@aragon-agent/cli 0.6.24 → 0.6.26

package/CHANGELOG.md

@@ -1,1786 +1,1924 @@
-# Changelog
-
-All notable changes to `@aragon-agent/cli` are documented here. Entries at
-`0.4.x` and earlier refer to the package by its former name, `@argon-agent/cli`
-— those releases really were published under that name, so they are left as they
-were written.
-
-## Unreleased
-
-### Fixed
-
-- **`Ctrl+I` was dead on Windows.** The win32-input-mode translator only
-  recognised the chord when the console reported it as `VK_TAB` (vk 9) with
-  ctrl - a record shape no real console was ever probed for. Physical
-  Ctrl+I keeps the LETTER vk (0x49, the same Ctrl+letter rule that reports
-  Ctrl+A as vk 65 and Ctrl+J as vk 74 in the Shift+Enter probe captures)
-  while `char` carries the Tab byte, so the record fell through to the
-  generic Ctrl-combo branch, came out as plain Tab, and the keybinding
-  silently did completion instead of opening the index confirm. The
-  translator now tells Ctrl+I apart from Tab by `char === 0x09 && ctrl`
-  whatever vk it rides, with Shift still mapping to Shift+Tab; both record
-  shapes are pinned by regression rows in `win32-input-mode.test.ts`.
-
-### Team dispatch supervisor (team-overseer)
-
-- **A time ceiling on a subagent is now an inspection trigger, not a death
-  sentence.** With the supervisor on (new key `team.overseer`, default
-  `true`, needs the fast tier), two soft timers per child - event silence at
-  `idleTimeoutMs` and a wall clock at `subagentTimeoutMs` (or 5 minutes when
-  no ceiling is configured) - fire one bounded fast-tier inspection. The
-  supervisor decides wait / nudge / replace / abandon; the runtime applies,
-  funnelling every lifecycle change through the worker loop that awaits the
-  child, so a replacement is never built for nobody.
-- **A nudge is steering; a replace is a rebuild with an amended brief.**
-  Guidance is queued at core's safe top-of-loop checkpoint; a replacement
-  carries the supervisor's reason, the dead child's `filesTouched`, and the
-  original brief bounded to the same prompt budget. A post-mortem inspection
-  diagnoses a child that died silently (watchdog death, stream loss) and may
-  re-assign it once.
-- **Budgets are structural, not settings**: 4 inspections and 1 replacement
-  per child per dispatch; past them the child runs under its own (lengthened
-  while supervised) idle watchdog and the supervisor goes quiet. An
-  unparseable or failing supervision call always becomes `wait` -
-  supervision must never kill a healthy child.
-- Every applied decision is visible three ways: the report's Supervisor
-  section (`[silence 30s] a1: nudge - ...`), the team panel and the `-p`
-  stderr line, and a `debug` log event. `aragon exec` emits it as an
-  additive `team` subtype `overseer` (`EXEC_SCHEMA_VERSION` unchanged).
-- New command `/team overseer on|off` (refused mid-dispatch, like every
-  other `/team` setting) and new config key `team.overseer`, also settable
-  as `aragon config set team.overseer false`. With the supervisor off - or
-  no fast tier - every path is byte-for-byte the pre-feature regime:
-  `subagentTimeoutMs` is a hard abort and silence is the child's own
-  watchdog.
-
-### 启动与缩窗不再误报 Frame diff lost sync
-
-- 启动即现的 "Frame diffing lost sync with the terminal and fell back to a
-  full repaint … `--no-diff-render`" 警告已消除：Ink 挂载/卸载经渲染 stdout 写出的
-  光标隐藏/恢复序列（`ESC[?25l` / `ESC[?25h`）由帧写入器直接透传——不进差分、
-  不失效缓存、不计 fallback，也不再消耗会话的首写豁免。启动序列恢复
-  "种子帧透传 → 首帧全量 → 之后纯 diff"。
-- 终端高度缩小（拖小窗口、改字号、面板重排）不再触发同一条警告：Ink 的
-  `clearTerminal` 整帧写入被差分器识别为 Ink 自身写法，透传并失效缓存但不计
-  fallback；该控制写同样不消耗首写豁免，其后的首个 log-update 帧仍按种子帧豁免，
-  下一帧自然全量重绘。
-- `/perf` 的 `fallbacks` 重新从 0 起步，"健康会话 fallbacks === 0" 的诊断信号
-  （AC-11）恢复可判。
-
-### Consistent TUI styling
-
-- Interactive sessions always use Unicode and default to truecolor, including
-  PowerShell sessions without terminal or UTF locale markers. Ink's actual color
-  output follows the same policy, with renderer state restored on exit.
-- Explicit color opt-outs keep Unicode borders and icons; the opening logo now
-  depends only on available space. Headless terminal detection is unchanged.
-- Reloading model settings preserves session display preferences. Syntax
-  highlighting follows the same color policy, and status icons use consistent
-  single-column glyphs to prevent narrow-frame overflow.
-- Hiding the editor caret while scrolling preserves draft geometry, preventing
-  a layout update loop when entering text above the full startup logo.
-- Register exit cleanup before mounting the application, so immediate exits and
-  initial render failures restore terminal and color state reliably. Failed
-  mounts no longer mark an auto-installed build as healthy.
-
-### Model context window
-
-- Discover context limits at startup and after connection changes without blocking
-  input. The status bar and compaction use the discovered limit; `/context` names
-  API versus catalog metadata. Unknown windows show `?`, not a fabricated 128k
-  capacity or percentage. Manual `contextWindow` overrides retain precedence.
-- Window overrides now apply to tail relief and compaction progress as well as
-  the gauge. Workers on different models/endpoints do not inherit the lead's
-  override. API-key changes refresh metadata, and ending a session cancels its
-  pending model queries. `/context` also hides unknown capacities/percentages.
-
-### TUI 输入与排队可靠性
-
-- Shift+Enter 与粘贴按输入顺序处理；普通 Enter 保持弹层兼容，混合输入最多提交一次，
-  余稿及拒绝提交的正文保留。结束输入流时完整排空，不泄漏内部控制标记。
-- 超限粘贴拒绝同一外部输入块的正文和 Enter，避免误提交旧稿；未闭合粘贴继续按正文消费。
-- 拖选后仅 Ctrl+C 复制；文字变化与丢失鼠标释放时取消无效高亮，避免卡住视口。
-- 固定 Queue 状态行按消息 ID 回执更新，支持窄屏、暂停与多条消息计数。
-- 保存和退出保留所有未发送全文；恢复先验证存档，旧 queued 作为警告展示且不自动重放。
-
-### Removed
-
-- Removed the inline terminal UI and its automatic downgrade rules. Interactive
-  sessions now always use the full-screen UI; small windows show a resize notice.
-- Removed `--fullscreen`, `--no-fullscreen`, `ARAGON_FULLSCREEN`, and the
-  `fullscreen` config setting. Old flags are rejected and old stored preferences
-  are ignored. Headless commands and piped input retain their existing behavior.
-- Interactive-only screens now reject non-TTY input/output with exit code 2.
-
-### Added
-
-- **`system/init` now lists what this build can do.** A new optional
-  `capabilities: string[]` field, carrying `["interrupt"]` for this release.
-
-  A wrapper cannot tell by observation whether an interrupt will settle
-  gracefully or kill the process outright, and guessing wrong costs the user
-  their conversation. Rather than have callers compare version numbers — which
-  breaks on forks, prereleases, and version shapes nobody anticipated — the
-  build now says so itself, in an event that already arrives on every start.
-  Absent on older builds, and absence should be read as "no", never as
-  "probably".
-
-  Additive, so `schemaVersion` does not move. Consumers that do not know the
-  field ignore it, exactly as the schema contract asks.
-
-- **The composer understands pasting.** Until now it did not have the concept at
-  all: every byte the terminal delivered on stdin was a keystroke, and a
-  clipboard block is thousands of bytes arriving in an unpredictable number of
-  chunks. Three separate failures fell out of that one omission, and on Windows
-  you hit all three at once.
-
-  A carriage return that landed alone in a chunk **sent the half-finished
-  message** — a fifteen-line paste could arrive as a dozen separate messages,
-  interleaved with the agent's replies. A carriage return in the middle of a
-  chunk was inserted into the draft as a literal `\r`, which the terminal obeyed
-  by **overwriting the row it was drawn on**. And a paste that survived both grew
-  the composer to the height of the pasted text, pushing the transcript out of a
-  frame whose row accounting still believed the composer was three rows tall.
-
-  Pasting now works the way it reads:
-
-  - **A paste never submits.** Line breaks become buffer newlines, exactly one
-    per source line break, `CRLF` and lone `CR` alike.
-  - **A short paste appears verbatim.** At or below 6 lines and 400 characters it
-    is inserted as-is, indentation intact, and renders exactly like typed text.
-  - **A large paste collapses to a placeholder** — `[Pasted text #1 +218 lines]`
-    — which occupies one line, carries its own size in its label, deletes as a
-    single unit (one `Backspace`, and `Ctrl+W` / `Ctrl+U` / `Ctrl+K` cannot leave
-    half of one behind), and expands back to the full text the moment you send.
-    The model receives every byte; the composer never has to draw them.
-  - **No pasted byte is executed as a terminal control.** `\r`, `\x1b`, NUL, DEL
-    and the C1 range are stripped from the payload — never from your keystrokes,
-    so `Ctrl+A` and Escape are untouched.
-  - **The composer has a ceiling.** A draft of any length occupies at most 3 / 6 /
-    10 rows depending on terminal height, scrolls internally to keep the caret
-    visible, shows how many rows are hidden above and below, and reports its true
-    height to the layout budget. Four hundred presses of `Shift+Enter` were
-    already able to destroy the frame; they no longer are.
-  - **A submitted message cannot make the transcript unresponsive.** A user entry
-    renders at most 40 rows plus a `... +N more lines` tail. The full text is kept
-    and the exit transcript is uncapped.
-
-  On terminals that support it, detection is **exact**: aragon asks for bracketed
-  paste (`DECSET 2004`) in full-screen mode and reads the markers the terminal
-  wraps your paste in. Where that is unavailable — inline mode, an old Windows
-  console — a heuristic on chunk shape catches the same cases, and it needs no
-  cooperation from the terminal at all. Both feed one code path.
-
-  Turn it off with `--no-paste`, `ARAGON_PASTE=0`, or
-  `aragon config set paste false`; the input path is then byte-for-byte what
-  0.6.2 did.
-
-### Fixed
-
-- **The right-edge scrollbar no longer disappears while output streams.** The
-  differential renderer ended every repainted row with `CSI K` (erase to end of
-  line). The scrollbar fills the last column of every viewport row, and on
-  terminals that keep the cursor on the last column after writing it (Windows
-  console hosts) that erase blanked the cell that had just been drawn, so the
-  scrollbar vanished whenever a row was redrawn. Rows that already fill the
-  terminal width are now repainted without the erase; narrower rows are
-  unchanged. `--no-diff-render` was never affected. `/perf` gained a `scrollbar:`
-  line for diagnosing visibility and drag problems.
-
-- **`{"type":"interrupt"}` now really does settle the run — on every platform.**
-  The stdin interrupt frame used to raise a SIGINT at our own process. On POSIX
-  that reaches the runner's terminator and the run winds down properly: it
-  breaks out of the loop, persists the session, and emits `result` with the
-  usage and cost. On Windows, Node does not deliver a self-directed SIGINT to
-  its own handlers — the process simply terminated with exit code 1, having run
-  none of it.
-
-  For anyone using `aragon exec` interactively that difference was invisible,
-  because an interrupt there is usually the last thing that happens. It is not
-  invisible to a **wrapper** that interrupts a turn and then respawns on the same
-  `--session-id` to keep the conversation going: on Windows it silently got back
-  a session missing everything since the previous save. `exec/index.ts` has
-  carried a comment about that caller for some time ("PERSIST FIRST, ANNOUNCE
-  SECOND … a caller that reads `result` and immediately respawns with the same
-  `--session-id` must not race the writer"), so the intent was there; the path
-  just could not be reached on Windows.
-
-  The frame now calls the runner directly instead of going through a signal.
-  POSIX behaviour is byte-for-byte unchanged — the same code runs, with the same
-  `stopReason`, the same exit code 130, and the same two-second window in which a
-  second signal still forces an immediate exit.
-
-- **A pasted API key is stored correctly again.** Ink broadcasts each stdin chunk
-  to every mounted input handler, so the settings screen, the question overlay
-  and the plan-review feedback field all see what the composer sees. Without
-  this, a ~100-character key pasted into Settings would have been stored wrapped
-  in invisible framing bytes, masked on screen by the password dots, and would
-  have failed to authenticate on every request afterwards with nothing anywhere
-  explaining why.
-
-- **`--no-mouse` still writes no mouse sequence.** Paste support installs a
-  stdin filter, and `paste` defaults to on, so the filter now exists in sessions
-  that disabled the mouse. Whether the mouse is on is a separate question from
-  whether the stream is wrapped, and the four places that used to conflate them
-  now ask the right one: `aragon --no-mouse` writes no `?1000h`, no `?1006h` and
-  no `?1002h`, leaves your terminal's own click-drag selection alone, and reports
-  `/mouse` as off — while pasting still works.
-
-### Changed
-
-- **The run status row moved above the input box and carries the `steer` hint.**
-  While a run is active, the spinner, the phrase and `steer / interrupt / exit`
-  share one row directly above the input, in place of the idle hint row below it
-  (the footer keeps its height). It scrolls with the input; when it leaves the
-  viewport the fixed bottom row shows the animation instead, so there is never
-  more than one spinner. Terminals under 20 rows and `--no-hints` keep the old
-  fixed row. The update notice is also held back for the whole run.
-
-- **Auto-compaction no longer reports compactions that did not happen.** A long
-  unattended run used to print a red `context not compacted #2: nothing_to_drop`
-  card into the transcript. The card was accurate; nobody had asked for it.
-  Auto-compaction evaluates itself at every turn boundary, and a checkpoint that
-  looks at the history and finds nothing worth dropping has changed nothing,
-  called nothing and spent nothing — a smoke alarm announcing "no fire" twice an
-  hour, in error colours, with a machine reason attached.
-
-  A compaction now announces itself **if and only if it commits to doing work**.
-  That is a position in the control flow, not a list of reason strings: everything
-  decided before the first spend is silent, and everything decided after it —
-  including every failure, the truncation fallback, an Esc, and a host fault — is
-  as loud as it ever was. A declined `pressure` checkpoint is now
-  indistinguishable from a turn on which the threshold was never crossed.
-
-  **`/compact` and overflow recovery are unchanged.** A compaction you asked for
-  is owed an answer, and a reactive compaction running after the provider has
-  already refused your request is the one moment the diagnosis matters most. Both
-  still report on every outcome.
-
-  **Quiet, not unrecorded.** Every attempt is still written to the diagnostic log
-  (its `compaction` line gains an `announced` field, so `announced: false` is one
-  grep), still charged to the anti-loop guards, and still counted by
-  `/compact status` — which now reads
-  `This session: 1 compaction (2 checkpoints declined), 94.2k tokens reclaimed`
-  when checkpoints declined, and is byte-identical to before when none did. If
-  declines keep happening, the self-disable notice still says so out loud, once,
-  naming the cause and the remedies. Card numbers are also contiguous again: `#N`
-  counts compactions that ran, not checkpoints that fired.
-
-  **Behaviour change for wrapper scripts.** A declined `pressure` compaction no
-  longer emits a `{ type: 'compaction', subtype: 'start' | 'end' }` pair on the
-  `--output-format json` stream, and no longer prints `[compaction] skipped:` on
-  `-p` stderr. The pairs stay balanced, so only consumers counting *attempts* are
-  affected; the diagnostic log reports those more accurately.
-
-### Added
-
-- **Context compaction, hardened — the trigger now counts the whole history, the
-  recent turns can no longer wedge a run, sub-agents get the same protection, and
-  a compaction is no longer irreversible.**
-
-  **The trigger was systematically late.** It fired on the provider's reported
-  usage for the PREVIOUS request, read at a checkpoint that runs AFTER the current
-  turn's tool results were appended — and tool results are the largest single
-  increment in this loop (100 KB per call, several per turn, ~25 k tokens each).
-  A session reading 88 % could be sending 110 %. Occupancy is now the measured
-  base PLUS an estimate of everything appended since that measurement, and the
-  status gauge reads the same number the trigger does, on every turn — including
-  the turns where a guard declines to compact, which is exactly when the number
-  matters most.
-
-  **The recent turns can no longer wedge a run.** Every previous rung of the
-  ladder operated on the head of the conversation. When the RETAINED tail was what
-  did not fit, each rung reported success or "nothing to drop" while the history
-  stayed un-sendable and the run died. Oversized `tool_result` bodies inside the
-  retained turns are now clipped, oldest first, only when the alternative is a
-  request the provider cannot accept. Nothing is removed — no message, no
-  tool-call id, no role — so a valid history stays valid. Every clip is announced
-  in the text the model reads, at the exact place the data went, and on the card,
-  in `/compact status`, on the JSON stream and in the archive.
-
-  **Sub-agents compact their own history** (`compaction.subagents`, default
-  `true`). A `task` child is bounded by 24 turns and fifteen minutes, and when one
-  overflowed its entire history was discarded and the dispatch reached the lead as
-  one partial sentence. Children run under tighter bounds than the lead: two
-  retained turns, at most two compactions, the fast tier preferred, and failure
-  forced to truncate because nobody is watching a background worker. No transcript
-  card — the lead's transcript describes the lead's context. Set
-  `compaction.subagents: false` to restore the previous behaviour exactly.
-
-  **A compaction is no longer irreversible** (`compaction.archive`, default
-  `true`). The dropped messages are written verbatim to `~/.aragon/compaction/`
-  before the splice is adopted. `/compact history` lists this run's archives,
-  `/compact show <n>` prints one's metadata and summary — never message bodies.
-  The directory is shared by every `aragon` on the machine, so files carry a
-  per-process run id and retention is two-tier: 20 files per run by count, plus a
-  seven-day sweep across all runs, so one session can never prune a live
-  neighbour's work. Writing is best-effort: a failure is logged and never affects
-  the compaction. This is deliberately **not** undo — restoring an over-full
-  history restores the condition that triggered the compaction.
-
-  **The long moment is legible.** The live card counts elapsed seconds and says
-  `esc to cancel`, which `AgentController.abort()` has always made true and no
-  surface has ever stated. A compaction that could only clip says so and names no
-  model, because none was called.
-
-  Two new config keys (`compaction.subagents`, `compaction.archive`), two new
-  environment variables (`ARAGON_COMPACTION_SUBAGENTS`,
-  `ARAGON_COMPACTION_ARCHIVE`), two new `/compact` verbs (`history`, `show <n>`),
-  and two new settings-screen rows. No flags were added.
-
-- **Context compaction — a long run no longer dies when the window fills, and it
-  is ON by default.** A session's conversation only grows: nothing anywhere
-  removed a message, so a 40-turn tool-heavy run had exactly one ending available
-  to it once the context window filled — an HTTP 400 and a banner reading
-  *"Context length exceeded — start a new conversation with /reset."* The
-  product's answer to running out of context was "throw away everything you were
-  doing."
-
-  Now, at **90 %** occupancy, the run pauses at a turn boundary, hands the older
-  part of the history to a summarizer, and swaps it for a structured record. Your
-  original task stays **verbatim**, the most recent turns stay **verbatim**, and
-  the run continues in the same loop with the same tools, the same todo list and
-  the same working directory.
-
-  **This changes behaviour for existing users, in the direction of not dying.**
-  `--no-compaction`, or `compaction.enabled: false`, restores the old behaviour
-  exactly: no manager is constructed, nothing is subscribed, no registry is
-  allocated, and a full window ends the run as it did before. Both flag forms are
-  declared, so a stored `false` survives a run that passes no flags.
-
-  Three ways it fires. **Proactively** on measured occupancy — the case that
-  happens almost every time. **Reactively** when the provider returns
-  `context_overflow` anyway, because the model is not in the static window table
-  or you are pointed at a proxy in front of a smaller one: one compaction, one
-  re-send, run continues. And **manually**, via `/compact`, which runs
-  immediately when idle and queues to the next turn boundary when running.
-
-  The trigger has two terms. 90 % of a 200 k window leaves 20 k, which is plenty;
-  90 % of a 32 k window leaves 3.2 k, which is less than a single `max_tokens` of
-  8192 — the request is already impossible. So compaction also fires whenever the
-  remaining headroom drops below one full response.
-
-  **The default summarizer is your own model, and that costs real money.**
-  `compaction.useFastTier` is `true`, but it is a *preference*: the fast tier is
-  off unless you configured it, so out of the box the first compaction is a
-  ~30 k-token call on the session's own model. It happens at most 5 times per
-  run, the alternative is losing the run entirely, and the spend is visible — in
-  the session total, priced with the summarizer's own table, and reported as
-  `pricing unknown` rather than `$0.00` when that table has never seen the model.
-  Configure the fast tier to make it cheap.
-
-  Summarization fails sometimes. The ladder is one retry (on your **main** model
-  if the first attempt used the fast tier), then a **truncation** that says so on
-  the card, on the JSON event stream, and inside the block the model itself
-  reads. `compaction.onFailure: "stop"` makes the other trade.
-
-  Four guards stop it looping on your money: never two on consecutive turns,
-  never more than 5 per run, a progress requirement, and a self-disable after two
-  no-progress attempts with a notice naming the cause and the fix. The guards are
-  **trigger-aware** — the reactive path stays armed even after a self-disable,
-  because a compaction answering a request the provider has already refused is
-  the difference between a live run and a dead one.
-
-  Three UI surfaces: `Compacting context…` on the activity row while it runs, a
-  `compacting` chip plus threshold-aligned gauge colours as state, and a
-  transcript card with the before/after counts, the summarizer, the duration and
-  an expandable summary as the record. The gauge falls **immediately** rather
-  than waiting for the next turn. `/compact status` is the guaranteed surface at
-  any width.
-
-  New: `compaction` config section (6 keys), `--compaction` / `--no-compaction` /
-  `--compaction-threshold`, `ARAGON_COMPACTION*` (3 vars), `/compact` (6 forms),
-  three settings-screen rows, a `compaction` event on the `aragon exec` JSON
-  stream (`EXEC_SCHEMA_VERSION` **unchanged** — adding an event type never bumps
-  it), and two stderr lines under `-p`.
-
-### Changed
-
-- **The context gauge now counts cache tokens**, through the same
-  `occupiedTokens` the compaction trigger reads. Anthropic reports `input_tokens`
-  *excluding* cached tokens and puts them in two separate fields, so the old
-  `inputTokens + outputTokens` silently under-reported occupancy for anyone
-  behind a caching gateway. That was already a defect in the number on the status
-  bar; it becomes a much worse one the moment a trigger fires on a different
-  number from the one you are watching.
-
-- **`context_overflow` now names `/compact` alongside `/reset`.** That remedy
-  list was complete until this release, because throwing the conversation away
-  really was the only exit.
-
-### Known limitation
-
-- **A summarizer that is 90 % faithful will, eventually, drop something the run
-  needed,** and no amount of prompt engineering closes that. The design bounds it
-  — verbatim task, verbatim recent turns, a block that tells the model the record
-  is partial, and three surfaces that make it visible so a human can step in —
-  but it does not eliminate it. That is inherent to compaction as a technique, it
-  is the same trade Claude Code and Codex make, and it is preferable by a wide
-  margin to the previous behaviour, which was to lose the entire run.
-
-## 0.6.0
-
-### Added
-
-- **`aragon exec` — a second, machine-facing face on the same binary.** The
-  package already installed from npm and already ran from a terminal, so "it has
-  a CLI" was true. It was nonetheless not *callable* by another program:
-  everything it exposed was shaped for a person, and a build script, a CI job, a
-  web backend or a second agent had to spawn it, capture free text, and guess.
-  There was no way to read the answer without also reading the narration, no way
-  to carry a conversation across two invocations, no way to say "you may read but
-  not write" when there is no human to answer `--confirm`, and no way to bound
-  what a run may spend.
-
-  `aragon exec` adds a **contract** rather than a feature: a versioned JSON event
-  schema on stdout, a session identifier the caller may mint itself, a tool
-  permission policy enforced at the tool boundary, caller-set budgets with their
-  own exit code, and a discovery command so a wrapper can ask what it just
-  installed before depending on it. Everything underneath is reused unchanged —
-  the same controller, the same tools, the same config resolution, the same
-  skills, teams, todo and fast-tier subsystems.
-
-  - `--output-format text | json | stream-json`. `stream-json` is NDJSON: one
-    object per line, `system/init` first, exactly one `result` last, on every
-    exit path including error, budget and signal. `text` is byte-identical to
-    `aragon -p`, because it **delegates** to the same renderer rather than
-    reimplementing it.
-  - `--session-id` / `--resume` / `--continue`, backed by `~/.aragon-agent/
-    sessions/` and an exclusive lock, so two runs cannot interleave and lose a
-    turn. A second run on a held id exits 2 with `session_busy` rather than
-    blocking.
-  - `--permission-mode auto|plan|strict` plus `--allow-tool` / `--deny-tool`.
-    A denied tool is **not registered at all**, so the model never sees it and
-    never plans around it — and the policy reaches subagents, so one `task` call
-    cannot bypass it.
-  - `--max-turns` / `--max-duration`, both aborting through the same path `Esc`
-    uses, both reporting `isError: false` and exit **3**: a budget firing is not
-    a malfunction, but it is not a complete answer either, and a shell script has
-    to be able to tell without parsing JSON.
