;
+}
+
+async function convertAndNormalizeMessages(
+	messages: AgentMessage[],
+	context: AgentContext,
+	config: AgentLoopConfig,
+): Promise<Context["messages"]> {
+	const keys = buildConvertedContextCacheKeys(messages, context, config);
+	if (!keys) {
+		return normalizeMessagesForProvider(await config.convertToLlm(messages), config.model);
+	}
+	const previous = convertedContextCache.get(config);
+	const canReuse =
+		previous &&
+		previous.convertToLlm === config.convertToLlm &&
+		previous.transformContext === config.transformContext &&
+		previous.modelKey === keys.modelKey &&
+		previous.toolKey === keys.toolKey &&
+		previous.intentTracing === keys.intentTracing;
+
+	if (canReuse) {
+		const stablePrefixLength = findStablePrefixLength(previous.messageHashes, keys.messageHashes);
+		if (stablePrefixLength === keys.messageHashes.length && stablePrefixLength === previous.messageHashes.length) {
+			return previous.normalizedMessages;
+		}
+		// Append-only fast path: convert only the new suffix and concatenate.
+		// CONTRACT: `convertToLlm` must be per-message (each output message
+		// derived solely from its input message). The bundled converters
+		// satisfy this — they map/filter message-by-message. A converter that
+		// merges adjacent messages or pairs across the suffix boundary would
+		// diverge from a full rebuild; such converters must not be combined
+		// with appendOnlyContext. Covered by the suffix-equivalence test in
+		// agent-loop-context-cache.test.ts.
+		if (
+			config.appendOnlyContext &&
+			stablePrefixLength === previous.messageHashes.length &&
+			keys.messageHashes.length > previous.messageHashes.length
+		) {
+			const suffix = messages.slice(stablePrefixLength);
+			const convertedSuffix = await config.convertToLlm(suffix);
+			const llmMessages = [...previous.llmMessages, ...convertedSuffix];
+			const normalizedMessages = normalizeMessagesForProvider(llmMessages, config.model);
+			convertedContextCache.set(config, {
+				...keys,
+				convertToLlm: config.convertToLlm,
+				transformContext: config.transformContext,
+				llmMessages,
+				normalizedMessages,
+			});
+			return normalizedMessages;
+		}
+	}
+
+	const llmMessages = await config.convertToLlm(messages);
+	const normalizedMessages = normalizeMessagesForProvider(llmMessages, config.model);
+	convertedContextCache.set(config, {
+		...keys,
+		convertToLlm: config.convertToLlm,
+		transformContext: config.transformContext,
+		llmMessages,
+		normalizedMessages,
+	});
+	return normalizedMessages;
+}
+
 export const INTENT_FIELD = "_i";
 
-function injectIntentIntoSchema(schema: unknown, mode: "require" | "optional" = "require"): unknown {
+function injectIntentIntoSchema(schema: unknown, mode: "require" | "optional" = "optional"): unknown {
 	if (!schema || typeof schema !== "object" || Array.isArray(schema)) return schema;
 	const schemaRecord = schema as Record<string, unknown>;
 	const propertiesValue = schemaRecord.properties;
@@ -400,7 +527,7 @@ export function normalizeTools(tools: AgentContext["tools"], injectIntent: boole
 function resolveIntentMode(intent: AgentTool["intent"]): "require" | "optional" | "omit" {
 	if (typeof intent === "function") return "omit";
 	if (intent === "optional" || intent === "omit") return intent;
-	return "require";
+	return intent === "require" ? "require" : "optional";
 }
 
 function extractIntent(args: Record<string, unknown>): { intent?: string; strippedArgs: Record<string, unknown> } {
@@ -711,9 +838,9 @@ async function streamAssistantResponse(
 		messages = await config.transformContext(messages, signal);
 	}
 
-	// Convert to LLM-compatible messages (AgentMessage[] → Message[])
-	const llmMessages = await config.convertToLlm(messages);
-	const normalizedMessages = normalizeMessagesForProvider(llmMessages, config.model);
+	// Convert to LLM-compatible messages (AgentMessage[] → Message[]) and normalize at the LLM boundary.
+	// Cache hits are keyed by provider-visible content hashes, never message object identity.
+	const normalizedMessages = await convertAndNormalizeMessages(messages, context, config);
 
 	// Build LLM context — append-only mode caches system prompt + tools
 	// AND keeps an append-only message log so prior-turn bytes are stable.