@ferris1225/pi-subagents 2.0.1 → 2.0.2

package/README.md

@@ -88,9 +88,10 @@ index.
 - **Direct fallback with real thinking capabilities** — each agent has at most
   one selected model. An unavailable selection, rate limit, invalid key, quota,
   missing model, or provider failure hands directly to the current main model.
-  A child-only provider adapter forces request retries to zero, and the RPC
-  parent cancels Pi's outer turn retry before another call, without changing user
-  settings. Auto thinking clamps the agent preference to the
+  A child-only provider adapter forces inner request retries to zero; transient
+  stream drops still use Pi's outer turn retry, and only a settled model-level
+  failure hands off, without changing user settings. Auto thinking clamps the
+  agent preference to the
   effective model's real `thinkingLevelMap`; manual setup shows only levels that
   model supports.
 - **Resumes, retargets, and forks preserve context** — every run is session-backed.
@@ -331,12 +332,13 @@ available catalog is skipped. Any model-level runtime failure — rate limit,
 quota, invalid key/auth, missing model, provider error, or idle model stream —
 hands directly to current main, including stream errors that retain partial text.
 A child-only Pi extension wraps the selected provider's registered API stream
-with `maxRetries: 0`; if Pi schedules its separate outer turn retry, the RPC parent
-immediately sends `abort_retry` before another provider call. This uses supported
-extension/RPC surfaces in Node and standalone/Bun builds, never rewrites global or
-project settings, and does not alter descendant tool environments. Tool/test
-failures stay on the same model because they are task failures, not model
-availability failures. Only a truly
+with `maxRetries: 0` so a deterministic auth/quota miss fails fast. Transient
+stream drops such as xAI `terminated` still use Pi's outer turn retry — the
+parent does not `abort_retry` them — and only a settled model-level failure
+hands off to current main. This uses supported extension/RPC surfaces in Node
+and standalone/Bun builds, never rewrites global or project settings, and does
+not alter descendant tool environments. Tool/test failures stay on the same
+model because they are task failures, not model availability failures. Only a truly
 zero-activity process startup race can retry; an accepted prompt or any
 agent/turn/stream/tool activity forbids replay.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ferris1225/pi-subagents",
-  "version": "2.0.1",
+  "version": "2.0.2",
   "description": "Controllable background sub-agent threads for pi: specialized roles, capability-aware thinking, direct main-model fallback, auto-fix chains, and Git worktree isolation.",
   "type": "module",
   "license": "MIT",

package/src/dispatch.ts

@@ -1483,7 +1483,8 @@ export function registerSubagentTool(pi: ExtensionAPI, runtime: SubagentRuntime)
 								triggerTurn: completionTriggersTurn(result, runConfig.notifyOnReviewPass),
 							};
 							if (modelLevel) {
-								runCtx.ui.notify(`✗ ${result.agent} dispatch failed: model unavailable or broken — task handed to the main window`, "error");
+								const detail = result.errorMessage?.trim() || "model unavailable or broken";
+								runCtx.ui.notify(`✗ ${result.agent} dispatch failed: ${detail} — task handed to the main window`, "error");
 							} else if (dispatchFailed) {
 								runCtx.ui.notify(`✗ ${result.agent} dispatch failed: ${result.errorMessage ?? "dispatch crashed"}`, "error");
 							}

package/src/format.ts

@@ -155,11 +155,15 @@ export function formatCompletionBlock(
  * fresh (which would re-scan everything). */
 export function modelLevelTakeoverNote(result: SingleResult, opts?: { runId?: number }): string {
 	const retry = result.modelFallbackFrom ? ", and the current main model also failed" : "";
+	const detail = result.errorMessage?.trim();
+	const cause = detail
+		? `its model/provider call failed (${detail})`
+		: "its model was unavailable or failed (or the run stalled)";
 	const sessionPreserved = Boolean(result.sessionDir && result.sessionId) && opts?.runId !== undefined;
 	const recovery = sessionPreserved
 		? ` The sub-agent's earlier work in this run is preserved. Once a model is available again, call subagent_control with { action: "resume", id: ${opts!.runId} } to CONTINUE it in-context (it keeps the same run id and does not re-scan), or execute the task in the main window with your own tools.`
 		: ` Please execute this task in the main window with your own tools; do not re-dispatch it as a sub-agent.`;
-	return `The sub-agent could not complete this task: its model was unavailable or failed (or the run stalled)${retry}.${recovery}`;
+	return `The sub-agent could not complete this task: ${cause}${retry}.${recovery}`;
 }
 
 /** Resolve a run-id request to actual ids: an exact numeric match always wins

package/src/rpc-run.ts

@@ -813,12 +813,11 @@ export async function runRpcAgentAttempt(options: RunRpcAttemptOptions): Promise
 			result.rpcActivity = true;
 		}
 
-		// The child provider adapter disables request-level retries. Cancel Pi's
-		// separate outer turn retry the instant it is scheduled, before another
-		// same-model provider call can begin.
-		if (event.type === "auto_retry_start") {
-			void send({ type: "abort_retry" }).catch(() => undefined);
-		}
+		// Let Pi's outer turn retry run. Grok/xAI long streams commonly drop with
+		// a retryable `terminated` mid-turn; aborting that retry was misread as
+		// "model unavailable" and handed a still-working model back to the parent.
+		// After retries exhaust, dispatch still classifies a settled model-level
+		// failure and hands off.
 
 		// Child RPC mode exposes extension dialogs. Sub-agents are non-interactive:
 		// cancel blocking dialogs so an unrelated child extension cannot deadlock.