[Bug] Orphaned background tasks block slot allocation after session reset

[randomm]

Problem

After a session reset/resume, background task slots remain occupied by stale/orphaned task records, preventing new tasks from being spawned even though fewer tasks are actually running.

Reproduction

1. Have tasks running when session resets/crashes
2. Resume or restart the same session
3. Spawn new tasks → hits "exceeded concurrent task limit (5)" prematurely
4. Check running tasks → fewer than 5 are actually running

Key Detail

Session restart/resume does NOT fix the issue. The orphaned slot count persists across session restarts.

Only workaround: Start a completely NEW session (losing all session context/history).

Expected Behavior

• Task slots should be released when parent session resets
• Or: On session resume, detect and clean up orphaned task records
• Or: Task completion events should properly free slots regardless of session state

Observed Behavior

• Task slots remain occupied by ghost/orphaned tasks
• check_task on old task IDs returns completed or not_found, but slots aren't freed
• Problem persists across session restart/resume
• User must abandon session entirely to recover full slot capacity

Impact

• High - Forces users to lose session context to recover task capacity
• Blocks parallel work orchestration

Potential Fix Areas

• packages/opencode/src/tool/task.ts - task tracking/counting logic
• Session initialization - cleanup of stale task records on resume
• Background task state persistence - ensure slot counts are accurate
• Consider: slots tied to session ID vs global count

[randomm]

Root Cause Analysis Complete

Architecture

Two files manage task slots:

session/index.ts:

• reservedTaskSlots Map (line 92) - tracks reserved slots per session
• pendingTaskMetadata Map (line 90) - tracks running task metadata
• getSessionTaskCount() (line 109) - counts from both maps

tool/task.ts:

• releaseCallbacks Map (line 32) - stores slot release functions
• tryIncrementSessionCount() - reserves slot, stores release callback
• decrementSessionCount() - calls release callback

The Bug

When a task completes normally:

1. task.ts line 326 calls decrementSessionCount()
2. This calls the stored release callback
3. Slot is properly freed ✅

When a task times out or crashes:

1. session/index.ts trackBackgroundTask catches the error
2. Finally block (line 284-287) cleans pendingBackgroundTasks and pendingTaskMetadata
3. BUT reservedTaskSlots is NOT cleaned ❌
4. The releaseCallbacks in task.ts are never called
5. Slot remains permanently occupied

Why Session Restart Doesn't Help

• These are in-memory Maps in the Node process
• Session restart doesn't restart the Node process
• Maps persist with stale data

Confirmed Impact

• 5 tasks spawned, 4 timed out
• Only 1 slot available despite 0 tasks actually running
• Completely blocks parallel orchestration

[randomm]

Fixed in PR #49