Agent sandbox run should fail or timeout when nested agent remains processing with pending tools

[chubes4]

Problem

WP Codebox wp-codebox.agent-sandbox-run can complete the recipe step with exit code 0 even when the nested Data Machine agents/chat result is still processing, has pending tools, and reached max turns without a final answer or file changes.

This lets parent Homeboy logic see runtime artifacts and attempt late outcome classification, even though the agent did not complete the task.

Evidence

Parent tracker: Extra-Chill/homeboy#3378

Overlay run:

• Run id: homeboy-3378-release-channel-fanout-overlay-20260603015451
• Transcript A: /var/folders/lr/c_cmmt7s0592m4njz99v5yb40000gn/T/opencode/homeboy-3378-release-channel-fanout-overlay-20260603015451-a-artifacts/runtime-mpxez1y8-ubvf6z/files/transcript.json
• Agent result A: /var/folders/lr/c_cmmt7s0592m4njz99v5yb40000gn/T/opencode/homeboy-3378-release-channel-fanout-overlay-20260603015451-a-artifacts/runtime-mpxez1y8-ubvf6z/files/agent-result.json

Nested agents/chat metadata:

status: processing
current_turn: 20
has_pending_tools: true
completed: null/false

WP Codebox agent result:

{
  "schema": "wp-codebox/agent-result/v1",
  "status": "completed",
  "actionable": false,
  "summary": "Agent sandbox completed without actionable file changes.",
  "changedFiles": { "count": 0 },
  "patch": { "bytes": 0 },
  "noOpReason": "no_file_changes"
}

The recipe command itself had exit code 0, and the parent Homeboy run later preserved empty patch artifacts.

Expected behavior

A sandbox agent run should not be classified as completed when the nested conversation is still processing or has pending tools at max turns.

Acceptance criteria

• Inspect nested agent_runtime.result.completed, metadata.status, metadata.has_pending_tools, metadata.current_turn, and max-turn diagnostics.
• If nested state is processing with pending tools or max turns reached, classify the WP Codebox agent result as timeout, incomplete, or failed, not completed.
• Emit a structured diagnostic such as agent_runtime.incomplete_pending_tools.
• Produce non-zero recipe outcome when the nested agent cannot complete and no explicit no-op final answer exists.
• Add smoke coverage using a fake nested agents/chat result with status: processing, has_pending_tools: true, and no file changes.

Related

• Sandbox perception should inject mounted workspace context like world mode #533
• Sandbox workspace adoption rejects linked worktree mounts #518
• agent-task: prove release-channel Lab Codebox fanout end to end Extra-Chill/homeboy#3378
• agent-task: do not classify zero-byte patch artifacts as actionable success Extra-Chill/homeboy#3383
• AI loop: model-facing tool results must preserve useful payloads Extra-Chill/data-machine#2470

AI assistance

• AI assistance: Yes
• Tool(s): OpenCode (openai/gpt-5.5)
• Used for: Reproducing the nested agent pending-tools outcome and drafting this tracker. Chris remains responsible for review and prioritization.