@poncho-ai/harness 0.52.0 → 0.52.2

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/harness@0.52.0 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
+> @poncho-ai/harness@0.52.2 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
 > node scripts/embed-docs.js && tsup src/index.ts --format esm --dts
 
 [embed-docs] Generated poncho-docs.ts with 4 topics
@@ -8,9 +8,9 @@
 CLI tsup v8.5.1
 CLI Target: es2022
 ESM Build start
-ESM dist/index.js            536.18 KB
+ESM dist/index.js            540.60 KB
 ESM dist/isolate-F2PPSUL6.js 53.82 KB
-ESM ⚡️ Build success in 234ms
+ESM ⚡️ Build success in 204ms
 DTS Build start
-DTS ⚡️ Build success in 7615ms
-DTS dist/index.d.ts 92.18 KB
+DTS ⚡️ Build success in 7242ms
+DTS dist/index.d.ts 93.60 KB

package/CHANGELOG.md

@@ -1,5 +1,85 @@
 # @poncho-ai/harness
 
+## 0.52.2
+
+### Patch Changes
+
+- [#124](https://github.com/cesr/poncho-ai/pull/124) [`4ae26e0`](https://github.com/cesr/poncho-ai/commit/4ae26e0d8d2788f57411f9c17e10766769514f9b) Thanks [@cesr](https://github.com/cesr)! - harness: postgres retry covers exec/transaction + 3 attempts + tighter idle
+
+  Follow-up to the previous `idle_timeout`/`max_lifetime`/retry patch.
+  Live testing on Railway showed the previous values weren't tight
+  enough — `write CONNECTION_ENDED postgres.railway.internal:5432`
+  still surfaced both during user-facing chat turns and during
+  subagent auto-callback reruns, despite the new config and the
+  one-shot retry.
+
+  Two failure modes the previous version didn't cover:
+  1. The retry only wrapped `private query()` (executor.run/get/all),
+     but `executor.exec` (`sql.unsafe`) and `executor.transaction`
+     (`sql.begin`) called the postgres.js client directly. A pg drop
+     inside a transaction or migration write threw straight through.
+  2. After an idle period the pool can have multiple stale sockets;
+     a single retry can checkout a second stale socket from the pool
+     and fail again. One-shot retry exhausted into an error visible
+     to the caller.
+
+  Fixes:
+  - All three executor paths (`run/get/all`, `exec`, `transaction`)
+    now go through the same `runWithRetry` wrapper. Transactions
+    only retry the connection-level `CONNECTION_ENDED` reject from
+    the postgres.js client — actual SQL errors mid-transaction
+    surface as a different error class and bypass the retry,
+    preserving atomic semantics.
+  - Three attempts with light exponential backoff (0, 50ms, 200ms).
+    Enough to ride out a typical staleness wave; if all three fail
+    the network is genuinely broken.
+  - `CONNECT_TIMEOUT` and `ECONNRESET` added to the retry-eligible
+    error codes.
+
+  Config knobs tightened:
+  - `idle_timeout: 5` (was 20). Empirically Railway's pg drops
+    sockets well before 20s; 5s wins the race in practice while
+    staying long enough for bursty workloads to reuse connections.
+  - `max_lifetime: 300` (was 600). Same reasoning — recycle more
+    aggressively.
+  - `connect_timeout: 10` (was 30 default). Faster failure during
+    incidents lets callers shed load instead of stacking up.
+
+- [#144](https://github.com/cesr/poncho-ai/pull/144) [`28d640b`](https://github.com/cesr/poncho-ai/commit/28d640b2f82ea780f8e0be90965972d9903c01d7) Thanks [@cesr](https://github.com/cesr)! - orchestrator: make subagent result delivery reliable
+
+  Subagent results could silently never reach the parent agent. Several
+  plumbing bugs in `runSubagent` / `runSubagentContinuation`:
+  - **Emit-before-persist race.** `subagent:completed` / `subagent:error`
+    were emitted to the parent's event stream _before_ the result was
+    written to the store, so a consumer reacting to the event (the parent
+    callback, the streaming client) could race the write. Now the result
+    is persisted first, then the event is emitted.
+  - **Silently swallowed writes.** Two `appendSubagentResult(...).catch(() => {})`
+    call sites (the error path and the continuation-error path) dropped the
+    result with no trace on a transient store failure. Replaced with a
+    shared `appendSubagentResultReliable` helper that retries once and then
+    logs loudly — a dropped result is the worst failure mode (the parent
+    waits forever on a subagent it thinks is still running).
+  - **Un-awaited eventSink.** The subagent-callback run path was the lone
+    `this.eventSink(...)` call site that didn't `await` (every other site
+    does), so callback-turn events could interleave out of order. Now awaited.
+  - **Spawn rejections went to a bare `console.error`.** A background
+    `runSubagent` that rejected outside its own try/catch left the parent
+    hanging. Both fire-and-forget spawn paths now route to a
+    `handleSpawnFailure` that marks the child errored and hands the parent
+    an error result so the turn can resume.
+  - **`recoverStaleSubagents` now also drains undelivered results.** It
+    previously only rescued children stuck in `running`; it now also
+    re-triggers the parent callback for any parent that has results sitting
+    in the store with no active run (e.g. a result persisted just before a
+    process restart, whose in-memory callback trigger was lost).
+
+## 0.52.1
+
+### Patch Changes
+
+- [`0e8fff1`](https://github.com/cesr/poncho-ai/commit/0e8fff12aed9d5efe1821ed3560ead48a16113c1) Thanks [@cesr](https://github.com/cesr)! - Only send `temperature` to the model when the agent explicitly sets one. The harness previously defaulted to `temperature: 0.2` and always passed it to `streamText`, which returns a 400 ("`temperature` is deprecated for this model") on models that removed sampling params (Fable 5, Opus 4.7+). `temperature` is now omitted from the request when undefined — the same treatment `maxTokens` already had — and `defaultAgentDefinition` no longer hard-codes a `temperature` line into the generated frontmatter (pass `temperature` explicitly to set one).
+
 ## 0.52.0
 
