v1.1.0

Changed

• Bring feedback into your agent with a command. Run
  r3 feedback fetch <artifact-id> in your harness to load new feedback and
  replies. Supported agents then listen for future feedback automatically. In the
  web UI, a new button replaces Copy prompt and helps you copy this command
  with the current artifact ID prefilled.

Fixed

• Feedback now has at-least-once delivery. Fetch and watch mark feedback
  delivered only after stdout succeeds, so failed output cannot be mistaken for
  delivery. Retries may repeat feedback.
• Resolution updates survive edits. Agents still receive resolution updates
  for previously delivered notes that were edited before being resolved.

Removed

• The r3 prompt alias has been removed. Use r3 feedback fetch instead.

v1.0.1

Changed

• A release can be retried one step at a time. Publishing the GitHub
  release, the per-platform packages, and the launcher are separate steps. The
  launcher waits up to ten minutes for those platform packages to appear on
  npm, and a failed step can be run again without rebuilding or repeating a
  step that already succeeded.

v1.0.0

r3 1.0 is redesigned around HTML artifacts: agents publish interactive pages,
and humans review them and leave feedback directly on page elements. Agents can
reply to comments and publish new versions, keeping the artifact and its
conversation together. Files and diff artifact types retain the document and code
review features from previous versions.

v0.13.0

Added

• r3 can actively wake up session to send notifications. r3 listen <id>
  registers the current Claude Code or Codex session and returns. When the
  reviewer submits feedback, approves, or abandons the review, r3 sends a
  notification into that agent session; the agent can then fetch the feedback
  with r3 prompt. Previously, agents must keep an r3 watch long running
  process to receive feedback.
• Session ids are recorded automatically. Reviews and work-in-progress
  badges identify the agent session involved. The UI shows a short id and lets
  you copy the full one, making it easier to find or contact the same session.

Changed

• Large reviews stay steadier while loading and typing. Deferred file cards
  reserve their expected height so the scrollbar does not keep shifting as you
  move through the review, and desktop paint containment roughly halves the
  measured composer-keystroke cost on a 200-file review.
• The agent guide better explains how to use r3. It now starts with which
  kind of review to create and walks through receiving feedback, claiming work,
  replying, and sending another revision.
• Anchoring feedback in scratch reviews is simpler. Agent commands can refer
  to a scratch file by its displayed filename, such as notes.md.

Fixed

• A failed listener hand-off no longer looks successful. The UI reports when
  Submit reached nobody, repeated Submits cannot be deduplicated into silence,
  and stale registrations no longer lock another agent out of the review.
• Marking a rendered Markdown table row keeps its columns aligned. The
  feedback bar no longer inserts an apparent empty cell ahead of the row.

v0.12.0

Added

• Fold the feedback dock to a rail. Collapse the panel to a strip that keeps the
  open count and a status glyph, and floats the composer next to the code when
  you anchor a note there.
• Resolved feedback is unmistakable. A resolved card takes a success wash
  and a ✓ resolved pill.

Changed

• Large reviews got a lot faster. A 200-file review scrolls smoothly,
  highlighted code arrives 59% lighter per line, a keystroke in a feedback
  composer costs one box instead of the whole page, and a multi-megabyte
  response compresses off the daemon's event loop so live updates and an agent
  blocked on r3 watch keep flowing.

Fixed

• A files review whose files have since been deleted stays quiet. The card
  reads "not in the working tree" instead of re-requesting the missing file
  every time you scroll past it.
• Returning to a backgrounded tab re-renders the review you are on, not
  every review that tab has visited.

v0.11.0

Added

• Mermaid diagrams render in a reviewed .md. A ```mermaid flowchart
  or sequence diagram now shows as a diagram — the server renders it to SVG —
  instead of highlighted source. Any other diagram kind falls back to an ordinary
  code fence.

