jcode, state by state
One small task, a flaky test, walked through every state the harness passes on the way to done. Every screenshot is a real frame rendered by jcode's own replay engine, the same machinery that plays back any saved session as video.
01Prompt in, memory recalled
Before the model reads the prompt, the memory system has already run: each turn is embedded and matched against a graph of past memories, and the hits are injected inline. Here it recalls why the backoff constant is 50ms and how to run this crate's tests, context the model would otherwise burn a turn rediscovering.
02Streaming begins
The model starts working. The context gauge in the corner tracks exactly what this turn costs; the header shows the provider, model, and MCP state at a glance.
03A tool runs
A live tool card: the spinner, the agent's stated intent ("Reproduce the flaky
test"), the exact command, and elapsed time. Alt+B would push it to the
background and let the agent keep thinking while it runs.
04The failure is the finding
The test fails, and that is the point: the card flips to an error state with the failing assertion inline, and the agent reads the timing math straight off it. No scrolling through a buried log, the harness renders failure as a first-class state.
05Todos with confidence
The agent files its plan through the todo tool, and every item carries a confidence score the harness makes it defend: confidence may only rise in evidence-backed steps, and a completion that spikes straight to 100% gets challenged to go validate.
06The edit, as a diff
File changes render as compact diff cards, +1 -1 on src/retry.rs,
expandable to the full diff inline. The git widget on the right tracks the working
tree as it changes.
07Verification, not vibes
The fix isn't claimed, it's hammered: 50 consecutive test runs in a loop. The harness is built around this pattern, agents that design their own verification instead of asking you to check.
08Done means verified
The full transcript at rest: prompt, memories, the failing run, the plan, the diff, the 50-run verification, and a final answer that states what was proven rather than what was attempted. All todos closed with completion confidence, or the auto-poke would have sent the agent back to work.
Frames rendered headlessly by jcode replay --video from a hand-authored
timeline in a scratch project, so the transcript is safe to publish and reproducible. Capture pipeline: tools/harness-walkthrough in the site repo.