SuperDiscussion v0.2.0

SuperDiscussion v0.2.0

中文

v0.2.0 将 SuperDiscussion 从“实现后的代码审核”扩展为“实现前的设计审核 + 实现后的代码审核”，同时完成原计划中的本机可观察性与外发透明度加固。

一眼看懂这次更新

• 先审设计，再写代码。 现在可以直接用自然语言发起 Design Review，检查产品逻辑、用户流程、前端交互、状态模型和技术架构。
• 联网前先看清楚。 新增 preview-request，可以预览外部 Agent 将收到的上下文、不会收到的内容、目标 Provider 和请求哈希。
• 审核进度不再是黑盒。 新增 task-status，集中显示授权、冷却时间、审核轮次、候选版本、阻塞原因和下一责任方。
• 审核结果更可信。 候选、请求、reviewer 结果、Codex 裁定与最终归档均进行一致性校验；代码一旦改变，旧 approval 不会继续算数。
• Windows 上更稳定。 改进并发状态访问、短暂文件占用和账本追加过程，降低安全软件或同时读取造成的误失败。
• 升级无需重新配置。 配置与 runtime 继续使用 v2，现有 API Key、Provider 配置和正式审核记录保持兼容。

查看完整更新日志

自然语言 Design Review

• 用户不需要输入 design-review 命令。Codex 根据“先别开发”“审一下交互逻辑”“比较架构方案”等自然语言意图选择设计审核。
• 只有用户同时明确邀请 SuperDiscussion、外部 reviewer、Kimi、ZCode 或多 Agent 时才产生外部授权；普通设计问题仍由 Codex 单独回答。
• Design Review 支持一至三轮单审或双盲双审。每轮冻结设计候选，Codex 逐条裁定并只修改设计工件，不在该阶段写实现代码。
• 设计输出必须覆盖问题定义、假设、替代方案、权衡、产品逻辑、前端交互、架构与简化机会。缺少浏览器原型时必须如实记录。

可观察性与安全

• task-status 只读显示授权、Provider 冷却、审核轮次、候选哈希、缺失 reviewer、阻塞原因和下一责任方。
• preview-request 在联网前展示精确上下文投影、排除项、Provider 目标和请求哈希，不创建授权、不访问 Provider、不输出 API Key。
• invoke 与 review-team 可绑定预览得到的请求文件 SHA-256；调用前和重试前发现漂移即关闭失败。
• 新增两轮 Design Review 假 Provider 黑盒流程，覆盖真实设计工件、目标不可切换、最终归档，以及首次发送和结构化修复重试前的请求漂移。
• Reviewer 失败原因会持久化到只读状态视图；README 引用的二维码作为显式审计二进制资源进入确定性发布包。
• Windows JSON 原子替换与读取对短暂共享冲突进行有限重试，并保留失败临时文件清理与并发回归测试。
• 只读 ledger 可区分并发追加中的短暂尾行与持续损坏；reviewer waiver 会清除对应的旧 blocker，保持 runtime 可验证。
• 三轮部署前双审补强：review 只能经状态机派发，设计候选必须提供哈希绑定的完整可见工件，预览区分授权与实际调度就绪状态，账本与发布输入使用原子/不可变快照，审核状态记录实际发送请求的快照身份，外部 HTTP 后的漂移仍留下失败审计，v2 quorum、裁决和降级语义严格校验，降级失败摘要在归档及后续加载时复核，最终来源清单必须精确导出最终候选哈希，二维码按长度与 SHA-256 固定。

配置格式与 runtime 继续使用 v2，旧 review-team 代码审核请求、授权链、候选哈希、归档和最多三轮限制保持兼容；独立 review 不再允许绕过 review-team 直接调用。

SuperDiscussion 可以显著提高设计与代码质量，但会增加 Token 消耗和完成时间，请按任务价值选择使用。

English

v0.2.0 expands SuperDiscussion from post-implementation code review into pre-implementation Design Review plus post-implementation code review, while delivering the planned local observability and outbound-context transparency improvements.