 ### Minor Changes

package/dist/index.d.ts

@@ -737,7 +737,12 @@ interface DefaultAgentDefinitionOptions {
     modelProvider?: "anthropic" | "openai" | "openai-codex";
     /** Model name. Default: "claude-opus-4-5". */
     modelName?: string;
-    /** Sampling temperature. Default: 0.2. */
+    /**
+     * Sampling temperature. When unset, it is omitted from the generated
+     * frontmatter entirely and the harness sends no temperature (provider
+     * default). Newer models (Fable 5, Opus 4.7+) reject `temperature` — leave
+     * this unset for them.
+     */
     temperature?: number;
     /** Max tool-call steps per run. Default: 20. */
     maxSteps?: number;
@@ -1797,22 +1802,27 @@ declare class PostgresEngine extends SqlStorageEngine {
     private patchVfs;
     private query;
     /**
-     * Single retry on a transient connection-layer failure. The
-     * `idle_timeout` / `max_lifetime` config above prevents *most*
-     * stale-connection cases, but a query can still race a
-     * provider-initiated drop in flight — the postgres.js client
-     * rejects with `code: "CONNECTION_ENDED"` and the next attempt
-     * checks out a fresh connection from the pool. One retry is
-     * enough; if it fails again the host-side network is genuinely
-     * broken and the caller should see the error.
+     * Retry on transient connection-layer failures. Three attempts
+     * with exponential-ish backoff (0, 50ms, 200ms) — the pool may
+     * have multiple stale sockets accumulated during an idle period
+     * (especially on managed Postgres after boot when no traffic
+     * has flowed for a while), so a single retry can land on a
+     * second stale socket and still fail. Three attempts virtually
+     * always exhausts the staleness wave; if all three throw, the
+     * failure is real and the caller should see it.
      *
-     * Only retries reads + the standard exec/run paths in `query`;
-     * `sql.unsafe(sql)` calls in `executeRaw` (migration DDL) and
-     * `sql.begin(...)` transactions are unwrapped — those are
-     * idempotent-by-construction (DDL is `IF NOT EXISTS`) or
-     * atomically scoped (transactions roll back cleanly), and adding
-     * a retry around them would complicate the transaction
-     * semantics.
+     * Applied to every pg path the executor exposes:
+     *  - `query()` (run/get/all)  — natural retry: queries are
+     *    idempotent at the connection-failure boundary because the
+     *    server-side rollback runs cleanly on socket close.
+     *  - `exec(sql)` for DDL      — `CREATE TABLE IF NOT EXISTS` and
+     *    friends are idempotent by construction.
+     *  - `transaction(fn)`        — only retried when the
+     *    CONNECTION_ENDED reject arrives *before* the transaction
+     *    body started executing on the connection; if it errors
+     *    mid-transaction, the postgres.js client surfaces a
+     *    different error class (the inner SQL error) and bypasses
+     *    this retry, preserving the all-or-nothing semantics.
      */
     private runWithRetry;
     private addToPathCache;
@@ -2140,6 +2150,22 @@ declare class AgentOrchestrator {
     processSubagentCallback(conversationId: string, skipLockCheck?: boolean): Promise<void>;
     runSubagentContinuation(conversationId: string, conversation: Conversation, continuationMessages: Message[]): AsyncGenerator<AgentEvent>;
     createSubagentManager(): SubagentManager;
+    /**
+     * Append a subagent result to its parent, retrying once on a transient
+     * store failure before giving up loudly. A silently dropped result is the
+     * worst subagent failure mode — the parent waits forever on a subagent it
+     * thinks is still running — so this never swallows the error the way the
+     * old `.catch(() => {})` call sites did. Returns whether the result landed.
+     */
+    private appendSubagentResultReliable;
+    /**
+     * A subagent's fire-and-forget background run rejected outside its own
+     * error handling (e.g. it threw before entering its try block, or the
+     * catch block itself threw). Without this the parent is left waiting on a
+     * subagent that will never report back. Record the failure on the child
+     * and hand the parent an error result so the turn can resume.
+     */
+    private handleSpawnFailure;
     recoverStaleSubagents(): Promise<void>;
 }
 