-  - `--append-system-prompt` / `--append-system-prompt-file`. Appending only;
-    there is deliberately no full `--system-prompt` replacement, because the
-    builtin prompt carries tool discipline, todo planning, the skills catalog and
-    the plan-mode block, and replacing it silently disables half the product.
-  - `--input-format stream-json`, so one process can serve several turns from a
-    pipe. A malformed line is skipped with a non-fatal `error` event; an unknown
-    message type is ignored.
-  - **`aragon sessions`** (`list` / `show` / `rm` / `prune` / `path`),
-    **`aragon info [--json]`** and **`aragon doctor [--json] [--probe]`**.
-
-  Documented in full under
-  [Use it from another project](./README.md#use-it-from-another-project),
-  including an honest Limits section: there is no filesystem sandbox, process
-  exit is terminal on Windows whether or not a `result` arrived, an exec session
-  resumed in the TUI shows no scrollback, and `sessions prune` shares a directory
-  with the TUI's `/save`.
-
-### Changed
-
-- **Exit code `3` is new, and the signal row is now stated as `128 + signo`.**
-  `3` means "stopped by a budget you set" and is reachable only from
-  `aragon exec`. The signal codes are unchanged in behaviour — this CLI has
-  always exited `128 + signo`, so `SIGTERM` has always been `143` — but the
-  README previously simplified all of them to `130`, which is fine for prose
-  about a CLI a human interrupts and not good enough for a wire contract.
-- `interrupt` handling now settles the run before exiting when `aragon exec`
-  owns it, so a wrapper reading NDJSON gets a final `result` on Ctrl-C instead of
-  having to time out to learn the run ended. `aragon`, `aragon -p` and every
-  other subcommand keep the previous bare-exit behaviour byte for byte.
-
-### Unchanged, and asserted rather than claimed
-
-- `aragon`, `aragon -p`, every existing flag, every slash command and every byte
-  of their output. `aragon exec --output-format text` and `runHeadless` are
-  compared for byte equality on **both** stdout and stderr by a regression test,
-  and `createBuiltinTools` with no permission policy is proven by object identity
-  to return the tools it returned before.
-
-## Unreleased
-
-### Project index on Ctrl+I (bundled `project-indexer` skill)
-
-- The `project-indexer` skill now ships **bundled** with the package
-  (`<pkg>/skills/project-indexer`, read-only, alongside `skill-creator`): it
-  generates the `.claude-index/` navigation map and injects the Google-style
-  Clean Code Guidelines section into `CLAUDE.md`, adapted from the standalone
-  Claude Code skill (ask_user instead of AskUserQuestion; body condensed under
-  the 30 KB invocation budget with the four templates bundled as files).
-- New keybinding `Ctrl+I`: opens a Yes/No confirm that names every write
-  target, then dispatches `/skill:project-indexer` through the ordinary
-  command path - the indexing run is a normal agent turn whose tool calls
-  stream into the transcript. While a run is in flight the key warns instead
-  of queueing; with skills disabled or the skill shadowed it explains instead
-  of failing silently.
-- `Ctrl+I` is legacy-encoded as the Tab byte, so the stdin filter's two
-  translators (win32-input-mode on Windows, kitty CSI-u / modifyOtherKeys
-  elsewhere) now emit a distinct NUL-delimited `INDEX_KEY_FRAME` for it -
-  the same framing technique `Shift+Enter` uses. On terminals with neither
-  protocol the key keeps its legacy Tab meaning (completion).
-- The header names the key next to the detail toggle: `^I Index` collapsed,
-  `^I build index` in the `Ctrl+G` expanded state; hidden below 72 columns and
-  while an overlay is open. `/skill:project-indexer` and `/project-indexer`
-  remain available as typed commands.
-
-### Added
-
-- **Drag the mouse to select text, and release to copy it.** Full-screen mode
-  takes the alternate screen, turns on mouse reporting so the wheel can scroll a
-  transcript the terminal no longer knows how to scroll, and draws every row
-  itself. Each of those is right on its own; together they took your terminal's
-  own drag-to-select away and put nothing in its place. The only thing shipped
-  in its place was a one-line notice naming a bypass key that is Shift on some
-  emulators, Fn on others and a preference elsewhere — a workaround, told once,
-  in a line that scrolls away.
-
-  Press the left button and drag: the cells you drag across are highlighted as
-  you go, and releasing copies exactly them. The copy goes out over **OSC 52**
-  first, so it reaches the clipboard of the machine you are *sitting at* rather
-  than the one you are ssh'd into, with the platform clipboard binary attempted
-  alongside it for terminals that ignore the sequence. Neither mechanism is
-  detectable, so the toast names what was sent rather than claiming a success it
-  cannot verify. `/copy` now goes through the same path and gained ssh support on
-  the way.
-
-  The highlight is painted in the frame pipeline rather than in React, so a drag
-  costs one to three re-addressed rows instead of a re-render of the whole tree,
-  and the text that is copied is read from the same mirrored rows the highlight
-  was painted over — what you see is what you get by construction, not by a
-  staleness check that can be wrong.
-
-  Two switches, both reversible without a restart: `/mouse off` hands the mouse
-  back to your terminal for the session (and says plainly that the wheel goes
-  inert while it is off), `/mouse on` takes it back, and
-  `aragon config set mouseSelect false` keeps wheel scrolling while leaving
-  drag-select off — which also stops the CLI asking the terminal for motion
-  reports at all, so that session puts exactly the bytes on the wire that it did
-  before this release. `--no-mouse` remains the switch that captures nothing.
-
-### Changed
-
-- **Scrolling up now actually stops the transcript.** The viewport followed the
-  newest output by holding a *fixed distance from the bottom*, so scrolling up
-  during a run did not pause anything: twelve new rows of output moved the
-  window twelve rows down, and whatever you had scrolled up to read slid off the
-  top while you were reading it. The faster the agent talked, the faster your
-  reading position ran away.
-
-  While you are scrolled up, the transcript is now anchored to the **content**:
-  new output accumulates below the window instead of dragging the window along
-  with it, and the rows on screen hold still. After five seconds with no
-  scrolling the viewport returns to the newest line by itself — but only if
-  something actually arrived while you were paused, so reading a *finished*
-  transcript is never interrupted by a jump to a bottom that has not moved.
-  Every gesture restarts the countdown, a selection drag suspends it, and
-  `aragon config set scrollResumeMs 0` turns the return off entirely
-  (`ARAGON_SCROLL_RESUME_MS` sets it too).
-
-  The number the hint shows means something sharper as a result: it is now
-  literally "rows of newer output below your reading position", which is what
-  `PgDn` traverses and what the status bar's `↑N` has always shown.
-
-- **The "N new lines · PgDn" hint moved inside the input box.** It used to be
-  rendered as the viewport's own last row, which spent a line of your transcript
-  on chrome at exactly the moment the transcript was longest — and, because the
-  hint's presence shrank the viewport, it fed back into the very number it was
-  displaying. It is now a right-aligned chip on the composer's input row, inside
-  the border, where it costs no rows at all. The chip sits in a fixed-width cell
-  so the number gaining a digit cannot re-wrap a draft, it never truncates (the
-  draft wraps first), and unlike the teaching row below the box it neither fades
-  with experience nor disappears on a short terminal: how far behind you are is
-  state, not a tutorial.
-
-### Fixed
-
-- **The working spinner no longer disappears when an ack takes its row.** Steer a
-  run, approve a plan, or press `Ctrl+T` while the agent is working, and the row
-  above the composer traded its spinner for the acknowledgement — and for that
-  ack's full two and a half seconds, nothing anywhere on screen moved. The status
-  bar still read `running` and the tool card still read `running`, so no
-  information had left the frame; only the evidence that the process was alive
-  had, during exactly the long tool calls where "is it stuck?" is the question you
-  are already asking. Consecutive acks extended the window, and steering is
-  something the composer invites on every run.
-
-  That row was made the sole owner of animation in this same release, and
-  everything else on screen holds still for as long as the row is believed to be
-  showing one. Believed, not checked: an ack won the row outright, so the belief
-  and the frame disagreed, and the disagreement was invisible to every test
-  because no component was wrong on its own.
-
-  The ack still owns the row and keeps every word of its text. The spinner now
-  rides beside it as a bare glyph — no phrase, no clock — so there is exactly one
-  animation at every instant of a run, with an ack up or without one: never none,
-  and never the several this release set out to remove. Terminals without Unicode
-  and anyone running with `reducedMotion` come out byte for byte unchanged; they
-  had no animation to lose, and nothing is prefixed to their acks.
-
-- **`/clear` now takes the todo panel with it.** Finish a plan, type `/clear`,
-  and the transcript emptied while the `7/7 done` rail stayed exactly where it
-  was — on a screen you had just asked to be cleared. It stayed until your next
-  real message, which is precisely the pause you take after clearing to decide
-  what to ask next, so essentially everyone met it.
-
-  This was deliberate once. The list is a projection of what the model believes,
-  and `/clear` wipes the transcript rather than the conversation, so the model
-  still believed in the plan and a panel contradicting the model was judged worse
-  than no panel. That reasoning holds for everything the CLI does on its own; it
-  does not hold against you typing the command. An explicit instruction was
-  always this rule's exception — `/todo clear` has used the same door from the
-  start — and `/clear` now goes through it.
-
-  The list itself is cleared, not just the pixels, so `/save`, `/todo status`,
-  `/todo continue` and auto follow-through all agree with the screen: no plan is
-  written into a saved session, nothing offers to continue one, and nothing spends
-  a follow-through on it. `/reset` is unchanged and still the only command that
-  clears the conversation. Nothing is stranded either way — the model's next
-  `todo_write` is a full replacement, so the panel comes back the moment there is
-  a plan to show.
-
-### Added
-
-- **An update that will not start is undone automatically.** The updater shipped
-  in this same release could install a version that fails to *import* — an
-  undeclared dependency, a Node API that moved, a syntax error past the transpile
-  target — and the only recovery was to know the package name and type
-  `npm i -g @aragon-agent/cli@<some older version>` from memory. None of those
-  failures reach any line inside `main()`, so no amount of care there helps.
-
-  The `aragon` command is now a small launcher that runs **before** the CLI. It
-  reads one ~300-byte file, and on every launch except the first one after an
-  auto-update that is all it does — no write at all. But if the version it is
-  about to start is one **this updater installed**, and that version has already
-  exited non-zero twice, it reinstalls the version that was running before the
-  update and marks the bad one so nothing puts it back:
-
-  ```
-  aragon 0.6.0 failed to start twice; rolling back to 0.5.9.
-  ```
-
-  It boots anyway. If the crash was transient, refusing to start would lock you
-  out of a working CLI on our guess; the reinstall is detached and finishes
-  either way. The next session says once what happened and then stops. A **newer**
-  release clears the mark and installs normally — you were never opted out of
-  updates, only out of the loop.
-
-  It never touches a version **you** installed yourself, and it does not try to
-  detect a release that starts and is merely wrong. `aragon update --rollback` is
-  the same mechanism run deliberately, for the case a human spots in one second
-  and a heuristic would get wrong on a bad day. Exit `4` means there was nothing
-  recorded to roll back to.
-
-  **`update.mode: off` disables this too.** The guard is armed only by an
-  auto-install, so a machine that never auto-installs never arms it — the kill
-  switch turns off the whole subsystem, recovery included.
-
-- **A blocked install now says what is actually blocking it.** Every install
-  failure used to print the same line: `update failed · npm i -g
-  @aragon-agent/cli`. On Windows the commonest cause is that the shim which
-  launched us is held open by the shell running it, so that advice hits the same
-  handle and fails identically — the one visible failure was also a wrong
-  instruction. Failures are now classified, and three of them name the real
-  remedy and appear immediately rather than after three strikes, because each is
-  something you can fix and none of them fixes itself:
-
-  ```
-    ! update blocked · close other aragon windows, then: npm i -g @aragon-agent/cli
-    ! update failed · no disk space
-    ! update needs write access · npm i -g @aragon-agent/cli
-  ```
-
-  Network failures are deliberately **not** in that set: you cannot fix the
-  registry and the next check may well succeed, so they stay silent and back off.
-
-- **Updates work behind a corporate proxy.** Node's `fetch` ignores `HTTP_PROXY`,
-  so on a proxied machine the check failed silently forever. When the direct
-  request fails, `aragon` now asks the npm you already have instead, which
-  honours `proxy`, `https-proxy` and your `.npmrc`. It runs *only* after a
-  request that has already failed, so on an ordinary machine it never runs.
-  `/update status` shows `probe: npm` when this is what answered. No credentials
-  are read or forwarded — we run the command you could have typed.
-
-- `/update status` gained `rollback`, `lastgood` and `probe` lines, so what the
-  guard did and what `--rollback` would reinstall are answerable without reading
-  the log.
-
-### Changed
-
-- **`aragon update` exit codes: a blocked install is now `2`, not `1`.** This is
-  the one change here a script can notice. `1` means "try again later" and `2`
-  means "this machine will never auto-update, run the command we printed" — and
-  for `EPERM` / `EBUSY` / `ENOSPC` the second is the accurate one: waiting does
-  not unblock a held file handle or an empty disk. A caller shaped like
-  `aragon update || sleep 300 && retry` changes behaviour. `0`, and `1` for an
-  ordinary failed check or install, are unchanged.
-
-- `bin.aragon` now points at `dist/launcher.js` rather than `dist/cli.js`.
-  `dist/cli.js` is unchanged, keeps its shebang and stays executable, so
-  `npm start` and any script invoking it directly keep working — only the mapping
-  moved.
-
-- **`aragon` keeps itself up to date.** Until now the installed copy was frozen
-  forever: nothing in the CLI ever looked at the registry, so someone who
-  installed 0.4.x eighteen releases ago was still running it and the only signal
-  that anything had moved was a README they were never going to re-read.
-
-  There is now a background updater. At most once every four hours, and at most
-  once per machine per interval, it asks npm what the `latest` dist-tag resolves
-  to; if that version is genuinely newer, stable, not deprecated and runnable on
-  this Node, it installs it silently into the same global installation you
-  already have and says so on **one line above the composer**:
-
-  ```
-    * 0.6.0 installed - restart aragon to apply
-  ```
-
-  **The running process is never mutated in any way it can observe.** The new
-  version lands on disk and takes effect on the *next* launch. A restart in the
-  middle of a turn would destroy a live run, a subagent dispatch and an open plan
-  review, so the line waits and the human decides when.
-
-  The notice does not get its own row — it shares the row the toast stack and the
-  working line already share, behind both of them. A toast is a response to
-  something you did and wins outright; the working line is transient and would be
-  lost if deferred; the update line is persistent and loses nothing by waiting.
-  The practical effect is that **someone who is actually working never sees the
-  updater until they stop.**
-
-  It only ever runs `npm i -g` for one kind of installation: a global one where
-  no project manifest claims to own the version. pnpm, yarn, bun and volta are
-  each detected and get the correct command for *that* manager printed instead
-  of a command run on their behalf; `npx`, a project's `node_modules` and a
-  clone of this repo produce no line at all, because none of them is out of date
-  in any sense the user can act on. A global directory that cannot be written is
-  detected before anything is spawned, and npm exiting `0` into a *different*
-  prefix — which happens on any machine with `prefix=` in `.npmrc` or two Node
-  installations — is caught by re-reading the version from the directory we
-  classified, rather than being believed.
-
-  Every failure is silent: it logs, backs off exponentially, and says nothing
-  until the third consecutive one, at which point it says something actionable.
-  Four config keys (`update.mode`, `update.checkIntervalMs`, `update.registry`,
-  `update.distTag`), the flags `--update` / `--no-update`, the environment
-  variables `ARAGON_UPDATE` and `ARAGON_UPDATE_REGISTRY`, the slash command
-  `/update [status|now|skip|off]`, and a foreground `aragon update
-  [--check] [--to <version>] [--json]` for headless users. `update.mode: off`
-  and `ARAGON_UPDATE=0` are kill switches, and with either set the CLI executes
-  zero lines of update code.
-
-  Auto-update widens the window in which a compromised release reaches users
-  without a human decision. The README says so plainly, along with what is and
-  is not mitigated. No credentials are ever attached to the check.
-
-- **A running command shows what it is doing.** `bash` is the only tool here that
-  can run for minutes — `npm test`, a `git clone`, a training job — and while it
-  ran, its card was one line of text that never changed. Nothing in that frame
-  distinguished a test suite at 40% from a child process blocked forever on a
-  password prompt this CLI will never show. Its card now carries the last eight
-  lines of what the command is printing, plus one footer row that reads
-  `(running)` or, once the child has been quiet for ten seconds,
-  `no output for 45s` — with the seconds advancing.
-
-  A `\r`-driven progress bar (npm, pip, curl, docker) is *one* line rewritten a
-  thousand times, and it renders as **one row counting up in place** rather than
-  a thousand rows of nearly the same string — including the ordinary case where
-  the bar emits no newline at all for its entire run.
-
-  Everything the tail shows is **stripped, not interpreted**: ANSI colour, cursor
-  motion, `\r`, `\b`, tabs and every other control byte are removed at the
-  boundary before a row is stored, so nothing can move a cursor inside a frame
-  the renderer believes it owns. It is bounded twice by construction — eight rows
-  per call, sixteen concurrent calls — so a command that writes 400 MB to stdout
-  costs a fixed number of retained rows.
-
-  It is a **display and nothing else**. The text the model receives is
-  byte-for-byte what it was; `Ctrl+O` does nothing while a command runs, because
-  expansion is a promise about stored content and there is no stored rest of a
-  400 MB stream; `/save` during a run writes no tail, since a resumed one would
-  describe a process that died with the session; and at settle the card is
-  replaced by the authoritative result exactly as before.
-
-  New config key `liveToolOutput`, defaulting to **`true`** — unlike
-  `showThinking` below, this adds information rather than hiding it.
-  `aragon config set liveToolOutput false` (or `--no-live-tool-output`, or
-  `ARAGON_LIVE_TOOL_OUTPUT=0`, or the **Live output** row in `/settings`) turns it
-  off, and with it off no store is allocated and no recorder is attached: `bash`
-  is byte-identical to the build before this entry.
-
-- **The working line stops guessing while a tool runs.** It reads
-  `⠙ Running bash` instead of rotating thinking-words. `running` covers the whole
-  turn, including the minutes the model is idle and a child process is doing the
-  work, so "Pondering" during a build was not calm — it was wrong. One row, no
-  digits, no clock: everything the row already promised still holds.
-
-### Changed
-
-- **Reasoning blocks are now hidden by default.** `showThinking` is a new
-  top-level config key and it defaults to `false`. If you have been reading
-  thinking every day, this is the entry that concerns you: `Ctrl+T` still shows
-  it for the session exactly as before, `aragon config set showThinking true`
-  (or `--show-thinking`, or `ARAGON_SHOW_THINKING=1`) restores the old behaviour
-  permanently, and `/settings` has a **Show thinking** row directly under
-  **Thinking**.
-
-  Nothing is hidden silently. A settled turn that thought leaves one muted row
-  where the block used to be — `✱ thought for 12s · ctrl+t to show` — so the fact
-  that reasoning happened, how long it took, and how to see it are all still on
-  screen. In **inline** mode that row omits the key hint and the toast reads
-  `Thinking shown for new output.`, because a settled entry there has already
-  been printed into your terminal's scrollback and cannot be repainted; the
-  toggle applies to the live entry and to everything drawn after it.
-
-  `thinkingLevel` is untouched and means what it always did — the effort the
-  provider is asked to spend. `showThinking` is only whether the terminal draws
-  what came back. They are one word apart in the config file, which is why the
-  flag is `--show-thinking` rather than the shorter name.
-
-### Added
-
-- **A live working line while a run is in flight.** One row above the composer,
-  carrying a spinner and a word that rotates every four seconds:
-  `⠋ Percolating…`. Take reasoning off the screen and the gap between submitting
-  and the first token is otherwise completely blank, which reads as dead rather
-  than as busy — the two changes are one change.
-
-  It carries no clock and no token count, deliberately: the status bar one row
-  below already renders both, under the same running condition and driven by the
-  same 200 ms ticker this row reuses. Under `reducedMotion` the spinner is a
-  static glyph and the word is chosen once and holds, because rotating text is
-  motion too. An idle frame is unchanged, and so is the viewport's row budget —
-  the line shares the row the toast strip already owned rather than adding one.
-
-- **`write_file` and `edit_file` render a real diff.** Hunks with old/new line
-  numbers, a `+N -M` summary, add/remove/context colouring, and `(truncated)` or
-  a `binary` / `too-large` / `unreadable` note when the old side could not be
-  read. Diffs collapse at 12 rows behind the existing `Ctrl+O` — no new
-  keybinding. A resumed session written by an older build, or a card whose patch
-  the bounded store already evicted, falls back to the previous text preview;
-  nothing renders blank.
-
-  `write_file` previously reported creating a 300-line file as
-  `Wrote 8214 bytes to path`, and `edit_file`'s diff trimmed the common prefix
-  and suffix and rendered *everything* in between as one removed block followed
-  by one added block — for a `replace_all` touching line 40 and line 900, that
-  was 860 removed rows followed by 860 nearly identical added ones.
-
-  **The model's input got smaller, not larger.** The structured patch reaches the
-  screen on a CLI-local channel rather than through the tool result, and the text
-  the model reads gained a 4 000-character ceiling it did not have before (its
-  only previous bound was the executor's 100 KB backstop). A single-region edit
-  is byte-for-byte what it was plus one `@@` header; a multi-region edit is
-  strictly smaller. `write_file`'s result text is unchanged.
-
-### Changed
-
-- **A stream that stops mid-answer is now a loud failure instead of a silent
-  truncated one.** All three provider adapters used to end their SSE loop by
-  yielding "done" with whatever had arrived, so a connection closed cleanly
-  half-way through an answer — a proxy idle cut, a load-balancer timeout, an
-  HTTP/2 GOAWAY — was indistinguishable from a completed reply. You got a
-  truncated answer and were told nothing.
-
-  Each adapter now reports the truncation, which is what lets it be retried and
-  recovered. An **abort** is explicitly excluded and still completes normally, and
-  a model that legitimately says nothing and stops is still a success.
-
-  This is deliberately **not** behind `--no-retry`: an adapter that reports a
-  truncated stream honestly should do so whether or not retry is enabled. A switch
-  that also disabled telling you the truth would be a mute button, not a kill
-  switch.
-
-- **The mouse wheel now only ever scrolls.** A notch scrolls the transcript
-  three rows (a page with `Shift`), or the open overlay when there is one, no
-  matter where the pointer is sitting. Prompt history is recalled by `↑` / `↓`
-  alone, exactly as in Claude Code and Codex.
-
-  **This supersedes the 0.5.0 entry below**, which said the wheel "scrolls what
-  is under the pointer" and stepped prompt history one entry per notch when the
-  pointer was over the input box. That turned out to be wrong in use rather than
-  in principle: the pointer is not an instrument you aim in a TUI — it rests
-  wherever you last left it, which in a chat-shaped app is next to the thing you
-  type into. So the ordinary "scroll back and see what just happened" gesture
-  landed on the composer and replaced the draft with an old prompt, with no
-  error, no undo and nothing visibly connecting the gesture to the outcome. The
-  band that decided this was invisible too: its boundary was a measured row that
-  moved as the toast row, the hint row and the completion popup came and went, so
-  the same gesture could do two different things a second apart.
-
-  Nothing else about the wheel changes — same three rows per notch, same
-  `Shift`+wheel page, same `--no-mouse` / `ARAGON_MOUSE=0` / `mouse: false`
-  escape hatches, same `Shift`+drag for selecting text.
-
-- **A long session stays as fast as a short one.** The TUI used to get slower as
-  a session grew and, past a threshold, stopped making progress at all. That was
-  structural rather than incidental: Ink has no output caching, so every frame
-  re-walked the entire mounted tree, re-measured every text node and
-  re-serialised the frame. The per-frame cost was therefore proportional to every
-  character the session had ever produced — and `transcriptWindow` bounded the
-  entry *count*, which bounds nothing when one entry can be fifty thousand lines.
-
-  Entries outside the visible band are no longer mounted. The viewport renders
-  what you can see plus a couple either side and replaces the rest with two
-  spacers of the right height, so the frame costs what the screen costs.
-  Scrolling, the `↑N` readout and the wheel routing are untouched. Because the
-  horizon is no longer a rendering budget, `transcriptWindow` rises from 300 to
-  1000 and its ceiling from 2000 to 20000; an existing setting is still valid
-  and still clamps.
-
-  Syntax highlighting and markdown parsing are memoised on their input, so a
-  200-line code block is tokenised once rather than thirty times a second for as
-  long as it stays on screen. Only the *parsed* form is cached and never a themed
-  one, so `/theme` still takes effect on the next frame.
-
-  Under genuine load the streaming interval widens — 33 ms up to a 320 ms
-  ceiling — so keystrokes and `Ctrl+C` do not queue behind renders. This is
-  visible rather than silent: a muted **`eco`** chip appears in the status bar
-  while it is in effect, because a user who cannot see why the stream got
-  chunkier will conclude the model got slower. `--no-render-governor`,
-  `renderGovernor: false` or `ARAGON_RENDER_GOVERNOR=0` turns it off;
-  `--max-render-interval 33` flattens the ladder without disabling the mechanism.
-
-- **A live thinking block no longer grows without limit.** While an answer is
-  streaming, visible thinking is clamped to its newest 24 rows behind a
-  `... N earlier lines` marker, and the whole block renders again the moment the
-  answer settles. Thinking is scratch reasoning you read from its newest end, and
-  an unbounded one was the easiest way for a single live entry to fill the
-  viewport on its own.
