ck-pi-zen-session0.1.26

OpenCode Zen session auto-refresh and free-tier model catalog synchronization for the Pi coding agent.

cknb666 Published 4 days ago MIT ~3.2K downloads/mo

Untested on omp

This is a pi package. No one has reported whether it works on omp yet. If you try it, add your result below.

Reports 0

No reports yet.

Readme

ck-pi-zen-session

中文说明 | English Documentation

OpenCode Zen 免费模型同步、双协议路由 (Completions & Responses)、请求头伪装与会话自愈插件，专为 Pi 编码智能体打造。

赞助与支持

如果您觉得本项目提升了您的开发效率，欢迎赞助支持！您的赞助将直接用于持续维护、网络节点验证、逆向协议对齐与新特性的快速迭代。

赞助渠道

GitHub Sponsors: https://github.com/sponsors/1837620622
企业赞助与商务合作:
- 欢迎 AI 基础设施提供商、代理网关服务商或团队提供赞助。
- 赞助权益包含：在项目 README 显著位置展示企业 Logo 与链接、优先处理定制化需求与专属支持。

一、核心特性

/zen 指令一键激活与动态免费模型全量探测：输入一次 API Key (/zen oc_sk_...)，全自动在线探测所有带有 free / zero / pickle 标识的模型，全量标记 0 额度消耗 (cost: { input: 0, output: 0 }) 并动态热挂载进 Pi 与 CC-Switch。
双供应商双协议分流架构：将 OpenCode Zen 拆分为 opencode-zen-free (openai-completions) 与 opencode-zen-free-responses (openai-responses)，彻底解决 Muse Spark 系列模型报错 400 ModelProtocolUnsupported 的协议冲突问题。
官方算法深度逆向：完整对齐 OpenCode 客户端 Identifier.descending("ses") 与 Identifier.ascending("msg") 算法，生成 30 位降序会话 ID 与升序请求 ID，支持毫秒级时间戳反解。
全套官方级伪装请求头：自动注入全部 7 个官方客户端对齐请求头：
- User-Agent: opencode/1.18.32 ai-sdk/provider-utils/4.0.23 runtime/bun/1.3.14
- x-opencode-client: cli
- x-opencode-session: ses_<12位反转时间戳十六进制><14位Base62>
- x-opencode-request: msg_<12位升序时间戳十六进制><14位Base62>（每次 HTTP 请求唯一）
- x-opencode-project: prj_<16位工作区哈希>（与官方项目哈希行为对齐）
- x-session-affinity: 与 session 严格同步
- X-Session-Id: 与 session 严格同步
25 分钟平滑轮换与 401/403 自动换新自愈：官方后端免费会话限制约 1 小时（超时直接 403 FreeTierError）。本插件在底层拦截器和生命周期钩子中设置 25 分钟主动换新阈值，且在收到 401/403 异常时当场自动轮换并重试，彻底杜绝死锁。
全局 Fetch 底层拦截与会话压缩 403 根治： Pi 执行上下文压缩时常发起纯文本无工具请求，导致网关拦截。底层拦截器自动注入官方 6 大核心工具链（bash, edit, glob, grep, read, write）与 tool_choice: "none"，并对 ChatCompletions 执行 SSE-to-JSON 汇聚反序列化，彻底解决压缩卡死。
同模型 3 次透明重试自愈：遇网络抖动、网关超时（500/502/503/504/520-525/533）或上游伪 200 空响应（Provider returned an empty response）时，底层无损流预读并在内存中自动进行 3 次指数退避重试，严禁擅自跨模型降级。
严格供应商隔离：所有生命周期钩子与拦截逻辑严格限定于 opencode-zen-free 和 opencode-zen-free-responses，对 relayhub、deepseek、anthropic、openai 及其他插件完全静默放行，零副作用。

二、双协议架构与 Issue #1 修复

1. 问题背景 (Issue #1)

OpenCode Zen 上游网关对不同的模型系列启用了不同的传输协议：

标准 Chat 模型（Xiaomi MiMo, NVIDIA Nemotron, Big Pickle, Ling, Space Bunny 等）：运行在 /zen/v1/chat/completions（OpenAI Chat Completions 协议）；
Meta Muse Spark 系列（muse-spark-1.3-contributor-free, muse-spark-1.2-contributor-free）：仅部署在 /zen/v1/responses（OpenAI Responses API 协议）。