package/dist/index.js

@@ -588,7 +588,8 @@ var defaultAgentDefinition = (opts = {}) => {
   const description = opts.description ?? DEFAULT_AGENT_DESCRIPTION;
   const modelProvider = opts.modelProvider ?? DEFAULT_MODEL_PROVIDER;
   const modelName = opts.modelName ?? DEFAULT_MODEL_NAME;
-  const temperature = opts.temperature ?? DEFAULT_TEMPERATURE;
+  const temperatureLine = opts.temperature !== void 0 ? `
+  temperature: ${opts.temperature}` : "";
   const maxSteps = opts.maxSteps ?? DEFAULT_MAX_STEPS;
   const timeout = opts.timeout ?? DEFAULT_TIMEOUT;
   return `---
@@ -597,8 +598,7 @@ id: ${id}
 description: ${description}
 model:
   provider: ${modelProvider}
-  name: ${modelName}
-  temperature: ${temperature}
+  name: ${modelName}${temperatureLine}
 limits:
   maxSteps: ${maxSteps}
   timeout: ${timeout}
@@ -4415,12 +4415,12 @@ var PostgresEngine = class extends SqlStorageEngine {
         return rows;
       },
       exec: async (sql) => {
-        await this.sql.unsafe(sql);
+        await this.runWithRetry(() => this.sql.unsafe(sql));
       },
       transaction: async (fn) => {
-        await this.sql.begin(async () => {
+        await this.runWithRetry(() => this.sql.begin(async () => {
           await fn();
-        });
+        }));
       }
     };
   }
@@ -4438,25 +4438,34 @@ var PostgresEngine = class extends SqlStorageEngine {
       prepare: false,
       // Connection-pool resilience. Managed Postgres providers
       // (Railway, Neon, Heroku, etc.) routinely drop idle TCP
-      // connections server-side after a few minutes. Without these
-      // knobs, porsager/postgres keeps stale sockets in the pool;
-      // the next query on one rejects with
-      // `write CONNECTION_ENDED <host>:5432` at `durMs=0`, surfacing
-      // as a hard failure to the caller. Two complementary settings:
+      // connections server-side after a few minutes — and on
+      // Railway in particular, mid-stream drops within a few
+      // seconds of inactivity are common. Without these knobs,
+      // porsager/postgres keeps stale sockets in the pool; the
+      // next query on one rejects with
+      // `write CONNECTION_ENDED <host>:5432` at `durMs=0`,
+      // surfacing as a hard failure to the caller.
       //
-      //   - `idle_timeout: 20` closes idle connections client-side
-      //     after 20s, before any reasonable provider-side timer
-      //     fires. Fresh connection on next checkout = no stale
-      //     socket race.
-      //   - `max_lifetime: 600` (10 min) recycles long-lived
-      //     connections defensively even if they've stayed busy,
-      //     which sidesteps a separate class of provider-side
-      //     "max connection age" limits.
+      //   - `idle_timeout: 5` closes idle connections client-side
+      //     aggressively. Empirically Railway's pg drops sockets
+      //     well before the 20s value that managed-provider docs
+      //     suggest; 5s is short enough to win the race in
+      //     practice while staying long enough that bursty
+      //     workloads still get connection reuse.
+      //   - `max_lifetime: 300` (5 min) recycles long-lived
+      //     connections defensively. Even with idle_timeout, a
+      //     connection that's been actively serving small queries
+      //     for an hour can hit provider-side max-age limits.
+      //   - `connect_timeout: 10` — slightly less patient on
+      //     initial connect than the 30s default. Combined with
+      //     the retry below, "connection refused" surfaces faster
+      //     during incidents and the caller can shed load instead
+      //     of stacking up.
       //
-      // Defaults remain `max: 10`, `connect_timeout: 30` — leaving
-      // pool size + initial connect behavior unchanged.
-      idle_timeout: 20,
-      max_lifetime: 60 * 10
+      // Pool size (`max: 10`) unchanged.
+      idle_timeout: 5,
+      max_lifetime: 60 * 5,
+      connect_timeout: 10
     });
   }
   async initialize() {
@@ -4505,33 +4514,47 @@ var PostgresEngine = class extends SqlStorageEngine {
     );
   }
   /**
-   * Single retry on a transient connection-layer failure. The
-   * `idle_timeout` / `max_lifetime` config above prevents *most*
-   * stale-connection cases, but a query can still race a
-   * provider-initiated drop in flight — the postgres.js client
-   * rejects with `code: "CONNECTION_ENDED"` and the next attempt
-   * checks out a fresh connection from the pool. One retry is
-   * enough; if it fails again the host-side network is genuinely
-   * broken and the caller should see the error.
+   * Retry on transient connection-layer failures. Three attempts
+   * with exponential-ish backoff (0, 50ms, 200ms) — the pool may
+   * have multiple stale sockets accumulated during an idle period
+   * (especially on managed Postgres after boot when no traffic
+   * has flowed for a while), so a single retry can land on a
+   * second stale socket and still fail. Three attempts virtually
+   * always exhausts the staleness wave; if all three throw, the
+   * failure is real and the caller should see it.
    *
-   * Only retries reads + the standard exec/run paths in `query`;
-   * `sql.unsafe(sql)` calls in `executeRaw` (migration DDL) and
-   * `sql.begin(...)` transactions are unwrapped — those are
-   * idempotent-by-construction (DDL is `IF NOT EXISTS`) or
-   * atomically scoped (transactions roll back cleanly), and adding
-   * a retry around them would complicate the transaction
-   * semantics.
+   * Applied to every pg path the executor exposes:
+   *  - `query()` (run/get/all)  — natural retry: queries are
+   *    idempotent at the connection-failure boundary because the
+   *    server-side rollback runs cleanly on socket close.
+   *  - `exec(sql)` for DDL      — `CREATE TABLE IF NOT EXISTS` and
+   *    friends are idempotent by construction.
+   *  - `transaction(fn)`        — only retried when the
+   *    CONNECTION_ENDED reject arrives *before* the transaction
+   *    body started executing on the connection; if it errors
+   *    mid-transaction, the postgres.js client surfaces a
+   *    different error class (the inner SQL error) and bypasses
+   *    this retry, preserving the all-or-nothing semantics.
    */
   async runWithRetry(fn) {
-    try {
-      return await fn();
-    } catch (err) {
-      const code = err?.code;
-      if (code === "CONNECTION_ENDED" || code === "CONNECTION_CLOSED" || code === "CONNECTION_DESTROYED") {
+    const backoffs = [0, 50, 200];
+    let lastErr;
+    for (let attempt = 0; attempt < backoffs.length; attempt++) {
+      if (backoffs[attempt] > 0) {
+        await new Promise((r) => setTimeout(r, backoffs[attempt]));
+      }
+      try {
         return await fn();
+      } catch (err) {
+        lastErr = err;
+        const code = err?.code;
+        if (code === "CONNECTION_ENDED" || code === "CONNECTION_CLOSED" || code === "CONNECTION_DESTROYED" || code === "CONNECT_TIMEOUT" || code === "ECONNRESET") {
+          continue;
+        }
+        throw err;
       }
-      throw err;
     }
+    throw lastErr;
   }
   addToPathCache(tenantId, path) {
     const paths = this.pathCache.get(tenantId);
@@ -10832,7 +10855,7 @@ ${textContent}` };
           cachedCoreMessages = [...cachedCoreMessages, ...newCoreMessages];
           convertedUpTo = messages.length;
           const coreMessages = cachedCoreMessages;
-          const temperature = agent.frontmatter.model?.temperature ?? 0.2;
+          const temperature = agent.frontmatter.model?.temperature;
           const maxTokens = agent.frontmatter.model?.maxTokens;
           const cachedMessages = skipTailCache ? coreMessages : addPromptCacheBreakpoints(
             coreMessages,
@@ -10860,7 +10883,7 @@ ${textContent}` };
             ...useStaticCache ? {} : { system: systemPrompt },
             messages: messagesForStep,
             tools: toolsForStep,
-            temperature,
+            ...typeof temperature === "number" ? { temperature } : {},
             abortSignal: input.abortSignal,
             ...typeof maxTokens === "number" ? { maxTokens } : {},
             experimental_telemetry: {
@@ -13077,12 +13100,6 @@ var AgentOrchestrator = class {
         };
         await this.conversationStore.update(conv);
       }
-      this.hooks?.onStreamEnd?.(childConversationId);
-      await this.eventSink(parentConversationId, {
-        type: "subagent:completed",
-        subagentId: childConversationId,
-        conversationId: childConversationId
-      });
       let gathered = realResponseText(runResult?.response) || realResponseText(draft.assistantResponse);
       if (!gathered) {
         const freshSubConv = await this.conversationStore.get(childConversationId);
@@ -13104,7 +13121,13 @@ var AgentOrchestrator = class {
         ...abnormal ? { error: { code: runError?.code ?? "SUBAGENT_INCOMPLETE", message: runError?.message ?? "subagent ended without a result" } } : {},
         timestamp: Date.now()
       };
-      await this.conversationStore.appendSubagentResult(parentConversationId, pendingResult);
+      await this.appendSubagentResultReliable(parentConversationId, pendingResult);
+      this.hooks?.onStreamEnd?.(childConversationId);
+      await this.eventSink(parentConversationId, {
+        type: "subagent:completed",
+        subagentId: childConversationId,
+        conversationId: childConversationId
+      });
       this.triggerParentCallback(parentConversationId).catch(
         (err) => console.error(`[poncho][subagent] Parent callback failed:`, err instanceof Error ? err.message : err)
       );
@@ -13121,13 +13144,6 @@ var AgentOrchestrator = class {
         conv.updatedAt = Date.now();
         await this.conversationStore.update(conv);
       }
-      this.hooks?.onStreamEnd?.(childConversationId);