-
-### Fixed
-
-- **Inline mode could hard-freeze the terminal, and now cannot.** Once a live
-  entry grew taller than the terminal, Ink took a branch that writes
-  `clearTerminal` plus the *entire accumulated session history* on every
-  subsequent frame — bypassing both of its own dedupe gates, and growing without
-  bound for the rest of the session. Any inline user (`--no-fullscreen`,
-  `TERM=dumb`, CI, a terminal under 12 rows or 40 columns) whose model wrote a
-  long answer could reach it. The live region is now clamped below that
-  threshold, with a `... N earlier lines` marker naming what is deferred. Nothing
-  is lost: the entry prints in full into the terminal's own scrollback the moment
-  it settles.
-
-- **The view state no longer grows without a ceiling.** A single assistant answer
-  is capped at 256 KiB in the view (thinking at 64 KiB, streamed tool arguments
-  at 16 KiB) by eliding the MIDDLE, so both the opening and the live cursor
-  survive; a marker states how many characters went. `transcriptRetain`
-  (default 1000) caps the entries kept in memory, and what it removed is counted
-  and reported both by `/perf` and at the top of the exit replay. A model that
-  emits five megabytes no longer makes every later frame a five-megabyte scan.
-
-### Added
-
-- **A fast model tier: a second, cheaper model you can delegate to and be
-  reviewed by.** Off by default — with `fast.enabled: false` the system prompt,
-  the `task` schema, the transcript and the request payloads are byte-identical
-  to a build without it — and it does two things once you turn it on.
-
-  **Delegation.** `task` gains one optional field per subagent, `model: "fast"`.
-  A child marked that way runs on the fast model and is otherwise identical:
-  same tools, same working directory, same `--confirm` gate, same plan-mode
-  gate, same skills ceiling. Delegating to a cheaper model is a *model* choice,
-  never a permission boundary. It is where "not complicated, but very expensive
-  in context" gets paid for — a child that reads nine files and reports four
-  sentences moves the file bodies out of the lead's window and onto a model that
-  costs a fraction as much per token.
-
-  **Periodic review.** Every few turns a small digest of what the lead has been
-  doing goes to the fast model with one question — is this still on track? — and
-  a critique that is not "on track" is injected into the running loop as a
-  `<fast_review>` block, so the lead sees a second opinion while it can still
-  act on it:
-
-  ```
-    fast review #2 · claude-haiku-4-5 · turn 12 · 1.4s
-    You have edited config/schema.ts three times without running the tests.
-  ```
-
-  The digest is built from a bounded ring of recent turns and **never** from the
-  message history: sending a session's accumulated file bodies to the "cheap"
-  model would make a review cost more than the turn it reviews, which is the
-  inversion the whole feature exists to prevent.
-
-  Spend is attributed **per tier** — a fast child's tokens are priced at the
-  fast model's table, not the lead's, and the dispatch report states the two
-  separately. A model the built-in price table has never seen is reported as
-  `unknown` rather than as `$0.00`, because a feature that looks free while it
-  is spending money is worse than one that admits it does not know.
-
-  Turn it on with `/fast model <id>` then `/fast on`, with
-  `aragon config set fast.enabled true`, with `--fast --fast-model <id>`, or with
-  `ARAGON_FAST=1 ARAGON_FAST_MODEL=<id>`. `/fast status` is the guaranteed
-  readout: the resolved model, the budget, and this session's reviews, delegated
-  children, tokens and cost. Eleven `fast.*` config keys, four flags, four
-  environment variables and five settings-screen rows; see the README.
-
-  **The review has a session budget.** A session is not one run — it is one run
-  per message you send — so a cadence bounds the pace and nothing bounds the
-  total. `fast.reviewMaxPerSession` (default **40**, raisable live with
-  `/fast budget <n>`) is the number of reviews a session may *start*; reaching it
-  prints one `info` notice that names the remedy and says delegation is
-  unaffected, and `/fast status` reads `12/40` from the first review rather than a
-  bare count you can only size after the fact. It is denominated in **reviews,
-  not currency**, because the fast tier is exactly where a model the price table
-  has never seen lives — a dollar ceiling would silently never fire for the
-  configuration that most needs one. Delegation is never budgeted: it is spend the
-  lead chose in service of a message you sent, it is already bounded by `team.*`,
-  and it exists to *reduce* total cost.
-
-  **The review also fails fast, so it cannot compete with you.** It calls through
-  its own provider registry with **one** retry rather than the ten
-  `retry.maxRetries` defaults to for the lead, and it observes neither `retry.*`
-  nor `/retry` — that policy says how hard to fight for *your* answer, and
-  inheriting it for an unrequested background call is how an advisory ends up
-  spending a rate-limited provider's quota on itself. Its own wall-clock bound
-  is 20 s.
-
-  Two failure modes are handled out loud rather than silently. If the tier stops
-  resolving mid-session — you switch the main provider and `fast.provider` was
-  inheriting it — `model:"fast"` is downgraded to the main model, **counted, and
-  stated in the report**, because an agent that silently gets an expensive child
-  has no way to learn its cost model is wrong. And three consecutive
-  *misconfiguration* failures disable the reviewer for the session with exactly
-  one warning, while a merely busy provider (rate-limited or overloaded) is
-  treated as transient and never counts — that condition fixes itself, and a
-  session-long silent shutdown over it would not. A review you *cancelled* is
-  neither: Esc, a run that ends first, or quitting leaves a `cancelled` card and
-  no strike against the tier.
-
-  Reviews log under their own `fast` scope rather than under `agent`, so
-  `fast_review_*` records filter cleanly away from the lead's.
-
-- **API calls are retried with backoff, and the waiting is legible.** A transient
-  provider failure — HTTP 429, 529, 5xx, 408, or a dropped connection — used to end
-  the turn. One red notice, an idle prompt, and forty tool calls of work thrown
-  away. It is now retried up to **10 times** with an increasing interval
-  (1 s → 2 → 4 → 8 → 16 → 30 s, equal-jittered, 30 s per wait at most) under a
-  4-minute budget measured from the first failure.
-
-  What you see while it waits is **one line, rewritten in place** — never ten
-  notices — carrying what failed, which attempt this is, and a countdown to the
-  next one:
-
-  ```
-    ↻  Provider overloaded · retry 3/10 in 7s          Esc to cancel
-    ↻  recovered after 3 retries · 12.4s
-  ```
-
-  plus a `retry 3/10 7s` chip on the status bar (`[r3]` on a narrow terminal). The
-  card settles into the transcript as history — `recovered`, `gave up`, or
-  `interrupted` — and shows up in the exit replay, because a turn that spent three
-  minutes retrying is exactly the kind of cost an exported transcript should
-  account for.
-
-  **Esc ends the wait within a frame**, not at the end of the countdown, and it
-  never produces a fabricated error: an aborted request is silent. That is the one
-  detail everything else here depends on, because the provider layer marks an
-  `AbortError` as retryable — so the signal is checked *before* the error, and a
-  user pressing Esc can never be mistaken for a timeout worth retrying.
-
-  Never retried: `auth_error` (401/403), `invalid_request`, `context_overflow`.
-  Ten retries of a bad API key is three minutes of looking broken.
-
-  `Retry-After` is honoured as a **floor**, never a replacement — a `Retry-After: 1`
-  on the ninth retry does not undo eight retries of backoff — and a stated wait
-  longer than a minute is surfaced immediately with the provider's own number
-  rather than silently wedging the terminal.
-
-  A connection that dies **mid-answer** is also covered, and this is the case that
-  needed more than a retry loop: the partial text is discarded, the affected tool
-  cards are removed, and the answer restarts from the top. Nothing is duplicated,
-  because the engine only writes history on a completed turn.
-
-- **`/retry [show|on|off|max <n>]`, `--retry` / `--no-retry`, `--retry-max <n>`,
-  `ARAGON_RETRY`, `ARAGON_RETRY_MAX`, and a `retry` config section** with nine
-  keys (`aragon config set retry.*`). `/retry off` takes effect **in the running
-  session** — subagents included, since they share the lead's provider registry —
-  not only at the next launch. `retry.maxRetries: 0` is the kill switch and is a
-  real value at every one of those entry points. `/settings` gains one read-only
-  `API retries` row.
-
-- **A subagent's retry is visible on the team card.** A child in backoff reports
-  `retry 3/10` on its own row. Children inherit the retry policy for free, but
-  their events never reach the lead's view, so without this a subagent spending
-  three minutes waiting looked exactly like a hung dispatch.
-
-- **Retry lines in `-p` mode go to stderr, never stdout.** `[retry 3/10]
-  overloaded — waiting 7s` and `[retry] recovered after 3 retries` are on stderr,
-  so `aragon -p … | jq` stays valid; a discarded partial answer is announced there
-  too, because a piped consumer otherwise cannot tell a replay from a model that
-  repeated itself. They are not suppressed by `--quiet`: a three-minute pause with
-  no explanation is not a compact run, it is a hang.
-
-- **`/perf`.** Reports what the renderer is actually doing: the frame interval
-  and the last commit cost, how many entries exist / are retained / were dropped
-  / are mounted, height-cache and render-cache occupancy, and the viewport
-  geometry. `/perf reset` clears the caches and returns the governor to its
-  fastest rung. Silent degradation is the one thing a claim of robustness cannot
-  afford, and this is the place that names it.
-
-- **Three configuration keys**, each with a flag and an environment override:
-  `transcriptRetain` (`--transcript-retain`, `ARAGON_TRANSCRIPT_RETAIN`),
-  `renderGovernor` (`--render-governor` / `--no-render-governor`,
-  `ARAGON_RENDER_GOVERNOR`) and `maxRenderIntervalMs`
-  (`--max-render-interval`, `ARAGON_MAX_RENDER_INTERVAL_MS`). No migration and no
-  config version bump: a file written before they existed loads unchanged.
-  `transcriptRetain` is raised to `transcriptWindow` when the two conflict —
-  retaining less than you can scroll to would make part of the horizon
-  unreachable — and a startup notice names both numbers, because overriding a
-  value you typed and saying nothing is the one thing none of this may do.
-
-- **Following the plan through, on a budget you can do arithmetic on.** A run
-  that ended with steps outstanding used to print one sentence and stop. It now
-  asks three questions the CLI already knew the answers to — did you abort it,
-  did it error, did the list move — and does one of three things.
-
-  With the default `notify` the sentence is unchanged, except that a run **you**
-  cancelled with `Esc` now says nothing at all: being told what you just
-  interrupted is the CLI answering a question nobody asked.
-
-  With `/todo follow auto` it continues the plan itself, after announcing
-  `Continuing with 4 remaining steps in 3s - Esc to stop.` Auto-continuation was
-  refused last release as "an unbounded cost loop wearing a helpful hat", and
-  that judgement stands — what changed is that the loop is now bounded first.
-  A continuation that completes no step buys exactly one more attempt; a single
-  plan is continued at most 25 times however productive it looks; and a re-plan
-  forgives the first counter but never the second, so a model that reshapes its
-  list every turn cannot lift the ceiling. A run that ended with an error never
-  auto-continues in any mode. `Esc` cancels, and so does typing anything.
-
-  Whatever sends the continuation — the timer or `/todo continue` — now sends
-  the numbered list of open steps rather than the words "continue with the
-  remaining todo items", so a plan whose `todo_write` has scrolled out of the
-  context window is still followed exactly. Neither path records that text in
-  your prompt history any more; it is not something you typed.
-
-  Under `-p` there is no grace window, because there is nobody there to use it,
-  which is precisely why the two counters rather than the human are the
-  protection there. `[todo] continuing (3 steps left)` and
-  `[todo] 3 steps unfinished` go to stderr, suppressed by `--quiet`, and an
-  unfinished plan still never changes the exit code.
-
-  New surfaces: `/todo follow <notify|auto|off>`; `--todo-follow <mode>`;
-  `ARAGON_TODO_FOLLOW`; and `todo.followThrough` in `config.json`. The default
-  is `notify`, so a session that does not ask for any of this behaves as it did.
-
-- **A plan strip for inline mode.** Inline mode has no fixed frame and so no
-  rail, which left every pipe, every `CI=1`, every `TERM=dumb` terminal and
-  every window under twelve rows able to see `todo 3/7` but not *which* step was
-  running. There is now a one-row strip under the transcript —
-  `todo 3/7  >  Adding the rail to AppShell        +2 done` — windowed on the
-  same step the rail would have anchored on. It truncates rather than wrapping
-  and is exactly one row at every width, so the composer does not move.
-  `--no-todo-panel` hides it along with the rail.
-
-### Changed
-
-- `npm run typecheck` now covers the test tree in both packages. `tsconfig.json`
-  excludes `__tests__` so the build never emits fixtures, and vitest transpiles
-  without typechecking — between the two, a widened required field could break
-  ten test files while `npm run build` and `npm test` were both green. Twelve
-  such errors had accumulated; all twelve were fixture drift and all twelve are
-  repaired, with no suppressions and no production type relaxed to accommodate a
-  fixture.
-
-- **TODO planning, and a rail to watch it in.** For work with three or more
-  distinct steps the agent now keeps a visible plan: it writes a checklist with
-  a new `todo_write` tool, marks exactly one item in progress, and ticks it off
-  before starting the next. In full-screen mode the list is rendered as a
-  right-hand rail about a fifth of the terminal wide, windowed around the step
-  being worked on rather than sliced from the top. Ask for something small and
-  nothing appears — the prompt tells the model not to plan work it can finish in
-  one or two steps, and a one-item list against an empty plan is refused
-  outright, so a session that never plans has byte-identical layout to a build
-  without this feature.
-
-  The list is a projection of what the model believes, and that single rule
-  decides everything else: `/clear` keeps it (the conversation is still there),
-  `/reset` drops it, `/save` writes it and `/resume` restores it — and resuming
-  a session that carries no list CLEARS the current one rather than leaving a
-  plan on screen whose conversation has just been replaced. A finished plan
-  survives to the start of your next message so you get to see `7/7 done`; an
-  unfinished one survives "continue" and "now do the rest", then expires after
-  three unrelated turns rather than holding a fifth of the screen for the rest
-  of the session. A run that ends with items left over produces a notice and
-  `/todo continue`; nothing auto-continues on your behalf.
-
-  "One step at a time" is enforced by structure rather than asked for in a
-  prompt: any payload with two in-progress items is repaired down to one, and a
-  list with none promotes the first unfinished item. Whatever the model sends,
-  the panel has exactly one current row.
-
-  New surfaces: the `todo_write` tool; `/todo [status|on|off|panel on|off|clear|
-  continue]`; `--todo` / `--no-todo` and `--todo-panel` / `--no-todo-panel`;
-  `ARAGON_TODO=0`; and `todo.enabled` / `todo.panel` in `config.json` (a third
-  key, `todo.followThrough`, arrived with the follow-through entry above).
-  `panel` is
-  separate from `enabled` on purpose — a screen-reader user wants the planning
-  discipline without the column, and the system prompt varies one sentence to
-  match. Under `-p` the plan goes to stderr as `[todo] 3/7 <step>` lines,
-  suppressed by `--quiet`. Below 80 columns, behind an overlay, in inline mode
-  or on a short terminal the rail is not mounted and the counter appears in the
-  status bar instead; the transcript keeps at least 62 columns at every width.
-
-- **Output token limits that cannot themselves fail a run.** Every model and
-  every provider now defaults to 64000 output tokens, and that number is
-  configurable from the TUI, from a flag, from `ARAGON_MAX_TOKENS`, and by hand
-  in `~/.aragon-agent/config.json`. Leave it alone and it stays 64000.
-
-  What is new is that the number is an *ambition*. `gpt-4o` caps at 16384 and
-  `gemini-1.5-pro` at 8192; sending 64000 to either is an HTTP 400 that ends the
-  turn with a raw provider string in the transcript. The cap is now clamped down
-  to whatever the target model actually accepts, from a per-model table plus
-  whatever the provider's own model list reports. For a model nothing knows — a
-  proxy, a `--base-url` endpoint, something that shipped last week — the request
-  goes out at 64000, and if the provider objects, `aragon` reads the real ceiling
-  out of the error text, repairs the request, sends it exactly once more, and
-  remembers the answer for the rest of the session. One warning line; the run
-  continues.
-
-  Two combinations that were a guaranteed 400 on every request are fixed with
-  it: `--thinking xhigh` asks for a 65536-token budget while Anthropic requires
-  `max_tokens` to be strictly greater than it, and OpenAI's reasoning family
-  (`o1`, `o3`, `gpt-5`) rejects the `max_tokens` parameter name outright in
-  favour of `max_completion_tokens`.
-
-  New surfaces: `/max-tokens [<n>|auto|default]`, which with no argument reports
-  the setting *and* the effective cap for the live model; an `Effective:` line
-  under the settings screen's `Max tokens` field, so a value the model will not
-  honour is visible before you save it; `--max-tokens auto`; and
-  `aragon config set maxTokens <n|auto|default>`, which echoes the value that was
-  actually stored rather than the one you typed.
-
-- **Team subagents.** The agent can now hand independent parts of a job to up to
-  five short-lived subagents that run in parallel, then read one combined report
-  of what each of them found. It decides for itself whether to delegate, so a
-  session that never calls the new `task` tool is byte-for-byte the session that
-  existed before. On by default; `--no-team` or `team.enabled: false` turns it
-  off, and `aragon config set team.maxSubagents <n>` sets the width (1-10; the
-  10 is a hard ceiling no config file can raise).
-
-  A subagent is a full agent with the same tools and working directory but no
-  memory of your conversation, so its brief has to be self-contained. It cannot
-  delegate further and it cannot ask you anything. Plan mode and a skill's
-  `allowed-tools` ceiling both reach one level down, so delegating is never a
-  way around either; `--confirm` queues subagent confirmations one at a time and
-  labels each with the subagent that is asking.
-
-  Subagents can message each other with `team_send` / `team_wait`, delivered by
-  attaching the message to the recipient's next tool result rather than
-  interrupting it. The limits are deliberate — 6 messages each, one every 15
-  seconds, 24 per dispatch — and every refusal is a plain result naming the
-  limit rather than an error, because an error invites the retry the limits
-  exist to prevent.
-
-  While a dispatch runs, a roster sits between the transcript and the composer
-  and the status bar carries `agents 3/5`; both disappear when it finishes and
-  the transcript keeps a card you can expand with `Ctrl+O`. `Esc` aborts every
-  subagent and returns the partial report. Under `-p` the transitions go to
-  stderr and the subagents' tokens are folded into the `[usage]` footer, so the
-  cost readout cannot under-report what the team spent.
-
-  Full documentation, including the limits table and how team mode interacts
-  with plan mode, skills and `--confirm`, is in the README.
-
-- **The subagent roster now shows what each subagent is doing, and shows the
-  ones that are working.** Where every row used to say `thinking` for minutes at
-  a time, it now says `bash: npm test -w cli`, `read: src/api/routes.ts`,
-  `waiting for a2`, or the tail of the sentence the subagent is currently
-  writing. On a narrow terminal the same row falls back to a file name and a
-  program name rather than wrapping, and the elapsed clock keeps its place at
-  every width.
-
-  The five visible rows are now the five that matter: whatever is running comes
-  first, then whatever is queued, then the most recent results. Above five
-  subagents the old panel showed the five that had finished first and hid
-  everything still working behind `+3 more`, which is the opposite of what the
-  roster is for. `+3 more` now says how many of the hidden ones are still going.
-
-- **A subagent no longer waits two minutes for a message nobody can send.**
-  `team_wait` now checks whether anyone could actually answer before it blocks:
-  if the teammate it named has already finished, does not exist, or is itself
-  waiting in a way that cannot be released in time, it returns straight away and
-  says which. Nothing is refused that could still have been answered — including
-  by a third teammate further down the chain — and the subagent is told what to
-  do instead rather than handed an error.
-
-- **A subagent whose very first request dies on a rate limit is retried once.**
-  Only a genuine cold start qualifies: no turn completed and no tool ran, so
-  there is nothing to undo. It restarts after two seconds, the roster says
-  `starting (retry)`, and the report notes `retried 1x`. Pressing `Esc`, a
-  timeout, or any subagent that had already begun work is never retried.
-
-### Changed
-
-- **`"maxTokens": null` in `config.json` now means AUTO** — use each model's own
-  ceiling, never above 64000. It previously resolved to 64000 by accident (the
-  layered resolver skipped `null`), and there was no way to express AUTO through
-  any supported write path. If you have a `null` there today, the only difference
-  is a downward clamp on models whose real ceiling is below 64000 — the direction
-  that prevents failures rather than causing them. Absent still means 64000, and
-  an explicit number still means that number.
-
-- **`config.json` holds settings again.** The prompts you type are no longer
-  stored in it. They move to `~/.aragon-agent/prompt-history.jsonl` — one JSON
-  object per line, `0600` on POSIX — and the submit counter and the "notice
-  already shown" flag move to `~/.aragon-agent/state.json`. `↑` and `↓` behave
-  exactly as before.
-
-  Two things get better. Opening `config.json` to change your model no longer
-  means scrolling past a transcript of your own typing, and copying it to
-  another machine no longer carries your prompts along. And submitting a message
-  no longer rewrites the file that holds your API keys: `config.json` is written
-  when you change a setting, and at no other time.
-
-  **Your history and counters are migrated automatically, once, at the next
-  launch.** Nothing is printed, because nothing changes for you. The dead
-  `recentModels` key — which nothing has ever written — is dropped.
-
-  If you go back to 0.5.x after upgrading, that version does not know about the
-  new file and `↑` will look empty. Nothing has been lost: the entries are still
-  in `prompt-history.jsonl`, and upgrading again brings them back.
-
-### Added
-
-- **`aragon history`** — `path`, `list` (`-n`, `--json`) and `clear --yes` for
-  the prompt history, alongside the existing `aragon logs` commands. Entries are
-  stored verbatim, so `clear` is how you erase them; `aragon config set
-  historyEnabled false` stops new ones being recorded without deleting or hiding
-  what is already there.
-
-- **The mouse wheel scrolls what is under the pointer.** In full-screen mode the
-  wheel now scrolls the transcript three rows per notch (a page with `Shift`),
-  scrolls an open overlay when the pointer is over one, and steps prompt history
-  one entry per notch when the pointer is over the input box — the same code
-  path `↑` / `↓` use, so it inherits the rule that an in-progress draft is never
-  overwritten. A slim rail on the viewport's right edge shows where you are in
-  the conversation.
-
-  **This fixes a destructive bug.** Previously the wheel recalled prompt history
-  no matter where the pointer was, because the alternate screen makes terminals
-  translate each notch into arrow keys. Wheeling while composing a message
-  silently replaced the draft with an old prompt, and nothing scrolled.
-
-  **Selecting text with the mouse now requires `Shift`+drag** in most terminals.
-  That is the standard convention for "an application is tracking the mouse" and
-  it applies to everyone, including people who never scroll with the wheel — so
-  the first session says so once, in the transcript, and never again. Turn the
-  whole feature off with `--no-mouse`, `ARAGON_MOUSE=0`, or
-  `aragon config set mouse false`; the wheel then does nothing rather than
-  reverting to the old behaviour. Inline mode (`--no-fullscreen`) is untouched:
-  your terminal's own scrollback already does the right thing there.
-
-  Under `tmux` with `mouse on`, tmux keeps the wheel for itself and the app
-  never sees it. `set -g mouse off` hands it back.
-
-### Changed
-
-- **Your files moved to `~/.aragon-agent/`.** Config, sessions, installed skills
-  and the usage counters now live in one directory in your home — on Windows
-  `C:\Users\<you>\.aragon-agent\`, and the same path on macOS and Linux. They
-  used to be split across `%APPDATA%\aragon-agent-nodejs\Config` and
-  `%LOCALAPPDATA%\aragon-agent-nodejs\Data` (or the XDG equivalents), which are
-  correct locations and terrible answers to "where do I change the model?".
-
-  **The move happens once, automatically, on first launch, and prints a line on
-  stderr when it does.** Nothing is deleted: `config.json` is COPIED rather than
-  moved, and the old directories are left where they are, so reinstalling 0.5.0
-  remains lossless. `aragon config home` prints the new location and names the
-  old ones. Once you are satisfied, you can delete them yourself.
-
-  `ARAGON_HOME` relocates the whole tree. It has to be a real environment
-  variable — the path is resolved before any `.env` is read, so putting it in a
-  `.env` will not work. `aragon config home` always prints what is actually in
-  use.
-
-### Added
-
-- **Logging.** Records go to `~/.aragon-agent/logs/aragon-<date>.log` as JSON
-  Lines: startup and exit, config writes, turn boundaries with token counts,
-  tool names with durations, skill loads, and every uncaught error. Files rotate
-  by day and by size (`log.maxFileBytes`, default 5 MiB) and the oldest beyond
-  `log.maxFiles` (default 10) are deleted.
-
-  **`info`, the default, records no conversation content**, so attaching a log
-  to a bug report is safe. `--log-level debug` (or `--verbose`) adds prompts,
-  replies and tool arguments truncated to `log.previewChars`; `trace` adds them
-  in full. Those two write what you typed — check before sharing one.
-
-  API keys are removed at every level by two independent passes: fields whose
-  name looks like a credential are blanked whatever they hold, and every string
-  is scanned both for known key formats and for the literal keys this process is
-  holding — which is what covers a custom endpoint whose key looks like nothing
-  in particular. `log.redactSecrets false` turns both off and warns you on
-  stdout when you set it.
-
-  Nothing is ever uploaded. There is no network path out of this subsystem.
