v1.2.5

Headroom 1.2.5 (build 198).

What changed

Added

• Coding-agent approvals show the agent's actual request. Every field the
  provider sent is listed in reading order with its own label — an Edit shows
  the file, the text being replaced and the replacement, tinted so the pair
  reads as a before and after. Bulk fields sit behind Show request so the
  feed stays scannable.
• Why carries Claude's own stated reasons for asking.
• A value the host had to clip says Shortened to fit, and dropped fields
  are counted, so a prefix of a command is never mistaken for the whole one.
• Each agent row says how long it has been waiting — same words and placement
  as an activity row's age, because they are two halves of one feed. A request
  that has sat for six minutes reads very differently from one that just
  arrived, and the permission hook gives up at around five.
• A third answer, Always allow this exact request, saves a permission rule
  so Claude stops asking. Headroom writes only the exact command or path it
  showed you — Claude's own "Yes, don't ask again" widens a command to a
  prefix, and a grant made from a phone outlives the request that prompted it.
  The row prints the rule under the buttons before you tap, and glob
  characters in paths are escaped so a folder named [2024-06] Reports cannot
  match its siblings. Questions are never offered it.
• Each row carries the agent's own mark in its brand colour instead of one
  generic speech bubble, so a Claude row and a Codex row stop looking alike.
• Notices that can only be dismissed can be swiped away. Rows carrying a real
  answer cannot: a swipe that denied a permission would send Claude a decision
  by accident.
• You can answer Claude's questions from the phone. Its options become the
  buttons, and tapping one sends the choice back so Claude carries on without
  you touching the Mac. No hook can hand AskUserQuestion a selection — but a
  denied PreToolUse call is documented to show Claude the reason, so the
  choice travels as the reason. It is a workaround and behaves like one:
  Claude sees a blocked tool plus your words rather than a clean result, so it
  may occasionally acknowledge the block. Headroom answers only a single
  question of two to six options, never a multiSelect one, and everything
  else — a timeout, an odd shape, or Ask on Mac — defers, which leaves the
  question to appear on the Mac exactly as before. See
  docs/agent-attention.md.
• Installed hooks are now version 2, adding a PreToolUse entry scoped to
  AskUserQuestion. Settings reports Outdated until you reinstall them.

Fixed

• An AskUserQuestion row is readable. Claude's questions arrive through the
  permission hook, and the nested questions array reached the phone as a wall
  of raw JSON — the question was in there, but nobody was going to find it. The
  row now leads with the question itself instead of "Use AskUserQuestion", and
  each option is one control carrying the reason you would pick it. The first
  pass listed the descriptions above a row of buttons repeating the same
  labels, which said everything twice.
• Answer buttons take the account's accent instead of the system blue, so they
  belong to the agent that asked. Short answers stay bordered pills; a
  question's options are sentences with a reason under each, and tinting those
  turned every one into a large coloured slab. They read as plain rows with a
  divider and a chevron now — the shape a grouped list uses everywhere else on
  the system — with the colour on the chevron. Ask on Mac sits below the
  divider rather than among the answers, because it is the way out rather than
  another option.
• The provider mark moved to the top-right corner of an agent row. Which agent
  asked is a property of the row, not the first thing to read in the sentence.
• Claude finished responding no longer stacks. A session's finished or idle
  notice replaces the one it makes untrue, so the feed carries at most one per
  session instead of a wall of identical rows burying the approvals that
  actually want an answer. Superseding is scoped by session and kind, so a
  notice arriving can never close a permission request you have not answered.

Changed

• The phone used to decode four fields of a request and drop the rest, which
  made an Edit approval read as "Use Edit" and a Write show a path but
  never the content. detail.request now carries typed fields end to end; an
  unrecognised tool renders without an app update.
• The Claude adapter reads permission_reasons (the documented field) as well
  as the older permission_suggestions, and keeps tool_use_id / prompt_id
  for correlation.
• docs/agent-attention.md corrects its claim that structured questions are
  notify-only pending provider support. Against Claude Code 2.1.220 the
  Elicitation hook returns real form values, updatedPermissions makes
  "always allow" answerable, and decision.interrupt exists — all four are
  wiring gaps on our side, now written down as such.

macOS

Download Headroom-macOS.zip, unzip, open Headroom.app.

Notarized — Gatekeeper should accept it on first open.

On first launch the Welcome sheet starts the bundled host automatically. Confirm Claude / Codex / Cursor, then Continue.

iPhone

Install via TestFlight.

Pairing uses the mobile token (Mac Settings → Copy mobile token), not the host token used by the ESP32.

Full Changelog: v1.2.4...v1.2.5