-      await this.eventSink(parentConversationId, {
-        type: "subagent:error",
-        subagentId: childConversationId,
-        conversationId: childConversationId,
-        error: errMsg
-      });
       const pendingResult = {
         subagentId: childConversationId,
         task,
@@ -13135,7 +13151,13 @@ var AgentOrchestrator = class {
         error: { code: "SUBAGENT_ERROR", message: errMsg },
         timestamp: Date.now()
       };
-      await this.conversationStore.appendSubagentResult(parentConversationId, pendingResult).catch(() => {
+      await this.appendSubagentResultReliable(parentConversationId, pendingResult);
+      this.hooks?.onStreamEnd?.(childConversationId);
+      await this.eventSink(parentConversationId, {
+        type: "subagent:error",
+        subagentId: childConversationId,
+        conversationId: childConversationId,
+        error: errMsg
       });
       this.triggerParentCallback(parentConversationId).catch(
         (err2) => console.error(`[poncho][subagent] Parent callback failed:`, err2 instanceof Error ? err2.message : err2)
@@ -13251,12 +13273,12 @@ ${resultBody}`,
         },
         initialContextTokens: conversation.contextTokens ?? 0,
         initialContextWindow: conversation.contextWindow ?? 0,
-        onEvent: (event) => {
+        onEvent: async (event) => {
           if (event.type === "run:started") {
             const active = this.activeConversationRuns.get(conversationId);
             if (active) active.runId = event.runId;
           }
-          this.eventSink(conversationId, event);
+          await this.eventSink(conversationId, event);
         }
       });
       flushTurnDraft(execution.draft);
@@ -13442,11 +13464,6 @@ ${resultBody}`,
         await this.conversationStore.update(conv);
       }
       this.activeSubagentRuns.delete(conversationId);
-      await this.eventSink(parentConversationId, {
-        type: "subagent:completed",
-        subagentId: conversationId,
-        conversationId
-      });
       let gathered = realResponseText(runResult?.response) || realResponseText(draft.assistantResponse);
       if (!gathered) {
         const freshSubConv = await this.conversationStore.get(conversationId);
@@ -13464,7 +13481,14 @@ ${resultBody}`,
           ...abnormal ? { error: { code: runError?.code ?? "SUBAGENT_INCOMPLETE", message: runError?.message ?? "subagent ended without a result" } } : {},
           timestamp: Date.now()
         };
-        await this.conversationStore.appendSubagentResult(parentConversationId, result);
+        await this.appendSubagentResultReliable(parentConversationId, result);
+      }
+      await this.eventSink(parentConversationId, {
+        type: "subagent:completed",
+        subagentId: conversationId,
+        conversationId
+      });
+      if (parentConv) {
         if (this.isServerless) {
           this.hooks.dispatchBackground("subagent-callback", parentConversationId);
         } else {
@@ -13492,11 +13516,6 @@ ${resultBody}`,
         conv.updatedAt = Date.now();
         await this.conversationStore.update(conv);
       }
-      await this.eventSink(conversation.parentConversationId, {
-        type: "subagent:completed",
-        subagentId: conversationId,
-        conversationId
-      });
       const parentConv = await this.conversationStore.get(conversation.parentConversationId);
       if (parentConv) {
         const result = {
@@ -13506,12 +13525,20 @@ ${resultBody}`,
           error: { code: "CONTINUATION_ERROR", message: err instanceof Error ? err.message : String(err) },
           timestamp: Date.now()
         };
-        await this.conversationStore.appendSubagentResult(conversation.parentConversationId, result);
+        await this.appendSubagentResultReliable(conversation.parentConversationId, result);
+      }
+      await this.eventSink(conversation.parentConversationId, {
+        type: "subagent:completed",
+        subagentId: conversationId,
+        conversationId
+      });
+      if (parentConv) {
         if (this.isServerless) {
           this.hooks.dispatchBackground("subagent-callback", conversation.parentConversationId);
         } else {
-          this.processSubagentCallback(conversation.parentConversationId).catch(() => {
-          });
+          this.processSubagentCallback(conversation.parentConversationId).catch(
+            (err2) => console.error(`[poncho][subagent] Continuation-error callback failed:`, err2 instanceof Error ? err2.message : err2)
+          );
         }
       }
     }
@@ -13555,7 +13582,7 @@ ${resultBody}`,
             opts.parentConversationId,
             opts.task,
             opts.ownerId
-          ).catch((err) => console.error(`[poncho][subagent] Background spawn failed:`, err instanceof Error ? err.message : err));
+          ).catch((err) => this.handleSpawnFailure(conversation.conversationId, opts.parentConversationId, opts.task, err));
         }
         return { subagentId: conversation.conversationId };
       },
@@ -13588,7 +13615,7 @@ ${resultBody}`,
             conversation.parentConversationId,
             message,
             conversation.ownerId
-          ).catch((err) => console.error(`[poncho][subagent] Background sendMessage failed:`, err instanceof Error ? err.message : err));
+          ).catch((err) => this.handleSpawnFailure(subagentId, conversation.parentConversationId, message, err));
         }
         return { subagentId };
       },
@@ -13667,6 +13694,67 @@ ${resultBody}`,
     };
   }
   // ── Stale subagent recovery ──
+  /**
+   * Append a subagent result to its parent, retrying once on a transient
+   * store failure before giving up loudly. A silently dropped result is the
+   * worst subagent failure mode — the parent waits forever on a subagent it
+   * thinks is still running — so this never swallows the error the way the
+   * old `.catch(() => {})` call sites did. Returns whether the result landed.
+   */
+  async appendSubagentResultReliable(parentConversationId, result) {
+    try {
+      await this.conversationStore.appendSubagentResult(parentConversationId, result);
+      return true;
+    } catch (firstErr) {
+      try {
+        await this.conversationStore.appendSubagentResult(parentConversationId, result);
+        return true;
+      } catch (secondErr) {
+        console.error(
+          `[poncho][subagent] FAILED to persist result for subagent ${result.subagentId} to parent ${parentConversationId} after 2 attempts \u2014 the parent will not see this result:`,
+          secondErr instanceof Error ? secondErr.message : secondErr,
+          `(first attempt: ${firstErr instanceof Error ? firstErr.message : firstErr})`
+        );
+        return false;
+      }
+    }
+  }
+  /**
+   * A subagent's fire-and-forget background run rejected outside its own
+   * error handling (e.g. it threw before entering its try block, or the
+   * catch block itself threw). Without this the parent is left waiting on a
+   * subagent that will never report back. Record the failure on the child
+   * and hand the parent an error result so the turn can resume.
+   */
+  async handleSpawnFailure(childConversationId, parentConversationId, task, err) {
+    const message = err instanceof Error ? err.message : String(err);
+    console.error(`[poncho][subagent] Background run failed for ${childConversationId}:`, message);
+    try {
+      const conv = await this.conversationStore.get(childConversationId);
+      if (conv?.subagentMeta && conv.subagentMeta.status === "running") {
+        conv.subagentMeta = {
+          ...conv.subagentMeta,
+          status: "error",
+          error: { code: "SUBAGENT_SPAWN_FAILED", message }
+        };
+        conv.updatedAt = Date.now();
+        await this.conversationStore.update(conv);
+      }
+    } catch {
+    }
+    const appended = await this.appendSubagentResultReliable(parentConversationId, {
+      subagentId: childConversationId,
+      task,
+      status: "error",
+      error: { code: "SUBAGENT_SPAWN_FAILED", message },
+      timestamp: Date.now()
+    });
+    if (appended) {
+      this.triggerParentCallback(parentConversationId).catch(
+        (e) => console.error(`[poncho][subagent] Parent callback failed after spawn failure:`, e instanceof Error ? e.message : e)
+      );
+    }
+  }
   async recoverStaleSubagents() {
     const allSummaries = await this.conversationStore.listSummaries();
     const subagentSummaries = allSummaries.filter((s) => s.parentConversationId);
@@ -13692,11 +13780,20 @@ ${resultBody}`,
             error: conv.subagentMeta.error,
             timestamp: Date.now()
           };
-          await this.conversationStore.appendSubagentResult(conv.parentConversationId, pendingResult);
+          await this.appendSubagentResultReliable(conv.parentConversationId, pendingResult);
           parentsToCallback.add(conv.parentConversationId);
         }
       }
     }
+    const parentIds = new Set(
+      subagentSummaries.map((s) => s.parentConversationId).filter((id) => !!id)
+    );
+    for (const parentId of parentIds) {
+      if (parentsToCallback.has(parentId)) continue;
+      if (this.activeConversationRuns.has(parentId)) continue;
+      const parent = await this.conversationStore.get(parentId);
+      if (parent?.pendingSubagentResults?.length) parentsToCallback.add(parentId);
+    }
     for (const parentId of parentsToCallback) {
       this.processSubagentCallback(parentId).catch(
         (err) => console.error(`[poncho][subagent] Recovery callback failed for ${parentId}:`, err instanceof Error ? err.message : err)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/harness",
-  "version": "0.52.0",
+  "version": "0.52.2",
   "description": "Agent execution runtime - conversation loop, tool dispatch, streaming",
   "repository": {
     "type": "git",

