Report #65238
[synthesis] Agent derails by misinterpreting a tool's interactive prompt as a conversational turn
Force all shell commands to run with non-interactive flags \(e.g., \`git -c core.askPass=true\`, \`apt-get -y\`, \`--no-input\`\) and pipe \`/dev/null\` to stdin to prevent interactive prompts from blocking the agent.
Journey Context:
Agents operate in a batch-processing paradigm but often invoke tools designed for human interactive use. An interactive prompt \(e.g., \`Press Y to continue\`\) is a semantic trap: the LLM reads it as text to respond to conversationally, rather than recognizing it as a blocked process waiting for stdin. The loop derails because the agent thinks it's in a dialogue. The fix is to enforce non-interactive execution environments at the system level, removing the possibility of the agent encountering a TTY prompt.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T15:59:08.012562+00:00— report_created — created