agentfootprint 2.11.1 → 2.11.2

package/dist/esm/core/Agent.js

@@ -13,22 +13,10 @@
  *            agentfootprint.stream.tool_start / tool_end
  *            agentfootprint.context.* (via ContextRecorder)
  */
-import { FlowChartExecutor, flowChart, } from 'footprintjs';
-// ArrayMergeMode lives on footprintjs's `advanced` subpath, not its
-// main barrel. Used to set `arrayMerge: Replace` on subflow output
-// mapping for the Tools slot — the slot's deduped tool list must
-// REPLACE the parent's `dynamicToolSchemas` rather than concatenate
-// with it (default behavior re-introduces duplicate tool names that
-// LLM providers reject).
-import { ArrayMergeMode } from 'footprintjs/advanced';
+import { FlowChartExecutor, } from 'footprintjs';
 import { cacheDecisionSubflow } from '../cache/CacheDecisionSubflow.js';
 import { cacheGateDecide, updateSkillHistory as updateSkillHistoryStage, } from '../cache/CacheGateDecider.js';
 import { getDefaultCacheStrategy } from '../cache/strategyRegistry.js';
-import { isPauseRequest } from './pause.js';
-import { emitCostTick } from './cost.js';
-import { STAGE_IDS, SUBFLOW_IDS } from '../conventions.js';
-import { defaultCommentaryTemplates } from '../recorders/observability/commentary/commentaryTemplates.js';
-import { defaultThinkingTemplates } from '../recorders/observability/thinking/thinkingTemplates.js';
 import { ContextRecorder } from '../recorders/core/ContextRecorder.js';
 import { streamRecorder } from '../recorders/core/StreamRecorder.js';
 import { agentRecorder } from '../recorders/core/AgentRecorder.js';
@@ -37,21 +25,24 @@ import { permissionRecorder } from '../recorders/core/PermissionRecorder.js';
 import { evalRecorder } from '../recorders/core/EvalRecorder.js';
 import { memoryRecorder } from '../recorders/core/MemoryRecorder.js';
 import { skillRecorder } from '../recorders/core/SkillRecorder.js';
-import { typedEmit } from '../recorders/core/typedEmit.js';
-import { memoryInjectionKey } from '../memory/define.types.js';
-import { unwrapMemoryFlowChart } from '../memory/define.js';
-import { mountMemoryRead, mountMemoryWrite } from '../memory/wire/mountMemoryPipeline.js';
 import { buildSystemPromptSlot } from './slots/buildSystemPromptSlot.js';
 import { buildMessagesSlot } from './slots/buildMessagesSlot.js';
 import { buildToolsSlot } from './slots/buildToolsSlot.js';
 import { buildInjectionEngineSubflow } from '../lib/injection-engine/buildInjectionEngineSubflow.js';
-import { buildReadSkillTool } from '../lib/injection-engine/skillTools.js';
-import { defineInstruction } from '../lib/injection-engine/factories/defineInstruction.js';
-import { applyOutputFallback, validateCannedAgainstSchema, } from './outputFallback.js';
+import { applyOutputFallback } from './outputFallback.js';
 import { buildCheckpoint, classifyFailurePhase, RunCheckpointError, validateCheckpoint, } from './runCheckpoint.js';
-import { applyOutputSchema, buildDefaultInstruction, OutputSchemaError, } from './outputSchema.js';
+import { applyOutputSchema, OutputSchemaError } from './outputSchema.js';
 import { RunnerBase, makeRunId } from './RunnerBase.js';
-import { clampIterations, safeStringify, validateMemoryIdUniqueness, validateToolNameUniqueness, } from './agent/validators.js';
+import { clampIterations, validateMemoryIdUniqueness, validateToolNameUniqueness, } from './agent/validators.js';
+import { iterationStartStage } from './agent/stages/iterationStart.js';
+import { routeDeciderStage } from './agent/stages/route.js';
+import { buildSeedStage } from './agent/stages/seed.js';
+import { buildCallLLMStage } from './agent/stages/callLLM.js';
+import { buildToolCallsHandler } from './agent/stages/toolCalls.js';
+import { buildAgentChart } from './agent/buildAgentChart.js';
+import { buildToolRegistry } from './agent/buildToolRegistry.js';
+import { AgentBuilder } from './agent/AgentBuilder.js';
+export { AgentBuilder };
 // Public types (AgentOptions, AgentInput, AgentOutput) extracted to
 // ./agent/types.ts and re-exported above (v2.11.1).
 // AgentState extracted to ./agent/types.ts (v2.11.1).
@@ -509,168 +500,33 @@ export class Agent extends RunnerBase {
         const systemPromptCachePolicy = this.systemPromptCachePolicy;
         const cachingDisabled = this.cachingDisabledValue;
         const cacheStrategy = this.cacheStrategy;
-        const seed = (scope) => {
-            const args = scope.$getArgs();
-            scope.userMessage = args.message;
-            // If `resumeOnError(...)` set the side channel, restore the
-            // checkpointed conversation history. The next iteration sees
-            // the prior messages and continues from the failure point.
-            // We always clear the field after reading so subsequent runs
-            // (without resumeOnError) start fresh.
-            if (this.pendingResumeHistory && this.pendingResumeHistory.length > 0) {
-                scope.history = [...this.pendingResumeHistory];
+        // seed extracted to ./agent/stages/seed.ts (v2.11.2). Factory takes
+        // chart-build-time constants + per-run mutable accessors so the
+        // resume side-channel and current run id remain dynamic.
+        // toolSchemas is finalized further down; pass a getter that reads
+        // the eventual const at stage-execution time.
+        let toolSchemasResolved = [];
+        const seed = buildSeedStage({
+            maxIterations,
+            cachingDisabled,
+            get toolSchemas() {
+                return toolSchemasResolved;
+            },
+            consumePendingResumeHistory: () => {
+                const h = this.pendingResumeHistory;
                 this.pendingResumeHistory = undefined;
-            }
-            else {
-                scope.history = [{ role: 'user', content: args.message }];
-            }
-            // Default identity uses the runId so multi-run isolation works
-            // without consumer changes; explicit identity (multi-tenant)
-            // overrides via `agent.run({ identity })`.
-            scope.runIdentity = args.identity ?? {
-                conversationId: this.currentRunContext?.runId ?? 'default',
-            };
-            scope.newMessages = [];
-            scope.turnNumber = 1;
-            // Permissive default — explicit cap will land when PricingTable
-            // gets a context-window field. Memory pickByBudget treats anything
-            // ≥ minimumTokens as "fits", so this just enables the budget path.
-            scope.contextTokensRemaining = 32_000;
-            scope.iteration = 1;
-            scope.maxIterations = maxIterations;
-            scope.finalContent = '';
-            scope.totalInputTokens = 0;
-            scope.totalOutputTokens = 0;
-            scope.turnStartMs = Date.now();
-            scope.systemPromptInjections = [];
-            scope.messagesInjections = [];
-            scope.toolsInjections = [];
-            scope.llmLatestContent = '';
-            scope.llmLatestToolCalls = [];
-            scope.pausedToolCallId = '';
-            scope.pausedToolName = '';
-            scope.pausedToolStartMs = 0;
-            scope.cumTokensInput = 0;
-            scope.cumTokensOutput = 0;
-            scope.cumEstimatedUsd = 0;
-            scope.costBudgetHit = false;
-            scope.activeInjections = [];
-            scope.activatedInjectionIds = [];
-            scope.dynamicToolSchemas = toolSchemas;
-            // Cache layer state (v2.6) — initialized to inert defaults.
-            // CacheDecision subflow populates `cacheMarkers` per iteration;
-            // UpdateSkillHistory + CacheGate consume `cachingDisabled`,
-            // `recentHitRate`, `skillHistory`. Empty defaults mean the
-            // CacheGate falls through to 'apply-markers' on iter 1 (no
-            // history yet → no churn detected; recentHitRate undefined →
-            // hit-rate floor doesn't fire).
