@warmdrift/kgauto-compiler 2.0.0-alpha.32 → 2.0.0-alpha.34

package/dist/{chunk-WXCFWUCN.mjs → chunk-XJZRDGHF.mjs}

@@ -328,7 +328,8 @@ var STARTER_CHAINS_GROUNDED = {
     { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Reasoning + cost balance \u2014 engineer pick" },
     { id: "claude-opus-4-7", grounding: "judgment", reason: 'Same-provider walk-UP on 429 (rare exception to "always cheaper")' },
     { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor" },
-    { id: "deepseek-v4-pro", grounding: "judgment", reason: "Tier 3 cost floor \u2014 no brain evidence yet" }
+    { id: "deepseek-v4-pro", grounding: "judgment", reason: "Tier 3 cost floor \u2014 no brain evidence yet" },
+    { id: "gpt-5.4", grounding: "judgment", reason: "alpha.33: third-provider tail (archetypePerf=7) \u2014 closes openai-in-default-fallback-chains" }
   ],
   // Quality + cost match.
   generate: [
@@ -359,7 +360,8 @@ var STARTER_CHAINS_GROUNDED = {
     { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Quality anchor \u2014 engineer pick" },
     { id: "claude-haiku-4-5", grounding: "judgment", reason: "Same-provider step-down" },
     { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor" },
-    { id: "gemini-2.5-flash", grounding: "judgment", reason: "Cost floor \u2014 forgiving archetype tolerates Flash" }
+    { id: "gemini-2.5-flash", grounding: "judgment", reason: "Cost floor \u2014 forgiving archetype tolerates Flash" },
+    { id: "gpt-5.4-mini", grounding: "judgment", reason: "alpha.33: third-provider tail (archetypePerf=7) \u2014 closes openai-in-default-fallback-chains" }
   ],
   // Parallel-tool throughput champion — Flash leads on the L-040 cliff
   // (capability-fact: Flash 15-75 parallel calls/step vs DeepSeek 7-8).
@@ -367,21 +369,24 @@ var STARTER_CHAINS_GROUNDED = {
     { id: "gemini-2.5-flash", grounding: "capability-fact", reason: "L-040 parallel-tool throughput champion (15-75 calls/step)" },
     { id: "gemini-2.5-pro", grounding: "capability-fact", reason: "Cross-provider tier 1 with strong parallel-tool support" },
     { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Quality safety net for blocked-Flash case" },
-    { id: "claude-haiku-4-5", grounding: "judgment", reason: "Reduced tool budget \u2014 cliff at 16 fires" }
+    { id: "claude-haiku-4-5", grounding: "judgment", reason: "Reduced tool budget \u2014 cliff at 16 fires" },
+    { id: "gpt-5.4-mini", grounding: "judgment", reason: "alpha.33: third-provider tail \u2014 OpenAI parallel-tool capable archetype" }
   ],
   // Cost-sensitive + tolerant. DeepSeek brain-evidence tier 1.
   summarize: [
     { id: "gemini-2.5-flash", grounding: "judgment", reason: "Cost-sensitive primary \u2014 engineer pick" },
     { id: "deepseek-v4-flash", grounding: "measured", reason: "Brain-validated tier 1 for cost-sensitive summarize workloads", n: 169 },
     { id: "claude-haiku-4-5", grounding: "judgment", reason: "Quality safety net" },
-    { id: "gemini-2.5-flash-lite", grounding: "judgment", reason: "Emergency floor \u2014 onboarded s22, no brain evidence yet" }
+    { id: "gemini-2.5-flash-lite", grounding: "judgment", reason: "Emergency floor \u2014 onboarded s22, no brain evidence yet" },
+    { id: "gpt-5.4-nano", grounding: "judgment", reason: "alpha.33: third-provider tail (archetypePerf=6) \u2014 cheapest OpenAI, matches summarize cost-band" }
   ],
   // Brain-validated DeepSeek tier 1 (169 rows, 0% empty rate).
   classify: [
     { id: "gemini-2.5-flash", grounding: "judgment", reason: "Cost-sensitive primary \u2014 engineer pick" },
     { id: "deepseek-v4-flash", grounding: "measured", reason: "Brain-validated tier 1 (169 rows, 0% empty rate)", n: 169 },
     { id: "claude-haiku-4-5", grounding: "judgment", reason: "Quality safety net" },
-    { id: "gemini-2.5-flash-lite", grounding: "judgment", reason: "Cache-discount 10\xD7 floor for repeat-prompt workloads" }
+    { id: "gemini-2.5-flash-lite", grounding: "judgment", reason: "Cache-discount 10\xD7 floor for repeat-prompt workloads" },
+    { id: "gpt-5.4-nano", grounding: "judgment", reason: "alpha.33: third-provider tail (archetypePerf=6) \u2014 cheapest OpenAI for classify" }
   ]
 };
 var STARTER_CHAINS = (() => {
@@ -436,7 +441,7 @@ function resolveStarterForMode(archetype, toolOrchestration, allChains) {
   return allChains[archetype];
 }
 function getDefaultFallbackChain(opts) {
-  const { archetype, primary, maxDepth = 3, policy, reachability, toolOrchestration } = opts;
+  const { archetype, primary, maxDepth = 4, policy, reachability, toolOrchestration } = opts;
   if (maxDepth < 1) {
     throw new Error(
       `getDefaultFallbackChain: maxDepth must be >= 1, got ${maxDepth}`