package/src/default-agent.ts

@@ -26,7 +26,12 @@ export interface DefaultAgentDefinitionOptions {
   modelProvider?: "anthropic" | "openai" | "openai-codex";
   /** Model name. Default: "claude-opus-4-5". */
   modelName?: string;
-  /** Sampling temperature. Default: 0.2. */
+  /**
+   * Sampling temperature. When unset, it is omitted from the generated
+   * frontmatter entirely and the harness sends no temperature (provider
+   * default). Newer models (Fable 5, Opus 4.7+) reject `temperature` — leave
+   * this unset for them.
+   */
   temperature?: number;
   /** Max tool-call steps per run. Default: 20. */
   maxSteps?: number;
@@ -55,7 +60,10 @@ export const defaultAgentDefinition = (
   const description = opts.description ?? DEFAULT_AGENT_DESCRIPTION;
   const modelProvider = opts.modelProvider ?? DEFAULT_MODEL_PROVIDER;
   const modelName = opts.modelName ?? DEFAULT_MODEL_NAME;
-  const temperature = opts.temperature ?? DEFAULT_TEMPERATURE;
+  // Opt-in: only emit a `temperature:` line when explicitly provided, so the
+  // harness sends no temperature otherwise (newer models reject it).
+  const temperatureLine =
+    opts.temperature !== undefined ? `\n  temperature: ${opts.temperature}` : "";
   const maxSteps = opts.maxSteps ?? DEFAULT_MAX_STEPS;
   const timeout = opts.timeout ?? DEFAULT_TIMEOUT;
 
@@ -65,8 +73,7 @@ id: ${id}
 description: ${description}
 model:
   provider: ${modelProvider}
-  name: ${modelName}
-  temperature: ${temperature}
+  name: ${modelName}${temperatureLine}
 limits:
   maxSteps: ${maxSteps}
   timeout: ${timeout}

package/src/harness.ts

@@ -2907,7 +2907,12 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
         convertedUpTo = messages.length;
         const coreMessages = cachedCoreMessages;
 
