v0.0.11

v0.0.11（2026-07-17）

🇬🇧 English

Flagship fuse upgraded to Kimi K3 (top-tier frontier model) — quality ceiling raised to the front of the pack.

Model (P1)

• flag-fuse (旗舰·融合) model changed qwen3.7-max → kimi-k3 (2.8T params, 1M context)
• README model table + intro updated; OpenCode Go quota 2x from 7/24 (140 → 280 / 5h)

---

🇨🇳 中文

旗舰融合升级至 Kimi K3（顶级前沿模型）—— 质量天花板推到第一梯队最前。

模型（P1）

• flag-fuse（旗舰·融合）模型由 qwen3.7-max 改为 kimi-k3（2.8T 参数，1M 上下文）
• README 模型表与介绍同步更新；OpenCode Go 额度 7/24 起翻倍（140 → 280 / 5h）

---

v0.0.10

v0.0.10（2026-07-16）

🇬🇧 English

Consistency & verification hardening (no agent/model changes).

Docs (P1)

• README.md / README.zh.md: added conditional phrasing to the "~90% cost saving" claim — savings are real vs all-flagship baseline but depend on task mix; existing "80/18/2 is design ratio, not measured" caveat preserved.

Fix (P1)

• install.ps1 / install.sh: verification no longer hard-codes agent/command/skill counts (22/5/3); now reports the actual dynamic counts, so adding agents won't cause false failures.

Tests (P1)

• T0-static-verify.ps1: fixed a PowerShell parse error (AGENT_COUNT\" -eq 19); made skill count dynamic; added README core-fact anchor checks for README.md (22 agents · 5 commands · 3 skills) and README.zh.md (22 个 agent · 5 个命令 · 3 个 skill). Now PASS: 48 FAIL: 0.

Chore (P2)

• Added docs/benchmark/ skeleton (task set B01–B15 + YAML template) to gather real-task evidence for cost/quality claims.

Chore (P2)

• TRANSLATION.md: promoted to "Translation Governance" — source-of-truth rules for multilingual docs, now states 3 core skills.
• docs/opengraph.png / docs/opengraph-zh.png: optimized (smaller byte size), same content.

🇨🇳 中文

一致性与验证加固（无 agent / 模型变更）。

文档 (P1)

• README.md / README.zh.md：为「~90% 成本节省」补充条件限定——对比全程旗舰基线属实，但实际取决于任务结构；保留既有「80/18/2 为设计比例非实测」说明。

修复 (P1)

• install.ps1 / install.sh：验证不再硬编码 agent/命令/skill 数量（22/5/3），改为报告实际动态计数，新增 agent 不再导致误报失败。

测试 (P1)

• T0-static-verify.ps1：修复 PowerShell 语法错误（AGENT_COUNT\" -eq 19）；skill 计数改为动态；新增 README 核心事实锚点检查（英文 22 agents · 5 commands · 3 skills、中文 22 个 agent · 5 个命令 · 3 个 skill）。现 PASS: 48 FAIL: 0。

杂项 (P2)

• 新增 docs/benchmark/ 骨架（任务集 B01–B15 + YAML 模板），用于收集真实任务的成本/质量证据。

杂项 (P2)

• TRANSLATION.md：升级为「翻译治理规范」——多语言文档的 source-of-truth 规则，现写明 3 个 core skills。
• docs/opengraph.png / docs/opengraph-zh.png：优化（体积变小），内容不变。

---

v0.0.9

v0.0.9（2026-07-15）

🇬🇧 English

MoA 2.0: Residual-enhanced fusion + confidence routing + 22 agents.

New agents (+2)

• 残差提取者 (Flash): analyzes divergence between 3 parallel plans; triggers only when consensus < 95%. Enriches fusion with structured difference data.
• 置信度评估者 (DS Pro): evaluates fusion output across 4 dimensions (hallucination, requirement alignment, spec conflict, feasibility), returns a 0–100 confidence score.