package/dist/glassbox/index.d.mts

@@ -1,6 +1,6 @@
-import { G as GlassboxEvent } from '../types-BjrIFPGe.mjs';
-export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-BjrIFPGe.mjs';
-import '../ir-De2AQtlr.mjs';
+import { G as GlassboxEvent } from '../types-BoqPXoUb.mjs';
+export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-BoqPXoUb.mjs';
+import '../ir-BAAHLfFL.mjs';
 import '../dialect.mjs';
 
 /**

package/dist/glassbox/index.d.ts

@@ -1,6 +1,6 @@
-import { G as GlassboxEvent } from '../types-D_JAhCv4.js';
-export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-D_JAhCv4.js';
-import '../ir-BIAT9gJk.js';
+import { G as GlassboxEvent } from '../types-CXyKj6-K.js';
+export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-CXyKj6-K.js';
+import '../ir-B_ygNURd.js';
 import '../dialect.js';
 
 /**

package/dist/glassbox-routes/index.d.mts

@@ -1,5 +1,5 @@
-import { G as GlassboxEvent } from '../types-BjrIFPGe.mjs';
-import { h as Adapter, u as SectionKind } from '../ir-De2AQtlr.mjs';
+import { G as GlassboxEvent } from '../types-BoqPXoUb.mjs';
+import { i as Adapter, v as SectionKind } from '../ir-BAAHLfFL.mjs';
 import '../dialect.mjs';
 
 /**

package/dist/glassbox-routes/index.d.ts

@@ -1,5 +1,5 @@
-import { G as GlassboxEvent } from '../types-D_JAhCv4.js';
-import { h as Adapter, u as SectionKind } from '../ir-BIAT9gJk.js';
+import { G as GlassboxEvent } from '../types-CXyKj6-K.js';
+import { i as Adapter, v as SectionKind } from '../ir-B_ygNURd.js';
 import '../dialect.js';
 
