@qwen-code/arena-agent 0.0.10-dev.8-test.0 → 0.0.10-test.0

package/.env.example

@@ -13,6 +13,9 @@ OPENAI_API_KEY=
 # ARENA_OPENAI_API=auto
 # ARENA_OPENAI_API=responses
 # ARENA_OPENAI_API=chat
+# Set false to omit reasoning encrypted_content and id from Responses input.
+# Other reasoning fields and the complete local history are preserved.
+# ARENA_RESPONSES_REPLAY_REASONING_STATE=true
 
 # Unified runtime mode. Canonical values are shown; full_stack/fullstack and
 # single_html/single-html remain accepted compatibility aliases.
@@ -20,6 +23,14 @@ ARENA_HARNESS_MODE=arena
 # ARENA_HARNESS_MODE=react
 # ARENA_HARNESS_MODE=html
 # ARENA_HARNESS_MODE=full-stack
+# All modes use incremental message events for headless stream-json.
+# Set full for consumers requiring a full messages snapshot on every event.
+# ARENA_STREAM_HISTORY=full
+# ARENA_CODE_STREAM_HISTORY remains a lower-priority compatibility alias.
+
+# Headless configuration: a redacted startup record and actual request settings
+# on stderr and in the session's startup-config.jsonl. Set false to disable both.
+# ARENA_CONFIG_LOGGING=true
 
 # Native provider fallbacks used when OPENAI_BASE_URL/API_KEY are unset.
 # ANTHROPIC_BASE_URL=https://api.anthropic.com/v1
@@ -46,7 +57,8 @@ ARENA_HARNESS_MODE=arena
 # HARBOR_REMOTE_COMMAND_API_KEY=
 
 # Trace-compatible model settings.
-# ARENA_MAX_TURNS=500
+# Agent Arena defaults to 250 model steps per turn; send a follow-up to continue.
+# ARENA_MAX_TURNS=250
 # Code Arena compatibility alias used only when ARENA_MAX_TURNS is unset:
 # CODE_HARNESS_MAX_STEPS=50
 # ARENA_MAX_TOKENS=131072
@@ -57,6 +69,23 @@ ARENA_HARNESS_MODE=arena
 # ARENA_TOOL_CHOICE=auto
 # JSON array of exact tool names hidden from the model and rejected if called.
 # ARENA_DISALLOWED_TOOLS=["bash","write_file"]
+# Agent Arena connectors through Merge MCP; opt-in, ignored by Code Arena.
+# Enable the captured catalog in the Tool Pack; connect apps for the Registered
+# User through Merge Link. The key and user must belong to the same environment.
+# One configured Registered User is shared by all sessions of this process.
+# Slugs: notion, google_calendar, google_drive, google_docs, google_sheets,
+# google_slides, gmail, dropbox, supabase, vercel, linear.
+# ARENA_CONNECTORS=["notion","google_drive","google_docs","google_sheets"]
+# Optional host-provided authenticated-app snapshot, separate from conversation
+# enablement. This does not authorize apps or query Merge. Explicit [] enables
+# discovery with the captured no-connected-apps prompt. Omit both arrays to keep
+# the connector-off baseline.
+# ARENA_CONNECTED_CONNECTORS=["notion","google_drive","google_docs","google_sheets","gmail"]
+# ARENA_MERGE_ACCESS_KEY=
+# ARENA_MERGE_TOOL_PACK_ID=
+# ARENA_MERGE_REGISTERED_USER_ID=
+# ARENA_MERGE_API_BASE=https://ah-api.merge.dev
+# ARENA_CONNECTOR_TIMEOUT_MS=60000
 # ARENA_THINKING_BUDGET=81920
 # ARENA_ENABLE_THINKING=true
 # ARENA_PRESERVE_THINKING=true
@@ -64,16 +93,22 @@ ARENA_HARNESS_MODE=arena
 # prompt fixtures and Code Arena prompts are unchanged.
 # ARENA_EXCLUDE_QWEN_FROM_SYSTEM_PROMPT=false
 # CLI equivalent: --exclude-qwen-from-system-prompt
-# Agent Arena hard-compacts at 250000 tokens for every model. It uses the most
-# recent provider-reported input usage plus a local estimate of newer messages,
-# falling back to a full local estimate when no reliable usage is available.
-# These compatibility settings remain available; Code Arena still uses
-# context window multiplied by the compaction ratio, which defaults to 0.7.
+# Agent Arena compaction: rolling (default) or summary (historical full summary).
+# Code Arena always uses summary and ignores this setting.
+# ARENA_COMPACTION_MODE=rolling
+# Agent Arena: threshold = min(250000, context window * ratio), default ratio 0.8.
+# Each setting can be configured independently; invalid values use its default.
+# Code Arena keeps independent controls (default window 1000000, ratio 0.7).
+# Compaction counts use completed-request usage plus newer messages.
 # ARENA_CONTEXT_WINDOW=1000000
-# ARENA_CONTEXT_COMPACTION_RATIO=0.7
+# ARENA_CONTEXT_COMPACTION_RATIO=0.8
+# Restore Agent Arena background memory by default; set false to disable.
+# CLI overrides: --restore-session-memory / --no-restore-session-memory.
+# ARENA_RESTORE_SESSION_MEMORY=true
 # Provisional Agent Arena background-memory budget for newly archivable content.
+# Uses request-usage deltas with local supplements; local counts omit tool-call IDs.
 # ARENA_MEMORY_UPDATE_TOKENS=24000
-# Agent Arena default: 131072; Code Arena default: 8192.
+# Default: rolling memory 131072; summary mode and Code Arena 8192.
 # ARENA_COMPACTION_SUMMARY_MAX_TOKENS=131072
 # ARENA_EXTRA_BODY={"enable_thinking":true}
 # Transient model failures use full-jitter exponential backoff. The retry delay
@@ -134,9 +169,16 @@ ARENA_HARNESS_MODE=arena
 