-
-  New commands: `aragon logs path | list | tail [-n N] [--follow] [--level lv]
-  [--json] | clear --yes | open`, and `/logs` inside the TUI. New flags:
-  `--log-level`, `--verbose`, `--log-file` / `--no-log-file`, `--log-dir`. New
-  environment variables: `ARAGON_LOG_LEVEL`, `ARAGON_LOG_FILE`, `ARAGON_LOG_DIR`
-  — these three, unlike `ARAGON_HOME`, can come from a `.env`.
-
-- **`aragon config get <key>` / `list [--json]` / `edit` / `home`.** `get` and
-  `list` mask every secret, `--json` included. `edit` opens `config.json` in
-  `$VISUAL` / `$EDITOR` (`notepad` on Windows), takes a `config.json.bak` first,
-  and re-parses on exit — an invalid edit is reported immediately, and the
-  backup is named in the message.
-
-- **`/reload`** re-reads `config.json` into the running session, for when you
-  have edited it in another window. Refused while a run is in progress: swapping
-  the model mid-turn would let the second half of an answer come from different
-  settings than the first, with nothing afterwards able to explain the result.
-
-- **Log level in the settings screen**, alongside provider / model / thinking.
-  The rest of the log settings — and `redactSecrets` above all — stay
-  command-line-only, so the one switch that can put credentials on disk cannot
-  be flipped with an arrow key.
-
-### Fixed
-
-- **A malformed `config.json` no longer resets everything in silence.** A stray
-  comma used to send the model, theme, timeouts, skills and (as far as you could
-  tell) your API key back to defaults, with nothing anywhere explaining it. The
-  fallback is unchanged — a bad config still never stops the CLI from starting —
-  but it now says so, and records why.
-
-- **Ctrl+C now flushes pending log records in every mode.** The signal handlers
-  only existed in the full-screen branch, so `-p`, the inline renderer and every
-  subcommand took Node's default termination, which does not run exit handlers.
-
-### Added
-
-- **Plan mode.** `Shift+Tab` toggles between `BUILD` (everything as before) and
-  `PLAN`, a read-only research-and-design posture. `/plan [on|off|status]` is a
-  first-class equivalent for terminals that swallow `CSI Z`, and `--plan` /
-  `--no-plan` / `ARAGON_PLAN=1` / `planModeDefault` cover the startup case.
-
-  In `PLAN`, `write_file`, `edit_file`, `bash`, `skill_install` and
-  `skill_create` are refused **at the tool boundary**, so read-only is a
-  property of the wiring rather than a promise in a prompt. `bash` is refused in
-  full, `git status` included: a shell string cannot be classified reliably, and
-  a gate that is right most of the time is worse than one that is always right,
-  because you stop trusting the badge. The refusal names `read_file` / `glob` /
-  `grep` as the alternatives.
-
-  Two tools exist only while a human is attached. `ask_user` renders a
-  keyboard-driven wizard - 1-5 questions, 2-4 options each, exactly one marked
-  `RECOMMENDED`, plus an `Other...` free-text option - where pressing `Enter`
-  through the wizard accepts every recommendation. `submit_plan` renders the
-  plan as a scrollable card: `a` approves, `r` sends written feedback (`<-`
-  returns to the card), `Esc` dismisses. **Approving flips the session to
-  `BUILD` immediately and the same run continues into implementation.**
-
-  Dismissing a plan asks for a better one rather than stopping - the agent is
-  told to refine and resubmit - so the way out of a run is `Esc` twice: the
-  first closes the card, the second aborts the run.
-
-  Toggling mid-run is asymmetric on purpose: `build -> plan` applies at once,
-  `plan -> build` is deferred to the end of the run (the status bar shows
-  `PLAN -> BUILD` meanwhile) so a run launched under a read-only guarantee
-  cannot start writing files because of one stray keypress.
-
-  `aragon -p --plan "..."` registers no question or plan tool - there is nobody
-  to ask - but the read-only gate still applies, so it means "tell me how you
-  would do this, and do not touch my repository".
-
-- Three flat config keys: `planModeDefault` (default `false`),
-  `planModeMaxAskRounds` (default `4`, clamped `[1, 10]`) and
-  `planModeHumanTimeoutMs` (default `1800000`, clamped `[60000, 7200000]`). All
-  three are settable with `aragon config set`. Additive, as always: a config
-  file written by an earlier version loads unchanged.
-
-### Fixed
-
-- A `--confirm` prompt no longer races the idle watchdog. Taking longer than
-  `idleTimeoutMs` over `Proceed? (y/N)` used to abort the run with no
-  explanation; every human wait now suspends the watchdog and restarts it in a
-  `finally`.
-
-- Honor an explicit `NODE_TLS_REJECT_UNAUTHORIZED=0` setting in the CLI and the
-  official npm release workflow while hiding only Node's standard insecure-TLS
-  warning. Other process warnings remain visible.
-
-## 0.5.0
-
-### Breaking
-
-- **The package is renamed `@argon-agent/cli` → `@aragon-agent/cli`.** The brand
-  was always meant to be *Aragon*, matching AragonMesh; the missing `a` was a
-  typo that reached the registry. Install with:
-
-  ```
-  npm uninstall -g @argon-agent/cli
-  npm install -g @aragon-agent/cli
-  ```
-
-  `@argon-agent/cli` is deprecated and receives no further releases. **The
-  executable is still `aragon`** — the command you type does not change.
-
-- **Environment variables are `ARAGON_*`, not `ARGON_*`.** There is no
-  compatibility fallback: `ARGON_MODEL`, `ARGON_PROVIDER`, `ARGON_BASE_URL`,
-  `ARGON_THINKING`, `ARGON_MAX_TOKENS`, `ARGON_THEME`, `ARGON_FULLSCREEN`,
-  `ARGON_SKILLS`, `ARGON_SKILLS_DISABLED`, `ARGON_SKILLS_PATH` and
-  `ARGON_REDUCED_MOTION` are all ignored from this release on. Rename them in
-  your shell profile and CI. Provider keys (`ANTHROPIC_API_KEY`,
-  `OPENAI_API_KEY`, `GOOGLE_API_KEY`, `GEMINI_API_KEY`) are untouched.
-
-- **The project-level skills directory is `.aragon/skills`, not
-  `.argon/skills`.** This one lives in your repository, so rename it yourself:
-
-  ```
-  git mv .argon .aragon
-  ```
-
-  A config file that already persisted the old default keeps working — an
-  explicitly saved `skills.projectDirs` still wins over the default — but the
-  shipped default is now `['.aragon/skills', '.claude/skills']`. Update it with
-  `aragon config set` if you want the new default.
-
-### Added
-
-- **Automatic migration of your user-level state.** On first run, 0.5.0 moves
-  the `env-paths` config and data directories from `argon-agent` to
-  `aragon-agent` — your API key, sessions and installed skills come with it, and
-  each installed skill's `.argon-skill.json` manifest is renamed to
-  `.aragon-skill.json` so `skills update` and `skills uninstall` still recognise
-  it. The migration prints one line to **stderr**, runs once, and is skipped
-  entirely if the new directory already exists. If it fails for any reason the
-  CLI still starts; the old directory is never deleted when a cross-volume copy
-  was needed.
-
-- The opening wordmark now reads `ARAGON` and is 52 columns wide (was 44).
-  Terminals between 52 and 55 columns show the single-line banner rather than a
-  clipped wordmark.
-
-- **Skills.** A skill is a directory holding a `SKILL.md` (YAML frontmatter +
-  Markdown body) plus optional `reference/`, `scripts/` and `assets/`. Skills
-  reach the model through three levels of progressive disclosure: a name +
-  one-line "when to use" catalog in the system prompt (≤ 6 000 bytes total),
-  the full body only when the model calls the `skill` tool, and bundled files
-  only when it reads them with `read_file` / `bash`. A skill therefore costs a
-  line of context rather than its whole body. How many fit in Level 1 depends on
-  description length and language — roughly 30 with terse English descriptions,
-  roughly 10 with Chinese ones at the character cap; past that the catalog
-  truncates and `skill_find` covers the remainder.
-- Four model-facing tools: `skill` (load), `skill_find` (search installed
-  skills), `skill_install` (install from a local directory, git repo, or https
-  `SKILL.md` / `.zip`) and `skill_create` (author a new one from a procedure the
-  agent just worked out).
-- **`aragon skills update <name> | --all`**, with `--dry-run`, `--force` and
-  `--check`. Re-fetches from the source recorded in the install manifest,
-  preserving `installedAt` and recording `updatedAt` / `previousVersion`. Refuses
-  before downloading anything when the installed copy has local edits, and
-  re-validates the source against the *current* `skills.allowedHosts` rather than
-  the one in force at install time. Not exposed to the model.
-- **Catalog ranking.** Level 1 entries are ordered by scope, then by how recently
-  and often each skill was used, so truncation drops what you do not use rather
-  than what sorts last alphabetically. With no usage data the order is unchanged.
-- **`skill_find`**, advertised in the prompt only when the catalog actually
-  truncated. Searches installed skills by name, description and keyword; never
-  reaches the network, never returns `activation: manual` skills, and tells the
-  model to ask for a source rather than invent one when nothing matches.
-- **Cross-process install lock.** Install / create / remove / update take a
-  per-directory advisory lock, closing a race in which two concurrent installs
-  could interleave their atomic-replace steps and delete a skill with no error.
-  Self-expires after 60 s; waits at most 5 s; only ever released by its owner.
-- **Load-time integrity.** Discovery re-checks `SKILL.md` against the sha256
-  recorded at install, at no extra I/O. `skills.integrity` selects
-  `off` / `warn` (default) / `strict`.
-- **Local usage counters** at `<data>/skill-usage.json` — skill name, count and
-  timestamp only, never transmitted, disabled with `skills.usageTracking=false`.
-  Automatically injected `activation: always` skills are not counted.
-- Config keys `skills.integrity` and `skills.usageTracking`; `aragon skills list
-  --sort=recent` and `/skills list --sort=recent`.
-- Every usable skill also becomes a slash command: `/<skill-name> [args]`, with
-  `$ARGUMENTS` / `$1..$9` substitution. On a name clash the built-in command
-  always wins and the skill is reachable as `/skill:<name>`.
-- `/skills` management: `list`, `info`, `install`, `remove`, `enable`,
-  `disable`, `reload`, `create`, `trust`, `untrust`.
-- `aragon skills …` non-interactive equivalents plus `path` and `doctor`.
-  Exit codes follow `config set`: `0` ok, `1` run-time failure, `2` usage.
-- Four discovery scopes in ascending precedence: bundled (`<pkg>/skills`), user
-  (`<data>/skills`), project (`<cwd>/.aragon/skills` and, read-only,
-  `<cwd>/.claude/skills` for Claude Code interop), and `ARAGON_SKILLS_PATH`.
-- New flags `--no-skills`, `--skill <name>` (repeatable) and `--skills-yes`;
-  env `ARAGON_SKILLS`, `ARAGON_SKILLS_PATH`, `ARAGON_SKILLS_DISABLED`; config keys
-  `skills.enabled`, `skills.requireApproval`, `skills.catalogMaxBytes`,
-  `skills.bodyMaxBytes`.
-- One bundled skill, `skill-creator`, documenting how to write a good one.
-- New runtime dependency: `fflate` (~30 KB, MIT, zero transitive deps) for zip
-  extraction. `@aragon-agent/core` gains no dependency.
-- **Release tooling for the scope change.** `npm run verify:brand` scans every
-  file that `npm pack` would put in the tarball — README and `dist/` included —
-  and blocks the release on any surviving `argon` reference, with only the four
-  documented exceptions (the two CHANGELOGs and the legacy-state migration
-  module, which hold the old names by definition). `packages/cli` also gains a
-  `prepublishOnly: npm run build`, so a publish that bypasses
-  `publish-latest.ps1` can no longer ship a `dist/` that is out of step with the
-  sources. `deprecate-legacy.ps1` retires `@argon-agent/*` on the registry, and
-  refuses to do so until the replacements are actually published.
-
-### Changed
-
-- The globally installed executable is now `aragon`. The previous `argon` and
-  `argon-agent` aliases are no longer installed.
-- `config/store.ts` now deep-merges the `skills` section on both read and write,
-  the same treatment `apiKeys` already received. Without it a partial patch such
-  as the one `/skills disable` sends would replace the whole section and
-  silently discard `trustedProjectDirs`, `allowedHosts` and `requireApproval`.
-- `slashSuggestions` accepts `-` and `:`, so kebab-case and `skill:`-namespaced
-  commands autocomplete. Previously `/my-sk` produced no popup at all.
-- `AgentController` now has a single `agent.setSystemPrompt()` call site, so a
-  `/cwd` change can no longer drop the skill catalog out of the prompt.
-- With skills off or none installed, the system prompt and the 7-tool array are
-  byte-identical to the previous release.
-- Added a repository-level PowerShell release workflow that automatically
-  updates CLI and Core versions, synchronizes their dependency and lockfile,
-  verifies both tarballs, and publishes them in dependency order.
-
-### Security
-
-- Installing a skill requires an explicit human approval that **probes for a
-  human first and refuses when there is none**. Under `-p` there is no TUI and
-  therefore no approver, so `skill_install` is refused with an actionable
-  message rather than writing a third-party directory unattended. Use
-  `--skills-yes` (or `skills.requireApproval=false`) to opt out deliberately.
-- Project skill directories require a one-time trust confirmation before they
-  are loaded — cloning an unfamiliar repository does not grant it the ability
-  to inject instructions into your session.
-- Git is invoked only via `execFile` with an argv array and `shell: false`, and
-  owner / repo / ref / subdirectory fragments are whitelisted and refused if
-  they begin with `-`. A leading dash matters even with a pure argv array:
-  `--upload-pack=<program>` makes git execute that program.
-- Zip extraction is streamed and gated on **actually written** bytes, per-entry
-  size, compression ratio and entry count, all of which can abort mid-archive.
-  The size a zip *declares* is written by whoever built it and is never trusted.
-  Path traversal, absolute entries, symlinks and Windows device names are
-  refused, and a rejected archive is removed rather than left half-extracted.
-- Network fetches are `https:` only, restricted to `skills.allowedHosts`,
-  re-checked on every redirect (max 3), size-capped, and refuse IP literals and
-  loopback.
-- Nothing a skill ships is executed at install time, and the executable bit is
-  not preserved. Scripts run only when the model invokes `bash` explicitly.
-  There is no sandbox: approval blocks *silent* installation, it does not make
-  a skill you approved safe.
-
-## 0.3.0
-
-A full-screen TUI release. `@argon-agent/core` is untouched, no new runtime
-dependency is added, and the headless/print contract, exit codes, and
-slash-command semantics are unchanged. The 0.2.0 inline renderer remains
-available in full via `--no-fullscreen`.
-
-### Added
-
-- **Full-screen mode (default).** `aragon` now takes over the terminal's
-  alternate screen buffer — the same mechanism `vim` / `htop` / `lazygit` use.
-  Your shell history is covered, not erased, and comes back untouched on exit.
-  The frame is a fixed `rows - 1` tall, so **the composer and the status bar sit
-  at the physical bottom of the screen from the very first frame**: empty
-  session, long session, mid-scroll, or overlay open.
-- **Brand region.** A six-row gradient ASCII wordmark on a roomy empty session,
-  degrading to a banner, then a single-line bar, then a bare wordmark as space
-  runs out. Row 1 always begins with the brand glyph.
-- **Self-drawn scrolling.** `PgUp` / `PgDn` by the page, `Shift+↑` / `Shift+↓` by
-  the line. The viewport auto-follows new output while pinned to the bottom;
-  scrolling away shows `↑N` in the status bar plus a "N new lines" hint, and
-  submitting a message always re-pins.
-- **Exit replay.** After leaving the alternate screen the session is replayed
-  into the normal buffer as plain, colorless text, so the conversation does not
-  evaporate with the screen. Turn it off with `--no-exit-transcript`.
-- **New flags** `--fullscreen` / `--no-fullscreen` / `--no-exit-transcript`, the
-  env var `ARGON_FULLSCREEN=0|1`, and the config keys `fullscreen`,
-  `exitTranscript`, and `transcriptWindow`.
-- **Automatic downgrade.** A non-TTY stdout, `TERM=dumb`, CI, and terminals under
-  12 rows / 40 columns fall back to the inline renderer. Only the non-TTY gate is
-  un-overridable — `--fullscreen` beats the rest.
-
-### Changed
-
-- The status bar lost its round border and is now exactly one row; it had been
-  spending an eighth of a 24-row terminal framing a single line of text.
-- The toast strip holds a fixed row in full-screen mode, so the transcript no
-  longer jumps as toasts appear and expire.
-- The composer is a rounded frame whose border color reports state (idle, has a
-  draft, running, blurred by an overlay), with the completion popup above it.
-- `Ctrl+L` repaints the frame instead of writing a clear sequence. Ink dedupes
-  identical output at two separate gates, so under a fixed frame the old
-  approach erased the screen and then declined to redraw it.
-- `Shift+↑` / `Shift+↓` no longer recall prompt history — they scroll.
-- Full-screen mode routes `console.*` into the transcript as notices rather than
-  letting Ink write them straight to stdout, which corrupts the frame's line
-  accounting.
-
-### Known trade-offs
-
-- The terminal's native scrollback and mouse wheel do not scroll the transcript
-  in full-screen mode; use `PgUp` / `PgDn`, or run with `--no-fullscreen`.
-- Mouse tracking is deliberately left off: enabling it costs text selection and
-  copy in most terminals.
-- Should a crash ever strand your terminal on the alternate screen, `reset`
-  restores it. Four independent restore paths exist to prevent that: normal
-  exit, `process.exit`, `SIGINT` / `SIGTERM` / `SIGHUP`, and Ink's own signal
-  handling.
-
-## 0.2.0
-
-A design-elevation & HCI-hardening release. No breaking changes: `@argon-agent/core`
-is untouched, no new runtime dependency is added, and the headless/print contract,
-exit codes, config keys, and slash-command semantics from 0.1.0 are unchanged.
-
-### Added
-
-- **Real theming engine.** Distinct `dark` / `light` palettes with an `auto`
-  resolution (`auto → dark`), color degradation across terminal depth
-  (truecolor → 256 → 16 → monochrome), and an ASCII glyph fallback for terminals
-  without Unicode. Honors `NO_COLOR`, `FORCE_COLOR`, `--no-color`, and
-  `TERM=dumb`. Switch live with `/theme <auto|dark|light>` (persisted).
-- **Elegant chrome.** A gradient wordmark header that collapses to a slim sticky
-  bar after the first turn, a first-run **Welcome / empty-state** card, and a
-  status bar with a colored **context gauge** (threshold colors), a **tokens/sec**
-  read-out while running, and a contextual hint.
-- **Legible tool output.** Colored unified diffs (green adds / red removes) and
-  per-tool rich previews (`edit_file`/`write_file`/`read_file`/`bash`/`list_dir`/
-  `glob`). Long previews collapse to 8 lines and expand with `Ctrl+O` or
-  `/expand [n]`.
-- **Autocomplete.** A slash-command palette and `@file` path completion popup
-  (`Tab`/`→` to complete, `Up`/`Down` to move, `Esc` to close).
-- **Full line editing.** Home/End, word jump/delete, kill-to-start/end, and
-  multi-line vertical cursor navigation, with prompt-history recall at the edges.
-  Unknown control/escape sequences are dropped instead of inserted.
-- **Reliability under load.** Ink `<Static>` finalization for settled entries and
-  a streaming delta coalescer, so long/fast sessions stay flicker-free.
-- **Calm feedback.** Ephemeral, auto-dismissing **toasts** for transient acks;
-  the transcript keeps only durable content (messages, tool cards, run errors).
-- `reducedMotion` config (also `ARGON_REDUCED_MOTION`; implied by `--no-color`)
-  replaces spinners with a static glyph.
-
-### Changed
-
-- History scrollback now uses the **native terminal scrollback** that `<Static>`
-  prints into (this supersedes the never-implemented `PgUp`/`PgDn` promise).
-
-## 0.1.0
-
-Initial release — a Claude-Code / Codex-style interactive terminal UI (TUI) on
-top of `@argon-agent/core`.
-
-### Added
-
-- **Interactive TUI** (React + Ink): gradient header, streaming transcript with
-  markdown + syntax highlighting, collapsible thinking blocks, live tool-call
-  cards, a status bar (model / context % / tokens / cost / elapsed), and modal
-  overlays (settings, model picker, help, confirm).
-- **Built-in toolset** wired into the core agent at full permission:
-  `read_file`, `write_file`, `edit_file`, `list_dir`, `glob`, `grep`, `bash`.
-- **Layered configuration** (defaults → user file → env/`.env` → flags) with
-  live in-TUI editing of provider / model / base URL / API keys / thinking level
-  / max tokens. Secrets are masked and the config file is written `0600` on POSIX.
-- **One-shot / print mode** (`argon -p`) for scripting and piping, plus
-  `argon config` / `argon config set` / `argon config path` / `argon models`.
-- **Session control**: abort (Esc), steer-while-running, follow-up queue,
-  double-Ctrl-C exit, `/save` / `/resume` JSON sessions.
-- **Timeout invariant** (`idleTimeout ≥ toolTimeout`) so a long single tool run
-  is never killed by the idle watchdog.
-- **Never-silent failures**: pre-flight API-key validation, error-StreamEvent
-  rendering, and a guard that surfaces a swallowed agent error instead of a
-  blank no-op (headless exits `1`/`2` accordingly).
+# Changelog
+
+All notable changes to `@aragon-agent/cli` are documented here. Entries at
+`0.4.x` and earlier refer to the package by its former name, `@argon-agent/cli`
+— those releases really were published under that name, so they are left as they
+were written.
+
+## Unreleased
+
+### Added
+- **`Ctrl+L` branches this session into a new terminal - and `Ctrl+R` now
+  redraws (KEY BINDING CHANGE).** One keystroke opens a PowerShell (Windows) or
+  bash (macOS/Linux) window in the session's current `/cwd`, running
+  `aragon --resume <id>` there: the child derives its own id and only ever writes
+  its own file, so the TUI you branched from keeps running untouched. The session
+  is force-flushed to disk BEFORE anything is spawned, because the child reads
+  that file first and the normal checkpoint is debounced. Works mid-run too - the
+  child inherits the last settled turn. Platforms: a new Windows Terminal tab
+  (`wt -d <cwd> ... --`), a new console window (`cmd /c start`), a Terminal.app
+  window via `osascript`, or the first available Linux emulator; each rung carries
+  the working directory itself, because `--resume` does not restore one. No
+  terminal (SSH, container) prints a pasteable `aragon --resume <id>` instead of
+  failing silently. `ARAGON_BRANCH_COMMAND` / `ARAGON_BRANCH_SHELL` override what
+  runs and which shell; presses are throttled to one window per two seconds and a
+  failed attempt never starts the cooldown. New `/branch` is the same code path.
+  The header (top right) now shows `^L branch` next to `^I Index` and `^G more`,
+  spending a fixed cell budget: `^I` therefore appears from 52 columns
+  (collapsed) / 62 columns (expanded) instead of 72, and on a very narrow terminal
+  the branch hint degrades to the bare `^L` before it is dropped.
+  **Breaking:** the redraw previously bound to `Ctrl+L` moved to `Ctrl+R` - same
+  behaviour, same overlay guard (the settings screen keeps its own `Ctrl+R`).
+
+- **A TUI session now has an id, and another terminal can continue it.** Every
+  interactive session is written to `~/.aragon-agent/sessions/<id>.json` as it
+  runs and shows that id at the right end of the status bar (`id a4k2m9`, or
+  `a4k2m9` on a narrow terminal - the id itself is never truncated). Checkpoints
+  are written after the first frame, after every settled turn (debounced), on
+  `/session pin`, and once on exit; each one is an atomic replace, because a
+  second terminal may be reading. `aragon --resume <id|path>` continues a saved
+  session under a NEW id and never writes the file it came from, so branching
+  while the original TUI keeps running is safe - including
+  `aragon exec --resume <tui-id>`, which now writes a session of its own instead
+  of overwriting yours. `aragon --save <id>` names this run's id and pins it
+  against `sessions prune`. New `/session [copy|pin]` reports the id, the file and
+  the branch parent, copies the resume command through the same clipboard door as
+  `/copy` (OSC 52 included), and pins the file. A checkpoint taken mid-tool is
+  repaired on the saved copy only, so resuming it cannot produce an illegal
+  history, and a tool card that was still running is settled so nothing spins
+  after a resume. `aragon -p` writes one checkpoint at exit: context only, no
+  transcript, because nothing was displayed.
+
+- **A model profile can be verified where it is written.** The profile editor
+  gains two action rows: **Test connection** (`Ctrl+R`) and **Discover models**
+  (`Ctrl+F`). A test is one real completion against the UNSAVED draft - the
+  provider, model, base URL and the key `keep`/`replace`/`shared` actually
+  resolves to - and it answers with a verdict (`OK 812 ms | 18 in / 4 out`) or a
+  classified failure plus the one thing to fix (`Key rejected by anthropic (HTTP
+  401).` / `Fix: edit API key, or set Credential mode to Shared.`). A discovery
+  lists what the key can really reach, so the model is picked instead of typed.
+  Both run against the draft, cost at most one short request (discovery costs no
+  tokens), never retry, and never write `ModelRegistry`'s five-minute discovery
+  cache. Every result is bound to a target fingerprint, so an `OK` can never
+  describe a model, endpoint or key the user has since edited. The progress line
+  is static text: no new animated site, so `single-spinner` still holds.
