zames_pro 2.82.139 → 2.82.141

package/CHANGELOG.md

@@ -7,6 +7,63 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [2.82.141] - 2026-10-26
+
+### Fixed
+
+- **i18n:** /rewind help names --yes and the default point
+- **i18n:** /web help lists the /websearch alias
+- **i18n:** /config help lists its subcommand aliases
+- **i18n:** /queue help lists its list/clear aliases
+- **i18n:** /goal help lists the off and on forms
+- **i18n:** json and jsonl are the same event stream
+- **i18n:** the /loop floor is not a 15s provider throttle
+- **help:** list the /session and /usage aliases
+- **i18n:** drop the duplicate backslash newline hint
+- **tui:** enable the kitty keyboard protocol so Shift+Enter works
+- **i18n:** drop the duplicate status.review_none key
+- **i18n:** /status says MCP tools is a count, not a list
+- **i18n:** the empty-screen hint names a working newline key
+- **prompt:** say that plan mode also hides mutating MCP tools
+- **prompt:** list every *_base64 escape hatch
+- **i18n:** plan mode lists every dropped tool, MCP included
+- **prompt:** decode the terminal bytes across reads, not per read
+- **i18n:** the clipboard hint names the real Wayland and Windows tools
+- **i18n:** the doctor auth row describes the sign-in marker it actually reads
+- **i18n:** the Qwen guest notice says /provider login opens the window itself
+- **i18n:** each full-screen key hint names its own mouse mode
+- **i18n:** the TUI key hint does not advertise a wheel that is off
+- **i18n:** Esc in the AskUser modal means no answer, not dismiss
+- **i18n:** keep prompt.answer_language English in both locales
+- **mcp:** localize the floating-version warning
+- **browser:** localize the send-pause fallback line
+- **browser:** DeepSeek login notices go through the UI hook
+- **compact:** fill the {provider} placeholder in compact.start
+- **i18n:** AskUser free-text mode explains how to send
+- **llms:** generated header says GLM runs headed too
+- **i18n:** suggestion hint says Ctrl+N/P select, not scroll
+- **i18n:** startup hint shows Ctrl+V for clipboard images too
+- **i18n:** /last says there is no session for THIS directory
+- **i18n:** drop the unverified ~10 messages/day guest quota claim
+- **i18n:** keep the agent-facing ONLY TOOL CALLS block English in both locales
+- **i18n:** MCP hint lists the .mcp.json path too
+- **i18n:** help shows the real clipboard paste binding (Ctrl+V)
+- **i18n:** GLM guest_limit message says the quota is exhausted
+- **i18n:** deepseek send_no_start states the real ~50s budget
+- **i18n:** describe both mouseScroll states and the real default
+- **i18n:** translate the English /status labels into Russian
+
+### Changed
+
+- **i18n:** drop the duplicate doctor.unknown and msg.not_dir keys
+
+## [2.82.140] - 2026-10-25
+
+### Fixed
+
+- **scheduler:** stop the ticker on exit; correct the abort help line
+- **config:** say when a project config overrides the home one
+
 ## [2.82.139] - 2026-10-25
 
 ### Changed

package/README.md

@@ -127,7 +127,9 @@ Key pieces:
 - **Subagents** — the `Task` tool delegates a self-contained sub-task to an
   agent in its own separate chat (isolated context), which returns only a
   report, so heavy research does not flood the main context. On by default;
-  tune with `browser.subagents` / `browser.maxSubagents`.
+  `browser.subagents` turns it off and `browser.maxSubagents` caps the runs
+  per TASK (default 3). Calls are sequential, not parallel, and the subtype
+  matters: `explore` is read-only, `general` may edit files.
 - **MCP support** — plug in external tool servers (e.g. `@playwright/mcp`).
 - **Scheduling** — `/loop`, `/cron` and `/jobs` repeat tasks on a timer.
 - **Bilingual UI** — Russian / English (`/config lang`).
@@ -287,8 +289,12 @@ The prompt is a small line editor with persistent history:
 - `Ctrl+_` — undo the last edit in the input line (a fat-fingered `Ctrl+U` /
   `Ctrl+K` is recoverable).
 - `Ctrl+U` — clear the line, `Ctrl+K` — delete to end of line, `Ctrl+W` —
-  delete the word before the cursor, `Ctrl+←`/`Ctrl+→` — move by words.
-- `\` + `Enter`, `Ctrl+J`, `Ctrl+Enter` or `Shift+Enter` — insert a newline.
+  delete the word before the cursor, `Ctrl+←`/`Ctrl+→` — move by words,
+  `Ctrl+Delete`/`Alt+Delete` — delete the word after the cursor.
+- `\` + `Enter` and `Ctrl+J` — insert a newline; these two work in
+  every terminal. `Ctrl+Enter` and `Shift+Enter` do it too, but only where the
+  terminal implements the kitty keyboard protocol: without it the terminal
+  sends a PLAIN Enter for Shift+Enter and the message is sent.
 - `/` + `Tab` — slash-command hints and completion.
 - `!command` — run a shell command directly, bypassing the model (like
   Claude Code's bash mode). The same sandbox guard as the `Bash` tool applies,
@@ -302,9 +308,10 @@ selected without leaving the terminal:
   wheel never scrolls the input history — history is arrow keys only.
 - `ui.mouseScroll=false` — the terminal owns the mouse again: normal text
   selection works and the wheel is inert.
-- `PgUp` / `PgDn` — move up/down the dialog half a screen.
-- `Home` / `End` — jump to the very top / bottom of the dialog (only while the
-  input line is empty; with text typed they move the caret as usual).
+- the mouse wheel, `PgUp` / `PgDn` (half a screen) and `Home` / `End` (the very
+  top / bottom) scroll the dialog; the wheel needs `ui.mouseScroll=true` (the default), and
+  Home/End jump only while the input line is empty (with text typed they
+  move the caret as usual).
 - To copy a LARGE block, use `/export` (writes the whole session as Markdown)
   or the transcript file (`/transcript` shows the path) — the terminal's own
   drag-select cannot reach past one screen.
@@ -313,8 +320,8 @@ Set `NO_COLOR=1` to disable colors (a calm default palette is used otherwise).
 