-        const temperature = agent.frontmatter.model?.temperature ?? 0.2;
+        // Only send temperature when the agent explicitly set one. Newer
+        // models (Fable 5, Opus 4.7+) removed sampling params entirely and
+        // return a 400 ("`temperature` is deprecated for this model") on any
+        // value — forcing a default here broke them. Treated like maxTokens
+        // below: omitted from the request when undefined.
+        const temperature = agent.frontmatter.model?.temperature;
         const maxTokens = agent.frontmatter.model?.maxTokens;
         // Place the tail breakpoint before any untruncated tool-result so
         // we cache only the stable prefix when prior-run tool results are
@@ -2971,7 +2976,7 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
           ...(useStaticCache ? {} : { system: systemPrompt }),
           messages: messagesForStep,
           tools: toolsForStep,
-          temperature,
+          ...(typeof temperature === "number" ? { temperature } : {}),
           abortSignal: input.abortSignal,
           ...(typeof maxTokens === "number" ? { maxTokens } : {}),
           experimental_telemetry: {

package/src/orchestrator/orchestrator.ts

@@ -1012,13 +1012,6 @@ export class AgentOrchestrator {
         await this.conversationStore.update(conv);
       }
 
-      this.hooks?.onStreamEnd?.(childConversationId);
-      await this.eventSink(parentConversationId, {
-        type: "subagent:completed",
-        subagentId: childConversationId,
-        conversationId: childConversationId,
-      });
-
       // Recover the subagent's real output: prefer the run response, then the
       // streamed draft, then walk the transcript — discarding the synthetic
       // "[Error: ...]" placeholder at each step.
@@ -1051,7 +1044,18 @@ export class AgentOrchestrator {
           : {}),
         timestamp: Date.now(),
       };
-      await this.conversationStore.appendSubagentResult(parentConversationId, pendingResult);
+      // Persist the result BEFORE emitting subagent:completed: a consumer
+      // reacting to the event (the parent callback, the streaming client)
+      // must find the result already durable in the store, not race its write.
+      await this.appendSubagentResultReliable(parentConversationId, pendingResult);
+
+      this.hooks?.onStreamEnd?.(childConversationId);
+      await this.eventSink(parentConversationId, {
+        type: "subagent:completed",
+        subagentId: childConversationId,
+        conversationId: childConversationId,
+      });
+
       this.triggerParentCallback(parentConversationId).catch(err =>
         console.error(`[poncho][subagent] Parent callback failed:`, err instanceof Error ? err.message : err),
       );
@@ -1070,6 +1074,16 @@ export class AgentOrchestrator {
         await this.conversationStore.update(conv);
       }
 
+      const pendingResult: PendingSubagentResult = {
+        subagentId: childConversationId,
+        task,
+        status: "error",
+        error: { code: "SUBAGENT_ERROR", message: errMsg },
+        timestamp: Date.now(),
+      };
+      // Persist before emitting (see the success path); never swallow.
+      await this.appendSubagentResultReliable(parentConversationId, pendingResult);
+
       this.hooks?.onStreamEnd?.(childConversationId);
       await this.eventSink(parentConversationId, {
         type: "subagent:error",
@@ -1078,14 +1092,6 @@ export class AgentOrchestrator {
         error: errMsg,
       });
 
-      const pendingResult: PendingSubagentResult = {
-        subagentId: childConversationId,
-        task,
-        status: "error",
-        error: { code: "SUBAGENT_ERROR", message: errMsg },
-        timestamp: Date.now(),
-      };
-      await this.conversationStore.appendSubagentResult(parentConversationId, pendingResult).catch(() => {});
       this.triggerParentCallback(parentConversationId).catch(err2 =>
         console.error(`[poncho][subagent] Parent callback failed:`, err2 instanceof Error ? err2.message : err2),
       );
@@ -1221,12 +1227,15 @@ export class AgentOrchestrator {
         },
         initialContextTokens: conversation.contextTokens ?? 0,
         initialContextWindow: conversation.contextWindow ?? 0,
-        onEvent: (event) => {
+        onEvent: async (event) => {
           if (event.type === "run:started") {
             const active = this.activeConversationRuns.get(conversationId);
             if (active) active.runId = event.runId;
           }
-          this.eventSink(conversationId, event);
+          // Await so the event is fully sunk before the next step's events,
+          // matching every other eventSink call site (the callback run path
+          // was the lone fire-and-forget exception).
+          await this.eventSink(conversationId, event);
         },
       });
       flushTurnDraft(execution.draft);
@@ -1436,11 +1445,6 @@ export class AgentOrchestrator {
       }
 
       this.activeSubagentRuns.delete(conversationId);
-      await this.eventSink(parentConversationId, {
-        type: "subagent:completed",
-        subagentId: conversationId,
-        conversationId,
-      });
 
       let gathered = realResponseText(runResult?.response) || realResponseText(draft.assistantResponse);
       if (!gathered) {
@@ -1464,8 +1468,17 @@ export class AgentOrchestrator {
             : {}),
           timestamp: Date.now(),
         };
-        await this.conversationStore.appendSubagentResult(parentConversationId, result);
+        // Persist before emitting completion (see runSubagent).
+        await this.appendSubagentResultReliable(parentConversationId, result);
+      }
 
+      await this.eventSink(parentConversationId, {
+        type: "subagent:completed",
+        subagentId: conversationId,
+        conversationId,
+      });
+
+      if (parentConv) {
         if (this.isServerless) {
           this.hooks!.dispatchBackground!("subagent-callback", parentConversationId);
         } else {
@@ -1490,12 +1503,6 @@ export class AgentOrchestrator {
         await this.conversationStore.update(conv);
       }
 
-      await this.eventSink(conversation.parentConversationId!, {
-        type: "subagent:completed",
-        subagentId: conversationId,
-        conversationId,
-      });
-
       const parentConv = await this.conversationStore.get(conversation.parentConversationId!);
       if (parentConv) {
         const result: PendingSubagentResult = {
@@ -1505,11 +1512,23 @@ export class AgentOrchestrator {
           error: { code: "CONTINUATION_ERROR", message: err instanceof Error ? err.message : String(err) },
           timestamp: Date.now(),
         };
-        await this.conversationStore.appendSubagentResult(conversation.parentConversationId!, result);
+        // Persist before emitting; never swallow (was `.catch(() => {})`).
+        await this.appendSubagentResultReliable(conversation.parentConversationId!, result);
+      }
+
+      await this.eventSink(conversation.parentConversationId!, {
+        type: "subagent:completed",
+        subagentId: conversationId,
+        conversationId,
+      });
+
+      if (parentConv) {
         if (this.isServerless) {
           this.hooks!.dispatchBackground!("subagent-callback", conversation.parentConversationId!);
         } else {
-          this.processSubagentCallback(conversation.parentConversationId!).catch(() => {});
+          this.processSubagentCallback(conversation.parentConversationId!).catch(err2 =>
+            console.error(`[poncho][subagent] Continuation-error callback failed:`, err2 instanceof Error ? err2.message : err2),
+          );
         }
       }
     }
@@ -1559,7 +1578,7 @@ export class AgentOrchestrator {
             opts.parentConversationId,
             opts.task,
             opts.ownerId,
-          ).catch(err => console.error(`[poncho][subagent] Background spawn failed:`, err instanceof Error ? err.message : err));
+          ).catch(err => this.handleSpawnFailure(conversation.conversationId, opts.parentConversationId, opts.task, err));
         }
 