+
+### Changed
+
+- **`aragon sessions list` gained an `origin` column** and shows TUI checkpoints,
+  which are managed files now. `sessions prune` defaults to `--origin exec`, so it
+  behaves exactly as before for every file that existed before this release; the
+  new `--origin tui|all` selects the other writer. `prune` never deletes a pinned
+  session or one whose owner process is still alive, and `--all` now means only
+  "also consider files with no session metadata" - it can no longer reach a TUI
+  checkpoint. `/resume` inside the TUI moves the session to a new id and restarts
+  the spend counter, because a new id's file must not bill the previous
+  conversation.
+
+### Fixed
+
+- **A conversation saved mid-tool could be resumed into an illegal history.**
+  A `tool_call` with no matching `tool_result` makes the next request fail. The
+  checkpoint writer closes every dangling call on the copy it writes and re-checks
+  sendability before replacing the file; a history it cannot make sendable is not
+  written at all, so the previous good checkpoint survives.
+- **A checkpoint could carry a tool card that never stopped spinning.** A card
+  saved while its tool was running kept `ToolCard` animated and kept the
+  transcript's settled boundary from ever passing it, which re-rendered the tail
+  on every frame for the rest of the resumed session. Cards are now settled on the
+  saved copy only, which leaves `/save` (an explicit export you are watching) and
+  its documented behaviour untouched.
+- **`aragon exec --resume <tui-id>` overwrote the session it resumed.** The
+  write-back id preferred the file's own metadata, so exec took the TUI's id and
+  replaced the checkpoint - silent, unrecoverable loss of the conversation still
+  running in the other terminal. A TUI file now always yields a fresh id, and the
+  same rule covers `--session-id`, which used to return the typed id verbatim.
+- **`sessions prune --origin tui --yes` could delete a running session's
+  checkpoint.** The `live` marker was written by the TUI but the process lock it
+  is checked against was only ever taken by `exec`, so the liveness guard answered
+  "nobody owns this" for exactly the files it was written for. The TUI checkpoint
+  writer now holds the session lock for as long as it owns the file, releases it
+  on exit, and moves it when `/resume` rotates the id.
+- **`aragon --save <id>` said "already exists" for an id that does not exist.**
+  A session directory that cannot be written (read-only home, full disk) was
+  reported as a name collision, which sends the user to `--resume <id>` and then
+  to "no such session". It now names the directory instead.
+- **A hand-written `/save` file resumed as if it were a branch of itself.**
+  `--resume ./notes.json` recorded `notes` as `meta.branchOf` and announced it as
+  the parent session, though it is not a session id. Lineage is now recorded only
+  for ids this CLI minted, and a collision while deriving a branch id re-mints the
+  suffix instead of dropping the branch shape.
+- **`/session` answered `pinned  no` right after `/session pin`.** The file was
+  marked and the report was not. The identity is now updated with the pin.
+- **An id rotation could write one turn into the old id's file.** For the window
+  between `/resume` claiming a new id and the next render, the writer still aimed
+  at the previous file while recording the new parent. The rotation now republishes
+  the writer's id in the same step.
+- **`/clear` did not clear the context.** It reset the visible transcript and the
+  todo panel and left `agent.state.messages` - the history that is actually sent,
+  billed, measured by `/context` and compacted - completely untouched. A long
+  session looked restarted and behaved like it had never been cleared: the next
+  request still carried every old turn, so the model remembered the previous task
+  and the occupancy gauge did not move. `/clear` now runs the one real
+  new-conversation boundary (`controller.resetConversation()`) in the only safe
+  order: interrupt the running turn (`forceStop`, which also drops the events the
+  unwinding engine keeps emitting), wait for it to settle inside a bounded budget,
+  archive the old conversation, drop both queues, then clear history, compaction
+  identity, context measurement and skill activation frames. A reset that cannot
+  stop the engine clears **nothing** and says so.
+- **`/reset`'s "cancelled N queued messages" notice is preserved.** `/reset` is now
+  an alias of `/clear`, and the count is read **before** the queues are cleared -
+  read afterwards it would always be zero, which is a notice that lies.
+
+### Changed
+
+- **`/clear` means "start a new conversation"; the old behaviour moved to
+  `/clear screen`.** The previous semantics (wipe the transcript and the todo
+  panel, keep the history and the queue) are still available, spelled explicitly,
+  and are byte-identical to what `/clear` used to do. `/clear` no longer refuses
+  while a run is in flight: it interrupts, which removes the dead end where the
+  only command that worked mid-run was the one that kept the context.
+- **`/reset` is an alias of `/clear`, one implementation.** Two bodies could drift
+  back into "one clears the screen, one clears the session"; one body cannot.
+- **The conversation being cleared is archived first**, to
+  `~/.aragon-agent/sessions/cleared/cleared-<yyyyMMdd>-<HHmmss>-<pid>.json`
+  (newest 20 kept, `<ARAGON_HOME>` honoured). The path is reported in the
+  transcript and restores with `/resume <path>`. The file carries no `meta`, so
+  `sessions list` / `sessions prune` and `--continue` are unaffected; a failed
+  archive only logs and never blocks the clear.
+
+- **`Ctrl+I` was dead on Windows.** The win32-input-mode translator only
+  recognised the chord when the console reported it as `VK_TAB` (vk 9) with
+  ctrl - a record shape no real console was ever probed for. Physical
+  Ctrl+I keeps the LETTER vk (0x49, the same Ctrl+letter rule that reports
+  Ctrl+A as vk 65 and Ctrl+J as vk 74 in the Shift+Enter probe captures)
+  while `char` carries the Tab byte, so the record fell through to the
+  generic Ctrl-combo branch, came out as plain Tab, and the keybinding
+  silently did completion instead of opening the index confirm. The
+  translator now tells Ctrl+I apart from Tab by `char === 0x09 && ctrl`
+  whatever vk it rides, with Shift still mapping to Shift+Tab; both record
+  shapes are pinned by regression rows in `win32-input-mode.test.ts`.
+
+### Team dispatch supervisor (team-overseer)
+
+- **A time ceiling on a subagent is now an inspection trigger, not a death
+  sentence.** With the supervisor on (new key `team.overseer`, default
+  `true`, needs the fast tier), two soft timers per child - event silence at
+  `idleTimeoutMs` and a wall clock at `subagentTimeoutMs` (or 5 minutes when
+  no ceiling is configured) - fire one bounded fast-tier inspection. The
+  supervisor decides wait / nudge / replace / abandon; the runtime applies,
+  funnelling every lifecycle change through the worker loop that awaits the
+  child, so a replacement is never built for nobody.
+- **A nudge is steering; a replace is a rebuild with an amended brief.**
+  Guidance is queued at core's safe top-of-loop checkpoint; a replacement
+  carries the supervisor's reason, the dead child's `filesTouched`, and the
+  original brief bounded to the same prompt budget. A post-mortem inspection
+  diagnoses a child that died silently (watchdog death, stream loss) and may
+  re-assign it once.
+- **Budgets are structural, not settings**: 4 inspections and 1 replacement
+  per child per dispatch; past them the child runs under its own (lengthened
+  while supervised) idle watchdog and the supervisor goes quiet. An
+  unparseable or failing supervision call always becomes `wait` -
+  supervision must never kill a healthy child.
+- Every applied decision is visible three ways: the report's Supervisor
+  section (`[silence 30s] a1: nudge - ...`), the team panel and the `-p`
+  stderr line, and a `debug` log event. `aragon exec` emits it as an
+  additive `team` subtype `overseer` (`EXEC_SCHEMA_VERSION` unchanged).
+- New command `/team overseer on|off` (refused mid-dispatch, like every
+  other `/team` setting) and new config key `team.overseer`, also settable
+  as `aragon config set team.overseer false`. With the supervisor off - or
+  no fast tier - every path is byte-for-byte the pre-feature regime:
+  `subagentTimeoutMs` is a hard abort and silence is the child's own
+  watchdog.
+
+### 启动与缩窗不再误报 Frame diff lost sync
+
+- 启动即现的 "Frame diffing lost sync with the terminal and fell back to a
+  full repaint … `--no-diff-render`" 警告已消除：Ink 挂载/卸载经渲染 stdout 写出的
+  光标隐藏/恢复序列（`ESC[?25l` / `ESC[?25h`）由帧写入器直接透传——不进差分、
+  不失效缓存、不计 fallback，也不再消耗会话的首写豁免。启动序列恢复
+  "种子帧透传 → 首帧全量 → 之后纯 diff"。
+- 终端高度缩小（拖小窗口、改字号、面板重排）不再触发同一条警告：Ink 的
+  `clearTerminal` 整帧写入被差分器识别为 Ink 自身写法，透传并失效缓存但不计
+  fallback；该控制写同样不消耗首写豁免，其后的首个 log-update 帧仍按种子帧豁免，
+  下一帧自然全量重绘。
+- `/perf` 的 `fallbacks` 重新从 0 起步，"健康会话 fallbacks === 0" 的诊断信号
+  （AC-11）恢复可判。
+
+### Consistent TUI styling
+
+- Interactive sessions always use Unicode and default to truecolor, including
+  PowerShell sessions without terminal or UTF locale markers. Ink's actual color
+  output follows the same policy, with renderer state restored on exit.
+- Explicit color opt-outs keep Unicode borders and icons; the opening logo now
+  depends only on available space. Headless terminal detection is unchanged.
+- Reloading model settings preserves session display preferences. Syntax
+  highlighting follows the same color policy, and status icons use consistent
+  single-column glyphs to prevent narrow-frame overflow.
+- Hiding the editor caret while scrolling preserves draft geometry, preventing
+  a layout update loop when entering text above the full startup logo.
+- Register exit cleanup before mounting the application, so immediate exits and
+  initial render failures restore terminal and color state reliably. Failed
+  mounts no longer mark an auto-installed build as healthy.
+
+### Model context window
+
+- Discover context limits at startup and after connection changes without blocking
+  input. The status bar and compaction use the discovered limit; `/context` names
+  API versus catalog metadata. Unknown windows show `?`, not a fabricated 128k
+  capacity or percentage. Manual `contextWindow` overrides retain precedence.
+- Window overrides now apply to tail relief and compaction progress as well as
+  the gauge. Workers on different models/endpoints do not inherit the lead's
+  override. API-key changes refresh metadata, and ending a session cancels its
+  pending model queries. `/context` also hides unknown capacities/percentages.
+
+### TUI 输入与排队可靠性
+
+- Shift+Enter 与粘贴按输入顺序处理；普通 Enter 保持弹层兼容，混合输入最多提交一次，
+  余稿及拒绝提交的正文保留。结束输入流时完整排空，不泄漏内部控制标记。
+- 超限粘贴拒绝同一外部输入块的正文和 Enter，避免误提交旧稿；未闭合粘贴继续按正文消费。
+- 拖选后仅 Ctrl+C 复制；文字变化与丢失鼠标释放时取消无效高亮，避免卡住视口。
+- 固定 Queue 状态行按消息 ID 回执更新，支持窄屏、暂停与多条消息计数。
+- 保存和退出保留所有未发送全文；恢复先验证存档，旧 queued 作为警告展示且不自动重放。
+
+### Removed
+
+- Removed the inline terminal UI and its automatic downgrade rules. Interactive
+  sessions now always use the full-screen UI; small windows show a resize notice.
+- Removed `--fullscreen`, `--no-fullscreen`, `ARAGON_FULLSCREEN`, and the
+  `fullscreen` config setting. Old flags are rejected and old stored preferences
+  are ignored. Headless commands and piped input retain their existing behavior.
+- Interactive-only screens now reject non-TTY input/output with exit code 2.
+
+### Added
+
+- **`system/init` now lists what this build can do.** A new optional
+  `capabilities: string[]` field, carrying `["interrupt"]` for this release.
+
+  A wrapper cannot tell by observation whether an interrupt will settle
+  gracefully or kill the process outright, and guessing wrong costs the user
+  their conversation. Rather than have callers compare version numbers — which
+  breaks on forks, prereleases, and version shapes nobody anticipated — the
+  build now says so itself, in an event that already arrives on every start.
+  Absent on older builds, and absence should be read as "no", never as
+  "probably".
+
+  Additive, so `schemaVersion` does not move. Consumers that do not know the
+  field ignore it, exactly as the schema contract asks.
+
+- **The composer understands pasting.** Until now it did not have the concept at
+  all: every byte the terminal delivered on stdin was a keystroke, and a
+  clipboard block is thousands of bytes arriving in an unpredictable number of
+  chunks. Three separate failures fell out of that one omission, and on Windows
+  you hit all three at once.
+
+  A carriage return that landed alone in a chunk **sent the half-finished
+  message** — a fifteen-line paste could arrive as a dozen separate messages,
+  interleaved with the agent's replies. A carriage return in the middle of a
+  chunk was inserted into the draft as a literal `\r`, which the terminal obeyed
+  by **overwriting the row it was drawn on**. And a paste that survived both grew
+  the composer to the height of the pasted text, pushing the transcript out of a
+  frame whose row accounting still believed the composer was three rows tall.
+
+  Pasting now works the way it reads:
+
+  - **A paste never submits.** Line breaks become buffer newlines, exactly one
+    per source line break, `CRLF` and lone `CR` alike.
+  - **A short paste appears verbatim.** At or below 6 lines and 400 characters it
+    is inserted as-is, indentation intact, and renders exactly like typed text.
+  - **A large paste collapses to a placeholder** — `[Pasted text #1 +218 lines]`
+    — which occupies one line, carries its own size in its label, deletes as a
+    single unit (one `Backspace`, and `Ctrl+W` / `Ctrl+U` / `Ctrl+K` cannot leave
+    half of one behind), and expands back to the full text the moment you send.
+    The model receives every byte; the composer never has to draw them.
+  - **No pasted byte is executed as a terminal control.** `\r`, `\x1b`, NUL, DEL
+    and the C1 range are stripped from the payload — never from your keystrokes,
+    so `Ctrl+A` and Escape are untouched.
+  - **The composer has a ceiling.** A draft of any length occupies at most 3 / 6 /
+    10 rows depending on terminal height, scrolls internally to keep the caret
+    visible, shows how many rows are hidden above and below, and reports its true
+    height to the layout budget. Four hundred presses of `Shift+Enter` were
+    already able to destroy the frame; they no longer are.
+  - **A submitted message cannot make the transcript unresponsive.** A user entry
+    renders at most 40 rows plus a `... +N more lines` tail. The full text is kept
+    and the exit transcript is uncapped.
+
+  On terminals that support it, detection is **exact**: aragon asks for bracketed
+  paste (`DECSET 2004`) in full-screen mode and reads the markers the terminal
+  wraps your paste in. Where that is unavailable — inline mode, an old Windows
+  console — a heuristic on chunk shape catches the same cases, and it needs no
+  cooperation from the terminal at all. Both feed one code path.
+
+  Turn it off with `--no-paste`, `ARAGON_PASTE=0`, or
+  `aragon config set paste false`; the input path is then byte-for-byte what
+  0.6.2 did.
+
+### Fixed
+
+- **The right-edge scrollbar no longer disappears while output streams.** The
+  differential renderer ended every repainted row with `CSI K` (erase to end of
+  line). The scrollbar fills the last column of every viewport row, and on
+  terminals that keep the cursor on the last column after writing it (Windows
+  console hosts) that erase blanked the cell that had just been drawn, so the
+  scrollbar vanished whenever a row was redrawn. Rows that already fill the
+  terminal width are now repainted without the erase; narrower rows are
+  unchanged. `--no-diff-render` was never affected. `/perf` gained a `scrollbar:`
+  line for diagnosing visibility and drag problems.
+
+- **`{"type":"interrupt"}` now really does settle the run — on every platform.**
+  The stdin interrupt frame used to raise a SIGINT at our own process. On POSIX
+  that reaches the runner's terminator and the run winds down properly: it
+  breaks out of the loop, persists the session, and emits `result` with the
+  usage and cost. On Windows, Node does not deliver a self-directed SIGINT to
+  its own handlers — the process simply terminated with exit code 1, having run
+  none of it.
+
+  For anyone using `aragon exec` interactively that difference was invisible,
+  because an interrupt there is usually the last thing that happens. It is not
+  invisible to a **wrapper** that interrupts a turn and then respawns on the same
+  `--session-id` to keep the conversation going: on Windows it silently got back
+  a session missing everything since the previous save. `exec/index.ts` has
+  carried a comment about that caller for some time ("PERSIST FIRST, ANNOUNCE
+  SECOND … a caller that reads `result` and immediately respawns with the same
+  `--session-id` must not race the writer"), so the intent was there; the path
+  just could not be reached on Windows.
+
+  The frame now calls the runner directly instead of going through a signal.
+  POSIX behaviour is byte-for-byte unchanged — the same code runs, with the same
+  `stopReason`, the same exit code 130, and the same two-second window in which a
+  second signal still forces an immediate exit.
+
+- **A pasted API key is stored correctly again.** Ink broadcasts each stdin chunk
+  to every mounted input handler, so the settings screen, the question overlay
+  and the plan-review feedback field all see what the composer sees. Without
+  this, a ~100-character key pasted into Settings would have been stored wrapped
+  in invisible framing bytes, masked on screen by the password dots, and would
+  have failed to authenticate on every request afterwards with nothing anywhere
+  explaining why.
+
+- **`--no-mouse` still writes no mouse sequence.** Paste support installs a
+  stdin filter, and `paste` defaults to on, so the filter now exists in sessions
+  that disabled the mouse. Whether the mouse is on is a separate question from
+  whether the stream is wrapped, and the four places that used to conflate them
+  now ask the right one: `aragon --no-mouse` writes no `?1000h`, no `?1006h` and
+  no `?1002h`, leaves your terminal's own click-drag selection alone, and reports
+  `/mouse` as off — while pasting still works.
+
+### Changed
+
+- **The run status row moved above the input box and carries the `steer` hint.**
+  While a run is active, the spinner, the phrase and `steer / interrupt / exit`
+  share one row directly above the input, in place of the idle hint row below it
+  (the footer keeps its height). It scrolls with the input; when it leaves the
+  viewport the fixed bottom row shows the animation instead, so there is never
+  more than one spinner. Terminals under 20 rows and `--no-hints` keep the old
+  fixed row. The update notice is also held back for the whole run.
+
+- **Auto-compaction no longer reports compactions that did not happen.** A long
+  unattended run used to print a red `context not compacted #2: nothing_to_drop`
+  card into the transcript. The card was accurate; nobody had asked for it.
+  Auto-compaction evaluates itself at every turn boundary, and a checkpoint that
+  looks at the history and finds nothing worth dropping has changed nothing,
+  called nothing and spent nothing — a smoke alarm announcing "no fire" twice an
+  hour, in error colours, with a machine reason attached.
+
+  A compaction now announces itself **if and only if it commits to doing work**.
+  That is a position in the control flow, not a list of reason strings: everything
+  decided before the first spend is silent, and everything decided after it —
+  including every failure, the truncation fallback, an Esc, and a host fault — is
+  as loud as it ever was. A declined `pressure` checkpoint is now
+  indistinguishable from a turn on which the threshold was never crossed.
+
+  **`/compact` and overflow recovery are unchanged.** A compaction you asked for
+  is owed an answer, and a reactive compaction running after the provider has
+  already refused your request is the one moment the diagnosis matters most. Both
+  still report on every outcome.
+
+  **Quiet, not unrecorded.** Every attempt is still written to the diagnostic log
+  (its `compaction` line gains an `announced` field, so `announced: false` is one
+  grep), still charged to the anti-loop guards, and still counted by
+  `/compact status` — which now reads
+  `This session: 1 compaction (2 checkpoints declined), 94.2k tokens reclaimed`
+  when checkpoints declined, and is byte-identical to before when none did. If
+  declines keep happening, the self-disable notice still says so out loud, once,
+  naming the cause and the remedies. Card numbers are also contiguous again: `#N`
+  counts compactions that ran, not checkpoints that fired.
+
+  **Behaviour change for wrapper scripts.** A declined `pressure` compaction no
+  longer emits a `{ type: 'compaction', subtype: 'start' | 'end' }` pair on the
+  `--output-format json` stream, and no longer prints `[compaction] skipped:` on
+  `-p` stderr. The pairs stay balanced, so only consumers counting *attempts* are
+  affected; the diagnostic log reports those more accurately.
+
+### Added
+
+- **Context compaction, hardened — the trigger now counts the whole history, the
+  recent turns can no longer wedge a run, sub-agents get the same protection, and
+  a compaction is no longer irreversible.**
+
+  **The trigger was systematically late.** It fired on the provider's reported
+  usage for the PREVIOUS request, read at a checkpoint that runs AFTER the current
+  turn's tool results were appended — and tool results are the largest single
+  increment in this loop (100 KB per call, several per turn, ~25 k tokens each).
+  A session reading 88 % could be sending 110 %. Occupancy is now the measured
+  base PLUS an estimate of everything appended since that measurement, and the
+  status gauge reads the same number the trigger does, on every turn — including
+  the turns where a guard declines to compact, which is exactly when the number
+  matters most.
+
+  **The recent turns can no longer wedge a run.** Every previous rung of the
+  ladder operated on the head of the conversation. When the RETAINED tail was what
+  did not fit, each rung reported success or "nothing to drop" while the history
+  stayed un-sendable and the run died. Oversized `tool_result` bodies inside the
+  retained turns are now clipped, oldest first, only when the alternative is a
+  request the provider cannot accept. Nothing is removed — no message, no
+  tool-call id, no role — so a valid history stays valid. Every clip is announced
+  in the text the model reads, at the exact place the data went, and on the card,
+  in `/compact status`, on the JSON stream and in the archive.
+
+  **Sub-agents compact their own history** (`compaction.subagents`, default
+  `true`). A `task` child is bounded by 24 turns and fifteen minutes, and when one
+  overflowed its entire history was discarded and the dispatch reached the lead as
+  one partial sentence. Children run under tighter bounds than the lead: two
+  retained turns, at most two compactions, the fast tier preferred, and failure
+  forced to truncate because nobody is watching a background worker. No transcript
+  card — the lead's transcript describes the lead's context. Set
+  `compaction.subagents: false` to restore the previous behaviour exactly.
+
+  **A compaction is no longer irreversible** (`compaction.archive`, default
+  `true`). The dropped messages are written verbatim to `~/.aragon/compaction/`
+  before the splice is adopted. `/compact history` lists this run's archives,
+  `/compact show <n>` prints one's metadata and summary — never message bodies.
+  The directory is shared by every `aragon` on the machine, so files carry a
+  per-process run id and retention is two-tier: 20 files per run by count, plus a
+  seven-day sweep across all runs, so one session can never prune a live
+  neighbour's work. Writing is best-effort: a failure is logged and never affects
+  the compaction. This is deliberately **not** undo — restoring an over-full
+  history restores the condition that triggered the compaction.
+
+  **The long moment is legible.** The live card counts elapsed seconds and says
+  `esc to cancel`, which `AgentController.abort()` has always made true and no
+  surface has ever stated. A compaction that could only clip says so and names no
+  model, because none was called.
+
+  Two new config keys (`compaction.subagents`, `compaction.archive`), two new
+  environment variables (`ARAGON_COMPACTION_SUBAGENTS`,
+  `ARAGON_COMPACTION_ARCHIVE`), two new `/compact` verbs (`history`, `show <n>`),
+  and two new settings-screen rows. No flags were added.
+
+- **Context compaction — a long run no longer dies when the window fills, and it
+  is ON by default.** A session's conversation only grows: nothing anywhere
+  removed a message, so a 40-turn tool-heavy run had exactly one ending available
+  to it once the context window filled — an HTTP 400 and a banner reading
+  *"Context length exceeded — start a new conversation with /reset."* The
+  product's answer to running out of context was "throw away everything you were
+  doing."
+
+  Now, at **90 %** occupancy, the run pauses at a turn boundary, hands the older
+  part of the history to a summarizer, and swaps it for a structured record. Your
+  original task stays **verbatim**, the most recent turns stay **verbatim**, and
+  the run continues in the same loop with the same tools, the same todo list and
+  the same working directory.
+
+  **This changes behaviour for existing users, in the direction of not dying.**
+  `--no-compaction`, or `compaction.enabled: false`, restores the old behaviour
+  exactly: no manager is constructed, nothing is subscribed, no registry is
+  allocated, and a full window ends the run as it did before. Both flag forms are
+  declared, so a stored `false` survives a run that passes no flags.
+
+  Three ways it fires. **Proactively** on measured occupancy — the case that
+  happens almost every time. **Reactively** when the provider returns
+  `context_overflow` anyway, because the model is not in the static window table
+  or you are pointed at a proxy in front of a smaller one: one compaction, one
+  re-send, run continues. And **manually**, via `/compact`, which runs
+  immediately when idle and queues to the next turn boundary when running.
+
+  The trigger has two terms. 90 % of a 200 k window leaves 20 k, which is plenty;
+  90 % of a 32 k window leaves 3.2 k, which is less than a single `max_tokens` of
+  8192 — the request is already impossible. So compaction also fires whenever the
+  remaining headroom drops below one full response.
+
+  **The default summarizer is your own model, and that costs real money.**
+  `compaction.useFastTier` is `true`, but it is a *preference*: the fast tier is
+  off unless you configured it, so out of the box the first compaction is a
+  ~30 k-token call on the session's own model. It happens at most 5 times per
+  run, the alternative is losing the run entirely, and the spend is visible — in
+  the session total, priced with the summarizer's own table, and reported as
+  `pricing unknown` rather than `$0.00` when that table has never seen the model.
+  Configure the fast tier to make it cheap.
+
+  Summarization fails sometimes. The ladder is one retry (on your **main** model
+  if the first attempt used the fast tier), then a **truncation** that says so on
+  the card, on the JSON event stream, and inside the block the model itself
+  reads. `compaction.onFailure: "stop"` makes the other trade.