 ### Images and files
 
-You can paste an image or a file into the input line (Ctrl+Shift+V / Shift+Insert
-or the terminal's own paste). zames saves it under `<project>/tmp` and shows a
+You can paste an image or a file into the input line (Shift+Insert or the
+terminal's own paste; for a FILE from the clipboard press Ctrl+V). zames saves it under `<project>/tmp` and shows a
 marker in the line - `[image#1]` for images, `[file#1]` for other files - then
 uploads the real file to the chat together with your message. These are the
 same formats the DeepSeek / Qwen / GLM web chat accepts (PNG, JPEG, GIF, WEBP, BMP,
@@ -384,6 +391,8 @@ The agent has the same style of tools as Claude Code / Codex CLI:
 - **Git** — `GitStatus`, `GitDiff`, `GitLog`, `GitShow`, `GitBranchList`,
   `GitAdd`, `GitCommit`, `GitPush`.
 - **Web** — `WebFetch`, `WebSearch`.
+- **Subagents** — `Task` (delegate a self-contained sub-task to an agent in
+  its own chat; `explore` is read-only, `general` may edit; capped per task).
 - **Operator** — `AskUser` (a modal question with options; skipped in plan
   mode so a plan is never interrupted).
 - **Service** — `respond` (final answer to the operator, ends the task).
@@ -431,8 +440,9 @@ commands (/new or /clear, /sessions, /chats, /resume-id, /last, /chat, /cd,
 - /review [focus] [--staged] — ask the agent to review uncommitted changes
   and report findings (no code changes).
 - /plan [on|off] — plan (read-only) mode. While it is on, the mutating tools
-  (Write/Edit/MultiEdit/ApplyPatch/Bash, GitAdd/GitCommit/GitPush) are removed
-  from the tool set, so the agent can investigate without touching the tree.
+  (Write/Edit/MultiEdit/ApplyPatch/Bash/BashOutput, GitAdd/GitCommit/GitPush,
+  AskUser) are removed from the tool set, and so are MCP tools with a mutating
+  name, so the agent can investigate without touching the tree.
   Start in it with `--plan`.
 - /compact — ask the active backend to compress the current chat into a handover
   summary, then open a NEW chat, resend the system prompt and post the summary
@@ -440,9 +450,10 @@ commands (/new or /clear, /sessions, /chats, /resume-id, /last, /chat, /cd,
 - /goal <condition> — set a **completion condition** (Claude Code parity). The
   agent starts working toward it at once and keeps starting new turns until a
   separate evaluator model judges the condition met. The goal clears itself
-  automatically when it is met (or judged impossible), or with `/goal clear`.
-  The condition is stored in `<project>/.zames-goal` (git-ignored) and restored
-  on the next launch. `/goal` with no argument shows the current state.
+  automatically when it is met (or judged impossible), or with `/goal clear`
+  (or its synonym `/goal off`). The condition is stored in `<project>/.zames-goal`
+  (git-ignored) and restored on the next launch. `/goal` with no argument — or
+  `/goal on` — shows the current state.
 - /loop <interval> <task> — repeat a task periodically (e.g. `/loop 10m run the
 tests`). /cron "<min> <hour> <dom> <month> <dow>" <task> — run on a schedule.
   /jobs [rm <id>|clear] lists and stops them. A fired job is put into the same
@@ -456,8 +467,9 @@ tests`). /cron "<min> <hour> <dom> <month> <dow>" <task> — run on a schedule.
   "Web search" item in the "+" menu.
   These (and /queue, /jobs, /goal, `/config <sub>`) can be used WHILE the agent
   is working — they do not touch the in-flight generation.
-- /queue [clear] — list or clear the messages waiting to be sent after the
-  current task.
+- /queue [list|ls|clear|c|clean] — list (`list`/`ls`) or clear
+  (`clear`/`c`/`clean`) the messages waiting to be sent after the current
+  task.
 
