Skip to content

CHANGELOG.md

Latest commit

 

History

History
398 lines (312 loc) · 30.6 KB

CHANGELOG.md

File metadata and controls

398 lines (312 loc) · 30.6 KB
Raw

Changelog

v3.0.0 (2026-06-08) — Plugin Architecture: Deterministic Policy Engine

Breaking Changes

  • Architecture rewrite: Governance rules moved from prompt-embedded NEVER rules (47KB, 66+ rules) to deterministic YAML policy engine (5 files, 31 rules). Plugin enforces rules at infrastructure layer — NEVER rules are no longer "polite requests" to LLMs.
  • Orchestrator agent: Reduced from 1130 lines (47KB) to 80 lines (3KB). All governance logic extracted to policies/base/*.yaml.
  • write/edit PHYSICALLY DENIED: Orchestrator's write and edit tools blocked at plugin layer (I1). Structural enforcement, not prompt-based. Inspired by Microsoft AGT's "Actions the kernel denies are STRUCTURALLY IMPOSSIBLE."

Added

  • Policy Engine (src/engine/policy-engine.mjs): Deterministic YAML policy evaluator with PolicyDocument/PolicyRule/PolicyCondition/PolicyEvaluator/PolicyDecision classes. Supports deny_overrides conflict resolution, 6 operators (eq/neq/in/contains/matches/gt/lt), priority-based rule ordering.
  • Audit Trail (src/audit/audit-trail.mjs): JSONL audit log with cryptographic hash chaining. Each entry linked to previous via MD5. Includes integrity verification and markdown export.
  • Gate System (src/gates/gate-system.mjs): File-based stage gates (.approved JSON) with MD5 content hashes, previous stage chain links, and composite hash for full custody chain.
  • 5 Policy Files (policies/base/):
    • never-rules.yaml: I1 (write/edit/bash denial), I4 (fullstack→deep), I2/I6 (dispatch verification), I3 (discovery), I5 (auditor)
    • stage-gates.yaml: Stage transition gates with hash verification (discovery→designer→planner→builder)
    • dispatch-rules.yaml: Builder count ≥ subsystem count, Visual Lead Protocol for UI
    • depth-rules.yaml: Depth classification (quick/standard/deep) + fullstack FORCE deep + auto-upgrade rules
    • quality-gates.yaml: GWT coverage, UI Visual Concept + Delight + Interaction States, Polish Round, API Contract, Auditor Pass 0
  • Pre-Builder Gate: Pattern-matched UI validation blocks Builder dispatch if Visual Concept, ≥2 Delight Elements, or Interaction States missing.
  • Plugin Hook: tool.execute.before: I1 write/edit structural denial + Pre-Builder Gate + edit reliability (3-level matching).
  • Plugin Hook: tool.execute.after: Post-edit verification.
  • Plugin Hook: experimental.chat.messages.transform: Bootstrap injection (carried forward, enhanced with v3.0 architecture summary).

Changed

  • package.json: v3.0.0, added files field for npm publishing, enhanced keywords, proper repository URL.
  • README.md: Updated for v3.0 plugin architecture with OpenCode installation methods.
  • Orchestrator agent (agents/forcoding.md): Simplified from 1130 lines to 80 lines. References plugin for enforcement. Pre-Flight Gate retained as checkpoint, not sole enforcement.
  • forcoding-core SKILL.md: Updated description to v3.0 with policy-engine reference.

Inspired By

  • Microsoft Agent Governance Toolkit (AGT) — Policy Engine design, pre-execution interception, deny_overrides strategy
  • NLAH (arXiv:2603.25723) — File-backed state, verifier separation
  • SAL (arXiv:2604.22136) — Evidence Chain, content hash verification
  • FORGE (arXiv:2602.16708) — Deterministic correctness guarantee via reference monitor pattern
  • SARC (arXiv:2605.07728) — Constraints as first-class architectural objects

Migration from v2.9

  • All 13 improvements (I1-I13) from v2.7-v2.9 are now structurally enforced by the Policy Engine rather than prompt-based.
  • Agent files are backward compatible (same skill loading, same task dispatch model).
  • Gate files (I11) and hash chains (I13) are now managed by the Gate System module.
  • HITL tiers (I7) remain documented in agent prompts; permission.ask hook planned for v3.1.