Fusion layer overhaul (residual-enhanced)

• 旗舰·融合 (Qwen3.7 Max, 16384 tokens): upgraded to 3-stage fusion — consensus identification → residual analysis → integration. Includes a "fast path" that skips residual when consensus ≥ 95%.
• 中级·融合 (Kimi, 16384 tokens): 2-stage simplified fusion (consensus + residual → output).
• 前端·总工 (GLM-5.2, 16384 tokens): adds confidence scoring dimension to frontend plan selection.
• 融合·保底 (DS V4 Pro, 16384 tokens): generic fallback fusion agent that inherits the residual-enhanced fusion flow. Triggered automatically when any primary fusion agent fails.
• All 3 fusion agents now have hidden: true (not shown in @ menu).

Quality assurance upgrade

• 旗舰·质检 (DS Pro, 16384 tokens): expanded to dual-role — plan review (with confidence assessment) AND code acceptance. Adds "learning records" (logs typical problem tags after each review).

Router upgrade

• 门童路由员 (Flash): added confidence estimation (4 dimensions: clarity, familiarity, context sufficiency, risk level) and VOC (Value of Computation) escalation rules. Expanded task whitelist to include the 2 new agents.

Token limit fixes

• All production-layer agents bumped from 8192 → 16384 max_tokens (旗舰·实现, 旗舰·质检, 旗舰·融合, 旗舰·架构, 旗舰·规划, 旗舰·工程, 中级·工程, 中级·创意, 中级·码农, 中级·融合, 前端·总工, 前端·逻辑, 前端·动效, 融合·保底).
• Frontend·还原 bumped from 8192 → 16384; added hidden: true.
• Analysis-layer agents (残差提取者, 置信度评估者): 4096 max_tokens.
• Tool-layer agents remain at 2048 (cost optimization).

Command updates

• /moa-flagship: reflects new flow (残差提取者 + 质检方案审查).
• /moa-medium: reflects new flow (残差提取者 + 可选质检).
• /moa-quick: marked as deprecated/equivalent to "small task" shortcut.

Skill updates

• architecture-moa/SKILL.md: reflects new flow (残差提取者 + 质检方案审查).

Documentation fixes

• README: agent count updated 19 → 22 (header, deploy section, FAQ).
• README: clarified "MCP permission isolation" — noted it refers to agent-level read: deny / bash: deny, not actual MCP server isolation (no MCP servers are configured).
• README: 80/18/2 ratio annotated as "designed call volume distribution, not measured cost proportion".
• T0 test script: updated to expect 22 agents, correct model assignments, 22 reasoningEffort entries.

Abandoned items (intentionally not done)

• P0-4: Clean up invalid bash allows in opencode.json — The grep/ls/cat allow entries on bash: deny agents are noise but harmless; they're still useful for bash: ask agents.
• P0-5: Unify 4 fusion agent prompts — Different input sources make unified prompts counterproductive; separate prompts allow fine-tuning per chain.
• P1-3: Deduplicate 3-document set (README + 2 manuals) — Each doc serves a different purpose (quick-start guide vs AI installer vs English installer); cross-referencing would increase maintenance burden.

Verification

• T0 static verification: 44/44 PASS (previously 40/40 at v0.0.8, +4 new checks for new agents/models).

🇨🇳 中文

MoA 2.0：残差增强融合 + 置信度路由 + 22 个 agent。

新增 agent（+2）

• 残差提取者 (Flash)：分析 3 份并行方案的差异；仅在共识覆盖率 < 95% 时触发，为融合层提供结构化差异数据。
• 置信度评估者 (DS Pro)：从 4 个维度评估融合输出（幻觉检测、要求对齐、规范冲突、可行性），返回 0–100 置信度评分。

融合层重构（残差增强）

• 旗舰·融合 (Qwen3.7 Max, 16384 token)：升级为三阶段融合 — 共识识别 → 残差分析 → 整合。含「快速通道」：共识 ≥ 95% 时跳过残差分析。
• 中级·融合 (Kimi, 16384 token)：两阶段简化融合（共识 + 残差 → 输出）。
• 前端·总工 (GLM-5.2, 16384 token)：新增置信度评分维度。
• 融合·保底 (DS V4 Pro, 16384 token)：通用保底融合 agent，继承残差增强融合流程。任一主融合 agent 失败时自动触发。
• 3 个融合 agent 均设 hidden: true（不在 @ 菜单显示）。

质检升级

• 旗舰·质检 (DS Pro, 16384 token)：扩展为双重职责 — 方案审查（含置信度评估）+ 代码验收。新增「学习记录」（每次审查后记录典型问题标签）。