 The token context is also shown live: the status line above the input has the
 spinner/text on the left and the context on the right (e.g. `125k · 13%`). The
@@ -563,8 +575,7 @@ too):
 
 - **Memory** (`@modelcontextprotocol/server-memory`) — a local knowledge graph
   (9 tools) for durable facts across chats. NOTE: zames already has `MEMORY.md`
-
-* `AGENTS.md`; add this only if you want a queryable graph.
+  + `AGENTS.md`; add this only if you want a queryable graph.
 
 Keep the stack SMALL: every server loads its tool definitions into the model's
 context. Add one when you feel its absence, not before.
@@ -580,13 +591,14 @@ You can view and change settings without leaving the agent — use the
 toggle in place; numbers and strings open an input prompt.
 
 ```
-/config                       interactive settings menu
-/config list                  print all editable settings
+/config                       interactive settings menu (alias: menu)
+/config list                  print all editable settings (aliases: show, ls)
 /config get <path>            show a setting
 /config set <path> <value>    change a setting
-/config reset <path>          reset a setting to its default
-/config path                  show config file paths
+/config reset <path>          reset a setting to its default (alias: unset)
+/config path                  show config file paths (alias: paths)
 /config lang <ru|en>          switch interface and agent language
+                              (aliases: language, язык)
 ```
 
 Examples:
@@ -821,7 +833,10 @@ zames не вызывает API модели. Он запускает headless C
 - **Субагенты** — инструмент `Task` поручает самодостаточную подзадачу агенту
   в отдельном чате (изолированный контекст), который возвращает только отчёт,
   так что тяжёлая разведка не засоряет основной контекст. Включены по
-  умолчанию; настраиваются через `browser.subagents` / `browser.maxSubagents`.
+  умолчанию; `browser.subagents` выключает их, а `browser.maxSubagents` ограничивает
+  число запусков на ОДНУ задачу (по умолчанию 3). Вызовы последовательные,
+  не параллельные, и подтип важен: `explore` только читает, `general` может
+  править файлы.
 - **Поддержка MCP** — подключение внешних серверов инструментов (например,
   `@playwright/mcp`).
 - **Планировщик** — `/loop`, `/cron` и `/jobs` повторяют задачи по таймеру.
@@ -980,9 +995,13 @@ zames
 - `Ctrl+_` — отменить последнюю правку строки ввода (случайно нажатые
   `Ctrl+U` / `Ctrl+K` можно вернуть).
 - `Ctrl+U` — очистить строку, `Ctrl+K` — удалить до конца строки, `Ctrl+W` —
-  удалить слово перед курсором, `Ctrl+←`/`Ctrl+→` — переход по словам.
-- `\` + `Enter`, `Ctrl+J`, `Ctrl+Enter` или `Shift+Enter` — вставить новую
-  строку.
+  удалить слово перед курсором, `Ctrl+←`/`Ctrl+→` — переход по словам,
+  `Ctrl+Delete`/`Alt+Delete` — удалить слово после курсора.
+- `\` + `Enter` и `Ctrl+J` — вставить новую строку; эти две
+  работают в любом терминале. `Ctrl+Enter` и `Shift+Enter` тоже вставляют
+  строку, но только там, где терминал поддерживает kitty-протокол
+  клавиатуры: без него Shift+Enter приходит как обычный Enter, и
+  сообщение отправляется.
 - `/` + `Tab` — подсказки и автодополнение слэш-команд.
 - `!команда` — выполнить shell-команду напрямую, минуя модель (как bash-режим
   Claude Code). Действует та же песочница, что и для инструмента `Bash`, так
@@ -996,9 +1015,10 @@ zames
   никогда не листает историю ввода — история только стрелками.
 - `ui.mouseScroll=false` — мышью снова владеет терминал: обычное выделение
   текста работает, колесо инертно.
-- `PgUp` / `PgDn` — вверх/вниз по диалогу на пол-экрана.
-- `Home` / `End` — прыжок в самый верх / низ диалога (только когда строка ввода
-  пуста; при набранном тексте они двигают курсор, как обычно).
+- колесо мыши, `PgUp` / `PgDn` (пол-экрана) и `Home` / `End` (самый верх / низ)
+  прокручивают диалог; колесо работает при `ui.mouseScroll=true` (по умолчанию), а
+  Home/End прыгают только когда строка ввода пуста (при набранном тексте
+  они двигают курсор, как обычно).
 - Чтобы скопировать БОЛЬШОЙ кусок, используйте `/export` (пишет всю сессию в
   Markdown) или файл транскрипта (`/transcript` покажет путь) — собственным
   выделением терминала за пределы одного экрана не выделить.
@@ -1008,8 +1028,8 @@ zames
 
