micro-models-agent 0.60.0 → 0.61.0

package/CHANGELOG.md

@@ -4,6 +4,23 @@ All notable changes to Micro Models Agent (MMA) will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/), and this project adheres to [Semantic Versioning](https://semver.org/).
 
+## [0.61.0] - 2026-09-15
+
+### Added
+- **Budget share controls**: `mma context --system <f>` / `--reserve <f>` and REPL `/context system <f>` / `reserve <f>` set `contextBudget.systemPrompt` / `responseReserve` (validated 0.05–0.9, system + reserve < 0.95). `mma context` and `/context` with no argument now print the budget breakdown (system / reserve / history in tokens and %). `mma config set contextBudget.systemPrompt <f>` keeps working.
+
+### Changed
+- **Overflow fix advice**: the "system prompt exceeds budget" warning no longer only snaps to a coarse standard context size — a 32K project was jumping straight to `131072`, which looked arbitrary and was often unusable on a local model. The hint (`prompt.overflow.hint_*`, en+ru) now states the raw requirement (`needs ~N system tokens; current budget: W × f = B`) and both levers: raise `contextWindow` to the exact minimum (with the standard size) **or** keep the window and raise `contextBudget.systemPrompt` to the required share (`mma context --system <f>`).
+
+### Fixed
+- **`mma context` ignored `MMA_CONFIG_DIR`**: it saved to a hardcoded `~/.mma/config.json`; it now writes to the config dir resolved by bootstrap.
+
+## [0.60.1] - 2026-09-15
+
+### Fixed
+- **Silent prompt-overflow summarization**: the startup dry-run warns that AGENTS.md / the project map "will be summarized before the first run", but the actual summarization ran with no notice — the first prompt appeared to hang while a local model rewrote a large AGENTS.md. `resolvePromptOverflow` now emits (en+ru, `prompt.overflow.*`): a "compressing N block(s)…" line before it starts, a per-block "summarizing \"label\" (N → M tok)…" line before the model call, a "truncating …" line on the fallback, and a failure line before falling back to truncation.
+- **Hardcoded English startup logs**: model auto-load success/failure, reasoning probe result, project indexing start/finish/failure, unreadable index entries and background-processes-killed-on-shutdown were untranslated; they now use `env.*` keys (en+ru). Project indexing now announces start and completion (`Indexed N file(s)`), so a slow first walk no longer pauses startup silently.
+
 ## [0.60.0] - 2026-09-15
 
 ### Added

package/dist/i18n/en.json

@@ -52,6 +52,16 @@
   "env.playwright_missing": "Playwright package is not installed — the browser tool will fail. Install with: bun add playwright",
   "env.playwright_browsers_missing": "Playwright browsers not downloaded ({dir}). The browser tool will fail. Install with: bunx playwright install chromium",
   "env.crash_stderr": "MMA crashed ({type}): {message} — crash report written to ~/.mma/logs/crash.jsonl",
+  "env.model_autoload_failed": "Model auto-load failed: {error}. Continuing with manual load.",
+  "env.model_loaded": "Model {model} loaded in {seconds}s",
+  "env.reasoning_probe": "Reasoning probe: strategy={strategy} → {result}",
+  "env.reasoning_respected": "mechanism respected",
+  "env.reasoning_ignored": "mechanism IGNORED",
+  "env.indexing": "Indexing project files…",
+  "env.indexed": "Indexed {files} file(s) for the project map",
+  "env.indexing_failed": "Project indexing failed: {error}",
+  "env.index_skip_unreadable": "indexer: skipping unreadable entry \"{path}\"",
+  "env.killed_processes": "Killed {count} background process(es) on shutdown",
   "session.started": "Session started: {id}",
   "session.ended": "Session ended: {id}",
   "session.not_found": "Session not found: {id}",
@@ -269,6 +279,14 @@
   "cli.manage_context": "Manage context window",
   "cli.invalid_context_size": "Invalid context size. Must be a number >= 1024",
   "cli.context_set": "Context window set to: {size} tokens",
+  "cli.context_budget_header": "Context budget (window {window} tokens):",
+  "cli.context_system_line": "  system prompt: {tokens} tokens ({percent}%)",
+  "cli.context_reserve_line": "  response reserve: {tokens} tokens ({percent}%)",
+  "cli.context_history_line": "  history: {tokens} tokens ({percent}%)",
+  "cli.context_system_fraction": "System-prompt share of the context window (0.05–0.9)",
+  "cli.context_reserve_fraction": "Response-reserve share of the context window (0.05–0.9)",
