opencode-task-result-guard

OpenCode server plugin that repairs missing or incomplete foreground Task handoffs before the parent agent sees them.

OpenCode can currently mark a child Task complete while returning an empty <task_result> when the child's terminal assistant message has no text. A separate race can let an auto-fallback plugin continue the child after the parent already received that empty result. Long-running critic agents can also compact and return a progress note instead of their required final report.

The guard runs at tool.execute.after and leaves normal Task output unchanged.

Behavior

• Reuses the newest non-synthetic child text when the terminal Task result is empty.
• Waits a bounded interval for a late report after ContentFilterError.
• Returns a visible <task_error> with the same task_id when no text can be recovered.
• Optionally recognizes incomplete agents by name prefix and gives them one synthesis-only finalizer turn.
• Injects a hidden finalizer agent with all tools denied.
• Bounds repaired output while preserving its beginning and tail.

Install

opencode plugin opencode-task-result-guard --global

Or add it to global opencode.json:

{
  "plugin": ["opencode-task-result-guard"]
}

Restart OpenCode after installation.

Requires an OpenCode build with modern server plugin entrypoints (exports["./server"]). Tested against 0.0.0-beta-202607101924.

Options

Plugin options use an OpenCode plugin tuple:

{
  "plugin": [
    [
      "opencode-task-result-guard",
      {
        "lateResultWaitMs": 180000,
        "finalizeTimeoutMs": 180000,
        "pollMs": 1000,
        "maxBytes": 49152,
        "finalizerAgent": "task-result-guard-finalizer",
        "finalizerModel": "openai/gpt-5.4-mini",
        "finalizerReasoningEffort": "high",
        "finalizeAgentPrefixes": ["eye-"],
        "finalMarker": "## Verdict"
      }
    ]
  ]
}

finalizerModel and finalizerReasoningEffort are optional. Without a model, the finalizer inherits the child session's model. Set finalizeAgentPrefixes to an empty array to disable semantic finalization while keeping empty-result recovery.

Set lateResultWaitMs to 0 to disable waiting for post-error fallback output.

Environment fallbacks are also supported:

• OPENCODE_TASK_RESULT_WAIT_MS
• OPENCODE_TASK_FINALIZE_TIMEOUT_MS
• OPENCODE_TASK_RESULT_POLL_MS
• OPENCODE_TASK_RESULT_MAX_BYTES
• OPENCODE_TASK_FINALIZER_AGENT
• OPENCODE_TASK_FINALIZER_MODEL
• OPENCODE_TASK_FINALIZER_REASONING_EFFORT

Tuple options take precedence over environment variables.

Limits

This covers synchronous foreground Task results. Background Task completion is injected as a synthetic parent message outside tool.execute.after and is not repaired by this plugin.

Semantic completeness is unknowable for arbitrary agents. By default, only agents whose names start with eye- are checked for a line equal to ## Verdict. Empty-result recovery applies to every agent.

The guard never exposes reasoning parts as a result. If recovery fails, it returns an explicit error and resumable task ID instead.

Relevant upstream reports include #24447, #27210, and #32132.

Development

bun install --frozen-lockfile
bun test
bun run build
npm run test:package
npm pack --dry-run --json --ignore-scripts