Added

  • Stage 1.5: Tech Stack Detection (orchestrator): 7 步检测流程（前端/状态管理/后端/基础设施/测试/AI·ML/架构模式）。输出 Tech Stack Profile 注入所有子智能体。
  • Builder Tech Stack Adaptive Skills: 16 项检测信号 → 技能映射表。
  • Planner Tech Stack Adaptive Skills: 12 项检测信号 → 技能映射表（含 Planning Impact 列）。
  • Designer Tech Stack Adaptive Skills: 10 项检测信号 → 技能映射表（含 Design Impact 列）。
  • Auditor Tech Stack Adaptive Skills: 8 项检测信号 → 技能映射表（含 Audit Focus 列）。
  • Drafter Tech Stack Adaptive Skills: 6 项检测信号 → 技能映射表（含 Spec Impact 列）。
  • Stage Routing Gate: 编导器根据深度（quick/standard/deep）跳过不必要的阶段。禁止线性执行全部 6 阶段。
  • forcoding-scout: 添加到权限块 + 编导器提示中，作为可选预探索步骤。

Fixed

  • Auditor checklist 缺失 3 项: Mobile (+"no hover-only on mobile")、Design System (+"animation curve consistent")、Delight (+"visual+functional not decorative")。与 Designer 的 23 项完全对齐。
  • 16 处 clean-code 引用: 全局 agents 目录下 8 个文件修复为 code-quality-category-pointer。同步修复全局 PROMPT-TEMPLATES.md 3 处。
  • 全局 agents 目录过期 (🔥 P0): ~/.config/opencode/agents/ 中 6 个 forcoding 子智能体文件全部为旧版本（最大差距 16KB→3.2KB）。forcoding-designer.md 完全缺失。OpenCode 的 task 工具只能从该目录发现子智能体，导致 Designer 无法被 dispatch。

Changed

  • 同步流程修正: 全局同步命令 Copy-Item 现已扩展，不仅同步 ~/.config/opencode/forcoding/，也同步 ~/.config/opencode/agents/ 下的 forcoding agent 文件。
  • forcoding-core RSI Reality Check 新增 "Registration Integrity": 验证所有子智能体 agent .md 文件是否同时存在于插件目录和全局 agents 目录。
  • forcoding-core Self-Check Table 新增 "Registration integrity": 门禁级 check，用于每次同步后的自检。
  • forcoding-core Capability Scan 新增第 5 项: Registration integrity audit。
  • forcoding.md Post-Sync Validation: 新增 PowerShell 验证脚本 + 3 条规则。注册完整性是一级门禁，同 Pre-Flight Gate 级别。

事故永久记录

2026-06-08 P0 事故: ~/.config/opencode/agents/ 目录中 6 个 forcoding agent .md 全部为旧版本（最大差距 16KB→3.2KB），forcoding-designer.md 完全缺失。根因是验证模型只检查插件目录 forcoding/ 的文件一致性，从未验证全局 agents 目录 agents/ 的注册完整性。修复措施：全局同步命令增加 agents 步骤 + RSI Reality Check 增加 Registration Integrity 项 + Self-Check Table 增加 Registration integrity 门禁。