Changed

• A large file no longer stalls the daemon. Syntax highlighting runs off the
  daemon's thread, so nothing holds up live updates or an agent blocked on
  r3 watch while a big blob is highlighted. A diff review also loads only the
  round you are looking at, expand-context reuses the diff it already derived,
  and marking lines or dragging the gutter stays smooth in long files.
• Snapshot and expand-context diffs use Myers, the algorithm git itself uses.
  A small edit between two large files now reads as a small edit rather than a
  full rewrite.

Fixed

• A binary or oversize file in a files review is marked as such, instead of
  being reported as deleted or quietly dropped from a snapshot.
• Feedback handed off to r3 watch can't be lost. The hand-off is drained in
  one step, so two agents on the same review can't both claim it and a note
  whose text resembles r3's own output no longer reads as an empty round.
• Removing a diff review's last round leaves it empty, instead of falling
  back to whatever the working tree holds now.
• r3 reanchor checks where you point it — the path and line range are
  validated on the way in, and the next automatic pass re-verifies the result
  rather than trusting it.
• A round or snapshot never reuses a number an older reply still refers to,
  so a @path:Lx-y ref keeps pointing at the code it was written against.
• r3 guide matches what the CLI accepts for files rm, snapshot rm, and
  scratch reviews.
• Forgetting a repo clears its reviews from open tabs.
• Editing a note you already handed off marks it undelivered right away.
• Approve says why it is disabled when you hover it.
• Jumping to a file keeps the current-file marker moving instead of freezing
  it where it was.

v0.10.1

Added

• One agent watches a review at a time. Two r3 watch clients on the
  same review no longer race the round. A second watcher is refused.

Fixed

• Editing a sent reply sends the new wording to the agent. The next
  Copy prompt, Submit, or r3 watch includes the correction.

v0.10.0

Added

• Claim work in flight. r3 claim <feedback_id>... shows in the UI
  which notes an agent is handling, so a long edit no longer looks like
  the agent wandered off.
• Fenced code in a reviewed .md is highlighted with the same syntax
  theme as the code view.
• Large files reviews stay scrollable. Bodies load as you approach
  them, instead of all at once.

Changed

• Approve no longer waits on open notes you chose not to chase. It
  blocks on unread content, or on an item the agent has claimed.

Fixed

• Re-anchoring keeps the quote you marked, instead of rewriting the
  note to whatever lines it was pointed at.
• r3 stop kills a stuck daemon instead of abandoning it, so later
  starts can recover.
• A spawn timeout names a sandbox that can't reach localhost, instead
  of only blaming a stuck daemon.
• Directory rows in the file tree line up with the files at the same
  depth.

v0.9.3

Changed

• Focusing a note no longer moves the file pane. Resolving or replying
  advances to the next card — and j/k walk the list — with each anchor
  ringing where it already is; the pane jumps only on an explicit locate: a
  card's file:line header, a reply's pin, or o.

Fixed

• Selecting rendered Markdown anchors exactly what you picked.
  • A quote crossing a link, a table row, bold text, or an   is matched
    against the text the browser shows, so it no longer comes back "the text
    this refers to changed" the moment it's saved.
  • When a doc repeats a phrase, the note — and its highlight — lands on the
    copy you selected, not a near-duplicate elsewhere in the file.
  • A selection longer than the quote cap keeps its full line range instead of
    being rewritten to the quote's first lines.
  • A triple-clicked paragraph no longer extends the anchor into the block
    after it, and a drag released over the feedback panel keeps the paragraph's
    whole text in the quote.
• Relative links between reviewed docs work. A [setup.md](setup.md) link
  in a rendered .md opened a URL that never existed; it now jumps to that
  file's card in the review — #heading fragments included — and a link to a
  file outside the review renders as dead instead of looking live and doing
  nothing.

v0.9.2

Fixed

• Feedback on rendered Markdown anchors to what you selected. A note on a
  bullet, a table cell, or a phrase mid-paragraph was recorded as the whole
  list, table, or paragraph, and now lands on the lines its quote occupies.
• Clicking prose in a .md review no longer hijacks the page. Only a click
  on the highlighted quote focuses its feedback, instead of a click anywhere in
  the block around it.