路由升级

• 门童路由员 (Flash)：新增置信度估计（4 维度：需求清晰度、技术熟悉度、上下文充分性、风险等级）和 VOC（计算价值）升级规则。task 白名单扩展至包含 2 个新 agent。

Token 上限修复

• 所有产出层 agent 从 8192 → 16384 max_tokens（旗舰·实现/质检/融合/架构/规划/工程，中级·工程/创意/码农/融合，前端·总工/逻辑/动效，融合·保底）。
• 前端·还原 从 8192 → 16384；添加 hidden: true。
• 分析层 agent（残差提取者、置信度评估者）：4096 max_tokens。
• 工具层 agent 保持 2048（成本优化）。

命令文件更新

• /moa-flagship：反映新流程（残差提取者 + 质检方案审查）。
• /moa-medium：反映新流程（残差提取者 + 可选质检）。
• /moa-quick：标记为废弃/等价于「小任务」快捷方式。

Skill 更新

• architecture-moa/SKILL.md：反映新流程（残差提取者 + 质检方案审查）。

文档修正

• README：agent 计数 19 → 22（头部、部署节、FAQ）。
• README：澄清「MCP 权限隔离」— 实际指 agent 级别的 read: deny / bash: deny，并非真实 MCP 服务器隔离（项目未配置 MCP 服务器）。
• README：80/18/2 比例标注为「设计调用量占比，非实测成本占比」。
• T0 测试脚本：更新为期望 22 个 agent、正确的模型分配、22 个 reasoningEffort。

放弃项（有意不做）

• P0-4: 清理 opencode.json 中无效的 bash allow — 对 bash: deny agent 无效但对 bash: ask agent 仍有用，保留无害。
• P0-5: 统一 4 个融合 agent 提示词 — 输入来源不同，统一反而失去针对性。
• P1-3: 三件套文档去重 — README/部署手册/英文手册各有定位，改为互相引用增加维护负担。

验证

• T0 静态验证：44/44 PASS（v0.0.8 时为 40/40，新增 4 项对新 agent/模型的检查）。

---

v0.0.8

v0.0.8（2026-07-11）

🇬🇧 English

Deploy-deviation fixes and 信达雅 (faithful / expressive / elegant) wrap-up across both language manuals.

• docs/opencode-moa.en.md moved from docs/en/ to docs/; README download link updated.
• Both manuals: fixed two deploy-blocking bugs — command file moa-Medium.md → moa-medium.md (case-sensitivity on Linux/WSL); Provider key handling now explicitly creates .opencode/local/opencode-go.key and references it via {file:} (was ambiguous → silent 401 on all 19 agents).
• Both manuals: unified router name (EN doorman → concierge-router); added a "Customization" note — agent names and model assignments are not hard-bound and can be swapped freely, with a warning that renaming an agent must sync the task: whitelist, permission.task, and all cross-agent @/task calls.
• READMEs (EN/ZH): 信达雅 fixes — qualified "~90%" savings (vs all-flagship baseline); unified concierge-router naming; fixed free-model name (Big Pickle → North Mini Code Free); fixed ZH typo (hero 的 → 这里说的).
• Doc footer version v0.0.6 → v0.0.8 in both manuals.

🇨🇳 中文

部署偏差修复与信达雅收口（两份语言手册）。

• docs/en/opencode-moa.en.md → docs/opencode-moa.en.md（移出 docs/en/）；README 下载链接同步。
• 两份手册修复两个部署阻断 bug：命令文件 moa-Medium.md → moa-medium.md（Linux/WSL 大小写敏感）；密钥落盘改为显式创建 .opencode/local/opencode-go.key 并经 {file:} 引用（原表述歧义 → 全 19 agent 静默 401）。
• 两份手册统一路由名（英文 doorman → concierge-router）；新增「自定义」备注——agent 名称与 model 不捆绑死、可自由替换，并提醒改名须同步 task: 白名单、permission.task 与跨 agent 调用。
• 中英文 README 信达雅修正：限定「省 90%」（对比全程旗舰基线）；统一 concierge-router 命名；修正免费模型名（Big Pickle → North Mini Code Free）；修正中文错别字（hero 的 → 这里说的）。
• 两份手册文档页脚版本 v0.0.6 → v0.0.8。

---

v0.0.7

v0.0.7（2026-07-11）

🇬🇧 English

Cross-platform keybindings and version-floor consistency wrap-up (extension of the v0.0.6 doc fixes, covering code / config / agent files not yet aligned).