  • Stage 编号歧义: 新增 Stage Routing Gate 明确 quick 跳 Stage 2-3、standard 跳 Stage 3。

Changed

  • forcoding.md: Stage 1.5 技术栈检测节 + ctx_reduce 禁令节 + Stage Routing Gate + Merge Protocol + Parallel Dispatch 更新 + Designer/Planner dispatch 行注入 Tech Stack Profile + scout 权限 + Post-Sync Validation
  • forcoding-builder.md: 新增 Tech Stack Adaptive Skills 节（16 项）+ Section Output Rule
  • forcoding-planner.md: 新增 Tech Stack Adaptive Skills 节（12 项）+ Same-File Parallel Planning
  • forcoding-designer.md: 新增 Tech Stack Adaptive Skills 节（10 项） + 版本 2.6.2→2.6.3
  • forcoding-auditor.md: 新增 Tech Stack Adaptive Skills 节（8 项） + checklist 补齐 3 项
  • forcoding-drafter.md: 新增 Tech Stack Adaptive Skills 节（6 项）
  • forcoding-core SKILL.md: Context Drop Protection 增强（阶段级约束）
  • skills.json: 版本 2.6.2 → 2.6.3
  • ~/.config/opencode/agents/: 8 个 agent 文件 clean-code → code-quality-category-pointer
  • ~/.config/opencode/PROMPT-TEMPLATES.md: 3 处 clean-code 修复
  • Section Marker Standard (forcoding-core): 全新信标格式 <!-- SECTION: name TYPE:type ORDER:n -->，支持同文件并行时 Builder 输出隔离
  • Merge Protocol (forcoding.md): 6 步合并流程（收集→分组→排序→验证→组装→输出），含 MERGE FAILED 回退机制
  • Parallel Dispatch Decision (forcoding.md): 新增 PARALLEL SAME-FILE 分支（Q2=N 但子系统独立时启用）
  • Builder Section Output Rule (forcoding-builder.md): Same-File Parallel 模式下强制使用 SECTION 标记包装
  • Same-File Parallel Planning (forcoding-planner.md): Plan 可定义同文件并行组 + section name/TYPE/ORDER
  • Same-File Parallel 规则 (forcoding-gate): Step 3 新增条件判断 + Step 4 新增回退机制
  • Prompt Quality Gate (forcoding-core): 新增 [PARALLEL SAME-FILE] 检查项

全链路技能矩阵（v2.6.3 最终版）

Agent task_type 技能 tech_stack 技能 checklist 项
Designer 5 task types × N skills 10 项检测 5 区 23 项
Drafter 7 task types × N skills 6 项检测 —
Planner 7 task types × N skills 12 项检测 26 项视觉标准
Builder 9 task types × N skills 16 项检测 26 项 visual-review
Auditor 8 task types × N skills 8 项检测 5 区 23 项 + 26 项 visual-review

Fixed

  • Planner design 技能缺口 (P0): 从 3 技能（design/uxui-principles/accessibility-a11y）扩展到 9 技能，对齐 Designer 和 Builder 的设计能力。新增：brainstorming/design-taste-frontend/popular-web-designs/ux-flow/forcoding-ui-checklist/ui-ux-pro-max。确保 Planner 在分解任务时了解视觉约束、移动端适配、设计系统令牌。
  • Auditor Pre-Delivery Checklist 过时 (P0): 从 9 项旧清单升级到 5 区 23 项清单（Visual 9 + Mobile 4 + DesignSystem 5 + Delight 2 + Polish 3）。Auditor 现在完整验证 Designer v2.6.1 定义的所有检查区。
  • 视觉审计数字不一致 (P1): 3 处 "22 项" 改为 "26 项"（Builder/Planner/Auditor）。部分通过区间从 12-15/22 改为 12-18/26。
  • Builder Polish 阈值不一致 (P1): "X < 16" 改为 "X < 19"，与 Auditor 的 ≥ 19/26 Pass 阈值对齐。防止 Builder 自认为修到 16 分即可但被 Auditor INVALIDATED。

Changed

  • 版本号统一升至 2.6.2: 7 个文件（skills.json / forcoding.md / designer / drafter / forcoding-core / forcoding-gate / AGENTS.md）
  • CHANGELOG: 新增 v2.6.2 条目

