FlashBob

The Problem AI coding agents running inside editors like IBM Bob or VS Code operate with a finite context window — typically 100K to 270K tokens, meaningful effective context is even lesser. When an agent is given a task in a large codebase, it must decide which files to read before it can act. The naive approach — loading all files — saturates the context and leaves little room for reasoning, generation, and tool execution. The careful approach — having the agent itself scan and triage files — burns tokens on classification work before the real task even begins. Either way, the agent's most valuable resource is spent on a selection problem, not the problem it was hired to solve. There is no clean separation today between "figure out what's relevant" and "do the work." Both jobs fall on the same model, in the same context window, at the same time. The Solution: FlashBob FlashBob is a cloud workspace mirror with an embedded MCP server and a Jev-powered context selection layer. When a developer opens a folder in IBM Bob, a lightweight sync script (or extension) pushes a full file snapshot to the FlashBob web app over a single HTTP POST. From that point on, every file save streams a delta patch — keeping the cloud mirror in persistent, live sync with the local editor state. The mirror is exposed over two surfaces. First, a clean web UI lets developers browse the synced workspace, inspect any file, and interactively run Jev relevance checks against it. Second, and more importantly, an MCP server runs as a streamable HTTP endpoint at /mcp — directly inside the React Router app, no separate process required. IBM Bob agents connect to this URL in their MCP configuration with a single line. When the agent receives a task, instead of loading files itself it calls get_relevant_files(workspaceId, task) on the FlashBob MCP server. That tool calls Jev — TypeSafe's System One decision model accessed via OpenRouter (typesafe/jev-1.13) — which evaluates every file in the mirror against the task description using two typed questions: a noul (yes/no probability) for relevance and a score (0–2 scale) for importance. Jev returns structured probabilities, not generated text, making it fast and cheap. FlashBob filters and ranks the results, returning only the files that matter. The agent's context window receives a ranked shortlist, not a full codebase dump. Target Users Developers using IBM Bob or VS Code who work on non-trivial codebases and regularly delegate implementation tasks to AI agents. What Makes It Unique The separation of concerns is the key insight: context selection happens outside the agent's context window, in a dedicated classification layer (Jev), while context execution stays clean inside Bob. The cloud mirror provides persistent, always-fresh state that any number of agent sessions can query without re-reading the disk. The MCP interface means zero changes to how developers prompt their agents — they just get better results.