-            scope.cacheMarkers = [];
-            scope.cachingDisabled = cachingDisabled;
-            scope.recentHitRate = undefined;
-            scope.skillHistory = [];
-            typedEmit(scope, 'agentfootprint.agent.turn_start', {
-                turnIndex: 0,
-                userPrompt: args.message,
-            });
-        };
-        // Tool registry composition — three sources:
-        //
-        //   1. Static registry: tools registered via `.tool()`. Always
-        //      visible to the LLM; always executable.
-        //   2. `read_skill` (auto-attached when ≥1 Skill is registered):
-        //      activation tool for LLM-guided Skills.
-        //   3. Skill-supplied tools (`Skill.inject.tools[]`): visible only
-        //      when the Skill is active (filtered by tools slot subflow);
-        //      MUST always be in the executor registry so when the LLM
-        //      calls one, the tool-calls handler can dispatch.
-        //
-        // Tool-name uniqueness is enforced across all three sources at
-        // build time. The LLM only sees `tool.schema.name` (no ids), so
-        // names ARE the runtime dispatch key — collisions break the LLM's
-        // ability to call the right tool. Throw early instead of subtly
-        // shadowing.
-        const skills = this.injections.filter((i) => i.flavor === 'skill');
-        // Collect skill tools, deduping by name when the SAME Tool reference
-        // is shared across skills. Different Tool implementations under the
-        // same name throws (already validated upstream by
-        // validateToolNameUniqueness) — we keep the runtime check as
-        // belt-and-suspenders.
-        //
-        // Block C runtime — `autoActivate: 'currentSkill'` semantics:
-        //   When a skill's `defineSkill({ autoActivate: 'currentSkill' })`
-        //   is set, its tools are EXCLUDED from the static registry. They
-        //   flow into the LLM's tool list ONLY through `dynamicSchemas`
-        //   (the buildToolsSlot path that reads activeInjections), which
-        //   means they're visible ONLY on iterations after the skill is
-        //   activated by `read_skill('id')`. Without this, the LLM sees
-        //   every skill's tools on every iteration and the
-        //   per-skill-narrowing autoActivate promised in `defineSkill`
-        //   doesn't actually narrow anything. Skills WITHOUT autoActivate
-        //   keep the v2.4 behavior (tools always visible) for back-compat.
-        const skillToolEntries = [];
-        const sharedSkillTools = new Map();
-        for (const skill of skills) {
-            const meta = skill.metadata;
-            const isAutoActivate = meta?.autoActivate === 'currentSkill';
-            const toolsFromSkill = skill.inject.tools ?? [];
-            for (const tool of toolsFromSkill) {
-                const name = tool.schema.name;
-                const existing = sharedSkillTools.get(name);
-                if (existing) {
-                    if (existing !== tool) {
-                        throw new Error(`Agent: tool name '${name}' is declared by multiple skills with different ` +
-                            `Tool implementations. Skills MAY share the SAME Tool reference; they may ` +
-                            `NOT register different functions under the same name.`);
-                    }
-                    continue; // dedupe — same reference already added
-                }
-                sharedSkillTools.set(name, tool);
-                // autoActivate skills: their tools come ONLY through
-                // dynamicSchemas (buildToolsSlot.ts pulls them from
-                // activeInjections.inject.tools when the skill is active).
-                // Don't pre-load them in the static registry.
-                if (isAutoActivate)
-                    continue;
-                skillToolEntries.push({ name, tool });
-            }
-        }
-        // buildReadSkillTool returns undefined when skills is empty; the
-        // length check below short-circuits so the non-null assertion is safe.
-        const readSkillEntries = skills.length > 0 ? [{ name: 'read_skill', tool: buildReadSkillTool(skills) }] : [];
-        const augmentedRegistry = [
-            ...registry,
-            ...readSkillEntries,
-            ...skillToolEntries,
-        ];
-        // Final cross-source name-uniqueness check: static .tool() vs
-        // read_skill vs (deduped) skill tools. After the dedupe above this
-        // catches collisions BETWEEN sources (e.g., a static .tool('foo')
-        // colliding with a Skill's foo) which are real bugs.
-        const seenNames = new Set();
-        for (const entry of augmentedRegistry) {
-            if (seenNames.has(entry.name)) {
-                throw new Error(`Agent: duplicate tool name '${entry.name}'. Tool names must be unique ` +
-                    `across .tool() registrations and Skills' inject.tools (after deduping ` +
-                    `same-reference shares across skills). The LLM dispatches by name; ` +
-                    `collisions break tool routing.`);
-            }
-            seenNames.add(entry.name);
-        }
-        const registryByName = new Map(augmentedRegistry.map((e) => [e.name, e.tool]));
-        // Block C runtime — autoActivate skill tools live OUTSIDE the LLM-
-        // visible registry (so they don't pollute the per-iteration tool
-        // list before the skill activates), but they MUST still be findable
-        // by the dispatch handler — the LLM calls them by name once the
-        // skill is active, and dispatch looks up by name. Add them to the
-        // dispatch map so `lookupTool` resolves correctly. Using the Map
-        // backing the static registryByName means autoActivate tools share
-        // the same `.execute` wiring as normal tools — no special path.
-        for (const [name, tool] of sharedSkillTools.entries()) {
-            if (!registryByName.has(name)) {
-                registryByName.set(name, tool);
-            }
-        }
-        const toolSchemas = augmentedRegistry.map((e) => e.tool.schema);
+                return h;
+            },
+            getCurrentRunId: () => this.currentRunContext?.runId,
+        });
+        // Tool registry composition extracted to ./agent/buildToolRegistry.ts.
+        // Composes static .tool() registry + auto-attached read_skill +
+        // skill-supplied tools (with autoActivate scoping); validates
+        // name uniqueness; produces the dispatch map.
+        const { registryByName, toolSchemas } = buildToolRegistry(registry, this.injections);
+        // Late-bind toolSchemas into the seed stage's deps (the factory was
+        // built earlier with a getter; this resolves the actual value).
+        toolSchemasResolved = toolSchemas;
         const injectionEngineSubflow = buildInjectionEngineSubflow({
             injections: this.injections,
         });
@@ -683,1064 +539,52 @@ export class Agent extends RunnerBase {
             tools: toolSchemas,
             ...(this.externalToolProvider && { toolProvider: this.externalToolProvider }),
         });
-        const iterationStart = (scope) => {
-            typedEmit(scope, 'agentfootprint.agent.iteration_start', {
-                turnIndex: 0,
-                iterIndex: scope.iteration,
-            });
-        };
-        const callLLM = async (scope) => {
-            const systemPromptInjections = scope.systemPromptInjections ?? [];
-            // `scope.messagesInjections` is read by ContextRecorder for
-            // observability; the LLM-wire path now reads scope.history
-            // directly (see below for rationale).
-            const iteration = scope.iteration;
-            const systemPrompt = systemPromptInjections
-                .map((r) => r.rawContent ?? '')
-                .filter((s) => s.length > 0)
-                .join('\n\n');
-            // Read the LLM message stream from `scope.history` directly.
-            // The `messagesInjections` projection is for observability
-            // (ContextRecorder, Lens) — it flattens InjectionRecords for
-            // event reporting and doesn't carry the full LLM-protocol
-            // shape (assistant `toolCalls[]`, etc.). For Anthropic's API
-            // contract we need the original LLMMessage with `toolCalls`
-            // intact so tool_use → tool_result correlation survives.
-            const messages = scope.history ?? [];
-            typedEmit(scope, 'agentfootprint.stream.llm_start', {
-                iteration,
-                provider: provider.name,
-                model,
-                systemPromptChars: systemPrompt.length,
-                messagesCount: messages.length,
-                toolsCount: toolSchemas.length,
-                ...(temperature !== undefined && { temperature }),
-            });
-            const startMs = Date.now();
-            // Use dynamic schemas — registry tools + injection-supplied
-            // tools (Skills' `inject.tools` when their Injection is active).