全链路技能对比

阶段 修复前 修复后 对齐
Designer 9 skills 9 skills ✅
Planner 3 skills 9 skills ✅
Builder 10 skills 10 skills ✅
Auditor 5 skills 5 skills —
Auditor checklist 9 项 23 项 (5区) ✅

Incident

专注森林流程事故 (2026-06-08): 8 子系统的 UI 任务被当作"一个 HTML 文件"处理，编排器完全绕过 Pre-Flight Gate、Designer、Planner、Auditor，自己写了 36.6KB 代码。14 项质量差距（配色/视觉效果/交互细节）vs 参考实现。

根因:

  • Pre-Flight Gate 被跳过（不数子系统、不定深度、不承诺委派）
  • Auto-Inject Manifest 技能知识在 dispatch 前被 ctx_reduce 删除（Context Drop Protection 违反）
  • Designer/Planner/Auditor 三道品质防线全部缺失
  • 编排器写了代码 — 核心规则违反

这与 2026-06-06 天气卡片 NaN bug 事故同根同源。

Added

  • forcoding-gate (新技能): 强制 Pre-Flight Gate。在 build/create/design/UI 任务时自动触发。4 步自检（数子系统→定深度→选流程→承诺委派）。如果 Builder count > 0，禁止编辑/写入工具。输出 Gate Report。包含辩解防护表（Rationalization Shield）。≤30 秒执行。
  • AGENTS.md 强制路由规则: build/design/UI/持久化/多交互/特效 任务必须走 @forcoding 工作流。激活步骤 4 步。违规警示含两次事故记录。
  • 三层反违规防护: L1 系统层 (forcoding-gate 技能) → L2 Agent 层 (AGENTS.md 路由规则) → L3 行为层 (Pre-Flight Gate 嵌入 agent)。

Changed

  • forcoding-core SKILL.md: Self-Check 表新增 forcoding-gate 条目。Skill Priority 列表新增 forcoding-gate 为第一 Process skill。
  • forcoding.md agent: 版本号 v2.6 → v2.6.1。描述增加 gate enforced + delegate-only 语义。
  • skills.json: 版本 2.6.0 → 2.6.1。forcoding_builtin 列表新增 forcoding-gate。

Fixed

  • Context Drop Protection 漏洞: 明确规则 — 提取 Manifest 前禁止 ctx_reduce 技能内容。AGENTS.md 新增对应条目。
  • Rationalization Trap 闭合: forcoding-gate 包含完整的辩解→真相对照表，覆盖本次事故中出现的所有借口。

Added

  • forcoding-design-md-bridge (新技能): ForCoding ↔ DESIGN.md 双向桥梁。支持 Google Labs design.md 规范的 YAML front matter 生成、{path.to.token} 引用语法、VoltAgent/awesome-design-md 73 品牌参考库索引、Builder prompt 令牌注入协议、CLI 工具集成。
  • DESIGN.md 生成流程: Phase 0 Visual Concept 同时输出 DESIGN.md YAML 令牌（colors/typography/rounded/spacing/components）。
  • Reference Analysis 升级: 使用 awesome-design-md 73 品牌按类别索引匹配项目风格。
  • DESIGN.md Token Compliance (Auditor Pass 5): 令牌使用率检查(颜色≥95%/圆角≥90%/间距≥80%)、引用完整性验证。
  • ui-ux-pro-max v2.5 深度集成: Design System Generator 作为 Phase 0 首选数据源、--persist 分层检索、Pre-Delivery Checklist (9项) 注入 Designer+Auditor。
  • Plan Self-Review (Planner): 输出前 5 项自审(spec coverage/placeholders/type consistency/task granularity/Visual Lead)。
  • Verification Gate (Auditor): 无新运行证据不能 VALIDATED。禁止"should pass"/"probably works"。
  • Anti-Rationalization Tables (Builder+Auditor): Superpowers-inspired 借口 vs 真实情况对照 + Red Flags 清单。
  • Subagent Context Isolation: NEVER pass session history to sub-agents。精而准构建每 agent 的上下文。
  • Continuous Execution: 自主模式不暂停、不发送进度更新、仅在 true BLOCKED 暂停。

