floe-guard 0.2.0 → 0.3.0

package/dist/index.d.ts

@@ -23,8 +23,9 @@ interface PricedModel {
 /**
  * Resolve a model to its per-token price, or `null` if it cannot be priced.
  *
- * Overrides win, then the bundled cost map (looked up by bare name, then the raw
- * field). Fail-closed: both prices must be finite, else `null`.
+ * Per specificity group (exact forms first, date-stripped fallbacks second):
+ * overrides win, then the bundled cost map. Fail-closed: the first matching
+ * entry must have finite prices, else `null`.
  */
 declare function resolvePrice(model: string, overrides?: Record<string, ManualPrice>): PricedModel | null;
 /** USD cost for token usage. Negative counts are clamped to zero. */
@@ -63,6 +64,32 @@ declare namespace pricing {
  * same epsilon handling, same fail-closed default.
  */
 
+/**
+ * One priced spend event in the guard's per-call ledger.
+ *
+ * Every {@link BudgetGuard.record} / {@link BudgetGuard.settle} /
+ * {@link BudgetGuard.recordTool} that accrues spend appends exactly one event, so
+ * the ledger's costs sum to `spentUsd` (unless a `maxLogEvents` ring buffer has
+ * evicted old events). The schema is identical in the Python
+ * package (`SpendEvent` in `src/floe_guard/guard.py`) and
+ * {@link BudgetGuard.exportLog} serialises it with the same snake_case keys in
+ * both languages, so every agent emits the same shape regardless of stack.
+ */
+interface SpendEvent {
+    /** Unix epoch seconds (UTC). */
+    readonly timestamp: number;
+    readonly kind: "llm" | "tool";
+    readonly modelOrTool: string;
+    /** `null` for tool events. */
+    readonly promptTokens: number | null;
+    /** `null` for tool events. */
+    readonly completionTokens: number | null;
+    readonly costUsd: number;
+    /** Caller-supplied tag (agent/task name). */
+    readonly label?: string;
+    /** The reservation settled by this call, if any. */
+    readonly reserved?: number;
+}
 interface BudgetGuardOptions {
     /** Per-model manual prices for models the bundled cost map cannot price. */
     priceOverrides?: Record<string, ManualPrice>;
@@ -83,6 +110,13 @@ interface BudgetGuardOptions {
      * flags `nearLimit` so an agent can taper before the hard-stop. Default 8000.
      */
     nearLimitBps?: number;
+    /**
+     * Optional cap on the per-call spend ledger ({@link BudgetGuard.spendLog}).
+     * When set, the ledger is a ring buffer keeping the most recent N events so a
+     * long-running agent's memory stays bounded; the running totals are
+     * unaffected. Default: keep every event.
+     */
+    maxLogEvents?: number;
 }
 /**
  * A context-aware spend signal for the single local budget.
@@ -119,6 +153,9 @@ declare class BudgetGuard {
     private lastCost;
     /** USD held for in-flight calls (reserved, not yet settled). Counts toward the ceiling. */
     private reserved;
+    /** Per-call ledger, oldest first; a ring buffer when maxLogEvents is set. */
+    private readonly spendEvents;
+    private readonly maxLogEvents?;
     /**
      * @param limitUsd the spend ceiling, in USD. `0` blocks the very first call.
      */
@@ -151,11 +188,15 @@ declare class BudgetGuard {
      * Release a reservation and record the actual cost. `record` is `settle` with
      * no reservation. Returns the USD cost of this call; unpriceable-model handling
      * matches {@link BudgetGuard.record}, and any held reservation is released even
-     * on the warn-and-skip path.
+     * on the warn-and-skip path. A priced call appends one {@link SpendEvent} to
+     * {@link BudgetGuard.spendLog} (`label` tags it, e.g. with an agent/task name);
+     * the warn-and-skip path accrues nothing and logs nothing, so the ledger stays
+     * in lockstep with `spentUsd`.
      */
     settle(model: string, promptTokens: number, completionTokens: number, options?: {
         reserved?: number;
         price?: ManualPrice;
+        label?: string;
     }): number;
     /**
      * Price one response's tokens offline and add the cost to the total.