 /**

package/dist/glassbox-routes/index.js

@@ -1414,7 +1414,8 @@ var STARTER_CHAINS_GROUNDED = {
     { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Reasoning + cost balance \u2014 engineer pick" },
     { id: "claude-opus-4-7", grounding: "judgment", reason: 'Same-provider walk-UP on 429 (rare exception to "always cheaper")' },
     { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor" },
-    { id: "deepseek-v4-pro", grounding: "judgment", reason: "Tier 3 cost floor \u2014 no brain evidence yet" }
+    { id: "deepseek-v4-pro", grounding: "judgment", reason: "Tier 3 cost floor \u2014 no brain evidence yet" },
+    { id: "gpt-5.4", grounding: "judgment", reason: "alpha.33: third-provider tail (archetypePerf=7) \u2014 closes openai-in-default-fallback-chains" }
   ],
   // Quality + cost match.
   generate: [
@@ -1445,7 +1446,8 @@ var STARTER_CHAINS_GROUNDED = {
     { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Quality anchor \u2014 engineer pick" },
     { id: "claude-haiku-4-5", grounding: "judgment", reason: "Same-provider step-down" },
     { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor" },
-    { id: "gemini-2.5-flash", grounding: "judgment", reason: "Cost floor \u2014 forgiving archetype tolerates Flash" }
+    { id: "gemini-2.5-flash", grounding: "judgment", reason: "Cost floor \u2014 forgiving archetype tolerates Flash" },
+    { id: "gpt-5.4-mini", grounding: "judgment", reason: "alpha.33: third-provider tail (archetypePerf=7) \u2014 closes openai-in-default-fallback-chains" }
   ],
   // Parallel-tool throughput champion — Flash leads on the L-040 cliff
   // (capability-fact: Flash 15-75 parallel calls/step vs DeepSeek 7-8).
@@ -1453,21 +1455,24 @@ var STARTER_CHAINS_GROUNDED = {
     { id: "gemini-2.5-flash", grounding: "capability-fact", reason: "L-040 parallel-tool throughput champion (15-75 calls/step)" },
     { id: "gemini-2.5-pro", grounding: "capability-fact", reason: "Cross-provider tier 1 with strong parallel-tool support" },
     { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Quality safety net for blocked-Flash case" },
-    { id: "claude-haiku-4-5", grounding: "judgment", reason: "Reduced tool budget \u2014 cliff at 16 fires" }
+    { id: "claude-haiku-4-5", grounding: "judgment", reason: "Reduced tool budget \u2014 cliff at 16 fires" },
+    { id: "gpt-5.4-mini", grounding: "judgment", reason: "alpha.33: third-provider tail \u2014 OpenAI parallel-tool capable archetype" }
   ],
   // Cost-sensitive + tolerant. DeepSeek brain-evidence tier 1.
   summarize: [
     { id: "gemini-2.5-flash", grounding: "judgment", reason: "Cost-sensitive primary \u2014 engineer pick" },
     { id: "deepseek-v4-flash", grounding: "measured", reason: "Brain-validated tier 1 for cost-sensitive summarize workloads", n: 169 },
     { id: "claude-haiku-4-5", grounding: "judgment", reason: "Quality safety net" },
-    { id: "gemini-2.5-flash-lite", grounding: "judgment", reason: "Emergency floor \u2014 onboarded s22, no brain evidence yet" }
+    { id: "gemini-2.5-flash-lite", grounding: "judgment", reason: "Emergency floor \u2014 onboarded s22, no brain evidence yet" },
+    { id: "gpt-5.4-nano", grounding: "judgment", reason: "alpha.33: third-provider tail (archetypePerf=6) \u2014 cheapest OpenAI, matches summarize cost-band" }
   ],
   // Brain-validated DeepSeek tier 1 (169 rows, 0% empty rate).
   classify: [
     { id: "gemini-2.5-flash", grounding: "judgment", reason: "Cost-sensitive primary \u2014 engineer pick" },
     { id: "deepseek-v4-flash", grounding: "measured", reason: "Brain-validated tier 1 (169 rows, 0% empty rate)", n: 169 },
     { id: "claude-haiku-4-5", grounding: "judgment", reason: "Quality safety net" },
-    { id: "gemini-2.5-flash-lite", grounding: "judgment", reason: "Cache-discount 10\xD7 floor for repeat-prompt workloads" }
+    { id: "gemini-2.5-flash-lite", grounding: "judgment", reason: "Cache-discount 10\xD7 floor for repeat-prompt workloads" },
+    { id: "gpt-5.4-nano", grounding: "judgment", reason: "alpha.33: third-provider tail (archetypePerf=6) \u2014 cheapest OpenAI for classify" }
   ]
 };
 var STARTER_CHAINS = (() => {
@@ -1522,7 +1527,7 @@ function resolveStarterForMode(archetype, toolOrchestration, allChains) {
   return allChains[archetype];
 }
 function getDefaultFallbackChain(opts) {
-  const { archetype, primary, maxDepth = 3, policy, reachability, toolOrchestration } = opts;
+  const { archetype, primary, maxDepth = 4, policy, reachability, toolOrchestration } = opts;
   if (maxDepth < 1) {
     throw new Error(
       `getDefaultFallbackChain: maxDepth must be >= 1, got ${maxDepth}`

package/dist/glassbox-routes/index.mjs

@@ -1,7 +1,7 @@
 import {
   ARCHETYPE_FLOOR_DEFAULT,
   getDefaultFallbackChain
-} from "../chunk-WXCFWUCN.mjs";
+} from "../chunk-XJZRDGHF.mjs";
 import {
   tryGetProfile
 } from "../chunk-JQGRWJZO.mjs";