-            // Falls back to the static schemas at startup before the tools
-            // slot has run for the first time.
-            const activeToolSchemas = scope.dynamicToolSchemas ?? toolSchemas;
-            const baseRequest = {
-                ...(systemPrompt.length > 0 && { systemPrompt }),
-                messages,
-                ...(activeToolSchemas.length > 0 && { tools: activeToolSchemas }),
-                model,
-                ...(temperature !== undefined && { temperature }),
-                ...(maxTokens !== undefined && { maxTokens }),
-            };
-            // v2.6+ — call cache strategy to attach provider-specific cache
-            // hints. CacheGate has already routed (apply-markers / no-markers)
-            // and populated scope.cacheMarkers accordingly. Strategy.prepareRequest
-            // is a pass-through for empty markers.
-            const cacheMarkers = scope.cacheMarkers ?? [];
-            const cachePrepared = await cacheStrategy.prepareRequest(baseRequest, cacheMarkers, {
-                iteration,
-                iterationsRemaining: Math.max(0, maxIterations - iteration),
-                recentHitRate: scope.recentHitRate,
-                cachingDisabled: scope.cachingDisabled ?? false,
-            });
-            const llmRequest = cachePrepared.request;
-            // Streaming-first: when the provider implements `stream()` we
-            // consume chunk-by-chunk so consumers (Lens commentary, chat
-            // UIs) see tokens as they arrive instead of waiting for the
-            // full LLM call to finish. Each non-terminal chunk fires
-            // `agentfootprint.stream.token` with the token text + index.
-            //
-            // The terminal chunk SHOULD carry the authoritative
-            // `LLMResponse` (toolCalls + usage + stopReason); when it does
-            // we use it directly. When it doesn't (older providers, partial
-            // implementations) we fall back to `complete()` for the
-            // authoritative payload — keeping the ReAct loop deterministic.
-            let response;
-            if (provider.stream) {
-                for await (const chunk of provider.stream(llmRequest)) {
-                    if (chunk.done) {
-                        if (chunk.response)
-                            response = chunk.response;
-                        break;
-                    }
-                    if (chunk.content.length > 0) {
-                        typedEmit(scope, 'agentfootprint.stream.token', {
-                            iteration,
-                            tokenIndex: chunk.tokenIndex,
-                            content: chunk.content,
-                        });
-                    }
-                }
-            }
-            if (!response) {
-                // No `stream()` OR stream finished without a response payload.
-                response = await provider.complete(llmRequest);
-            }
-            const durationMs = Date.now() - startMs;
-            scope.totalInputTokens = scope.totalInputTokens + response.usage.input;
-            scope.totalOutputTokens = scope.totalOutputTokens + response.usage.output;
-            scope.llmLatestContent = response.content;
-            scope.llmLatestToolCalls = response.toolCalls;
-            typedEmit(scope, 'agentfootprint.stream.llm_end', {
-                iteration,
-                content: response.content,
-                toolCallCount: response.toolCalls.length,
-                usage: response.usage,
-                stopReason: response.stopReason,
-                durationMs,
-            });
-            emitCostTick(scope, pricingTable, costBudget, model, response.usage);
-        };
-        /** Decides the next branch: 'tool-calls' or 'final'. */
-        const routeDecider = (scope) => {
-            const toolCalls = scope.llmLatestToolCalls;
-            const iteration = scope.iteration;
-            const chosen = toolCalls.length > 0 && iteration < scope.maxIterations ? 'tool-calls' : 'final';
-            typedEmit(scope, 'agentfootprint.agent.route_decided', {
-                turnIndex: 0,
-                iterIndex: iteration,
-                chosen,
-                rationale: chosen === 'tool-calls'
-                    ? `LLM requested ${toolCalls.length} tool call(s)`
-                    : iteration >= scope.maxIterations
-                        ? 'maxIterations reached — forcing final'
-                        : 'LLM produced no tool calls — final answer',
-            });
-            return chosen;
-        };
-        /**
-         * Pausable tool-call handler.
-         *
-         * `execute` iterates the LLM-requested tool calls. If a tool throws
-         * `PauseRequest` via `pauseHere()`, we save the remaining work into
-         * scope and return the pause data — footprintjs captures a checkpoint
-         * and bubbles it up. The outer `Agent.run()` surfaces it as a
-         * `RunnerPauseOutcome`.
-         *
-         * `resume` is called when the consumer provides the human's answer.
-         * We treat that answer as the paused tool's result and append it to
-         * history, then continue the ReAct iteration loop.
-         */
-        const toolCallsHandler = {
-            execute: async (scope) => {
-                const toolCalls = scope.llmLatestToolCalls;
-                const iteration = scope.iteration;
-                const newHistory = [...scope.history];
-                // ALWAYS push the assistant turn when there are tool calls — even
-                // if the content was empty — so providers (Anthropic, OpenAI) can
-                // round-trip the tool_use blocks via `LLMMessage.toolCalls`.
-                // Without this, the next iteration's request lacks the assistant
-                // turn that initiated the tool call, and the API rejects the
-                // following tool_result with "preceding tool_use missing".
-                if (scope.llmLatestContent || toolCalls.length > 0) {
-                    newHistory.push({
-                        role: 'assistant',
-                        content: scope.llmLatestContent ?? '',
-                        ...(toolCalls.length > 0 && { toolCalls }),
-                    });
-                }
-                // Resolve a tool by name, consulting the external ToolProvider
-                // if one was wired via `.toolProvider()` and the static
-                // registry doesn't carry the tool. The provider sees the same
-                // ctx the Tools slot used, so dispatch + visibility stay
-                // consistent within the iteration.
-                const externalToolProvider = this.externalToolProvider;
-                const lookupTool = (toolName) => {
-                    const fromRegistry = registryByName.get(toolName);
-                    if (fromRegistry)
-                        return fromRegistry;
-                    if (!externalToolProvider)
-                        return undefined;
-                    const activatedIds = scope.activatedInjectionIds ?? [];
-                    const identity = scope.runIdentity;
-                    const ctx = {
-                        iteration: scope.iteration,
-                        ...(activatedIds.length > 0 && {
-                            activeSkillId: activatedIds[activatedIds.length - 1],
-                        }),
-                        ...(identity && { identity }),
-                    };
-                    const visible = externalToolProvider.list(ctx);
-                    return visible.find((t) => t.schema.name === toolName);
-                };
-                for (const tc of toolCalls) {
-                    const tool = lookupTool(tc.name);
-                    typedEmit(scope, 'agentfootprint.stream.tool_start', {
-                        toolName: tc.name,
-                        toolCallId: tc.id,
-                        args: tc.args,
-                        ...(toolCalls.length > 1 && { parallelCount: toolCalls.length }),
-                    });
-                    const startMs = Date.now();
-                    let result;
-                    let error;
-                    // Permission gate — when a checker is configured, evaluate BEFORE
-                    // executing the tool. Emits `permission.check` with the decision.
-                    // On 'deny', the tool is not executed and its result is a
-                    // synthetic denial string; on 'allow'/'gate_open', execution
-                    // proceeds normally (the gate is informational — the consumer's
-                    // checker is responsible for any gate-open side effects).
-                    let denied = false;
-                    if (permissionChecker) {
-                        try {
-                            const decision = await permissionChecker.check({
-                                capability: 'tool_call',
-                                actor: 'agent',
-                                target: tc.name,
-                                context: tc.args,
-                            });
-                            typedEmit(scope, 'agentfootprint.permission.check', {
-                                capability: 'tool_call',
-                                actor: 'agent',
-                                target: tc.name,
-                                result: decision.result,
-                                ...(decision.policyRuleId !== undefined && { policyRuleId: decision.policyRuleId }),
-                                ...(decision.rationale !== undefined && { rationale: decision.rationale }),
-                            });
-                            if (decision.result === 'deny') {
-                                denied = true;
-                                result = `[permission denied: ${decision.rationale ?? 'policy'}]`;
-                            }
-                        }
-                        catch (permErr) {
-                            // A checker that throws is treated as deny-by-default. The
-                            // denial message records the thrown error so consumers can
-                            // debug policy-adapter failures without losing the run.