 ### Картинки и файлы
 
-В строку ввода можно вставить картинку или файл (Ctrl+Shift+V / Shift+Insert
-или собственной вставкой терминала). zames сохраняет его в `<project>/tmp` и
+В строку ввода можно вставить картинку или файл (Shift+Insert или
+собственной вставкой терминала; файл из буфера обмена — по Ctrl+V). zames сохраняет его в `<project>/tmp` и
 показывает маркер в строке — `[image#1]` для картинок, `[file#1]` для прочих
 файлов — затем загружает реальный файл в чат вместе с вашим сообщением. Это те
 же форматы, что принимает веб-чат DeepSeek / Qwen / GLM (PNG, JPEG, GIF, WEBP, BMP,
@@ -1079,6 +1099,9 @@ zames --task "..." --output-format jsonl \
 - **Git** — `GitStatus`, `GitDiff`, `GitLog`, `GitShow`, `GitBranchList`,
   `GitAdd`, `GitCommit`, `GitPush`.
 - **Web** — `WebFetch`, `WebSearch`.
+- **Субагенты** — `Task` (делегировать самодостаточную подзадачу агенту в
+  отдельном чате; `explore` только читает, `general` может править; число
+  запусков на задачу ограничено).
 - **Оператор** — `AskUser` (модальный вопрос с вариантами; в режиме плана не
   предлагается, чтобы не прерывать план).
 - **Служебный** — `respond` (финальный ответ оператору, завершает задачу).
@@ -1127,9 +1150,10 @@ zames --task "..." --output-format jsonl \
 - /review [focus] [--staged] — попросить агента проверить незакоммиченные
   изменения и сообщить находки (без правок кода).
 - /plan [on|off] — режим плана (только чтение). Пока он включён, мутирующие
-  инструменты (Write/Edit/MultiEdit/ApplyPatch/Bash, GitAdd/GitCommit/GitPush)
-  убраны из набора, чтобы агент изучал код, не трогая дерево. Стартовать в нём
-  можно с `--plan`.
+  инструменты (Write/Edit/MultiEdit/ApplyPatch/Bash/BashOutput,
+  GitAdd/GitCommit/GitPush, AskUser) убраны из набора вместе с MCP-инструментами
+  с мутирующим именем, чтобы агент изучал код, не трогая дерево. Стартовать в
+  нём можно с `--plan`.
 - /compact — попросить активный бэкенд сжать текущий чат в передаточное резюме,
   затем открыть НОВЫЙ чат, заново отправить системный промпт и положить резюме как
   перенесённый контекст. Используйте, когда контекст разрастается.
@@ -1137,9 +1161,10 @@ zames --task "..." --output-format jsonl \
   сразу начинает работать над ним и продолжает запускать новые ходы, пока
   отдельная модель-оценщик не признает условие выполненным. Цель снимается
   автоматически при достижении (или если оценщик признал её недостижимой),
-  либо командой `/goal clear`. Условие хранится в `<project>/.zames-goal`
-  (git-ignored) и восстанавливается при следующем запуске. `/goal` без
-  аргумента показывает текущее состояние.
+  либо командой `/goal clear` (синоним — `/goal off`). Условие
+  хранится в `<project>/.zames-goal` (git-ignored) и восстанавливается при
+  следующем запуске. `/goal` без аргумента — или `/goal on` —
+  показывает текущее состояние.
 - /loop <interval> <task> — повторять задачу периодически (например,
   `/loop 10m run the tests`). /cron "<min> <hour> <dom> <month> <dow>" <task> —
   запуск по расписанию. /jobs [rm <id>|clear] перечисляет и останавливает их.
@@ -1153,8 +1178,8 @@ zames --task "..." --output-format jsonl \
   в меню «+».
   Их (а также /queue, /jobs, /goal, `/config <sub>`) можно использовать, ПОКА
   агент работает — они не трогают текущую генерацию.
-- /queue [clear] — показать или очистить сообщения, ждущие отправки после
-  текущей задачи.
+- /queue [list|ls|clear|c|clean] — показать (`list`/`ls`) или очистить
+  (`clear`/`c`/`clean`) сообщения, ждущие отправки после текущей задачи.
 
