@warlock.js/ai-panoptic 5.30.0 → 5.32.0

package/CHANGELOG.md

@@ -4,6 +4,27 @@ All notable changes to `@warlock.js/ai-panoptic` are documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). `@warlock.js/*` packages are released in lockstep — every package shares the same version number, so a version below may list only the changes that affected this package.
 
+## 5.32.0 - 2026-10-18
+
+### Changed
+
+- Pin the package development compiler to TypeScript 7.0.2 and use explicit Node types and source roots with supported module resolution.
+
+### Fixed
+
+- Use the canonical usage rollup for trace aggregates, retaining hosted-tool counters and historical request fees alongside cache lifetime and mixed-provider cache semantics.
+
+## 5.31.0 - 2026-10-18
+
+### Fixed
+
+- Trace cost totals (exporters, the dashboard and trace filters) include the new cache-write cost channel.
+
+### Added
+
+- Tool spans carry the provider's tool-call id as the `tool.callId` attribute.
+- Agent runs with `status: "incomplete"` (their final answer was cut off) are exported as errors: OTel `ERROR` status, Langfuse level `ERROR`, the console exporter's error stream (marked `INCOMPLETE`), the dashboard's error filter, and the `failed` count of trace-store aggregates. `?status=incomplete` is accepted by the dashboard API.
+
 ## 5.30.0 - 2026-10-16
 
 ### Changed
@@ -81,6 +102,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 ### Changed
 
 - Lockstep release maintenance and dependency refresh.
+
 ## 5.20.0 - 2026-09-24
 
 ### Changed
@@ -146,9 +168,9 @@ changelogs.
 
 - **Dashboard stored XSS via `javascript:` markdown links — fixed.** The dashboard's markdown link renderer (`mdInline` in `ui.html.ts`) rewrote `[text](url)` into a live `<a href>` without validating the URL scheme, so a `javascript:` URL embedded in captured trace content (`span.input` / `span.output` under `captureContent` — i.e. prompt-injected model output or attacker-controlled tool results) became a stored XSS that fired when an operator clicked the link; the page's CSP (`script-src 'unsafe-inline'`) permits `javascript:` URI execution and does not restrict top-level navigation, putting the in-page bearer token in reach. Link URLs are now checked against a scheme allowlist (`http:`, `https:`, `mailto:`, plus relative/anchor URLs) after normalizing the URL the way a browser will — attribute-entity decode, strip of ignored control chars/whitespace (`java\tscript:`), lowercase — and scheme-relative `//host` links are rejected too; a rejected URL renders its label as plain text with no `href` at all. Regression-tested by executing the actual inlined client renderer against `javascript:`/`data:`/`vbscript:` payloads and their case/whitespace/entity obfuscations (`ui.html.md-links.spec.ts`)
 - **Dashboard token-handling hardening (3 residual findings from the same audit as the XSS above).** None of these are exploitable on their own today, but each widened the blast radius of a future bug:
-  - *Bearer token no longer a page-global.* `ui.html.ts`'s client script kept `TOKEN` as a `var` shared across its whole ~1000-line closure. It's now sealed inside its own inner IIFE that exposes only the `fetchAuthed` helper — code elsewhere in that closure (present or future) can no longer read the raw token by name.
-  - *Constant-time token comparison.* `serve.ts`'s `isAuthorized` compared the header/query token with plain `===`, which short-circuits on the first differing byte and is a textbook timing side-channel — the one scenario the package's own docs call out as realistic (`authToken` is required specifically when bound off-loopback, i.e. reachable over a network). Both the header and query-string checks now go through `crypto.timingSafeEqual` on length-checked, equal-length buffers.
-  - *Token-in-URL narrowed to the one route that needs it.* `?token=` is no longer accepted on the JSON API routes, only on the HTML page route. The page's own polling already re-sends the token as an `Authorization` header (fixed in 4.8.2), so the only request that structurally *can't* carry a header is the initial browser navigation that loads the HTML shell — that's the sole remaining query-string exposure, narrowing the token's footprint in access/proxy logs and browser history from every poll to one request.
+  - _Bearer token no longer a page-global._ `ui.html.ts`'s client script kept `TOKEN` as a `var` shared across its whole ~1000-line closure. It's now sealed inside its own inner IIFE that exposes only the `fetchAuthed` helper — code elsewhere in that closure (present or future) can no longer read the raw token by name.
+  - _Constant-time token comparison._ `serve.ts`'s `isAuthorized` compared the header/query token with plain `===`, which short-circuits on the first differing byte and is a textbook timing side-channel — the one scenario the package's own docs call out as realistic (`authToken` is required specifically when bound off-loopback, i.e. reachable over a network). Both the header and query-string checks now go through `crypto.timingSafeEqual` on length-checked, equal-length buffers.
+  - _Token-in-URL narrowed to the one route that needs it._ `?token=` is no longer accepted on the JSON API routes, only on the HTML page route. The page's own polling already re-sends the token as an `Authorization` header (fixed in 4.8.2), so the only request that structurally _can't_ carry a header is the initial browser navigation that loads the HTML shell — that's the sole remaining query-string exposure, narrowing the token's footprint in access/proxy logs and browser history from every poll to one request.
 
