[Bug]: Completed subagent results remain unreconciled and Orchestrator stays idle until user sends “Wake up”

[ProphetOfDoom-PoD]

What happened, and what did you expect?

When the Orchestrator delegates work to specialist agents, the specialists can complete successfully but the parent Orchestrator does not resume, reconcile their results, or provide any update.

The session appears completely inactive. It remains idle indefinitely until I send another user message such as “Wake up”.

This happens across multiple OpenCode Desktop tabs and sessions.

I reproduced it in a session without restarting OpenCode:

1. The Orchestrator dispatched Librarian and Oracle tasks.
2. Both tasks completed.
3. Oracle remained marked completed/unreconciled.
4. The parent Orchestrator produced no response and did not continue.
5. Sending “Wake up” caused the parent to resume and process the results.

I also observed a related case after restarting OpenCode. A child Fixer session was aborted during restart, but the parent retained the stale belief that the child was still running. backgroundJobs.continueOnIdle sent its one automatic continuation, but the parent stopped again while waiting for a child that no longer existed.

This does not appear to be an empty model response or provider failure. Completed assistant responses had finish: "stop" and non-zero output tokens.

Expected behavior:

• Completing a child task should reliably wake the parent.
• The parent should reconcile every terminal result.
• If a child fails or is aborted, the parent should be informed and recover or report the failure.
• The Orchestrator should continue incomplete todos or clearly explain why it cannot.
• No manual “Wake up” message should be required.

Steps to reproduce

Environment:

• Windows
• OpenCode Desktop 1.18.5
• oh-my-opencode-slim 2.2.8
• Agent: Orchestrator
• Active preset: openai
• backgroundJobs.continueOnIdle: true

Clean-cache preparation:

1. Fully close OpenCode.
2. Delete ~/.cache/opencode.
3. Restart OpenCode.

Reproduction:

1. Open a fresh Orchestrator session.

2. Give it a task that delegates two independent investigations, for example:

   “Ask Librarian and Oracle in parallel to investigate this problem, then reconcile and summarize both results.”

3. Wait for both child agents to complete.

4. Observe that the parent sometimes becomes idle without reconciling or responding.

5. Check the background-job board: one or more children may remain completed/unreconciled.

6. Wait several minutes. No automatic continuation occurs.

7. Send “Wake up”.

8. Observe that the parent immediately resumes and processes the results.

The issue also occurs during larger multi-step tasks with incomplete todos.

Relevant sessions from one occurrence:

• Parent: ses_0623148c3ffes3JcB2TTm4WldA
• Aborted child after restart: ses_05d70e336ffe7ylUMs4E8gREYL
• Completed/unreconciled Oracle from a no-restart occurrence: ses_05d3d39c8ffeyEReiz3Q5Q3XaP

oh-my-opencode.json

{
  "$schema": "https://unpkg.com/oh-my-opencode-slim@latest/oh-my-opencode-slim.schema.json",
  "preset": "openai",
  "presets": {
    "openai": {
      "orchestrator": {
        "model": "openai/gpt-5.6-terra",
        "variant": "xhigh",
        "skills": ["*"],
        "mcps": ["*", "!context7"]
      },
      "oracle": {
        "model": "openai/gpt-5.6-sol",
        "variant": "xhigh",
        "skills": ["simplify"],
        "mcps": []
      },
      "librarian": {
        "model": "openai/gpt-5.6-luna",
        "variant": "low",
        "skills": [],
        "mcps": ["websearch", "context7", "gh_grep"]
      },
      "explorer": {
        "model": "openai/gpt-5.6-luna",
        "variant": "low",
        "skills": [],
        "mcps": []
      },
      "designer": {
        "model": "openai/gpt-5.6-luna",
        "variant": "medium",
        "skills": [],
        "mcps": []
      },
      "fixer": {
        "model": "openai/gpt-5.6-luna",
        "variant": "xhigh",
        "skills": [],
        "mcps": []
      },
      "observer": {
        "model": "openai/gpt-5.6-luna",
        "variant": "low",
        "skills": [],
        "mcps": []
      }
    }
  },
  "disabled_agents": [],
  "image_routing": "auto",
  "companion": {
    "enabled": false,
    "position": "bottom-left",
    "size": "medium",
    "gifPack": "default",
    "loopStyle": "classic",
    "speed": 1,
    "debug": false
  },
  "fallback": {
    "enabled": true,
    "maxRetries": 1,
    "retry_on_empty": true
  },
  "multiplexer": {
    "type": "zellij",
    "zellij_pane_mode": "agent-tab"
  },
  "backgroundJobs": {
    "continueOnIdle": true
  }
}

OpenCode version

OpenCode Desktop 1.18.5

oh-my-opencode-slim version

oh-my-opencode-slim 2.2.8

Operating system

Windows 11

Logs, screenshots, or extra context

[mhenke]

This was generated by AI during triage.

Cross-referencing from #917 and #924, which describe the opposite symptom of the same reconciliation failure family: instead of the orchestrator staying idle after subagent completion (your report), those issues show it looping repeatedly when unreconciled jobs exist on the board.

Your report is useful because it pins down a second failure mode. The idle reconciliation path (idle-reconciliation.ts:32) checks hasInputWait and skips scheduling when wait_for_user is active. But when continueOnIdle is true and no wait_for_user is called, the evaluator should fire on idle. Your case suggests that path sometimes doesn't fire either.

Three PRs are in flight that touch the reconciliation and foreground-fallback logic: #925, #921, #913. Worth verifying your repro against those once they land, particularly #925 which changes when terminal jobs get reconciled.

[brucemead]

Hi @mhenke,

Cross-checked against #917, #924, and the open fixes in #913, #921, and #925. They cover the same reconciliation/foreground-fallback area, and nothing here conflicts with those changes. No additional fix is required from this feedback; the repo should be retested after #925 lands.

Thanks,
Bruce