• .opencode/skills/opencode-moa/SKILL.md: env-check version floor >= 1.1.1 → >= 1.3.4; deploy-success check "Ctrl+. to switch agent" → cross-platform "Tab cycle (Windows desktop also Ctrl+.)"; verify expectation "41 PASS" → "all PASS / FAIL=0 (with system-level key, WARN also counts as pass)"; removed the no-upload entry for the no-longer-generated user_config.json.
• install.sh / install.ps1: post-install keybinding hint changed from hardcoded Ctrl+. to cross-platform "Tab (Windows desktop also Ctrl+.)".
• .opencode/agents/concierge-router.md: free-model switch hint Ctrl+. (actually the agent key, conflicting with the keybinding convention) → /models and pick the Free tag (Windows desktop Ctrl+'); plan-agent switch Ctrl+. → cross-platform "Tab (Windows desktop Ctrl+.)".
• .github/ISSUE_TEMPLATE/bug-report.yml: version placeholder 1.1.1 → 1.3.4.

🇨🇳 中文

跨平台键位与版本门槛一致性收口（v0.0.6 文档修正的延伸，覆盖尚未对齐的代码 / 配置 / agent 文件）。

• .opencode/skills/opencode-moa/SKILL.md：环境检查版本门槛 >= 1.1.1 → >= 1.3.4；部署成功判定「Ctrl+. 切 agent」→ 跨平台「Tab 循环（Win 桌面端亦可用 Ctrl+.）」；验证预期「41 PASS」→「全部 PASS / FAIL=0（系统级 key 时 WARN 也算过）」；「不上传」清单移除已不生成的 user_config.json。
• install.sh / install.ps1：安装完成提示的切换键由写死的 Ctrl+. 改为跨平台「Tab（Win 桌面端亦可用 Ctrl+.）」。
• .opencode/agents/concierge-router.md：免费模型切换指引 Ctrl+.（实为 agent 键，与键位约定冲突）→ /models 选 Free 标签（Win 桌面端 Ctrl+'）；plan agent 切换 Ctrl+. → 跨平台「Tab（Win 桌面端 Ctrl+.）」。
• .github/ISSUE_TEMPLATE/bug-report.yml：版本占位符 1.1.1 → 1.3.4。

---

v0.0.6

v0.0.6（2026-07-11）

🇬🇧 English

This release = manual forceReasoning premise correction + i18n completion (en promoted from skeleton to full) + README rework.

1. Correct the manual's wrong premise that "custom provider must add forceReasoning: true"

Background

The v0.0.5 manual stated: "opencode ≥ 1.3.4, @ai-sdk/openai-compatible providers no longer pass reasoning params by default; you must add forceReasoning: true, otherwise the matrix silently fails." Per OpenCode issue #20815, that premise is false.

Correction (based on issue #20815 testing)

• The v1.3.4 reasoning passthrough regression only affects custom providers using "npm": "@ai-sdk/openai" (AI SDK v6 validates against a "known reasoning model list"; if not in the list it silently drops reasoningEffort).
• The same issue confirms @ai-sdk/openai-compatible is unaffected — reasoningEffort is correctly passed through as reasoning_effort.
• forceReasoning is a switch to bypass the list validation for @ai-sdk/openai; our provider uses openai-compatible, so it should not (and does not need to) be added (adding it is a no-op and misleading). Only when you change npm to @ai-sdk/openai (e.g. to use the responses API) should you add forceReasoning: true in options.

Changes

• docs/opencode-moa.md: Provider config warning, example JSON (remove forceReasoning line), matrix prereq, troubleshooting "reasoning intensity unchanged" row, version note, and footer all corrected to the SDK-distinction wording; doc version v0.0.5 → v0.0.6.
• Kept consistent with the existing provider blocks in install.ps1 / install.sh (which never had forceReasoning).

Empirical evidence

v0.0.4 already recorded reasoning_effort rising from medium→max gave reasoning tokens 359→536, proving passthrough works natively on openai-compatible, consistent with the issue conclusion.