Changed

  • Visual Concept 输出格式: 从自由文本改为 DESIGN.md YAML front matter + Markdown prose。
  • Phase 0 → Designer 流向: ui-ux-pro-max search.py → style/color/type → Visual Concept → DESIGN.md。
  • Designer 输出: 增加 DESIGN.md 格式规范 + Pre-Delivery Checklist。
  • Auditor 审查维度: 新增 Pass 5 DESIGN.md Token Compliance + Pre-Delivery Checklist 验证。
  • visual-review 标准: 从 16 项(PASS ≥11/16) → 22 项(PASS ≥16/22) → 26 项(PASS ≥19/26), 新增 G段(2026标准6项) + H段(Token Fidelity 4项)。
  • Pre-Builder Gate: 从 9 项扩展为 10 项(增加 DESIGN.md tokens)。
  • NEVER rules: 新增 6 条(Subagent Context Isolation / Continuous Execution / Plan Self-Review / Verification Gate / DESIGN.md tokens / DESIGN.md library)。
  • 工作流深度: standard 深度强制包含 Polish Round + 视觉审计。流程图更新为 Visual Lead + JS Specialists。

Fixed

  • Prompt Quality Gate 引用不一致: "16 项/11/16" → "26 项/19/26"。
  • Pre-Builder Gate 覆盖不全: 从 4 项扩展为 10 项。
  • Designer 未更新 2026 标准: 增加 WCAG 2.2/motion tokens/dark mode/触摸目标。
  • Planner 不知道 Visual Lead Protocol: 增加 VISUAL LEAD/JS ONLY/POLISH 任务标记。
  • Anti-Emoji 规则与用户需求冲突: 增加优先级规则(用户需求 > 技能规则)。
  • ctx_reduce 漏洞: 提取 Manifest 后才可 drop 技能内容(Context Drop Protection)。

v2.5.2 (2026-06-08) — Priority Triage over Edit-Down

Changed

  • 改进10重写: Edit-Down Pass（砍功能换预算）→ Priority Triage（功能分级）。AI 开发没有代码量瓶颈，所有用户需求全部保留。改为 P0/P1/P2/P3 分级告诉 Builder 精力分配，不是砍功能。
  • 改进13重写: Code Budget（代码预算百分比）→ Priority Triage 精力分配（P0 40%/P1 35%/P2 25%）。不再设硬性代码比例，而是精力优先级。
  • Pre-Builder Gate 更新: 去掉 Code Budget 检查，改为 Priority Triage 检查。
  • NEVER 规则更新: 去掉"必须有 code budget"，改为"必须有 Priority Triage"。
  • 所有文件同步: forcoding-core/forcoding.md/designer/builder 四个文件的"砍功能"语义全部替换为"功能分级"。