package/dist/index.d.mts

@@ -1,5 +1,5 @@
-import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SectionRewrite, R as RecordInput, e as RecordOutcomeInput, O as OutcomeResult, f as OracleScore, g as CompileResult, B as BestPracticeAdvisory, h as Adapter, i as PerAxisMetrics, j as Provider, k as ChainEntry, G as Grounding } from './ir-De2AQtlr.mjs';
-export { l as CallAttempt, m as CallError, n as ChainWithGrounding, o as Constraints, F as FallbackReason, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, p as MutationApplied, q as NormalizedTokens, r as OutcomeKind, s as PerAxisMetricsByModel, t as PromptSection, u as SectionKind, T as ToolCall, v as ToolDefinition } from './ir-De2AQtlr.mjs';
+import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, B as BestPracticeAdvisory, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-BAAHLfFL.mjs';
+export { m as CallAttempt, n as CallError, o as ChainWithGrounding, p as Constraints, F as FallbackReason, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, q as MutationApplied, r as NormalizedTokens, s as OutcomeKind, t as PerAxisMetricsByModel, u as PromptSection, v as SectionKind, T as ToolCall, w as ToolDefinition } from './ir-BAAHLfFL.mjs';
 import { ModelProfile } from './profiles.mjs';
 export { ALIASES, CacheStrategy, CliffRule, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, allProfiles, getProfile, profilesByProvider, tryGetProfile } from './profiles.mjs';
 import { IntentArchetypeName } from './dialect.mjs';
@@ -54,6 +54,15 @@ interface ExecuteOptions {
     apiKeys?: ApiKeys;
     fetchImpl?: typeof fetch;
     providerOverrides?: ProviderOverrides;
+    /**
+     * alpha.34. When supplied AND the chosen provider supports streaming, the
+     * execute() per-provider branch switches to SSE wire format and invokes
+     * `onChunk(delta)` once per text-delta event. Single-shot behavior
+     * preserved when omitted (current consumers unaffected). Wired from
+     * `CallOptions.onChunk` by `call()` after gating on `profile.streaming`
+     * and `!opts.noStream`.
+     */
+    onChunk?: (chunk: string) => void;
 }
 interface ExecuteOk {
     ok: true;
@@ -94,6 +103,89 @@ declare function execute(request: CompiledRequest, opts?: ExecuteOptions): Promi
  */
 declare function call(ir: PromptIR, opts?: CallOptions): Promise<CallResult>;
 
