@aragon-agent/cli 0.6.15 → 0.6.17

package/README.md

@@ -409,7 +409,7 @@ signal ends the run (`130` SIGINT, `143` SIGTERM, `129` SIGHUP).
 | `/team [on\|off\|max <n>]` | Team subagents: status and the live roster, the on/off switch, the fan-out width |
 | `/todo [on\|off\|panel on\|off\|follow <mode>\|clear\|continue]` | Todo planning: status, the on/off switch, the rail, follow-through, and picking an unfinished plan back up |
 | `/retry [show\|on\|off\|max <n>]` | API retry: the effective ladder, the on/off switch, the retry count. Takes effect **in the running session**, children included |
-| `/fast [on\|off\|model <id>\|provider <id>\|same\|review <n\|off>\|delegate on\|off]` | Fast model tier: status and this session's totals, the on/off switch, which model, the review cadence |
+| `/fast [on\|off\|model <id>\|provider <id>\|same\|review <n\|off>]` | Fast model tier: status and this session's totals, the on/off switch, which model, the review cadence |
 | `/compact [status\|on\|off\|threshold <n>\|keep <n>\|history\|show <n>\|<instructions>]` | Context compaction: occupancy and this session's totals, the on/off switch, when it fires, how much it keeps. With free text, compacts now and tells the summarizer what to pay attention to |
 | `/context` | What the context gauge is showing: occupancy split into measured and estimated, the window and **where it came from**, how stale the measurement is, compaction's actual state, and what the session-spend readout includes |
 | `/bg [list\|logs <id> [n]\|stop <id\|all>\|status]` | Background services: what is running, its log tail, and how to stop it. `list` is the default |
@@ -433,7 +433,7 @@ completes, `Up` / `Down` moves the selection, `Esc` closes it.
 | Key | Action |
 | --- | --- |
 | `Enter` | 空闲时发送，运行时加入队列；收到 Core 已纳入会话的接收凭据后才移除 Queue 提示 |
-| `Ctrl+J` / `Alt+Enter` / `Shift+Enter` | 插入换行。Ctrl+J 已通过输入链路测试；另外两种组合键需要宿主发送独立序列，Alt+Enter 可能被宿主拦截。参见 [终端设置](#terminal-setup-for-shiftenter) |
+| `Ctrl+J` / `Alt+Enter` / `Shift+Enter` | 插入换行。应用默认请求终端区分修饰键（Windows 用 win32-input-mode，macOS/Linux 用 kitty/modifyOtherKeys），Shift+Enter 在支持的终端开箱即用；不支持的终端用 Ctrl+J。参见 [终端设置](#terminal-setup-for-shiftenter) |
 | `Shift+Tab` | Toggle **plan mode** (`BUILD` <-> `PLAN`). Equivalent: `/plan`. On Windows this needs a console flag Node < 22.17 does not set; aragon sets it at startup — see [Requirements](#requirements) |
 | `Ctrl+P` | The same toggle, over a channel no Windows console can swallow. Use it if `Shift+Tab` does nothing |
 | `Esc` | Close an overlay or popup; otherwise ask for interruption confirmation. While idle, cancel pending auto-continuation |
@@ -490,10 +490,13 @@ is what it used to do. Inline mode is unaffected either way.
 
 ## Terminal setup for Shift+Enter
 
-输入支持 `Ctrl+J`、`Alt+Enter` 和可独立编码的 `Shift+Enter` 换行。如果终端把
-Shift+Enter 与 Enter 都发送为 CR，应用无法区分；请用 `/terminal-setup` 查看绑定，
-让 Shift+Enter 发送 `\u001b[13;2u`。合并到已有终端配置，不要覆盖其他快捷键；
-终端拦截某个组合键时可使用另外两种方式。参考绑定如下：
+应用现在会主动请求终端区分修饰键：Windows 上推送 win32-input-mode（conhost 与
+Windows Terminal 均支持），macOS/Linux 上推送 kitty disambiguate 与
+`modifyOtherKeys=2`（iTerm2、kitty、ghostty、WezTerm、foot、Alacritty 实现其一），
+并在 stdin 过滤器里把新编码翻译回传统字节，Ink 的按键解析不受影响。上述推送不满足时
+（macOS Terminal.app、旧 Node 下的 conhost、或不支持任何协议的终端）Shift+Enter 仍与
+Enter 同为 CR，此时用 `Ctrl+J` 换行，或按 `/terminal-setup` 给终端加一条让
+Shift+Enter 发送 `\u001b[13;2u` 的绑定（合并进已有配置，不要覆盖其他快捷键）。参考绑定如下：
 
 - **Windows Terminal** — `settings.json` 的 `actions` 与 `keybindings` 分别定义动作和按键；完整示例见下文“输入、待处理队列与底部状态区”。
 - **VS Code integrated terminal** — `keybindings.json`:
@@ -507,10 +510,14 @@ Shift+Enter 与 Enter 都发送为 CR，应用无法区分；请用 `/terminal-s
 - **macOS Terminal.app / legacy conhost** — cannot rebind `Enter`; use
   `Ctrl+J` or `Alt+Enter`
 
-aragon deliberately does not push kitty keyboard mode or `modifyOtherKeys=2`
-onto the terminal itself: mode 2 re-encodes *every* modified key as CSI-u and
-would blind this app's (and Ink's) key parsing for the whole session. One
-user-side binding reaches the same result for the one key that needs it.
+The push and its translation ship as one unit: `input/win32-input-mode.ts` and
+`input/csiu-keys.ts` translate every enhanced encoding back to the legacy
+bytes before Ink parses them, so a translated stream is byte-identical to a
+pre-enhancement one (Enter family excepted: a modified Enter becomes the
+newline frame). Windows additionally requires Node >= 22.17 / >= 24.2 --
+older versions never turn on virtual-terminal input. Opt out with
+`aragon config set keyboardEnhancement false`, `--no-keyboard-enhancement`,
+or `ARAGON_KEYBOARD_ENHANCEMENT=0`.
 