+  "cli.context_fraction_set": "{key} share set to {value}",
+  "cli.context_invalid_fraction": "Invalid share. Use a number between 0.05 and 0.9 (system + reserve must stay below 0.95).",
   "repl.reload": "Reload agent with current config",
   "repl.reload_usage": "Usage: /reload",
   "repl.reloading": "Reloading agent...",
@@ -432,7 +450,7 @@
   "repl.sysprompt_desc": "Show system prompt",
   "repl.sysprompt_tokens": "System prompt ({count} tokens):",
   "repl.excluded_blocks": "\nExcluded blocks: {count}",
-  "repl.context_usage": "Usage: /context <size>  (min 1024)",
+  "repl.context_usage": "Usage: /context [<size> | system <fraction> | reserve <fraction>]  (min 1024; fractions 0.05–0.9)",
   "repl.msgs": "msgs",
   "repl.max_iters": "Max iters:",
   "repl.stuck_thresh": "Stuck thresh:",
@@ -715,6 +733,13 @@
   "moe.partial_result": "Partial result",
   "moe.scope_approved": "🔓 Scope expanded for \"{subtask}\": {files}",
   "moe.scope_rejected": "⛔ Scope request rejected for \"{subtask}\"",
+  "prompt.overflow.compressing": "System prompt overflow: compressing {count} block(s) before the first run — this may take a while on a local model.",
+  "prompt.overflow.summarizing": "System prompt overflow: summarizing \"{label}\" ({original} tok → {budget} tok budget) with the model…",
+  "prompt.overflow.truncating": "System prompt overflow: truncating \"{label}\" ({original} tok) — summarization unavailable.",
+  "prompt.overflow.summarize_failed": "System prompt overflow: summarization failed for \"{label}\": {error} — falling back to truncation.",
+  "prompt.overflow.hint_needed": "the prompt needs ~{needed} system tokens (current budget: {window} × {fraction} = {budget})",
+  "prompt.overflow.hint_window": "raise contextWindow to at least {required} (standard size {recommended}) — {how}",
+  "prompt.overflow.hint_fraction": "or keep the window and raise contextBudget.systemPrompt to ~{fraction} (run: mma context --system {fraction})",
   "prompt.overflow.exceeded": "Prompt overflow: \"{label}\" ({original} tok) exceeded the system-prompt budget ({budget} tok) — {mode} to {resolved} tok. To fit fully, {hint}.",
   "prompt.overflow.failed": "Prompt overflow resolution failed: {error} — oversized blocks will be dropped.",
   "prompt.overflow.startup": "Startup check: {block} ({original} tok) exceeds the system-prompt budget ({budget} tok) — it will be summarized/truncated before the first run. To include it fully, {hint}.",

package/dist/i18n/ru.json

@@ -51,6 +51,16 @@
   "env.playwright_missing": "Пакет Playwright не установлен — браузерный инструмент не будет работать. Установите: bun add playwright",
   "env.playwright_browsers_missing": "Браузеры Playwright не скачаны ({dir}). Браузерный инструмент не будет работать. Установите: bunx playwright install chromium",
   "env.crash_stderr": "MMA упал ({type}): {message} — отчёт о падении записан в ~/.mma/logs/crash.jsonl",
+  "env.model_autoload_failed": "Не удалось автоматически загрузить модель: {error}. Продолжаю с ручной загрузкой.",
+  "env.model_loaded": "Модель {model} загружена за {seconds}с",
+  "env.reasoning_probe": "Проверка reasoning: стратегия={strategy} → {result}",
+  "env.reasoning_respected": "механизм соблюдается",
+  "env.reasoning_ignored": "механизм ИГНОРИРУЕТСЯ",
+  "env.indexing": "Индексация файлов проекта…",
+  "env.indexed": "Проиндексировано файлов: {files} (карта проекта)",
+  "env.indexing_failed": "Ошибка индексации проекта: {error}",
+  "env.index_skip_unreadable": "индексатор: пропускаю недоступный файл \"{path}\"",
+  "env.killed_processes": "Остановлено фоновых процессов при завершении: {count}",
   "session.started": "Сессия начата: {id}",
   "session.ended": "Сессия завершена: {id}",
   "session.not_found": "Сессия не найдена: {id}",
@@ -267,6 +277,14 @@
   "cli.manage_context": "Управление контекстным окном",
   "cli.invalid_context_size": "Некорректный размер контекста. Должно быть число >= 1024",
   "cli.context_set": "Контекстное окно установлено: {size} токенов",
+  "cli.context_budget_header": "Бюджет контекста (окно {window} токенов):",
+  "cli.context_system_line": "  системный промпт: {tokens} токенов ({percent}%)",
+  "cli.context_reserve_line": "  резерв ответа: {tokens} токенов ({percent}%)",
+  "cli.context_history_line": "  история: {tokens} токенов ({percent}%)",