 Контекст в токенах показывается и вживую: в строке статуса над вводом слева
 спиннер/текст, справа — контекст (например, `125k · 13%`). Процент считается от
@@ -1276,13 +1301,14 @@ MCP-браузер и браузер агента подерутся за оди
 enum переключаются на месте; числа и строки открывают ввод.
 
 ```
-/config                       интерактивное меню настроек
-/config list                  показать все изменяемые настройки
+/config                       интерактивное меню настроек (алиас: menu)
+/config list                  показать все изменяемые настройки (алиасы: show, ls)
 /config get <path>            показать настройку
 /config set <path> <value>    изменить настройку
-/config reset <path>          сбросить настройку к дефолту
-/config path                  показать пути файлов конфига
+/config reset <path>          сбросить настройку к дефолту (алиас: unset)
+/config path                  показать пути файлов конфига (алиас: paths)
 /config lang <ru|en>          переключить язык интерфейса и агента
+                              (алиасы: language, язык)
 ```
 
 Примеры:

package/SECURITY.md

@@ -19,8 +19,8 @@ is not a sandbox:
   your user. Treat every task you give the agent as code you are willing to
   run.
 - The chat-provider credentials in `~/.zames/config.json` and the browser
-  profiles (`~/.zames/profile`, `~/.zames/qwen-profile`) are sensitive. They
-  never leave your machine. Files
+  profiles (`~/.zames/profile`, `~/.zames/qwen-profile`, `~/.zames/glm-profile`)
+  are sensitive. They never leave your machine. Files
   written by the agent's own state/config writers are created with `0600`
   permissions, but files created by your own tools or by `Bash` follow your
   umask.
@@ -40,7 +40,10 @@ is not a sandbox:
   tools from the model's tool set. Since this release it also drops MCP tools
   whose names look mutating (click/type/write/...); the classification is a
   heuristic deny-list, so treat plan mode as a strong hint rather than a hard
-  guarantee for third-party MCP servers.
+  guarantee for third-party MCP servers. It is NOT a security boundary:
+  it only makes those tools INVISIBLE to the model for that task —
+  `.zames/permissions.json` is unchanged, no OS isolation is added, and a
+  server whose tool name the heuristic does not recognise is still offered.
 - `Write`/`Edit` and the multi-file tools keep a backup (`/undo`,
   `/undo-list`), but that is an undo safety net, not a security boundary. Give
   the agent only tasks you are willing to run.

package/dist/browser.js

@@ -616,7 +616,11 @@ export class DeepSeekBrowser {
                 catch { }
             },
             persistSession: () => this._persistSessionIfNeeded(),
-            notice: (key) => console.log(theme.system(this._t(key))),
+            // Route login notices through the SAME notice hook as every other
+            // message: a raw console.log is painted OVER by the LineEditor status
+            // line, so the operator saw a frozen spinner instead of "a browser
+            // window opened" / "login timed out" (see DESIGN-NOTES, onNotice).
+            notice: (key) => this._notice(this._t(key)),
             isAborted: () => this._abort,
             windowClosed: () => this._windowClosed,
             sleep: (ms) => new Promise((r) => setTimeout(r, ms)),
@@ -1932,7 +1936,9 @@ export class DeepSeekBrowser {
                 catch { }
             }
             else {
-                console.error(theme.warn(`⏳ send pause ${secs}s`));
+                // Fallback when no UI hook is wired (a direct ask() from a script):
+                // localize it — the pause is operator-facing text.
+                console.error(theme.warn(translate(this.locale)('spinner.pause', { n: String(secs) })));
             }
         };
         report(gap);

package/dist/commands.js

@@ -336,7 +336,7 @@ export function renderDoctor(d, t = translate('en')) {
         : t('doctor.not_repo'));
     row(d.configOk, 'config', d.configOk
         ? t('doctor.loaded')
-        : t('doctor.error', { v: d.configError || t('doctor.unknown') }));
+        : t('doctor.error', { v: d.configError || t('common.unknown') }));
     row(true, 'browser', d.browserChannel ? d.browserChannel : t('doctor.bundled'));
     row(!!d.authSaved, 'auth', d.authSaved ? t('doctor.auth_saved') : t('doctor.auth_none'));
     row(!!d.clipboardTool, 'clipboard', d.clipboardTool || t('doctor.clipboard_none'));
@@ -1203,6 +1203,8 @@ export const SLASH_COMMANDS = [
     { name: '/new', key: 'help.cmd.new' },
     { name: '/clear', key: 'help.cmd.new' },
     { name: '/sessions', key: 'help.cmd.sessions' },
+    // ALIAS: /session is handled by the same branch in src/index.ts.
+    { name: '/session', key: 'help.cmd.sessions' },
     { name: '/chats', key: 'help.cmd.chats' },
     { name: '/resume-id', key: 'help.cmd.resume_id' },
     { name: '/last', key: 'help.cmd.last' },
@@ -1223,6 +1225,8 @@ export const SLASH_COMMANDS = [
     { name: '/diff', key: 'help.cmd.diff' },
     { name: '/diffstat', key: 'help.cmd.diffstat' },
     { name: '/cost', key: 'help.cmd.cost' },
+    // ALIAS: /usage takes the same branch as /cost.
+    { name: '/usage', key: 'help.cmd.cost' },
     { name: '/limits', key: 'help.cmd.limits' },
     { name: '/export', key: 'help.cmd.export' },
     { name: '/doctor', key: 'help.cmd.doctor' },

