@salesforce/sfdx-agent-harness-mastra 0.88.0 → 0.90.0

package/CHANGELOG.md

@@ -3,6 +3,19 @@
 All notable changes to `@salesforce/sfdx-agent-harness-mastra` are documented in this file.
 Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
+## [0.90.0] - 2026-09-23
+
+### Fixes
+- **harness-mastra**: emit full llm-request body in wire-communication events (remove 256KB truncation) @W-24280587@ ([#839](https://github.com/forcedotcom/agentic-dx/pull/839))
+
+### Chores
+- **deps**: bump @ai-sdk/provider from 4.0.4 to 4.0.14 ([#838](https://github.com/forcedotcom/agentic-dx/pull/838))
+
+## [0.89.0] - 2026-09-21
+
+### Features
+- **agent-sdk,harness-mastra**: add GPT-5.6 + Claude Sonnet 5 to model registry; fix stale output caps + compaction max_tokens @W-23949616@ ([#833](https://github.com/forcedotcom/agentic-dx/pull/833))
+
 ## [0.88.0] - 2026-09-18
 
 ### Fixes

package/dist/pass-through-anthropic.js

@@ -308,6 +308,16 @@ function applyThinkingConfig(params, model, providerHint) {
 }
 async function doGenerate(ctx, options) {
     const { client, commonParams, toolSchemas } = await buildAnthropicRequest(ctx, options);
+    // Clamp max_tokens to a summary-safe ceiling on this non-streaming path. The
+    // Anthropic SDK refuses a non-streaming request whose max_tokens implies a
+    // completion that could exceed 10 minutes (`calculateNonstreamingTimeout`,
+    // threshold ~21k tokens). doGenerate's only caller is compactThread's
+    // summarize() (see the note below) — a conversation summary never needs the
+    // model's full output budget — so a model with a large cap (Opus 4.x @ 128k,
+    // Sonnet 4.x @ 64k) would otherwise throw here. `Math.min` leaves a smaller
+    // cap untouched; any thinking budget buildAnthropicRequest applied is <= 4096,
+    // safely under this floor.
+    const summaryMaxTokens = Math.min(commonParams.max_tokens, SUMMARY_MAX_OUTPUT_TOKENS);
     // Forward the turn's AbortSignal into the SDK's RequestOptions so firing the
     // controller (POST /cancel, idle timeout, shutdown abortAll) actually aborts
     // the outbound gateway fetch, not just harness bookkeeping (W-23988781).
@@ -318,7 +328,7 @@ async function doGenerate(ctx, options) {
     // as COMPACTION_FAILED. Error classification here targets the streaming turn
     // path (8d) — a raw create() rejection propagating uncoded is correct for the
     // summarize-only path.
-    const message = await client.messages.create({ ...commonParams, stream: false }, { signal: options.abortSignal });
+    const message = await client.messages.create({ ...commonParams, max_tokens: summaryMaxTokens, stream: false }, { signal: options.abortSignal });
     const content = [];
     let sawToolCalls = false;
     for (const block of message.content) {
@@ -564,6 +574,17 @@ async function doStream(ctx, options) {
  * API requires the field; Mastra leaves it unset for many turns. 4096 mirrors
  * the `@ai-sdk/anthropic` provider's default. */
 const DEFAULT_MAX_TOKENS = 4096;
+/**
+ * Ceiling for the non-streaming summary request `doGenerate` issues for `compactThread`.
+ * The Anthropic SDK refuses a non-streaming request whose max_tokens implies a >10-minute
+ * completion — `expectedTime = 3.6e6 * maxTokens / 128000 ms`, so the guard trips above
+ * ~21,333 tokens. 16384 sits safely under that (~461s est., ~23% margin), stays above the
+ * max thinking budget (4096), and matches the historical working summary budget for the
+ * previous default model (Sonnet 4.6 @ 16384) — so a large per-model output cap
+ * (Opus 4.x @ 128k, Sonnet 4.x @ 64k) is clamped here instead of throwing. A conversation
+ * summary never needs more.
+ */
+const SUMMARY_MAX_OUTPUT_TOKENS = 16384;
 /**
  * Translates a V3 `LanguageModelV3Prompt` into Anthropic's Messages API shape
  * — top-level `system` string + `MessageParam[]` of role-tagged content blocks.

package/dist/v3-stream-mapping.d.ts

@@ -280,11 +280,13 @@ export type GatewayFetchWireHooks = {
  * fetch resolution to HTTP-body terminality: an owned tee branch observes only
  * recognized assistant text deltas without blocking the provider SDK's branch.
  * Wire-communication events are best-effort: a failed body decode does NOT
- * prevent the call from completing. Body capture is
- * intentionally bounded — only requests with a JSON body shorter than 256 KB
- * are decoded; longer bodies are recorded as `{ truncated: true, byteLength:
- * <n> }` so wire-communication captures don't blow the consumer's log file
- * budget.
+ * prevent the call from completing. When a subscriber is present the request
+ * body is captured in full with no size cap (W-24280587) — the channel is
+ * opt-in and diagnostic, so fidelity wins; with no subscriber the decode is
+ * skipped entirely (the event is dropped by the wire bus anyway), so a
+ * subscriber-less production process pays no full-body parse on the hot path.
+ * (The streaming *response* observer separately bounds its in-memory text
+ * capture at `TRACE_BODY_BYTE_BUDGET`.)
  */
 export declare function createGatewayFetch(getHeaders: () => Record<string, string> | Promise<Record<string, string>>, wireHooks?: GatewayFetchWireHooks, innerFetch?: typeof fetch, 
 /**

