fix(agent): native and narrated tool calls reach the client, runaway text-channel calls are cut on both /v1/messages and /v1/chat/completions; MODEL_MAP alias mapping with a dashboard editor

.env.example

@@ -18,6 +18,20 @@ SERVICE_PORT=3000
 # retry regenerates from scratch and is the longest such window.
 # ANTHROPIC_PING_INTERVAL_MS=15000
 
+# Anthropic (/v1/messages) 与 OpenAI (/v1/chat/completions) 两条路径共用：一轮 Agent
+# 回合里文本通道工具调用的上限（默认 24，钳在 4..256）。
+# 模型在叙述的 [TOOL CALL] 之后失控（同一调用重复上百次、幻想整段 agent 会话）时，
+# 第 N 个已放行的调用一到就终止上游，已放行的调用照常交付
+# （stop_reason=tool_use / finish_reason=tool_calls）。更早的 delta 里已放行过调用之后，
+# 再出现重复调用 / 被拒绝的调用 / 正文或思考同样立刻截断，不受此上限影响。
+# Both agent paths (/v1/messages and /v1/chat/completions): cap on text-channel tool
+# calls per agent turn (default 24, clamped to 4..256). When the model runs away after a
+# narrated [TOOL CALL] (repeats the same call hundreds of times, hallucinates a whole
+# agentic session), the upstream is cut right after the N-th admitted call and the
+# admitted calls are still delivered (stop_reason=tool_use / finish_reason=tool_calls).
+# Once a call was admitted in an earlier delta, a duplicate, a rejected call, or
+# prose/thinking after it also cuts the turn, regardless of this cap.
+# AGENT_TURN_MAX_TOOL_CALLS=24
 
 # 监听地址(非必填)
 # Listen address (optional)
@@ -169,15 +183,45 @@ QWEN_CLI_PROXY_URL=
 # 示例 / Example: PROXY_URL=socks5://127.0.0.1:1080
 PROXY_URL=
 
-# ========== Claude Code 兼容配置 / Claude Code Compatibility ==========
-
-# Claude-to-Qwen 模型映射（可选，内置默认映射已覆盖主流 Claude 模型）
-# 格式: claude-model-name=qwen-model-id,claude-model-name2=qwen-model-id2
-# 未匹配的 claude-* 模型自动回退到 qwen3-coder-plus
-# Claude-to-Qwen model mapping (optional; built-in defaults cover mainstream Claude models)
-# Format: claude-model-name=qwen-model-id,claude-model-name2=qwen-model-id2
-# Unmatched claude-* models fall back to qwen3-coder-plus
-# CLAUDE_MODEL_MAP=claude-sonnet-5=qwen3-coder-plus,claude-opus-4=qwen3-max
+# ========== 入站模型名映射 / Incoming model name mapping ==========
+
+# 把客户端发来的模型名映射成 Qwen 模型 id，只作用于 /v1/chat/completions 和 /v1/messages
+# （图片、视频、CLI 端点不走这里）。Claude Code 的子代理会发 claude-opus-5 / claude-haiku-*
+# （除非客户端自己设置了 ANTHROPIC_DEFAULT_OPUS/SONNET/HAIKU_MODEL 或 CLAUDE_CODE_SUBAGENT_MODEL；
+# 服务端映射不需要任何客户端配置），OpenAI 风格客户端会发 gpt-*；不映射时上游返回 "Model not found"。
+# 规则（两个端点相同）：
+#   1. 精确匹配优先，不区分大小写，末尾的 [..] 后缀先去掉（claude-opus-5[1m] 按 claude-opus-5 处理）
+#   2. 上游已存在的 Qwen id（含 -thinking 等变体）原样透传，不受 * 影响
+#   3. 其余名字用 * 条目；没有 * 时用上游第一个 t2t 模型并打印一条 warn。
+#      仅在上游模型列表可用时生效：列表取不到时名字原样转发（打印一条 warn），不套用 *
+# 目标 id 可带 -thinking 等后缀，后缀照常生效（会打开思考）。响应里的 model 字段回显解析后的
+# Qwen id，不是别名。别名不需要出现在 /v1/models 里。落到回退目标的名字记录在进程内存中
+# （每个 PM2 worker 一份，最多 100 个）。
+# Maps incoming model names to Qwen model ids; applies to /v1/chat/completions and /v1/messages
+# only (not images/videos/cli). Claude Code subagents send claude-opus-5 / claude-haiku-* (unless
+# the client sets ANTHROPIC_DEFAULT_OPUS/SONNET/HAIKU_MODEL or CLAUDE_CODE_SUBAGENT_MODEL; the
+# server-side map needs no client config), OpenAI-style clients send gpt-*; without a map the
+# upstream answers "Model not found". Rule (same on both endpoints):
+#   1. exact entry wins, case-insensitive; a trailing [..] suffix is stripped first (claude-opus-5[1m] = claude-opus-5)
+#   2. names that already exist upstream (incl. -thinking variants) pass through, even with *
+#   3. everything else uses the * entry; with no * the first upstream t2t model is used with a warn.
+#      Only while the upstream model list is available: if it cannot be fetched the name is
+#      forwarded unchanged (one warn) and * is not applied
+# Targets may carry suffixes such as -thinking; they apply as usual (thinking switches on). The
+# response `model` field echoes the resolved Qwen id, not the alias. Aliases work without being
+# listed in /v1/models. Names that fell to the fallback are recorded in process memory (one list
+# per PM2 worker, 100 max).
+# Dashboard：系统设置里的「模型映射」卡片可在线编辑。dashboard 保存过的映射优先于本变量（重启后仍生效）；
+#   DATA_SAVE_MODE=none 时 dashboard 的修改只在内存里生效，重启即丢；「恢复 env 映射」会清掉保存的映射，
+#   本变量重新生效。PM2 多 worker 时，保存的映射只在处理请求的 worker 立即生效，其他 worker 重启后才读到。
+# Dashboard: the "Model mapping" card in Settings edits this at runtime. A dashboard-saved map takes
+#   precedence over this variable (and survives restarts); with DATA_SAVE_MODE=none dashboard changes
+#   live in memory only and are lost on restart; "Restore env map" clears the saved map so this variable
+#   applies again. With several PM2 workers a saved map applies at once only in the worker that handled
+#   the save; the others pick it up at their next restart.
+# 格式 / Format: alias=qwen-model-id,alias2=qwen-model-id2,*=fallback-qwen-model-id
+# 示例 / Example: MODEL_MAP=claude-opus-5=qwen3.8-max,*=qwen3.8-max-thinking
+MODEL_MAP=
 