 ## Queued messages while the agent runs
 
@@ -803,9 +810,11 @@ still exits the session outright.
 
 ### Where the mode is shown
 
-The status bar always names a non-default mode (`PLAN`, or `PLAN → BUILD` while
-a switch is pending), including on short terminals and with `--no-hints`. The composer additionally shows a `PLAN` chip on its hint row and
-tints its border, where there is room for it.
+The status bar's first row always names the current mode (`BUILD` or `PLAN`,
+an accent word right after the run state; `PLAN>BUILD` while a switch is
+pending; `B` / `P` / `P>B` when the terminal is narrow), including on short
+terminals and with `--no-hints`. The Ctrl+G details row repeats it, and the
+composer tints its border in plan mode.
 
 ### Switching mid-run
 
@@ -814,7 +823,7 @@ during a run takes effect immediately — you pressed it *because* the agent is
 about to do something. Pressing it toward `BUILD` during a run is **deferred to
 the end of the run**, so a session you launched under a read-only guarantee
 cannot start writing files because of one stray keypress. The status bar shows
-`PLAN → BUILD` while that is pending. Approving a plan is the one exception, and
+`PLAN>BUILD` while that is pending. Approving a plan is the one exception, and
 it is an informed act rather than a stray keypress.
 
 ### If `Shift+Tab` does nothing
@@ -832,7 +841,7 @@ node -e "process.stdin.setRawMode(true);process.stdin.resume();process.stdin.on(
 | It prints | What it means |
 | --- | --- |
 | `09` | The console is swallowing the modifier — the key is arriving as an ordinary `Tab`. Check `node -v`: below 22.17.0 (or on 24.0/24.1) this is expected, and aragon's own attempt to fix it was blocked. See [Requirements](#requirements) |
-| `1b5b5a` | The key reaches the process fine, so something inside aragon is eating it — most often an overlay is open, or the run is still going and the switch is pending (the status bar shows `BUILD → PLAN`) |
+| `1b5b5a` | The key reaches the process fine, so something inside aragon is eating it — most often an overlay is open, or the run is still going and the switch is pending (the status bar shows `PLAN>BUILD`) |
 | nothing | Your terminal, multiplexer or remote-desktop stack never sent it; some never send `CSI Z` at all |
 
 On the first row the wheel is dead for the same reason, and with the `/` palette
@@ -1177,25 +1186,24 @@ children's tokens are folded into the `[usage]` footer.
 ## Fast model tier
 
 A second, cheaper model can be made a first-class citizen of a session. It is
-**off by default** — with `fast.enabled: false` the system prompt, the `task`
-tool schema, the transcript and the request payloads are exactly what they were
-before this feature existed — and it buys you two things once you turn it on.
-
-**Delegation (the agent decides).** `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. This is where "not complicated, but
-very expensive in context" is actually paid for: a child that reads nine files
-and reports four sentences moves ~200 KB of file bodies out of the lead's window
-and onto a model that costs a fraction as much per token.
+**off by default** — with `fast.enabled: false` the system prompt, the
+transcript and the request payloads are exactly what they were before this
+feature existed — and it buys you two things once you turn it on.
+
+**Every `task` child runs the lead's own model.** Per-child fast dispatch
+(`model: "fast"`) was removed for main-agent parity: a delegation is never a
+downgrade, and all the gates (tools, working directory, `--confirm`, plan mode,
+skills ceiling, compaction policy) are the lead's own.
 
