@salesforce/sfdx-agent-harness-mastra 0.87.0 → 0.89.0

package/CHANGELOG.md

@@ -3,6 +3,16 @@
 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/).
 
+## [0.89.0] - 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
+- defer previous-workspace destroy while a coordinator is parked so a remembered-approval resume doesn't crash @W-24233008@ ([#834](https://github.com/forcedotcom/agentic-dx/pull/834))
+
 ## [0.87.0] - 2026-09-17
 
 ### Chores

package/dist/mastra-approval-coordinator.js

@@ -242,7 +242,10 @@ export class MastraApprovalCoordinator {
         });
     }
     onTeardown(callback) {
-        if (this.aborted || this.disposed) {
+        // If already terminal, fire now: a callback added after teardown ran
+        // would land in an already-cleared set and never fire. `isSettled()`
+        // covers natural-finish/timeout/pump-error settles (no abort/dispose flag).
+        if (this.aborted || this.disposed || this.isSettled()) {
             queueMicrotask(callback);
             return;
         }

package/dist/mastra-harness.js

@@ -1257,8 +1257,45 @@ export class MastraHarness {
             state.toolSearchProcessor = toolSearchProcessor;
             state.skillSearchProcessor = skillSearchProcessor;
             state.hooks = options?.hooks ?? {};
-            previousSkillSearchProcessor?.dispose();
-            await previousWorkspace?.destroy().catch(() => undefined);
+            // Defer tearing down the previous Agent's workspace (+ its skill-search
+            // processor) until every coordinator still bound to that Agent settles;
+            // do NOT destroy synchronously. A coordinator parked on a tool-approval
+            // resumes against the old Agent and reads `workspace.skills` (in
+            // resumeStream, and via the skill-search processor's getter) — if the
+            // workspace is already destroyed, @mastra/core throws
+            // WorkspaceNotReadyError. Common trigger: an "Allow always" approval,
+            // whose rule-persist calls updateAgentConfig (→ this method) before the
+            // approval settles.
+            //
+            // Two things must BOTH stay for this to be correct: the filter +
+            // `onTeardown` registration run in one synchronous tick (so no
+            // coordinator settles between them and is missed), and `onTeardown`'s
+            // own `isSettled()` arm handles an already-settled coordinator. Bounded
+            // because `destroyAgent()`/`shutdown()` force-dispose coordinators, so
+            // the deferred teardown always eventually runs.
+            const disposePreviousAgentResources = async () => {
+                previousSkillSearchProcessor?.dispose();
+                await previousWorkspace?.destroy();
+            };
+            const onTeardownFailed = (error) => {
+                // Guarded: a deferred teardown can reject after shutdown() disposed
+                // the buses; logging there must not become an unhandled rejection.
+                try {
+                    this.buses.getLogBus()?.debug('Deferred previous-Agent resource teardown failed', { error });
+                }
+                catch {
+                    /* buses already disposed */
+                }
+            };
+            const unsettledCoordinators = [...state.coordinators.values()].filter((c) => !c.isSettled());
+            if (unsettledCoordinators.length === 0) {
+                await disposePreviousAgentResources().catch(onTeardownFailed);
+            }
+            else {
+                void Promise.all(unsettledCoordinators.map((c) => new Promise((resolve) => c.onTeardown(resolve))))
+                    .then(() => disposePreviousAgentResources())
+                    .catch(onTeardownFailed);
+            }
         }
         catch (error) {
             // Phase 2 failed after Phase 1 mutated state.mcpServers. We don't

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/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@salesforce/sfdx-agent-harness-mastra",
-  "version": "0.87.0",
+  "version": "0.89.0",
   "description": "Mastra-backed AgentHarness implementation for @salesforce/sfdx-agent-sdk",
   "type": "module",
   "main": "dist/index.js",
@@ -42,12 +42,12 @@
     "openai": "^7.17.0"
   },
   "peerDependencies": {
-    "@salesforce/sfdx-agent-sdk": "0.88.0"
+    "@salesforce/sfdx-agent-sdk": "0.90.0"
   },
   "devDependencies": {
     "@eslint/js": "^10.0.1",
     "@salesforce/harness-conformance": "0.0.1",
-    "@salesforce/sfdx-agent-sdk": "0.88.0",
+    "@salesforce/sfdx-agent-sdk": "0.90.0",
     "@types/node": "^22.20.1",
     "@vitest/coverage-istanbul": "^5.0.1",
     "@vitest/eslint-plugin": "^1.6.27",