         return { subagentId: conversation.conversationId };
@@ -1596,7 +1615,7 @@ export class AgentOrchestrator {
             conversation.parentConversationId,
             message,
             conversation.ownerId,
-          ).catch(err => console.error(`[poncho][subagent] Background sendMessage failed:`, err instanceof Error ? err.message : err));
+          ).catch(err => this.handleSpawnFailure(subagentId, conversation.parentConversationId!, message, err));
         }
 
         return { subagentId };
@@ -1684,6 +1703,79 @@ export class AgentOrchestrator {
 
   // ── Stale subagent recovery ──
 
+  /**
+   * Append a subagent result to its parent, retrying once on a transient
+   * store failure before giving up loudly. A silently dropped result is the
+   * worst subagent failure mode — the parent waits forever on a subagent it
+   * thinks is still running — so this never swallows the error the way the
+   * old `.catch(() => {})` call sites did. Returns whether the result landed.
+   */
+  private async appendSubagentResultReliable(
+    parentConversationId: string,
+    result: PendingSubagentResult,
+  ): Promise<boolean> {
+    try {
+      await this.conversationStore.appendSubagentResult(parentConversationId, result);
+      return true;
+    } catch (firstErr) {
+      try {
+        await this.conversationStore.appendSubagentResult(parentConversationId, result);
+        return true;
+      } catch (secondErr) {
+        console.error(
+          `[poncho][subagent] FAILED to persist result for subagent ${result.subagentId} ` +
+            `to parent ${parentConversationId} after 2 attempts — the parent will not see this result:`,
+          secondErr instanceof Error ? secondErr.message : secondErr,
+          `(first attempt: ${firstErr instanceof Error ? firstErr.message : firstErr})`,
+        );
+        return false;
+      }
+    }
+  }
+
+  /**
+   * A subagent's fire-and-forget background run rejected outside its own
+   * error handling (e.g. it threw before entering its try block, or the
+   * catch block itself threw). Without this the parent is left waiting on a
+   * subagent that will never report back. Record the failure on the child
+   * and hand the parent an error result so the turn can resume.
+   */
+  private async handleSpawnFailure(
+    childConversationId: string,
+    parentConversationId: string,
+    task: string,
+    err: unknown,
+  ): Promise<void> {
+    const message = err instanceof Error ? err.message : String(err);
+    console.error(`[poncho][subagent] Background run failed for ${childConversationId}:`, message);
+    try {
+      const conv = await this.conversationStore.get(childConversationId);
+      if (conv?.subagentMeta && conv.subagentMeta.status === "running") {
+        conv.subagentMeta = {
+          ...conv.subagentMeta,
+          status: "error",
+          error: { code: "SUBAGENT_SPAWN_FAILED", message },
+        };
+        conv.updatedAt = Date.now();
+        await this.conversationStore.update(conv);
+      }
+    } catch {
+      // best-effort: the result append below is what the parent actually needs
+    }
+    const appended = await this.appendSubagentResultReliable(parentConversationId, {
+      subagentId: childConversationId,
+      task,
+      status: "error",
+      error: { code: "SUBAGENT_SPAWN_FAILED", message },
+      timestamp: Date.now(),
+    });
+    if (appended) {
+      this.triggerParentCallback(parentConversationId).catch(e =>
+        console.error(`[poncho][subagent] Parent callback failed after spawn failure:`, e instanceof Error ? e.message : e),
+      );
+    }
+  }
+
   async recoverStaleSubagents(): Promise<void> {
     const allSummaries = await this.conversationStore.listSummaries();
     const subagentSummaries = allSummaries.filter((s) => s.parentConversationId);
@@ -1711,11 +1803,26 @@ export class AgentOrchestrator {
             error: conv.subagentMeta.error,
             timestamp: Date.now(),
           };
-          await this.conversationStore.appendSubagentResult(conv.parentConversationId, pendingResult);
+          await this.appendSubagentResultReliable(conv.parentConversationId, pendingResult);
           parentsToCallback.add(conv.parentConversationId);
         }
       }
     }
+
+    // Also drain parents that already have results sitting in the store but
+    // no active run to deliver them — e.g. a result persisted just before a
+    // process restart, whose in-memory callback trigger was lost. Without
+    // this the parent stays stuck even though its result landed durably.
+    const parentIds = new Set(
+      subagentSummaries.map(s => s.parentConversationId).filter((id): id is string => !!id),
+    );
+    for (const parentId of parentIds) {
+      if (parentsToCallback.has(parentId)) continue;
+      if (this.activeConversationRuns.has(parentId)) continue;
+      const parent = await this.conversationStore.get(parentId);
+      if (parent?.pendingSubagentResults?.length) parentsToCallback.add(parentId);
+    }
+
     for (const parentId of parentsToCallback) {
       this.processSubagentCallback(parentId).catch(err =>
         console.error(`[poncho][subagent] Recovery callback failed for ${parentId}:`, err instanceof Error ? err.message : err),