 **Periodic review (the harness decides).** Every *N* completed 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 if the answer is not "on track" it is injected
 back into the running loop as a `<fast_review>` block. The lead sees a short
 second opinion mid-run, when it can still act on it, for a few hundred tokens
-instead of a second full-size turn.
+instead of a second full-size turn. **Subagents get the same review**: when the
+tier is live and `fast.review` is on, each dispatched child runs with its own
+reviewer — same cadence, same budget per child, same fail-fast transport — so a
+team is as capable as the agent that dispatched it.
 
 ```
 > refactor the config loader and keep the tests green
@@ -1217,7 +1225,7 @@ separately. If the fast model is not in the built-in price table the number is
 shown as `unknown` rather than as `$0.00` — a feature that looks free while it
 is spending money is worse than one that admits it does not know.
 
-**The review has a session budget, and delegation deliberately does not.** A
+**The review has a session budget.** A
 session is not one run — it is one run per message you send — so a cadence alone
 bounds nothing over a working afternoon. `fast.reviewMaxPerSession` (default
 **40**) is the total number of reviews a session may *start*; reaching it prints
@@ -1226,9 +1234,7 @@ first review rather than a bare count you can only size afterwards. Raise it liv
 with `/fast budget 80`. The budget is denominated in **reviews, not currency**,
 because the fast tier is precisely 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.
+most needs one.
 
 ### Turning it on
 
@@ -1248,9 +1254,8 @@ cannot grow a field mid-session.
 | `fast.provider` | `''` | Empty inherits the main provider. |
 | `fast.model` | `''` | The fast model id. Empty means *not configured* — it never falls back to the main model, so "same as main" stays distinguishable from a typo. |
 | `fast.baseUrl` | `''` | Empty inherits the session's base URL **only** when the provider matches. |
-| `fast.thinkingLevel` | `'off'` | Applied to fast children and to reviews. |
-| `fast.delegate` | `true` | Allow `model:"fast"` on `task`. |
-| `fast.review` | `true` | Run the periodic review. |
+| `fast.thinkingLevel` | `'off'` | Applied to reviews (lead and subagents) and to fast-tier compaction summaries. |
+| `fast.review` | `true` | Run the periodic review — for the lead and for every `task` child. |
 | `fast.reviewEveryTurns` | `5` | Turns between reviews (1–50). |
 | `fast.reviewContextTurns` | `3` | Turns included in one digest (1–10). |
 | `fast.reviewMaxChars` | `280` | Ceiling on an injected critique (80–600). |
@@ -1274,7 +1279,6 @@ reporting one:
 /fast same                 run the fast tier on the main model
 /fast review <n> | off
 /fast budget [n]           the session review budget; bare form reads it
-/fast delegate on | off
 ```
 
 `/fast budget` with no argument is a **read**, so it is answerable while a `task`
@@ -1283,9 +1287,8 @@ like every other `/fast` setting.
 
 **What you see.** A `fast` chip in the status bar's right cluster while the tier
 is live (`fast*` while a review is in flight, dropped below 100 columns), a
-one-line card per review in the transcript, and `n fast` plus a trailing `~` on
-fast children in the team roster. `/fast status` reports the session totals the
-status bar cannot: reviews run, children delegated, tokens and cost. Under `-p`
+one-line card per review in the transcript. `/fast status` reports the session totals the
+status bar cannot: reviews run, tokens and cost. Under `-p`
 nothing extra is written to stdout or stderr — the review spend is folded into
 the `[usage]` footer and the reviews are in the log.
 
