v1.3.2

Added

• tailr demo starts a small sample application, proxies it, and prints a
  review URL: the whole round trip against something that is not your own work,
  with nothing installed into a project and no agent involved. The sample app
  ships with the package now, so npx -y @gcrft123/tailr demo needs nothing
  cloned first.

Changed

• The line printed when a session starts names the key the reviewer will
  actually be holding, instead of always saying Alt at someone who has changed
  it. It also stops handing the reviewer's instructions and the agent's next
  command to the same reader, since only one of those two people has a terminal.
• tailr init closes by telling whoever ran it what to do next, rather than
  what the agent should do next. It used to end on tailr wait and tailr pull, the two commands that are not a person's to run, which is a fair way
  to leave someone thinking they had misread who does what. It now says where
  the rules landed, that the agent re-reads them from there, and leaves the one
  command a person might actually want, which is starting the session.
• The README lists everything init edits where init is first mentioned —
  four things, where it used to name three — and PROMPT.md has the agent say so
  before running it. Requirements now covers what Tailr asks of a dev server
  rather than only which Node it needs, including that a mark's source address
  is null on a project whose tooling emits none, and that this is a fallback
  rather than a failure.
• The skills installed without a marketplace are now tailr-start,
  tailr-review and tailr-config, so the commands are /tailr-start and
  /tailr-config. That covers npx skills add gcrft123/tailr -g and the
  Antigravity and Gemini extension installs alike: all of them read the skills
  straight out of the repository and add no namespace of their own, so the old
  names arrived bare and /start and /config landed on top of commands the
  agent already had — Antigravity's own /config among them. A marketplace
  install is unaffected and still reads /tailr:start, because that path adds
  the prefix itself. Anyone on one of the unprefixed paths should install again;
  the old names are gone rather than aliased.

Fixed

• A dev server that compresses its HTML regardless of what was asked of it now
  gets the overlay. Tailr requests uncompressed HTML; the servers that send gzip,
  deflate or brotli anyway had their pages passed through unrewritten, which put
  the reviewer on a page where holding the key did nothing and said so nowhere.
  An encoding Tailr still cannot undo passes through intact, as it did before.

v1.3.1

Fixed

• --target localhost:5173, --target 5173 and --target 127.0.0.1:3000 now
  mean what they say. A scheme-less address used to start a session that could
  only answer 502, and a bare port crashed with a stack trace; what still cannot
  be a dev server URL is refused in a sentence.
• A second tailr started in a project that already has a live session — on
  another port, say — is turned away with that session's URL, instead of
  overwriting .tailr/session.json and then deleting it on its way out, which
  left the first session running but unfindable by status, wait and the
  MCP tools.
• tailr progress with a ref that is not in the batch is a 400 that names the
  refs that are, rather than a silent success that counted a mark nobody made
  as landed.
• tailr init registers the MCP server as npx -y @gcrft123/tailr mcp, the
  same as the plugin, so a client launching it in a --no-install project
  never stalls on npx asking permission to fetch.

Changed

• PROMPT.md tells the agent to run init again with --file when its own
  instruction file is not one the project already had: on a fresh project
  init writes AGENTS.md, which Claude Code does not read on its own, and the
  rules were landing somewhere the agent would never look again.
• The landing page's agent fan carries Antigravity in place of Gemini CLI —
  its own mark, and agy plugin install as the command it hands over.
• The README's Gemini CLI section is now an Antigravity CLI one
  (agy plugin install https://github.com/gcrft123/tailr), since Google turns
  individual accounts away from Gemini CLI, and the npx skills add paragraph
  says where that installer actually puts things now and that the skills land
  there without the tailr: prefix.

v1.3.0

Added

• Sound, on every action. A mark made, kept, reopened, dropped or discarded;
  versions or a slider asked for; a batch sent; a version picked, a value kept
  or reset; the batch taken back, re-sent or reloaded into; the session ended;
  the island landing in a new corner. Nine cues, each with its own shape rather
  than the same click at different volumes, and synthesized live through the Web
  Audio API — nothing is fetched and no audio file ships. Two things stay quiet:
  a key going down — bar the Enter that commits a comment, which hands the agent
  something to do — and a single mark coming back applied, which you watch empty
  out on screen; the run closing is what you hear. On by default; sfx:false
  turns the lot off.
• /tailr:config (/config in Cursor), tailr config, and the tailr_config
  MCP tool: read or change your settings, e.g.
  /tailr:config sfx:false modifier:cmd. Settings belong to the person rather
  than the project, so they are kept in ~/.tailr/config.json and hold across
  every project — nothing new appears in your repository. A change made while a
  session is up is pushed to the open review page, so it takes effect without a
  reload.
• modifier, so the key held to arm marking can be alt (the default), ctrl
  or cmd when an app or an OS wants Alt back. Every hint on screen names
  whichever key is yours, and tailr_status reports it so the agent never has
  to guess which one to tell you about.