+  "cli.context_system_fraction": "Доля системного промпта в контекстном окне (0.05–0.9)",
+  "cli.context_reserve_fraction": "Доля резерва ответа в контекстном окне (0.05–0.9)",
+  "cli.context_fraction_set": "Доля {key} установлена: {value}",
+  "cli.context_invalid_fraction": "Некорректная доля. Нужно число от 0.05 до 0.9 (системный + резерв должны оставаться ниже 0.95).",
   "cli.set_model": "Установить модель по умолчанию",
   "cli.model_set": "Модель установлена: {name}",
   "cli.certify": "Запустить сертификационный набор для модели",
@@ -426,7 +444,7 @@
   "repl.sysprompt_desc": "Показать системный промпт",
   "repl.sysprompt_tokens": "Системный промпт ({count} токенов):",
   "repl.excluded_blocks": "\nИсключённые блоки: {count}",
-  "repl.context_usage": "Использование: /context <размер>  (мин 1024)",
+  "repl.context_usage": "Использование: /context [<размер> | system <доля> | reserve <доля>]  (мин 1024; доли 0.05–0.9)",
   "repl.msgs": "сообщ.",
   "repl.max_iters": "Макс итераций:",
   "repl.stuck_thresh": "Порог зависания:",
@@ -715,6 +733,13 @@
   "moe.partial_result": "Частичный результат",
   "moe.scope_approved": "🔓 Scope расширен для \"{subtask}\": {files}",
   "moe.scope_rejected": "⛔ Запрос scope отклонён для \"{subtask}\"",
+  "prompt.overflow.compressing": "Переполнение системного промпта: сжимаю {count} блок(ов) перед первым запуском — на локальной модели это может занять время.",
+  "prompt.overflow.summarizing": "Переполнение системного промпта: суммаризирую \"{label}\" ({original} токенов → бюджет {budget} токенов) моделью…",
+  "prompt.overflow.truncating": "Переполнение системного промпта: обрезаю \"{label}\" ({original} токенов) — суммаризация недоступна.",
+  "prompt.overflow.summarize_failed": "Переполнение системного промпта: не удалось суммаризировать \"{label}\": {error} — перехожу к обрезке.",
+  "prompt.overflow.hint_needed": "промпту нужно ~{needed} токенов системного бюджета (сейчас: {window} × {fraction} = {budget})",
+  "prompt.overflow.hint_window": "подними contextWindow минимум до {required} (стандартный размер {recommended}) — {how}",
+  "prompt.overflow.hint_fraction": "или оставь окно и подними contextBudget.systemPrompt до ~{fraction} (выполни: mma context --system {fraction})",
   "prompt.overflow.exceeded": "Промпт переполнен: \"{label}\" ({original} токенов) превысил бюджет системного промпта ({budget} токенов) — {mode} до {resolved} токенов. Чтобы включить полностью, {hint}.",
   "prompt.overflow.failed": "Не удалось разрешить переполнение промпта: {error} — слишком большие блоки будут отброшены.",
   "prompt.overflow.startup": "Проверка при старте: {block} ({original} токенов) превышает бюджет системного промпта ({budget} токенов) — будет суммаризован/обрезан перед первым запуском. Чтобы включить полностью, {hint}.",

package/dist/main.js

@@ -2492,6 +2492,16 @@ The path was joined onto the working directory because it does not exist as give
     "env.playwright_missing": "Playwright package is not installed — the browser tool will fail. Install with: bun add playwright",
     "env.playwright_browsers_missing": "Playwright browsers not downloaded ({dir}). The browser tool will fail. Install with: bunx playwright install chromium",
     "env.crash_stderr": "MMA crashed ({type}): {message} — crash report written to ~/.mma/logs/crash.jsonl",
+    "env.model_autoload_failed": "Model auto-load failed: {error}. Continuing with manual load.",
+    "env.model_loaded": "Model {model} loaded in {seconds}s",
+    "env.reasoning_probe": "Reasoning probe: strategy={strategy} → {result}",
+    "env.reasoning_respected": "mechanism respected",
+    "env.reasoning_ignored": "mechanism IGNORED",
+    "env.indexing": "Indexing project files…",
+    "env.indexed": "Indexed {files} file(s) for the project map",
+    "env.indexing_failed": "Project indexing failed: {error}",
+    "env.index_skip_unreadable": 'indexer: skipping unreadable entry "{path}"',
+    "env.killed_processes": "Killed {count} background process(es) on shutdown",
     "session.started": "Session started: {id}",
     "session.ended": "Session ended: {id}",
     "session.not_found": "Session not found: {id}",
@@ -2718,6 +2728,14 @@ Fix the error and re-edit the file (a clean write clears the failure), or mark t
     "cli.manage_context": "Manage context window",
     "cli.invalid_context_size": "Invalid context size. Must be a number >= 1024",