2. i18n completion (en promoted from skeleton to full)

• Added README.md (English homepage): a full English version aligned with the Chinese README (README.zh.md).
• docs/en/opencode-moa.en.md: promoted from ~skeleton to a full English version section-by-section matching the Chinese manual (Provider config, matrix, troubleshooting, FAQ, etc.); version synced to v0.0.6.
• docs/TRANSLATION.md: zh/en bilingual status flipped from "🚧 skeleton (to translate)" to "✅ complete".

3. README rework

• Restructured the deploy prerequisites table and the system-level key-path warning: added a "read before deploy: don't misplace the key path" alert box (project-level self-contained vs system-level shared, pick one; pin %USERPROFILE%\.config\opencode not %APPDATA%\opencode).
• Moved Q&A from the middle of the body to a unified "FAQ" at the end, and fixed several cross-references (e.g. "can't see the doorman", "where are free models" now point to the right sections).
• Added a "why ~90% cheaper" cost breakdown: weighted by call volume (tool layer 80% / opinion layer 18% / fusion layer 2%) estimates effective output price ≈ $0.69/1M, ~90% cheaper than the flagship baseline, ~80% cheaper than a single mid-tier.
• hero / architecture diagram model names updated to concrete versions (e.g. Qwen3.7 Max, MiniMax M3), aligned with the cost table; unified Qwen3.7Max/Qwen3.7Plus to the spaced spelling across the doc.

4. Other

• Deleted docs/PLAN-moa-hardening.md (plan consolidated into per-version notes) and removed v0.0.5's reference to it.
• Keybinding wording changed to cross-platform: agent switching unified to Tab cycle (or Ctrl+x a to open the agent list); free models via /models and pick the Free tag; your tested Windows desktop shortcuts Ctrl+. (switch agent) / Ctrl+' (switch free model) kept as "Windows desktop" annotations. Fixed the earlier bug of hardcoding these two keys, which conflicted with the official defaults (Tab//models/Ctrl+t; and Ctrl+. is officially input redo, no Ctrl+' binding) — covering README.md (EN) / README.zh.md (ZH) / docs/opencode-moa.md / docs/en/opencode-moa.en.md everywhere relevant.
• Added a cross-platform warning: do not manually switch "variant / reasoning tier" in the TUI — OpenCode's variant selection overrides the agent's reasoningEffort and writes to the model-selection cache (~/.local/state/opencode/model.json on Linux/macOS, %USERPROFILE%\.local\state\opencode\model.json on Windows); it persists after restart and is cross-platform consistent, silently overriding this plan's low→xhigh matrix. Added one line each in the reasoning-matrix section and the "reasoning intensity feels unchanged" troubleshooting row of docs/opencode-moa.md / docs/en/opencode-moa.en.md.
• /connect command-palette key fix: Method B originally said "press Ctrl+K in TUI to open command palette", but the official command_list / command palette = Ctrl+P, and no Ctrl+K binding exists in the full official keybinds table (Ctrl+K in the input box is actually "delete to end of line"). Changed to "type /connect in TUI (or press Ctrl+P to open command palette)", covering docs/opencode-moa.md / docs/en/opencode-moa.en.md.

🇨🇳 中文

本期 = 手册 forceReasoning 前提纠正 + 多语言补全（en 由骨架升至完整）+ README 重构。

1. 纠正手册「自定义 provider 必须加 forceReasoning: true」的错误前提

背景

v0.0.5 手册写入：「opencode ≥ 1.3.4 起，@ai-sdk/openai-compatible provider 默认不再透传 reasoning 参数，必须加 forceReasoning: true，否则矩阵静默失效」。经核实 OpenCode issue #20815，该前提不成立。

纠正（基于 issue #20815 实测）

• v1.3.4 的 reasoning 透传回归只影响 "npm": "@ai-sdk/openai" 的自定义 provider（AI SDK v6 起按「已知推理模型列表」校验，不在表中就吞掉 reasoningEffort）。
• 同一 issue 确认 @ai-sdk/openai-compatible 不受影响，reasoningEffort 会正确透传为 reasoning_effort。
• forceReasoning 是给 @ai-sdk/openai 绕过列表校验的开关；本项目 provider 用的是 openai-compatible，无需也不应加（加了是 no-op，且会误导）。仅当把 npm 改成 @ai-sdk/openai（如用 responses API）时才需在 options 加 forceReasoning: true。