What changed at a glance

• Review the design before writing code. Natural-language Design Review now covers product logic, user flow, frontend interaction, state models, and architecture.
• See outbound context before networking. preview-request shows what an external agent will receive, what is excluded, the destination provider, and the request hash.
• Review progress is no longer a black box. task-status reports authorization, cooldowns, review rounds, candidate identity, blockers, and the next responsible actor.
• Review conclusions are harder to misrepresent. Candidates, requests, reviewer results, Codex adjudication, and final archives are consistency-checked; an old approval cannot survive a code change.
• More reliable on Windows. State access, transient file sharing, and concurrent ledger appends are handled more robustly.
• No configuration migration. Config and runtime remain at v2, preserving existing API keys, providers, and formal review records.

Read the complete changelog

Natural-language Design Review

• Users do not need a design-review command. Codex infers the design phase from ordinary requests to pause implementation, challenge an interaction, or compare architectures.
• External authorization still requires an explicit invitation to SuperDiscussion, external reviewers, Kimi, ZCode, or multiple agents. Ordinary design questions remain Codex-only.
• Design Review supports one to three single or blind-dual rounds. Each round freezes the design candidate; Codex adjudicates findings and revises design artifacts without editing implementation code during the phase.
• Structured results cover problem framing, assumptions, alternatives, trade-offs, product logic, frontend interaction, architecture, and simplification. Missing browser prototypes are reported honestly.

Observability and safety

• task-status reads local grants, provider cooldowns, review rounds, candidate hashes, missing reviewers, blockers, and next actors.
• preview-request shows the exact context projection, exclusions, destination, and request hash before networking without creating a grant, contacting a provider, or exposing an API key.
• invoke and review-team can bind the previewed request-file SHA-256 and fail closed before calls or retries when content drifts.
• Adds a two-round fake-provider Design Review black-box flow covering real design artifacts, immutable targets, final archive binding, and request drift before initial sends and structured repair retries.
• Reviewer failure reasons persist in the read-only status view, and the README-linked QR image ships as an explicitly audited binary in the deterministic archive.
• Windows atomic JSON replacement and reads now perform bounded retries for transient sharing violations while retaining temporary-file cleanup and concurrency regression coverage.
• Read-only ledger inspection distinguishes a transient append tail from persistent corruption, and reviewer waiver clears the corresponding stale blocker so runtime state remains valid.
• Three pre-deployment dual-review rounds harden state-machine-only review dispatch, hash-bound full design artifacts, authorization-versus-dispatch preview readiness, atomic ledger and immutable release snapshots, identity of the request snapshot actually sent, failed auditing after post-contact drift, strict v2 quorum/adjudication/degradation semantics, degradation-evidence revalidation, exact final-source candidate identity, and length/SHA-256 pinning for the QR asset.

Config and runtime remain v2. Existing review-team code-review requests, grants, candidate hashes, archives, and the three-round ceiling remain compatible; standalone review calls can no longer bypass review-team.

SuperDiscussion can materially improve design and code quality at the cost of additional tokens and completion time. Use it selectively.

SuperDiscussion v0.1.2

SuperDiscussion v0.1.2

中文

本版本完成审核状态机与本机安全边界重构，不会自动启用任何外部 Agent。

审核架构

• Codex 固定担任 coder 和最终裁决者。常规审核中的 Kimi/ZCode 只担任 reviewer；显式外部 judge 仅提供咨询意见。
• 支持单审或双审 × 一至三轮共六种策略。用户已经明确要求审核但未指定策略时，默认一轮双盲审核；普通对话仍然只使用 Codex。
• 每轮绑定精确候选 SHA-256。只有 reviewer 对最终未修改的同一候选完整通过，才产生 reviewer_approved 或 dual_approved。
• 新增 Codex 逐 finding JSON 裁定，以及 codex_finalized、codex_finalized_with_dissent、blocked 状态。旧 self_accepted 只读兼容。
• 双审缺席不会静默降级。持续 quota/overload 后，只有用户明确同意才能记录 degraded_to_single_by_user，且永远不能显示为双审通过。

