@miller-tech/uap 1.72.1 → 1.74.0

package/dist/utils/concurrency-pool.d.ts

@@ -33,8 +33,9 @@
 export declare function concurrentMap<T, R>(items: T[], fn: (item: T, index: number) => Promise<R>, options?: {
     /** Maximum concurrent operations. Overrides auto-detection. */
     maxConcurrent?: number;
-    /** 'cpu' reserves cores for OS/inference, 'io' allows higher concurrency */
-    mode?: 'cpu' | 'io';
+    /** 'cpu' reserves cores for OS/inference, 'io' allows higher concurrency,
+     *  'model' caps to the inference backend's model-slot budget */
+    mode?: 'cpu' | 'io' | 'model';
 }): Promise<R[]>;
 /**
  * Map over items with bounded concurrency, settling all promises.
@@ -46,6 +47,6 @@ export declare function concurrentMap<T, R>(items: T[], fn: (item: T, index: num
  */
 export declare function concurrentMapSettled<T, R>(items: T[], fn: (item: T, index: number) => Promise<R>, options?: {
     maxConcurrent?: number;
-    mode?: 'cpu' | 'io';
+    mode?: 'cpu' | 'io' | 'model';
 }): Promise<PromiseSettledResult<R>[]>;
 //# sourceMappingURL=concurrency-pool.d.ts.map

package/dist/utils/concurrency-pool.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"concurrency-pool.d.ts","sourceRoot":"","sources":["../../src/utils/concurrency-pool.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAIH;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,wBAAsB,aAAa,CAAC,CAAC,EAAE,CAAC,EACtC,KAAK,EAAE,CAAC,EAAE,EACV,EAAE,EAAE,CAAC,IAAI,EAAE,CAAC,EAAE,KAAK,EAAE,MAAM,KAAK,OAAO,CAAC,CAAC,CAAC,EAC1C,OAAO,CAAC,EAAE;IACR,+DAA+D;IAC/D,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,4EAA4E;IAC5E,IAAI,CAAC,EAAE,KAAK,GAAG,IAAI,CAAC;CACrB,GACA,OAAO,CAAC,CAAC,EAAE,CAAC,CAkBd;AAED;;;;;;;GAOG;AACH,wBAAsB,oBAAoB,CAAC,CAAC,EAAE,CAAC,EAC7C,KAAK,EAAE,CAAC,EAAE,EACV,EAAE,EAAE,CAAC,IAAI,EAAE,CAAC,EAAE,KAAK,EAAE,MAAM,KAAK,OAAO,CAAC,CAAC,CAAC,EAC1C,OAAO,CAAC,EAAE;IACR,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,IAAI,CAAC,EAAE,KAAK,GAAG,IAAI,CAAC;CACrB,GACA,OAAO,CAAC,oBAAoB,CAAC,CAAC,CAAC,EAAE,CAAC,CAuBpC"}
+{"version":3,"file":"concurrency-pool.d.ts","sourceRoot":"","sources":["../../src/utils/concurrency-pool.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAKH;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,wBAAsB,aAAa,CAAC,CAAC,EAAE,CAAC,EACtC,KAAK,EAAE,CAAC,EAAE,EACV,EAAE,EAAE,CAAC,IAAI,EAAE,CAAC,EAAE,KAAK,EAAE,MAAM,KAAK,OAAO,CAAC,CAAC,CAAC,EAC1C,OAAO,CAAC,EAAE;IACR,+DAA+D;IAC/D,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;oEACgE;IAChE,IAAI,CAAC,EAAE,KAAK,GAAG,IAAI,GAAG,OAAO,CAAC;CAC/B,GACA,OAAO,CAAC,CAAC,EAAE,CAAC,CAoBd;AAED;;;;;;;GAOG;AACH,wBAAsB,oBAAoB,CAAC,CAAC,EAAE,CAAC,EAC7C,KAAK,EAAE,CAAC,EAAE,EACV,EAAE,EAAE,CAAC,IAAI,EAAE,CAAC,EAAE,KAAK,EAAE,MAAM,KAAK,OAAO,CAAC,CAAC,CAAC,EAC1C,OAAO,CAAC,EAAE;IACR,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,IAAI,CAAC,EAAE,KAAK,GAAG,IAAI,GAAG,OAAO,CAAC;CAC/B,GACA,OAAO,CAAC,oBAAoB,CAAC,CAAC,CAAC,EAAE,CAAC,CAyBpC"}