 ## 4.12.0
 

package/README.md

@@ -81,8 +81,8 @@ Four exporters ship today, each a factory returning an `ExporterContract`:
 ```ts
 import {
   consoleExporter, // zero-dep — print trace summary or full span tree
-  fileExporter,    // zero-dep — append traces as JSON Lines, buffered
-  otelExporter,    // OpenTelemetry — gen_ai.* semantic conventions (lazy @opentelemetry/api)
+  fileExporter, // zero-dep — append traces as JSON Lines, buffered
+  otelExporter, // OpenTelemetry — gen_ai.* semantic conventions (lazy @opentelemetry/api)
   langfuseExporter, // Langfuse — traces + generations (lazy langfuse)
 } from "@warlock.js/ai-panoptic";
 

package/cjs/index.cjs

@@ -114,10 +114,7 @@ function extractSpanAttributes(report) {
 		case "orchestrator":
 			addOrchestratorAttributes(attributes, extensions);
 			break;
-		case "tool":
-			addToolAttributes(attributes, extensions);
-			break;
-		default: break;
+		case "tool": addToolAttributes(attributes, extensions);
 	}
 	if (Object.keys(attributes).length === 0) return;
 	return attributes;
@@ -146,6 +143,7 @@ function addOrchestratorAttributes(attributes, extensions) {
 }
 function addToolAttributes(attributes, extensions) {
 	if (extensions.tripIndex !== void 0) attributes["tool.tripIndex"] = extensions.tripIndex;
+	if (extensions.callId !== void 0) attributes["tool.callId"] = extensions.callId;
 	if (extensions.recoveredFrom !== void 0) attributes["tool.recoveredFrom"] = extensions.recoveredFrom;
 }
 
@@ -501,48 +499,14 @@ function toEpoch(bound) {
 
 //#endregion
 //#region ../../packages/ai-panoptic/src/store/sum-usage.ts
-/**
-* Fold a child {@link Usage} into a running accumulator. Token channels
-* (`input` / `output` / `total`) always sum; the optional cache /
-* reasoning channels (`cachedTokens` / `cacheWriteTokens` /
-* `reasoningTokens`) sum only when at least one side reported them, so
-* a provider that never meters a channel doesn't fabricate a `0` for
-* it. The `cost` breakdown is merged with the core framework's
-* {@link accumulateCost}, keeping cost-rollup semantics identical to a
-* native report tree — an unpriced contributor never erases a priced
-* one.
-*
-* Pure: returns a fresh `Usage`, never mutates either argument. Seed an
-* aggregation with {@link emptyUsage}.
-*
-* @example
-* let total = emptyUsage();
-* for (const trace of traces) {
-*   total = sumUsage(total, trace.usage);
-* }
-*/
+/** Return an independent all-channel usage rollup, preserving hosted-tool counters. */
 function sumUsage(accumulator, next) {
-	const merged = {
-		input: accumulator.input + next.input,
-		output: accumulator.output + next.output,
-		total: accumulator.total + next.total
-	};
-	const cachedTokens = sumOptional(accumulator.cachedTokens, next.cachedTokens);
-	if (cachedTokens !== void 0) merged.cachedTokens = cachedTokens;
-	const cacheWriteTokens = sumOptional(accumulator.cacheWriteTokens, next.cacheWriteTokens);
-	if (cacheWriteTokens !== void 0) merged.cacheWriteTokens = cacheWriteTokens;
-	const reasoningTokens = sumOptional(accumulator.reasoningTokens, next.reasoningTokens);
-	if (reasoningTokens !== void 0) merged.reasoningTokens = reasoningTokens;
-	const cost = (0, _warlock_js_ai.accumulateCost)(accumulator.cost, next.cost);
-	if (cost !== void 0) merged.cost = cost;
+	const merged = emptyUsage();
+	(0, _warlock_js_ai.mergeUsage)(merged, accumulator);
+	(0, _warlock_js_ai.mergeUsage)(merged, next);
 	return merged;
 }
-/**
-* A zero-valued {@link Usage} to seed an aggregation. Only the required