本机与网络安全

• runtime/config/ledger 使用跨平台内核锁与持久化原子写；进程崩溃不会遗留伪所有者锁。
• 每个 reviewer 槽位在外部调用期间持有独立跨进程内核锁；重复子请求与并发跳过操作必须等待并重新验证周期状态，不能在周期关闭后继续外呼。输出占位也使用内核锁，进程崩溃不会遗留永久阻塞同一路径的结果锁。
• 授权新增目的地指纹，绑定端点、模型、网络范围和直接/代理路径；旧授权升级后要求重新确认。
• 环境代理不再隐式继承。DNS 全部结果和实际连接地址均执行 public/loopback/private 策略，拒绝 metadata、link-local、CGNAT、保留、组播与 mapped-address 绕过。
• 连接建立默认 30 秒；300 秒默认值只用于流式无进展检测；非流式深度审核默认无读取总时限，也没有输出 token 上限。
• structuredTemperature 与 discussionTemperature 可分别设置，null 表示不发送。

更新与兼容

• 新增 scripts/update.py check|prepare|apply|rollback 官方稳定版更新链，以及 detached Ed25519 签名、逐文件哈希、安全解包、备份、journal 和回滚。
• 每次目标文件替换会先同步文件与目标目录，再推进 durable journal；already_applied 会重新核对全部签名文件，发现缺失或篡改即安全回滚并明确报错。
• README 中英文均提供一句可复制的签名更新 Prompt；更新不会读取 API Key、修改用户状态或调用外部 Agent。
• v0.1.1 配置和运行状态会事务迁移并备份。未完成的旧审核周期不会被静默套用新 approval 语义。
• CI 覆盖 Windows、macOS、Linux 上的 Python 3.11 与 3.13。
• 配置说明按当前对话选择中文或英文；Windows 命令显式启用 UTF-8，不再受系统 GBK/CP936 默认编码影响。

本 Skill 能显著提高代码质量，但会增加 Token 消耗与完成时间，请按任务风险酌情使用。

English

This release rebuilds the review state machine and local security boundaries without automatically enabling any external agent.

Review architecture

• Codex is fixed as coder and final adjudicator. Kimi/ZCode are reviewers in the normal review flow; an explicitly requested external judge is advisory only.
• Supports all six policies: single or dual review × one to three rounds. An explicitly authorized review with no selected policy defaults to one blind dual round; ordinary conversation still uses Codex alone.
• Every round binds the exact candidate SHA-256. reviewer_approved or dual_approved requires complete approval of the unchanged final candidate.
• Adds structured per-finding Codex adjudication and honest codex_finalized, codex_finalized_with_dissent, and blocked states. Legacy self_accepted is read-only history.
• Dual review never degrades silently. After persistent quota/overload failure, only explicit user consent may record degraded_to_single_by_user, which can never become dual approval.

Local and network security

• Runtime/config/ledger use cross-platform kernel locks and durable atomic writes; a crashed process cannot leave a false lock owner.
• Each reviewer slot holds a dedicated cross-process kernel lock during external work. Duplicate children and racing waiver operations must wait and revalidate cycle state, so they cannot contact a reviewer after closure. Output reservations also use kernel locks, so a crashed process cannot permanently block the same result path.
• Grants bind destination fingerprints covering endpoint, model, network scope, and direct/proxy route. Migrated old grants require reauthorization.
• Environment proxies are no longer inherited. Every DNS result and actual connected address must satisfy the public/loopback/private policy; metadata, link-local, CGNAT, reserved, multicast, and mapped-address bypasses are denied.
• Connection setup defaults to 30 seconds. The 300-second default applies only to streaming no-progress detection. Non-streaming deep review has no default read-duration cap, and no output-token cap is imposed.
• structuredTemperature and discussionTemperature are independent; null omits the field.

Updates and compatibility

• Adds the official scripts/update.py check|prepare|apply|rollback stable update path with detached Ed25519 signatures, per-file hashes, safe extraction, backup, journaling, and rollback.
• Every target replacement syncs both file data and its destination directory before advancing the durable journal. already_applied rechecks every signed file and safely rolls back with an explicit error on missing or changed content.
• Both READMEs include one copyable signed-update prompt. Updating neither reads API keys nor changes user state nor calls an external agent.
• v0.1.1 config and runtime state migrate transactionally with backups. Incomplete legacy review cycles never silently acquire the new approval semantics.
• CI covers Python 3.11 and 3.13 on Windows, macOS, and Linux.
• Setup instructions select Chinese or English for the current conversation, and Windows commands explicitly enable UTF-8 instead of depending on the GBK/CP936 system default.