+
+  Four guards stop it looping on your money: never two on consecutive turns,
+  never more than 5 per run, a progress requirement, and a self-disable after two
+  no-progress attempts with a notice naming the cause and the fix. The guards are
+  **trigger-aware** — the reactive path stays armed even after a self-disable,
+  because a compaction answering a request the provider has already refused is
+  the difference between a live run and a dead one.
+
+  Three UI surfaces: `Compacting context…` on the activity row while it runs, a
+  `compacting` chip plus threshold-aligned gauge colours as state, and a
+  transcript card with the before/after counts, the summarizer, the duration and
+  an expandable summary as the record. The gauge falls **immediately** rather
+  than waiting for the next turn. `/compact status` is the guaranteed surface at
+  any width.
+
+  New: `compaction` config section (6 keys), `--compaction` / `--no-compaction` /
+  `--compaction-threshold`, `ARAGON_COMPACTION*` (3 vars), `/compact` (6 forms),
+  three settings-screen rows, a `compaction` event on the `aragon exec` JSON
+  stream (`EXEC_SCHEMA_VERSION` **unchanged** — adding an event type never bumps
+  it), and two stderr lines under `-p`.
+
+### Changed
+
+- **The context gauge now counts cache tokens**, through the same
+  `occupiedTokens` the compaction trigger reads. Anthropic reports `input_tokens`
+  *excluding* cached tokens and puts them in two separate fields, so the old
+  `inputTokens + outputTokens` silently under-reported occupancy for anyone
+  behind a caching gateway. That was already a defect in the number on the status
+  bar; it becomes a much worse one the moment a trigger fires on a different
+  number from the one you are watching.
+
+- **`context_overflow` now names `/compact` alongside `/reset`.** That remedy
+  list was complete until this release, because throwing the conversation away
+  really was the only exit.
+
+### Known limitation
+
+- **A summarizer that is 90 % faithful will, eventually, drop something the run
+  needed,** and no amount of prompt engineering closes that. The design bounds it
+  — verbatim task, verbatim recent turns, a block that tells the model the record
+  is partial, and three surfaces that make it visible so a human can step in —
+  but it does not eliminate it. That is inherent to compaction as a technique, it
+  is the same trade Claude Code and Codex make, and it is preferable by a wide
+  margin to the previous behaviour, which was to lose the entire run.
+
+## 0.6.0
+
+### Added
+
+- **`aragon exec` — a second, machine-facing face on the same binary.** The
+  package already installed from npm and already ran from a terminal, so "it has
+  a CLI" was true. It was nonetheless not *callable* by another program:
+  everything it exposed was shaped for a person, and a build script, a CI job, a
+  web backend or a second agent had to spawn it, capture free text, and guess.
+  There was no way to read the answer without also reading the narration, no way
+  to carry a conversation across two invocations, no way to say "you may read but
+  not write" when there is no human to answer `--confirm`, and no way to bound
+  what a run may spend.
+
+  `aragon exec` adds a **contract** rather than a feature: a versioned JSON event
+  schema on stdout, a session identifier the caller may mint itself, a tool
+  permission policy enforced at the tool boundary, caller-set budgets with their
+  own exit code, and a discovery command so a wrapper can ask what it just
+  installed before depending on it. Everything underneath is reused unchanged —
+  the same controller, the same tools, the same config resolution, the same
+  skills, teams, todo and fast-tier subsystems.
+
+  - `--output-format text | json | stream-json`. `stream-json` is NDJSON: one
+    object per line, `system/init` first, exactly one `result` last, on every
+    exit path including error, budget and signal. `text` is byte-identical to
+    `aragon -p`, because it **delegates** to the same renderer rather than
+    reimplementing it.
+  - `--session-id` / `--resume` / `--continue`, backed by `~/.aragon-agent/
+    sessions/` and an exclusive lock, so two runs cannot interleave and lose a
+    turn. A second run on a held id exits 2 with `session_busy` rather than
+    blocking.
+  - `--permission-mode auto|plan|strict` plus `--allow-tool` / `--deny-tool`.
+    A denied tool is **not registered at all**, so the model never sees it and
+    never plans around it — and the policy reaches subagents, so one `task` call
+    cannot bypass it.
+  - `--max-turns` / `--max-duration`, both aborting through the same path `Esc`
+    uses, both reporting `isError: false` and exit **3**: a budget firing is not
+    a malfunction, but it is not a complete answer either, and a shell script has
+    to be able to tell without parsing JSON.
+  - `--append-system-prompt` / `--append-system-prompt-file`. Appending only;
+    there is deliberately no full `--system-prompt` replacement, because the
+    builtin prompt carries tool discipline, todo planning, the skills catalog and
+    the plan-mode block, and replacing it silently disables half the product.
+  - `--input-format stream-json`, so one process can serve several turns from a
+    pipe. A malformed line is skipped with a non-fatal `error` event; an unknown
+    message type is ignored.
+  - **`aragon sessions`** (`list` / `show` / `rm` / `prune` / `path`),
+    **`aragon info [--json]`** and **`aragon doctor [--json] [--probe]`**.
+
+  Documented in full under
+  [Use it from another project](./README.md#use-it-from-another-project),
+  including an honest Limits section: there is no filesystem sandbox, process
+  exit is terminal on Windows whether or not a `result` arrived, an exec session
+  resumed in the TUI shows no scrollback, and `sessions prune` shares a directory
+  with the TUI's `/save`.
+
+### Changed
+
+- **Exit code `3` is new, and the signal row is now stated as `128 + signo`.**
+  `3` means "stopped by a budget you set" and is reachable only from
+  `aragon exec`. The signal codes are unchanged in behaviour — this CLI has
+  always exited `128 + signo`, so `SIGTERM` has always been `143` — but the
+  README previously simplified all of them to `130`, which is fine for prose
+  about a CLI a human interrupts and not good enough for a wire contract.
+- `interrupt` handling now settles the run before exiting when `aragon exec`
+  owns it, so a wrapper reading NDJSON gets a final `result` on Ctrl-C instead of
+  having to time out to learn the run ended. `aragon`, `aragon -p` and every
+  other subcommand keep the previous bare-exit behaviour byte for byte.
+
+### Unchanged, and asserted rather than claimed
+
+- `aragon`, `aragon -p`, every existing flag, every slash command and every byte
+  of their output. `aragon exec --output-format text` and `runHeadless` are
+  compared for byte equality on **both** stdout and stderr by a regression test,
+  and `createBuiltinTools` with no permission policy is proven by object identity
+  to return the tools it returned before.
+
+## Unreleased
+
+### Project index on Ctrl+I (bundled `project-indexer` skill)
+
+- The `project-indexer` skill now ships **bundled** with the package
+  (`<pkg>/skills/project-indexer`, read-only, alongside `skill-creator`): it
+  generates the `.claude-index/` navigation map and injects the Google-style
+  Clean Code Guidelines section into `CLAUDE.md`, adapted from the standalone
+  Claude Code skill (ask_user instead of AskUserQuestion; body condensed under
+  the 30 KB invocation budget with the four templates bundled as files).
+- New keybinding `Ctrl+I`: opens a Yes/No confirm that names every write
+  target, then dispatches `/skill:project-indexer` through the ordinary
+  command path - the indexing run is a normal agent turn whose tool calls
+  stream into the transcript. While a run is in flight the key warns instead
+  of queueing; with skills disabled or the skill shadowed it explains instead
+  of failing silently.
+- `Ctrl+I` is legacy-encoded as the Tab byte, so the stdin filter's two
+  translators (win32-input-mode on Windows, kitty CSI-u / modifyOtherKeys
+  elsewhere) now emit a distinct NUL-delimited `INDEX_KEY_FRAME` for it -
+  the same framing technique `Shift+Enter` uses. On terminals with neither
+  protocol the key keeps its legacy Tab meaning (completion).
+- The header names the key next to the detail toggle: `^I Index` collapsed,
+  `^I build index` in the `Ctrl+G` expanded state; hidden below 72 columns and
+  while an overlay is open. `/skill:project-indexer` and `/project-indexer`
+  remain available as typed commands.
+
+### Added
+
+- **Drag the mouse to select text, and release to copy it.** Full-screen mode
+  takes the alternate screen, turns on mouse reporting so the wheel can scroll a
+  transcript the terminal no longer knows how to scroll, and draws every row
+  itself. Each of those is right on its own; together they took your terminal's
+  own drag-to-select away and put nothing in its place. The only thing shipped
+  in its place was a one-line notice naming a bypass key that is Shift on some
+  emulators, Fn on others and a preference elsewhere — a workaround, told once,
+  in a line that scrolls away.
+
+  Press the left button and drag: the cells you drag across are highlighted as
+  you go, and releasing copies exactly them. The copy goes out over **OSC 52**
+  first, so it reaches the clipboard of the machine you are *sitting at* rather
+  than the one you are ssh'd into, with the platform clipboard binary attempted
+  alongside it for terminals that ignore the sequence. Neither mechanism is
+  detectable, so the toast names what was sent rather than claiming a success it
+  cannot verify. `/copy` now goes through the same path and gained ssh support on
+  the way.
+
+  The highlight is painted in the frame pipeline rather than in React, so a drag
+  costs one to three re-addressed rows instead of a re-render of the whole tree,
+  and the text that is copied is read from the same mirrored rows the highlight
+  was painted over — what you see is what you get by construction, not by a
+  staleness check that can be wrong.
+
+  Two switches, both reversible without a restart: `/mouse off` hands the mouse
+  back to your terminal for the session (and says plainly that the wheel goes
+  inert while it is off), `/mouse on` takes it back, and
+  `aragon config set mouseSelect false` keeps wheel scrolling while leaving
+  drag-select off — which also stops the CLI asking the terminal for motion
+  reports at all, so that session puts exactly the bytes on the wire that it did
+  before this release. `--no-mouse` remains the switch that captures nothing.
+
+### Changed
+
+- **Scrolling up now actually stops the transcript.** The viewport followed the
+  newest output by holding a *fixed distance from the bottom*, so scrolling up
+  during a run did not pause anything: twelve new rows of output moved the
+  window twelve rows down, and whatever you had scrolled up to read slid off the
+  top while you were reading it. The faster the agent talked, the faster your
+  reading position ran away.
+
+  While you are scrolled up, the transcript is now anchored to the **content**:
+  new output accumulates below the window instead of dragging the window along
+  with it, and the rows on screen hold still. After five seconds with no
+  scrolling the viewport returns to the newest line by itself — but only if
+  something actually arrived while you were paused, so reading a *finished*
+  transcript is never interrupted by a jump to a bottom that has not moved.
+  Every gesture restarts the countdown, a selection drag suspends it, and
+  `aragon config set scrollResumeMs 0` turns the return off entirely
+  (`ARAGON_SCROLL_RESUME_MS` sets it too).
+
+  The number the hint shows means something sharper as a result: it is now
+  literally "rows of newer output below your reading position", which is what
+  `PgDn` traverses and what the status bar's `↑N` has always shown.
+
+- **The "N new lines · PgDn" hint moved inside the input box.** It used to be
+  rendered as the viewport's own last row, which spent a line of your transcript
+  on chrome at exactly the moment the transcript was longest — and, because the
+  hint's presence shrank the viewport, it fed back into the very number it was
+  displaying. It is now a right-aligned chip on the composer's input row, inside
+  the border, where it costs no rows at all. The chip sits in a fixed-width cell
+  so the number gaining a digit cannot re-wrap a draft, it never truncates (the
+  draft wraps first), and unlike the teaching row below the box it neither fades
+  with experience nor disappears on a short terminal: how far behind you are is
+  state, not a tutorial.
+
+### Fixed
+
+- **The working spinner no longer disappears when an ack takes its row.** Steer a
+  run, approve a plan, or press `Ctrl+T` while the agent is working, and the row
+  above the composer traded its spinner for the acknowledgement — and for that
+  ack's full two and a half seconds, nothing anywhere on screen moved. The status
+  bar still read `running` and the tool card still read `running`, so no
+  information had left the frame; only the evidence that the process was alive
+  had, during exactly the long tool calls where "is it stuck?" is the question you
+  are already asking. Consecutive acks extended the window, and steering is
+  something the composer invites on every run.
+
+  That row was made the sole owner of animation in this same release, and
+  everything else on screen holds still for as long as the row is believed to be
+  showing one. Believed, not checked: an ack won the row outright, so the belief
+  and the frame disagreed, and the disagreement was invisible to every test
+  because no component was wrong on its own.
+
+  The ack still owns the row and keeps every word of its text. The spinner now
+  rides beside it as a bare glyph — no phrase, no clock — so there is exactly one
+  animation at every instant of a run, with an ack up or without one: never none,
+  and never the several this release set out to remove. Terminals without Unicode
+  and anyone running with `reducedMotion` come out byte for byte unchanged; they
+  had no animation to lose, and nothing is prefixed to their acks.
+
+- **`/clear` now takes the todo panel with it.** Finish a plan, type `/clear`,
+  and the transcript emptied while the `7/7 done` rail stayed exactly where it
+  was — on a screen you had just asked to be cleared. It stayed until your next
+  real message, which is precisely the pause you take after clearing to decide
+  what to ask next, so essentially everyone met it.
+
+  This was deliberate once. The list is a projection of what the model believes,
+  and `/clear` wipes the transcript rather than the conversation, so the model
+  still believed in the plan and a panel contradicting the model was judged worse
+  than no panel. That reasoning holds for everything the CLI does on its own; it
+  does not hold against you typing the command. An explicit instruction was
+  always this rule's exception — `/todo clear` has used the same door from the
+  start — and `/clear` now goes through it.
+
+  The list itself is cleared, not just the pixels, so `/save`, `/todo status`,
+  `/todo continue` and auto follow-through all agree with the screen: no plan is
+  written into a saved session, nothing offers to continue one, and nothing spends
+  a follow-through on it. `/reset` is unchanged and still the only command that
+  clears the conversation. Nothing is stranded either way — the model's next
+  `todo_write` is a full replacement, so the panel comes back the moment there is
+  a plan to show.
+
+### Added
+
+- **An update that will not start is undone automatically.** The updater shipped
+  in this same release could install a version that fails to *import* — an
+  undeclared dependency, a Node API that moved, a syntax error past the transpile
+  target — and the only recovery was to know the package name and type
+  `npm i -g @aragon-agent/cli@<some older version>` from memory. None of those
+  failures reach any line inside `main()`, so no amount of care there helps.
+
+  The `aragon` command is now a small launcher that runs **before** the CLI. It
+  reads one ~300-byte file, and on every launch except the first one after an
+  auto-update that is all it does — no write at all. But if the version it is
+  about to start is one **this updater installed**, and that version has already
+  exited non-zero twice, it reinstalls the version that was running before the
+  update and marks the bad one so nothing puts it back:
+
+  ```
+  aragon 0.6.0 failed to start twice; rolling back to 0.5.9.
+  ```
+
+  It boots anyway. If the crash was transient, refusing to start would lock you
+  out of a working CLI on our guess; the reinstall is detached and finishes
+  either way. The next session says once what happened and then stops. A **newer**
+  release clears the mark and installs normally — you were never opted out of
+  updates, only out of the loop.
+
+  It never touches a version **you** installed yourself, and it does not try to
+  detect a release that starts and is merely wrong. `aragon update --rollback` is
+  the same mechanism run deliberately, for the case a human spots in one second
+  and a heuristic would get wrong on a bad day. Exit `4` means there was nothing
+  recorded to roll back to.
+
+  **`update.mode: off` disables this too.** The guard is armed only by an
+  auto-install, so a machine that never auto-installs never arms it — the kill
+  switch turns off the whole subsystem, recovery included.
+
+- **A blocked install now says what is actually blocking it.** Every install
+  failure used to print the same line: `update failed · npm i -g
+  @aragon-agent/cli`. On Windows the commonest cause is that the shim which
+  launched us is held open by the shell running it, so that advice hits the same
+  handle and fails identically — the one visible failure was also a wrong
+  instruction. Failures are now classified, and three of them name the real
+  remedy and appear immediately rather than after three strikes, because each is
+  something you can fix and none of them fixes itself:
+
+  ```
+    ! update blocked · close other aragon windows, then: npm i -g @aragon-agent/cli
+    ! update failed · no disk space
+    ! update needs write access · npm i -g @aragon-agent/cli
+  ```
+
+  Network failures are deliberately **not** in that set: you cannot fix the
+  registry and the next check may well succeed, so they stay silent and back off.
+
+- **Updates work behind a corporate proxy.** Node's `fetch` ignores `HTTP_PROXY`,
+  so on a proxied machine the check failed silently forever. When the direct
+  request fails, `aragon` now asks the npm you already have instead, which
+  honours `proxy`, `https-proxy` and your `.npmrc`. It runs *only* after a
+  request that has already failed, so on an ordinary machine it never runs.
+  `/update status` shows `probe: npm` when this is what answered. No credentials
+  are read or forwarded — we run the command you could have typed.
+
+- `/update status` gained `rollback`, `lastgood` and `probe` lines, so what the
+  guard did and what `--rollback` would reinstall are answerable without reading
+  the log.
+
+### Changed
+
+- **`aragon update` exit codes: a blocked install is now `2`, not `1`.** This is
+  the one change here a script can notice. `1` means "try again later" and `2`
+  means "this machine will never auto-update, run the command we printed" — and
+  for `EPERM` / `EBUSY` / `ENOSPC` the second is the accurate one: waiting does
+  not unblock a held file handle or an empty disk. A caller shaped like
+  `aragon update || sleep 300 && retry` changes behaviour. `0`, and `1` for an
+  ordinary failed check or install, are unchanged.
+
+- `bin.aragon` now points at `dist/launcher.js` rather than `dist/cli.js`.
+  `dist/cli.js` is unchanged, keeps its shebang and stays executable, so
+  `npm start` and any script invoking it directly keep working — only the mapping
+  moved.
+
+- **`aragon` keeps itself up to date.** Until now the installed copy was frozen
+  forever: nothing in the CLI ever looked at the registry, so someone who
+  installed 0.4.x eighteen releases ago was still running it and the only signal
+  that anything had moved was a README they were never going to re-read.
+
+  There is now a background updater. At most once every four hours, and at most
+  once per machine per interval, it asks npm what the `latest` dist-tag resolves
+  to; if that version is genuinely newer, stable, not deprecated and runnable on
+  this Node, it installs it silently into the same global installation you
+  already have and says so on **one line above the composer**:
+
+  ```
+    * 0.6.0 installed - restart aragon to apply
+  ```
+
+  **The running process is never mutated in any way it can observe.** The new
+  version lands on disk and takes effect on the *next* launch. A restart in the
+  middle of a turn would destroy a live run, a subagent dispatch and an open plan
+  review, so the line waits and the human decides when.
+
+  The notice does not get its own row — it shares the row the toast stack and the
+  working line already share, behind both of them. A toast is a response to
+  something you did and wins outright; the working line is transient and would be
+  lost if deferred; the update line is persistent and loses nothing by waiting.
+  The practical effect is that **someone who is actually working never sees the
+  updater until they stop.**
+
+  It only ever runs `npm i -g` for one kind of installation: a global one where
+  no project manifest claims to own the version. pnpm, yarn, bun and volta are
+  each detected and get the correct command for *that* manager printed instead
+  of a command run on their behalf; `npx`, a project's `node_modules` and a
+  clone of this repo produce no line at all, because none of them is out of date
+  in any sense the user can act on. A global directory that cannot be written is
+  detected before anything is spawned, and npm exiting `0` into a *different*
+  prefix — which happens on any machine with `prefix=` in `.npmrc` or two Node
+  installations — is caught by re-reading the version from the directory we
+  classified, rather than being believed.
+
+  Every failure is silent: it logs, backs off exponentially, and says nothing
+  until the third consecutive one, at which point it says something actionable.
+  Four config keys (`update.mode`, `update.checkIntervalMs`, `update.registry`,
+  `update.distTag`), the flags `--update` / `--no-update`, the environment
+  variables `ARAGON_UPDATE` and `ARAGON_UPDATE_REGISTRY`, the slash command
+  `/update [status|now|skip|off]`, and a foreground `aragon update
+  [--check] [--to <version>] [--json]` for headless users. `update.mode: off`
+  and `ARAGON_UPDATE=0` are kill switches, and with either set the CLI executes
+  zero lines of update code.
+
+  Auto-update widens the window in which a compromised release reaches users
+  without a human decision. The README says so plainly, along with what is and
+  is not mitigated. No credentials are ever attached to the check.
+
+- **A running command shows what it is doing.** `bash` is the only tool here that
+  can run for minutes — `npm test`, a `git clone`, a training job — and while it
+  ran, its card was one line of text that never changed. Nothing in that frame
+  distinguished a test suite at 40% from a child process blocked forever on a
+  password prompt this CLI will never show. Its card now carries the last eight
+  lines of what the command is printing, plus one footer row that reads
+  `(running)` or, once the child has been quiet for ten seconds,
+  `no output for 45s` — with the seconds advancing.
+
+  A `\r`-driven progress bar (npm, pip, curl, docker) is *one* line rewritten a
+  thousand times, and it renders as **one row counting up in place** rather than
+  a thousand rows of nearly the same string — including the ordinary case where
+  the bar emits no newline at all for its entire run.
+
+  Everything the tail shows is **stripped, not interpreted**: ANSI colour, cursor
+  motion, `\r`, `\b`, tabs and every other control byte are removed at the
+  boundary before a row is stored, so nothing can move a cursor inside a frame
+  the renderer believes it owns. It is bounded twice by construction — eight rows
+  per call, sixteen concurrent calls — so a command that writes 400 MB to stdout
+  costs a fixed number of retained rows.
+
+  It is a **display and nothing else**. The text the model receives is
+  byte-for-byte what it was; `Ctrl+O` does nothing while a command runs, because
+  expansion is a promise about stored content and there is no stored rest of a
+  400 MB stream; `/save` during a run writes no tail, since a resumed one would
+  describe a process that died with the session; and at settle the card is
+  replaced by the authoritative result exactly as before.
+
+  New config key `liveToolOutput`, defaulting to **`true`** — unlike
+  `showThinking` below, this adds information rather than hiding it.
+  `aragon config set liveToolOutput false` (or `--no-live-tool-output`, or
+  `ARAGON_LIVE_TOOL_OUTPUT=0`, or the **Live output** row in `/settings`) turns it
+  off, and with it off no store is allocated and no recorder is attached: `bash`
+  is byte-identical to the build before this entry.
+
+- **The working line stops guessing while a tool runs.** It reads
+  `⠙ Running bash` instead of rotating thinking-words. `running` covers the whole
+  turn, including the minutes the model is idle and a child process is doing the
+  work, so "Pondering" during a build was not calm — it was wrong. One row, no
+  digits, no clock: everything the row already promised still holds.
+
+### Changed
+
+- **Reasoning blocks are now hidden by default.** `showThinking` is a new
+  top-level config key and it defaults to `false`. If you have been reading
+  thinking every day, this is the entry that concerns you: `Ctrl+T` still shows
+  it for the session exactly as before, `aragon config set showThinking true`
+  (or `--show-thinking`, or `ARAGON_SHOW_THINKING=1`) restores the old behaviour
+  permanently, and `/settings` has a **Show thinking** row directly under
+  **Thinking**.
+
+  Nothing is hidden silently. A settled turn that thought leaves one muted row
+  where the block used to be — `✱ thought for 12s · ctrl+t to show` — so the fact
+  that reasoning happened, how long it took, and how to see it are all still on
+  screen. In **inline** mode that row omits the key hint and the toast reads
+  `Thinking shown for new output.`, because a settled entry there has already
+  been printed into your terminal's scrollback and cannot be repainted; the
+  toggle applies to the live entry and to everything drawn after it.
+
+  `thinkingLevel` is untouched and means what it always did — the effort the
+  provider is asked to spend. `showThinking` is only whether the terminal draws
+  what came back. They are one word apart in the config file, which is why the
+  flag is `--show-thinking` rather than the shorter name.
+
+### Added
+
+- **A live working line while a run is in flight.** One row above the composer,
+  carrying a spinner and a word that rotates every four seconds:
+  `⠋ Percolating…`. Take reasoning off the screen and the gap between submitting
+  and the first token is otherwise completely blank, which reads as dead rather
+  than as busy — the two changes are one change.
+
+  It carries no clock and no token count, deliberately: the status bar one row
+  below already renders both, under the same running condition and driven by the
+  same 200 ms ticker this row reuses. Under `reducedMotion` the spinner is a
+  static glyph and the word is chosen once and holds, because rotating text is
+  motion too. An idle frame is unchanged, and so is the viewport's row budget —
+  the line shares the row the toast strip already owned rather than adding one.
+
+- **`write_file` and `edit_file` render a real diff.** Hunks with old/new line
+  numbers, a `+N -M` summary, add/remove/context colouring, and `(truncated)` or
+  a `binary` / `too-large` / `unreadable` note when the old side could not be
+  read. Diffs collapse at 12 rows behind the existing `Ctrl+O` — no new
+  keybinding. A resumed session written by an older build, or a card whose patch
+  the bounded store already evicted, falls back to the previous text preview;
+  nothing renders blank.
+
+  `write_file` previously reported creating a 300-line file as
+  `Wrote 8214 bytes to path`, and `edit_file`'s diff trimmed the common prefix
+  and suffix and rendered *everything* in between as one removed block followed
+  by one added block — for a `replace_all` touching line 40 and line 900, that
+  was 860 removed rows followed by 860 nearly identical added ones.
+
+  **The model's input got smaller, not larger.** The structured patch reaches the
+  screen on a CLI-local channel rather than through the tool result, and the text
+  the model reads gained a 4 000-character ceiling it did not have before (its
+  only previous bound was the executor's 100 KB backstop). A single-region edit
+  is byte-for-byte what it was plus one `@@` header; a multi-region edit is
+  strictly smaller. `write_file`'s result text is unchanged.