Rationale

AI 生成中功能数量不是质量瓶颈，prompt 详细程度才是。 编排器的核心职责 = 提供尽量详细可执行的指示，不是砍掉用户要的功能。

v2.5.1 (2026-06-08) — Quality Improvement Batch Integration

Added

  • Pre-Flight Gate (forcoding.md) — Mandatory 4-step self-check before ANY action: Count Subsystems → Determine Workflow Depth → Select Flow → Commit to Not Write Code. Cannot be skipped.
  • Delegation Report template (forcoding.md) — Orchestrator must output a structured report before dispatching Builders, with explicit subsystem count, workflow depth, and delegation targets.
  • Same-File Parallelism (forcoding-core) — Documented mechanism for dispatching multiple Builders to the same file when subsystems are independent.
  • Incident Report (docs/forcoding/learnings/2026-06-08-orchestrator-code-violation.md) — Permanent warning for the 58KB dream travel journal violation.
  • UI Task Quality Definition (改进9) — Separate quality model for UI tasks: visual impact + consistency + emotional + restraint. Code budget: visual+delight ≥ 45%.
  • Edit-Down Pass (改进10) — Feature audit step before Builder: "砍掉会有人抱怨吗？" checklist. Deliberate omissions list. Code budget allocation.
  • Visual Concept Phase (改进11) — Mandatory phase between Intent and Designer. Visual Concept Definition becomes the NORTH STAR for all downstream decisions.
  • 5-Phase UI Workflow (改进12) — Restructured: Visual Concept → Designer → Edit-Down → Builder+Polish → Auditor.
  • Code Budget Visualization (改进13) — Budget allocation section in Builder prompt: structure ≤15%, logic ≤30%, visual ≥30%, delight ≥15%, defense ≤10%.
  • Delight Floor Enhancement (改进14) — Expanded delight element list (sensory/visual/gamification/empowerment). Builder must explicitly select ≥2 items.
  • Auto-Inject Manifest Protocol (改进4) — Orchestrator reads skill → extracts manifest → embeds verbatim in Builder prompt. Must be imperative (must/must not).
  • Prompt Quality Self-Check Gate (改进2) — 7-item pre-dispatch checklist in forcoding-core. Orchestrator verifies before every Builder dispatch.
  • Auditor Pass 4 Expansion (改进7) — UI quality sub-checklist: CSS vars, safe-area, fonts, micro-interactions, IIFE, scroll-into-view, sample data.
  • User Value Level Gate (改进15) — Level 1-5 hierarchy. Target: Level 4 Delightful. Less than Level 4 → INVALIDATED.
  • forcoding-ui-checklist skill (改进1) — New skill with mobile UI best practices. Manifest auto-injected for design/build tasks.
  • Self-Check Table upgraded (改进8) — Added 5 new rows: prompt quality gate, polish round, auto-inject, code budget, delight selection.
  • NEVER rules expanded (改进5-6) — Added 5 new prohibitions: skip Visual Concept, skip Edit-Down, no code budget, no Polish Round, skip Pre-Flight Gate.

Fixed

  • Rationalization traps expanded — Added 6 new traps covering: "user gave complete spec" → delegate anyway, "loaded skill = executed" → loading ≠ executing, "single file = one builder" → subsystems are independent even in same file.
  • Quick workflow tightened — Added 3 more conditions: no interaction states, no touch/gesture, no platform adaptation.
  • Delegation mandate hardened — Added concrete violation examples from history (weather card, dream travel journal) as explicit warnings.
  • Cross-file consistency verified — 6 files checked for 11 keywords. All gaps identified and closed.
  • Reference Analysis phase (改进6) — Designer spec includes 3-minute reference scan: user-provided refs, websearch, project memory, ui-ux-pro-max.
  • Inspiration Pattern Library (改进16) — 8 proven high-impact patterns (sweep glow, conic-gauge, haptic, stagger, glass depth, breathing glow, number bounce, achievement badge). Builder injects 2-3 selected patterns.
  • Design System Seed (改进17) — Enhanced Phase 0 with full ui-ux-pro-max integration: search → curate → output Visual Concept. 3-step protocol.
  • Design System Persistence (改进18) — Multi-page persistence via --persist: MASTER.md (global) + pages/[page].md (overrides). Orchestrator archives then Builders consume.

v2.5.0 (2026-06-07) — Design Thinking + Flash-Only Architecture

Major: Design Thinking Double Diamond Workflow

v2.x completes the transition from task-execution to design-thinking. ForCoding now follows the full Double Diamond model: Insight (Empathize) → Define+Explore (Designer) → Plan (Planner) → Build (Builder) → Verify (Auditor) → RSI Check.

