claude-code - 💡(How to fix) Fix [Bug] Background Agent produces no deliverable when parent session hits usage limit [2 comments, 3 participants]

Bug Description Claude Code Bug Report — Background Agent + Session Limit Date: 2026-05-04 Reporter: [email protected] Product: Claude Code CLI Summary Background Agent tool reports completion but produces no deliverable file when the parent session hits its usage limit mid-run. User loses work and has no way to recover agent output. Environment OS: Windows (path C:\Users\giris...) Mode: Auto mode active Plugins: context-mode, aws-serverless, document-skills, anthropic-skills, others Session usage at 98% when background agent launched Reproduction Parent session at 98% usage. Launch background research agent via Agent tool with run_in_background: true. Prompt instructs agent to Write final report to a specified absolute path. Receive notification: agent completed. Verify file at expected path → file does not exist. Re-launch second background agent with same instructions. Parent session hits usage limit before second agent completes. No SendMessage tool available in current toolset to query/resume agent. Expected Background agent that completes successfully should leave its Write artifacts on disk. If agent fails, parent should see error/stderr, not "completed" status. Parent session should be able to retrieve agent results even after usage limit (e.g. agent output persisted to a known location). Actual Agent reports complete; no file written. No mechanism to inspect agent failure without reading the JSONL transcript (which the harness explicitly forbids — "would overflow your context"). When parent session ends, background agent state and any in-flight work appears lost. Impact Silent data loss. User cannot trust "agent completed" notifications. Background agents are most valuable precisely when context is tight — but that is exactly when they fail opaquely. Suggested Fixes Persist background agent's final tool-call summary (especially Write paths) to a small status file the parent CAN safely read. Distinguish "agent finished its turn" from "agent achieved its goal" in notifications. Provide a SendMessage / AgentStatus tool by default so parent can query a completed background agent for its deliverables. When usage limit nears, surface a warning before background agents are launched so user can decide whether to defer. Agent IDs (for log lookup) a93ff8663f50186cc — first run, "completed" but no file written a66a22198a3656213 — second run, parent hit limit before completion

Environment Info

• Platform: win32
• Terminal: mintty
• Version: 2.1.126
• Feedback ID: eb4bff33-cc96-48e7-89ec-18ec4c04c175

Errors

[]