AgentCrash Resume Harness

A persistent harness that lets long AI coding tasks resume after a crash

Problem

When an agent runs long jobs like analyzing a repo, editing many files and running tests, a process crash can wipe the conversation and progress that only lived in memory. Restarting means guessing where it stopped and usually re-running from scratch, burning tokens and time.

Solution

A durable agent harness that commits conversation state, model turns, tool calls, queued input, checkpoints and agent status to local SQLite first, then surfaces them in a TUI and web console. After a crash or manual abort, it resumes from the last checkpoint and clearly shows completed vs. pending steps.

Users

Indie devs and small teams running long Claude Code / Codex style tasks; engineers doing multi-step local automation.

MVP

Store sessions and checkpoints in local SQLite; list resumable jobs and resume with one command; show completed vs pending steps and last tool result; ship a CLI plus a minimal web view; no cloud or multi-machine sync.

Revenue

Open-source core with a paid Pro tier (multi-project management, shared checkpoints, hosted sync); one-time personal license around $29–49.

Why now

Coding agents increasingly own long multi-step workflows, yet official runtimes still lack durable resume. People are already hacking their own harnesses in the open, so the need is real and the tooling is not settled.