@salesforce/sfdx-agent-harness-mastra 0.88.0 → 0.89.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -3,6 +3,11 @@
3
3
  All notable changes to `@salesforce/sfdx-agent-harness-mastra` are documented in this file.
4
4
  Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
5
5
 
6
+ ## [0.89.0] - 2026-09-21
7
+
8
+ ### Features
9
+ - **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))
10
+
6
11
  ## [0.88.0] - 2026-09-18
7
12
 
8
13
  ### Fixes
@@ -308,6 +308,16 @@ function applyThinkingConfig(params, model, providerHint) {
308
308
  }
309
309
  async function doGenerate(ctx, options) {
310
310
  const { client, commonParams, toolSchemas } = await buildAnthropicRequest(ctx, options);
311
+ // Clamp max_tokens to a summary-safe ceiling on this non-streaming path. The
312
+ // Anthropic SDK refuses a non-streaming request whose max_tokens implies a
313
+ // completion that could exceed 10 minutes (`calculateNonstreamingTimeout`,
314
+ // threshold ~21k tokens). doGenerate's only caller is compactThread's
315
+ // summarize() (see the note below) — a conversation summary never needs the
316
+ // model's full output budget — so a model with a large cap (Opus 4.x @ 128k,
317
+ // Sonnet 4.x @ 64k) would otherwise throw here. `Math.min` leaves a smaller
318
+ // cap untouched; any thinking budget buildAnthropicRequest applied is <= 4096,
319
+ // safely under this floor.
320
+ const summaryMaxTokens = Math.min(commonParams.max_tokens, SUMMARY_MAX_OUTPUT_TOKENS);
311
321
  // Forward the turn's AbortSignal into the SDK's RequestOptions so firing the
312
322
  // controller (POST /cancel, idle timeout, shutdown abortAll) actually aborts
313
323
  // the outbound gateway fetch, not just harness bookkeeping (W-23988781).
@@ -318,7 +328,7 @@ async function doGenerate(ctx, options) {
318
328
  // as COMPACTION_FAILED. Error classification here targets the streaming turn
319
329
  // path (8d) — a raw create() rejection propagating uncoded is correct for the
320
330
  // summarize-only path.
321
- const message = await client.messages.create({ ...commonParams, stream: false }, { signal: options.abortSignal });
331
+ const message = await client.messages.create({ ...commonParams, max_tokens: summaryMaxTokens, stream: false }, { signal: options.abortSignal });
322
332
  const content = [];
323
333
  let sawToolCalls = false;
324
334
  for (const block of message.content) {
@@ -564,6 +574,17 @@ async function doStream(ctx, options) {
564
574
  * API requires the field; Mastra leaves it unset for many turns. 4096 mirrors
565
575
  * the `@ai-sdk/anthropic` provider's default. */
566
576
  const DEFAULT_MAX_TOKENS = 4096;
577
+ /**
578
+ * Ceiling for the non-streaming summary request `doGenerate` issues for `compactThread`.
579
+ * The Anthropic SDK refuses a non-streaming request whose max_tokens implies a >10-minute
580
+ * completion — `expectedTime = 3.6e6 * maxTokens / 128000 ms`, so the guard trips above
581
+ * ~21,333 tokens. 16384 sits safely under that (~461s est., ~23% margin), stays above the
582
+ * max thinking budget (4096), and matches the historical working summary budget for the
583
+ * previous default model (Sonnet 4.6 @ 16384) — so a large per-model output cap
584
+ * (Opus 4.x @ 128k, Sonnet 4.x @ 64k) is clamped here instead of throwing. A conversation
585
+ * summary never needs more.
586
+ */
587
+ const SUMMARY_MAX_OUTPUT_TOKENS = 16384;
567
588
  /**
568
589
  * Translates a V3 `LanguageModelV3Prompt` into Anthropic's Messages API shape
569
590
  * — top-level `system` string + `MessageParam[]` of role-tagged content blocks.
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@salesforce/sfdx-agent-harness-mastra",
3
- "version": "0.88.0",
3
+ "version": "0.89.0",
4
4
  "description": "Mastra-backed AgentHarness implementation for @salesforce/sfdx-agent-sdk",
5
5
  "type": "module",
6
6
  "main": "dist/index.js",
@@ -42,12 +42,12 @@
42
42
  "openai": "^7.17.0"
43
43
  },
44
44
  "peerDependencies": {
45
- "@salesforce/sfdx-agent-sdk": "0.89.0"
45
+ "@salesforce/sfdx-agent-sdk": "0.90.0"
46
46
  },
47
47
  "devDependencies": {
48
48
  "@eslint/js": "^10.0.1",
49
49
  "@salesforce/harness-conformance": "0.0.1",
50
- "@salesforce/sfdx-agent-sdk": "0.89.0",
50
+ "@salesforce/sfdx-agent-sdk": "0.90.0",
51
51
  "@types/node": "^22.20.1",
52
52
  "@vitest/coverage-istanbul": "^5.0.1",
53
53
  "@vitest/eslint-plugin": "^1.6.27",