+/**
+ * alpha.33 — AI-SDK streamText helpers.
+ *
+ * The cross-consumer pattern surfaced 2026-05-20 (`streamText-cache-marker-
+ * propagation-gap` on `command-center/interfaces/kgauto.md ## Requested`):
+ * AI-SDK consumers using `streamText({ system: <string>, messages:
+ * convertToModelMessages(messages), model })` silently drop the Anthropic
+ * `cache_control` markers the compiler emits on `result.request.messages`.
+ * Two losses on that line:
+ *   (a) `convertToModelMessages()` reads raw inputs not the lowered output,
+ *       dropping per-message providerOptions; and
+ *   (b) `streamText({ system: '<string>' })` strips any providerOptions
+ *       on the system prefix.
+ *
+ * Net effect prior to this surface: `cache_read_input_tokens = 0` across
+ * every measured production call by IC + tt-intel even though both
+ * consumers had `historyCachePolicy: { strategy: 'all-but-latest' }` wired
+ * since alpha.5. The compiler did the right thing; the consumer integration
+ * pattern lost the markers at the wire boundary.
+ *
+ * This helper closes the gap. Pass the consumer's already-converted messages
+ * + the kgauto CompileResult and get back a `{ system, messages }` pair
+ * with cacheControl re-attached at the deterministic indices the compiler
+ * computed (`diagnostics.historyCacheMarkIndex` + `systemMessages` with
+ * its own cacheControl markers).
+ *
+ * No I/O. Pure transform. Idempotent — calling twice with the same inputs
+ * produces structurally-identical output.
+ */
+
+/**
+ * AI-SDK-shaped converted-message structure. Matches what
+ * `convertToModelMessages()` returns at the wire layer. We don't import
+ * the AI-SDK type to avoid pulling that dep into the kgauto-compiler
+ * package — consumers pass their own already-converted messages.
+ */
+interface AISDKConvertedMessage {
+    role: 'system' | 'user' | 'assistant' | 'tool';
+    content: unknown;
+    providerOptions?: {
+        anthropic?: {
+            cacheControl?: {
+                type: 'ephemeral';
+            };
+        };
+    };
+}
+/**
+ * Result of `attachCacheControlToStreamTextInput()` — ready to spread into
+ * `streamText(...)` directly.
+ *
+ * The `system` field is either the structured `SystemModelMessage[]` form
+ * (when the compiler computed at least one cacheable section + chose
+ * Anthropic) or the flat-string concatenation (when no cacheable sections
+ * exist OR the provider isn't Anthropic). Consumers can pass either form
+ * to `streamText({ system })` — the AI-SDK accepts both.
+ */
+interface AttachCacheControlResult {
+    system: string | SystemModelMessage[];
+    messages: AISDKConvertedMessage[];
+}
+/**
+ * Attach Anthropic cache_control markers to a streamText input that the
+ * consumer has already prepared via `convertToModelMessages()`. Reads
+ * `result.diagnostics.historyCacheMarkIndex` for the position-on-history
+ * marker and `result.systemMessages` for the per-section system markers.
+ *
+ * The flat-string `system` form is returned when:
+ *   - the compiler produced no `systemMessages` (e.g., zero sections kept), or
+ *   - the chosen provider is not Anthropic, so cacheControl wouldn't
+ *     achieve anything at the wire layer (other providers cache implicitly
+ *     based on prefix shape, no marker needed).
+ *
+ * Otherwise the structured `SystemModelMessage[]` form is returned —
+ * preserves the per-section cacheControl markers across `streamText`.
+ *
+ * @param result - kgauto CompileResult from `compile(ir, opts)`
+ * @param convertedMessages - the consumer's already-converted messages
+ *   (from `convertToModelMessages(rawMessages)` in their AI-SDK call site)
+ * @returns `{ system, messages }` ready to spread into `streamText(...)`
+ */
+declare function attachCacheControlToStreamTextInput(result: CompileResult, convertedMessages: AISDKConvertedMessage[]): AttachCacheControlResult;
+
 /**
  * alpha.11 — opt-in nested config for brain-query mode (chains / archetype
  * perf / pricing / models registry). Enabled by default when endpoint is
@@ -1419,4 +1511,4 @@ declare const loadAliasesFromBrain: () => Record<string, string>;
  */
 declare function compile(ir: PromptIR, opts?: CompileOptions): CompileResult;
 
-export { ABSOLUTE_FLOOR, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, BestPracticeAdvisory, type BrainConfig, type BrainQueryConfig, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type FallbackPosture, type GetActionableAdvisoriesOptions, type GetDefaultFallbackChainOpts, type GetPerAxisMetricsOpts, Grounding, IntentArchetypeName, type LLMJudgeOptions, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, PROVIDER_ENV_KEYS, PerAxisMetrics, type PricingRow, type ProfileToRowOptions, PromptIR, Provider, ProviderOverrides, type ProviderReachability, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, SectionRewrite, type SupportedProvider, TRANSLATOR_FLOOR, applySectionRewrites, buildLLMJudge, call, clearBrain, compile, configureBrain, countTokens, execute, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStarterChain, getStarterChainWithGrounding, isBrainQueryActiveFor, isModelReachable, isProviderReachable, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, markAdvisoryResolved, profileToRow, record, recordOutcome, resetTokenizer, resolvePricingAt, resolveProviderKey, runAdvisor, setTokenizer };
+export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, type AttachCacheControlResult, BestPracticeAdvisory, type BrainConfig, type BrainQueryConfig, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type FallbackPosture, type GetActionableAdvisoriesOptions, type GetDefaultFallbackChainOpts, type GetPerAxisMetricsOpts, Grounding, IntentArchetypeName, type LLMJudgeOptions, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, PROVIDER_ENV_KEYS, PerAxisMetrics, type PricingRow, type ProfileToRowOptions, PromptIR, Provider, ProviderOverrides, type ProviderReachability, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, SectionRewrite, type SupportedProvider, SystemModelMessage, TRANSLATOR_FLOOR, applySectionRewrites, attachCacheControlToStreamTextInput, buildLLMJudge, call, clearBrain, compile, configureBrain, countTokens, execute, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStarterChain, getStarterChainWithGrounding, isBrainQueryActiveFor, isModelReachable, isProviderReachable, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, markAdvisoryResolved, profileToRow, record, recordOutcome, resetTokenizer, resolvePricingAt, resolveProviderKey, runAdvisor, setTokenizer };