     "cli.context_set": "Context window set to: {size} tokens",
+    "cli.context_budget_header": "Context budget (window {window} tokens):",
+    "cli.context_system_line": "  system prompt: {tokens} tokens ({percent}%)",
+    "cli.context_reserve_line": "  response reserve: {tokens} tokens ({percent}%)",
+    "cli.context_history_line": "  history: {tokens} tokens ({percent}%)",
+    "cli.context_system_fraction": "System-prompt share of the context window (0.05–0.9)",
+    "cli.context_reserve_fraction": "Response-reserve share of the context window (0.05–0.9)",
+    "cli.context_fraction_set": "{key} share set to {value}",
+    "cli.context_invalid_fraction": "Invalid share. Use a number between 0.05 and 0.9 (system + reserve must stay below 0.95).",
     "repl.reload": "Reload agent with current config",
     "repl.reload_usage": "Usage: /reload",
     "repl.reloading": "Reloading agent...",
@@ -2883,7 +2901,7 @@ Available commands:`,
     "repl.sysprompt_tokens": "System prompt ({count} tokens):",
     "repl.excluded_blocks": `
 Excluded blocks: {count}`,
-    "repl.context_usage": "Usage: /context <size>  (min 1024)",
+    "repl.context_usage": "Usage: /context [<size> | system <fraction> | reserve <fraction>]  (min 1024; fractions 0.05–0.9)",
     "repl.msgs": "msgs",
     "repl.max_iters": "Max iters:",
     "repl.stuck_thresh": "Stuck thresh:",
@@ -3192,6 +3210,13 @@ Apply a matching solution from these results. If none is relevant — do NOT rep
     "moe.partial_result": "Partial result",
     "moe.scope_approved": '🔓 Scope expanded for "{subtask}": {files}',
     "moe.scope_rejected": '⛔ Scope request rejected for "{subtask}"',
+    "prompt.overflow.compressing": "System prompt overflow: compressing {count} block(s) before the first run — this may take a while on a local model.",
+    "prompt.overflow.summarizing": 'System prompt overflow: summarizing "{label}" ({original} tok → {budget} tok budget) with the model…',
+    "prompt.overflow.truncating": 'System prompt overflow: truncating "{label}" ({original} tok) — summarization unavailable.',
+    "prompt.overflow.summarize_failed": 'System prompt overflow: summarization failed for "{label}": {error} — falling back to truncation.',
+    "prompt.overflow.hint_needed": "the prompt needs ~{needed} system tokens (current budget: {window} × {fraction} = {budget})",
+    "prompt.overflow.hint_window": "raise contextWindow to at least {required} (standard size {recommended}) — {how}",
+    "prompt.overflow.hint_fraction": "or keep the window and raise contextBudget.systemPrompt to ~{fraction} (run: mma context --system {fraction})",
     "prompt.overflow.exceeded": 'Prompt overflow: "{label}" ({original} tok) exceeded the system-prompt budget ({budget} tok) — {mode} to {resolved} tok. To fit fully, {hint}.',
     "prompt.overflow.failed": "Prompt overflow resolution failed: {error} — oversized blocks will be dropped.",
     "prompt.overflow.startup": "Startup check: {block} ({original} tok) exceeds the system-prompt budget ({budget} tok) — it will be summarized/truncated before the first run. To include it fully, {hint}.",
@@ -3282,6 +3307,16 @@ var init_ru = __esm(() => {
     "env.playwright_missing": "Пакет Playwright не установлен — браузерный инструмент не будет работать. Установите: bun add playwright",
     "env.playwright_browsers_missing": "Браузеры Playwright не скачаны ({dir}). Браузерный инструмент не будет работать. Установите: bunx playwright install chromium",
     "env.crash_stderr": "MMA упал ({type}): {message} — отчёт о падении записан в ~/.mma/logs/crash.jsonl",
+    "env.model_autoload_failed": "Не удалось автоматически загрузить модель: {error}. Продолжаю с ручной загрузкой.",
+    "env.model_loaded": "Модель {model} загружена за {seconds}с",
+    "env.reasoning_probe": "Проверка reasoning: стратегия={strategy} → {result}",
+    "env.reasoning_respected": "механизм соблюдается",
+    "env.reasoning_ignored": "механизм ИГНОРИРУЕТСЯ",
+    "env.indexing": "Индексация файлов проекта…",
+    "env.indexed": "Проиндексировано файлов: {files} (карта проекта)",
+    "env.indexing_failed": "Ошибка индексации проекта: {error}",
+    "env.index_skip_unreadable": 'индексатор: пропускаю недоступный файл "{path}"',
+    "env.killed_processes": "Остановлено фоновых процессов при завершении: {count}",
     "session.started": "Сессия начата: {id}",