@@ -1353,7 +1356,8 @@ Context
 ```
 
 Two lines are worth reading closely. **`Window`** always names where the number
-came from — the model API, the model table, your own `contextWindow`, or the invented
+came from — the model API, the model table, your own `contextWindow`, a
+`model-windows.json` entry, or the invented
 placeholder — because those are indistinguishable everywhere else. And
 **`Compaction`** reports what is actually true of this session, including
 `off - not registered for this session (started with --no-compaction)`; it never
@@ -1361,8 +1365,9 @@ promises a threshold that nothing will act on.
 
 ### `contextWindow`
 
-Window resolution uses your explicit override first, then API metadata from the
-configured endpoint, then the built-in catalog (including known aliases and dated
+Window resolution uses your explicit override first, then per-model declarations
+in `model-windows.json`, then API metadata from the configured endpoint, then the
+built-in catalog (including known aliases and dated
 snapshots). Startup, connection changes, and subsequent prompts refresh discovery
 in the background, with a five-minute cache. OpenAI-compatible endpoints may expose
 `context_length` or `context_window`; Anthropic uses `max_input_tokens`; Google uses
@@ -1384,6 +1389,33 @@ Values are clamped to `[8000, 5000000]`; `auto` (the default) is stored as
 and `/context` still shows that it came from you, which is what makes a wrong
 value findable.
 
+#### `model-windows.json` — one window per model
+
+`contextWindow` describes ONE model: the active one. To pin several at once — a
+main model plus a fast tier plus a gateway alias — declare them by id in
+`~/.aragon-agent/model-windows.json`:
+
+```json
+{
+  "version": 1,
+  "windows": {
+    "kimi-k3": 1048576,
+    "glm-5.3": 1048576,
+    "glm-5.3-flash": 1048576
+  }
+}
+```
+
+A bare map without the wrapper works too. Keys are matched case-insensitively
+after stripping a `models/`-style prefix and a dated snapshot suffix, so
+`Kimi-K3-20260716` still resolves. Entries must be whole tokens between 1024 and
+1e9; at most 512 are kept. The file is read live (mtime-cached), never written
+by the CLI, and outranks API discovery and the catalog while still losing to the
+active model's `contextWindow` override. `/context` reports these windows as
+`from model-windows.json`, and a file that cannot be parsed is ignored with one
+log line rather than trusted. UTF-8 (with or without BOM) and UTF-16 saves are
+both accepted.
+
 ## Context compaction
 
 A long agentic run fills the model's context window. Before this existed, the
@@ -1522,19 +1554,20 @@ refusal is explained. Automatic tail clipping and failure truncation are disable
 
 ### Sub-agents get the same protection
 
-A `task` child is bounded by `team.maxTurnsPerSubagent` (24) and
-`team.dispatchTimeoutMs` (15 minutes) — the same profile as a lead run that fills
-a window. When a child overflowed, its history was discarded and the whole
-dispatch reached you as one partial sentence.
-
-Children now compact their own history, under tighter bounds: **two** retained
-turns instead of four, at most **two** compactions per child instead of five, the
-fast tier preferred whenever it resolves, and failures preserve history. Each
+Children compact their own history **under the lead's own policy**: the same
+`keepRecentTurns`, the same `onFailure`, the same archive setting, the same
+per-run compaction bound, the fast tier preferred whenever it resolves. Each
 child owns its own context meter and counts compaction only after Core accepts
 the candidate. There is no transcript card for a
 child — the lead's transcript describes the lead's context. Turn it off with
 `compaction.subagents: false`.
 
+Since main-agent parity, the turn cap, the per-child wall clock and the dispatch
+wall clock all default to **0 = no limit**: a child runs until it finishes, like
+the lead, with only its own idle watchdog and your Esc as ceilings. Set any of
+`team.subagentTimeoutMs` / `team.dispatchTimeoutMs` / `team.maxTurnsPerSubagent`
+to a positive value to bring a ceiling back.
+
 ### A compaction is no longer irreversible
 
 After a splice is adopted, the dropped messages are written verbatim to
@@ -1923,9 +1956,9 @@ env / `.env` → CLI flags**.
 | `team.enabled` | `true` | Master switch for team subagents. `false` means the `task` tool is never registered — `/team on` in such a session saves the setting for next launch and says so, rather than advertising a tool that is not there. |
 | `team.maxSubagents` | `5` | Subagents per dispatch, clamped to `[1, 10]`. The 10 is the hard ceiling and is enforced twice — on the config value and again on what the model asks for — so hand-editing this file cannot raise it. |
 | `team.maxConcurrent` | `3` | Subagents in flight at once, clamped to `[1, 10]` and capped at `maxSubagents`. A separate knob because provider rate limits are a different constraint from context economics. |
-| `team.subagentTimeoutMs` | `300000` | One subagent's wall clock before it is aborted, clamped to `[30000, 1800000]`. A wedged subagent dies on its own ceiling without taking the dispatch with it. |
-| `team.dispatchTimeoutMs` | `900000` | The whole dispatch's wall clock, clamped to `[60000, 3600000]`. |
-| `team.maxTurnsPerSubagent` | `24` | Runaway-loop cap, clamped to `[4, 100]`. A subagent stopped here is reported as `stopped: turn cap reached`, with whatever it had produced. |
+| `team.subagentTimeoutMs` | `0` | One subagent's wall clock before it is aborted, clamped to `[0, 1800000]`. `0` (the default) = no ceiling — a wedged subagent still dies on its own idle watchdog, exactly like the lead. |