package/src/storage/postgres-engine.ts

@@ -36,12 +36,25 @@ export class PostgresEngine extends SqlStorageEngine {
         return rows as T[];
       },
       exec: async (sql: string): Promise<void> => {
-        await this.sql.unsafe(sql);
+        // DDL is idempotent in our migrations (`CREATE TABLE IF NOT
+        // EXISTS`, etc.), so retrying on a stale-socket drop is
+        // safe — same idempotency as `query()` reads/writes.
+        await this.runWithRetry(() => this.sql.unsafe(sql));
       },
       transaction: async (fn: () => Promise<void>): Promise<void> => {
-        await this.sql.begin(async () => {
+        // Transactions are inherently retry-safe at the
+        // CONNECTION_ENDED boundary: if the connection dies before
+        // BEGIN takes effect server-side, no work was committed and
+        // re-running `fn` produces the correct end state. The retry
+        // only catches the connection-level reject from the
+        // postgres.js client; a partial-commit + drop scenario
+        // surfaces as a different error code and bypasses the
+        // retry, preserving the caller's expectation that a
+        // returned transaction either fully committed or fully
+        // rolled back.
+        await this.runWithRetry(() => this.sql.begin(async () => {
           await fn();
-        });
+        }));
       },
     };
   }
@@ -59,25 +72,34 @@ export class PostgresEngine extends SqlStorageEngine {
       prepare: false,
       // Connection-pool resilience. Managed Postgres providers
       // (Railway, Neon, Heroku, etc.) routinely drop idle TCP
-      // connections server-side after a few minutes. Without these
-      // knobs, porsager/postgres keeps stale sockets in the pool;
-      // the next query on one rejects with
-      // `write CONNECTION_ENDED <host>:5432` at `durMs=0`, surfacing
-      // as a hard failure to the caller. Two complementary settings:
+      // connections server-side after a few minutes — and on
+      // Railway in particular, mid-stream drops within a few
+      // seconds of inactivity are common. Without these knobs,
+      // porsager/postgres keeps stale sockets in the pool; the
+      // next query on one rejects with
+      // `write CONNECTION_ENDED <host>:5432` at `durMs=0`,
+      // surfacing as a hard failure to the caller.
       //
-      //   - `idle_timeout: 20` closes idle connections client-side
-      //     after 20s, before any reasonable provider-side timer
-      //     fires. Fresh connection on next checkout = no stale
-      //     socket race.
-      //   - `max_lifetime: 600` (10 min) recycles long-lived
-      //     connections defensively even if they've stayed busy,
-      //     which sidesteps a separate class of provider-side
-      //     "max connection age" limits.
+      //   - `idle_timeout: 5` closes idle connections client-side
+      //     aggressively. Empirically Railway's pg drops sockets
+      //     well before the 20s value that managed-provider docs
+      //     suggest; 5s is short enough to win the race in
+      //     practice while staying long enough that bursty
+      //     workloads still get connection reuse.
+      //   - `max_lifetime: 300` (5 min) recycles long-lived
+      //     connections defensively. Even with idle_timeout, a
+      //     connection that's been actively serving small queries
+      //     for an hour can hit provider-side max-age limits.
+      //   - `connect_timeout: 10` — slightly less patient on
+      //     initial connect than the 30s default. Combined with
+      //     the retry below, "connection refused" surfaces faster
+      //     during incidents and the caller can shed load instead
+      //     of stacking up.
       //
-      // Defaults remain `max: 10`, `connect_timeout: 30` — leaving
-      // pool size + initial connect behavior unchanged.
-      idle_timeout: 20,
-      max_lifetime: 60 * 10,
+      // Pool size (`max: 10`) unchanged.
+      idle_timeout: 5,
+      max_lifetime: 60 * 5,
+      connect_timeout: 10,
     });
   }
 
@@ -147,33 +169,53 @@ export class PostgresEngine extends SqlStorageEngine {
   }
 
   /**
-   * Single retry on a transient connection-layer failure. The
-   * `idle_timeout` / `max_lifetime` config above prevents *most*
-   * stale-connection cases, but a query can still race a
-   * provider-initiated drop in flight — the postgres.js client
-   * rejects with `code: "CONNECTION_ENDED"` and the next attempt
-   * checks out a fresh connection from the pool. One retry is
-   * enough; if it fails again the host-side network is genuinely
-   * broken and the caller should see the error.
+   * Retry on transient connection-layer failures. Three attempts
+   * with exponential-ish backoff (0, 50ms, 200ms) — the pool may
+   * have multiple stale sockets accumulated during an idle period
+   * (especially on managed Postgres after boot when no traffic
+   * has flowed for a while), so a single retry can land on a
+   * second stale socket and still fail. Three attempts virtually
+   * always exhausts the staleness wave; if all three throw, the
+   * failure is real and the caller should see it.
    *
-   * Only retries reads + the standard exec/run paths in `query`;
-   * `sql.unsafe(sql)` calls in `executeRaw` (migration DDL) and
-   * `sql.begin(...)` transactions are unwrapped — those are
-   * idempotent-by-construction (DDL is `IF NOT EXISTS`) or
-   * atomically scoped (transactions roll back cleanly), and adding
-   * a retry around them would complicate the transaction
-   * semantics.
+   * Applied to every pg path the executor exposes:
+   *  - `query()` (run/get/all)  — natural retry: queries are
+   *    idempotent at the connection-failure boundary because the
+   *    server-side rollback runs cleanly on socket close.
+   *  - `exec(sql)` for DDL      — `CREATE TABLE IF NOT EXISTS` and
+   *    friends are idempotent by construction.
+   *  - `transaction(fn)`        — only retried when the
+   *    CONNECTION_ENDED reject arrives *before* the transaction
+   *    body started executing on the connection; if it errors
+   *    mid-transaction, the postgres.js client surfaces a
+   *    different error class (the inner SQL error) and bypasses
+   *    this retry, preserving the all-or-nothing semantics.
    */
   private async runWithRetry<T>(fn: () => Promise<T>): Promise<T> {
-    try {
-      return await fn();
-    } catch (err) {
-      const code = (err as { code?: string } | null | undefined)?.code;
-      if (code === "CONNECTION_ENDED" || code === "CONNECTION_CLOSED" || code === "CONNECTION_DESTROYED") {
+    const backoffs = [0, 50, 200];
+    let lastErr: unknown;
+    for (let attempt = 0; attempt < backoffs.length; attempt++) {
+      if (backoffs[attempt] > 0) {
+        await new Promise((r) => setTimeout(r, backoffs[attempt]));
+      }
+      try {
         return await fn();
+      } catch (err) {
+        lastErr = err;
+        const code = (err as { code?: string } | null | undefined)?.code;
+        if (
+          code === "CONNECTION_ENDED" ||
+          code === "CONNECTION_CLOSED" ||
+          code === "CONNECTION_DESTROYED" ||
+          code === "CONNECT_TIMEOUT" ||
+          code === "ECONNRESET"
+        ) {
+          continue;
+        }
+        throw err;
       }
-      throw err;
     }
+    throw lastErr;
   }
 
   private addToPathCache(tenantId: string, path: string): void {