package/dist/utils/concurrency-pool.js

@@ -7,6 +7,7 @@
  * Uses getMaxParallel() for auto-detection with UAP_MAX_PARALLEL env override.
  */
 import { getMaxParallel } from './system-resources.js';
+import { getMaxModelConcurrency } from './model-slots.js';
 /**
  * Map over items with bounded concurrency.
  *
@@ -34,7 +35,8 @@ import { getMaxParallel } from './system-resources.js';
 export async function concurrentMap(items, fn, options) {
     if (items.length === 0)
         return [];
-    const max = options?.maxConcurrent ?? getMaxParallel(options?.mode ?? 'io');
+    const max = options?.maxConcurrent ??
+        (options?.mode === 'model' ? getMaxModelConcurrency() : getMaxParallel(options?.mode ?? 'io'));
     const results = new Array(items.length);
     let nextIndex = 0;
     const worker = async () => {
@@ -58,7 +60,8 @@ export async function concurrentMap(items, fn, options) {
 export async function concurrentMapSettled(items, fn, options) {
     if (items.length === 0)
         return [];
-    const max = options?.maxConcurrent ?? getMaxParallel(options?.mode ?? 'io');
+    const max = options?.maxConcurrent ??
+        (options?.mode === 'model' ? getMaxModelConcurrency() : getMaxParallel(options?.mode ?? 'io'));
     const results = new Array(items.length);
     let nextIndex = 0;
     const worker = async () => {

package/dist/utils/concurrency-pool.js.map

@@ -1 +1 @@
-{"version":3,"file":"concurrency-pool.js","sourceRoot":"","sources":["../../src/utils/concurrency-pool.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,EAAE,cAAc,EAAE,MAAM,uBAAuB,CAAC;AAEvD;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,MAAM,CAAC,KAAK,UAAU,aAAa,CACjC,KAAU,EACV,EAA0C,EAC1C,OAKC;IAED,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,CAAC;IAElC,MAAM,GAAG,GAAG,OAAO,EAAE,aAAa,IAAI,cAAc,CAAC,OAAO,EAAE,IAAI,IAAI,IAAI,CAAC,CAAC;IAC5E,MAAM,OAAO,GAAQ,IAAI,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;IAC7C,IAAI,SAAS,GAAG,CAAC,CAAC;IAElB,MAAM,MAAM,GAAG,KAAK,IAAmB,EAAE;QACvC,OAAO,SAAS,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC;YAChC,MAAM,CAAC,GAAG,SAAS,EAAE,CAAC;YACtB,OAAO,CAAC,CAAC,CAAC,GAAG,MAAM,EAAE,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;QACrC,CAAC;IACH,CAAC,CAAC;IAEF,MAAM,WAAW,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,MAAM,CAAC,CAAC;IAChD,MAAM,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,WAAW,EAAE,EAAE,GAAG,EAAE,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC;IAEvE,OAAO,OAAO,CAAC;AACjB,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACxC,KAAU,EACV,EAA0C,EAC1C,OAGC;IAED,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,CAAC;IAElC,MAAM,GAAG,GAAG,OAAO,EAAE,aAAa,IAAI,cAAc,CAAC,OAAO,EAAE,IAAI,IAAI,IAAI,CAAC,CAAC;IAC5E,MAAM,OAAO,GAA8B,IAAI,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;IACnE,IAAI,SAAS,GAAG,CAAC,CAAC;IAElB,MAAM,MAAM,GAAG,KAAK,IAAmB,EAAE;QACvC,OAAO,SAAS,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC;YAChC,MAAM,CAAC,GAAG,SAAS,EAAE,CAAC;YACtB,IAAI,CAAC;gBACH,MAAM,KAAK,GAAG,MAAM,EAAE,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;gBACpC,OAAO,CAAC,CAAC,CAAC,GAAG,EAAE,MAAM,EAAE,WAAW,EAAE,KAAK,EAAE,CAAC;YAC9C,CAAC;YAAC,OAAO,MAAM,EAAE,CAAC;gBAChB,OAAO,CAAC,CAAC,CAAC,GAAG,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,EAAE,CAAC;YAC9C,CAAC;QACH,CAAC;IACH,CAAC,CAAC;IAEF,MAAM,WAAW,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,MAAM,CAAC,CAAC;IAChD,MAAM,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,WAAW,EAAE,EAAE,GAAG,EAAE,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC;IAEvE,OAAO,OAAO,CAAC;AACjB,CAAC"}
+{"version":3,"file":"concurrency-pool.js","sourceRoot":"","sources":["../../src/utils/concurrency-pool.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,EAAE,cAAc,EAAE,MAAM,uBAAuB,CAAC;AACvD,OAAO,EAAE,sBAAsB,EAAE,MAAM,kBAAkB,CAAC;AAE1D;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,MAAM,CAAC,KAAK,UAAU,aAAa,CACjC,KAAU,EACV,EAA0C,EAC1C,OAMC;IAED,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,CAAC;IAElC,MAAM,GAAG,GACP,OAAO,EAAE,aAAa;QACtB,CAAC,OAAO,EAAE,IAAI,KAAK,OAAO,CAAC,CAAC,CAAC,sBAAsB,EAAE,CAAC,CAAC,CAAC,cAAc,CAAC,OAAO,EAAE,IAAI,IAAI,IAAI,CAAC,CAAC,CAAC;IACjG,MAAM,OAAO,GAAQ,IAAI,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;IAC7C,IAAI,SAAS,GAAG,CAAC,CAAC;IAElB,MAAM,MAAM,GAAG,KAAK,IAAmB,EAAE;QACvC,OAAO,SAAS,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC;YAChC,MAAM,CAAC,GAAG,SAAS,EAAE,CAAC;YACtB,OAAO,CAAC,CAAC,CAAC,GAAG,MAAM,EAAE,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;QACrC,CAAC;IACH,CAAC,CAAC;IAEF,MAAM,WAAW,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,MAAM,CAAC,CAAC;IAChD,MAAM,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,WAAW,EAAE,EAAE,GAAG,EAAE,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC;IAEvE,OAAO,OAAO,CAAC;AACjB,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACxC,KAAU,EACV,EAA0C,EAC1C,OAGC;IAED,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,CAAC;IAElC,MAAM,GAAG,GACP,OAAO,EAAE,aAAa;QACtB,CAAC,OAAO,EAAE,IAAI,KAAK,OAAO,CAAC,CAAC,CAAC,sBAAsB,EAAE,CAAC,CAAC,CAAC,cAAc,CAAC,OAAO,EAAE,IAAI,IAAI,IAAI,CAAC,CAAC,CAAC;IACjG,MAAM,OAAO,GAA8B,IAAI,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;IACnE,IAAI,SAAS,GAAG,CAAC,CAAC;IAElB,MAAM,MAAM,GAAG,KAAK,IAAmB,EAAE;QACvC,OAAO,SAAS,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC;YAChC,MAAM,CAAC,GAAG,SAAS,EAAE,CAAC;YACtB,IAAI,CAAC;gBACH,MAAM,KAAK,GAAG,MAAM,EAAE,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;gBACpC,OAAO,CAAC,CAAC,CAAC,GAAG,EAAE,MAAM,EAAE,WAAW,EAAE,KAAK,EAAE,CAAC;YAC9C,CAAC;YAAC,OAAO,MAAM,EAAE,CAAC;gBAChB,OAAO,CAAC,CAAC,CAAC,GAAG,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,EAAE,CAAC;YAC9C,CAAC;QACH,CAAC;IACH,CAAC,CAAC;IAEF,MAAM,WAAW,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,MAAM,CAAC,CAAC;IAChD,MAAM,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,WAAW,EAAE,EAAE,GAAG,EAAE,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC;IAEvE,OAAO,OAAO,CAAC;AACjB,CAAC"}