+| `team.dispatchTimeoutMs` | `0` | The whole dispatch's wall clock (and the `task` tool timeout), clamped to `[0, 3600000]`. `0` (the default) = no ceiling; only Esc ends a dispatch early. |
+| `team.maxTurnsPerSubagent` | `0` | Runaway-loop cap, clamped to `[0, 100]`. `0` (the default) = no cap, matching the lead. A subagent stopped at a positive cap is reported as `stopped: turn cap reached`, with whatever it had produced. |
 | `todo.enabled` | `true` | Master switch for todo planning. `false` means the `todo_write` tool is never registered — `/todo on` in such a session saves the setting for next launch and says so, rather than advertising a tool that is not there. |
 | `todo.panel` | `true` | Render the right-hand rail. Independent of `todo.enabled` on purpose: a screen-reader user wants the planning discipline without the column, and the system prompt varies one sentence accordingly. |
 | `todo.followThrough` | `"notify"` | What happens when a run ends with steps left. `notify` says so and stops; `auto` continues the plan after a 3-second grace window, bounded by two structural limits (one fruitless attempt, 25 continuations per plan) that are deliberately **not** settings; `off` is silent. An unrecognized value falls back to `notify`. |
@@ -1936,7 +1969,7 @@ env / `.env` → CLI flags**.
 | `retry.multiplier` | `2` | Growth factor, clamped to `[1, 5]`. The one **float** key in this section: `2.5` is a legitimate factor and is stored as typed. |
 | `retry.jitter` | `true` | Equal jitter (`d/2 + rand·d/2`), which de-correlates concurrent subagents hitting the same 429 while keeping the announced countdown within a factor of two of the truth. Full jitter can fire 40 ms after announcing 30 s, which reads as a bug. |
 | `retry.respectRetryAfter` | `true` | Honour `Retry-After` (and the provider rate-limit reset headers) as a **floor**, never a replacement — a `Retry-After: 1` on the ninth retry does not undo eight retries of backoff. A stated wait above 60 s is surfaced immediately instead of waited out. |
-| `retry.maxElapsedMs` | `240000` | Wall-clock budget from the first failure, clamped to `[10000, 1800000]`. Deliberately **below** `team.subagentTimeoutMs`: an equal budget means a subagent is killed at the same instant its ladder ends, and the user is then told "subagent timed out" for what was a provider outage. |
+| `retry.maxElapsedMs` | `240000` | Wall-clock budget from the first failure, clamped to `[10000, 1800000]`. When `team.subagentTimeoutMs` is positive, keep this below it: an equal budget means a subagent is killed at the same instant its ladder ends, and the user is then told "subagent timed out" for what was a provider outage. |
 | `retry.onPartialStream` | `true` | Allow a restart after content has already streamed to the transcript. The view is rewound and the discarded tool cards removed; engine history is provably clean at that point, because it is only written on a completed turn. `false` refuses the whole path. |
 | `skills.enabled` | `true` | Master switch for the skill system. |
 | `skills.requireApproval` | `true` | Require a human to approve `skill_install` / `skill_create`. |
@@ -2150,6 +2183,8 @@ Everything AragonAgent keeps for you lives under one directory in your home:
 ~/.aragon-agent/                    (Windows: C:\Users\<you>\.aragon-agent\)
 ├── config.json                     settings + API keys (0600 on POSIX)
 ├── config.json.bak                 written by `aragon config edit` before it opens
+├── model-windows.json              per-model context windows you declare by hand;
+│                                   read live, never written by the CLI
 ├── prompt-history.jsonl            prompts you submitted, for ↑ recall (0600 on POSIX)
 ├── state.json                      UI bookkeeping: submit count, one-shot notices
 ├── update-state.json               auto-update: last check, pending restart, skipped version,
@@ -2326,19 +2361,19 @@ node packages/cli/dist/cli.js --version
 - `echo "list files" | node dist/cli.js -p` prints an answer and exits `0`.
 - A missing/invalid key produces a visible, actionable message (never a blank
   screen), both in the TUI and headless.
-- `Shift+Tab` at 40, 80 and 120 columns - the chip, the hint and the status bar
-  are all legible and nothing wraps onto a second row.
+- `Shift+Tab` at 40, 80 and 120 columns - the status-bar mode word (`PLAN` / `P`)
+  stays legible and nothing wraps onto a second row.
 - The question wizard at 24 rows with 5 questions x 4 options - no clipping, and
   the `n/N` counter matches what is on screen.
 - Approve a plan and watch the SAME run start writing files, with no second
   prompt.
-- `--no-color` on a legacy `cmd.exe` - the chip renders `[PLAN]`, radios render
-  `*`/`o`, checkboxes render `[x]`/`[ ]`, no mojibake anywhere.
+- `--no-color` on a legacy `cmd.exe` - the mode word renders as plain `PLAN` text,
+  radios render `*`/`o`, checkboxes render `[x]`/`[ ]`, no mojibake anywhere.
 - A plan with a 600-character summary and a 400-character step detail - the text
   wraps and scrolls, the footer count matches the real number of rows, and
   nothing is truncated to a single line.
 - `--no-hints` on a 20-row terminal in plan mode - the status bar still says
-  `PLAN` with no chip on screen.
+  `PLAN`.
 - Dismiss a plan card with `Esc`, let the agent resubmit, then close the card and press `Esc` twice within 1.5 seconds:
   the run aborts and the session returns to an idle prompt. Do this one BY HAND
   as well as in a test - the point of the check is whether the exit is findable,
@@ -2366,7 +2401,9 @@ UTF-16 编辑和 ZWJ/组合字形仍有既有限制，本次修复保证视觉