package/dist/v3-stream-mapping.js

@@ -353,11 +353,13 @@ export function mapOpenAIUsage(usage) {
  * fetch resolution to HTTP-body terminality: an owned tee branch observes only
  * recognized assistant text deltas without blocking the provider SDK's branch.
  * Wire-communication events are best-effort: a failed body decode does NOT
- * prevent the call from completing. Body capture is
- * intentionally bounded — only requests with a JSON body shorter than 256 KB
- * are decoded; longer bodies are recorded as `{ truncated: true, byteLength:
- * <n> }` so wire-communication captures don't blow the consumer's log file
- * budget.
+ * prevent the call from completing. When a subscriber is present the request
+ * body is captured in full with no size cap (W-24280587) — the channel is
+ * opt-in and diagnostic, so fidelity wins; with no subscriber the decode is
+ * skipped entirely (the event is dropped by the wire bus anyway), so a
+ * subscriber-less production process pays no full-body parse on the hot path.
+ * (The streaming *response* observer separately bounds its in-memory text
+ * capture at `TRACE_BODY_BYTE_BUDGET`.)
  */
 export function createGatewayFetch(getHeaders, wireHooks, innerFetch = globalThis.fetch, 
 /**
@@ -411,6 +413,13 @@ logBus) {
         const startedAt = clock.now();
         const correlationId = (wireHooks.idGenerator ?? new UUIDGenerator()).getUniqueId();
         const requestStart = performance.now();
+        // Decode the (now uncapped) request body ONLY when a subscriber is
+        // listening — mirroring the response-observer gating below. In production
+        // `wireHooks` is always supplied but subscribers are usually absent, and a
+        // real turn's body is large, so an unconditional full-body JSON.parse would
+        // be pure hot-path waste for a subscriber-less process (the event is dropped
+        // by the wire bus anyway). Fidelity for actual subscribers is unchanged.
+        const hasWireSubscribers = wireHooks.hasWireCommunicationSubscribers?.() ?? true;
         wireHooks.emitWireCommunication({
             type: 'llm-request',
             timestamp: startedAt,
@@ -418,7 +427,7 @@ logBus) {
             method,
             model,
             correlationId,
-            body: tryDecodeBody(finalInit.body),
+            body: hasWireSubscribers ? tryDecodeBody(finalInit.body) : undefined,
         });
         try {
             const response = await innerFetch(input, finalInit);
@@ -791,10 +800,11 @@ function tryDecodeBody(body) {
         return undefined;
     if (typeof body !== 'string')
         return undefined; // Streams / FormData / Blob: skip.
-    const byteLength = Buffer.byteLength(body, 'utf8');
-    if (byteLength > TRACE_BODY_BYTE_BUDGET) {
-        return { truncated: true, byteLength };
-    }
+    // No size cap: the wire-communication channel is opt-in, subscriber-gated, and
+    // purely diagnostic — its whole point is fidelity, so emit the full request body
+    // (system / messages / tools) verbatim. (W-24280587; a 256 KB cap here used to
+    // replace large bodies with a `{ truncated, byteLength }` stub, hiding exactly
+    // the fields the channel exists to show.)
     try {
         const parsed = JSON.parse(body);
         return typeof parsed === 'object' && parsed !== null ? parsed : { value: parsed };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@salesforce/sfdx-agent-harness-mastra",
-  "version": "0.88.0",
+  "version": "0.90.0",
   "description": "Mastra-backed AgentHarness implementation for @salesforce/sfdx-agent-sdk",
   "type": "module",
   "main": "dist/index.js",
@@ -29,7 +29,7 @@
     "LICENSE.txt"
   ],
   "dependencies": {
-    "@ai-sdk/provider": "4.0.4",
+    "@ai-sdk/provider": "4.0.14",
     "@anthropic-ai/bedrock-sdk": "^0.33.6",
     "@anthropic-ai/sdk": "^0.126.0",
     "@mastra/core": "^1.67.0",
@@ -42,12 +42,12 @@
     "openai": "^7.17.0"
   },
   "peerDependencies": {
-    "@salesforce/sfdx-agent-sdk": "0.89.0"
+    "@salesforce/sfdx-agent-sdk": "0.91.0"
   },
   "devDependencies": {
     "@eslint/js": "^10.0.1",
     "@salesforce/harness-conformance": "0.0.1",
-    "@salesforce/sfdx-agent-sdk": "0.89.0",
+    "@salesforce/sfdx-agent-sdk": "0.91.0",
     "@types/node": "^22.20.1",
     "@vitest/coverage-istanbul": "^5.0.1",
     "@vitest/eslint-plugin": "^1.6.27",