package/dist/utils/model-slot-lease.d.ts

@@ -0,0 +1,33 @@
+/**
+ * Cross-process model-slot lease: a semaphore over the inference backend's
+ * concurrent slots, shared via the coordination DB so multiple agents AND
+ * multiple `uap` processes hitting the same endpoint never exceed the budget.
+ *
+ * `withModelSlot(holder, fn)` acquires a slot (waiting when the budget is full),
+ * runs `fn`, and releases — the right wrapper around any model call / agent spawn.
+ */
+import { CoordinationService } from '../coordination/service.js';
+export interface SlotLeaseOptions {
+    cwd?: string;
+    /** Budget override (else resolved from model-slots). */
+    budget?: number;
+    /** Lease TTL — a crashed holder is reaped after this. Default 120s. */
+    ttlMs?: number;
+    /** Max time to wait for a slot before proceeding anyway (fail-open). Default 5m. */
+    timeoutMs?: number;
+    /** Poll interval while the budget is full. Default 200ms. */
+    pollMs?: number;
+    /** Reuse an existing service (else one is constructed on the default DB). */
+    service?: CoordinationService;
+}
+/**
+ * Acquire a model slot, waiting (up to timeoutMs) when the budget is full.
+ * Returns a lease id, or null if it timed out (caller may proceed best-effort).
+ */
+export declare function acquireModelSlot(holder: string, opts?: SlotLeaseOptions): Promise<{
+    leaseId: number | null;
+    service: CoordinationService;
+}>;
+/** Run `fn` while holding a model slot (acquire → fn → release). */
+export declare function withModelSlot<T>(holder: string, fn: () => Promise<T>, opts?: SlotLeaseOptions): Promise<T>;
+//# sourceMappingURL=model-slot-lease.d.ts.map