+
+### Changed
+
+- **A stream that stops mid-answer is now a loud failure instead of a silent
+  truncated one.** All three provider adapters used to end their SSE loop by
+  yielding "done" with whatever had arrived, so a connection closed cleanly
+  half-way through an answer — a proxy idle cut, a load-balancer timeout, an
+  HTTP/2 GOAWAY — was indistinguishable from a completed reply. You got a
+  truncated answer and were told nothing.
+
+  Each adapter now reports the truncation, which is what lets it be retried and
+  recovered. An **abort** is explicitly excluded and still completes normally, and
+  a model that legitimately says nothing and stops is still a success.
+
+  This is deliberately **not** behind `--no-retry`: an adapter that reports a
+  truncated stream honestly should do so whether or not retry is enabled. A switch
+  that also disabled telling you the truth would be a mute button, not a kill
+  switch.
+
+- **The mouse wheel now only ever scrolls.** A notch scrolls the transcript
+  three rows (a page with `Shift`), or the open overlay when there is one, no
+  matter where the pointer is sitting. Prompt history is recalled by `↑` / `↓`
+  alone, exactly as in Claude Code and Codex.
+
+  **This supersedes the 0.5.0 entry below**, which said the wheel "scrolls what
+  is under the pointer" and stepped prompt history one entry per notch when the
+  pointer was over the input box. That turned out to be wrong in use rather than
+  in principle: the pointer is not an instrument you aim in a TUI — it rests
+  wherever you last left it, which in a chat-shaped app is next to the thing you
+  type into. So the ordinary "scroll back and see what just happened" gesture
+  landed on the composer and replaced the draft with an old prompt, with no
+  error, no undo and nothing visibly connecting the gesture to the outcome. The
+  band that decided this was invisible too: its boundary was a measured row that
+  moved as the toast row, the hint row and the completion popup came and went, so
+  the same gesture could do two different things a second apart.
+
+  Nothing else about the wheel changes — same three rows per notch, same
+  `Shift`+wheel page, same `--no-mouse` / `ARAGON_MOUSE=0` / `mouse: false`
+  escape hatches, same `Shift`+drag for selecting text.
+
+- **A long session stays as fast as a short one.** The TUI used to get slower as
+  a session grew and, past a threshold, stopped making progress at all. That was
+  structural rather than incidental: Ink has no output caching, so every frame
+  re-walked the entire mounted tree, re-measured every text node and
+  re-serialised the frame. The per-frame cost was therefore proportional to every
+  character the session had ever produced — and `transcriptWindow` bounded the
+  entry *count*, which bounds nothing when one entry can be fifty thousand lines.
+
+  Entries outside the visible band are no longer mounted. The viewport renders
+  what you can see plus a couple either side and replaces the rest with two
+  spacers of the right height, so the frame costs what the screen costs.
+  Scrolling, the `↑N` readout and the wheel routing are untouched. Because the
+  horizon is no longer a rendering budget, `transcriptWindow` rises from 300 to
+  1000 and its ceiling from 2000 to 20000; an existing setting is still valid
+  and still clamps.
+
+  Syntax highlighting and markdown parsing are memoised on their input, so a
+  200-line code block is tokenised once rather than thirty times a second for as
+  long as it stays on screen. Only the *parsed* form is cached and never a themed
+  one, so `/theme` still takes effect on the next frame.
+
+  Under genuine load the streaming interval widens — 33 ms up to a 320 ms
+  ceiling — so keystrokes and `Ctrl+C` do not queue behind renders. This is
+  visible rather than silent: a muted **`eco`** chip appears in the status bar
+  while it is in effect, because a user who cannot see why the stream got
+  chunkier will conclude the model got slower. `--no-render-governor`,
+  `renderGovernor: false` or `ARAGON_RENDER_GOVERNOR=0` turns it off;
+  `--max-render-interval 33` flattens the ladder without disabling the mechanism.
+
+- **A live thinking block no longer grows without limit.** While an answer is
+  streaming, visible thinking is clamped to its newest 24 rows behind a
+  `... N earlier lines` marker, and the whole block renders again the moment the
+  answer settles. Thinking is scratch reasoning you read from its newest end, and
+  an unbounded one was the easiest way for a single live entry to fill the
+  viewport on its own.
+
+### Fixed
+
+- **Inline mode could hard-freeze the terminal, and now cannot.** Once a live
+  entry grew taller than the terminal, Ink took a branch that writes
+  `clearTerminal` plus the *entire accumulated session history* on every
+  subsequent frame — bypassing both of its own dedupe gates, and growing without
+  bound for the rest of the session. Any inline user (`--no-fullscreen`,
+  `TERM=dumb`, CI, a terminal under 12 rows or 40 columns) whose model wrote a
+  long answer could reach it. The live region is now clamped below that
+  threshold, with a `... N earlier lines` marker naming what is deferred. Nothing
+  is lost: the entry prints in full into the terminal's own scrollback the moment
+  it settles.
+
+- **The view state no longer grows without a ceiling.** A single assistant answer
+  is capped at 256 KiB in the view (thinking at 64 KiB, streamed tool arguments
+  at 16 KiB) by eliding the MIDDLE, so both the opening and the live cursor
+  survive; a marker states how many characters went. `transcriptRetain`
+  (default 1000) caps the entries kept in memory, and what it removed is counted
+  and reported both by `/perf` and at the top of the exit replay. A model that
+  emits five megabytes no longer makes every later frame a five-megabyte scan.
+
+### Added
+
+- **A fast model tier: a second, cheaper model you can delegate to and be
+  reviewed by.** Off by default — with `fast.enabled: false` the system prompt,
+  the `task` schema, the transcript and the request payloads are byte-identical
+  to a build without it — and it does two things once you turn it on.
+
+  **Delegation.** `task` gains one optional field per subagent, `model: "fast"`.
+  A child marked that way runs on the fast model and is otherwise identical:
+  same tools, same working directory, same `--confirm` gate, same plan-mode
+  gate, same skills ceiling. Delegating to a cheaper model is a *model* choice,
+  never a permission boundary. It is where "not complicated, but very expensive
+  in context" gets paid for — a child that reads nine files and reports four
+  sentences moves the file bodies out of the lead's window and onto a model that
+  costs a fraction as much per token.
+
+  **Periodic review.** Every few turns a small digest of what the lead has been
+  doing goes to the fast model with one question — is this still on track? — and
+  a critique that is not "on track" is injected into the running loop as a
+  `<fast_review>` block, so the lead sees a second opinion while it can still
+  act on it:
+
+  ```
+    fast review #2 · claude-haiku-4-5 · turn 12 · 1.4s
+    You have edited config/schema.ts three times without running the tests.
+  ```
+
+  The digest is built from a bounded ring of recent turns and **never** from the
+  message history: sending a session's accumulated file bodies to the "cheap"
+  model would make a review cost more than the turn it reviews, which is the
+  inversion the whole feature exists to prevent.
+
+  Spend is attributed **per tier** — a fast child's tokens are priced at the
+  fast model's table, not the lead's, and the dispatch report states the two
+  separately. A model the built-in price table has never seen is reported as
+  `unknown` rather than as `$0.00`, because a feature that looks free while it
+  is spending money is worse than one that admits it does not know.
+
+  Turn it on with `/fast model <id>` then `/fast on`, with
+  `aragon config set fast.enabled true`, with `--fast --fast-model <id>`, or with
+  `ARAGON_FAST=1 ARAGON_FAST_MODEL=<id>`. `/fast status` is the guaranteed
+  readout: the resolved model, the budget, and this session's reviews, delegated
+  children, tokens and cost. Eleven `fast.*` config keys, four flags, four
+  environment variables and five settings-screen rows; see the README.
+
+  **The review has a session budget.** A session is not one run — it is one run
+  per message you send — so a cadence bounds the pace and nothing bounds the
+  total. `fast.reviewMaxPerSession` (default **40**, raisable live with
+  `/fast budget <n>`) is the number of reviews a session may *start*; reaching it
+  prints one `info` notice that names the remedy and says delegation is
+  unaffected, and `/fast status` reads `12/40` from the first review rather than a
+  bare count you can only size after the fact. It is denominated in **reviews,
+  not currency**, because the fast tier is exactly where a model the price table
+  has never seen lives — a dollar ceiling would silently never fire for the
+  configuration that most needs one. Delegation is never budgeted: it is spend the
+  lead chose in service of a message you sent, it is already bounded by `team.*`,
+  and it exists to *reduce* total cost.
+
+  **The review also fails fast, so it cannot compete with you.** It calls through
+  its own provider registry with **one** retry rather than the ten
+  `retry.maxRetries` defaults to for the lead, and it observes neither `retry.*`
+  nor `/retry` — that policy says how hard to fight for *your* answer, and
+  inheriting it for an unrequested background call is how an advisory ends up
+  spending a rate-limited provider's quota on itself. Its own wall-clock bound
+  is 20 s.
+
+  Two failure modes are handled out loud rather than silently. If the tier stops
+  resolving mid-session — you switch the main provider and `fast.provider` was
+  inheriting it — `model:"fast"` is downgraded to the main model, **counted, and
+  stated in the report**, because an agent that silently gets an expensive child
+  has no way to learn its cost model is wrong. And three consecutive
+  *misconfiguration* failures disable the reviewer for the session with exactly
+  one warning, while a merely busy provider (rate-limited or overloaded) is
+  treated as transient and never counts — that condition fixes itself, and a
+  session-long silent shutdown over it would not. A review you *cancelled* is
+  neither: Esc, a run that ends first, or quitting leaves a `cancelled` card and
+  no strike against the tier.
+
+  Reviews log under their own `fast` scope rather than under `agent`, so
+  `fast_review_*` records filter cleanly away from the lead's.
+
+- **API calls are retried with backoff, and the waiting is legible.** A transient
+  provider failure — HTTP 429, 529, 5xx, 408, or a dropped connection — used to end
+  the turn. One red notice, an idle prompt, and forty tool calls of work thrown
+  away. It is now retried up to **10 times** with an increasing interval
+  (1 s → 2 → 4 → 8 → 16 → 30 s, equal-jittered, 30 s per wait at most) under a
+  4-minute budget measured from the first failure.
+
+  What you see while it waits is **one line, rewritten in place** — never ten
+  notices — carrying what failed, which attempt this is, and a countdown to the
+  next one:
+
+  ```
+    ↻  Provider overloaded · retry 3/10 in 7s          Esc to cancel
+    ↻  recovered after 3 retries · 12.4s
+  ```
+
+  plus a `retry 3/10 7s` chip on the status bar (`[r3]` on a narrow terminal). The
+  card settles into the transcript as history — `recovered`, `gave up`, or
+  `interrupted` — and shows up in the exit replay, because a turn that spent three
+  minutes retrying is exactly the kind of cost an exported transcript should
+  account for.
+
+  **Esc ends the wait within a frame**, not at the end of the countdown, and it
+  never produces a fabricated error: an aborted request is silent. That is the one
+  detail everything else here depends on, because the provider layer marks an
+  `AbortError` as retryable — so the signal is checked *before* the error, and a
+  user pressing Esc can never be mistaken for a timeout worth retrying.
+
+  Never retried: `auth_error` (401/403), `invalid_request`, `context_overflow`.
+  Ten retries of a bad API key is three minutes of looking broken.
+
+  `Retry-After` is honoured as a **floor**, never a replacement — a `Retry-After: 1`
+  on the ninth retry does not undo eight retries of backoff — and a stated wait
+  longer than a minute is surfaced immediately with the provider's own number
+  rather than silently wedging the terminal.
+
+  A connection that dies **mid-answer** is also covered, and this is the case that
+  needed more than a retry loop: the partial text is discarded, the affected tool
+  cards are removed, and the answer restarts from the top. Nothing is duplicated,
+  because the engine only writes history on a completed turn.
+
+- **`/retry [show|on|off|max <n>]`, `--retry` / `--no-retry`, `--retry-max <n>`,
+  `ARAGON_RETRY`, `ARAGON_RETRY_MAX`, and a `retry` config section** with nine
+  keys (`aragon config set retry.*`). `/retry off` takes effect **in the running
+  session** — subagents included, since they share the lead's provider registry —
+  not only at the next launch. `retry.maxRetries: 0` is the kill switch and is a
+  real value at every one of those entry points. `/settings` gains one read-only
+  `API retries` row.
+
+- **A subagent's retry is visible on the team card.** A child in backoff reports
+  `retry 3/10` on its own row. Children inherit the retry policy for free, but
+  their events never reach the lead's view, so without this a subagent spending
+  three minutes waiting looked exactly like a hung dispatch.
+
+- **Retry lines in `-p` mode go to stderr, never stdout.** `[retry 3/10]
+  overloaded — waiting 7s` and `[retry] recovered after 3 retries` are on stderr,
+  so `aragon -p … | jq` stays valid; a discarded partial answer is announced there
+  too, because a piped consumer otherwise cannot tell a replay from a model that
+  repeated itself. They are not suppressed by `--quiet`: a three-minute pause with
+  no explanation is not a compact run, it is a hang.
+
+- **`/perf`.** Reports what the renderer is actually doing: the frame interval
+  and the last commit cost, how many entries exist / are retained / were dropped
+  / are mounted, height-cache and render-cache occupancy, and the viewport
+  geometry. `/perf reset` clears the caches and returns the governor to its
+  fastest rung. Silent degradation is the one thing a claim of robustness cannot
+  afford, and this is the place that names it.
+
+- **Three configuration keys**, each with a flag and an environment override:
+  `transcriptRetain` (`--transcript-retain`, `ARAGON_TRANSCRIPT_RETAIN`),
+  `renderGovernor` (`--render-governor` / `--no-render-governor`,
+  `ARAGON_RENDER_GOVERNOR`) and `maxRenderIntervalMs`
+  (`--max-render-interval`, `ARAGON_MAX_RENDER_INTERVAL_MS`). No migration and no
+  config version bump: a file written before they existed loads unchanged.
+  `transcriptRetain` is raised to `transcriptWindow` when the two conflict —
+  retaining less than you can scroll to would make part of the horizon
+  unreachable — and a startup notice names both numbers, because overriding a
+  value you typed and saying nothing is the one thing none of this may do.
+
+- **Following the plan through, on a budget you can do arithmetic on.** A run
+  that ended with steps outstanding used to print one sentence and stop. It now
+  asks three questions the CLI already knew the answers to — did you abort it,
+  did it error, did the list move — and does one of three things.
+
+  With the default `notify` the sentence is unchanged, except that a run **you**
+  cancelled with `Esc` now says nothing at all: being told what you just
+  interrupted is the CLI answering a question nobody asked.
+
+  With `/todo follow auto` it continues the plan itself, after announcing
+  `Continuing with 4 remaining steps in 3s - Esc to stop.` Auto-continuation was
+  refused last release as "an unbounded cost loop wearing a helpful hat", and
+  that judgement stands — what changed is that the loop is now bounded first.
+  A continuation that completes no step buys exactly one more attempt; a single
+  plan is continued at most 25 times however productive it looks; and a re-plan
+  forgives the first counter but never the second, so a model that reshapes its
+  list every turn cannot lift the ceiling. A run that ended with an error never
+  auto-continues in any mode. `Esc` cancels, and so does typing anything.
+
+  Whatever sends the continuation — the timer or `/todo continue` — now sends
+  the numbered list of open steps rather than the words "continue with the
+  remaining todo items", so a plan whose `todo_write` has scrolled out of the
+  context window is still followed exactly. Neither path records that text in
+  your prompt history any more; it is not something you typed.
+
+  Under `-p` there is no grace window, because there is nobody there to use it,
+  which is precisely why the two counters rather than the human are the
+  protection there. `[todo] continuing (3 steps left)` and
+  `[todo] 3 steps unfinished` go to stderr, suppressed by `--quiet`, and an
+  unfinished plan still never changes the exit code.
+
+  New surfaces: `/todo follow <notify|auto|off>`; `--todo-follow <mode>`;
+  `ARAGON_TODO_FOLLOW`; and `todo.followThrough` in `config.json`. The default
+  is `notify`, so a session that does not ask for any of this behaves as it did.
+
+- **A plan strip for inline mode.** Inline mode has no fixed frame and so no
+  rail, which left every pipe, every `CI=1`, every `TERM=dumb` terminal and
+  every window under twelve rows able to see `todo 3/7` but not *which* step was
+  running. There is now a one-row strip under the transcript —
+  `todo 3/7  >  Adding the rail to AppShell        +2 done` — windowed on the
+  same step the rail would have anchored on. It truncates rather than wrapping
+  and is exactly one row at every width, so the composer does not move.
+  `--no-todo-panel` hides it along with the rail.
+
+### Changed
+
+- `npm run typecheck` now covers the test tree in both packages. `tsconfig.json`
+  excludes `__tests__` so the build never emits fixtures, and vitest transpiles
+  without typechecking — between the two, a widened required field could break
+  ten test files while `npm run build` and `npm test` were both green. Twelve
+  such errors had accumulated; all twelve were fixture drift and all twelve are
+  repaired, with no suppressions and no production type relaxed to accommodate a
+  fixture.
+
+- **TODO planning, and a rail to watch it in.** For work with three or more
+  distinct steps the agent now keeps a visible plan: it writes a checklist with
+  a new `todo_write` tool, marks exactly one item in progress, and ticks it off
+  before starting the next. In full-screen mode the list is rendered as a
+  right-hand rail about a fifth of the terminal wide, windowed around the step
+  being worked on rather than sliced from the top. Ask for something small and
+  nothing appears — the prompt tells the model not to plan work it can finish in
+  one or two steps, and a one-item list against an empty plan is refused
+  outright, so a session that never plans has byte-identical layout to a build
+  without this feature.
+
+  The list is a projection of what the model believes, and that single rule
+  decides everything else: `/clear` keeps it (the conversation is still there),
+  `/reset` drops it, `/save` writes it and `/resume` restores it — and resuming
+  a session that carries no list CLEARS the current one rather than leaving a
+  plan on screen whose conversation has just been replaced. A finished plan
+  survives to the start of your next message so you get to see `7/7 done`; an
+  unfinished one survives "continue" and "now do the rest", then expires after
+  three unrelated turns rather than holding a fifth of the screen for the rest
+  of the session. A run that ends with items left over produces a notice and
+  `/todo continue`; nothing auto-continues on your behalf.
+
+  "One step at a time" is enforced by structure rather than asked for in a
+  prompt: any payload with two in-progress items is repaired down to one, and a
+  list with none promotes the first unfinished item. Whatever the model sends,
+  the panel has exactly one current row.
+
+  New surfaces: the `todo_write` tool; `/todo [status|on|off|panel on|off|clear|
+  continue]`; `--todo` / `--no-todo` and `--todo-panel` / `--no-todo-panel`;
+  `ARAGON_TODO=0`; and `todo.enabled` / `todo.panel` in `config.json` (a third
+  key, `todo.followThrough`, arrived with the follow-through entry above).
+  `panel` is
+  separate from `enabled` on purpose — a screen-reader user wants the planning
+  discipline without the column, and the system prompt varies one sentence to
+  match. Under `-p` the plan goes to stderr as `[todo] 3/7 <step>` lines,
+  suppressed by `--quiet`. Below 80 columns, behind an overlay, in inline mode
+  or on a short terminal the rail is not mounted and the counter appears in the
+  status bar instead; the transcript keeps at least 62 columns at every width.
+
+- **Output token limits that cannot themselves fail a run.** Every model and
+  every provider now defaults to 64000 output tokens, and that number is
+  configurable from the TUI, from a flag, from `ARAGON_MAX_TOKENS`, and by hand
+  in `~/.aragon-agent/config.json`. Leave it alone and it stays 64000.
+
+  What is new is that the number is an *ambition*. `gpt-4o` caps at 16384 and
+  `gemini-1.5-pro` at 8192; sending 64000 to either is an HTTP 400 that ends the
+  turn with a raw provider string in the transcript. The cap is now clamped down
+  to whatever the target model actually accepts, from a per-model table plus
+  whatever the provider's own model list reports. For a model nothing knows — a
+  proxy, a `--base-url` endpoint, something that shipped last week — the request
+  goes out at 64000, and if the provider objects, `aragon` reads the real ceiling
+  out of the error text, repairs the request, sends it exactly once more, and
+  remembers the answer for the rest of the session. One warning line; the run
+  continues.
+
+  Two combinations that were a guaranteed 400 on every request are fixed with
+  it: `--thinking xhigh` asks for a 65536-token budget while Anthropic requires
+  `max_tokens` to be strictly greater than it, and OpenAI's reasoning family
+  (`o1`, `o3`, `gpt-5`) rejects the `max_tokens` parameter name outright in
+  favour of `max_completion_tokens`.
+
+  New surfaces: `/max-tokens [<n>|auto|default]`, which with no argument reports
+  the setting *and* the effective cap for the live model; an `Effective:` line
+  under the settings screen's `Max tokens` field, so a value the model will not
+  honour is visible before you save it; `--max-tokens auto`; and
+  `aragon config set maxTokens <n|auto|default>`, which echoes the value that was
+  actually stored rather than the one you typed.
+
+- **Team subagents.** The agent can now hand independent parts of a job to up to
+  five short-lived subagents that run in parallel, then read one combined report
+  of what each of them found. It decides for itself whether to delegate, so a
+  session that never calls the new `task` tool is byte-for-byte the session that
+  existed before. On by default; `--no-team` or `team.enabled: false` turns it
+  off, and `aragon config set team.maxSubagents <n>` sets the width (1-10; the
+  10 is a hard ceiling no config file can raise).
+
+  A subagent is a full agent with the same tools and working directory but no
+  memory of your conversation, so its brief has to be self-contained. It cannot
+  delegate further and it cannot ask you anything. Plan mode and a skill's
+  `allowed-tools` ceiling both reach one level down, so delegating is never a
+  way around either; `--confirm` queues subagent confirmations one at a time and
+  labels each with the subagent that is asking.
+
+  Subagents can message each other with `team_send` / `team_wait`, delivered by
+  attaching the message to the recipient's next tool result rather than
+  interrupting it. The limits are deliberate — 6 messages each, one every 15
+  seconds, 24 per dispatch — and every refusal is a plain result naming the
+  limit rather than an error, because an error invites the retry the limits
+  exist to prevent.
+
+  While a dispatch runs, a roster sits between the transcript and the composer
+  and the status bar carries `agents 3/5`; both disappear when it finishes and
+  the transcript keeps a card you can expand with `Ctrl+O`. `Esc` aborts every
+  subagent and returns the partial report. Under `-p` the transitions go to
+  stderr and the subagents' tokens are folded into the `[usage]` footer, so the
+  cost readout cannot under-report what the team spent.
+
+  Full documentation, including the limits table and how team mode interacts
+  with plan mode, skills and `--confirm`, is in the README.
+
+- **The subagent roster now shows what each subagent is doing, and shows the
+  ones that are working.** Where every row used to say `thinking` for minutes at
+  a time, it now says `bash: npm test -w cli`, `read: src/api/routes.ts`,
+  `waiting for a2`, or the tail of the sentence the subagent is currently
+  writing. On a narrow terminal the same row falls back to a file name and a
+  program name rather than wrapping, and the elapsed clock keeps its place at
+  every width.
+
+  The five visible rows are now the five that matter: whatever is running comes
+  first, then whatever is queued, then the most recent results. Above five
+  subagents the old panel showed the five that had finished first and hid
+  everything still working behind `+3 more`, which is the opposite of what the
+  roster is for. `+3 more` now says how many of the hidden ones are still going.
+
+- **A subagent no longer waits two minutes for a message nobody can send.**
+  `team_wait` now checks whether anyone could actually answer before it blocks:
+  if the teammate it named has already finished, does not exist, or is itself
+  waiting in a way that cannot be released in time, it returns straight away and
+  says which. Nothing is refused that could still have been answered — including
+  by a third teammate further down the chain — and the subagent is told what to
+  do instead rather than handed an error.
+
+- **A subagent whose very first request dies on a rate limit is retried once.**
+  Only a genuine cold start qualifies: no turn completed and no tool ran, so
+  there is nothing to undo. It restarts after two seconds, the roster says
+  `starting (retry)`, and the report notes `retried 1x`. Pressing `Esc`, a
+  timeout, or any subagent that had already begun work is never retried.
+
+### Changed
+
+- **`"maxTokens": null` in `config.json` now means AUTO** — use each model's own
+  ceiling, never above 64000. It previously resolved to 64000 by accident (the
+  layered resolver skipped `null`), and there was no way to express AUTO through
+  any supported write path. If you have a `null` there today, the only difference
+  is a downward clamp on models whose real ceiling is below 64000 — the direction
+  that prevents failures rather than causing them. Absent still means 64000, and
+  an explicit number still means that number.
+
+- **`config.json` holds settings again.** The prompts you type are no longer
+  stored in it. They move to `~/.aragon-agent/prompt-history.jsonl` — one JSON
+  object per line, `0600` on POSIX — and the submit counter and the "notice
+  already shown" flag move to `~/.aragon-agent/state.json`. `↑` and `↓` behave
+  exactly as before.
+
+  Two things get better. Opening `config.json` to change your model no longer
+  means scrolling past a transcript of your own typing, and copying it to
+  another machine no longer carries your prompts along. And submitting a message
+  no longer rewrites the file that holds your API keys: `config.json` is written
+  when you change a setting, and at no other time.
+
+  **Your history and counters are migrated automatically, once, at the next
+  launch.** Nothing is printed, because nothing changes for you. The dead
+  `recentModels` key — which nothing has ever written — is dropped.