 # Search, page reading, image generation, and speech through ModelRouter/Routify.
 ROUTIFY_API_KEY=
+# Optional per-tool keys override ROUTIFY_API_KEY; unset or blank falls back.
+# ROUTIFY_WEB_SEARCH_API_KEY=
+# ROUTIFY_IMAGE_SEARCH_API_KEY=
+# ROUTIFY_GENERATE_IMAGE_API_KEY=
+# For an image proxy, keep the real MR key in ROUTIFY_API_KEY and set only
+# ROUTIFY_GENERATE_IMAGE_API_KEY to the proxy key, with the base URL below.
 # ROUTIFY_WEB_SEARCH_URL=https://routify-pub.alibaba-inc.com/protocol/compatible-mode/res/v1/llm/context
 # ROUTIFY_IMAGE_SEARCH_URL=https://routify-pub.alibaba-inc.com/protocol/brave/res/v1/images/search
 # ROUTIFY_VERTEX_BASE_URL=https://routify-pub.alibaba-inc.com/protocol/vertex/v1beta
+# ROUTIFY_OPENAI_BASE_URL=https://routify-pub.alibaba-inc.com/protocol/openai/v1
 # ARENA_WEB_SEARCH_MODEL=brave.llm-context
 # CODE_HARNESS_WEB_SEARCH_MODEL=brave.llm-context
 # Optional fetch_page/web_fetch path: Firecrawl first, then Jina Reader.
@@ -144,9 +186,18 @@ ROUTIFY_API_KEY=
 # Code Arena web_fetch content cap (UTF-8 bytes, including truncation notice).
 # Applies to local and Routify readers in all formats; must be a positive integer.
 # CODE_HARNESS_WEB_FETCH_MAX_BYTES=40000
-# ARENA_IMAGE_MODEL=gemini-3.1-flash-image
+# Agent Arena single images randomly use Gemini or gpt-image-2 by default.
+# Set this variable to pin a model (also overrides option one).
+# ARENA_IMAGE_MODEL=gpt-image-2
+# Second model for Agent Arena generate_image with offer_options=true.
+# Defaults to gpt-image-2 (or Gemini when that is the primary).
+# Without a user handler, save option one and continue; otherwise ask the user.
+# ARENA_IMAGE_ALTERNATIVE_MODEL=gpt-image-2
 # ARENA_IMAGE_SIZE=1024*1024
-# CODE_HARNESS_IMAGE_MODEL=gemini-3.1-flash-image
+# Code Arena randomly uses Sunburst or Flare.
+# Set this variable to pin a model; otherwise an explicit ARENA_IMAGE_MODEL
+# also pins Code Arena. Grok and Muse are not enabled.
+# CODE_HARNESS_IMAGE_MODEL=gpt-image-2.5-sunburst
 # CODE_HARNESS_IMAGE_SIZE=1024*1024
 # ARENA_SPEECH_MODEL=gemini-3.1-flash-tts-preview
 # ARENA_SPEECH_VOICE=Kore