package/dist/index.d.ts

@@ -1,5 +1,5 @@
-import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SectionRewrite, R as RecordInput, e as RecordOutcomeInput, O as OutcomeResult, f as OracleScore, g as CompileResult, B as BestPracticeAdvisory, h as Adapter, i as PerAxisMetrics, j as Provider, k as ChainEntry, G as Grounding } from './ir-BIAT9gJk.js';
-export { l as CallAttempt, m as CallError, n as ChainWithGrounding, o as Constraints, F as FallbackReason, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, p as MutationApplied, q as NormalizedTokens, r as OutcomeKind, s as PerAxisMetricsByModel, t as PromptSection, u as SectionKind, T as ToolCall, v as ToolDefinition } from './ir-BIAT9gJk.js';
+import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, B as BestPracticeAdvisory, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-B_ygNURd.js';
+export { m as CallAttempt, n as CallError, o as ChainWithGrounding, p as Constraints, F as FallbackReason, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, q as MutationApplied, r as NormalizedTokens, s as OutcomeKind, t as PerAxisMetricsByModel, u as PromptSection, v as SectionKind, T as ToolCall, w as ToolDefinition } from './ir-B_ygNURd.js';
 import { ModelProfile } from './profiles.js';
 export { ALIASES, CacheStrategy, CliffRule, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, allProfiles, getProfile, profilesByProvider, tryGetProfile } from './profiles.js';
 import { IntentArchetypeName } from './dialect.js';
@@ -54,6 +54,15 @@ interface ExecuteOptions {
     apiKeys?: ApiKeys;
     fetchImpl?: typeof fetch;
     providerOverrides?: ProviderOverrides;
+    /**
+     * alpha.34. When supplied AND the chosen provider supports streaming, the
+     * execute() per-provider branch switches to SSE wire format and invokes
+     * `onChunk(delta)` once per text-delta event. Single-shot behavior
+     * preserved when omitted (current consumers unaffected). Wired from
+     * `CallOptions.onChunk` by `call()` after gating on `profile.streaming`
+     * and `!opts.noStream`.
+     */
+    onChunk?: (chunk: string) => void;
 }
 interface ExecuteOk {
     ok: true;
@@ -94,6 +103,89 @@ declare function execute(request: CompiledRequest, opts?: ExecuteOptions): Promi
  */
 declare function call(ir: PromptIR, opts?: CallOptions): Promise<CallResult>;
 