 ## 输入、待处理队列与底部状态区
 
 Enter 在空闲时发送，运行时加入队列；启动阶段拒绝重复提交并保留草稿。
-Ctrl+J 换行；终端能发送独立序列时，Shift+Enter 与 Alt+Enter 也可换行。
+Ctrl+J 换行；应用默认推送键盘增强模式（Windows 用 win32-input-mode，其余平台用
+kitty disambiguate + modifyOtherKeys=2）并将新编码翻译回传统字节，因此 Shift+Enter
+与 Alt+Enter 在支持的终端开箱即用，Ctrl+J 始终可用。
 光标按完整字素移动，支持中文、组合字符和 emoji；输入最多占终端高度四分之一且不超过六行。
 括号粘贴保留换行，不触发发送。过滤器或粘贴支持关闭时，不保证未标记多行粘贴安全。
 过滤器不可用时 Delete 保守按 Backspace 处理，避免退格误删后文。
@@ -2375,14 +2412,14 @@ Ctrl+J 换行；终端能发送独立序列时，Shift+Enter 与 Alt+Enter 也
 原生剪贴板工具完成后显示确认，OSC 52 只表示请求已发送，终端可能拒绝。
 Shift+拖动或 `/mouse off` 后的选区由宿主管理，应用无法关闭宿主的自动复制。
 
-Queue 面板逐条显示待处理消息，数量在展开状态详情中按空间显示。只有 Core 已把消息纳入会话并发回接收凭据，
+Queue 面板逐条显示待处理消息，数量在展开状态详情中按空间显示。只有 Core 已把消息纳入会话并发回接收凭据，
 该条提示才会消失。中断时未接收消息显示暂停；`/queue` 可分页查看全部原文，Esc 关闭。
 `/clear` 不丢待处理消息；`/reset`、切换会话会明确报告取消数量；恢复保存的未发送历史不会自动重发。
 
-底部依次为固定输入框和一行核心状态；Ctrl+G 展开或收起第二行详情。五个核心字段按显示宽度缩写，始终保留。
-核心上下文显示当前主 Agent 的百分比；展开详情或 /context 可查看已用/上限 tokens；`~` 表示含估算，`?` 表示上限未知。
+底部依次为固定输入框和一行核心状态；Ctrl+G 展开或收起第二行详情。五个核心字段按显示宽度缩写，始终保留。
+核心上下文显示当前主 Agent 的百分比；展开详情或 /context 可查看已用/上限 tokens；`~` 表示含估算，`?` 表示上限未知。
 费用是含子代理、快速模型及压缩的会话估算 USD，恢复会话沿用累计，缺价格按零计，零值不保证免费。
-本轮均速只累计主 Agent 已报告的 output tokens，再除以本轮总耗时（含工具与等待）；排除辅助调用，并非瞬时生成速度。结束后保留耗时，速度显示 --。
+本轮均速只累计主 Agent 已报告的 output tokens，再除以本轮总耗时（含工具与等待）；排除辅助调用，并非瞬时生成速度。结束后保留耗时，速度显示 --。
 `/context` 提供统计范围和完整值。
 
 Windows Terminal：在设置 JSON **根级**添加 `"copyOnSelect": false`，把以下条目合并到已有数组：
@@ -2432,10 +2469,10 @@ VS Code 用户设置添加 `"terminal.integrated.copyOnSelection": false`，快
 
 配置库仍位于 `ARAGON_HOME` 指向目录中的 `config.json`，不写入会话文件。密钥按现有机制明文保存：POSIX 创建权限为 0600，Windows 依赖用户目录 ACL，不提供加密。`config get/list` 对方案密钥使用固定掩码；损坏或未来版本的配置段整体隐藏。既有 `config edit` 备份也包含凭据。关闭通用日志脱敏、使用短密钥或分享第三方 SDK 错误日志时，仍需自行检查日志内容。
 
-### 固定输入与状态详情人工验收
-
-在 Windows Terminal/PowerShell、VS Code 集成终端以及 macOS/Linux 终端分别检查
-40x12、80x24、120x40。保留草稿，连续输出思考、正文与工具日志至少 30 秒：
-输入上下边框不得移动；拖动消息滚动条不会隐藏输入。切换 Ctrl+G 仅改变一行，
-复制、队列、通知和更新不自动展开。验证补全、Unicode 草稿和光标在切换后保留，
-弹层期间 Ctrl+G 无效，小屏恢复后草稿与状态模式保留。自动化测试不能替代这些宿主观察。
+### 固定输入与状态详情人工验收
+
+在 Windows Terminal/PowerShell、VS Code 集成终端以及 macOS/Linux 终端分别检查
+40x12、80x24、120x40。保留草稿，连续输出思考、正文与工具日志至少 30 秒：
+输入上下边框不得移动；拖动消息滚动条不会隐藏输入。切换 Ctrl+G 仅改变一行，
+复制、队列、通知和更新不自动展开。验证补全、Unicode 草稿和光标在切换后保留，
+弹层期间 Ctrl+G 无效，小屏恢复后草稿与状态模式保留。自动化测试不能替代这些宿主观察。

package/dist/agent/controller.d.ts

@@ -15,6 +15,7 @@ import type { PromptOptions, PromptOutcome } from './prompt-options.js';
 import { ModelRegistry, type AgentEvent, type AgentTool, type Message, type ModelInfo, type ModelRef, type ProviderRegistry, type ThinkingLevel } from '@aragon-agent/core';
 import type { CliConfig, CompactionConfig, FastConfig, RetryConfig, TeamConfig, ThemeName, TodoConfig } from '../config/schema.js';
 import type { ModelRole } from '../config/model-profiles.js';
+import { ModelWindows } from '../config/model-windows.js';
 import { type ConfirmRequest } from '../tools/index.js';
 import type { ToolPermission } from '../exec/permission.js';
 import { type ToolOutputListener } from '../tools/tool-output-store.js';
@@ -102,6 +103,13 @@ export interface ControllerDeps {
      * once at construction and lost at the first `rebuildSystemPrompt()`.
      */
     appendSystemPrompt?: string;
+    /**
+     * The per-model window table from ~/.aragon-agent/model-windows.json.
+     * Injectable so tests can hand the controller a table without writing into
+     * the shared vitest home root (app-paths.ts TEST-ISOLATION CONTRACT);
+     * production leaves it absent and the constructor builds the real one.
+     */
+    modelWindows?: ModelWindows;
 }
 /** What `setAgentMode` actually adopted — see `AgentController.setAgentMode`. */
 export interface AgentModeState {
@@ -124,6 +132,8 @@ export declare class AgentController {
     private readonly agent;
     private readonly providerRegistry;
     private readonly modelRegistry;
+    /** Per-model user-declared context windows; see `config/model-windows.ts`. */
+    private readonly modelWindows;
     private modelMetadataEnabled;
     private modelMetadataGeneration;
     private modelMetadataAbort;

package/dist/agent/controller.js

@@ -7,10 +7,12 @@ import { clampCompactionConfig, clampFastConfig, clampRetryConfig, clampTeamConf
 import { resolveModelProfileState } from '../config/model-profile-resolution.js';
 import { makeGetApiKey } from '../config/load.js';
 import { updatePersistedConfig } from '../config/store.js';
+import { ModelWindows } from '../config/model-windows.js';
 import { createBuiltinTools } from '../tools/index.js';
 import { createFileChangeStore } from '../tools/file-change-store.js';
 import { createToolOutputStore, } from '../tools/tool-output-store.js';
 import { TeamRuntime } from '../team/runtime.js';
+import { createChildReviewer } from '../team/child-reviewer.js';
 import { TeamHumanQueue } from '../team/human-queue.js';
 import { createTaskTool } from '../team/task-tool.js';
 import { buildTeamBlock } from '../team/prompt.js';
@@ -53,6 +55,8 @@ export class AgentController {
     agent;
     providerRegistry;
     modelRegistry;
+    /** Per-model user-declared context windows; see `config/model-windows.ts`. */
+    modelWindows;
     modelMetadataEnabled = false;
     modelMetadataGeneration = 0;
     modelMetadataAbort = new AbortController();
@@ -158,14 +162,14 @@ export class AgentController {
     // Fast model tier (fast-model-tier §3.1 / §3.3)
     //
     // ONE FIELD AND FOUR THIN FORWARDERS. The flag pair, the tier cache, the
-    // reviewer's lifetime and the `resolveTier` closure all live in
+    // reviewer's lifetime and the shared fail-fast review transport all live in
     // `fast/wiring.ts`, which is a BUDGET decision rather than an aesthetic one