 # ========== CLI 配置 / CLI Configuration ==========
 

README-en.md

@@ -102,6 +102,7 @@ OUTPUT_THINK=true             # Whether to output thinking process (true/false)
 LEGACY_REASONING_IN_CONTENT=false # Reasoning format, false=reasoning_content field, true=legacy <think> inside content (true/false)
 SIMPLE_MODEL_MAP=false        # Simplify model mapping (true/false)
 MODELS_CACHE_TTL=3600         # Model list cache TTL in seconds, 0=never expires
+AGENT_TURN_MAX_TOOL_CALLS=24  # Anthropic path: text-channel tool_use cap per agent turn (4-256), upstream cut after it
 
 # 🌐 Proxy and Reverse Proxy Configuration
 QWEN_CHAT_PROXY_URL=          # Custom Chat API reverse proxy URL (default: https://chat.qwen.ai)
@@ -130,7 +131,9 @@ CACHE_MODE=default            # Image cache mode (default/file)
 | `OUTPUT_THINK` | Whether to show AI thinking process | `true` or `false` |
 | `LEGACY_REASONING_IN_CONTENT` | Reasoning output format. Default `false` = reasoning goes to a separate `reasoning_content` field; `true` = legacy behavior (`<think>` inside `content`) | `true` or `false` |
 | `SIMPLE_MODEL_MAP` | Simplify model mapping, return basic models without variants only | `true` or `false` |
+| `MODEL_MAP` | Incoming model name mapping: `alias=qwen-id,...,*=fallback`. Exact entry wins (trailing `[..]` stripped, case-insensitive), existing Qwen ids pass through, everything else uses `*`; applies to `/v1/chat/completions` and `/v1/messages` only. Also editable at runtime in the dashboard (Settings → Model mapping); a dashboard-saved map overrides this variable, see `.env.example` | `*=qwen3.8-max-thinking` |
 | `MODELS_CACHE_TTL` | Model list cache TTL in seconds; after expiry the next request refreshes it from upstream; `0` = never expires | `3600` |
+| `AGENT_TURN_MAX_TOOL_CALLS` | Anthropic path: cap on text-channel `tool_use` blocks per agent turn (4–256). When the model runs away after a narrated `[TOOL CALL]` (repeats the same call hundreds of times, hallucinates a whole session), the upstream is cut right after the N-th admitted call and the admitted calls are delivered with `stop_reason=tool_use`; once a call was admitted in an earlier delta, a duplicate, a rejected call or prose/thinking also cuts the turn | `24` |
 | `AGENT_CONTEXT_FILE_THRESHOLD_BYTES` | Externalize complete Agent tool definitions and history as a Qwen text document when the request body exceeds this size, avoiding the roughly 128 KiB WAF limit | `92160` (90 KiB) |
 | `AGENT_CONTEXT_LIVE_PROMPT_BYTES` | Maximum size of the tool protocol and current turn kept in the live request after context externalization | `49152` (48 KiB) |
 | `QWEN_CHAT_PROXY_URL` | Custom Chat API reverse proxy address | `https://your-proxy.com` |