-* token channels are set; optional channels stay absent until a
-* contributor reports them, preserving the "never reported" vs
-* "reported as 0" distinction.
-*/
+/** Seed an aggregation without claiming any optional counter was reported. */
 function emptyUsage() {
 	return {
 		input: 0,
@@ -550,15 +514,6 @@ function emptyUsage() {
 		total: 0
 	};
 }
-/**
-* Add two optional token counts, treating either side's `undefined` as
-* zero — but return `undefined` when both are absent, so an unreported
-* channel stays unreported rather than collapsing to `0`.
-*/
-function sumOptional(accumulator, next) {
-	if (accumulator === void 0 && next === void 0) return;
-	return (accumulator ?? 0) + (next ?? 0);
-}
 
 //#endregion
 //#region ../../packages/ai-panoptic/src/store/cache-trace-store.ts
@@ -781,12 +736,10 @@ var CacheTraceStore = class {
 				aggregate.completed += 1;
 				break;
 			case "failed":
+			case "incomplete":
 				aggregate.failed += 1;
 				break;
-			case "cancelled":
-				aggregate.cancelled += 1;
-				break;
-			default: break;
+			case "cancelled": aggregate.cancelled += 1;
 		}
 	}
 	/**
@@ -927,7 +880,8 @@ var InMemoryTraceStore = class {
 	* Non-terminal statuses (`awaiting-input`, `max-iterations`) are
 	* counted in `traces` but tracked by none of the three headline
 	* counters — intentional, those three answer the common
-	* "succeeded / errored / aborted" question.
+	* "succeeded / errored / aborted" question. An `incomplete` agent run
+	* (its answer was cut off) errored, so it counts as `failed`.
 	*/
 	countStatus(aggregate, trace) {
 		switch (trace.root.status) {
@@ -935,12 +889,10 @@ var InMemoryTraceStore = class {
 				aggregate.completed += 1;
 				break;
 			case "failed":
+			case "incomplete":
 				aggregate.failed += 1;
 				break;
-			case "cancelled":
-				aggregate.cancelled += 1;
-				break;
-			default: break;
+			case "cancelled": aggregate.cancelled += 1;
 		}
 	}
 	/**
@@ -985,10 +937,8 @@ function createInMemoryTraceStore(options) {
 //#region ../../packages/ai-panoptic/src/exporters/utils/total-cost.ts
 /**
 * Collapse a {@link Usage.cost} breakdown into a single USD scalar by
-* summing every populated field. Mirrors the formula documented on
-* `Usage.cost` (input + output + cachedInput + cachedOutput), and also
-* folds in `reasoning` for forward-safety when a provider prices
-* reasoning tokens as a separate channel. Returns `undefined` when no
+* summing every populated channel (input, output, cache reads, cache
+* writes, reasoning) via `totalCost` from `@warlock.js/ai`. Returns `undefined` when no
 * pricing was attached, so exporters can omit the cost attribute
 * entirely rather than reporting a misleading `0`.
 *
@@ -997,9 +947,7 @@ function createInMemoryTraceStore(options) {
 * // => 0.05
 */
 function totalCostUsd(usage) {
-	const cost = usage.cost;
-	if (!cost) return;
-	return (cost.input ?? 0) + (cost.output ?? 0) + (cost.cachedInput ?? 0) + (cost.cachedOutput ?? 0) + (cost.reasoning ?? 0);
+	return (0, _warlock_js_ai.totalCost)(usage.cost);
 }
 
 //#endregion
@@ -1171,6 +1119,7 @@ function statusMarker(status) {
 		case "completed": return "ok";
 		case "failed": return "ERR";
 		case "cancelled": return "cancel";
+		case "incomplete": return "INCOMPLETE";
 		case "max-iterations": return "max-iter";
 		case "awaiting-input": return "await";
 		default: return status;
@@ -1245,11 +1194,11 @@ function writeSpan(sink, span, depth, io, ioMaxChars) {
 	if (io) for (const line of formatSpanIO(span, depth, ioMaxChars)) writeAtSeverity(sink, span.status, line);
