Detailed Analysis
A developer has released a small but telling open-source tool that addresses a subtle but persistent friction point in working with Claude Code, Anthropic's agentic coding CLI: knowing whether the assistant has actually finished a task or is simply waiting on a response before it can proceed. The tool, published under an MIT license on GitHub as "claude-needs-input," changes the color of a user's iTerm2 terminal tab based on Claude Code's state—pulsing orange when it is blocked and needs input, solid green when a turn is complete, and untinted while it is actively working. The distinction matters because Claude Code's textual output doesn't always make clear whether a session has truly concluded or is stalled awaiting clarification, a problem familiar to anyone who has walked away from a terminal only to find an agent idle for minutes on a question it asked long ago.
The technical challenge the author describes is more nuanced than it first appears. Simple heuristics—like checking whether Claude's response contains a question mark—fail to reliably distinguish between a "blocking" question that halts progress (e.g., "Which file should I edit?") and a "closing" remark that merely wraps up a turn (e.g., "Anything else?"). To solve this, the developer built a hybrid classification system combining rule-based logic with Anthropic's lightweight Haiku model, balancing speed and accuracy in real time. Notably, the tool works even when Claude doesn't use the formal AskUserQuestion prompt structure, meaning it can infer blocking states from more general conversational patterns rather than relying solely on a structured API signal.
The project also surfaces a real, currently open pain point in Claude Code's design: a recently introduced 60-second timeout on AskUserQuestion prompts, tracked as GitHub issue #73125 in Anthropic's repository. Users have found this timeout disruptive because it can cause a prompt to auto-resolve or time out before a human has a chance to notice and respond, especially during long-running agentic sessions where attention naturally drifts. The author's workaround—disabling AskUserQuestion entirely via the `--disallowed-tools` flag and relying instead on the tab-coloring heuristic to signal blocking states—is a pragmatic patch that sidesteps the timeout issue altogether while preserving the ability to know when input is needed.
This small utility reflects a broader theme in the current wave of agentic AI tooling: as coding assistants like Claude Code become more autonomous and long-running, the interface problem of managing attention and interruption becomes increasingly important. When an AI agent can work independently for extended stretches, users need lightweight, low-friction signals to know when their involvement is required, without having to constantly monitor a terminal or parse ambiguous text output. The rise of community-built utilities—color-coded tabs, notification hooks, custom wrappers—around Claude Code's core functionality suggests a maturing ecosystem where the CLI itself is treated as a platform to be extended, and where practical usability issues (like a too-aggressive timeout) get identified and patched by the community faster than they may be addressed upstream. It's a small example of how developer tools increasingly evolve through open-source scaffolding built directly on top of foundation model products.
Read original article →