package/CHANGELOG.md

@@ -12,23 +12,146 @@
 
 ## [Unreleased]
 
-## [0.0.10-dev.8-test.0] - 2026-09-28
+## [0.0.10-test.0] - 2026-11-09
 
 ### Agent Arena
 
-- 修正上下文 token 重复估算。
-- 记忆更新支持 `ARENA_REASONING_EFFORT`。
+- 压缩阈值对齐为上下文窗口的 80% 与 250,000 tokens 中的较小值，单独设置 `ARENA_CONTEXT_WINDOW` 即可适配小窗口模型。
+- 达到压缩阈值但尚无可用记忆时继续运行，不再强制生成或等待记忆。
+- 每轮默认最多 250 个模型步骤，`ARENA_MAX_TURNS` 可覆盖；达到上限保存进度，发送消息即可继续。
+- `web_search` 各深度的候选数量、正文和片段预算对齐官方文档。
+- 普通附件单文件上限调整为 25 MB，PDF 为 10 MB，每条消息合计仍为 50 MB。
+- 按官方请求补齐会话记忆更新前缀及用户消息合并，修复长工具循环中更新被拒绝，并保留恢复记忆和图片上下文。
+
+### 通用
+
+- Responses、Chat Completions 和 Anthropic Messages 的断流、空回复及临时错误按配置重试，不再将不完整回复误报为成功或执行未完成的工具调用。
+- 摘要压缩、图片转文字及指标侧查询不再接受模型明确标记为未完成的内容。
+- 兼容更多网关限流错误并避免重试明确的鉴权失败，保留 Chat Completions 的明确拒答，拦截 Responses 中前后冲突的工具调用。
+
+## [0.0.10-dev.16] - 2026-11-04
+
+### Agent Arena
+
+- Repo 会话记住目标仓库，新增同级仓库后仍可继续和恢复；目标缺失或 Git 状态异常时明确报错。
+- 对齐近期日志中的工作区、Repo 和记忆提示词，以及已有工具的声明顺序。
+- `read_file` 支持 `offset`、`limit` 分页；`bash` 超时返回已有输出并允许命令继续运行，主动取消仍终止进程树。
+- 对齐工具参数的默认值、实时与恢复历史格式，以及同批工具并发执行和结果顺序。
+- 扩展 MCP 应用至 Google Drive、Docs、Sheets、Slides、Gmail、Dropbox、Supabase、Vercel 和 Linear，并对齐 Notion 工具、工作区文件传输及重复调用保护。
+- 新增 `ARENA_CONNECTED_CONNECTORS`，区分已连接但未开启和未连接的应用。
+- `generate_image` 单图默认随机使用 GPT Image 2 或 Gemini 3.1 Flash，`offer_options` 并行生成双图供选择，保存未选图片并支持中断后重试；第二模型可用 `ARENA_IMAGE_ALTERNATIVE_MODEL` 配置。
+- Headless 未配置用户 handler 时，`generate_image` 的 `offer_options=true` 默认保存第一张候选图并继续执行。
+
+### Code Arena
+
+- `generate_image` 默认随机使用 `gpt-image-2.5-sunburst` 或 `gpt-image-2.5-flare`。
+
+### 通用
+
+- 新增 `ROUTIFY_WEB_SEARCH_API_KEY`、`ROUTIFY_IMAGE_SEARCH_API_KEY`、`ROUTIFY_GENERATE_IMAGE_API_KEY`，支持搜索与生图独立密钥，未设置或留空时继续使用 `ROUTIFY_API_KEY`。
+
+## [0.0.10-dev.15] - 2026-10-28
+
+### Agent Arena
+
+- 后台记忆更新按服务端 token 用量差分判断新增预算，减少长签名造成的提前更新。
+- 默认恢复同 session 的后台记忆，避免多轮调用重复更新；可用 `ARENA_RESTORE_SESSION_MEMORY=false` 或 `--no-restore-session-memory` 关闭。
+
+### 通用
+
+- 本地上下文估算忽略工具调用 ID，避免内嵌长签名抬高计数。
+- Headless 记录生效配置、脱敏输入及实际模型请求参数和工具列表，输出到 stderr 与会话的 `startup-config.jsonl`，可用 `ARENA_CONFIG_LOGGING=false` 关闭。
+
+## [0.0.10-dev.14] - 2026-10-26
+
+### 通用
+
+- 修复 Qwen4 及后续代际的图片输入识别，默认仅支持文字和图片。
+- 修复 Claude Fable/Mythos、Vertex 日期型号及 DeepSeek Flash 视觉型号的输入模态识别。