如果将 Muse Spark 模型作为 openai-completions 提交到 /zen/v1/chat/completions，OpenCode 网关会立即拒绝并返回 HTTP 400 错误：

{"type": "ModelProtocolUnsupported", "message": "Model does not support this protocol."}

2. 双供应商架构设计

为从根本上解决该协议限制，ck-pi-zen-session 0.1.22 引入了双供应商路由体系：

opencode-zen-free (api: openai-completions)：
- 托管模型：fledge-alpha-free, ling-3.1-flash-free, mimo-v2.6-flash-free, nemotron-3.5-lightning-free, big-pickle, space-bunny-free, longcat-2.5-preview-free。
- 请求端点：/zen/v1/chat/completions。
- 工具格式：Nested function 结构 ({ type: "function", function: { ... } })。
- 携带参数：stream_options: { include_usage: true }, reasoning_effort。
opencode-zen-free-responses (api: openai-responses)：
- 托管模型：muse-spark-1.3-contributor-free, muse-spark-1.2-contributor-free。
- 请求端点：/zen/v1/responses。
- 工具格式：Flat 扁平结构 ({ type: "function", name, description, parameters })。
- 请求参数规约：使用 input（而非 messages）、max_output_tokens，在无工具/压缩模式下注入 tool_choice: "none"，严禁注入 Chat 专有字段（stream_options, reasoning_effort, max_tokens），避免网关抛出 400 校验异常。
双端点透明持久化与路由：
- 同步时自动在 ~/.pi/agent/models.json 和 auth.json 中配置两个独立 Provider；
- /zen refresh 与 /zen ping 自适应分流探测；
- /zen model muse-spark-1.3-contributor-free 自动切换底层 Provider 为 opencode-zen-free-responses。

三、指令用法与操作面板

在 Pi 终端交互界面中，可通过 /zen 体系指令进行全方位控制：

指令	说明	交互效果
`/zen`	无参直接运行：全自动检测并续期 Session ID，刷新请求头，重新对齐免费模型，双写同步 `models.json` / CC-Switch 并热载入 Pi	输出状态面板与已同步模型摘要
`/zen <key>` 或 `/zen key <key>`	更新 API Key：设置/更换 OpenCode Zen API Key，自动在线验证并全量同步模型与全套请求头	输出新 Key 验证结果与模型卡片
`/zen refresh` 或 `/zen sync`	强制换新 Session：强制生成全新的合法降序 Session ID 并更新所有请求头与落盘配置	输出全新 Session ID 与就绪状态
`/zen list` 或 `/zen models` 或 `/zen free`	查看模型列表：展示全部可用免费模型详细规格卡片（上下文、最大输出、思考等级、协议归属）	输出分级卡片列表
`/zen add <模型ID> [名称]`	登记自定义模型：支持将新上线的免费模型登记至本地扩展库并实时热载入	输出新增模型配置状态
`/zen status`	查看运行状态：查看当前 API Key 掩码、活跃会话精确存活时间、模型库与请求头保护状态	输出运行状态仪表盘
`/zen ping [模型ID]`	网络时延探针：实时探测 Zen 免费模型网络连通性与往返延迟 (RTT)	输出各模型网络时延与评级
`/zen model <模型ID>`	快速切换模型：一键切换当前 Pi 会话所使用的 Zen 模型（自动识别 Provider 归属）	即刻切换生效

面板输出示例：

[OpenCode Zen 已自动刷新并就绪]
- API Key: oc_sk_cb...yRSa
- 全新 Session: ses_f23f81105fferIvyODUuNI61BM (有效且已持久化)
- 自动对接免费模型: 已同步 10 个 0 额度消耗模型 (包含 Completions 与 Responses 双协议)
- 请求头保护: 25分钟自动轮换 + 7维官方签名 + 6大核心工具全注入