package/dist/utils/model-slot-lease.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-slot-lease.d.ts","sourceRoot":"","sources":["../../src/utils/model-slot-lease.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AACH,OAAO,EAAE,mBAAmB,EAAE,MAAM,4BAA4B,CAAC;AAGjE,MAAM,WAAW,gBAAgB;IAC/B,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,wDAAwD;IACxD,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,uEAAuE;IACvE,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,oFAAoF;IACpF,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,6DAA6D;IAC7D,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,6EAA6E;IAC7E,OAAO,CAAC,EAAE,mBAAmB,CAAC;CAC/B;AASD;;;GAGG;AACH,wBAAsB,gBAAgB,CACpC,MAAM,EAAE,MAAM,EACd,IAAI,GAAE,gBAAqB,GAC1B,OAAO,CAAC;IAAE,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IAAC,OAAO,EAAE,mBAAmB,CAAA;CAAE,CAAC,CAcnE;AAED,oEAAoE;AACpE,wBAAsB,aAAa,CAAC,CAAC,EACnC,MAAM,EAAE,MAAM,EACd,EAAE,EAAE,MAAM,OAAO,CAAC,CAAC,CAAC,EACpB,IAAI,GAAE,gBAAqB,GAC1B,OAAO,CAAC,CAAC,CAAC,CAOZ"}

package/dist/utils/model-slot-lease.js

@@ -0,0 +1,48 @@
+/**
+ * Cross-process model-slot lease: a semaphore over the inference backend's
+ * concurrent slots, shared via the coordination DB so multiple agents AND
+ * multiple `uap` processes hitting the same endpoint never exceed the budget.
+ *
+ * `withModelSlot(holder, fn)` acquires a slot (waiting when the budget is full),
+ * runs `fn`, and releases — the right wrapper around any model call / agent spawn.
+ */
+import { CoordinationService } from '../coordination/service.js';
+import { getModelSlotBudget } from './model-slots.js';
+const sleep = (ms) => new Promise((r) => setTimeout(r, ms));
+async function resolveBudget(opts) {
+    if (opts.budget && opts.budget > 0)
+        return opts.budget;
+    return (await getModelSlotBudget(opts.cwd)).budget;
+}
+/**
+ * Acquire a model slot, waiting (up to timeoutMs) when the budget is full.
+ * Returns a lease id, or null if it timed out (caller may proceed best-effort).
+ */
+export async function acquireModelSlot(holder, opts = {}) {
+    const service = opts.service ?? new CoordinationService();
+    const budget = await resolveBudget(opts);
+    const ttlMs = opts.ttlMs ?? 120_000;
+    const timeoutMs = opts.timeoutMs ?? 300_000;
+    const pollMs = opts.pollMs ?? 200;
+    const deadline = Date.now() + timeoutMs;
+    for (;;) {
+        const leaseId = service.acquireModelSlot(holder, budget, ttlMs);
+        if (leaseId !== null)
+            return { leaseId, service };
+        if (Date.now() >= deadline)
+            return { leaseId: null, service }; // fail-open
+        await sleep(pollMs);
+    }
+}
+/** Run `fn` while holding a model slot (acquire → fn → release). */
+export async function withModelSlot(holder, fn, opts = {}) {
+    const { leaseId, service } = await acquireModelSlot(holder, opts);
+    try {
+        return await fn();
+    }
+    finally {
+        if (leaseId !== null)
+            service.releaseModelSlot(leaseId);
+    }
+}
+//# sourceMappingURL=model-slot-lease.js.map