This Skill can substantially improve code quality at the cost of additional tokens and completion time. Use it according to task risk.

SuperDiscussion v0.1.1

SuperDiscussion v0.1.1

中文

本版本将项目从旧名称统一更名为 SuperDiscussion。Codex、Kimi 与 ZCode 的角色、显式授权边界、结构化审核、最多三轮复审和讨论能力均保持不变。

变更

• Skill 调用名改为 $superdiscussion。
• GitHub 仓库改为 supergitty520/superdiscussion。
• 运行脚本改为 scripts/superdiscussion.py。
• 用户配置目录改为 %APPDATA%\superdiscussion（Windows）或 ~/.config/superdiscussion（macOS/Linux）。
• 项目本机审计目录改为 .superdiscussion。
• 发布包改为 superdiscussion-0.1.1.zip。
• 对 v0.1.0 用户数据目录和环境变量保留迁移兼容，避免升级后丢失本机配置。
• 完全卸载可清理已识别的新旧用户数据目录；自定义目录需要精确路径确认，项目审核记录不受影响。
• 跨平台 HTTP 错误处理现在可容忍服务端在返回错误状态后立即重置连接，同时保留原始状态分类和凭据重定向保护。

使用成本提示：本 Skill 将极大提高代码质量，但代价是更高的 Token 消耗和更长的完成时间，请酌情使用。

English

This release adopts SuperDiscussion as the unified project name. Codex, Kimi, and ZCode roles, explicit authorization boundaries, structured review, the three-iteration review ceiling, and discussion behavior remain unchanged.

Changes

• The Skill invocation name is now $superdiscussion.
• The GitHub repository is now supergitty520/superdiscussion.
• The runtime entry point is now scripts/superdiscussion.py.
• User configuration moves to %APPDATA%\superdiscussion on Windows or ~/.config/superdiscussion on macOS/Linux.
• Project-local audit data now uses .superdiscussion.
• The release archive is now superdiscussion-0.1.1.zip.
• Migration compatibility preserves discovery of v0.1.0 user data and environment variables.
• Complete uninstall purges recognized current and legacy user data; custom paths require exact confirmation, and project audit records remain untouched.
• Cross-platform HTTP error handling now tolerates a peer reset after an error status while preserving the original classification and redirect credential protection.

Cost notice: This Skill can greatly improve code quality, at the cost of higher token usage and longer completion time. Use it selectively.

Agent Council v0.1.0

Agent Council v0.1.0

中文

这是 Agent Council 的首个公开版本。它作为 Codex Skill 工作，不替换或 fork Codex Harness。外部 Agent 默认关闭，只有用户明确授权后才会调用。

主要能力：

• Kimi K3 与 ZCode GLM-5.2 可配置为 coder、reviewer 或 judge。
• 支持一个 reviewer 或两个相互隔离的 blind reviewer。
• 自动复审最多三轮；本机状态机强制轮次、quorum、缺席 reviewer 重试和第四轮禁令，第三轮仍未达到 quorum 时由 Codex 完成最终修复并记录为 self_accepted。
• Review/Judge 使用结构化 JSON；群体讨论支持自然语言流式输出和结构化轮次状态。
• 上下文按角色裁剪，不会把整个聊天记录发送给外部模型。
• 支持授权 ledger、密钥扫描、quota 冷却、失败恢复和完整本机审核归档。
• 支持 Windows、macOS 和 Linux。
• README 提供一句可复制到 Codex 的安装指令；安装不会自动授权或调用外部 Agent。
• 提供跨平台安全卸载脚本：默认仅删除 Skill，完全卸载需明确授权删除用户级配置与状态，项目审核记录始终保留。
• Kimi Code、Moonshot 通用 API、Z.AI Coding Plan 与 Z.AI 通用 API 通过显式 endpointProfile 区分，避免密钥、余额和订阅额度跨产品误诊。
• 发布包由冻结清单生成，使用跨平台确定性的无压缩 ZIP，并对畸形 Provider 响应、讨论参与者和 quota/cooldown 边界进行黑盒验证。
• 发布扫描与运行时采用一致的凭据形状规则；审核归档会读取并校验所有基础请求、reviewer 子请求、结果和源码文件，绑定后持续复验归档哈希。
• 内建端点要求字面一致；畸形 message part、非法审核历史、注入后超限上下文会失败关闭，并兼容 Kimi 的 cached_tokens usage 字段。
• 16–19 位 Bearer 凭据也会被发布扫描拦截，最大长度讨论立场可安全注入后续轮次；连续 overload/quota 的 reviewer 仅在用户明确授权后留痕跳过，绝不伪造 verdict。
• 结构化讨论立场及跨 Agent 摘要上限提高到 16,000 字符，完整自然语言流仍保存在本机，请求总安全边界保持 200 万字符。
• 每条外部授权链由本机随机 grant 绑定当前 Codex 对话、项目目录与任务；支持单 Provider、任务、整段对话撤销和全局暂停，恢复必须获得新的明确授权。
• 连通性与实质协作 grant 用途隔离；每次 HTTP 重试和后续请求都重新检查授权，任务级恢复不会意外解除对话级封锁。

