Never thought I’d come to this, but the idea of leaving all my tech behind and disappearing into a cabin in the woods for few days/weeks, sounds so peaceful.
@pvncher You could actually ask Codex to enable and test computer use for itself. Got the binary and set up the MCP server.
Remember, if you feel like you are cut off from the magic, try wishing 🪄
Here’s the fun thought - the UK social media band will lead to kids getting more sophisticated in VPN usage. Some of them will figure out how to set up their own VPS with obfuscation protocols.
It’s like a free tech Boot Camp out there!
Encountered a bug in Claude Code @bcherny
Assistant responses are not rendered while a background subagent is running — UI shows "Waiting for 1 background agent to finish" instead, making the assistant appear unresponsive
Environment:
- Claude Code CLI v2.1.170, macOS (darwin 25.4.0), zsh
- Long-running interactive session using a self-paced loop (ScheduleWakeup) + the Agent tool with run_in_background: true and isolation: "worktree"
Repro steps:
1. In the main conversation, launch a subagent: Agent(..., run_in_background: true, isolation: "worktree").
2. End the assistant turn (in our case the turn ended with a ScheduleWakeup tool call — self-paced loop idle).
3. While the background agent is still running, type any user message in the terminal.
Expected:
The user message immediately invokes the main assistant; its text response is rendered in the terminal. Background agents are advertised as non-blocking ("You will be notified when it completes").
Actual:
The terminal prints only * Waiting for 1 background agent to finish after each user message. The main assistant actually receives each message and generates full text
responses in real time (they exist in the session transcript ~/.claude/projects/<project>/<session>.jsonl), but none of them are rendered to the terminal until the
background agent completes. From the user's perspective the assistant is frozen/ignoring them — in our session three consecutive user messages ("where is the
answer?", "why are you stuck?", "why are you ignoring me?") all got only the "Waiting…" line, while the transcript contains three complete, timely assistant replies.
Impact:
- Breaks the core promise of run_in_background (non-blocking).
- In autonomous/loop workflows this is a hard UX deadlock: a long-running subagent (ours ran 20+ minutes) silently holds the entire conversation's rendering hostage;
user cannot tell whether the assistant is alive, and conversation trust collapses.
- Notably, Workflow-tool background runs do NOT exhibit this — during a 12-agent Workflow run in the same session, user messages and assistant replies rendered normally. The blocking is specific to backgrounded Agent-tool tasks (possibly to the "Waiting for N background agents" gate that runs when a turn ends while an Agent task is alive).
Evidence available: session ID 8cd11d61-3087-4a44-9ded-0dbf4b9d2cdf, screenshot of the terminal showing three user messages each followed only by "Waiting for 1
background agent to finish", and the transcript JSONL containing the corresponding undelivered assistant replies with timestamps.
Took my half-open MacBook running Codex for a walk in my backpack for the first time.
It really is an odd feeling. Physical world in front of you, a digital one running on your back, and a phone in your hand as the bridge. 🌉