Every command, run against this repo.

❯gradient

the whole picture, no arguments

Every number here is counted, not estimated. The installed block is the uncomfortable one: both artifacts gradient generated for this repo have been used zero times, and it says so rather than quietly padding a coverage percentage.

Note what the cost lines do not claim. Nudges are reported as attention spent across the tokens you had to drive — automating them saves your attention, not those tokens. An earlier version framed that 3.1M figure as recoverable. It is not, and saying so was the fix.

❯gradient optimize

find what recurs and what has gone stale, then propose

Everything here is computed locally, with no model call and no network connection. Clustering is model-free; so is resolving the paths and scripts your instructions mention against the actual repository, and reading the frontmatter of every installed skill.

Read the rejection lines: this run threw away more than it kept, and said why. A pattern seen only within a single day is project history, not a habit. An approval phrase like lgtm is a nudge, not a workflow. An artifact that would only repeat the prompt it was mined from costs more to invoke than to retype.

Findings are ranked worst-first and grouped by family, each with an id, an evidence line you can check, and the exact change it would make. --json hands the whole set to your assistant, and every run writes a self-contained local page you click through.

The findings are about files. The block under them is what runs between your prompts, and what each switch that is off would do — the person deciding which findings to act on is the person deciding what to automate.

❯gradient optimize --apply

write exactly what you named

Every write to a file gradient does not own is snapshotted first and guarded by a content hash: if the file moved since the finding was computed, the change is refused rather than forced. Concurrent sessions take an advisory lock.

--apply does not re-mine. It is a follow-up to a run you are already looking at, so rebuilding findings from hundreds of transcripts would make approving a one-line change the slowest thing gradient does.

Every run that writes gets an id. gradient optimize --undo <runId> puts it back — except for a file that changed again afterwards, which is reported rather than restored. Undo must never be the thing that loses work.

❯gradient remove

undo exactly one artifact

Deletes one generated artifact — the file or hook it created and its manifest entry, nothing else. Adjacent hooks you wrote yourself, and your permissions block, are left untouched. Every experiment on this page was cleaned up with exactly this command, and the settings file was byte-compared afterwards.

❯gradient on|off

continuity · autopilot · board · optimize

continuity writes a progress file before a compaction and reads it back on resume, so a context death costs you a paragraph instead of a re-explanation. autopilot answers the nudges — the continue, lgtm, go on turns that carry no information — up to a budget you set, and never past what a project’s gradient.md allows.

board shows what your other sessions are doing in the same repo. optimize opens each session with the top fresh suggestion. Both write to the same local settings file, and turning either off removes precisely what it added.

❯the checkup page

written by every optimize run

A file:// document with no server, no port, and no external reference of any kind — no stylesheet, script, font, or image. It renders identically offline and cannot report what it is displaying.

Accept or deny each finding and it builds the gradient optimize --apply … line in the footer for you to copy back into the session you are already in.

❯gradient optimize --auto

headless, on your schedule

--auto applies only additive, reversible, gradient-owned changes. It never edits prose a person wrote, never installs a new artifact, and never writes a line carrying a command invocation — at any tier, in any family.

gradient on optimize installs both halves of the loop: a SessionEnd hook that re-checks after a session ends, debounced to at most once a day, and a SessionStart hook that surfaces the highest-severity pending finding at the next one.

gradient installs no daemon and owns no timer. --print-schedule prints the snippet for your platform; installing it — and removing it — stays yours.