四、内置免费模型库参数

插件自动同步官方免费模型并配置精准上限与思维链等级：

模型 ID	协议	供应商	上下文窗口	最大输出	推理思考等级	支持模态
`step-5-preview-free`	Completions	`opencode-zen-free`	200,000	32,768	全档位支持 (OpenAI low/med/high)	文本、图像 (多模态)
`exo-free`	Completions	`opencode-zen-free`	128,000	16,384	不支持 (快反系统/无需思考)	文本
`nemotron-3-ultra-free`	Completions	`opencode-zen-free`	1,000,000 (1M)	32,768	全档位支持 (OpenAI low/med/high)	文本
`ling-3.0-flash-fin-free`	Completions	`opencode-zen-free`	262,144	32,768	全档位支持 (OpenAI low/med/high)	文本
`ling-3.1-flash-free`	Completions	`opencode-zen-free`	262,144	32,768	全档位支持 (OpenAI low/med/high)	文本
`mimo-v2.6-flash-free`	Completions	`opencode-zen-free`	200,000	32,000	全档位支持 (OpenAI low/med/high)	文本、图像 (多模态)
`nemotron-3.5-lightning-free`	Completions	`opencode-zen-free`	262,144	32,768	全档位支持 (OpenAI low/med/high)	文本
`big-pickle`	Completions	`opencode-zen-free`	200,000	32,000	全档位支持 (OpenAI low/med/high)	文本
`space-bunny-free`	Completions	`opencode-zen-free`	200,000	32,000	全档位支持 (OpenAI low/med/high)	文本、图像 (多模态)
`longcat-2.5-preview-free`	Completions	`opencode-zen-free`	200,000	32,000	全档位支持 (OpenAI low/med/high)	文本
`jev-1.13-free`	Completions	`opencode-zen-free`	128,000	16,384	不支持 (快反系统/无需思考)	文本
`muse-spark-1.3-contributor-free`	Responses	`opencode-zen-free-responses`	1,048,576 (1M)	32,768	全档位支持 (OpenAI low/med/high)	文本、图像 (多模态)
`muse-spark-1.2-contributor-free`	Responses	`opencode-zen-free-responses`	1,048,576 (1M)	32,768	全档位支持 (OpenAI low/med/high)	文本、图像 (多模态)
`fledge-alpha-free`	Completions	`opencode-zen-free`	200,000	32,000	上游已下架 (已自动平滑重定向至 step-5-preview-free)	文本、图像 (多模态)

若 OpenCode 后端未来发布新免费模型，插件将自动在线探测发现并使用安全的自适应参数接入。

五、官方算法逆向解析

OpenCode Zen 客户端针对防刷和网关鉴权设计了基于时间戳编码的降序（descending）与升序（ascending）标识符体系：

// 1. Session ID (前缀 ses_, 降序排列)
export function generateZenSessionId(timestampMs = Date.now()): string {
  const combined = BigInt(timestampMs) * 0x1000n + BigInt(sequenceCounter);
  const inverted = ~combined; // 关键位反转：时间越新，字典序十六进制越小
  let hex = "";
  for (let i = 0; i < 6; i++) {
    const byte = Number((inverted >> BigInt(40 - 8 * i)) & 0xffn);
    hex += byte.toString(16).padStart(2, "0");
  }
  return `ses_${hex}${randomBase62(14)}`; // 30 字符
}

// 2. Request ID (前缀 msg_, 升序排列)
export function generateZenRequestId(timestampMs = Date.now()): string {
  const combined = BigInt(timestampMs) * 0x1000n + BigInt(sequenceCounter);
  let hex = "";
  for (let i = 0; i < 6; i++) {
    const byte = Number((combined >> BigInt(40 - 8 * i)) & 0xffn);
    hex += byte.toString(16).padStart(2, "0");
  }
  return `msg_${hex}${randomBase62(14)}`; // 30 字符
}