package/dist/utils/model-slot-lease.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-slot-lease.js","sourceRoot":"","sources":["../../src/utils/model-slot-lease.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AACH,OAAO,EAAE,mBAAmB,EAAE,MAAM,4BAA4B,CAAC;AACjE,OAAO,EAAE,kBAAkB,EAAE,MAAM,kBAAkB,CAAC;AAgBtD,MAAM,KAAK,GAAG,CAAC,EAAU,EAAiB,EAAE,CAAC,IAAI,OAAO,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,UAAU,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;AAEnF,KAAK,UAAU,aAAa,CAAC,IAAsB;IACjD,IAAI,IAAI,CAAC,MAAM,IAAI,IAAI,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,IAAI,CAAC,MAAM,CAAC;IACvD,OAAO,CAAC,MAAM,kBAAkB,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC;AACrD,CAAC;AAED;;;GAGG;AACH,MAAM,CAAC,KAAK,UAAU,gBAAgB,CACpC,MAAc,EACd,OAAyB,EAAE;IAE3B,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,IAAI,IAAI,mBAAmB,EAAE,CAAC;IAC1D,MAAM,MAAM,GAAG,MAAM,aAAa,CAAC,IAAI,CAAC,CAAC;IACzC,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,IAAI,OAAO,CAAC;IACpC,MAAM,SAAS,GAAG,IAAI,CAAC,SAAS,IAAI,OAAO,CAAC;IAC5C,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,IAAI,GAAG,CAAC;IAClC,MAAM,QAAQ,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;IAExC,SAAS,CAAC;QACR,MAAM,OAAO,GAAG,OAAO,CAAC,gBAAgB,CAAC,MAAM,EAAE,MAAM,EAAE,KAAK,CAAC,CAAC;QAChE,IAAI,OAAO,KAAK,IAAI;YAAE,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,CAAC;QAClD,IAAI,IAAI,CAAC,GAAG,EAAE,IAAI,QAAQ;YAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC,YAAY;QAC3E,MAAM,KAAK,CAAC,MAAM,CAAC,CAAC;IACtB,CAAC;AACH,CAAC;AAED,oEAAoE;AACpE,MAAM,CAAC,KAAK,UAAU,aAAa,CACjC,MAAc,EACd,EAAoB,EACpB,OAAyB,EAAE;IAE3B,MAAM,EAAE,OAAO,EAAE,OAAO,EAAE,GAAG,MAAM,gBAAgB,CAAC,MAAM,EAAE,IAAI,CAAC,CAAC;IAClE,IAAI,CAAC;QACH,OAAO,MAAM,EAAE,EAAE,CAAC;IACpB,CAAC;YAAS,CAAC;QACT,IAAI,OAAO,KAAK,IAAI;YAAE,OAAO,CAAC,gBAAgB,CAAC,OAAO,CAAC,CAAC;IAC1D,CAAC;AACH,CAAC"}

package/dist/utils/model-slots.d.ts

@@ -0,0 +1,32 @@
+export declare const DEFAULT_SLOTS = 2;
+export type BudgetSource = 'env' | 'config' | 'probe' | 'default';
+/** The inference base URL (without a trailing /v1) for probing `/slots`. */
+export declare function inferenceBase(cwd?: string): string;
+export declare function headroom(cwd?: string): number;
+/** Explicit slot count from env/config (no probing). */
+export declare function configuredSlots(cwd?: string): number | undefined;
+/** Probe a llama.cpp `/slots` endpoint; returns the slot count or null. */
+export declare function probeSlots(base: string, timeoutMs?: number): Promise<number | null>;
+/**
+ * Resolve the model-slot budget, probing the endpoint when no override is set.
+ * Cached for CACHE_TTL_MS. Call this once at the start of a long-running
+ * orchestrator (challenge run, deliver) to warm the cache for sync callers.
+ */
+export declare function getModelSlotBudget(cwd?: string, opts?: {
+    probe?: boolean;
+    force?: boolean;
+}): Promise<{
+    budget: number;
+    slots: number;
+    source: BudgetSource;
+}>;
+/** Warm the cache by probing — call at orchestrator startup. */
+export declare function warmModelSlotBudget(cwd?: string): Promise<number>;
+/**
+ * Synchronous budget for hot paths (no probe): env > config > cached probe >
+ * default, minus headroom. Use after warmModelSlotBudget() has populated the
+ * cache, or when an explicit override is set.
+ */
+export declare function getMaxModelConcurrency(cwd?: string): number;
+export declare function resetModelSlotCache(): void;
+//# sourceMappingURL=model-slots.d.ts.map

package/dist/utils/model-slots.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-slots.d.ts","sourceRoot":"","sources":["../../src/utils/model-slots.ts"],"names":[],"mappings":"AAeA,eAAO,MAAM,aAAa,IAAI,CAAC;AAY/B,MAAM,MAAM,YAAY,GAAG,KAAK,GAAG,QAAQ,GAAG,OAAO,GAAG,SAAS,CAAC;AAoBlE,4EAA4E;AAC5E,wBAAgB,aAAa,CAAC,GAAG,CAAC,EAAE,MAAM,GAAG,MAAM,CAOlD;AAED,wBAAgB,QAAQ,CAAC,GAAG,CAAC,EAAE,MAAM,GAAG,MAAM,CAE7C;AAED,wDAAwD;AACxD,wBAAgB,eAAe,CAAC,GAAG,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,CAEhE;AAED,2EAA2E;AAC3E,wBAAsB,UAAU,CAAC,IAAI,EAAE,MAAM,EAAE,SAAS,SAAO,GAAG,OAAO,CAAC,MAAM,GAAG,IAAI,CAAC,CAgBvF;AAMD;;;;GAIG;AACH,wBAAsB,kBAAkB,CACtC,GAAG,CAAC,EAAE,MAAM,EACZ,IAAI,GAAE;IAAE,KAAK,CAAC,EAAE,OAAO,CAAC;IAAC,KAAK,CAAC,EAAE,OAAO,CAAA;CAAO,GAC9C,OAAO,CAAC;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,YAAY,CAAA;CAAE,CAAC,CA6BlE;AAED,gEAAgE;AAChE,wBAAsB,mBAAmB,CAAC,GAAG,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAEvE;AAED;;;;GAIG;AACH,wBAAgB,sBAAsB,CAAC,GAAG,CAAC,EAAE,MAAM,GAAG,MAAM,CAK3D;AAED,wBAAgB,mBAAmB,IAAI,IAAI,CAE1C"}

