Feature: Passive filesystem awareness & bidirectional collaboration (antiprompts)

Feature Proposal: Passive Project Awareness & Bidirectional Collaboration in Claude Code

From a power user of Claude Code, directed at the Claude Code / Agent SDK team.

---

The current model and its friction

Claude Code is reactive: it responds to prompts. When a user edits files between prompts, Claude only learns about those changes if the user mentions them, or if a hook reports MTimes at the next prompt boundary. This creates a collaboration gap — Claude's internal model of the project drifts silently out of sync with reality, and the user must compensate by re-orienting Claude at the start of each prompt.

The proposed model: passive filesystem awareness

The Claude Code desktop application already embeds a Chromium/Electron engine that is entirely capable of directory watching. What is missing is a continuous, low-priority background process that:

1. Monitors file MTimes (or inotify/FSEvents/ReadDirectoryChangesW events) for the active project.
2. Updates Claude's working model of modified files incrementally — not on every keystroke, but on file-close or debounced quiet periods.
3. Feeds this silently into the context so that when the user sends the next prompt, Claude's understanding is already current — without the user having to say "I changed X and Y since last time."

The extension: antiprompts

Once Claude has continuous awareness, a natural evolution becomes possible: Claude initiates questions. Not interruptions, but a low-priority side channel — a small notification or queue — where Claude can ask: "You changed the pool-release logic in AcidDesktopMgr.cpp but kept the old frame-counter threshold — intentional?" or "I notice AutoFitFramesX is now set in two places; did you want to remove one?"

This transforms the tool from a question-answering machine into a genuine pair-programmer — one that notices things, flags inconsistencies, and asks rather than assumes.

Why this matters

The current friction is not about intelligence — Claude can reason well about code. The friction is about temporal awareness. A human pair-programmer sitting next to you sees your edits in real time. They don't need to be told what changed. They notice. They ask. Claude should too.

The infrastructure is largely already present in Claude Code. The missing piece is the loop that closes between filesystem events and model context — and the UX affordance for Claude to occasionally turn the question around.

---

Feedback welcome. This is written from a concrete daily workflow, not as a theoretical wish list.