安装后在 Codex 中发送 进行配置，按照说明在本机文件中填写 API Key。不要将密钥粘贴到聊天或提交到仓库。

本版本采用 AGPL-3.0-only：分发修改版或通过网络提供修改版服务时，必须向相应用户提供完整对应源代码。

English

This is the first public Agent Council release. It runs as a Codex Skill and does not replace or fork Codex Harness. External agents remain disabled until the user explicitly authorizes them for the current task.

Highlights:

• Kimi K3 and ZCode GLM-5.2 may be configured as coder, reviewer, or judge.
• Supports one reviewer or two mutually isolated blind reviewers.
• Automated remediation stops after at most three review iterations. Local machine state enforces quorum, missing-reviewer retries, and the fourth-review prohibition; if quorum still does not approve, Codex makes the final fixes and records self_accepted.
• Review and judge flows use structured JSON; group discussion supports natural-language streaming with structured round state.
• Context is projected by role rather than sending the complete conversation to external models.
• Includes authorization ledgers, secret scanning, quota cooldown, failure recovery, and complete local review archives.
• Supports Windows, macOS, and Linux.
• The README includes a single copyable Codex installation prompt; installation does not authorize or call external agents.
• Includes a cross-platform safe uninstaller: the default removes only the Skill, a complete purge requires explicit authorization for user-level data, and project audit records are always preserved.
• Explicit endpointProfile values distinguish Kimi Code, Moonshot general API, Z.AI Coding Plan, and Z.AI general API so credentials, balances, and subscription quotas are not diagnosed against the wrong product.
• Release archives come from a frozen manifest, use cross-platform deterministic stored ZIP entries, and add black-box coverage for malformed provider responses, discussion participants, and quota/cooldown boundaries.
• Release scanning mirrors runtime credential-shape rules. Review archives read and hash every base request, reviewer child request, result, and source file, then revalidate the bound archive hash on later loads.
• Built-in endpoints require literal matches; malformed message parts, impossible review histories, and oversized post-injection context fail closed, while Kimi cached_tokens usage metadata remains compatible.
• Release scanning rejects 16–19 character Bearer tokens, maximum-length discussion stances remain safe across injected rounds, and repeatedly unavailable reviewers are waived only with explicit user authorization and never receive a fabricated verdict.
• Structured discussion positions and cross-agent summaries now allow 16,000 characters; the full natural-language stream remains local and the total projected request stays bounded at two million characters.
• Every external authorization chain uses a random local grant bound to the current Codex conversation, project directory, and task. Provider-specific, task, conversation, and global stop controls require fresh explicit authorization before restoration.
• Connectivity and substantive grants are purpose-isolated. Every HTTP retry and follow-up request rechecks authorization, and task restoration cannot accidentally lift a conversation-wide block.

After installation, send setup in Codex and enter API keys in the local file described by the Skill. Never paste keys into chat or commit them to the repository.

This release uses AGPL-3.0-only: distributing a modified version or providing a modified version as a network service requires offering the complete Corresponding Source to the applicable recipients or remote users.