package/dist/utils/model-slots.js

@@ -0,0 +1,133 @@
+/**
+ * Model-slot budget: how many concurrent model calls the inference backend can
+ * actually serve. UAP's other concurrency limits are sized to CPU cores, which
+ * silently overruns a slot-limited server (llama.cpp `--parallel N`, or a remote
+ * API's concurrency ceiling) and causes timeouts / 429s / proxy doom-loops.
+ *
+ * Budget resolution (auto-detect + config override):
+ *   1. UAP_MODEL_SLOTS env                      (explicit override, no probe)
+ *   2. .uap.json modelConcurrency.slots         (config override, no probe)
+ *   3. probe the endpoint's llama.cpp `/slots`  (auto-detect)
+ *   4. DEFAULT_SLOTS                            (conservative fallback)
+ * minus `headroom`, floored at 1. Cached briefly so hot paths stay cheap.
+ */
+import { loadUapConfigRaw } from './config-loader.js';
+export const DEFAULT_SLOTS = 2; // matches llama-server-optimize's `--parallel 2`
+const CACHE_TTL_MS = 30_000;
+let _cache = null;
+function cfg(cwd) {
+    try {
+        const raw = loadUapConfigRaw(cwd ?? process.cwd());
+        return raw?.modelConcurrency ?? {};
+    }
+    catch {
+        return {};
+    }
+}
+function intEnv(name) {
+    const v = process.env[name];
+    if (!v)
+        return undefined;
+    const n = parseInt(v, 10);
+    return Number.isFinite(n) && n > 0 ? n : undefined;
+}
+/** The inference base URL (without a trailing /v1) for probing `/slots`. */
+export function inferenceBase(cwd) {
+    const raw = process.env.UAP_MODEL_ENDPOINT ||
+        process.env.UAP_INFERENCE_ENDPOINT ||
+        cfg(cwd).endpoint ||
+        'http://localhost:8080';
+    return raw.replace(/\/v1\/?$/, '').replace(/\/+$/, '');
+}
+export function headroom(cwd) {
+    return intEnv('UAP_MODEL_SLOT_HEADROOM') ?? cfg(cwd).headroom ?? 0;
+}
+/** Explicit slot count from env/config (no probing). */
+export function configuredSlots(cwd) {
+    return intEnv('UAP_MODEL_SLOTS') ?? cfg(cwd).slots;
+}
+/** Probe a llama.cpp `/slots` endpoint; returns the slot count or null. */
+export async function probeSlots(base, timeoutMs = 1500) {
+    try {
+        const ctrl = new AbortController();
+        const t = setTimeout(() => ctrl.abort(), timeoutMs);
+        try {
+            const res = await fetch(`${base}/slots`, { signal: ctrl.signal });
+            if (!res.ok)
+                return null;
+            const body = (await res.json());
+            if (Array.isArray(body) && body.length > 0)
+                return body.length;
+            return null;
+        }
+        finally {
+            clearTimeout(t);
+        }
+    }
+    catch {
+        return null;
+    }
+}
+function applyHeadroom(slots, cwd) {
+    return Math.max(1, slots - headroom(cwd));
+}
+/**
+ * Resolve the model-slot budget, probing the endpoint when no override is set.
+ * Cached for CACHE_TTL_MS. Call this once at the start of a long-running
+ * orchestrator (challenge run, deliver) to warm the cache for sync callers.
+ */
+export async function getModelSlotBudget(cwd, opts = {}) {
+    if (!opts.force && _cache && Date.now() - _cache.at < CACHE_TTL_MS) {
+        return { budget: _cache.budget, slots: _cache.slots, source: _cache.source };
+    }
+    let slots;
+    let source;
+    const explicit = configuredSlots(cwd);
+    if (intEnv('UAP_MODEL_SLOTS') !== undefined) {
+        slots = explicit;
+        source = 'env';
+    }
+    else if (explicit !== undefined) {
+        slots = explicit;
+        source = 'config';
+    }
+    else if (opts.probe !== false) {
+        const probed = await probeSlots(inferenceBase(cwd));
+        if (probed !== null) {
+            slots = probed;
+            source = 'probe';
+        }
+        else {
+            slots = DEFAULT_SLOTS;
+            source = 'default';
+        }
+    }
+    else {
+        slots = DEFAULT_SLOTS;
+        source = 'default';
+    }
+    const budget = applyHeadroom(slots, cwd);
+    _cache = { budget, slots, source, at: Date.now() };
+    return { budget, slots, source };
+}
+/** Warm the cache by probing — call at orchestrator startup. */
+export async function warmModelSlotBudget(cwd) {
+    return (await getModelSlotBudget(cwd, { probe: true, force: true })).budget;
+}
+/**
+ * Synchronous budget for hot paths (no probe): env > config > cached probe >
+ * default, minus headroom. Use after warmModelSlotBudget() has populated the
+ * cache, or when an explicit override is set.
+ */
+export function getMaxModelConcurrency(cwd) {
+    const explicit = configuredSlots(cwd);
+    if (explicit !== undefined)
+        return applyHeadroom(explicit, cwd);
+    if (_cache && Date.now() - _cache.at < CACHE_TTL_MS)
+        return _cache.budget;
+    return applyHeadroom(DEFAULT_SLOTS, cwd);
+}
+export function resetModelSlotCache() {
+    _cache = null;
+}
+//# sourceMappingURL=model-slots.js.map