-                            denied = true;
-                            const msg = permErr instanceof Error ? permErr.message : String(permErr);
-                            typedEmit(scope, 'agentfootprint.permission.check', {
-                                capability: 'tool_call',
-                                actor: 'agent',
-                                target: tc.name,
-                                result: 'deny',
-                                rationale: `permission-checker threw: ${msg}`,
-                            });
-                            result = `[permission denied: checker error: ${msg}]`;
-                        }
-                    }
-                    if (!denied) {
-                        try {
-                            if (!tool)
-                                throw new Error(`Unknown tool: ${tc.name}`);
-                            result = await tool.execute(tc.args, {
-                                toolCallId: tc.id,
-                                iteration,
-                            });
-                        }
-                        catch (err) {
-                            if (isPauseRequest(err)) {
-                                // Commit partial state so resume() can find history intact.
-                                scope.history = newHistory;
-                                scope.pausedToolCallId = tc.id;
-                                scope.pausedToolName = tc.name;
-                                scope.pausedToolStartMs = startMs;
-                                // Returning a defined value triggers footprintjs pause —
-                                // the returned object becomes the checkpoint's pauseData.
-                                return {
-                                    toolCallId: tc.id,
-                                    toolName: tc.name,
-                                    ...(typeof err.data === 'object' && err.data !== null
-                                        ? err.data
-                                        : { data: err.data }),
-                                };
-                            }
-                            error = true;
-                            result = err instanceof Error ? err.message : String(err);
-                        }
-                    }
-                    const durationMs = Date.now() - startMs;
-                    typedEmit(scope, 'agentfootprint.stream.tool_end', {
-                        toolCallId: tc.id,
-                        result,
-                        durationMs,
-                        ...(error === true && { error: true }),
-                    });
-                    const resultStr = typeof result === 'string' ? result : safeStringify(result);
-                    newHistory.push({
-                        role: 'tool',
-                        content: resultStr,
-                        toolCallId: tc.id,
-                        toolName: tc.name,
-                    });
-                    // ── Dynamic ReAct wiring ───────────────────────────────
-                    //
-                    // (1) `lastToolResult` drives `on-tool-return` Injection
-                    //     triggers — the InjectionEngine's NEXT pass will see
-                    //     this and activate any matching Instructions.
-                    scope.lastToolResult = { toolName: tc.name, result: resultStr };
-                    // (2) `read_skill` is the auto-attached activation tool.
-                    //     When the LLM calls it with a valid Skill id, append
-                    //     to `activatedInjectionIds` so the InjectionEngine's
-                    //     NEXT pass activates that Skill (lifetime: turn — stays
-                    //     active until the turn ends).
-                    if (tc.name === 'read_skill' && !error && !denied) {
-                        const requestedId = tc.args.id;
-                        if (typeof requestedId === 'string' && requestedId.length > 0) {
-                            const current = scope.activatedInjectionIds;
-                            if (!current.includes(requestedId)) {
-                                scope.activatedInjectionIds = [...current, requestedId];
-                            }
-                        }
-                    }
-                }
-                scope.history = newHistory;
-                typedEmit(scope, 'agentfootprint.agent.iteration_end', {
-                    turnIndex: 0,
-                    iterIndex: iteration,
-                    toolCallCount: toolCalls.length,
-                    history: scope.history,
-                });
-                scope.iteration = iteration + 1;
-                return undefined; // explicit: no pause, flow continues to loopTo
-            },
-            resume: (scope, input) => {
-                // Consumer-supplied resume input becomes the paused tool's result.
-                // The subflow's pre-pause scope is restored automatically by
-                // footprintjs 4.17.0 via `checkpoint.subflowStates`, so
-                // `scope.history` and `scope.pausedToolCallId` read back cleanly
-                // across same-executor AND cross-executor resume.
-                const toolCallId = scope.pausedToolCallId;
-                const toolName = scope.pausedToolName;
-                const startMs = scope.pausedToolStartMs;
-                const resultStr = typeof input === 'string' ? input : safeStringify(input);
-                const newHistory = [
-                    ...scope.history,
-                    {
-                        role: 'tool',
-                        content: resultStr,
-                        toolCallId,
-                        toolName,
-                    },
-                ];
-                scope.history = newHistory;
-                typedEmit(scope, 'agentfootprint.stream.tool_end', {
-                    toolCallId,
-                    result: input,
-                    durationMs: Date.now() - startMs,
-                });
-                const iteration = scope.iteration;
-                typedEmit(scope, 'agentfootprint.agent.iteration_end', {
-                    turnIndex: 0,
-                    iterIndex: iteration,
-                    toolCallCount: 1,
-                    history: scope.history,
-                });
-                scope.iteration = iteration + 1;
-                // Clear pause checkpoint fields.
-                scope.pausedToolCallId = '';
-                scope.pausedToolName = '';
-                scope.pausedToolStartMs = 0;
-            },
-        };
-        // Final branch is split so memory-write subflows can mount BETWEEN
-        // setting `finalContent` and breaking the ReAct loop. PrepareFinal
-        // captures the turn payload; BreakFinal terminates the loop.
-        const prepareFinalStage = (scope) => {
-            const iteration = scope.iteration;
-            scope.finalContent = scope.llmLatestContent;
-            // The turn payload memory writes persist: the user's message
-            // paired with the agent's final answer.
-            scope.newMessages = [
-                { role: 'user', content: scope.userMessage },
-                { role: 'assistant', content: scope.finalContent },
-            ];
-            typedEmit(scope, 'agentfootprint.agent.iteration_end', {
-                turnIndex: 0,
-                iterIndex: iteration,
-                toolCallCount: 0,
-            });
-            typedEmit(scope, 'agentfootprint.agent.turn_end', {
-                turnIndex: 0,
-                finalContent: scope.finalContent,
-                totalInputTokens: scope.totalInputTokens,
-                totalOutputTokens: scope.totalOutputTokens,
-                iterationCount: iteration,
-                durationMs: Date.now() - scope.turnStartMs,
-            });
-        };
-        const breakFinalStage = (scope) => {
-            // $break terminates the flow before loopTo fires, ending the
-            // ReAct iteration once memory writes (if any) have persisted.
-            scope.$break();
-            return scope.finalContent;
-        };
-        // Compose the final branch as its own subflow so memory write
-        // subflows mount as visible siblings in narrative + Lens.
-        let finalBranchBuilder = flowChart('PrepareFinal', prepareFinalStage, 'prepare-final', undefined, 'Capture turn payload (finalContent + newMessages)');
-        for (const m of this.memories) {
-            if (m.write) {
-                finalBranchBuilder = mountMemoryWrite(finalBranchBuilder, {
-                    pipeline: {
-                        read: unwrapMemoryFlowChart(m.read),
-                        write: unwrapMemoryFlowChart(m.write),
-                    },
-                    identityKey: 'runIdentity',
-                    turnNumberKey: 'turnNumber',
-                    contextTokensKey: 'contextTokensRemaining',
-                    newMessagesKey: 'newMessages',
-                    writeSubflowId: `sf-memory-write-${m.id}`,
-                });
-            }
-        }
-        const finalBranchChart = finalBranchBuilder
-            .addFunction('BreakFinal', breakFinalStage, 'break-final', 'Terminate the ReAct loop')
-            .build();
-        // Description prefix `Agent:` is a taxonomy marker — consumers
-        // (Lens + FlowchartRecorder) detect Agent-primitive subflows via
-        // this prefix and flag them as true agent boundaries (separate
-        // from LLMCall subflows which use `LLMCall:` prefix).