@@ -166,6 +207,21 @@ declare class BudgetGuard {
      */
     record(model: string, promptTokens: number, completionTokens: number, options?: {
         price?: ManualPrice;
+        label?: string;
+    }): number;
+    /**
+     * Accrue a non-LLM cost (a paid tool/API call) against the same ceiling.
+     *
+     * Tools with direct dollar costs — search APIs, scrapers, sandboxes — spend the
+     * same budget the LLM calls do; `recordTool` folds them into `spentUsd` (so
+     * `check()` / `reserve()` see them) and appends a `kind: "tool"`
+     * {@link SpendEvent} to {@link BudgetGuard.spendLog}. The caller supplies the
+     * cost: tools have no token usage to price. Deliberately does NOT update the
+     * next-call estimate — that predicts the next *LLM* call, and a tool's price
+     * would skew it. Returns `costUsd`.
+     */
+    recordTool(tool: string, costUsd: number, options?: {
+        label?: string;
     }): number;
     /**
      * Drop an in-flight reservation without recording spend (e.g. the call failed
@@ -174,6 +230,24 @@ declare class BudgetGuard {
     release(reserved: number): void;
     /** USD left before the ceiling, net of in-flight reservations (never negative). */
     get remainingUsd(): number;
+    /**
+     * The per-call spend ledger, oldest first — one {@link SpendEvent} per priced
+     * `record()` / `settle()` / `recordTool()`. Returns a snapshot copy: mutating
+     * it cannot corrupt the ledger.
+     */
+    get spendLog(): SpendEvent[];
+    /**
+     * The spend ledger as JSONL — one event per line, newline-terminated.
+     *
+     * The schema is stable and language-independent (snake_case keys, fixed order;
+     * optional fields omitted when absent), identical to the Python package's
+     * `export_log()`, so heterogeneous agents produce logs you can concatenate and
+     * analyse as one stream. (The *schema* is the contract, not the bytes: the two
+     * runtimes may render the same float differently, e.g. JS `0.0000025` vs
+     * Python `2.5e-06`.) Empty ledger yields `""`.
+     */
+    exportLog(): string;
+    private appendEvent;
     /**
      * Context-aware spend advisory for this budget — see {@link BudgetAdvisory}.
      *
@@ -289,4 +363,4 @@ interface BudgetGuardMiddleware {
  */
 declare function budgetGuardMiddleware(guard: BudgetGuard): BudgetGuardMiddleware;
 
-export { type BudgetAdvisory, BudgetExceeded, BudgetGuard, type BudgetGuardMiddleware, type BudgetGuardOptions, FloeGuardError, type ManualPrice, type PricedModel, UnpriceableModelError, budgetGuardMiddleware, priceTokens, pricing, resolvePrice };
+export { type BudgetAdvisory, BudgetExceeded, BudgetGuard, type BudgetGuardMiddleware, type BudgetGuardOptions, FloeGuardError, type ManualPrice, type PricedModel, type SpendEvent, UnpriceableModelError, budgetGuardMiddleware, priceTokens, pricing, resolvePrice };

package/dist/index.js

@@ -182,6 +182,12 @@ var cost_map_default = {
     litellm_provider: "anthropic",
     mode: "chat"
   },
+  "claude-sonnet-5": {
+    input_cost_per_token: 2e-6,
+    output_cost_per_token: 1e-5,
+    litellm_provider: "anthropic",
+    mode: "chat"
+  },
   "ft:gpt-3.5-turbo": {
     input_cost_per_token: 3e-6,
     output_cost_per_token: 6e-6,
@@ -632,6 +638,30 @@ var cost_map_default = {
     litellm_provider: "openai",
     mode: "chat"
   },
+  "gpt-5.6": {
+    input_cost_per_token: 5e-6,
+    output_cost_per_token: 3e-5,
+    litellm_provider: "openai",