README-ru.md

@@ -122,6 +122,7 @@ OUTPUT_THINK=true             # Выводить ли процесс размы
 LEGACY_REASONING_IN_CONTENT=false # Формат рассуждений: false=поле reasoning_content, true=старый режим <think> внутри content (true/false)
 SIMPLE_MODEL_MAP=false        # Упрощенное сопоставление моделей (true/false)
 MODELS_CACHE_TTL=3600         # Срок жизни кэша списка моделей (сек), 0=бессрочно
+AGENT_TURN_MAX_TOOL_CALLS=24  # Anthropic: лимит tool_use по текстовому каналу за ход агента (4-256), дальше обрыв upstream
 
 # 🌐 Прокси и обратный прокси
 QWEN_CHAT_PROXY_URL=          # Пользовательский URL обратного прокси Chat API (по умолчанию: https://chat.qwen.ai)
@@ -150,7 +151,9 @@ CACHE_MODE=default            # Режим кэширования изображ
 | `OUTPUT_THINK` | Отображать ли процесс размышления AI | `true` или `false` |
 | `LEGACY_REASONING_IN_CONTENT` | Формат вывода рассуждений. По умолчанию `false` = рассуждения в отдельном поле `reasoning_content`; `true` = старый режим (`<think>` внутри `content`) | `true` или `false` |
 | `SIMPLE_MODEL_MAP` | Упрощенное сопоставление моделей, возвращает только базовые модели без вариантов | `true` или `false` |
+| `MODEL_MAP` | Сопоставление входящих имён моделей: `alias=qwen-id,...,*=fallback`. Точное совпадение в приоритете (хвостовой `[..]` отбрасывается, без учёта регистра), существующие id Qwen проходят без изменений, остальное идёт в `*`; действует только для `/v1/chat/completions` и `/v1/messages`. Редактируется и в панели (Настройки → Сопоставление моделей); сохранённое в панели сопоставление имеет приоритет над этой переменной, см. `.env.example` | `*=qwen3.8-max-thinking` |
 | `MODELS_CACHE_TTL` | Срок жизни кэша списка моделей (в секундах); по истечении следующий запрос обновит список; `0` = бессрочный кэш | `3600` |
+| `AGENT_TURN_MAX_TOOL_CALLS` | Путь Anthropic: лимит блоков `tool_use` по текстовому каналу за один ход агента (4–256). Если модель «идёт вразнос» после нарративного `[TOOL CALL]` (сотни повторов одного вызова, выдуманная сессия целиком), upstream обрывается сразу после N-го принятого вызова, а принятые вызовы отдаются со `stop_reason=tool_use`; после уже принятого в более раннем delta вызова дубликат, отклонённый вызов или проза/размышление также обрывают ход | `24` |
 | `QWEN_CHAT_PROXY_URL` | Пользовательский адрес обратного прокси Chat API | `https://your-proxy.com` |
 | `QWEN_CLI_PROXY_URL` | Пользовательский адрес обратного прокси CLI API | `https://your-cli-proxy.com` |
 | `PROXY_URL` | Адрес прокси для исходящих запросов, поддержка HTTP/HTTPS/SOCKS5 | `http://127.0.0.1:7890` |