+
+## [0.0.10-dev.13] - 2026-10-23
+
+### Agent Arena
+
+- 修复远程 `ask_user` 在配置较长 `ARENA_ASK_USER_TIMEOUT_MS` 后仍于 300 秒提前超时的问题，并显示底层连接错误原因。
+
+### Code Arena
+
+- 对齐 Code Arena 官方 2026-10-19 更新日志：按官方契约声明工具的必填和可选参数，非 Responses 路径显式关闭严格模式；可选参数接受 `null` 并使用默认值。
+- 对齐 Code Arena 官方 2026-10-20 更新日志：直连 OpenAI Responses 使用严格工具 schema，可选字段填写 `null` 时使用默认值。
+- 根据最新回流修正 React 初始项目中文件列表的返回顺序。
+
+## [0.0.10-dev.12] - 2026-10-15
+
+### Agent Arena
+
+- 摘要压缩接入服务端 token 用量，并支持长单轮工具任务。
+
+### Code Arena
+
+- 摘要压缩接入服务端 token 用量，支持长单轮工具任务，并在续跑时恢复已应用的摘要。
+- 固定 React 和 Full Stack 模板的完整安装依赖，避免上游新版本发布导致初始化失败，并保留已有项目的依赖锁。
+- React 与 Full Stack 的命令执行支持毫秒级 `timeout`，默认 180 秒，并拒绝空命令；React 可执行 Codex 的 `shell_command` 调用。
+
+### 通用
+
+- 压缩计数改善 SVG 等大工具输出的估算，避免重复计入已压缩内容。
+- 模型因长度限制截断时保留部分输出并报告失败，避免将未完成的任务标为成功。
+- 修复 Responses 多工具调用的事件错配，确保工具结果、文件展示和用户输入请求对应正确调用。
+
+## [0.0.10-dev.11] - 2026-10-11
+
+### Agent Arena
+
+- 新增实验性 Notion、Google Calendar 接入，使用 `ARENA_CONNECTORS` 选择应用，并通过 `ARENA_MERGE_*` 配置已授权的 Merge 账号。
+- Headless Hook 新增 `before_compaction` 事件，可按压缩次数在摘要式或 rolling session memory 压缩前快照暂停，续跑时延续计数。
+
+### Code Arena
+
+- `generate_image` 默认模型改为 `gpt-image-2.5-sunburst`，继续通过 ModelRouter 调用。
+
+### 通用
+
+- 修复 Claude Haiku 5.5 开启 thinking 时因旧参数格式报 400 的问题。
+- 修复 Gemini 通过 Chat Completions 调用工具后因签名丢失而无法继续的问题。
+
+## [0.0.10-dev.10] - 2026-10-07
+
+### Agent Arena
+
+- 新增 `ARENA_COMPACTION_MODE=summary`，可切换回摘要式压缩，默认仍为 `rolling`。
+- 后台记忆写入 `history.json`，`--single-step` 默认恢复，普通续跑可用 `--restore-session-memory` 或 `ARENA_RESTORE_SESSION_MEMORY=true` 开启。
+- `ARENA_CONTEXT_WINDOW` 与 `ARENA_CONTEXT_COMPACTION_RATIO` 同时配置有效值时按乘积压缩，否则阈值固定为 250k。
+- 修复记忆中美元符号导致的提示词失真，补充更新失败与取消状态，并在任务结束和暂停时收束后台请求。
+
+### Code Arena
+
+- Full Stack 与 HTML 提示词对齐已核验的历史请求。
+
+### 通用
+
+- 新增 `ARENA_RESPONSES_REPLAY_REASONING_STATE=false`，可关闭 Responses 历史加密推理状态的回传。
+- 四种模式统一降低 headless `stream-json` 长轨迹的日志量与接收端内存占用；依赖每条事件完整历史的消费者可用 `ARENA_STREAM_HISTORY=full` 恢复旧输出，兼容旧变量 `ARENA_CODE_STREAM_HISTORY`。
+
+## [0.0.10-dev.9] - 2026-09-30
+
+### Agent Arena
+
+- 修正上下文 token 重复估算，记忆更新支持 `ARENA_REASONING_EFFORT`。
+- 修复工具参数 JSON 非法时后续请求报 400、无法继续纠错的问题。
+- 对齐空参数、搜索深度、URL 和重复选项 ID 的校验错误，修复旧会话续跑和 Responses 流中断后的恢复。
 