Changed

• The interaction sounds are Cuelume,
  copied into src/overlay/cuelume.js under its MIT licence rather than
  installed — its whole seventeen-recipe palette, so a version bump stays a copy
  rather than a merge. Tailr still has no dependencies.

v1.2.0

Added

• A slider companion to the 1×–4× variations toggle on the comment composer.
  Turn it on to ask the agent for a continuous parameter (glow, depth, scale…).
  The agent wires it behind data-tailr-slide-<ref>, reports the range with
  tailr slider / tailr_slider, and the reviewer scrubs it on the page before
  keeping a value — the same round trip versions use. The pill's track and thumb
  are drawn rather than left to the platform, and the number beside them has no
  spinner arrows: this is a value to scrub, not a form to fill in. Keep
  minimizes the pill to that value with its reference number, which reopens it,
  and Reset puts the parameter back to the default the agent reported.
  Keeping the value already kept takes the keep back, the way clicking the
  version already chosen does.
• /tailr:start is a slash-command skill. Typing it starts a session against
  the dev server and begins watching for the first batch; the model will not
  fire it on its own. The review loop stays a background skill, loaded when a
  batch is in play rather than offered as a second command. On Cursor it is a
  rule rather than a skill, because plugin skills all become slash commands
  there and /review is already a built-in. /start is a Cursor command file:
  pointing skills at the start folder finds nothing (Cursor wants a parent of
  skill folders), and disable-model-invocation hides plugin skills from the
  slash menu.
• Plugin catalogs for Cursor, Codex, GitHub Copilot CLI, and Gemini CLI, so the
  same bundle that Claude Code installs is what those agents install too. Agents
  without a marketplace of their own take the skills globally through
  npx skills add gcrft123/tailr -g.

Changed

• Mark reference numbers reset with each Tailr session. The server names the
  process; the overlay drops leftover marks from a previous process and starts
  the count at 01 again.
• Clicking the badge on a text mark reopens the inline editor, with a Delete /
  Done bar so the staged change can be revised or reverted. Double-clicking
  text that already has a mark, and commenting an element that already has a
  comment, reopen the existing mark instead of stacking a new one.
• A slider's row in the staged list carries what its pill carries — Keep, lit
  when that value is the one being kept, and Reset — plus the × that turns the
  slider down altogether, which stays in the list for the same reason a version
  set's does.

Fixed

• Version and slider pills on an element in the corner of the viewport no longer
  render offscreen. Both flip below the element, or slide along its edge, the way
  the comment composer already did. A slider pill also follows its element while
  the page scrolls, which it never did.
• The slider pill is no longer rebuilt when the reviewer keeps a value, so the
  button under the pointer survives being clicked and keyboard focus stays where
  it was. Both faces of the pill are now built once and swapped by class, which
  is what lets the change of shape animate at all — and it is the rule the
  version pill above it has always followed.
• A sent batch now disables a slider pill's controls rather than only greying
  them, and Keep reads as kept: its lit state used to be the same paper as its
  resting state, so it could not be told apart.
• The Claude Code app would not add Tailr from this repository. It looks for a
  plugin at the repo root (.claude-plugin/plugin.json, .mcp.json, skills)
  and the owner/repo marketplace shorthand clones over SSH, which the app
  cannot complete. The root now carries a plugin manifest that points into
  plugin/, the MCP server lives at .mcp.json where the app discovers it,
  and the README's add command is the HTTPS git URL.
• Latched markup read what someone was typing as commands. With the mode on, a
  caret in the application's own text box — or in the one Tailr opens to edit
  text in place — lost c, r and e to the comment, remove and edit verbs
  and the arrow keys to structural walking, and the characters never reached the
  field. A keystroke on its way into an input, a textarea, a select or a
  contenteditable is now typing rather than a shortcut. Escape still reaches
  Tailr from a field of the application's, where it is not swallowed and the
  field answers it too; the exception is Tailr's own inline editor, which
  answers Escape by putting the text back without also dropping the mode the
  reviewer was editing from.