六、安装方法

方式 1：npm 安装（推荐）

pi install npm:ck-pi-zen-session

方式 2：GitHub 源码安装

pi install git:github.com/1837620622/ck-pi-extension

七、更新日志

0.1.24：
- 全新对接与上线 OpenCode Zen 最新免费模型矩阵：
  - 全面收录并高精适配 13 款可用官方免费模型：包含最新旗舰 step-5-preview-free（阶跃星辰 Step-5 免费模型，200k 上下文，32k 输出，多模态视觉与全档位思考）、exo-free（128k 极速快反）、nemotron-3-ultra-free（1M 超长上下文）、ling-3.0-flash-fin-free（262k 金融与代码加速）、jev-1.13-free（128k 快反）；
  - 针对上游下架的 fledge-alpha-free，路由器自动平滑无感重定向至最新旗舰免费模型 step-5-preview-free，彻底避免 404 或内容过滤报错。
- 彻底攻克会话压缩截断与崩塌故障 (Summarization failed: generation hit the token cap / Provider finish_reason: content_filter)：
  - 针对 Pi 会话压缩默认传入的 1600 tokens 极小 token 预算，请求拦截层强制将其提升至 16384 tokens，并在压缩总结请求中强制压制深思考（reasoning_effort: "low"），从根本上避免深思考模型耗尽 token cap 触发截断；
  - 流式响应重构器（reconstructedStream）全面重构为行缓冲分包安全 TransformStream，逐帧捕获并自动规范化 finish_reason: "length" 与 finish_reason: "content_filter" 为 "stop"；
  - 在遇到安全策略阻断且尚未生成有效内容时，自动合成规范的结构化会话进度检查点，确保会话压缩 100% 成功完成并平滑进入后续对话。
0.1.23：
- 核心修复会话压缩失败与 Token Cap 崩溃问题 (generation hit the token cap and the summary is incomplete)：
  - 调优压缩输出上限至 16,384 tokens，并在压缩总结请求中强制设置 reasoning_effort: "low" 并剔除冗余思考，避免深度推理模型（如 fledge-alpha-free）耗尽 token 额度；
  - 针对上游返回的 finish_reason: "length"，在非空摘要场景下平滑改写为 finish_reason: "stop"，彻底根除 Pi 官方压缩模块在命中 token cap 时的崩溃中断。
- 全链路防御安全审核过滤阻断 (finish_reason: content_filter)：
  - 引入 sanitizeSensitiveAuditContent 审计脱敏引擎：针对 <conversation> 中的 SQL 注入特征、XSS 脚本块、漏洞利用载荷、密码散列和卡号等敏感特征进行合规占位脱敏，从源头杜绝触发上游内容审查拦截；
  - 在拦截器和流反序列化中捕获 finish_reason: "content_filter"：在压缩模式下自动合成规范的结构化会话检查点摘要并置 finish_reason: "stop"，在对话模式下保全已生成内容并附带友好安全调整提示。
- 动态优化免费模型检测算法与最新在线模型矩阵：
  - 剔除已下架不可用的历史模型（mimo-v2.5-free, ling-3.0-flash-fin-free, nemotron-3-ultra-free）；
  - 正式同步并支持 9 款在线可用免费模型（包含全新 fledge-alpha-free, ling-3.1-flash-free, longcat-2.5-preview-free 等）；
  - 优化短别名映射引擎（mimo, ultra, ling, fledge 等自动平滑重定向至最新存活模型）。
0.1.22：
- 支持 OpenAI Responses API 协议与双供应商路由架构 (彻底修复 Issue #1)：
  - 双供应商精准分流：将 OpenCode Zen 拆分为 opencode-zen-free（走 openai-completions，负责 Big Pickle, Xiaomi MiMo, NVIDIA Nemotron, Ling, Space Bunny）与 opencode-zen-free-responses（走 openai-responses，负责 Meta Muse Spark 1.3/1.2 Contributor Free）；
  - 彻底解决 400 ModelProtocolUnsupported 报错：由于 OpenCode Zen 网关仅在 /responses 端点提供 Muse Spark 系列模型，通过独立 provider 注册让 Pi 原生通过 Responses API 发送请求；
  - Responses 协议结构对齐：为 Responses API 注入扁平化工具定义（OPENCODE_OFFICIAL_RESPONSES_TOOLS），在 Compaction 场景下注入 tool_choice: "none"，严禁注入 ChatCompletions 专有字段（如 stream_options, reasoning_effort, max_tokens），防止网关参数校验失败；
  - Fetch 拦截器双端点适配：拦截器同时支持 /zen/v1/chat/completions 与 /zen/v1/responses，区分 SSE 解析逻辑，且仅对 ChatCompletions 端点执行 SSE-to-ChatCompletion 反序列化；
  - 持久化与运行态双写：syncZenConfiguration 自动双写双供应商配置至 ~/.pi/agent/models.json 与 auth.json，/zen model 与 /zen ping 指令自适应路由。
0.1.21：
- 深度修复上游空响应重试机制 (Provider returned an empty response)：
  - SSE 流式预读探测上限提升至 30 个数据包与 64KB，精准穿透网关 keepalive 帧；
  - 严格限定错误结构匹配，杜绝模型正常回复误判；
  - 真实指数退避带随机抖动，坚守 3 次同模型重试；
  - 运行态模型规范化校验 (sanitizePiModelDefinition)，杜绝畸变数据写入。
0.1.20：
- 供应商节点空响应自动同模型重试与隐身零额度模型识别：首创无损首包 Stream Peek 预读探测，当上游节点返回伪 200 空流时自动 3 次重试；自动识别 stealth/space-bunny-alpha 等隐身零消耗模型。