Added

  • forcoding-designer agent — New sub-agent replacing forcoding-scout+forcoding-drafter in the Define+Explore stage. Applies Kata 5問 + Given/When/Then + confidence tiers (confirmed/probable/hypothesis) + Green/Red zone classification. Generates NL VFs + Structured I/O contracts.
  • forcoding-edit-quality skill — Unified edit discipline + comment quality. Merges forcoding-reliable-edits + forcoding-clean-comments into one loadable unit.
  • forcoding-flash-engine skill — Flash model optimization engine. Documents Think Max as the only reasoning mode (Flash Max = 16/20, Flash Think = 9/20 in real-world 20-task test). 10 optimization principles with effect sizes.
  • SPOQ Wave-Based DAG Dispatch (Planner) — Complex tasks get wave-parallel scheduling with explicit dependency graphs.
  • VERIMAP Python VFs (Planner) — Deterministic Python verification functions generated per subtask for well-defined I/O.
  • VERIMAP NL VFs (Designer) — Natural language verification functions embedded in Given/When/Then acceptance criteria.
  • 4-Pass Auditor (Compiled AI-inspired): Pass 1 Security → Pass 2 Syntax → Pass 3 Execution (VFs) → Pass 4 Accuracy (UX+Quality)
  • Gate Metrics — Catch rate, escape rate, rework rate tracked per session → docs/forcoding/metrics.json
  • RSI Reality Check — 6-item self-audit after every task (Persistence, Acceleration, Auto-verify, Breadth, Economics, Bottleneck)
  • Insight-First Methodology — Constraint-driven discovery: "constraint + direction = creativity"
  • Interactive Questioning Protocol — Structured Q&A when uncertain, with "全自动" mode skipping
  • 4 Human Gates — Intent → Spec → Plan → Push, with skip rules per autonomy mode
  • Pro Escape Hatch — All-Flash default; Pro only on 3 Builder failures or 2 Auditor INVALIDATEDs
  • Longitudinal Quality Tracking — Per-session metrics with monthly review targets

Fixed

  • Flash-only architecture enforced — All stages use opencode-go/deepseek-v4-flash Think Max. Non-Think forbidden. Think High deprecated. Pro only via escape hatch.
  • Designer replaces Scout+Drafter — Design Thinking Define+Explore phase handled by single agent with structured spec format
  • Skill execution mandatory — All sub-agents must execute checklist items, not just load skills
  • UX injection at every stage — Designer writes user journey + interaction states, Builder implements all states, Auditor verifies completeness
  • Design taste enforcement — 20+ rule checklist from design-taste-frontend verified by Builder + Auditor

Changed

  • Workflow renamed: Standard 4-stage → 3-level depth (quick / standard / deep) with Design Thinking phases
  • Agent count: 6 → 7 (added forcoding-designer, kept forcoding-scout and forcoding-drafter as legacy)
  • Skill count: 7 → 9 (added forcoding-edit-quality, forcoding-flash-engine)
  • Skills.json: v2.5.0, restructured with install groups
  • All agents updated: New skill matching tables, mandatory skill execution, confidence declarations
  • forcoding-core expanded: 79→348 lines with Design Thinking, UX injection, VERIMAP, RSI, capability scan
  • forcoding-intent expanded: 69→98 lines with subsystem-based complexity classifier
  • Plugin bootstrap: Updated sub-agent list to include designer
  • Templates: Updated with Design Thinking flow references

Files Changed (21 files)

