Sub-agent Task silently returns empty result when the final response stream fails (e.g. 503) — no error, no retry, completed work lost

[octal-illumination]

Description

When a sub-agent (Task) finishes all its tool work but its final response stream then
fails — e.g. server [503] The request queue is full — the parent Task call returns an
empty result. No error is surfaced and no retry happens; the sub-agent's completed work
is silently discarded.

Reproduced on two explore sub-agents in one parent session (opencode 1.18.13). Both
completed 64 and 52 tool calls (data mining against a 151 MB DB), then hit [503] on
their last message (msg_fce23bb9..., msg_fce2648...). Both returned an empty
<task_result> to the parent. The work was only recoverable by re-extracting the
sub-agents' persisted tool outputs directly from opencode.db.

This is the known 503/free-tier overload family (#40254, #38051, #39276). The client-side
part I'm reporting is that a Task whose final stream dies should surface the stream error
to the parent (and/or retry), not return a success-looking empty result.

Plugins

@cortexkit/opencode-magic-context (context management); GitHub MCP; atomic-updater MCP

OpenCode version

1.18.13

Steps to reproduce

1. Start a Task (subagent) with a long multi-tool mining workload (many sqlite3/bash calls).
2. Hit a provider overload window so the final response stream fails with [503] ....
3. Observe: sub-agent completes all tool calls, final stream errors, sub-agent result is empty.
4. Parent receives empty result; no error shown; no retry; the mined data appears lost.

Screenshot and/or share link

No response

Operating System

Linux (Qubes OS 4.2.4) - Under Fedora (FC37) Standalone VM

Terminal

OpenCode TUI (CLI) - under Konsole

[github-actions]

This issue might be a duplicate of (or closely related to) existing issues. Please check:

• [BUG] Subagent terminates prematurely mid-execution but reports completed #32132: Subagent terminates prematurely mid-execution but reports completed — same symptom of no error surfaced and success-looking result when subagent work is cut short
• 502 provider_unavailable errors from OpenRouter are not retried, causing subagent/session aborts #22448: 502 provider_unavailable errors from OpenRouter are not retried, causing subagent/session aborts — covers the retry gap for 502/503 errors that leads to this exact lost-work scenario

Your issue adds a specific and well-documented angle: the final-stream failure path returns an empty success result rather than propagating the error to the parent. That distinction may warrant keeping this open alongside the others, but reviewing those issues first may be useful.