     // (C-12 / RV-5): `CLAUDE.md` caps a source file at 1000 lines and this one had
     // 40 to spare when the design was written.
     //
     // `null` when the tier could not be resolved at construction, which is the
-    // byte-identity branch: no reviewer is subscribed, `task` carries no `model`
-    // property, and `<fast_tier>` is never spliced (I-2).
+    // byte-identity branch: no reviewer is subscribed and `<fast_tier>` is
+    // never spliced (I-2).
     // -----------------------------------------------------------------------
     fast;
     // -----------------------------------------------------------------------
@@ -275,6 +279,7 @@ export class AgentController {
         // reuses this very registry instance.
         this.providerRegistry = initProviders({ retryPolicy: toRetryPolicy(config.retry) });
         this.modelRegistry = new ModelRegistry(this.providerRegistry);
+        this.modelWindows = deps.modelWindows ?? new ModelWindows();
         this.skillsEnabled = config.skills.enabled;
         this.onSkillsChanged = deps.onSkillsChanged;
         this.skills = new SkillService({
@@ -307,8 +312,8 @@ export class AgentController {
         this.todos = this.todoRegistered ? new TodoStore() : null;
         // --- Fast model tier ---------------------------------------------------
         //
-        // RESOLVED BEFORE THE TOOL ARRAY, because `fastRegistered` decides whether
-        // `task`'s schema carries a `model` property at all and the array is built
+        // RESOLVED BEFORE THE AGENT, because `fastRegistered` decides whether the
+        // `FastWiring` (reviewer, tier cache) exists at all and the wiring is built
         // ONCE (C-2). `FastWiring` re-runs the same pure function in its own
         // constructor; the two agree by construction because they read the same
         // config through the same resolver.
@@ -320,24 +325,35 @@ export class AgentController {
         this.teamRuntime = this.teamRegistered
             ? new TeamRuntime({
                 getConfig: () => this.config,
-                // Read LIVE through the wiring, exactly as `resolveTier` below is: the
-                // closure is only ever called from inside a dispatch, and
-                // `childFactory()` re-reads `compaction.subagents` on every call so a
-                // settings-screen edit reaches the next dispatch.
+                // Read LIVE: the closure is only ever called from inside a dispatch,
+                // and `childFactory()` re-reads `compaction.subagents` on every call
+                // so a settings-screen edit reaches the next dispatch.
                 contextManagerFor: (req) => this.compaction?.childFactory()?.(req),
-                // Read LIVE through the wiring, which does not exist yet — the closure
-                // is only ever called from inside a dispatch, long after construction.
-                resolveTier: (tier) => this.fast
-                    ? this.fast.resolveTier(tier)
-                    : {
-                        ref: {
-                            providerId: this.config.provider,
-                            modelId: this.config.model,
-                            ...(this.config.baseUrl ? { baseUrl: this.config.baseUrl } : {}),
-                        },
-                        thinkingLevel: this.config.thinkingLevel,
-                        role: 'main',
+                // PER-CHILD FAST REVIEW (main-agent parity): one reviewer per
+                // child, sharing the lead wiring's tier, policy and fail-fast
+                // transport. `this.fast` is null until after the Agent exists; these
+                // closures run only from inside a dispatch, long after construction.
+                review: {
+                    // EVERY `this.fast` READ LIVES INSIDE A CLOSURE: the wiring is
+                    // assigned AFTER this constructor block (it needs the Agent), and
+                    // a read in the constructor's own flow would be a use-before-assign.
+                    // The closures run only from inside a dispatch, long after
+                    // construction, when the field has its final value.
+                    active: () => this.fast !== null && this.fast.available() && this.config.fast.review,
+                    modelId: () => {
+                        const tier = this.fast?.getTier();
+                        return tier && tier.ok ? tier.ref.modelId : '';
                     },
+                    create: (request) => this.fast
+                        ? createChildReviewer(request, {
+                            getConfig: () => this.config,
+                            hasKey: (id, role) => this.hasApiKey(id, role),
+                            getApiKey: (id, role) => this.resolveKey(id, role),
+                            complete: (providerId, llmRequest) => this.fast.completeViaFastRegistry(providerId, llmRequest),
+                            available: () => this.fast?.available() === true,
+                        })
+                        : undefined,
+                },
                 providerRegistry: this.providerRegistry,
                 getCwd: () => this.cwd,
                 getMode: () => this.effectiveMode,
@@ -484,23 +500,12 @@ export class AgentController {
                         runtime: this.teamRuntime,
                         // Read LIVE: `/team off` flips the flag, it cannot unregister.
                         isTeamEnabled: () => this.teamEnabled,
-                        // FROZEN AT CONSTRUCTION (C-2). `/fast on` in a session launched