     "session.ended": "Сессия завершена: {id}",
     "session.not_found": "Сессия не найдена: {id}",
@@ -3507,6 +3542,14 @@ var init_ru = __esm(() => {
     "cli.manage_context": "Управление контекстным окном",
     "cli.invalid_context_size": "Некорректный размер контекста. Должно быть число >= 1024",
     "cli.context_set": "Контекстное окно установлено: {size} токенов",
+    "cli.context_budget_header": "Бюджет контекста (окно {window} токенов):",
+    "cli.context_system_line": "  системный промпт: {tokens} токенов ({percent}%)",
+    "cli.context_reserve_line": "  резерв ответа: {tokens} токенов ({percent}%)",
+    "cli.context_history_line": "  история: {tokens} токенов ({percent}%)",
+    "cli.context_system_fraction": "Доля системного промпта в контекстном окне (0.05–0.9)",
+    "cli.context_reserve_fraction": "Доля резерва ответа в контекстном окне (0.05–0.9)",
+    "cli.context_fraction_set": "Доля {key} установлена: {value}",
+    "cli.context_invalid_fraction": "Некорректная доля. Нужно число от 0.05 до 0.9 (системный + резерв должны оставаться ниже 0.95).",
     "cli.set_model": "Установить модель по умолчанию",
     "cli.model_set": "Модель установлена: {name}",
     "cli.certify": "Запустить сертификационный набор для модели",
@@ -3668,7 +3711,7 @@ var init_ru = __esm(() => {
     "repl.sysprompt_tokens": "Системный промпт ({count} токенов):",
     "repl.excluded_blocks": `
 Исключённые блоки: {count}`,
-    "repl.context_usage": "Использование: /context <размер>  (мин 1024)",
+    "repl.context_usage": "Использование: /context [<размер> | system <доля> | reserve <доля>]  (мин 1024; доли 0.05–0.9)",
     "repl.msgs": "сообщ.",
     "repl.max_iters": "Макс итераций:",
     "repl.stuck_thresh": "Порог зависания:",
@@ -3984,6 +4027,13 @@ var init_ru = __esm(() => {
     "moe.partial_result": "Частичный результат",
     "moe.scope_approved": '🔓 Scope расширен для "{subtask}": {files}',
     "moe.scope_rejected": '⛔ Запрос scope отклонён для "{subtask}"',
+    "prompt.overflow.compressing": "Переполнение системного промпта: сжимаю {count} блок(ов) перед первым запуском — на локальной модели это может занять время.",
+    "prompt.overflow.summarizing": 'Переполнение системного промпта: суммаризирую "{label}" ({original} токенов → бюджет {budget} токенов) моделью…',
+    "prompt.overflow.truncating": 'Переполнение системного промпта: обрезаю "{label}" ({original} токенов) — суммаризация недоступна.',
+    "prompt.overflow.summarize_failed": 'Переполнение системного промпта: не удалось суммаризировать "{label}": {error} — перехожу к обрезке.',
+    "prompt.overflow.hint_needed": "промпту нужно ~{needed} токенов системного бюджета (сейчас: {window} × {fraction} = {budget})",
+    "prompt.overflow.hint_window": "подними contextWindow минимум до {required} (стандартный размер {recommended}) — {how}",
+    "prompt.overflow.hint_fraction": "или оставь окно и подними contextBudget.systemPrompt до ~{fraction} (выполни: mma context --system {fraction})",
     "prompt.overflow.exceeded": 'Промпт переполнен: "{label}" ({original} токенов) превысил бюджет системного промпта ({budget} токенов) — {mode} до {resolved} токенов. Чтобы включить полностью, {hint}.',
     "prompt.overflow.failed": "Не удалось разрешить переполнение промпта: {error} — слишком большие блоки будут отброшены.",
     "prompt.overflow.startup": "Проверка при старте: {block} ({original} токенов) превышает бюджет системного промпта ({budget} токенов) — будет суммаризован/обрезан перед первым запуском. Чтобы включить полностью, {hint}.",
@@ -14629,6 +14679,9 @@ async function resolvePromptOverflow(opts) {
   const warnings = [];
   let used = includedTokens;
   let unresolved = [];
+  if (sorted.length > 0) {
+    logger?.warn(t("prompt.overflow.compressing", { count: String(sorted.length) }));
+  }
   for (const block of sorted) {
     const kind = block.kind;
     const maxTokens = systemBudget - used - HINT_BLOCK_TOKENS;
@@ -14646,6 +14699,11 @@ async function resolvePromptOverflow(opts) {
       logger?.debug(`Prompt overflow: cache hit for ${kind} (${key}, ${mode})`);
     } else if (provider) {
       const targetChars = Math.floor(maxTokens * (CHARS_PER_TOKEN - 1));
+      logger?.warn(t("prompt.overflow.summarizing", {
+        label: blockLabel2(block.content),
+        original: String(block.estimatedTokens),
+        budget: String(maxTokens)
+      }));