package/dist/compact.js

@@ -17,6 +17,11 @@ export const COMPACT_FALLBACK_LIMIT = 40;
 export async function performCompact(opts) {
     const { browser, workdir, locale, task, transcript = null, ui = null, buildSystemPrompt, tools, answerTimeoutMs, fallbackLimit = COMPACT_FALLBACK_LIMIT, skipUiLock = false, quiet = false, selfImprovement = false, contextExhausted = false, } = opts;
     const t = translate(locale);
+    // The active provider's LABEL for operator-facing text: performCompact is
+    // shared by every backend, so it cannot hardcode "DeepSeek". `providerArg`
+    // comes from the caller (the same registry label the rest of the UI shows);
+    // the browser is only a fallback for direct callers and tests.
+    const providerLabel = opts.providerLabel || 'DeepSeek';
     const log = (msg) => {
         if (!quiet)
             console.error(msg);
@@ -42,7 +47,7 @@ export async function performCompact(opts) {
         ui.lock(t('msg.input_locked'));
     try {
         if (!quiet)
-            log(t('compact.start'));
+            log(t('compact.start', { provider: providerLabel }));
         // PreCompact hook: runs BEFORE the chat is summarized/rewritten, so
         // a hook can persist state or notify. Best-effort.
         try {

package/dist/config-command.js

@@ -15,8 +15,22 @@ import { configGetValue, configValueHint } from './config-commands.js';
  * blindly from the main loop and from the live mid-run interception.
  */
 export async function handleConfigCommand(input, deps) {
-    const { t, print, printErr, getValue, setValue, resetField, showList, openMenu, currentLocale, normalizeLocale, localeDisplayName, homeConfigPath, projectConfigPath, } = deps;
+    const { t, print, printErr, getValue, setValue, resetField, showList, openMenu, currentLocale, normalizeLocale, localeDisplayName, homeConfigPath, projectConfigPath, shadowedKeys, } = deps;
     const schema = deps.schema;
+    // A HOME write is not always the value the next launch reads: the project
+    // `.zamesrc.json` is merged last, so a key present in both wins from the
+    // project file. Without this line the operator saw "saved" and then the OLD
+    // value after a restart (the stale `ui.locale: en` bug).
+    const warnShadowed = (key) => {
+        if (!shadowedKeys)
+            return;
+        if (!shadowedKeys().includes(key))
+            return;
+        printErr(theme.warn(t('cfg.shadowed', {
+            file: projectConfigPath,
+            keys: key,
+        })));
+    };
     // Accept BOTH spellings: the main loop passes the whole typed line
     // (`/config get ui.locale`), while tests and other callers pass the text
     // after the command (`get ui.locale`). Reading parts[1] blindly made the
@@ -60,6 +74,7 @@ export async function handleConfigCommand(input, deps) {
             return;
         }
         print(theme.assistant(t('cfg.lang_set', { v: localeDisplayName(loc) })));
+        warnShadowed('ui.locale');
         return;
     }
     if (sub === 'get') {
@@ -105,6 +120,7 @@ export async function handleConfigCommand(input, deps) {
             value: configGetValue(key, getValue(key), t),
             file: homeConfigPath,
         })));
+        warnShadowed(key);
         return;
     }
     if (sub === 'reset' || sub === 'unset') {