+
+  If you go back to 0.5.x after upgrading, that version does not know about the
+  new file and `↑` will look empty. Nothing has been lost: the entries are still
+  in `prompt-history.jsonl`, and upgrading again brings them back.
+
+### Added
+
+- **`aragon history`** — `path`, `list` (`-n`, `--json`) and `clear --yes` for
+  the prompt history, alongside the existing `aragon logs` commands. Entries are
+  stored verbatim, so `clear` is how you erase them; `aragon config set
+  historyEnabled false` stops new ones being recorded without deleting or hiding
+  what is already there.
+
+- **The mouse wheel scrolls what is under the pointer.** In full-screen mode the
+  wheel now scrolls the transcript three rows per notch (a page with `Shift`),
+  scrolls an open overlay when the pointer is over one, and steps prompt history
+  one entry per notch when the pointer is over the input box — the same code
+  path `↑` / `↓` use, so it inherits the rule that an in-progress draft is never
+  overwritten. A slim rail on the viewport's right edge shows where you are in
+  the conversation.
+
+  **This fixes a destructive bug.** Previously the wheel recalled prompt history
+  no matter where the pointer was, because the alternate screen makes terminals
+  translate each notch into arrow keys. Wheeling while composing a message
+  silently replaced the draft with an old prompt, and nothing scrolled.
+
+  **Selecting text with the mouse now requires `Shift`+drag** in most terminals.
+  That is the standard convention for "an application is tracking the mouse" and
+  it applies to everyone, including people who never scroll with the wheel — so
+  the first session says so once, in the transcript, and never again. Turn the
+  whole feature off with `--no-mouse`, `ARAGON_MOUSE=0`, or
+  `aragon config set mouse false`; the wheel then does nothing rather than
+  reverting to the old behaviour. Inline mode (`--no-fullscreen`) is untouched:
+  your terminal's own scrollback already does the right thing there.
+
+  Under `tmux` with `mouse on`, tmux keeps the wheel for itself and the app
+  never sees it. `set -g mouse off` hands it back.
+
+### Changed
+
+- **Your files moved to `~/.aragon-agent/`.** Config, sessions, installed skills
+  and the usage counters now live in one directory in your home — on Windows
+  `C:\Users\<you>\.aragon-agent\`, and the same path on macOS and Linux. They
+  used to be split across `%APPDATA%\aragon-agent-nodejs\Config` and
+  `%LOCALAPPDATA%\aragon-agent-nodejs\Data` (or the XDG equivalents), which are
+  correct locations and terrible answers to "where do I change the model?".
+
+  **The move happens once, automatically, on first launch, and prints a line on
+  stderr when it does.** Nothing is deleted: `config.json` is COPIED rather than
+  moved, and the old directories are left where they are, so reinstalling 0.5.0
+  remains lossless. `aragon config home` prints the new location and names the
+  old ones. Once you are satisfied, you can delete them yourself.
+
+  `ARAGON_HOME` relocates the whole tree. It has to be a real environment
+  variable — the path is resolved before any `.env` is read, so putting it in a
+  `.env` will not work. `aragon config home` always prints what is actually in
+  use.
+
+### Added
+
+- **Logging.** Records go to `~/.aragon-agent/logs/aragon-<date>.log` as JSON
+  Lines: startup and exit, config writes, turn boundaries with token counts,
+  tool names with durations, skill loads, and every uncaught error. Files rotate
+  by day and by size (`log.maxFileBytes`, default 5 MiB) and the oldest beyond
+  `log.maxFiles` (default 10) are deleted.
+
+  **`info`, the default, records no conversation content**, so attaching a log
+  to a bug report is safe. `--log-level debug` (or `--verbose`) adds prompts,
+  replies and tool arguments truncated to `log.previewChars`; `trace` adds them
+  in full. Those two write what you typed — check before sharing one.
+
+  API keys are removed at every level by two independent passes: fields whose
+  name looks like a credential are blanked whatever they hold, and every string
+  is scanned both for known key formats and for the literal keys this process is
+  holding — which is what covers a custom endpoint whose key looks like nothing
+  in particular. `log.redactSecrets false` turns both off and warns you on
+  stdout when you set it.
+
+  Nothing is ever uploaded. There is no network path out of this subsystem.
+
+  New commands: `aragon logs path | list | tail [-n N] [--follow] [--level lv]
+  [--json] | clear --yes | open`, and `/logs` inside the TUI. New flags:
+  `--log-level`, `--verbose`, `--log-file` / `--no-log-file`, `--log-dir`. New
+  environment variables: `ARAGON_LOG_LEVEL`, `ARAGON_LOG_FILE`, `ARAGON_LOG_DIR`
+  — these three, unlike `ARAGON_HOME`, can come from a `.env`.
+
+- **`aragon config get <key>` / `list [--json]` / `edit` / `home`.** `get` and
+  `list` mask every secret, `--json` included. `edit` opens `config.json` in
+  `$VISUAL` / `$EDITOR` (`notepad` on Windows), takes a `config.json.bak` first,
+  and re-parses on exit — an invalid edit is reported immediately, and the
+  backup is named in the message.
+
+- **`/reload`** re-reads `config.json` into the running session, for when you
+  have edited it in another window. Refused while a run is in progress: swapping
+  the model mid-turn would let the second half of an answer come from different
+  settings than the first, with nothing afterwards able to explain the result.
+
+- **Log level in the settings screen**, alongside provider / model / thinking.
+  The rest of the log settings — and `redactSecrets` above all — stay
+  command-line-only, so the one switch that can put credentials on disk cannot
+  be flipped with an arrow key.
+
+### Fixed
+
+- **A malformed `config.json` no longer resets everything in silence.** A stray
+  comma used to send the model, theme, timeouts, skills and (as far as you could
+  tell) your API key back to defaults, with nothing anywhere explaining it. The
+  fallback is unchanged — a bad config still never stops the CLI from starting —
+  but it now says so, and records why.
+
+- **Ctrl+C now flushes pending log records in every mode.** The signal handlers
+  only existed in the full-screen branch, so `-p`, the inline renderer and every
+  subcommand took Node's default termination, which does not run exit handlers.
+
+### Added
+
+- **Plan mode.** `Shift+Tab` toggles between `BUILD` (everything as before) and
+  `PLAN`, a read-only research-and-design posture. `/plan [on|off|status]` is a
+  first-class equivalent for terminals that swallow `CSI Z`, and `--plan` /
+  `--no-plan` / `ARAGON_PLAN=1` / `planModeDefault` cover the startup case.
+
+  In `PLAN`, `write_file`, `edit_file`, `bash`, `skill_install` and
+  `skill_create` are refused **at the tool boundary**, so read-only is a
+  property of the wiring rather than a promise in a prompt. `bash` is refused in
+  full, `git status` included: a shell string cannot be classified reliably, and
+  a gate that is right most of the time is worse than one that is always right,
+  because you stop trusting the badge. The refusal names `read_file` / `glob` /
+  `grep` as the alternatives.
+
+  Two tools exist only while a human is attached. `ask_user` renders a
+  keyboard-driven wizard - 1-5 questions, 2-4 options each, exactly one marked
+  `RECOMMENDED`, plus an `Other...` free-text option - where pressing `Enter`
+  through the wizard accepts every recommendation. `submit_plan` renders the
+  plan as a scrollable card: `a` approves, `r` sends written feedback (`<-`
+  returns to the card), `Esc` dismisses. **Approving flips the session to
+  `BUILD` immediately and the same run continues into implementation.**
+
+  Dismissing a plan asks for a better one rather than stopping - the agent is
+  told to refine and resubmit - so the way out of a run is `Esc` twice: the
+  first closes the card, the second aborts the run.
+
+  Toggling mid-run is asymmetric on purpose: `build -> plan` applies at once,
+  `plan -> build` is deferred to the end of the run (the status bar shows
+  `PLAN -> BUILD` meanwhile) so a run launched under a read-only guarantee
+  cannot start writing files because of one stray keypress.
+
+  `aragon -p --plan "..."` registers no question or plan tool - there is nobody
+  to ask - but the read-only gate still applies, so it means "tell me how you
+  would do this, and do not touch my repository".
+
+- Three flat config keys: `planModeDefault` (default `false`),
+  `planModeMaxAskRounds` (default `4`, clamped `[1, 10]`) and
+  `planModeHumanTimeoutMs` (default `1800000`, clamped `[60000, 7200000]`). All
+  three are settable with `aragon config set`. Additive, as always: a config
+  file written by an earlier version loads unchanged.
+
+### Fixed
+
+- A `--confirm` prompt no longer races the idle watchdog. Taking longer than
+  `idleTimeoutMs` over `Proceed? (y/N)` used to abort the run with no
+  explanation; every human wait now suspends the watchdog and restarts it in a
+  `finally`.
+
+- Honor an explicit `NODE_TLS_REJECT_UNAUTHORIZED=0` setting in the CLI and the
+  official npm release workflow while hiding only Node's standard insecure-TLS
+  warning. Other process warnings remain visible.
+
+## 0.5.0
+
+### Breaking
+
+- **The package is renamed `@argon-agent/cli` → `@aragon-agent/cli`.** The brand
+  was always meant to be *Aragon*, matching AragonMesh; the missing `a` was a
+  typo that reached the registry. Install with:
+
+  ```
+  npm uninstall -g @argon-agent/cli
+  npm install -g @aragon-agent/cli
+  ```
+
+  `@argon-agent/cli` is deprecated and receives no further releases. **The
+  executable is still `aragon`** — the command you type does not change.
+
+- **Environment variables are `ARAGON_*`, not `ARGON_*`.** There is no
+  compatibility fallback: `ARGON_MODEL`, `ARGON_PROVIDER`, `ARGON_BASE_URL`,
+  `ARGON_THINKING`, `ARGON_MAX_TOKENS`, `ARGON_THEME`, `ARGON_FULLSCREEN`,
+  `ARGON_SKILLS`, `ARGON_SKILLS_DISABLED`, `ARGON_SKILLS_PATH` and
+  `ARGON_REDUCED_MOTION` are all ignored from this release on. Rename them in
+  your shell profile and CI. Provider keys (`ANTHROPIC_API_KEY`,
+  `OPENAI_API_KEY`, `GOOGLE_API_KEY`, `GEMINI_API_KEY`) are untouched.
+
+- **The project-level skills directory is `.aragon/skills`, not
+  `.argon/skills`.** This one lives in your repository, so rename it yourself:
+
+  ```
+  git mv .argon .aragon
+  ```
+
+  A config file that already persisted the old default keeps working — an
+  explicitly saved `skills.projectDirs` still wins over the default — but the
+  shipped default is now `['.aragon/skills', '.claude/skills']`. Update it with
+  `aragon config set` if you want the new default.
+
+### Added
+
+- **Automatic migration of your user-level state.** On first run, 0.5.0 moves
+  the `env-paths` config and data directories from `argon-agent` to
+  `aragon-agent` — your API key, sessions and installed skills come with it, and
+  each installed skill's `.argon-skill.json` manifest is renamed to
+  `.aragon-skill.json` so `skills update` and `skills uninstall` still recognise
+  it. The migration prints one line to **stderr**, runs once, and is skipped
+  entirely if the new directory already exists. If it fails for any reason the
+  CLI still starts; the old directory is never deleted when a cross-volume copy
+  was needed.
+
+- The opening wordmark now reads `ARAGON` and is 52 columns wide (was 44).
+  Terminals between 52 and 55 columns show the single-line banner rather than a
+  clipped wordmark.
+
+- **Skills.** A skill is a directory holding a `SKILL.md` (YAML frontmatter +
+  Markdown body) plus optional `reference/`, `scripts/` and `assets/`. Skills
+  reach the model through three levels of progressive disclosure: a name +
+  one-line "when to use" catalog in the system prompt (≤ 6 000 bytes total),
+  the full body only when the model calls the `skill` tool, and bundled files
+  only when it reads them with `read_file` / `bash`. A skill therefore costs a
+  line of context rather than its whole body. How many fit in Level 1 depends on
+  description length and language — roughly 30 with terse English descriptions,
+  roughly 10 with Chinese ones at the character cap; past that the catalog
+  truncates and `skill_find` covers the remainder.
+- Four model-facing tools: `skill` (load), `skill_find` (search installed
+  skills), `skill_install` (install from a local directory, git repo, or https
+  `SKILL.md` / `.zip`) and `skill_create` (author a new one from a procedure the
+  agent just worked out).
+- **`aragon skills update <name> | --all`**, with `--dry-run`, `--force` and
+  `--check`. Re-fetches from the source recorded in the install manifest,
+  preserving `installedAt` and recording `updatedAt` / `previousVersion`. Refuses
+  before downloading anything when the installed copy has local edits, and
+  re-validates the source against the *current* `skills.allowedHosts` rather than
+  the one in force at install time. Not exposed to the model.
+- **Catalog ranking.** Level 1 entries are ordered by scope, then by how recently
+  and often each skill was used, so truncation drops what you do not use rather
+  than what sorts last alphabetically. With no usage data the order is unchanged.
+- **`skill_find`**, advertised in the prompt only when the catalog actually
+  truncated. Searches installed skills by name, description and keyword; never
+  reaches the network, never returns `activation: manual` skills, and tells the
+  model to ask for a source rather than invent one when nothing matches.
+- **Cross-process install lock.** Install / create / remove / update take a
+  per-directory advisory lock, closing a race in which two concurrent installs
+  could interleave their atomic-replace steps and delete a skill with no error.
+  Self-expires after 60 s; waits at most 5 s; only ever released by its owner.
+- **Load-time integrity.** Discovery re-checks `SKILL.md` against the sha256
+  recorded at install, at no extra I/O. `skills.integrity` selects
+  `off` / `warn` (default) / `strict`.
+- **Local usage counters** at `<data>/skill-usage.json` — skill name, count and
+  timestamp only, never transmitted, disabled with `skills.usageTracking=false`.
+  Automatically injected `activation: always` skills are not counted.
+- Config keys `skills.integrity` and `skills.usageTracking`; `aragon skills list
+  --sort=recent` and `/skills list --sort=recent`.
+- Every usable skill also becomes a slash command: `/<skill-name> [args]`, with
+  `$ARGUMENTS` / `$1..$9` substitution. On a name clash the built-in command
+  always wins and the skill is reachable as `/skill:<name>`.
+- `/skills` management: `list`, `info`, `install`, `remove`, `enable`,
+  `disable`, `reload`, `create`, `trust`, `untrust`.
+- `aragon skills …` non-interactive equivalents plus `path` and `doctor`.
+  Exit codes follow `config set`: `0` ok, `1` run-time failure, `2` usage.
+- Four discovery scopes in ascending precedence: bundled (`<pkg>/skills`), user
+  (`<data>/skills`), project (`<cwd>/.aragon/skills` and, read-only,
+  `<cwd>/.claude/skills` for Claude Code interop), and `ARAGON_SKILLS_PATH`.
+- New flags `--no-skills`, `--skill <name>` (repeatable) and `--skills-yes`;
+  env `ARAGON_SKILLS`, `ARAGON_SKILLS_PATH`, `ARAGON_SKILLS_DISABLED`; config keys
+  `skills.enabled`, `skills.requireApproval`, `skills.catalogMaxBytes`,
+  `skills.bodyMaxBytes`.
+- One bundled skill, `skill-creator`, documenting how to write a good one.
+- New runtime dependency: `fflate` (~30 KB, MIT, zero transitive deps) for zip
+  extraction. `@aragon-agent/core` gains no dependency.
+- **Release tooling for the scope change.** `npm run verify:brand` scans every
+  file that `npm pack` would put in the tarball — README and `dist/` included —
+  and blocks the release on any surviving `argon` reference, with only the four
+  documented exceptions (the two CHANGELOGs and the legacy-state migration
+  module, which hold the old names by definition). `packages/cli` also gains a
+  `prepublishOnly: npm run build`, so a publish that bypasses
+  `publish-latest.ps1` can no longer ship a `dist/` that is out of step with the
+  sources. `deprecate-legacy.ps1` retires `@argon-agent/*` on the registry, and
+  refuses to do so until the replacements are actually published.
+
+### Changed
+
+- The globally installed executable is now `aragon`. The previous `argon` and
+  `argon-agent` aliases are no longer installed.
+- `config/store.ts` now deep-merges the `skills` section on both read and write,
+  the same treatment `apiKeys` already received. Without it a partial patch such
+  as the one `/skills disable` sends would replace the whole section and
+  silently discard `trustedProjectDirs`, `allowedHosts` and `requireApproval`.
+- `slashSuggestions` accepts `-` and `:`, so kebab-case and `skill:`-namespaced
+  commands autocomplete. Previously `/my-sk` produced no popup at all.
+- `AgentController` now has a single `agent.setSystemPrompt()` call site, so a
+  `/cwd` change can no longer drop the skill catalog out of the prompt.
+- With skills off or none installed, the system prompt and the 7-tool array are
+  byte-identical to the previous release.
+- Added a repository-level PowerShell release workflow that automatically
+  updates CLI and Core versions, synchronizes their dependency and lockfile,
+  verifies both tarballs, and publishes them in dependency order.
+
+### Security
+
+- Installing a skill requires an explicit human approval that **probes for a
+  human first and refuses when there is none**. Under `-p` there is no TUI and
+  therefore no approver, so `skill_install` is refused with an actionable
+  message rather than writing a third-party directory unattended. Use
+  `--skills-yes` (or `skills.requireApproval=false`) to opt out deliberately.
+- Project skill directories require a one-time trust confirmation before they
+  are loaded — cloning an unfamiliar repository does not grant it the ability
+  to inject instructions into your session.
+- Git is invoked only via `execFile` with an argv array and `shell: false`, and
+  owner / repo / ref / subdirectory fragments are whitelisted and refused if
+  they begin with `-`. A leading dash matters even with a pure argv array:
+  `--upload-pack=<program>` makes git execute that program.
+- Zip extraction is streamed and gated on **actually written** bytes, per-entry
+  size, compression ratio and entry count, all of which can abort mid-archive.
+  The size a zip *declares* is written by whoever built it and is never trusted.
+  Path traversal, absolute entries, symlinks and Windows device names are
+  refused, and a rejected archive is removed rather than left half-extracted.
+- Network fetches are `https:` only, restricted to `skills.allowedHosts`,
+  re-checked on every redirect (max 3), size-capped, and refuse IP literals and
+  loopback.
+- Nothing a skill ships is executed at install time, and the executable bit is
+  not preserved. Scripts run only when the model invokes `bash` explicitly.
+  There is no sandbox: approval blocks *silent* installation, it does not make
+  a skill you approved safe.
+
+## 0.3.0
+
+A full-screen TUI release. `@argon-agent/core` is untouched, no new runtime
+dependency is added, and the headless/print contract, exit codes, and
+slash-command semantics are unchanged. The 0.2.0 inline renderer remains
+available in full via `--no-fullscreen`.
+
+### Added
+
+- **Full-screen mode (default).** `aragon` now takes over the terminal's
+  alternate screen buffer — the same mechanism `vim` / `htop` / `lazygit` use.
+  Your shell history is covered, not erased, and comes back untouched on exit.
+  The frame is a fixed `rows - 1` tall, so **the composer and the status bar sit
+  at the physical bottom of the screen from the very first frame**: empty
+  session, long session, mid-scroll, or overlay open.
+- **Brand region.** A six-row gradient ASCII wordmark on a roomy empty session,
+  degrading to a banner, then a single-line bar, then a bare wordmark as space
+  runs out. Row 1 always begins with the brand glyph.
+- **Self-drawn scrolling.** `PgUp` / `PgDn` by the page, `Shift+↑` / `Shift+↓` by
+  the line. The viewport auto-follows new output while pinned to the bottom;
+  scrolling away shows `↑N` in the status bar plus a "N new lines" hint, and
+  submitting a message always re-pins.
+- **Exit replay.** After leaving the alternate screen the session is replayed
+  into the normal buffer as plain, colorless text, so the conversation does not
+  evaporate with the screen. Turn it off with `--no-exit-transcript`.
+- **New flags** `--fullscreen` / `--no-fullscreen` / `--no-exit-transcript`, the
+  env var `ARGON_FULLSCREEN=0|1`, and the config keys `fullscreen`,
+  `exitTranscript`, and `transcriptWindow`.
+- **Automatic downgrade.** A non-TTY stdout, `TERM=dumb`, CI, and terminals under
+  12 rows / 40 columns fall back to the inline renderer. Only the non-TTY gate is
+  un-overridable — `--fullscreen` beats the rest.
+
+### Changed
+
+- The status bar lost its round border and is now exactly one row; it had been
+  spending an eighth of a 24-row terminal framing a single line of text.
+- The toast strip holds a fixed row in full-screen mode, so the transcript no
+  longer jumps as toasts appear and expire.
+- The composer is a rounded frame whose border color reports state (idle, has a
+  draft, running, blurred by an overlay), with the completion popup above it.
+- `Ctrl+L` repaints the frame instead of writing a clear sequence. Ink dedupes
+  identical output at two separate gates, so under a fixed frame the old
+  approach erased the screen and then declined to redraw it.
+- `Shift+↑` / `Shift+↓` no longer recall prompt history — they scroll.
+- Full-screen mode routes `console.*` into the transcript as notices rather than
+  letting Ink write them straight to stdout, which corrupts the frame's line
+  accounting.
+
+### Known trade-offs
+
+- The terminal's native scrollback and mouse wheel do not scroll the transcript
+  in full-screen mode; use `PgUp` / `PgDn`, or run with `--no-fullscreen`.
+- Mouse tracking is deliberately left off: enabling it costs text selection and
+  copy in most terminals.
+- Should a crash ever strand your terminal on the alternate screen, `reset`
+  restores it. Four independent restore paths exist to prevent that: normal
+  exit, `process.exit`, `SIGINT` / `SIGTERM` / `SIGHUP`, and Ink's own signal
+  handling.
+
+## 0.2.0
+
+A design-elevation & HCI-hardening release. No breaking changes: `@argon-agent/core`
+is untouched, no new runtime dependency is added, and the headless/print contract,
+exit codes, config keys, and slash-command semantics from 0.1.0 are unchanged.
+
+### Added
+
+- **Real theming engine.** Distinct `dark` / `light` palettes with an `auto`
+  resolution (`auto → dark`), color degradation across terminal depth
+  (truecolor → 256 → 16 → monochrome), and an ASCII glyph fallback for terminals
+  without Unicode. Honors `NO_COLOR`, `FORCE_COLOR`, `--no-color`, and
+  `TERM=dumb`. Switch live with `/theme <auto|dark|light>` (persisted).
+- **Elegant chrome.** A gradient wordmark header that collapses to a slim sticky
+  bar after the first turn, a first-run **Welcome / empty-state** card, and a
+  status bar with a colored **context gauge** (threshold colors), a **tokens/sec**
+  read-out while running, and a contextual hint.
+- **Legible tool output.** Colored unified diffs (green adds / red removes) and
+  per-tool rich previews (`edit_file`/`write_file`/`read_file`/`bash`/`list_dir`/
+  `glob`). Long previews collapse to 8 lines and expand with `Ctrl+O` or
+  `/expand [n]`.
+- **Autocomplete.** A slash-command palette and `@file` path completion popup
+  (`Tab`/`→` to complete, `Up`/`Down` to move, `Esc` to close).
+- **Full line editing.** Home/End, word jump/delete, kill-to-start/end, and
+  multi-line vertical cursor navigation, with prompt-history recall at the edges.
+  Unknown control/escape sequences are dropped instead of inserted.
+- **Reliability under load.** Ink `<Static>` finalization for settled entries and
+  a streaming delta coalescer, so long/fast sessions stay flicker-free.
+- **Calm feedback.** Ephemeral, auto-dismissing **toasts** for transient acks;
+  the transcript keeps only durable content (messages, tool cards, run errors).
+- `reducedMotion` config (also `ARGON_REDUCED_MOTION`; implied by `--no-color`)
+  replaces spinners with a static glyph.
+
+### Changed
+
+- History scrollback now uses the **native terminal scrollback** that `<Static>`
+  prints into (this supersedes the never-implemented `PgUp`/`PgDn` promise).
+
+## 0.1.0
+
+Initial release — a Claude-Code / Codex-style interactive terminal UI (TUI) on
+top of `@argon-agent/core`.
+
+### Added
+
+- **Interactive TUI** (React + Ink): gradient header, streaming transcript with
+  markdown + syntax highlighting, collapsible thinking blocks, live tool-call
+  cards, a status bar (model / context % / tokens / cost / elapsed), and modal
+  overlays (settings, model picker, help, confirm).
+- **Built-in toolset** wired into the core agent at full permission:
+  `read_file`, `write_file`, `edit_file`, `list_dir`, `glob`, `grep`, `bash`.
+- **Layered configuration** (defaults → user file → env/`.env` → flags) with
+  live in-TUI editing of provider / model / base URL / API keys / thinking level
+  / max tokens. Secrets are masked and the config file is written `0600` on POSIX.
+- **One-shot / print mode** (`argon -p`) for scripting and piping, plus
+  `argon config` / `argon config set` / `argon config path` / `argon models`.
+- **Session control**: abort (Esc), steer-while-running, follow-up queue,
+  double-Ctrl-C exit, `/save` / `/resume` JSON sessions.
+- **Timeout invariant** (`idleTimeout ≥ toolTimeout`) so a long single tool run
+  is never killed by the idle watchdog.
+- **Never-silent failures**: pre-flight API-key validation, error-StreamEvent
+  rendering, and a guard that surfaces a swallowed agent error instead of a
+  blank no-op (headless exits `1`/`2` accordingly).