       try {
         const prompt = [
           `Summarize the following ${KIND_LABEL[kind]} document.`,
@@ -14675,12 +14733,19 @@ async function resolvePromptOverflow(opts) {
 `), maxTokens);
         }
       } catch (err) {
-        logger?.warn(`Prompt overflow: summarization failed for ${kind}: ${err?.message ?? err}`);
+        logger?.warn(t("prompt.overflow.summarize_failed", {
+          label: blockLabel2(block.content),
+          error: String(err?.message ?? err)
+        }));
         text = "";
       }
     }
     if (!text) {
       mode = "truncate";
+      logger?.warn(t("prompt.overflow.truncating", {
+        label: blockLabel2(block.content),
+        original: String(block.estimatedTokens)
+      }));
       text = truncateToTokens(block.content, maxTokens);
     }
     if (estimateTokens(text) > maxTokens) {
@@ -14741,10 +14806,38 @@ function recommendContextSize(neededSystemTokens) {
   const sizes = [8192, 16384, 32768, 65536, 131072, 262144];
   return sizes.find((s) => Math.floor(s * 0.1) >= neededSystemTokens) ?? sizes[sizes.length - 1];
 }
+function requiredContextWindow(neededTokens, systemFraction) {
+  const fraction = systemFraction > 0 ? systemFraction : 0.1;
+  return Math.ceil(neededTokens / fraction);
+}
+function requiredSystemFraction(neededTokens, contextWindow) {
+  if (contextWindow <= 0)
+    return 1;
+  return Math.min(1, Math.ceil(neededTokens / contextWindow * 100) / 100);
+}
+function overflowHint(config, configDir, neededTokens) {
+  const fraction = config.contextBudget?.systemPrompt ?? 0.1;
+  const budget = Math.floor(config.contextWindow * fraction);
+  const recommended = recommendContextSize(neededTokens);
+  const requiredWindow = requiredContextWindow(neededTokens, fraction);
+  const requiredFraction = requiredSystemFraction(neededTokens, config.contextWindow);
+  const how = configDir ? contextWindowHint(config, configDir, recommended) : `increase contextWindow (e.g. to ${recommended})`;
+  return [
+    t("prompt.overflow.hint_needed", {
+      needed: neededTokens,
+      window: config.contextWindow,
+      fraction,
+      budget
+    }),
+    t("prompt.overflow.hint_window", { required: requiredWindow, recommended, how }),
+    t("prompt.overflow.hint_fraction", { fraction: requiredFraction })
+  ].join(" ");
+}
 var HINT_BLOCK_TOKENS = 60, CHARS_PER_TOKEN = 4, KIND_ORDER, KIND_LABEL, KIND_SOURCE_FILE;
 var init_prompt_overflow = __esm(() => {
   init_prompt_builder();
   init_token_counter();
+  init_i18n();
   KIND_ORDER = ["instructions", "project-map"];
   KIND_LABEL = {
     instructions: "project instructions (AGENTS.md)",
@@ -14920,8 +15013,8 @@ class Agent {
         this.promptOverrides.set(r.kind, r);
     }
     this.overflowHintBlock = res.hintBlock;
-    const recommended = recommendContextSize(dry.blocks.filter((b) => b.included).reduce((s, b) => s + b.tokens, 0) + overflow.reduce((s, b) => s + b.estimatedTokens, 0) + HINT_BLOCK_TOKENS);
-    const hint = this.deps.configDir ? contextWindowHint(cfg, this.deps.configDir, recommended) : `increase contextWindow (e.g. to ${recommended})`;
+    const needed = dry.blocks.filter((b) => b.included).reduce((s, b) => s + b.tokens, 0) + overflow.reduce((s, b) => s + b.estimatedTokens, 0) + HINT_BLOCK_TOKENS;
+    const hint = overflowHint(cfg, this.deps.configDir, needed);
     for (const w of res.warnings) {
       this.deps.logger.warn(t("prompt.overflow.exceeded", {
         label: w.label,
@@ -15483,7 +15576,7 @@ class Agent {
     contextManager.onCompact = null;
     const killed = processRegistry.killAll();
     if (killed > 0) {
-      logger.info(`Killed ${killed} background process(es) on shutdown`);
+      logger.info(t("env.killed_processes", { count: String(killed) }));
     }
     pluginManager.runOnSessionEnd({
       logger,
@@ -24926,7 +25019,7 @@ class Indexer {
             }
           }
         } catch (e) {
-          logger2.warn(`indexer: skipping unreadable entry "${relPath}"`, { error: String(e) });
+          logger2.warn(t("env.index_skip_unreadable", { path: relPath }), { error: String(e) });
           continue;
         }
       }
@@ -24947,6 +25040,7 @@ var logger2, IGNORE_DIRS;
 var init_walker = __esm(() => {
   init_app_logger();
   init_symbols();
+  init_i18n();
   logger2 = new Logger("warn", "indexer");
   IGNORE_DIRS = new Set([
     ".git",