-        let builder = flowChart('Seed', seed, STAGE_IDS.SEED, undefined, 'Agent: ReAct loop');
-        // Memory READ subflows — mounted between Seed and InjectionEngine
-        // for TURN_START timing (default). Each memory writes to its own
-        // scope key (`memoryInjection_${id}`) so multiple `.memory()`
-        // registrations layer without colliding.
-        for (const m of this.memories) {
-            builder = mountMemoryRead(builder, {
-                pipeline: {
-                    read: unwrapMemoryFlowChart(m.read),
-                    ...(m.write !== undefined && { write: unwrapMemoryFlowChart(m.write) }),
-                },
-                identityKey: 'runIdentity',
-                turnNumberKey: 'turnNumber',
-                contextTokensKey: 'contextTokensRemaining',
-                injectionKey: memoryInjectionKey(m.id),
-                readSubflowId: `sf-memory-read-${m.id}`,
-            });
-        }
-        builder = builder
-            // Injection Engine — evaluates every Injection's trigger once
-            // per iteration; writes activeInjections[] to parent scope for
-            // the slot subflows to consume. Skipped if no injections were
-            // registered (no observable difference, just one more no-op
-            // subflow boundary).
-            .addSubFlowChartNext(SUBFLOW_IDS.INJECTION_ENGINE, injectionEngineSubflow, 'Injection Engine', {
-            inputMapper: (parent) => ({
-                iteration: parent.iteration,
-                userMessage: parent.userMessage,
-                history: parent.history,
-                lastToolResult: parent.lastToolResult,
-                activatedInjectionIds: parent.activatedInjectionIds ?? [],
-            }),
-            outputMapper: (sf) => ({ activeInjections: sf.activeInjections }),
-            // CRITICAL: footprintjs's default `applyOutputMapping`
-            // CONCATENATES arrays from subflow output with the parent's
-            // existing array values. Without `Replace`, the parent's
-            // `activeInjections` from iter N gets CONCATENATED with the
-            // subflow's iter N+1 fresh evaluation — producing
-            // 8 → 16 → 24 → 32 cumulative injections per turn instead of
-            // the intended ~8-per-iter.
-            //
-            // The slot subflows below (SystemPrompt, Messages, Tools) all
-            // read `activeInjections` and render every entry, so without
-            // Replace the system prompt grows linearly with iteration
-            // count. This was the root-cause of Dynamic-mode costing
-            // ~2x more input tokens than Classic in the v2.5.0 Neo
-            // benchmarks — the InjectionEngine's intended per-iter
-            // recomposition wasn't happening; it was per-iter ACCUMULATION.
-            arrayMerge: ArrayMergeMode.Replace,
-        })
-            .addSubFlowChartNext(SUBFLOW_IDS.SYSTEM_PROMPT, systemPromptSubflow, 'System Prompt', {
-            inputMapper: (parent) => ({
-                userMessage: parent.userMessage,
-                iteration: parent.iteration,
-                activeInjections: parent.activeInjections,
-            }),
-            outputMapper: (sf) => ({ systemPromptInjections: sf.systemPromptInjections }),
-            // See Tools-subflow comment below — same array-concat hazard.
-            // Without Replace, iter N+1's systemPromptInjections gets
-            // CONCATENATED with iter N's, multiplying the system prompt
-            // each iteration.
-            arrayMerge: ArrayMergeMode.Replace,
-        })
-            .addSubFlowChartNext(SUBFLOW_IDS.MESSAGES, messagesSubflow, 'Messages', {
-            inputMapper: (parent) => ({
-                messages: parent.history,
-                iteration: parent.iteration,
-                activeInjections: parent.activeInjections,
-            }),
-            outputMapper: (sf) => ({ messagesInjections: sf.messagesInjections }),
-            // Same array-concat hazard. messagesInjections is consumer-
-            // facing observability metadata (ContextRecorder, Lens) — must
-            // reflect THIS iteration's history, not be appended to last
-            // iteration's. CallLLM no longer reads this for the wire
-            // request (uses scope.history directly), so the LLM-protocol
-            // bug is fixed independently — but consumers of the
-            // messagesInjections stream still expect the per-iteration
-            // semantics.
-            arrayMerge: ArrayMergeMode.Replace,
-        })
-            .addSubFlowChartNext(SUBFLOW_IDS.TOOLS, toolsSubflow, 'Tools', {
-            inputMapper: (parent) => ({
-                iteration: parent.iteration,
-                activeInjections: parent.activeInjections,
-                // The slot subflow reads these to build the per-iteration
-                // ToolDispatchContext when an external `.toolProvider()` is
-                // configured. Without them the provider sees activeSkillId
-                // = undefined every iteration, breaking skillScopedTools etc.
-                activatedInjectionIds: parent.activatedInjectionIds,
-                runIdentity: parent.runIdentity,
-            }),
-            outputMapper: (sf) => ({
-                toolsInjections: sf.toolsInjections,
-                // Pass merged tool schemas (registry + injection-supplied)
-                // back up so callLLM uses the right list for THIS iteration.
-                dynamicToolSchemas: sf.toolSchemas,
-            }),
-            // CRITICAL: footprintjs's default `applyOutputMapping`
-            // CONCATENATES arrays from subflow output with the parent's
-            // existing array values. Without `Replace`, the parent's
-            // `dynamicToolSchemas` (carrying the iter N value) gets
-            // concatenated with the slot's iter N+1 deduped list,
-            // re-introducing duplicate tool names that Anthropic's API
-            // rejects with "tools: Tool names must be unique." The slot's
-            // toolSchemas IS the authoritative list — replace, don't
-            // concatenate.
-            arrayMerge: ArrayMergeMode.Replace,
-        })
-            // ── Cache layer (v2.6) ─────────────────────────────────────
-            // CacheDecision subflow walks `activeInjections` + evaluates
-            // each `cache:` directive, emits provider-agnostic
-            // `CacheMarker[]` to scope. Pure transform; no IO.
-            //
-            // CRITICAL: arrayMerge: ArrayMergeMode.Replace — same lesson
-            // as the v2.5.1 InjectionEngine fix. The default footprintjs
-            // behavior CONCATENATES arrays from child to parent;
-            // `cacheMarkers` MUST replace each iteration, not accumulate.
-            .addSubFlowChartNext(SUBFLOW_IDS.CACHE_DECISION, cacheDecisionSubflow, 'CacheDecision', {
-            inputMapper: (parent) => ({
-                activeInjections: parent.activeInjections ?? [],
-                iteration: parent.iteration ?? 1,
-                maxIterations: parent.maxIterations ?? maxIterations,
-                userMessage: parent.userMessage ?? '',
-                ...(parent.lastToolResult !== undefined && {
-                    lastToolName: parent.lastToolResult?.toolName,
-                }),
-                cumulativeInputTokens: parent.totalInputTokens ?? 0,
-                systemPromptCachePolicy,
-                cachingDisabled: parent.cachingDisabled ?? false,
-            }),
-            outputMapper: (sf) => ({ cacheMarkers: sf.cacheMarkers }),
-            arrayMerge: ArrayMergeMode.Replace,
-        })
-            .addFunction('UpdateSkillHistory', updateSkillHistoryStage, STAGE_IDS.UPDATE_SKILL_HISTORY, 'Update skill-history rolling window for CacheGate churn detection')
-            .addDeciderFunction('CacheGate', cacheGateDecide, STAGE_IDS.CACHE_GATE, 'Gate cache-marker application: kill switch / hit-rate / skill-churn')
-            .addFunctionBranch(STAGE_IDS.APPLY_MARKERS, 'ApplyMarkers', 
-        // Pass-through stage — markers stay in scope as-is.
-        // BuildLLMRequest (Phase 7+) reads them on the next stage.
-        () => undefined, 'Proceed with cache markers from CacheDecision')
-            .addFunctionBranch(STAGE_IDS.SKIP_CACHING, 'SkipCaching', 
-        // Clear markers so BuildLLMRequest sees an empty list and
-        // makes the request unmodified.