+    mode: "chat"
+  },
+  "gpt-5.6-luna": {
+    input_cost_per_token: 1e-6,
+    output_cost_per_token: 6e-6,
+    litellm_provider: "openai",
+    mode: "chat"
+  },
+  "gpt-5.6-sol": {
+    input_cost_per_token: 5e-6,
+    output_cost_per_token: 3e-5,
+    litellm_provider: "openai",
+    mode: "chat"
+  },
+  "gpt-5.6-terra": {
+    input_cost_per_token: 25e-7,
+    output_cost_per_token: 15e-6,
+    litellm_provider: "openai",
+    mode: "chat"
+  },
   "gpt-audio": {
     input_cost_per_token: 25e-7,
     output_cost_per_token: 1e-5,
@@ -686,6 +716,18 @@ var cost_map_default = {
     litellm_provider: "openai",
     mode: "chat"
   },
+  "gpt-realtime-2.1": {
+    input_cost_per_token: 4e-6,
+    output_cost_per_token: 24e-6,
+    litellm_provider: "openai",
+    mode: "chat"
+  },
+  "gpt-realtime-2.1-mini": {
+    input_cost_per_token: 6e-7,
+    output_cost_per_token: 24e-7,
+    litellm_provider: "openai",
+    mode: "chat"
+  },
   "gpt-realtime-2025-08-28": {
     input_cost_per_token: 4e-6,
     output_cost_per_token: 16e-6,
@@ -776,6 +818,24 @@ var cost_map_default = {
     litellm_provider: "openai",
     mode: "chat"
   },
+  "openai/gpt-oss-120b": {
+    input_cost_per_token: 15e-8,
+    output_cost_per_token: 6e-7,
+    litellm_provider: "groq",
+    mode: "chat"
+  },
+  "openai/gpt-oss-20b": {
+    input_cost_per_token: 75e-9,
+    output_cost_per_token: 3e-7,
+    litellm_provider: "groq",
+    mode: "chat"
+  },
+  "openai/gpt-oss-safeguard-20b": {
+    input_cost_per_token: 75e-9,
+    output_cost_per_token: 3e-7,
+    litellm_provider: "groq",
+    mode: "chat"
+  },
   "qwen/qwen3-32b": {
     input_cost_per_token: 29e-8,
     output_cost_per_token: 59e-8,
@@ -810,39 +870,64 @@ var cost_map_default = {
 
 // src/pricing.ts
 var COST_MAP = cost_map_default;
-function bareModel(model) {
+var PROVIDER_PREFIXES = /* @__PURE__ */ new Set(["groq"]);
+var DATE_SUFFIX = /-(?:\d{8}|\d{4}-\d{2}-\d{2})$/;
+function candidateGroups(model) {
   const m = model.trim();
-  const slash = m.lastIndexOf("/");
-  return slash === -1 ? m : m.slice(slash + 1);
+  const base = [m];
+  const firstSlash = m.indexOf("/");
+  if (firstSlash !== -1 && PROVIDER_PREFIXES.has(m.slice(0, firstSlash))) {
+    base.push(m.slice(firstSlash + 1));
+  }
+  const lastSlash = m.lastIndexOf("/");
+  if (lastSlash !== -1) {
+    base.push(m.slice(lastSlash + 1));
+  }
+  const exact = [];
+  for (const cand of base) {
+    if (cand && !exact.includes(cand)) exact.push(cand);
+  }
+  const stripped = [];
+  for (const cand of exact) {
+    const c = cand.replace(DATE_SUFFIX, "");
+    if (c && !exact.includes(c) && !stripped.includes(c)) stripped.push(c);
+  }
+  return [exact, stripped];
 }
 function bothFinite(a, b) {
   return typeof a === "number" && typeof b === "number" && Number.isFinite(a) && Number.isFinite(b);
 }
 function resolvePrice(model, overrides) {
-  const bare = bareModel(model);
-  if (overrides) {
-    const ov = overrides[bare] ?? overrides[model.trim()];
-    if (ov !== void 0) {
-      if (bothFinite(ov.inputCostPerToken, ov.outputCostPerToken)) {
-        return {
-          inputCostPerToken: ov.inputCostPerToken,
-          outputCostPerToken: ov.outputCostPerToken,
-          source: "override"
-        };
+  for (const cands of candidateGroups(model)) {
+    if (overrides) {
+      for (const cand of cands) {
+        const ov = Object.prototype.hasOwnProperty.call(overrides, cand) ? overrides[cand] : void 0;