 }
 /**
-* Route a line to `error` when the span failed/cancelled, otherwise to
-* `log`. Keeps healthy traces out of the error stream.
+* Route a line to `error` when the span failed/cancelled/incomplete,
+* otherwise to `log`. Keeps healthy traces out of the error stream.
 */
 function writeAtSeverity(sink, status, line) {
-	if (status === "failed" || status === "cancelled") {
+	if (status === "failed" || status === "cancelled" || status === "incomplete") {
 		sink.error(line);
 		return;
 	}
@@ -1534,7 +1483,7 @@ function ownUsage(span) {
 * rolled-up subtree totals next to an own-usage `usage` block. Cost is
 * omitted for the same reason; Langfuse prices the own tokens itself.
 */
-const LANGFUSE_METADATA_OMIT = new Set([
+const LANGFUSE_METADATA_OMIT = /* @__PURE__ */ new Set([
 	GEN_AI_ATTRIBUTES.usageInputTokens,
 	GEN_AI_ATTRIBUTES.usageOutputTokens,
 	WARLOCK_ATTRIBUTES.totalTokens,
@@ -1556,10 +1505,10 @@ function langfuseMetadata(span) {
 }
 /**
 * Map the Panoptic span status onto a Langfuse observation level —
-* failed/cancelled spans surface as `ERROR`, everything else `DEFAULT`.
+* failed/cancelled/incomplete spans surface as `ERROR`, everything else `DEFAULT`.
 */
 function toLevel(span) {
-	if (span.status === "failed" || span.status === "cancelled") return "ERROR";
+	if (span.status === "failed" || span.status === "cancelled" || span.status === "incomplete") return "ERROR";
 	return "DEFAULT";
 }
 
@@ -1692,7 +1641,7 @@ function applyStatus(otelSpan, span) {
 		message: span.error.message,
 		stack: span.error.stack
 	});
-	if (span.status === "failed" || span.status === "cancelled") {
+	if (span.status === "failed" || span.status === "cancelled" || span.status === "incomplete") {
 		otelSpan.setStatus({
 			code: codes.ERROR,
 			message: span.error?.message
@@ -1934,7 +1883,9 @@ function panoptic(options = {}) {
 * the caller (the dashboard route) is expected to catch.
 */
 async function evaluateSystemPrompt(systemPrompt, config, instructionsOverride) {
-	return (0, _warlock_js_ai.judgePromptBody)(systemPrompt, typeof config.model === "function" ? await config.model() : config.model, instructionsOverride?.trim() || config.instructions);
+	const model = typeof config.model === "function" ? await config.model() : config.model;
+	const instructions = instructionsOverride?.trim() || config.instructions;
+	return (0, _warlock_js_ai.judgePromptBody)(systemPrompt, model, instructions);
 }
 
 //#endregion
@@ -1980,6 +1931,7 @@ const KNOWN_STATUSES = [
 	"completed",
 	"failed",
 	"cancelled",
+	"incomplete",
 	"max-iterations",
 	"awaiting-input",
 	"awaiting-approval"
@@ -2370,13 +2322,14 @@ function dashboardHtml(basePath, title, evaluateEnabled = false, evaluateDefault
 
   var STATUS_FILTERS = ["completed", "failed", "cancelled"];
   var TYPE_FILTERS = ["agent", "tool", "model", "supervisor", "team", "workflow", "orchestrator", "planner", "batch", "callback"];
-  var ERROR_STATUSES = { failed: 1, cancelled: 1 };
+  var ERROR_STATUSES = { failed: 1, cancelled: 1, incomplete: 1 };
   var NO_SESSION_KEY = "(no session)";
   var NO_PROMPT_KEY = "(no prompt)";
   var NO_TYPE_KEY = "(no type)";
 
   var statusClass = function (s) {
     if (s === "completed" || s === "failed" || s === "cancelled") return s;
+    if (s === "incomplete") return "failed";
     return "other";
   };
   var esc = function (v) {
@@ -2464,7 +2417,7 @@ function dashboardHtml(basePath, title, evaluateEnabled = false, evaluateDefault
   // Sum every priced lane of a cost object into one USD number.