-        (scope) => {
-            scope.cacheMarkers = [];
-        }, 'Skip caching this iteration')
-            .end()
-            .addFunction('IterationStart', iterationStart, 'iteration-start', 'Iteration begin marker')
-            .addFunction('CallLLM', callLLM, STAGE_IDS.CALL_LLM, 'LLM invocation')
-            .addDeciderFunction('Route', routeDecider, SUBFLOW_IDS.ROUTE, 'ReAct routing')
-            .addPausableFunctionBranch('tool-calls', 'ToolCalls', toolCallsHandler, 'Tool execution (pausable via pauseHere)')
-            .addSubFlowChartBranch('final', finalBranchChart, 'Final', {
-            // Pass through the read-only state the sub-chart needs;
-            // OMIT keys the sub-chart writes (finalContent, newMessages)
-            // — passing those via inputMapper would freeze them as args.
-            inputMapper: (parent) => {
-                const { finalContent: _f, newMessages: _nm, ...rest } = parent;
-                void _f;
-                void _nm;
-                return rest;
+        // iterationStart extracted to ./agent/stages/iterationStart.ts (v2.11.2).
+        const iterationStart = iterationStartStage;
+        // callLLM extracted to ./agent/stages/callLLM.ts (v2.11.2). Same
+        // late-binding pattern as seed for toolSchemas (computed below).
+        const callLLM = buildCallLLMStage({
+            provider,
+            model,
+            ...(temperature !== undefined && { temperature }),
+            ...(maxTokens !== undefined && { maxTokens }),
+            ...(pricingTable !== undefined && { pricingTable }),
+            ...(costBudget !== undefined && { costBudget }),
+            maxIterations,
+            cacheStrategy,
+            get toolSchemas() {
+                return toolSchemasResolved;
             },
-            outputMapper: (sf) => ({
-                finalContent: sf.finalContent,
-            }),
-            // BreakFinal's $break() must reach the outer loopTo so the
-            // ReAct iteration terminates; without this the inner break
-            // only exits the sub-chart and the outer loop continues.
-            propagateBreak: true,
-        })
-            .setDefault('final')
-            .end()
-            // Dynamic ReAct: loop back to the InjectionEngine so EVERY iteration
-            // re-evaluates triggers (rule predicates, on-tool-return, llm-activated)
-            // against the freshest context (the just-appended tool result).
-            // Without this, the InjectionEngine runs ONCE per turn and:
-            //   - on-tool-return predicates never fire on iter 2+
-            //   - read_skill('X') activations are never picked up next iteration
-            //   - autoActivate per-skill tool gating is structurally impossible
-            //   - tools / system-prompt slots stay frozen at iter 1 content
-            // The v2.4 default of loopTo(MESSAGES) bypassed all four — quietly
-            // breaking the framework's "Dynamic ReAct" claim. v2.5 restores the
-            // v1 behavior that documents promise.
-            .loopTo(SUBFLOW_IDS.INJECTION_ENGINE);
-        return builder.build();
-    }
-}
-/**
- * Fluent builder. `tool()` accepts any Tool<TArgs, TResult> and registers
- * it by its schema.name. Duplicate names throw at build time.
- */
-export class AgentBuilder {
-    opts;
-    systemPromptValue = '';
-    /**
-     * Cache policy for the base system prompt. Set via the optional
-     * 2nd argument to `.system(text, { cache })`. Default `'always'` —
-     * the base prompt is stable per-turn and an ideal cache anchor.
-     */
-    systemPromptCachePolicy = 'always';
-    /**
-     * Global cache kill switch. Set via `Agent.create({ caching: 'off' })`
-     * (handled in `AgentOptions` propagation). Defaults to `false`
-     * (caching enabled). When `true`, the CacheGate decider routes to
-     * `'no-markers'` every iteration regardless of other rules.
-     */
-    cachingDisabledValue = false;
-    /**
-     * Optional explicit CacheStrategy override. Default: undefined,
-     * which means the agent auto-resolves from
-     * `getDefaultCacheStrategy(provider.name)` at construction. Power
-     * users override here for custom backends or test mocks.
-     */
-    cacheStrategyOverride;
-    registry = [];
-    injectionList = [];
-    memoryList = [];
-    /**
-     * Optional terminal contract — see `outputSchema()`. Stored on the
-     * builder, propagated to the Agent at `.build()` time.
-     */
-    outputSchemaParser;
-    /** 3-tier output fallback chain — set via `.outputFallback({...})`.
-     *  Optional; absent = current throw-on-validation-failure behavior. */
-    outputFallbackCfg;
-    /**
-     * Optional `ToolProvider` set via `.toolProvider()`. Propagated to
-     * the Agent's Tools slot subflow + tool-call dispatcher; consulted
-     * per iteration so dynamic chains (`gatedTools`, `skillScopedTools`)
-     * react to current activation state.
-     */
-    toolProviderRef;
-    /**
-     * Optional override for `AgentOptions.maxIterations`. When set via
-     * the `.maxIterations()` builder method, takes precedence over the
-     * value passed to `Agent.create({ maxIterations })`.
-     */
-    maxIterationsOverride;
-    /**
-     * Recorders collected via `.recorder()`. Attached to the built Agent
-     * before `build()` returns (each via `agent.attach(rec)`).
-     */
-    recorderList = [];
-    // Voice config — defaults until the consumer calls .appName() /
-    // .commentaryTemplates() / .thinkingTemplates(). Stored as plain
-    // dicts (Record<string, string>) so the builder doesn't depend on
-    // the template-engine modules at compile time; the runtime types
-    // come from the agentfootprint barrel exports.
-    appNameValue = 'Chatbot';
-    commentaryOverrides = {};
-    thinkingOverrides = {};
-    constructor(opts) {
-        this.opts = opts;
-        // Cache layer: opts.caching === 'off' propagates to scope's
-        // `cachingDisabled` kill switch read by CacheGate. opts.cacheStrategy
-        // overrides the registry-resolved default.
-        if (opts.caching === 'off')
-            this.cachingDisabledValue = true;
-        if (opts.cacheStrategy !== undefined)
-            this.cacheStrategyOverride = opts.cacheStrategy;
-    }
-    /**
-     * Set the base system prompt.
-     *
-     * @param prompt - The system prompt text. Stable per-turn.
-     * @param options - Optional config. `cache` controls how the
-     *   CacheDecision subflow treats this prompt block:
-     *   - `'always'` (default) — cache the base prompt as a stable
-     *     prefix anchor. Highest cache-hit rate; recommended for
-     *     production agents whose system prompt rarely changes.
-     *   - `'never'` — skip caching. Use if the prompt contains volatile
-     *     content (timestamps, per-request user IDs).
-     *   - `'while-active'` — semantically equivalent to `'always'` for
-     *     the base prompt (it's always active by definition).
-     *   - `{ until }` — conditional invalidation (e.g., flush after iter 5).
-     */
-    system(prompt, options) {
-        this.systemPromptValue = prompt;
-        if (options?.cache !== undefined) {
-            this.systemPromptCachePolicy = options.cache;
-        }
-        return this;
-    }
-    tool(tool) {
-        const name = tool.schema.name;
-        if (this.registry.some((e) => e.name === name)) {
-            throw new Error(`Agent.tool(): duplicate tool name '${name}'`);
-        }
-        this.registry.push({ name, tool: tool });
-        return this;
-    }
-    /**
-     * Register many tools at once. Convenience for tool sources that
-     * return a list (e.g., `await mcpClient(...).tools()`). Each tool
-     * is registered via `.tool()` so duplicate-name validation still
-     * fires per-entry.
-     */
-    tools(tools) {
-        for (const t of tools)
-            this.tool(t);
-        return this;
-    }
-    /**
-     * Wire a chainable `ToolProvider` (from `agentfootprint/tool-providers`)
-     * as the agent's per-iteration tool source.