v1.1.0

Added

• Versions. A comment on an element or a spot can ask for up to four answers
  instead of one: the composer carries a 1× button beside Add that cycles to
  4×. The agent builds every version at once, each guarded on an attribute
  Tailr sets on <html>, and names it in a word or three. After the reload a
  pill sits on the element with a tab per version — hover one and it widens to
  the name while the page switches to that version live, so the comparison is
  between the real things rather than between two descriptions of them.
• Keeping a version is a mark like any other. It joins the batch, and sending it
  is what makes that version permanent and takes the losing ones — and the
  guards — out of the source. Turning the whole set down is the × on its row.
  An unresolved set keeps the island awake, because the alternative is versions
  sitting in someone's repository with nothing on screen that would remove them.
• tailr variants <ref> <names…>, and the matching tailr_variants MCP tool,
  for reporting what was built. The operating rules carry the whole contract:
  where the switch lives, that version 1 must also be what renders without it,
  and that a guard never outlives the choice that settles it.
• A way out. End session sits under the island's panel, and it asks before
  it acts: the card names what is being agreed to, one consequence per line —
  the marks that go unsent, the versions that get cleaned up, the address the
  application goes back to once Tailr stops proxying, and that the browser is
  cleared. Until now the only way to end a session was a terminal the reviewer
  does not have.
• A cleanup pass on the way out, because ending is the last moment anything can
  be done about what Tailr left behind. Versions nobody chose between go to the
  agent as one final batch that takes them, and the switches guarding them, out
  of the source; the switches come off the document; what Tailr kept in the
  browser is cleared; the server stops and the overlay takes itself off the
  page. If the agent never answers, End anyway leaves regardless and says
  the cleanup did not finish rather than implying it did.

Fixed

• A session was good for exactly one batch. The server keeps the last run
  indefinitely, so a reviewer who reloaded was handed a finished run again and
  the overlay re-entered its "Needs refresh" state — against changes they were
  already looking at — with Send locked and no way back. A closed run now only
  prompts the page that watched it close.

v1.0.0

Added

• A Claude Code plugin, with this repository as its marketplace:
  /plugin marketplace add gcrft123/tailr, then /plugin install tailr@tailr.
  It carries the MCP server, the operating rules as a skill, and a /tailr:start
  command — the same setup tailr init performs, without editing anything in the
  project, and updatable from /plugin rather than by re-running a script. The
  rules in it are generated from the same source init writes from, and the
  version it advertises is stamped when a release is cut; npm test fails if
  either drifts, so an installed plugin cannot quietly fall behind the protocol
  Tailr actually speaks.
• A release flow keyed to the tags: pushing vX.Y.Z verifies the tag against
  the manifest, runs the tests, publishes to npm with provenance, and cuts a
  GitHub Release from this file. Publishing authenticates as a trusted publisher
  over OIDC, so no npm token is stored anywhere. npm version is the whole
  interface to it.
• A test suite — npm test, no dependencies — covering the bridge state machine
  and its refusals, the proxy's injection and passthrough, the three outcomes of
  tailr wait, and the idempotence of tailr init. CI runs it on Node 18, 20
  and 22, and fails if Tailr ever takes on a runtime dependency.
• Source resolution reads more of what dev tooling already emits:
  data-v-inspector (Vue), data-inspector-relative-path (react-dev-inspector),
  data-astro-source-file (Astro), Svelte's __svelte_meta, and a generic
  data-source, alongside the React fibers and data-tailr-source it already
  read. Windows paths and trailing column numbers are handled.
• npm run demo starts the demo app itself instead of expecting one on a port
  nothing had started.

Fixed

• An https:// target crashed the session process on the first page request.
  https dev servers are now proxied, hot-reload upgrade included, and a
  self-signed certificate is accepted — a local dev server's always is.
• Rewritten HTML carried both the upstream's transfer-encoding: chunked and a
  freshly computed content-length. Strict clients refuse that combination
  outright, and every dev server streams its HTML. HTML arriving compressed is
  now passed through whole rather than decoded as utf-8 and injected into.
• tailr status exited 0 while the agent's own run was still in flight,
  reporting leased work back to it as a waiting batch. It now reports what the
  server calls pending.
• Closing a run that was already closed returned 500 Server error. rather than
  409 No open run. — an agent retrying tailr done after a dropped connection
  could not tell that apart from a real failure.
• A synchronous failure inside the proxy took the whole session down with it. It
  now costs one page, and the review URL stays up.

Changed

• PRODUCT.md no longer describes the project as having no implementation, and
  states exactly how far source resolution reaches. The overlay surface doc
  describes the Island that shipped rather than the withdrawn Callout direction.