TaskExecute spawn fails silently due to RPC envelope mismatch + ping race

[heyhuynhgiabuu]

Summary

TaskExecute fails with spawn failed — (empty error message) when RPC reply handlers return bare payloads instead of the expected RpcReply envelope. Additionally, checkSubagentsVersion() incorrectly marks subagents as "outdated" when it receives malformed {} ping replies from stale extension mirrors or legacy handlers.

Root Causes

1. rpcCall() only accepts envelope replies

rpcCall() casts all replies to RpcReply<T> and checks reply.success. If a handler returns a bare payload like { id: "abc" } (no envelope), reply.success is undefined (falsy), so it calls reject(new Error(reply.error)) — producing an empty error message and silent spawn failure.

2. checkSubagentsVersion() fails on first malformed reply

When multiple extensions respond to a ping (e.g., a stale mirror copy replies {} before the real handler replies with { success: true, data: { version: 2 } }), the function unsubscribes on the first reply. Since {} has no data.version, it immediately sets pendingWarning and never receives the valid reply.

3. No debounce on version checks

Each subagents:ready event triggers a new checkSubagentsVersion() call. If multiple ready events fire in rapid succession (e.g., during extension initialization), multiple concurrent pings race and can produce inconsistent state.

Impact

• TaskExecute returns spawn failed — for every task
• subagentsAvailable stays false even when subagents are loaded and functional
• The Agent tool works fine (bypasses RPC), so users see inconsistent behavior

Environment

• pi-tasks v0.4.2
• pi-subagents v0.5.1 (with custom bridge extensions)

Proposed Fix

See PR — fixes all three issues:

1. Make rpcCall() accept both envelope and bare payload replies
2. Make checkSubagentsVersion() ignore malformed replies and keep listening
3. Add versionCheckInFlight debounce guard
4. Also includes cascade data flow (inject prerequisite results) and model forwarding as improvements

[heyhuynhgiabuu]

can you take a look please? @tintinweb

[tintinweb]

the two versions you listed should be compatible. if they are not, you should see a warning. looks to me as if you're running two incompatible versions and I would really like to avoid having to add legacy fallback code.

[heyhuynhgiabuu]

Thanks for looking into this and for the screenshots — you're right.

After further investigation, the RPC failures in our environment were caused by our specific setup, not a bug in pi-tasks:

1. We had two competing pi-subagents packages installed side-by-side (@tintinweb/pi-subagents@0.5.1 + pi-subagents@0.11.2)
2. A custom bridge extension (and a stale mirror copy of it) was responding to ping events with {} before the native handler could reply with the proper envelope
3. This race condition made checkSubagentsVersion() see the malformed reply first and mark subagents as outdated

In a clean environment with just your two packages, everything works as expected — confirmed on your screenshots.

The RPC backward compatibility / legacy fallback part of the PR can be dropped. That said, the PR also includes a few improvements that are independently useful regardless of environment:

• Cascade data flow: buildTaskPrompt() injects completed dependency results into cascaded agent prompts (so downstream agents don't need to call TaskGet)
• Model forwarding: passes model from TaskExecute params to spawn + cascade calls
• Debounce guard: prevents concurrent checkSubagentsVersion() calls from stacking when multiple subagents:ready events fire

I'll update the PR to remove the legacy fallback code and keep only the feature additions. Apologies for the noise on the bug report — the root cause was entirely on our side.