-     *
-     * The provider is consulted EVERY iteration via `provider.list(ctx)`
-     * with `ctx = { iteration, activeSkillId, identity }`. Tools the
-     * provider emits flow into the Tools slot alongside any static
-     * tools registered via `.tool()` / `.tools()`. The tool-call
-     * dispatcher also consults the provider so dynamic chains
-     * (`gatedTools`, `skillScopedTools`) dispatch correctly when their
-     * visible-set changes mid-turn.
-     *
-     * Throws if called more than once on the same builder (avoids
-     * silent override surprises).
-     *
-     * @example  Permission-gated baseline
-     *   import { gatedTools, staticTools } from 'agentfootprint/tool-providers';
-     *   import { PermissionPolicy } from 'agentfootprint/security';
-     *
-     *   const policy = PermissionPolicy.fromRoles({
-     *     readonly: ['lookup', 'list_skills', 'read_skill'],
-     *     admin:    ['lookup', 'list_skills', 'read_skill', 'delete'],
-     *   }, 'readonly');
-     *
-     *   const provider = gatedTools(
-     *     staticTools(allTools),
-     *     (toolName) => policy.isAllowed(toolName),
-     *   );
-     *
-     *   const agent = Agent.create({ provider: llm, model })
-     *     .system('You answer.')
-     *     .toolProvider(provider)
-     *     .build();
-     */
-    toolProvider(provider) {
-        if (this.toolProviderRef) {
-            throw new Error('AgentBuilder.toolProvider: already set. Each agent has at most one external ToolProvider.');
-        }
-        this.toolProviderRef = provider;
-        return this;
-    }
-    /**
-     * Override the ReAct iteration cap set via `Agent.create({
-     * maxIterations })`. Convenience for builder-style code that prefers
-     * fluent setters over constructor opts. Last call wins.
-     *
-     * Throws if `n` is not a positive integer or exceeds the hard cap
-     * (`clampIterations`'s upper bound).
-     */
-    maxIterations(n) {
-        if (!Number.isInteger(n) || n <= 0) {
-            throw new Error(`AgentBuilder.maxIterations: expected a positive integer, got ${n}.`);
-        }
-        this.maxIterationsOverride = n;
-        return this;
-    }
-    /**
-     * Attach a footprintjs `CombinedRecorder` to the built Agent. Wired
-     * via `agent.attach(rec)` immediately after construction, so the
-     * recorder sees every event from the very first run.
-     *
-     * Equivalent to calling `agent.attach(rec)` post-build; the builder
-     * method is a convenience for codebases that prefer fully-fluent
-     * agent assembly. Multiple recorders are supported (each gets its
-     * own `attach()` call).
-     */
-    recorder(rec) {
-        this.recorderList.push(rec);
-        return this;
-    }
-    /**
-     * Set the agent's display name — substituted as `{{appName}}` in
-     * commentary + thinking templates. Same place to brand a tenant
-     * ("Acme Bot"), distinguish multi-agent roles ("Triage" vs
-     * "Reviewer"), or localize ("Asistente"). Default: `'Chatbot'`.
-     */
-    appName(name) {
-        this.appNameValue = name;
-        return this;
-    }
-    /**
-     * Override agentfootprint's bundled commentary templates. Spread on
-     * top of `defaultCommentaryTemplates`; missing keys fall back. Same
-     * `Record<string, string>` shape with `{{vars}}` substitution as
-     * the bundled defaults — see `defaultCommentaryTemplates` for the
-     * full key list.
-     *
-     * Use cases: i18n (`'agent.turn_start': 'El usuario...'`), brand
-     * voice ("You: {{userPrompt}}"), per-tenant customization.
-     */
-    commentaryTemplates(templates) {
-        this.commentaryOverrides = { ...this.commentaryOverrides, ...templates };
-        return this;
-    }
-    /**
-     * Override agentfootprint's bundled thinking templates. Same
-     * contract shape as commentary; different vocabulary — first-person
-     * status the chat bubble shows mid-call. Per-tool overrides go via
-     * `tool.<toolName>` keys (e.g., `'tool.weather': 'Looking up the
-     * weather…'`). See `defaultThinkingTemplates` for the full key list.
-     */
-    thinkingTemplates(templates) {
-        this.thinkingOverrides = { ...this.thinkingOverrides, ...templates };
-        return this;
-    }
-    // ─── Injection sugar — context engineering surface ───────────
-    //
-    // ALL of these push into the same `injectionList`. The Injection
-    // primitive is identical across flavors; the methods are just
-    // narrative-friendly aliases. Duplicate ids throw at build time.
-    /**
-     * Register any `Injection`. Use this for power-user / custom flavors;
-     * for built-in flavors use the typed sugar (`.skill`, `.steering`,
-     * `.instruction`, `.fact`).
-     */
-    injection(injection) {
-        if (this.injectionList.some((i) => i.id === injection.id)) {
-            throw new Error(`Agent.injection(): duplicate id '${injection.id}'`);
-        }
-        this.injectionList.push(injection);
-        return this;
-    }
-    /**
-     * Register a Skill — LLM-activated, system-prompt + tools.
-     * Auto-attaches the `read_skill` activation tool to the agent.
-     * Skill stays active for the rest of the turn once activated.
-     */
-    skill(injection) {
-        return this.injection(injection);
-    }
-    /**
-     * Bulk-register every Skill in a `SkillRegistry`. Use for shared
-     * skill catalogs across multiple Agents — register skills once on
-     * the registry; attach the same registry to every consumer Agent.
-     *
-     * @example
-     *   const registry = new SkillRegistry();
-     *   registry.register(billingSkill).register(refundSkill);
-     *   const supportAgent = Agent.create({ provider }).skills(registry).build();
-     *   const escalationAgent = Agent.create({ provider }).skills(registry).build();
-     */
-    skills(registry) {
-        for (const skill of registry.list())
-            this.injection(skill);
-        return this;
-    }
-    /**
-     * Register a Steering doc — always-on system-prompt rule.
-     * Use for invariant guidance: output format, persona, safety policies.
-     */
-    steering(injection) {
-        return this.injection(injection);
-    }
-    /**
-     * Register an Instruction — rule-based system-prompt guidance.
-     * Predicate runs each iteration. Use for context-dependent rules
-     * including the "Dynamic ReAct" `on-tool-return` pattern.
-     */
-    instruction(injection) {
-        return this.injection(injection);
-    }
-    /**
-     * Bulk-register many instructions at once. Convenience for consumer
-     * code that organizes its instruction set in a flat array (`const
-     * instructions = [outputFormat, dataRouting, ...]`). Each element
-     * is registered via `.instruction()` so duplicate-id checks still
-     * fire per-entry.
-     */
-    instructions(injections) {
-        for (const i of injections)
-            this.instruction(i);
-        return this;
-    }
-    /**
-     * Register a Fact — developer-supplied data the LLM should see.
-     * User profile, env info, computed summary, current time, …
-     * Distinct from Skills (LLM-activated guidance) and Steering
-     * (always-on rules) in INTENT — the engine treats them all alike.
-     */
-    fact(injection) {
-        return this.injection(injection);
-    }
-    /**
-     * Register a Memory subsystem — load/persist conversation context,
-     * facts, narrative beats, or causal snapshots across runs.
-     *
-     * The `MemoryDefinition` is produced by `defineMemory({ type, strategy,
-     * store })`. Multiple memories layer cleanly via per-id scope keys
-     * (`memoryInjection_${id}`):
-     *
-     * ```ts
-     * Agent.create({ provider })
-     *   .memory(defineMemory({ id: 'short', type: MEMORY_TYPES.EPISODIC,
-     *                          strategy: { kind: MEMORY_STRATEGIES.WINDOW, size: 10 },
-     *                          store }))
-     *   .memory(defineMemory({ id: 'facts', type: MEMORY_TYPES.SEMANTIC,
-     *                          strategy: { kind: MEMORY_STRATEGIES.EXTRACT,
-     *                                      extractor: 'pattern' }, store }))
-     *   .build();
-     * ```
-     *
-     * The READ subflow runs at the configured `timing` (default
-     * `MEMORY_TIMING.TURN_START`) and writes its formatted output to the
-     * `memoryInjection_${id}` scope key for the slot subflows to consume.