+        if (ov !== void 0) {
+          if (bothFinite(ov.inputCostPerToken, ov.outputCostPerToken)) {
+            return {
+              inputCostPerToken: ov.inputCostPerToken,
+              outputCostPerToken: ov.outputCostPerToken,
+              source: "override"
+            };
+          }
+          return null;
+        }
       }
-      return null;
+    }
+    for (const cand of cands) {
+      const entry = Object.prototype.hasOwnProperty.call(COST_MAP, cand) ? COST_MAP[cand] : void 0;
+      if (!entry) continue;
+      const input = entry.input_cost_per_token;
+      const output = entry.output_cost_per_token;
+      if (!bothFinite(input, output)) return null;
+      return {
+        inputCostPerToken: input,
+        outputCostPerToken: output,
+        source: "cost_map"
+      };
     }
   }
-  const entry = COST_MAP[bare] ?? COST_MAP[model.trim()];
-  if (!entry) return null;
-  const input = entry.input_cost_per_token;
-  const output = entry.output_cost_per_token;
-  if (!bothFinite(input, output)) return null;
-  return {
-    inputCostPerToken: input,
-    outputCostPerToken: output,
-    source: "cost_map"
-  };
+  return null;
 }
 function priceTokens(priced, promptTokens, completionTokens) {
   const p = Math.max(0, promptTokens);
@@ -867,6 +952,9 @@ var BudgetGuard = class {
   lastCost = 0;
   /** USD held for in-flight calls (reserved, not yet settled). Counts toward the ceiling. */
   reserved = 0;
+  /** Per-call ledger, oldest first; a ring buffer when maxLogEvents is set. */
+  spendEvents = [];
+  maxLogEvents;
   /**
    * @param limitUsd the spend ceiling, in USD. `0` blocks the very first call.
    */
@@ -882,7 +970,13 @@ var BudgetGuard = class {
         `nearLimitBps must be an integer in 0..10000, got ${nearLimitBps}`
       );
     }
+    if (options.maxLogEvents !== void 0 && (!Number.isInteger(options.maxLogEvents) || options.maxLogEvents < 0)) {
+      throw new RangeError(
+        `maxLogEvents must be a non-negative integer, got ${options.maxLogEvents}`
+      );
+    }
     this.limitUsd = limitUsd;
+    this.maxLogEvents = options.maxLogEvents;
     this.priceOverrides = options.priceOverrides;
     this.failClosed = options.failClosed ?? true;
     this.onBlock = options.onBlock ?? defaultOnBlock;
@@ -944,7 +1038,10 @@ var BudgetGuard = class {
    * Release a reservation and record the actual cost. `record` is `settle` with
    * no reservation. Returns the USD cost of this call; unpriceable-model handling
    * matches {@link BudgetGuard.record}, and any held reservation is released even
-   * on the warn-and-skip path.
+   * on the warn-and-skip path. A priced call appends one {@link SpendEvent} to
+   * {@link BudgetGuard.spendLog} (`label` tags it, e.g. with an agent/task name);
+   * the warn-and-skip path accrues nothing and logs nothing, so the ledger stays
+   * in lockstep with `spentUsd`.
    */
   settle(model, promptTokens, completionTokens, options = {}) {
     const reserved = options.reserved ?? 0;
@@ -981,6 +1078,18 @@ var BudgetGuard = class {
       this.spentUsd = this.limitUsd;
     }
     this.lastCost = cost;
+    this.appendEvent({
+      timestamp: Date.now() / 1e3,
+      kind: "llm",
+      modelOrTool: model,
+      promptTokens,
+      completionTokens,
+      costUsd: cost,
+      ...options.label !== void 0 ? { label: options.label } : {},
+      // 0 means "no reservation" (the plain record() path) — omit rather than
+      // log a meaningless zero.
