v0.1.1

Background-agent visibility & lifecycle improvements. Async delegation, steering, and the supervisor control plane are unchanged.

Changes

• Child-session cleanup. Once a delegation is terminal and its result has been read, its child session is deleted, so the TUI's child-session navigation (Ctrl+X ↓, then ←/→) cycles through live and not-yet-read operators instead of every finished one. The persisted .md artifact remains, so delegation_read still works after deletion. Opt out with BACKGROUND_AGENTS_KEEP_CHILD_SESSIONS=1.
• Dispatch toast. A toast appears the moment a background operator is launched, so a dispatch no longer looks like nothing happened while the supervisor goes idle.
• Identifiable child sessions. Each delegation's child session is titled <agent> · <id>, so operators are recognizable in the session list and the subagent footer. Navigate with Ctrl+X ↓ (open), ←/→ (switch), ↑ (return).
• Stop/timeout/delete paths mark their sessions as cleaned to avoid a redundant delete; tool-part titles are set via ToolResult.

Full Changelog: v0.1.0...v0.1.1

v0.1.0

Async, supervised background agent delegation for OpenCode — delegate, steer, stop, and persist sub-agent work.

Features

• Async delegation: delegate launches a sub-agent in the background and returns an ID immediately; the supervisor stays free to keep working.
• Interactive control: delegation_status, delegation_peek, delegation_steer, and delegation_stop to observe and course-correct running tasks without restarting them.
• Persistent output: every delegation result is written to disk and retrievable via delegation_read, surviving compaction.
• Strict / relaxed routing: route read-only vs. write-capable sub-agents through delegate or the native task tool (BACKGROUND_AGENTS_STRICT_READONLY).
• State persistence & crash recovery: active delegations are mirrored to disk and re-adopted after a process restart.
• Per-delegation model override: pick the model per task (provider/model-id), or fall back to the agent's configured model.

Fixes

• Server-side turn errors now surface as error: a turn that fails on the server (e.g. a bad model override → ProviderModelNotFoundError) resolves the prompt with the failure on the assistant message instead of rejecting. These runs are now finalized as error with the real message, rather than being mislabeled "complete with no output".
• Wake-on-completion: a delegation that completes while siblings are still running now wakes the supervisor immediately, so an idle supervisor no longer stays dormant until the whole batch settles. The last completion stays silent and the single all-complete notification delivers the final wake (no double-wake).

Verification

• TypeScript typecheck clean.
• 46 tests passing across 5 files.