-     */
-    memory(definition) {
-        if (this.memoryList.some((m) => m.id === definition.id)) {
-            throw new Error(`Agent.memory(): duplicate id '${definition.id}' — each memory needs a unique id ` +
-                'to keep its scope key (`memoryInjection_${id}`) collision-free.');
-        }
-        this.memoryList.push(definition);
-        return this;
-    }
-    /**
-     * Register a RAG retriever — semantic search over a vector-indexed
-     * corpus. Identical plumbing to `.memory()` (RAG resolves to a
-     * `MemoryDefinition` produced by `defineRAG()`); this alias exists
-     * so the consumer's intent reads clearly:
-     *
-     * ```ts
-     * agent
-     *   .memory(shortTermConversation)   // remembers what the USER said
-     *   .rag(productDocs)                // retrieves what the CORPUS says
-     *   .build();
-     * ```
-     *
-     * Both end up as memory subflows, but the alias separates "user
-     * conversation memory" from "document corpus retrieval" in code
-     * intent, ids, and Lens chips.
-     */
-    rag(definition) {
-        return this.memory(definition);
-    }
-    /**
-     * Declarative terminal contract. The agent's final answer must be
-     * JSON matching `parser`. Auto-injects a system-prompt instruction
-     * telling the LLM the shape, and exposes `agent.runTyped()` /
-     * `agent.parseOutput()` for parse + validate at the call site.
-     *
-     * The `parser` is duck-typed: any object with a `parse(unknown): T`
-     * method works (Zod, Valibot, ArkType, hand-written). The optional
-     * `description` field on the parser drives the auto-generated
-     * instruction; consumers can also override via `opts.instruction`.
-     *
-     * Throws if called more than once on the same builder (avoids
-     * silent override surprises).
-     *
-     * @param parser  Validation strategy that throws on shape failure.
-     * @param opts    Optional `{ name, instruction }` to customize.
-     *
-     * @example
-     *   import { z } from 'zod';
-     *   const Output = z.object({
-     *     status: z.enum(['ok', 'err']),
-     *     items: z.array(z.string()),
-     *   }).describe('A status enum + an array of strings.');
-     *
-     *   const agent = Agent.create({...})
-     *     .outputSchema(Output)
-     *     .build();
-     *
-     *   const typed = await agent.runTyped({ message: '...' });
-     *   typed.status; // narrowed to 'ok' | 'err'
-     */
-    outputSchema(parser, opts) {
-        if (this.outputSchemaParser) {
-            throw new Error('AgentBuilder.outputSchema: already set. Each agent has at most one terminal contract.');
-        }
-        this.outputSchemaParser = parser;
-        const instructionText = opts?.instruction ?? buildDefaultInstruction(parser);
-        const id = opts?.name ?? 'output-schema';
-        // Always-on system-slot instruction. Activates every iteration so
-        // long runs keep the contract present (recency-first redundancy).
-        this.injectionList.push(defineInstruction({
-            id,
-            activeWhen: () => true,
-            prompt: instructionText,
-        }));
-        return this;
-    }
-    /**
-     * 3-tier degradation for output-schema validation failures. Pairs
-     * with `.outputSchema()` — calling `.outputFallback()` without an
-     * `outputSchema` first throws (the fallback has nothing to validate).
-     *
-     * Three tiers:
-     *
-     *   1. **Primary** — LLM emitted schema-valid JSON. Caller gets it.
-     *   2. **Fallback** — `OutputSchemaError` thrown. The async
-     *      `fallback(error, raw)` runs; its return is re-validated.
-     *   3. **Canned** — static safety-net value. NEVER throws when set.
-     *
-     * `canned` is validated against the schema at builder time —
-     * fail-fast on misconfig (a `canned` that doesn't validate would
-     * defeat the fail-open guarantee).
-     *
-     * Two typed events fire on tier transitions for observability:
-     *   - `agentfootprint.resilience.output_fallback_triggered`
-     *   - `agentfootprint.resilience.output_canned_used`
-     *
-     * @example
-     * ```ts
-     * import { z } from 'zod';
-     * const Refund = z.object({ amount: z.number(), reason: z.string() });
-     *
-     * const agent = Agent.create({...})
-     *   .outputSchema(Refund)
-     *   .outputFallback({
-     *     fallback: async (err, raw) => ({ amount: 0, reason: 'manual review' }),
-     *     canned:   { amount: 0, reason: 'unable to process' },
-     *   })
-     *   .build();
-     * ```
-     */
-    outputFallback(options) {
-        if (!this.outputSchemaParser) {
-            throw new Error('AgentBuilder.outputFallback: call .outputSchema(parser) FIRST. ' +
-                'outputFallback supplements outputSchema; one without the other is incoherent.');
-        }
-        if (this.outputFallbackCfg) {
-            throw new Error('AgentBuilder.outputFallback: already set. Each agent has at most one fallback chain.');
-        }
-        // Build-time validation — canned MUST satisfy the schema.
-        if (options.canned !== undefined) {
-            validateCannedAgainstSchema(options.canned, this.outputSchemaParser);
-        }
-        this.outputFallbackCfg = {
-            fallback: options.fallback,
-            ...(options.canned !== undefined && { canned: options.canned }),
-            hasCanned: options.canned !== undefined,
-        };
-        return this;
-    }
-    build() {
-        // Resolve the voice config: bundled defaults + consumer overrides.
-        // Templates flow through the same barrel exports the rest of the
-        // library uses, so a future locale-pack swap is a single import.
-        const voice = {
-            appName: this.appNameValue,
-            commentaryTemplates: { ...defaultCommentaryTemplates, ...this.commentaryOverrides },
-            thinkingTemplates: { ...defaultThinkingTemplates, ...this.thinkingOverrides },
-        };
-        const opts = this.maxIterationsOverride !== undefined
-            ? { ...this.opts, maxIterations: this.maxIterationsOverride }
-            : this.opts;
-        const agent = new Agent(opts, this.systemPromptValue, this.registry, voice, this.injectionList, this.memoryList, this.outputSchemaParser, this.toolProviderRef, this.systemPromptCachePolicy, this.cachingDisabledValue, this.cacheStrategyOverride, this.outputFallbackCfg);
-        // Attach builder-collected recorders so they receive events from
-        // the very first run. Mirrors what consumers would do post-build
-        // via `agent.attach(rec)`; the builder method is purely sugar.
-        for (const rec of this.recorderList) {
-            agent.attach(rec);
-        }
-        return agent;
+        });
+        // routeDecider extracted to ./agent/stages/route.ts (v2.11.2).
+        const routeDecider = routeDeciderStage;
+        // toolCallsHandler extracted to ./agent/stages/toolCalls.ts (v2.11.2).
+        const toolCallsHandler = buildToolCallsHandler({
+            registryByName,
+            ...(this.externalToolProvider && { externalToolProvider: this.externalToolProvider }),
+            ...(permissionChecker && { permissionChecker }),
+        });
+        // Chart composition extracted to ./agent/buildAgentChart.ts (v2.11.2).
+        return buildAgentChart({
+            memories: this.memories,
+            systemPromptCachePolicy,
+            maxIterations,
+            seed,
+            iterationStart,
+            callLLM,
+            routeDecider,
+            toolCallsHandler,
+            injectionEngineSubflow,
+            systemPromptSubflow,
+            messagesSubflow,
+            toolsSubflow,
+            cacheDecisionSubflow,
+            updateSkillHistoryStage,
+            cacheGateDecide,
+        });
     }
 }
+// AgentBuilder extracted to ./agent/AgentBuilder.ts (v2.11.2).
+// Re-export so the 28+ existing import sites continue to work unchanged.
 // Validators + helpers extracted to ./agent/validators.ts (v2.11.1).
 //# sourceMappingURL=Agent.js.map