package/dist/utils/model-slots.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-slots.js","sourceRoot":"","sources":["../../src/utils/model-slots.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AACH,OAAO,EAAE,gBAAgB,EAAE,MAAM,oBAAoB,CAAC;AAEtD,MAAM,CAAC,MAAM,aAAa,GAAG,CAAC,CAAC,CAAC,iDAAiD;AACjF,MAAM,YAAY,GAAG,MAAM,CAAC;AAS5B,IAAI,MAAM,GAA+E,IAAI,CAAC;AAI9F,SAAS,GAAG,CAAC,GAAY;IACvB,IAAI,CAAC;QACH,MAAM,GAAG,GAAG,gBAAgB,CAAC,GAAG,IAAI,OAAO,CAAC,GAAG,EAAE,CAEzC,CAAC;QACT,OAAO,GAAG,EAAE,gBAAgB,IAAI,EAAE,CAAC;IACrC,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,EAAE,CAAC;IACZ,CAAC;AACH,CAAC;AAED,SAAS,MAAM,CAAC,IAAY;IAC1B,MAAM,CAAC,GAAG,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;IAC5B,IAAI,CAAC,CAAC;QAAE,OAAO,SAAS,CAAC;IACzB,MAAM,CAAC,GAAG,QAAQ,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;IAC1B,OAAO,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;AACrD,CAAC;AAED,4EAA4E;AAC5E,MAAM,UAAU,aAAa,CAAC,GAAY;IACxC,MAAM,GAAG,GACP,OAAO,CAAC,GAAG,CAAC,kBAAkB;QAC9B,OAAO,CAAC,GAAG,CAAC,sBAAsB;QAClC,GAAG,CAAC,GAAG,CAAC,CAAC,QAAQ;QACjB,uBAAuB,CAAC;IAC1B,OAAO,GAAG,CAAC,OAAO,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC;AACzD,CAAC;AAED,MAAM,UAAU,QAAQ,CAAC,GAAY;IACnC,OAAO,MAAM,CAAC,yBAAyB,CAAC,IAAI,GAAG,CAAC,GAAG,CAAC,CAAC,QAAQ,IAAI,CAAC,CAAC;AACrE,CAAC;AAED,wDAAwD;AACxD,MAAM,UAAU,eAAe,CAAC,GAAY;IAC1C,OAAO,MAAM,CAAC,iBAAiB,CAAC,IAAI,GAAG,CAAC,GAAG,CAAC,CAAC,KAAK,CAAC;AACrD,CAAC;AAED,2EAA2E;AAC3E,MAAM,CAAC,KAAK,UAAU,UAAU,CAAC,IAAY,EAAE,SAAS,GAAG,IAAI;IAC7D,IAAI,CAAC;QACH,MAAM,IAAI,GAAG,IAAI,eAAe,EAAE,CAAC;QACnC,MAAM,CAAC,GAAG,UAAU,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,KAAK,EAAE,EAAE,SAAS,CAAC,CAAC;QACpD,IAAI,CAAC;YACH,MAAM,GAAG,GAAG,MAAM,KAAK,CAAC,GAAG,IAAI,QAAQ,EAAE,EAAE,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC;YAClE,IAAI,CAAC,GAAG,CAAC,EAAE;gBAAE,OAAO,IAAI,CAAC;YACzB,MAAM,IAAI,GAAG,CAAC,MAAM,GAAG,CAAC,IAAI,EAAE,CAAY,CAAC;YAC3C,IAAI,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,IAAI,IAAI,CAAC,MAAM,GAAG,CAAC;gBAAE,OAAO,IAAI,CAAC,MAAM,CAAC;YAC/D,OAAO,IAAI,CAAC;QACd,CAAC;gBAAS,CAAC;YACT,YAAY,CAAC,CAAC,CAAC,CAAC;QAClB,CAAC;IACH,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,IAAI,CAAC;IACd,CAAC;AACH