八、许可

仓库: https://github.com/1837620622/ck-pi-extension
协议: MIT License

ck-pi-zen-session (English)

中文说明 | English Documentation

OpenCode Zen free-tier model sync, dual-protocol routing (Completions & Responses), request header disguise, and automatic session self-healing extension designed for Pi coding agent.

Sponsorship & Support

If you find this project helpful, please consider sponsoring. Your support directly funds continuous maintenance, endpoint verification, reverse-engineering protocol alignment, and rapid iteration.

Sponsorship Channels

GitHub Sponsors: https://github.com/sponsors/1837620622
Enterprise Sponsorship:
- AI infrastructure providers, proxy gateway services, and engineering teams are welcome to sponsor.
- Benefits: Prominent brand logo and link display in the README, priority issue response, and custom protocol optimizations.

1. Key Features
2. Dual-Protocol Architecture & Issue #1 Fix
3. Command Usage & Panel
4. Built-in Free Models
5. Reverse-Engineered Session Algorithm
6. Installation
7. Changelog
8. License

1. Key Features

One-click Activation & Dynamic Discovery: Discover all models marked with free, zero, or pickle with one key command (/zen oc_sk_...), tagging zero cost and mounting into Pi runtime and CC-Switch dynamically.
Dual-Provider Protocol Routing: Routes Chat models via opencode-zen-free (openai-completions) and Muse Spark models via opencode-zen-free-responses (openai-responses), eliminating 400 ModelProtocolUnsupported errors.
Reverse-Engineered Session Algorithm: Generates descending 30-char session IDs and ascending request IDs aligned with official OpenCode client algorithms.
Official Disguise Headers: Automatically injects all 7 official client-aligned headers:
- User-Agent: opencode/1.18.32 ai-sdk/provider-utils/4.0.23 runtime/bun/1.3.14
- x-opencode-client: cli
- x-opencode-session: ses_<12-hex-descending-ts><14-base62>
- x-opencode-request: msg_<12-hex-ascending-ts><14-base62>
- x-opencode-project: prj_<16-hex-hash>
- x-session-affinity: synchronized with session ID
- X-Session-Id: synchronized with session ID
Preemptive Renewal & Auto-Healing: Active 25-minute preemptive rotation with immediate auto-renewal on 401/403 gateway errors, preventing expiration deadlocks.
Global Fetch Interceptor & Compaction Defense: Injects required tool signatures during context compaction and handles SSE-to-JSON aggregation seamlessly without gateway rejection.
3-Attempt Same-Model Retry: Catches gateway timeouts (500/502/503/504/520-525/533) and empty responses (Provider returned an empty response) with 3-attempt exponential backoff same-model retry.
Strict Namespace Isolation: Operates strictly inside OpenCode Zen namespaces, preserving zero side effects for third-party providers.

2. Dual-Protocol Architecture & Issue #1 Fix

Background (Issue #1)

OpenCode Zen gateway serves models across different protocols:

Standard Chat models (Xiaomi MiMo, NVIDIA Nemotron, Big Pickle, Ling, Space Bunny): Served via /zen/v1/chat/completions (OpenAI Chat Completions protocol);
Meta Muse Spark models (muse-spark-1.3-contributor-free, muse-spark-1.2-contributor-free): Served exclusively via /zen/v1/responses (OpenAI Responses API protocol).