@@ -26524,11 +26618,11 @@ async function bootstrap(configDir, projectDir, noAgentsMd, exitOnComplete, reas
       autoLoad: config.modelLoad.autoLoad
     });
     if (!loadResult.success) {
-      logger4.warn(`Model auto-load failed: ${loadResult.error}. Continuing with manual load.`);
+      logger4.warn(t("env.model_autoload_failed", { error: String(loadResult.error ?? "") }));
     } else if (loadResult.alreadyLoaded) {
       logger4.debug(`Model ${config.model} already loaded`);
     } else {
-      logger4.info(`Model ${config.model} loaded in ${loadResult.loadTime}s`);
+      logger4.info(t("env.model_loaded", { model: config.model, seconds: String(loadResult.loadTime) }));
     }
   }
   const contextProbePromise = (async () => {
@@ -26564,7 +26658,10 @@ async function bootstrap(configDir, projectDir, noAgentsMd, exitOnComplete, reas
         const passed = cached !== undefined ? cached : await probeReasoningSupport2(llmProvider, reasoningStrategy);
         if (cached === undefined && passed !== null)
           setCachedProbeResult2(cacheK, passed);
-        logger4.info(`Reasoning probe: strategy=${reasoningStrategy} → ${passed === true ? "mechanism respected" : "mechanism IGNORED"}`);
+        logger4.info(t("env.reasoning_probe", {
+          strategy: reasoningStrategy,
+          result: passed === true ? t("env.reasoning_respected") : t("env.reasoning_ignored")
+        }));
         return passed === true;
       } catch {
         return false;
@@ -26588,9 +26685,15 @@ async function bootstrap(configDir, projectDir, noAgentsMd, exitOnComplete, reas
     cacheDir: projectMapCacheDir
   });
   try {
-    await indexerModule.buildIndex();
+    logger4.info(t("env.indexing"));
+    const index = await indexerModule.buildIndex();
+    if (index) {
+      logger4.info(t("env.indexed", { files: String(index.files.length) }));
+    } else {
+      logger4.warn(t("env.indexing_failed", { error: "no index produced" }));
+    }
   } catch (err) {
-    logger4.warn(`Project indexing failed: ${err.message}`);
+    logger4.warn(t("env.indexing_failed", { error: String(err?.message ?? err) }));
   }
   const skillsLoader = new SkillsLoader;
   const builtinDir = join46(import.meta.dirname, "skills", "builtin");
@@ -26782,8 +26885,7 @@ async function bootstrap(configDir, projectDir, noAgentsMd, exitOnComplete, reas
     const dry = dryRunOverflow([...promptBlocks, ...dynamicBlocks], systemBudget);
     if (dry.overflow.length > 0) {
       const overflowTokens = dry.overflow.reduce((s, b) => s + b.estimatedTokens, 0);
-      const recommended = recommendContextSize(dry.includedTokens + overflowTokens);
-      const hint = contextWindowHint(config, dir, recommended);
+      const hint = overflowHint(config, dir, dry.includedTokens + overflowTokens);
       for (const b of dry.overflow) {
         logger4.warn(t("prompt.overflow.startup", {
           block: b.kind === "instructions" ? "AGENTS.md" : "project map",
@@ -27688,6 +27790,49 @@ var init_map_command = __esm(() => {
   init_i18n();
 });
 
+// src/config/budget.ts
+function budgetBreakdown(config) {
+  const window = config.contextWindow;
+  const budget = config.contextBudget;
+  const system = Math.floor(window * budget.systemPrompt);
+  const reserve = Math.floor(window * budget.responseReserve);
+  const history = window - system - reserve;
+  return {
+    window,
+    system,
+    reserve,
+    history,
+    systemFraction: budget.systemPrompt,
+    reserveFraction: budget.responseReserve,
+    historyFraction: window > 0 ? history / window : 0
+  };
+}
+function setBudgetShare(config, key, value) {
+  if (!Number.isFinite(value) || value < MIN_SHARE || value > MAX_SHARE)
+    return "range";
+  const budget = config.contextBudget;
+  const nextSystem = key === "system" ? value : budget.systemPrompt;
+  const nextReserve = key === "reserve" ? value : budget.responseReserve;
+  if (nextSystem + nextReserve > MAX_COMBINED)
+    return "combined";
+  budget[key === "system" ? "systemPrompt" : "responseReserve"] = value;
+  return null;
+}
+function budgetBreakdownLines(config) {
+  const bd = budgetBreakdown(config);
+  const pct = (f) => Math.round(f * 100);
+  return [
+    t("cli.context_budget_header", { window: bd.window }),
+    t("cli.context_system_line", { tokens: bd.system, percent: pct(bd.systemFraction) }),
+    t("cli.context_reserve_line", { tokens: bd.reserve, percent: pct(bd.reserveFraction) }),