改动

• docs/opencode-moa.md：Provider 配置警告、示例 JSON（移除 forceReasoning 行）、矩阵前置依赖、排错表「推理强度没变」行、版本说明、页脚均更正为 SDK 区分表述；文档版本 v0.0.5 → v0.0.6。
• 与 install.ps1 / install.sh 既有 provider 块（本就无 forceReasoning）保持一致。

实测佐证

v0.0.4 已记录 reasoning_effort 从 medium→max 时 reasoning tokens 359→536，证明透传在 openai-compatible 上原生生效，与 issue 结论一致。

2. 多语言补全（en 由骨架升至完整）

• 新增英文 README（现 README.md 首页）：与中文 README（现 README.zh.md）结构对齐的完整英文版。
• docs/en/opencode-moa.en.md：由 ~骨架 补全为与中文手册逐节对应的完整英文版（Provider 配置、矩阵、排错表、FAQ 等），版本同步至 v0.0.6。
• docs/TRANSLATION.md：中 / 英双语状态由「🚧 骨架（待译）」翻为「✅ 完整」。

3. README 重构

• 部署前置条件表、系统级 key 路径提示重构：新增「部署前必读：key 路径别放错」警告框（项目级自包含 vs 系统级共享二选一，钉死 %USERPROFILE%\.config\opencode 而非 %APPDATA%\opencode）。
• Q&A 段落从正文中部移到文末「常见问题」统一收口，并修正若干交叉引用（如「看不到门童」「免费模型在哪」指向对应小节）。
• 新增「为什么省 ~90%」成本拆解：按调用量加权（工具层 80% / 意见层 18% / 融合层 2%）估算有效输出单价 ≈ $0.69/1M，对比旗舰基线省 ~90%、对比单中端省 ~80%。
• hero / 架构图模型名更新为具体版本（如 Qwen3.7 Max、MiniMax M3），与成本表对齐；全文档 Qwen3.7Max/Qwen3.7Plus 统一为带空格写法。

4. 其他

• 删除 docs/PLAN-moa-hardening.md（方案已整合进各版本变更说明），并移除 v0.0.5 对其的引用。
• 切换键说明改为跨平台写法：agent 切换统一用 Tab 循环（或 Ctrl+x a 打开 agent 列表），免费模型用 /models 打开模型列表选 Free 标签；你实测的 Windows 桌面端快捷键 Ctrl+.（切 agent）/ Ctrl+'（切免费模型）保留为「Win 桌面端」标注。修正了此前把这两个键写死、与官方默认键位（Tab//models/Ctrl+t，且 Ctrl+. 官方为 input redo、无 Ctrl+' 绑定）冲突的问题，覆盖 README.md（英文）/ README.zh.md（中文）/ docs/opencode-moa.md / docs/en/opencode-moa.en.md 全部相关处。
• 新增跨平台警告：不要在 TUI 手动切「变体/推理档」——OpenCode 的变体选择会覆盖 agent 配置的 reasoningEffort 并写入 model 选择缓存（Linux/macOS ~/.local/state/opencode/model.json、Windows %USERPROFILE%\.local\state\opencode\model.json），重启仍生效、跨平台一致，会静默顶掉本方案的 low→xhigh 矩阵。已在 docs/opencode-moa.md / docs/en/opencode-moa.en.md 的推理矩阵段与「推理强度感觉没变」排错行各补一条。
• /connect 命令面板键纠错：方式 B 原写「TUI 内按 Ctrl+K 打开命令面板」，但官方 command_list / 命令面板 = Ctrl+P，全官方 keybinds 表无 Ctrl+K 绑定（Ctrl+K 在输入框实为「删除到行尾」）。已改为「TUI 内输入 /connect（或按 Ctrl+P 打开命令面板）」，覆盖 docs/opencode-moa.md / docs/en/opencode-moa.en.md。

---

v0.0.5

Full Changelog: v0.0.4...v0.0.5

v0.0.4

Full Changelog: v0.0.3...v0.0.4

v0.0.3

Full Changelog: v0.0.2...v0.0.3

v0.0.2

Full Changelog: v0.0.1...v0.0.2