package/dist/config.js

@@ -165,19 +165,72 @@ export const DEFAULTS = {
 };
 export function loadConfig() {
     const merged = JSON.parse(JSON.stringify(DEFAULTS));
-    for (const file of [HOME_CONFIG, projectConfigPath()]) {
-        try {
-            const data = JSON.parse(fs.readFileSync(file, 'utf-8'));
-            deepMerge(merged, data);
-        }
-        catch (e) {
-            const err = e;
-            if (err.code !== 'ENOENT') {
-                console.error(`config: failed to read ${file}: ${err.message}`);
+    // Keys the PROJECT file overrides in the HOME file. Kept so the editor can
+    // say "your change was saved but the project file still wins" — a live
+    // symptom was `/config lang ru` answering "saved", then English after a
+    // restart, because a stale project `.zamesrc.json` pinned `ui.locale: en`.
+    shadowedKeys = [];
+    const homeData = readConfigObject(HOME_CONFIG);
+    const projectData = readConfigObject(projectConfigPath());
+    if (projectData) {
+        shadowedKeys = overriddenKeys(homeData, projectData);
+    }
+    for (const data of [homeData, projectData]) {
+        if (!data)
+            continue;
+        deepMerge(merged, data);
+    }
+    return merged;
+}
+/** Keys the lower-priority file also sets, from the last loadConfig(). */
+let shadowedKeys = [];
+export function shadowedConfigKeys() {
+    return shadowedKeys.slice();
+}
+/**
+ * Every leaf path ("ui.locale", "browser.auth.username") present in BOTH
+ * objects — i.e. a value the PROJECT config overrides over the HOME one.
+ * PURE so the precedence puzzle can be tested without touching the disk.
+ */
+export function overriddenKeys(lower, higher) {
+    if (!lower || !higher)
+        return [];
+    const out = [];
+    const walk = (a, b, prefix) => {
+        for (const key of Object.keys(b)) {
+            if (key === '__proto__' || key === 'constructor' || key === 'prototype') {
+                continue;
+            }
+            const path2 = prefix ? prefix + '.' + key : key;
+            const av = a[key];
+            const bv = b[key];
+            if (isPlainObject(av) && isPlainObject(bv)) {
+                walk(av, bv, path2);
+            }
+            else if (av !== undefined && !isPlainObject(bv)) {
+                out.push(path2);
             }
         }
+    };
+    walk(lower, higher, '');
+    return out.sort();
+}
+function isPlainObject(v) {
+    return !!v && typeof v === 'object' && !Array.isArray(v);
+}
+/** Parse one config file; null when it is missing or unreadable. */
+function readConfigObject(file) {
+    try {
+        const data = JSON.parse(fs.readFileSync(file, 'utf-8'));
+        return isPlainObject(data) ? data : null;
+    }
+    catch (e) {
+        const err = e;
+        if (err.code !== 'ENOENT') {
+            console.error(`config: failed to read ${file}: ${err.message}`);
+        }
+        return null;
     }
-    return merged;
 }
 // The agent-loop watchdog deadline per provider. The loop default (240s) is
 // tuned for DeepSeek; a Qwen GUEST turn regularly takes 7-8 minutes (cold