Posting Muse Spark models to /zen/v1/chat/completions results in HTTP 400:

{"type": "ModelProtocolUnsupported", "message": "Model does not support this protocol."}

Architecture Solution

In ck-pi-zen-session 0.1.22, a dual-provider architecture was introduced:

opencode-zen-free (api: openai-completions):
- Models: fledge-alpha-free, ling-3.1-flash-free, mimo-v2.6-flash-free, nemotron-3.5-lightning-free, big-pickle, space-bunny-free, longcat-2.5-preview-free.
- Endpoint: /zen/v1/chat/completions.
- Tools format: Nested function structure ({ type: "function", function: { ... } }).
- Parameters: stream_options: { include_usage: true }, reasoning_effort.
opencode-zen-free-responses (api: openai-responses):
- Models: muse-spark-1.3-contributor-free, muse-spark-1.2-contributor-free.
- Endpoint: /zen/v1/responses.
- Tools format: Flat structure ({ type: "function", name, description, parameters }).
- Parameters: Uses input and max_output_tokens; injects tool_choice: "none" on empty-tool requests; strictly excludes Chat-only fields (stream_options, reasoning_effort, max_tokens).
Transparent Persistence & Routing:
- Dual-writes both providers into ~/.pi/agent/models.json and auth.json.
- /zen refresh and /zen ping route probes dynamically according to protocol.
- /zen model muse-spark-1.3-contributor-free switches active provider to opencode-zen-free-responses.

3. Command Usage & Panel

Control all features in the Pi interactive interface using /zen commands:

Command	Description	Output
`/zen`	No args: check/renew session ID, refresh headers, sync `models.json` / CC-Switch, and hot reload	Status panel and synced models summary
`/zen <key>` or `/zen key <key>`	Update OpenCode Zen API key, verify online, and sync all models	Verification result and model cards
`/zen refresh` or `/zen sync`	Force generate new descending session ID and refresh headers	New session ID and ready state
`/zen list` or `/zen models` or `/zen free`	View catalog of available free models (context, max tokens, thinking levels, protocol)	Formatted model catalog cards
`/zen add <model-id> [name]`	Register custom free model into local catalog and hot reload	Registration status
`/zen status`	View API key mask, session age, and header protection status	Operational status dashboard
`/zen ping [model-id]`	Probe Zen free model network connectivity and round-trip time (RTT)	Latency and grade report
`/zen model <model-id>`	Quickly switch active Zen model (auto-detects provider namespace)	Immediate activation

Example Panel Output:

[OpenCode Zen 已自动刷新并就绪]
- API Key: oc_sk_cb...yRSa
- 全新 Session: ses_f23f81105fferIvyODUuNI61BM (有效且已持久化)
- 自动对接免费模型: 已同步 10 个 0 额度消耗模型 (包含 Completions 与 Responses 双协议)
- 请求头保护: 25分钟自动轮换 + 7维官方签名 + 6大核心工具全注入

4. Built-in Free Models

The extension syncs official free models with verified context windows and thinking levels:

Model ID	Protocol	Provider	Context Window	Max Output	Thinking Levels	Modality
`step-5-preview-free`	Completions	`opencode-zen-free`	200,000	32,768	Full support (OpenAI low/med/high)	Text, Vision
`exo-free`	Completions	`opencode-zen-free`	128,000	16,384	Unsupported (fast reflex / no thinking)	Text
`nemotron-3-ultra-free`	Completions	`opencode-zen-free`	1,000,000 (1M)	32,768	Full support (OpenAI low/med/high)	Text
`ling-3.0-flash-fin-free`	Completions	`opencode-zen-free`	262,144	32,768	Full support (OpenAI low/med/high)	Text
`ling-3.1-flash-free`	Completions	`opencode-zen-free`	262,144	32,768	Full support (OpenAI low/med/high)	Text
`mimo-v2.6-flash-free`	Completions	`opencode-zen-free`	200,000	32,000	Full support (OpenAI low/med/high)	Text, Vision
`nemotron-3.5-lightning-free`	Completions	`opencode-zen-free`	262,144	32,768	Full support (OpenAI low/med/high)	Text
`big-pickle`	Completions	`opencode-zen-free`	200,000	32,000	Full support (OpenAI low/med/high)	Text
`space-bunny-free`	Completions	`opencode-zen-free`	200,000	32,000	Full support (OpenAI low/med/high)	Text, Vision
`longcat-2.5-preview-free`	Completions	`opencode-zen-free`	200,000	32,000	Full support (OpenAI low/med/high)	Text
`jev-1.13-free`	Completions	`opencode-zen-free`	128,000	16,384	Unsupported (fast reflex / no thinking)	Text
`muse-spark-1.3-contributor-free`	Responses	`opencode-zen-free-responses`	1,048,576 (1M)	32,768	Full support (OpenAI low/med/high)	Text, Vision
`muse-spark-1.2-contributor-free`	Responses	`opencode-zen-free-responses`	1,048,576 (1M)	32,768	Full support (OpenAI low/med/high)	Text, Vision
`fledge-alpha-free`	Completions	`opencode-zen-free`	200,000	32,000	Delisted upstream (aliased to step-5-preview-free)	Text, Vision

