How it works

From a hook to a notification

1. The agent reports what it is doing

   Claude Code, Codex and Cursor each run runmarq's hook when something happens: a prompt is accepted, a tool starts or finishes, approval is needed, the turn ends. The hook prints only what the agent's hook protocol requires and always exits successfully.

2. The hook keeps only timing and categories

   In the hook process, the prompt is classified (bug fix, feature, migration…) and dropped. Commands are inspected only to tell tests from builds and to recognise a repeated command by a salted hash. What leaves the hook is a small event: times, types, tool names, categories and salted hashes, never prompt text.

3. The collector follows the run

   A background process on this computer turns events into runs with an explicit state: running, waiting for you, finished, failed, cancelled or unknown. It separates active time from time spent waiting for you, and never marks an unclear ending as finished.

4. The estimator predicts what is left

   Each meaningful event, and every fifteen seconds of quiet work, produces a new estimate: an answer to whether to wait, and a likely range for the time left. Several models trained on this machine's runs contribute, weighted, once enough runs have finished, by how well each has predicted your recent ones. When the work checks out, the agent stops starting tools and your earlier runs show what that means, it says “Finishing up”. Nothing is sent anywhere to compute it.

5. You find out without looking

   The tray shows every active run's estimate. A notification arrives when a run finishes, fails, needs your approval or looks stuck, and each run's timeline stays in history. In a terminal, runmarq wait returns when the run ends and runmarq statusline keeps it in your prompt.