Bands

Claude knows what you changed between turns: files edited outside Claude are named (with small diffs) beside your next prompt, and an optional cheap review flags slips in your own edits.

Runs the tests related to each change (Claude's edits and yours) in the background; a failure caused by Claude's edit reaches Claude mid-turn, and the band offers Show and Fix it.

Say "this error", "the tests", "this file", "what you just did" or "the plan" and what it points at goes along with your prompt: the failing output, the test failures, the file, the diff, the plan. No model calls, size-capped, never twice, and Remove takes it back.

Keeps work flowing: when Claude stops mid-task ("Next I'll…") or asks an obvious yes ("Shall I continue?"), a 5-second countdown you can stop sends a short continue. Never for choices, missing info or risky steps.

Every prompt you type gets a quiet second read: without changing your words, a hidden note gives Claude the goal in one line, the likely scope, the project's checks to keep green, and what done means. Haiku only for vague prompts; See, Undo and Off in the band.

A quick second pair of eyes after edits: reviews the turn's diff in the background and offers to fix what it finds.

A prompt queue that runs itself: line up prompts with /queue and they run one after another, pausing when Claude asks you something or a turn fails.

Sharper retries after a failed turn: when checks still fail, Claude gives up, errors loop, or you say "that didn't work", haiku drafts a retry with the lesson, a different approach and the exact failing output. One click sends it.

No lazy placeholders: flags TODOs, '... rest of code', NotImplemented stubs and empty bodies an edit introduces, nudges Claude to finish them, and lists what is left.

One prompt, several jobs: when a prompt you type holds independent tasks (a list, numbered steps, "do A, then B, and also C"), offers to run them one by one as their own turns, stopping at the first that fails or asks you something.

Runs your project's checks after Claude edits, shows pass/fail above the prompt, and closes the loop with one-press (or automatic, bounded) fixes.

Jev (TypeSafe's System One model) as Claude's instant co-processor: bulk filter/classify/rank/semantic-grep tools, risk notes under permission dialogs, a check on unverified 'it passes' claims, and optional per-prompt effort.

Docs that went stale because of this turn's changes: changed names, flags, env vars and config keys are looked up in the docs, and Jev reads each section against the new code.

A definition of done Jev checks after every turn: per-criterion probabilities that demand evidence, a band above the prompt, optional auto-continue.

Keeps changes on task: every edit is judged by Jev against your last request (and Claude's plan) in a few hundred ms; drive-by changes go back to Claude.

Catches tests being weakened to make them pass: every test-file edit is judged by Jev in a few hundred ms against what you asked for.

One-tap follow-ups: after each answer, three short next prompts above the prompt, on keys 1-3.

Standing context you control: pin notes and files once, and every prompt you send carries them.

Park a half-written prompt, ask something else, and pop it back into the box later.

Notices the commands you keep approving by hand and offers a one-click standing allow rule (never for risky or interpreter-wide commands).

During long turns, haiku writes one line above the prompt every ~30 s on what Claude is doing and what comes next, from a compact log of its tool calls; a one-line recap toast after turns over two minutes.

Opt-in: after a turn that adds or changes functions, sonnet (or haiku) drafts focused unit tests in the background, in the repo's framework and style; Review, Add & run (only that file), or Dismiss; never writes without a press; daily budget cap.

Every subagent, live: a /subagents pane with status, model, elapsed time, steps, tokens, last tool and answer, plus a band while any run.

Per-prompt effort autopilot: classifies each prompt (Jev, Claude Haiku, or heuristics) and sets that turn's effort, and optionally its model, only when the switch beats the prompt-cache rewrite it causes.