package/dist/context.js

@@ -66,7 +66,7 @@ const BUILTIN_SKILLS = [
     // them out of the box; a project skill of the same name overrides them.
     {
         name: 'brainstorming',
-        description: 'Use when the request is a GOAL, not a spec — "add X", "make it better", "integrate Y" — and it is unclear what "done" means.',
+        description: 'Use when the operator\'s request is a GOAL, not a spec — "add X", "make it better", "integrate Y" — and it is unclear what "done" means. Draws out the intent and plays it back before building.',
         body: 'When a request is a goal ("make scrolling feel right"), building ' +
             'immediately produces the wrong thing fast. First make the intent concrete.\n' +
             'Pin down: (a) the REAL outcome — what should be TRUE after that is not ' +
@@ -83,7 +83,7 @@ const BUILTIN_SKILLS = [
     },
     {
         name: 'writing-plans',
-        description: 'Use when a task needs more than one or two tool calls — a multi-file change, a new feature, a refactor. Breaks it into small ordered steps with exact paths and a verification for each.',
+        description: 'Use when a task is large enough to need more than one or two tool calls — a multi-file change, a new feature, a refactor. Breaks the work into small ordered steps with exact paths and a verification for each.',
         body: 'A plan is a sequence of steps a careful engineer with NO project ' +
             'context could follow. Each step (2-5 minutes) names: the EXACT file ' +
             'path(s); the concrete edit; how to verify it (command, test name, ' +
@@ -97,11 +97,16 @@ const BUILTIN_SKILLS = [
             'plan instead of improvising; a 30-step plan means the task should be ' +
             'split. In zames the plan lives in your reasoning and, for a long task, ' +
             'in a TodoWrite list — one entry per step. Do not narrate it between tool ' +
-            'calls; the todo list IS the visible plan.',
+            'calls; the todo list IS the visible plan. A plan is not a report: the ' +
+            'operator sees ONLY the last respond, so keep the per-step evidence for ' +
+            'that one message. Subagents do not parallelise the plan either — a Task ' +
+            'call is a fresh isolated chat and zames has ONE browser and ONE send ' +
+            'slot, so subagents are sequential and only keep a broad read-only ' +
+            'investigation out of the main context.',
     },
     {
         name: 'test-driven-development',
-        description: 'Use when writing or changing testable behaviour — a bug fix, a new function, a parser, a formatter. Write the failing test first, watch it fail, then make it pass.',
+        description: 'Use when writing or changing behaviour that can be tested — a bug fix, a new function, a parser, a formatter. Write the failing test first, watch it fail, then make it pass. Prevents tests that only pass because they were written against the finished code.',
         body: 'RED -> GREEN -> REFACTOR. The order is the point, not a formality.\n' +
             '1. RED — write the smallest test that expresses the new behaviour and ' +
             'RUN it. It must FAIL for the RIGHT reason ("not implemented", not a ' +
@@ -114,11 +119,16 @@ const BUILTIN_SKILLS = [
             'for ONE reason with a message that says what broke. In this project a ' +
             'bug fix ALWAYS gets a regression test reproducing the original bug. ' +
             'When a test needs fixing, ask whether the TEST or the CODE was wrong; ' +
-            'if the code was written before its test, rewrite it driven by the test.',
+            'if the code was written before its test, rewrite it driven by the test.\n' +
+            'A test that writes a temp dir or file cleans it up itself: in this repo ' +
+            'always go through test/tmp-dir.ts (makeTmpDir(t, prefix) / ' +
+            'makeTmpDirSync), which registers the t.after cleanup. On WSL a leaked ' +
+            'os.tmpdir() entry lives inside ext4.vhdx and never returns its space to ' +
+            'the Windows host, so an abandoned temp dir fragments the disk permanently.',
     },
     {
         name: 'systematic-debugging',
-        description: 'Use when something is broken and the cause is not obvious — a failing test, a wrong result, a hang, a crash. Four phases: reproduce, isolate, hypothesize, fix.',
+        description: 'Use when something is broken and the cause is not obvious — a failing test, a wrong result, a hang, a crash. Four phases: reproduce, isolate, hypothesize, fix. Prevents guessing.',
         body: 'Do NOT start by editing code. Work the four phases in order; each ' +
             'produces an artifact you can point at.\n' +
             '1. REPRODUCE — get a command that FAILS the way the operator describes; ' +
@@ -127,7 +137,12 @@ const BUILTIN_SKILLS = [
             '2. ISOLATE — shrink to the smallest failing unit (bisect, stub, ' +
             'binary-search). Read the code the failing path ACTUALLY executes. Check ' +
             'the obvious first: wrong file, stale build, cached state, different ' +
-            'config, an error swallowed by a catch.\n' +
+            'config, an error swallowed by a catch. In zames the evidence is already ' +
+            'on disk: the transcript (~/.zames/logs/*.jsonl, /transcript prints this ' +
+            "run's path), the raw answer bodies in ~/.zames/net-log with " +
+            'ZAMES_NET_DEBUG=1 (npm run dev sets it), and the per-send telemetry ' +
+            '(~/.zames/logs/sends-<provider>.jsonl, dev mode or ' +
+            'browser.sendTelemetry). Read them BEFORE guessing.\n' +
             '3. HYPOTHESIZE — state it in one sentence and predict what you should ' +
             'see if it is right; test exactly that. One change at a time; prefer an ' +
             'observation (log, test assertion) over an edit. If the prediction is ' +
@@ -141,7 +156,7 @@ const BUILTIN_SKILLS = [
     },
     {
         name: 'verification-before-completion',
-        description: 'Use BEFORE declaring a task done — a feature, a bug fix, a generated file. Forces you to run the verification you promised instead of trusting the code looks right.',
+        description: 'Use BEFORE declaring a task done or reporting success — a feature, a bug fix, a file you generated. Forces you to run the verification you promised instead of trusting that the code looks right.',
         body: '"It should work" is not done. Before reporting success, produce ' +
             'EVIDENCE from the real artifact, not from your intent.\n' +
             '1. Re-read the original request; tick every requirement (including the ' +
@@ -155,7 +170,10 @@ const BUILTIN_SKILLS = [
             'result — so the operator can reproduce it. "Tests pass" without the ' +
             'command is not evidence.\n' +
             'If you cannot run the verification, SAY SO explicitly and describe what ' +
-            'you did instead; never report "done" implying a check you did not run.',
+            'you did instead; never report "done" implying a check you did not run.\n' +
+            'In zames the report is ONE respond call: the operator reads only the ' +
+            'LAST respond, so put the complete evidence there and never end a turn ' +
+            'on a bare text update with work still pending.',
     },
 ];
 // Per-file budgets. AGENTS.md/MEMORY.md are hand-maintained and grow over