+      ...reserved ? { reserved } : {}
+    });
     return cost;
   }
   /**
@@ -993,9 +1102,40 @@ var BudgetGuard = class {
   record(model, promptTokens, completionTokens, options = {}) {
     return this.settle(model, promptTokens, completionTokens, {
       reserved: 0,
-      price: options.price
+      price: options.price,
+      label: options.label
     });
   }
+  /**
+   * Accrue a non-LLM cost (a paid tool/API call) against the same ceiling.
+   *
+   * Tools with direct dollar costs — search APIs, scrapers, sandboxes — spend the
+   * same budget the LLM calls do; `recordTool` folds them into `spentUsd` (so
+   * `check()` / `reserve()` see them) and appends a `kind: "tool"`
+   * {@link SpendEvent} to {@link BudgetGuard.spendLog}. The caller supplies the
+   * cost: tools have no token usage to price. Deliberately does NOT update the
+   * next-call estimate — that predicts the next *LLM* call, and a tool's price
+   * would skew it. Returns `costUsd`.
+   */
+  recordTool(tool, costUsd, options = {}) {
+    if (!Number.isFinite(costUsd) || costUsd < 0) {
+      throw new RangeError(`costUsd must be a finite, non-negative number, got ${costUsd}`);
+    }
+    this.spentUsd += costUsd;
+    if (this.spentUsd - this.limitUsd > 0 && this.spentUsd - this.limitUsd < EPS) {
+      this.spentUsd = this.limitUsd;
+    }
+    this.appendEvent({
+      timestamp: Date.now() / 1e3,
+      kind: "tool",
+      modelOrTool: tool,
+      promptTokens: null,
+      completionTokens: null,
+      costUsd,
+      ...options.label !== void 0 ? { label: options.label } : {}
+    });
+    return costUsd;
+  }
   /**
    * Drop an in-flight reservation without recording spend (e.g. the call failed
    * before producing usage). Safe to call with `0`.
@@ -1011,6 +1151,46 @@ var BudgetGuard = class {
   get remainingUsd() {
     return Math.max(0, this.limitUsd - this.spentUsd - this.reserved);
   }
+  /**
+   * The per-call spend ledger, oldest first — one {@link SpendEvent} per priced
+   * `record()` / `settle()` / `recordTool()`. Returns a snapshot copy: mutating
+   * it cannot corrupt the ledger.
+   */
+  get spendLog() {
+    return [...this.spendEvents];
+  }
+  /**
+   * The spend ledger as JSONL — one event per line, newline-terminated.
+   *
+   * The schema is stable and language-independent (snake_case keys, fixed order;
+   * optional fields omitted when absent), identical to the Python package's
+   * `export_log()`, so heterogeneous agents produce logs you can concatenate and
+   * analyse as one stream. (The *schema* is the contract, not the bytes: the two
+   * runtimes may render the same float differently, e.g. JS `0.0000025` vs
+   * Python `2.5e-06`.) Empty ledger yields `""`.
+   */
+  exportLog() {
+    return this.spendEvents.map((e) => {
+      const row = {
+        timestamp: e.timestamp,
+        kind: e.kind,
+        model_or_tool: e.modelOrTool,
+        prompt_tokens: e.promptTokens,
+        completion_tokens: e.completionTokens,
+        cost_usd: e.costUsd
+      };
+      if (e.label !== void 0) row.label = e.label;
+      if (e.reserved !== void 0) row.reserved = e.reserved;
+      return `${JSON.stringify(row)}
+`;
+    }).join("");
+  }
+  appendEvent(event) {
+    this.spendEvents.push(Object.freeze(event));
+    if (this.maxLogEvents !== void 0 && this.spendEvents.length > this.maxLogEvents) {
+      this.spendEvents.splice(0, this.spendEvents.length - this.maxLogEvents);
+    }
+  }
   /**
    * Context-aware spend advisory for this budget — see {@link BudgetAdvisory}.
    *