5. Reverse-Engineered Session Algorithm

OpenCode Zen client implements timestamp-encoded descending and ascending identifiers:

// 1. Session ID (prefix ses_, descending order)
export function generateZenSessionId(timestampMs = Date.now()): string {
  const combined = BigInt(timestampMs) * 0x1000n + BigInt(sequenceCounter);
  const inverted = ~combined; // Key bit inversion: newer timestamps produce smaller hex values
  let hex = "";
  for (let i = 0; i < 6; i++) {
    const byte = Number((inverted >> BigInt(40 - 8 * i)) & 0xffn);
    hex += byte.toString(16).padStart(2, "0");
  }
  return `ses_${hex}${randomBase62(14)}`; // 30 chars
}

// 2. Request ID (prefix msg_, ascending order)
export function generateZenRequestId(timestampMs = Date.now()): string {
  const combined = BigInt(timestampMs) * 0x1000n + BigInt(sequenceCounter);
  let hex = "";
  for (let i = 0; i < 6; i++) {
    const byte = Number((combined >> BigInt(40 - 8 * i)) & 0xffn);
    hex += byte.toString(16).padStart(2, "0");
  }
  return `msg_${hex}${randomBase62(14)}`; // 30 chars
}

6. Installation

Method 1: npm (Recommended)

pi install npm:ck-pi-zen-session

Method 2: From GitHub

pi install git:github.com/1837620622/ck-pi-extension

7. Changelog

0.1.24:
- Full integration & synchronization with OpenCode Zen latest free models:
  - Registered 13 active official free models: step-5-preview-free (StepFun Step-5, 200k context, 32k output, multimodal vision & full thinking levels), exo-free (128k fast reflex), nemotron-3-ultra-free (1M long context), ling-3.0-flash-fin-free (262k finance/code accelerated), jev-1.13-free (128k fast reflex);
  - Delisted model graceful migration: fledge-alpha-free smoothly redirected to step-5-preview-free without 404s or moderation crashes.
- Auto-compaction token cap & truncation crash permanent fix (Summarization failed: generation hit the token cap / Provider finish_reason: content_filter):
  - Compaction max tokens boost: forced minimum 16,384 tokens with reasoning_effort: "low" during summarization tasks, eliminating token cap exhaustion;
  - Robust SSE stream transformer: upgraded reconstructedStream to a line-buffered TransformStream that intercepts and normalizes both finish_reason: "length" and finish_reason: "content_filter" to "stop";
  - Synthesized structured progress checkpoints when content filter is encountered with empty text, guaranteeing 100% successful session compaction.
0.1.22:
- Support OpenAI Responses API protocol & dual-provider architecture (Resolves Issue #1):
  - Dual-provider routing: Partitioned OpenCode Zen into opencode-zen-free (openai-completions) and opencode-zen-free-responses (openai-responses);
  - Fix 400 ModelProtocolUnsupported error: Serves Muse Spark models via native Responses API;
  - Responses protocol alignment: Flat tool definitions (OPENCODE_OFFICIAL_RESPONSES_TOOLS), tool_choice: "none" on compaction, no Chat-only fields;
  - Dual-endpoint fetch interceptor: Intercepts both /zen/v1/chat/completions and /zen/v1/responses;
  - Configuration persistence: Dual-writes both providers into models.json and auth.json.
0.1.21:
- Deep fix for upstream empty response retry (Provider returned an empty response):
  - Peek buffer increased to 30 packets and 64KB, penetrating initial gateway keepalive/processing comments;
  - Strict error boundary matching;
  - Exponential backoff retry with jitter;
  - Model definition sanitization (sanitizePiModelDefinition).
0.1.20:
- 3-attempt same-model retry on empty response or zero-byte stream;
- Auto-discovery of stealth zero-cost models (space-bunny-alpha, etc.).

8. License

Repository: https://github.com/1837620622/ck-pi-extension
License: MIT License