+/**
+ * alpha.33 — AI-SDK streamText helpers.
+ *
+ * The cross-consumer pattern surfaced 2026-05-20 (`streamText-cache-marker-
+ * propagation-gap` on `command-center/interfaces/kgauto.md ## Requested`):
+ * AI-SDK consumers using `streamText({ system: <string>, messages:
+ * convertToModelMessages(messages), model })` silently drop the Anthropic
+ * `cache_control` markers the compiler emits on `result.request.messages`.
+ * Two losses on that line:
+ *   (a) `convertToModelMessages()` reads raw inputs not the lowered output,
+ *       dropping per-message providerOptions; and
+ *   (b) `streamText({ system: '<string>' })` strips any providerOptions
+ *       on the system prefix.
+ *
+ * Net effect prior to this surface: `cache_read_input_tokens = 0` across
+ * every measured production call by IC + tt-intel even though both
+ * consumers had `historyCachePolicy: { strategy: 'all-but-latest' }` wired
+ * since alpha.5. The compiler did the right thing; the consumer integration
+ * pattern lost the markers at the wire boundary.
+ *
+ * This helper closes the gap. Pass the consumer's already-converted messages
+ * + the kgauto CompileResult and get back a `{ system, messages }` pair
+ * with cacheControl re-attached at the deterministic indices the compiler
+ * computed (`diagnostics.historyCacheMarkIndex` + `systemMessages` with
+ * its own cacheControl markers).
+ *
+ * No I/O. Pure transform. Idempotent — calling twice with the same inputs
+ * produces structurally-identical output.
+ */
+
+/**
+ * AI-SDK-shaped converted-message structure. Matches what
+ * `convertToModelMessages()` returns at the wire layer. We don't import
+ * the AI-SDK type to avoid pulling that dep into the kgauto-compiler
+ * package — consumers pass their own already-converted messages.
+ */
+interface AISDKConvertedMessage {
+    role: 'system' | 'user' | 'assistant' | 'tool';
+    content: unknown;
+    providerOptions?: {
+        anthropic?: {
+            cacheControl?: {
+                type: 'ephemeral';
+            };
+        };
+    };
+}
+/**
+ * Result of `attachCacheControlToStreamTextInput()` — ready to spread into
+ * `streamText(...)` directly.
+ *
+ * The `system` field is either the structured `SystemModelMessage[]` form
+ * (when the compiler computed at least one cacheable section + chose
+ * Anthropic) or the flat-string concatenation (when no cacheable sections
+ * exist OR the provider isn't Anthropic). Consumers can pass either form
+ * to `streamText({ system })` — the AI-SDK accepts both.
+ */
+interface AttachCacheControlResult {
+    system: string | SystemModelMessage[];
+    messages: AISDKConvertedMessage[];
+}
+/**
+ * Attach Anthropic cache_control markers to a streamText input that the
+ * consumer has already prepared via `convertToModelMessages()`. Reads
+ * `result.diagnostics.historyCacheMarkIndex` for the position-on-history
+ * marker and `result.systemMessages` for the per-section system markers.
+ *
+ * The flat-string `system` form is returned when:
+ *   - the compiler produced no `systemMessages` (e.g., zero sections kept), or
+ *   - the chosen provider is not Anthropic, so cacheControl wouldn't
+ *     achieve anything at the wire layer (other providers cache implicitly
+ *     based on prefix shape, no marker needed).
+ *
+ * Otherwise the structured `SystemModelMessage[]` form is returned —
+ * preserves the per-section cacheControl markers across `streamText`.
+ *
+ * @param result - kgauto CompileResult from `compile(ir, opts)`
+ * @param convertedMessages - the consumer's already-converted messages
+ *   (from `convertToModelMessages(rawMessages)` in their AI-SDK call site)
+ * @returns `{ system, messages }` ready to spread into `streamText(...)`
+ */
+declare function attachCacheControlToStreamTextInput(result: CompileResult, convertedMessages: AISDKConvertedMessage[]): AttachCacheControlResult;
+
 /**
  * alpha.11 — opt-in nested config for brain-query mode (chains / archetype
  * perf / pricing / models registry). Enabled by default when endpoint is
@@ -1419,4 +1511,4 @@ declare const loadAliasesFromBrain: () => Record<string, string>;
  */
 declare function compile(ir: PromptIR, opts?: CompileOptions): CompileResult;
 
-export { ABSOLUTE_FLOOR, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, BestPracticeAdvisory, type BrainConfig, type BrainQueryConfig, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type FallbackPosture, type GetActionableAdvisoriesOptions, type GetDefaultFallbackChainOpts, type GetPerAxisMetricsOpts, Grounding, IntentArchetypeName, type LLMJudgeOptions, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, PROVIDER_ENV_KEYS, PerAxisMetrics, type PricingRow, type ProfileToRowOptions, PromptIR, Provider, ProviderOverrides, type ProviderReachability, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, SectionRewrite, type SupportedProvider, TRANSLATOR_FLOOR, applySectionRewrites, buildLLMJudge, call, clearBrain, compile, configureBrain, countTokens, execute, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStarterChain, getStarterChainWithGrounding, isBrainQueryActiveFor, isModelReachable, isProviderReachable, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, markAdvisoryResolved, profileToRow, record, recordOutcome, resetTokenizer, resolvePricingAt, resolveProviderKey, runAdvisor, setTokenizer };
+export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, type AttachCacheControlResult, BestPracticeAdvisory, type BrainConfig, type BrainQueryConfig, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type FallbackPosture, type GetActionableAdvisoriesOptions, type GetDefaultFallbackChainOpts, type GetPerAxisMetricsOpts, Grounding, IntentArchetypeName, type LLMJudgeOptions, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, PROVIDER_ENV_KEYS, PerAxisMetrics, type PricingRow, type ProfileToRowOptions, PromptIR, Provider, ProviderOverrides, type ProviderReachability, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, SectionRewrite, type SupportedProvider, SystemModelMessage, TRANSLATOR_FLOOR, applySectionRewrites, attachCacheControlToStreamTextInput, buildLLMJudge, call, clearBrain, compile, configureBrain, countTokens, execute, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStarterChain, getStarterChainWithGrounding, isBrainQueryActiveFor, isModelReachable, isProviderReachable, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, markAdvisoryResolved, profileToRow, record, recordOutcome, resetTokenizer, resolvePricingAt, resolveProviderKey, runAdvisor, setTokenizer };