-                        // without the tier cannot add a schema property, and it says so
-                        // rather than advertising one that is not there.
-                        fastRegistered,
-                        // Read LIVE at dispatch time (RV-3): the tier can stop resolving
-                        // mid-session, and the normalizer and the subagent factory must
-                        // agree about that at the same instant.
-                        fastAvailable: () => this.fast?.delegationAvailable() === true,
                         maxSubagents: () => this.config.team.maxSubagents,
                         hasApiKey: () => this.hasApiKey(),
                         activeProvider: () => this.config.provider,
+                        // Every child runs the lead's own model, so the lead's cost table
+                        // is the one honest price for the dispatch total.
                         modelCost: () => this.getModelInfo().cost,
-                        fastModelCost: () => {
-                            const ref = this.fast?.fastRef();
-                            return ref ? this.getModelInfoFor(ref).cost : undefined;
-                        },
-                        fastPricingUnknown: () => this.fast?.snapshot().pricingUnknown === true,
                         withPausedWatchdog: (fn) => this.withPausedWatchdog(fn),
                     }),
                 ]
@@ -599,6 +604,10 @@ export class AgentController {
                     // number only bites BECAUSE `task` races `ctx.signal` — the
                     // executor's timeout is cooperative, so an override is inert for any
                     // tool that does not listen (I-2).
+                    //
+                    // `dispatchTimeoutMs: 0` (the default) reaches the executor as an
+                    // override of `0` = NO TIMER AT ALL: the dispatch lasts as long as its
+                    // slowest child, and only an Esc ends it early (main-agent parity).
                     ...(this.teamRegistered ? { task: config.team.dispatchTimeoutMs } : {}),
                 },
             },
@@ -668,13 +677,6 @@ export class AgentController {
         });
         if (this.fast) {
             this.rebuildSystemPrompt();
-            // Fast-tier children are counted from the outcome rather than from the
-            // specs, so a downgraded child is not miscounted as a saving (§3.6).
-            this.teamRuntime?.subscribe((event) => {
-                if (event.type !== 'dispatch_end')
-                    return;
-                this.fast?.noteDelegated(event.outcome.runs.filter((r) => r.tier === 'fast').length);
-            });
         }
         // ONE LINE FOR THE TWO CONFIGURATION MISTAKES THAT ARE OTHERWISE INVISIBLE
         // (§3.3): the tier is on but unnamed, or named against a provider with no
@@ -734,11 +736,6 @@ export class AgentController {
                 ? buildTeamBlock({
                     maxSubagents: config.team.maxSubagents,
                     maxConcurrent: config.team.maxConcurrent,
-                    // One cross-reference sentence, present only when the capability
-                    // is (§3.8). `this.fast` is null during the FIRST compose (the
-                    // wiring needs the Agent), which is why the constructor composes
-                    // again once it exists.
-                    fastDelegation: !!fastAvailable && config.fast.delegate,
                 })
                 : '',
             // BOTH FLAGS, exactly as team mode uses both, and for the identical
@@ -758,7 +755,6 @@ export class AgentController {
             fastBlock: fastAvailable
                 ? buildFastBlock({
                     model: tier.ok ? tier.ref.modelId : '',
-                    delegate: config.fast.delegate,
                     review: config.fast.review,
                 })
                 : '',
@@ -1737,9 +1733,18 @@ export class AgentController {
         const found = this.modelRegistry.getModel(ref.providerId, ref.modelId);
         const baseUrl = ref.baseUrl ?? (ref.providerId === this.config.provider && ref.modelId === this.config.model
             ? this.config.baseUrl : undefined);
+        const userWindow = this.modelWindows.lookup(ref.modelId);
         return {
             ...(found ?? this.modelRegistry.buildRuntimeModel(ref.providerId, ref.modelId)),
             ...this.modelRegistry.getContextWindow(ref.providerId, ref.modelId, baseUrl),
+            // A USER-DECLARED WINDOW OUTRANKS EVERY TABLE (model-windows.json): the
+            // user asserted the number for exactly this id, while api/catalog are
+            // what this process managed to look up about it. It still LOSES to
+            // config.json's `contextWindow`, which `ContextMeter.resolveWindow`
+            // consults before it ever reaches here.
+            ...(userWindow !== undefined
+                ? { contextWindow: userWindow, contextWindowSource: 'user' }
+                : {}),
         };
     }
     /** Best-effort metadata lookup; never block typing or a model request. */

package/dist/agent/reducer.js

@@ -860,10 +860,6 @@ export function viewReducer(state, action) {
             const runs = action.specs.map((spec) => ({
                 label: spec.label,
                 description: spec.description,
-                // Already downgraded by the normalizer when the tier was unavailable, so
-                // the card and the report agree from the first frame (fast-model-tier
-                // §3.4).
-                tier: spec.tier,
                 phase: 'queued',
                 turns: 0,
                 toolCalls: 0,