   function costSumObj(c) {
     if (!c) return 0;
-    return (c.input || 0) + (c.output || 0) + (c.cachedInput || 0) + (c.cachedOutput || 0) + (c.reasoning || 0);
+    return (c.input || 0) + (c.output || 0) + (c.cachedInput || 0) + (c.cacheWrite || 0) + (c.cachedOutput || 0) + (c.reasoning || 0);
   }
   // The cost a single span directly carries (on its rolled-up usage).
   function usageCost(usage) { return usage ? costSumObj(usage.cost) : 0; }
@@ -2798,7 +2751,7 @@ function dashboardHtml(basePath, title, evaluateEnabled = false, evaluateDefault
   }
 
   function renderStats(a) {
-    var cost = a.cost ? (((a.cost.input || 0) + (a.cost.output || 0) + (a.cost.cachedInput || 0) + (a.cost.cachedOutput || 0)).toFixed(4)) : "—";
+    var cost = a.cost ? (((a.cost.input || 0) + (a.cost.output || 0) + (a.cost.cachedInput || 0) + (a.cost.cacheWrite || 0) + (a.cost.cachedOutput || 0) + (a.cost.reasoning || 0)).toFixed(4)) : "—";
     var u = a.usage || {};
     var cells = [
       ["traces", num(a.traces), null],
@@ -3055,7 +3008,7 @@ function dashboardHtml(basePath, title, evaluateEnabled = false, evaluateDefault
   // "Supervisor Terminated By", "span id" → "Span ID", "agent.trips" →
   // "Agent Trips". Underlying attribute keys are untouched.
   // NB: this whole script is a template literal — regex backslash classes
-  // MUST be double-escaped (\\s, not \s) or "\\s" collapses to a literal "s".
+  // MUST be double-escaped (\\s) so the rendered script retains its backslash.
   function humanizeKey(key) {
     return String(key).split(/[.\\s]+/).map(function (seg) {
       return seg.replace(/([a-z0-9])([A-Z])/g, "$1 $2").split(/\\s+/).map(function (w) {
@@ -3521,7 +3474,7 @@ function encodeForInlineScript(value) {
 //#endregion
 //#region ../../packages/ai-panoptic/src/dashboard/serve.ts
 /** Hard cap on the evaluate route's request body — well beyond a real instructions string. */
-const MAX_EVALUATE_BODY_BYTES = 64 * 1024;
+const MAX_EVALUATE_BODY_BYTES = 65536;
 /**
 * Security response headers added to every dashboard response (S4):
 * block MIME-sniffing, framing, referrer leakage, and lock the page's
@@ -3803,13 +3756,14 @@ function dashboard(store, options = {}) {
 	const basePath = normalizeBasePath(options.basePath);
 	if (!isLoopbackHost(host) && !options.authToken) return Promise.reject(/* @__PURE__ */ new Error(`Panoptic dashboard: binding to a non-loopback host ("${host}") requires an \`authToken\` (the dashboard exposes raw prompt content). Pass \`authToken\`, or bind to 127.0.0.1.`));
 	const allowedHosts = (options.allowedHosts ?? defaultAllowedHosts(host)).map((h) => h.toLowerCase());
-	const server = (0, node_http.createServer)(createRequestHandler(store, {
+	const handler = createRequestHandler(store, {
 		basePath,
 		title,
 		authToken: options.authToken,
 		allowedHosts,
 		evaluate: options.evaluate
-	}));
+	});
+	const server = (0, node_http.createServer)(handler);
 	return new Promise((resolve, reject) => {
 		const onError = (error) => {
 			server.off("error", onError);
@@ -4124,7 +4078,7 @@ function aggregateByType(traces) {
 */
 function costSumObj(cost) {
 	if (cost === void 0) return 0;
-	return (cost.input ?? 0) + (cost.output ?? 0) + (cost.cachedInput ?? 0) + (cost.cachedOutput ?? 0) + (cost.reasoning ?? 0);
+	return (cost.input ?? 0) + (cost.output ?? 0) + (cost.cachedInput ?? 0) + (cost.cacheWrite ?? 0) + (cost.cachedOutput ?? 0) + (cost.reasoning ?? 0);
 }
 /**
 * Rollup-aware subtree cost: a node's own cost when it carries one