 ### Code Arena
 
-- React 默认步数上限从 50 提高到 75。
-- 视觉模型可通过 `read_file` 查看 10 MiB 以内的 PNG/JPEG 图片，工具提示同步修正。
+- React 默认步数上限提高到 75，并改用 `bash` 安装依赖和构建。
+- 视觉模型可通过 `read_file` 查看 10 MiB 以内的 PNG/JPEG 图片，并对齐文件读取描述与 MIME。
+- 对齐工具参数校验和隐藏计划工具，修复 Full Stack 模板与依赖锁保留。
 
 ### 通用
 
 - Headless Hook 支持重复输出检测，可选择继续、快照后继续或快照退出。
 - 新增 `--single-step`，执行一轮后保存退出，可用 `--continue` 续跑。
-- 修复续跑时历史模型覆盖当前模型配置的问题。
+- 修复续跑时历史模型覆盖当前模型配置、历史中途修正未及时保存的问题。
 
 ## [0.0.10-dev.8] - 2026-09-20
 
@@ -194,14 +317,22 @@
 - `bash` 用于一次性命令，区分超时与信号终止，并清理遗留后台进程；需要持续运行
   的任务改用进程工具。预览端口支持 E2B Host 探测及按需返回告警。
 
-[Unreleased]: https://gitlab.alibaba-inc.com/know-how/arena-harness/-/compare/v0.0.10-dev.8-test.0...master
-[0.0.10-dev.8-test.0]: https://gitlab.alibaba-inc.com/know-how/arena-harness/-/compare/v0.0.10-dev.8...v0.0.10-dev.8-test.0
-[0.0.10-dev.8]: https://gitlab.alibaba-inc.com/know-how/arena-harness/-/compare/v0.0.10-dev.7...v0.0.10-dev.8
-[0.0.10-dev.7]: https://gitlab.alibaba-inc.com/know-how/arena-harness/-/compare/v0.0.10-dev.6...v0.0.10-dev.7
-[0.0.10-dev.6]: https://gitlab.alibaba-inc.com/know-how/arena-harness/-/compare/v0.0.10-dev.5...v0.0.10-dev.6
-[0.0.10-dev.5]: https://gitlab.alibaba-inc.com/know-how/arena-harness/-/compare/v0.0.10-dev.4...v0.0.10-dev.5
-[0.0.10-dev.4]: https://gitlab.alibaba-inc.com/know-how/arena-harness/-/compare/v0.0.10-dev.3...v0.0.10-dev.4
-[0.0.10-dev.3]: https://gitlab.alibaba-inc.com/know-how/arena-harness/-/compare/v0.0.10-dev.2...v0.0.10-dev.3
-[0.0.10-dev.2]: https://gitlab.alibaba-inc.com/know-how/arena-harness/-/compare/v0.0.10-dev.1...v0.0.10-dev.2
-[0.0.10-dev.1]: https://gitlab.alibaba-inc.com/know-how/arena-harness/-/compare/v0.0.10-dev.0...v0.0.10-dev.1
-[0.0.10-dev.0]: https://gitlab.alibaba-inc.com/know-how/arena-harness/-/compare/v0.0.9-dev.34...v0.0.10-dev.0
+[Unreleased]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-test.0&to=master&mergeBase=true
+[0.0.10-test.0]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.16&to=v0.0.10-test.0&mergeBase=true
+[0.0.10-dev.16]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.15&to=v0.0.10-dev.16&mergeBase=true
+[0.0.10-dev.15]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.14&to=v0.0.10-dev.15&mergeBase=true
+[0.0.10-dev.14]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.13&to=v0.0.10-dev.14&mergeBase=true
+[0.0.10-dev.13]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.12&to=v0.0.10-dev.13&mergeBase=true
+[0.0.10-dev.12]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.11&to=v0.0.10-dev.12&mergeBase=true
+[0.0.10-dev.11]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.10&to=v0.0.10-dev.11&mergeBase=true
+[0.0.10-dev.10]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.9&to=v0.0.10-dev.10&mergeBase=true
+[0.0.10-dev.9]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.8&to=v0.0.10-dev.9&mergeBase=true
+[0.0.10-dev.8]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.7&to=v0.0.10-dev.8&mergeBase=true
+[0.0.10-dev.7]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.6&to=v0.0.10-dev.7&mergeBase=true
+[0.0.10-dev.6]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.5&to=v0.0.10-dev.6&mergeBase=true
+[0.0.10-dev.5]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.4&to=v0.0.10-dev.5&mergeBase=true
+[0.0.10-dev.4]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.3&to=v0.0.10-dev.4&mergeBase=true
+[0.0.10-dev.3]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.2&to=v0.0.10-dev.3&mergeBase=true
+[0.0.10-dev.2]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.1&to=v0.0.10-dev.2&mergeBase=true
+[0.0.10-dev.1]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.10-dev.0&to=v0.0.10-dev.1&mergeBase=true
+[0.0.10-dev.0]: https://code.alibaba-inc.com/know-how/arena-harness/compare?from=v0.0.9-dev.34&to=v0.0.10-dev.0&mergeBase=true