Repository navigation
Releases
v1.2.0
Three changes to the survey and the chapter card.
- "What this change is for" opens on a press. The reasons a branch
exists sit behind their own heading, shut on every draw, with a turning
arrow and a "press to expand" hint. Heading and reasons both went up a type
step — they used to be the smallest type in the app. - Mechanical chapters are last everywhere. The swept chapters are
sunk to the end of the chapter array once, upstream of every renderer, so
the progress bar, the survey lane, "Chapter n of m", Previous/Next and the
chapter a finished one moves on to all name one order. Previously the bar
could show a mechanical chapter second while the survey drew it last. - A shut chapter is drawn as a card. It narrows to a reading measure,
centres, keeps its words near its edges and lifts off the page; the text
got wider on the way. Pressing through gives the full width back to the
diff.
Upgrade: git pull && ./install.sh in your clone.
v1.1.0
The chapter card, reworked from a round of review notes.
- Centred. The card's column stands in the middle of the chapter; its
lines stay flush left. - Named once. The chapter's name is on the card while the card is up,
and moves to the focus bar only once the diff is open — never on screen
twice. - A tick on every card. "Approve chapter" is offered on every
chapter's card, so the last chapter can be settled from its card like any
other. It moves on to the next unread chapter, as the tick under the diff
does. - The tick is its own press. Unticking on a card leaves the card
standing; it no longer opens the diff underneath. The card body and "Read
the diff" are the only ways in.
Upgrade: git pull && ./install.sh in your clone.
v1.0.0
lightspeed-review v1.0.0
Semantic diff review with targeted agent feedback — first public release.
An agent opens a review of a branch. A human reads the grouped diff in the
browser, selects the exact lines that are wrong and writes a comment. The agent
polls, receives the comment together with the selected code, and fixes it.
Then the agent can get all feedback together with the context the comments
were made about and improve the docs (AGENTS.md. CLAUDE.md, SKILLS etc.)
so that recurring patterns are fed back into future tasks.
The review
- Chapters, not files. A model groups the branch diff into chapters ordered
for reading: the change carrying the stated intent first, cause before
consequence, a contract before its uses, security-critical changes and new
dependencies high, tests last. Each chapter carries a one-sentence rationale;
a grouping that asks the reviewer questions or gives orders is sent back for
one repair round. - Study and sweep. Chapters with nothing to decide — renames, moves,
generated output, formatting — sink into a mechanical lane the survey
approves in one press. Guardrail files and tests are always read. - One chapter at a time. The page opens on the survey, with no diff drawn.
A chapter opens alone behind a card that says what it is for; ticks close
what they mark — a file under its diff, the chapter under its last file.
Finishing a chapter moves you on to the next one still to read; the last tick
puts an End review card over the page. - Targeted feedback. Select lines, write a comment, queue it. Send delivers
the comment with the selected code verbatim. The queue, a half-typed comment
and your place in the review all survive a reload. - Rounds. Re-run
startafter every round of fixes. A new round is offered
in the header rather than replacing the diff under you; between rounds a
replay shows what became of each comment. Approval carries across rounds for
files whose blob sha proves them untouched — edit one and its tick is gone. - Unified or side-by-side, light and dark, renames traced, the densest new
branching badged,prefers-reduced-motionrespected.
The agent side
lightspeed start <branch> [base] --intent "<why>"opens the review;
pollblocks until feedback arrives;approvals,end,stop,serve,
login/logout, andfeedbackfor the ledger.- Every command prints TOON on stdout, failures included:
error: {code, message, detail}plushelp[]. lightspeed init --agent <id>writes the integration skill where that agent
reads it: pi, Claude Code, Codex, OpenCode, VS Code.- A feedback ledger keeps every comment across sessions with a verdict —
addressed,ignored,repeated,unknown— re-judged on every round.
Configuration and security
.lightspeed.conf.jsonin each reviewed repository names the provider,
model and thinking level; no silent defaults. Providers are pi-ai's
(anthropic,openai,google, …), plus proxies and gateways with their
own headers, plus subscription sign-in vialightspeed login.- The server binds
127.0.0.1only, checks Host and Origin, reads the
repository through git with argument lists (never a shell) and serves only
the files a session names. No telemetry. SeeSECURITY.md.
Install
git clone https://github.com/JakobKrummeich/lightspeed-review.git
./lightspeed-review/install.sh --agent pi # or claude-code, codex, opencode, vscodeRequires Node ≥ 22.18 and git. Not on the npm registry; the package is
lightspeed-review, the command is lightspeed.