+    t("cli.context_history_line", { tokens: bd.history, percent: pct(bd.historyFraction) })
+  ];
+}
+var MIN_SHARE = 0.05, MAX_SHARE = 0.9, MAX_COMBINED = 0.95;
+var init_budget = __esm(() => {
+  init_i18n();
+});
+
 // node_modules/yaml/dist/nodes/identity.js
 var require_identity = __commonJS((exports) => {
   var ALIAS = Symbol.for("yaml.alias");
@@ -36239,11 +36384,28 @@ function registerProviderCommands(ctx) {
   ctx.registerCommand({
     name: "context",
     description: t("cli.manage_context"),
-    usage: "/context <size>",
+    usage: t("repl.context_usage"),
     action: async (args) => {
+      const save = async () => {
+        const configPath = join54(ctx.configDir, "config.json");
+        saveConfig(ctx.config, configPath, dirname24(configPath));
+        await ctx.agent.reconfigure(ctx.config);
+      };
       if (args.length === 0) {
-        console.log(`/context ${t("repl.context")}  ${ctx.config.contextWindow}`);
-        console.log(t("repl.context_usage"));
+        for (const line of budgetBreakdownLines(ctx.config))
+          console.log(line);
+        console.log(pc2.dim(t("repl.context_usage")));
+        return;
+      }
+      if (args[0] === "system" || args[0] === "reserve") {
+        const key = args[0];
+        const value = Number(args[1]);
+        if (setBudgetShare(ctx.config, key, value) !== null) {
+          console.log(pc2.yellow(t("cli.context_invalid_fraction")));
+          return;
+        }
+        await save();
+        console.log(pc2.green(t("cli.context_fraction_set", { key, value })));
         return;
       }
       const size = parseInt(args[0], 10);
@@ -36252,9 +36414,7 @@ function registerProviderCommands(ctx) {
         return;
       }
       ctx.config.contextWindow = size;
-      const configPath = join54(ctx.configDir, "config.json");
-      saveConfig(ctx.config, configPath, dirname24(configPath));
-      await ctx.agent.reconfigure(ctx.config);
+      await save();
       console.log(pc2.green(t("cli.context_set", { size })));
     }
   });
@@ -36641,6 +36801,7 @@ var init_repl_commands = __esm(() => {
   init_i18n();
   init_setup();
   init_config2();
+  init_budget();
   init_token_counter();
   init_map_command();
   init_utils();
@@ -37372,6 +37533,7 @@ function targetsFromProviders(entries) {
 // src/cli/commands.ts
 init_version();
 init_map_command();
+init_budget();
 
 // src/modules/updater/changelog-reader.ts
 import { readFileSync as readFileSync35, existsSync as existsSync53 } from "fs";
@@ -37632,9 +37794,25 @@ function buildModelCommands(program2) {
   });
 }
 function buildContextCommand(program2) {
-  program2.command("context").description(t("cli.manage_context")).argument("<size>", "Context window size in tokens").action(async (size) => {
-    const configPath = join53(homedir17(), ".mma", "config.json");
-    const { config } = await bootstrap();
+  program2.command("context").description(t("cli.manage_context")).argument("[size]", "Context window size in tokens (omit to show the budget breakdown)").option("--system <fraction>", t("cli.context_system_fraction")).option("--reserve <fraction>", t("cli.context_reserve_fraction")).action(async (size, opts) => {
+    const { config, configDir } = await bootstrap();
+    const configPath = join53(configDir, "config.json");
+    if (opts.system !== undefined || opts.reserve !== undefined) {
+      const key = opts.system !== undefined ? "system" : "reserve";
+      const value = Number(opts.system ?? opts.reserve);
+      if (setBudgetShare(config, key, value) !== null) {
+        console.log(t("cli.context_invalid_fraction"));
+        return;
+      }
+      saveConfig(config, configPath, dirname23(configPath));
+      console.log(t("cli.context_fraction_set", { key, value }));
+      return;
+    }
+    if (size === undefined) {
+      for (const line of budgetBreakdownLines(config))
+        console.log(line);
+      return;
+    }
     const contextWindow = parseInt(size, 10);
     if (isNaN(contextWindow) || contextWindow < 1024) {
       console.log(t("cli.invalid_context_size"));

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "micro-models-agent",
-  "version": "0.60.0",
+  "version": "0.61.0",
   "description": "Micro Models Agent (MMA) — LLM agent harness for small models (Qwen3.5-9B, 32K-64K context)",
   "type": "module",
   "bin": {