README.md

@@ -103,6 +103,7 @@ LEGACY_REASONING_IN_CONTENT=false # 推理输出格式，false=reasoning_content
 SIMPLE_MODEL_MAP=false        # 简化模型映射 (true/false)
 MODELS_CACHE_TTL=3600         # 模型列表缓存有效期（秒），0=永不过期
 AGENT_TURN_MAX_ATTEMPTS=3     # 单个 Agent 回合生成有效工具调用/最终态的最大尝试数（2-6）
+AGENT_TURN_MAX_TOOL_CALLS=24  # Anthropic 路径单轮文本通道 tool_use 上限（4-256），到数即截断上游
 AGENT_TURN_ALLOW_PROSE_WITH_TOOLS=false  # 允许工具调用回合同时带可见正文（Anthropic 客户端）
 AGENT_TURN_ACCEPT_BARE_FINAL=false       # 允许没有 <agent_final> 包装的可见正文作为正常结束
 AGENT_CONTEXT_FILE_THRESHOLD_BYTES=92160 # 超过阈值时外置完整 Agent 上下文
@@ -135,10 +136,12 @@ CACHE_MODE=default            # 图片缓存模式 (default/file)
 | `OUTPUT_THINK` | 是否显示 AI 思考过程 | `true` 或 `false` |
 | `LEGACY_REASONING_IN_CONTENT` | 推理输出格式。默认 `false`=推理走独立的 `reasoning_content` 字段；`true`=旧版行为（`<think>` 并入 `content`） | `true` 或 `false` |
 | `SIMPLE_MODEL_MAP` | 简化模型映射，只返回基础模型不包含变体 | `true` 或 `false` |
+| `MODEL_MAP` | 入站模型名映射：`alias=qwen-id,...,*=fallback`。精确匹配优先（末尾 `[..]` 先去掉、不区分大小写），上游已有的 Qwen id 原样透传，其余走 `*`；只作用于 `/v1/chat/completions` 与 `/v1/messages`。也可在管理面板「系统设置 → 模型映射」里在线编辑，面板保存的映射优先于本变量；详见 `.env.example` | `*=qwen3.8-max-thinking` |
 | `MODELS_CACHE_TTL` | 模型列表缓存有效期（秒），过期后下次请求自动向上游刷新；`0` 表示永不过期 | `3600` |
 | `AGENT_TURN_ALLOW_PROSE_WITH_TOOLS` | 放宽回合门禁：允许同一回合既有有效工具调用又有可见正文。Anthropic Messages API 允许 `text` 与 `tool_use` 共存，Claude Code 等客户端因此会被严格模式反复判为 `invalid_tool_call` | `false` |
 | `AGENT_TURN_ACCEPT_BARE_FINAL` | 放宽回合门禁：把有可见正文但缺少 `<agent_final>` 包装的回合按 `finish_reason=stop` 接受，而不是判为 `bare` 并重试 | `false` |
 | `AGENT_TURN_MAX_ATTEMPTS` | 工具请求在一次 HTTP 回合内生成有效 `tool_calls`、明确完成态或阻塞态的最大尝试数；范围 2–6，耗尽后非流式请求返回 HTTP 429/503，SSE 请求返回显式错误帧，绝不伪装成正常 `stop` | `3` |
+| `AGENT_TURN_MAX_TOOL_CALLS` | Anthropic 路径：一轮 Agent 回合里文本通道 `tool_use` 的上限（4–256）。模型在叙述的 `[TOOL CALL]` 之后失控（同一调用重复上百次、幻想整段会话）时，第 N 个已放行的调用之后立刻终止上游，已放行的调用以 `stop_reason=tool_use` 交付；更早的 delta 里已放行过调用之后再出现重复、被拒绝的调用或正文/思考同样截断 | `24` |
 | `AGENT_CONTEXT_FILE_THRESHOLD_BYTES` | Agent 请求体超过此大小时，将完整工具定义和历史自动外置为 Qwen 文本文档，避免触发约 128 KiB 的 WAF 限制 | `92160`（90 KiB） |
 | `AGENT_CONTEXT_LIVE_PROMPT_BYTES` | 上下文外置后，实时请求中保留的工具协议、system/developer 指令、原始任务、最近工具进度和当前结果的最大大小 | `49152`（48 KiB） |
 | `QWEN_CHAT_PROXY_URL` | 自定义 Chat API 反代地址 | `https://your-proxy.com` |