,CAAC;AAED,SAAS,aAAa,CAAC,KAAa,EAAE,GAAY;IAChD,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC;AAC5C,CAAC;AAED;;;;GAIG;AACH,MAAM,CAAC,KAAK,UAAU,kBAAkB,CACtC,GAAY,EACZ,OAA6C,EAAE;IAE/C,IAAI,CAAC,IAAI,CAAC,KAAK,IAAI,MAAM,IAAI,IAAI,CAAC,GAAG,EAAE,GAAG,MAAM,CAAC,EAAE,GAAG,YAAY,EAAE,CAAC;QACnE,OAAO,EAAE,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,KAAK,EAAE,MAAM,CAAC,KAAK,EAAE,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,CAAC;IAC/E,CAAC;IACD,IAAI,KAAa,CAAC;IAClB,IAAI,MAAoB,CAAC;IACzB,MAAM,QAAQ,GAAG,eAAe,CAAC,GAAG,CAAC,CAAC;IACtC,IAAI,MAAM,CAAC,iBAAiB,CAAC,KAAK,SAAS,EAAE,CAAC;QAC5C,KAAK,GAAG,QAAS,CAAC;QAClB,MAAM,GAAG,KAAK,CAAC;IACjB,CAAC;SAAM,IAAI,QAAQ,KAAK,SAAS,EAAE,CAAC;QAClC,KAAK,GAAG,QAAQ,CAAC;QACjB,MAAM,GAAG,QAAQ,CAAC;IACpB,CAAC;SAAM,IAAI,IAAI,CAAC,KAAK,KAAK,KAAK,EAAE,CAAC;QAChC,MAAM,MAAM,GAAG,MAAM,UAAU,CAAC,aAAa,CAAC,GAAG,CAAC,CAAC,CAAC;QACpD,IAAI,MAAM,KAAK,IAAI,EAAE,CAAC;YACpB,KAAK,GAAG,MAAM,CAAC;YACf,MAAM,GAAG,OAAO,CAAC;QACnB,CAAC;aAAM,CAAC;YACN,KAAK,GAAG,aAAa,CAAC;YACtB,MAAM,GAAG,SAAS,CAAC;QACrB,CAAC;IACH,CAAC;SAAM,CAAC;QACN,KAAK,GAAG,aAAa,CAAC;QACtB,MAAM,GAAG,SAAS,CAAC;IACrB,CAAC;IACD,MAAM,MAAM,GAAG,aAAa,CAAC,KAAK,EAAE,GAAG,CAAC,CAAC;IACzC,MAAM,GAAG,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,EAAE,EAAE,IAAI,CAAC,GAAG,EAAE,EAAE,CAAC;IACnD,OAAO,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,CAAC;AACnC,CAAC;AAED,gEAAgE;AAChE,MAAM,CAAC,KAAK,UAAU,mBAAmB,CAAC,GAAY;IACpD,OAAO,CAAC,MAAM,kBAAkB,CAAC,GAAG,EAAE,EAAE,KAAK,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC;AAC9E,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,sBAAsB,CAAC,GAAY;IACjD,MAAM,QAAQ,GAAG,eAAe,CAAC,GAAG,CAAC,CAAC;IACtC,IAAI,QAAQ,KAAK,SAAS;QAAE,OAAO,aAAa,CAAC,QAAQ,EAAE,GAAG,CAAC,CAAC;IAChE,IAAI,MAAM,IAAI,IAAI,CAAC,GAAG,EAAE,GAAG,MAAM,CAAC,EAAE,GAAG,YAAY;QAAE,OAAO,MAAM,CAAC,MAAM,CAAC;IAC1E,OAAO,aAAa,CAAC,aAAa,EAAE,GAAG,CAAC,CAAC;AAC3C,CAAC;AAED,MAAM,UAAU,mBAAmB;IACjC,MAAM,GAAG,IAAI,CAAC;AAChB,CAAC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miller-tech/uap",
-  "version": "1.72.1",
+  "version": "1.74.0",
   "description": "Autonomous AI agent memory system with CLAUDE.md protocol enforcement",
   "type": "module",
   "main": "dist/index.js",