File Changes
agents/forcoding.md Full rewrite: Design Thinking, HARD-GATE, 6 stages, RSI, NEVER rules, Interactive Questioning
agents/forcoding-designer.md New: Kata 5問, Given/When/Then, confidence tiers, zone classification
agents/forcoding-builder.md Skill execution mandatory, pre-task context injection, doc sync required
agents/forcoding-auditor.md 4-Pass review, VF execution, gate metrics, PR template
agents/forcoding-planner.md SPOQ DAG wave dispatch, VERIMAP Python VFs, skill execution mandatory
agents/forcoding-drafter.md Flash optimization embedding table, skill execution mandatory
agents/forcoding-scout.md Updated skill matching, parallel exploration marker
skills/forcoding-core/SKILL.md 79→348 lines: Design Thinking, VERIMAP, RSI, UX injection, capability scan
skills/forcoding-intent/SKILL.md 69→98 lines: subsystem-based complexity, "single file ≠ trivial" hardening
skills/forcoding-edit-quality/SKILL.md New: unified edit discipline + comment quality
skills/forcoding-flash-engine/SKILL.md New: Flash Max only, 10 optimization principles
skills.json v2.5.0 restructured with install groups
package.json v2.5.0
CHANGELOG.md v1.2.0, v2.0.0, v2.5.0 entries
README.md v2.5 feature updates, designer agent
REQUIREMENTS.md Updated agent/skill counts, designer entry
.opencode/plugins/forcoding.js Bootstrap updated with designer
templates/AGENTS.md Added designer reference
templates/PROMPT-TEMPLATES.md Updated with Design Thinking flow

v1.2.0 (2026-06-07) — Enforcement Infrastructure

Major: From Passive Rules to Active Gates

v1.1 had rules. v1.2 adds enforcement. After the weather-cards-v2 accident (orchestrator wrote all code instead of delegating to 4 parallel builders), we audited every file for "rules exist but constraints are weak" patterns. 12 fixes applied.

Core problem: Passive rules ("should", "when you find yourself writing code") don't stop rationalization. Active gates ("before writing, answer these questions") do.

Added

  • <HARD-GATE> implementation gate (forcoding.md Stage 4) — Mandatory decision tree before ANY code writing
  • Two-stage review (forcoding-auditor.md) — Spec Compliance + Code Quality
  • Quality gates (forcoding.md Stages 1→2, 2→3, 3→4) — Each stage transition requires explicit verification
  • Parallel dispatch decision tree (forcoding.md Stage 4) — Visual decision tree replaces passive text rule
  • "Never" rules summary (forcoding.md end) — 10 absolute prohibitions
  • Skill checklist enforcement (forcoding-drafter.md, forcoding-planner.md) — Mandatory skill execution sections
  • Phase 5 audit consequences (forcoding.md) — Violations accumulate in session.json
  • Upward feedback loop (forcoding.md) — Sub-agents can signal structural defects upstream
  • Session persistence recovery (forcoding.md) — Unfinished sessions auto-detected on startup

Fixed

  • Auditor escape hatch eliminated — "Self-review if timeout" removed
  • Pre-review escape hatch fixed (forcoding-planner.md) — Timeout no longer skips pre-review
  • Doc sync mandatory (forcoding-builder.md) — Builder must output doc sync status after every commit
  • Builder context injection verified (forcoding.md) — Orchestrator checks pre-task context injection
  • Complexity classifier hardened — "Single file = trivial" rationalization explicitly blocked

Changed

  • All files converted to English — 9 files, all Chinese text translated
  • Self-check table expanded — Added HARD-GATE and sub-agent count rows
  • Orchestrator prompt restructured — Delegation mandate strengthened to "MUST" / "NEVER"

Superpowers Integration

Cross-audited against superpowers agent system. Adopted: <HARD-GATE> XML tag, two-stage review, decision tree visuals, "Never" rule list.

Files Changed

File Changes
agents/forcoding.md Full rewrite: HARD-GATE, quality gates, parallel tree, Never rules, session recovery
agents/forcoding-builder.md Eliminate self-review escape, mandatory doc sync
agents/forcoding-auditor.md Two-stage review structure
agents/forcoding-drafter.md Skill checklist enforcement section
agents/forcoding-planner.md Fix pre-review escape, skill checklist section
agents/forcoding-scout.md English conversion
skills/forcoding-core/SKILL.md HARD-GATE in self-check table
skills/forcoding-intent/SKILL.md English conversion

v1.1.1 (2026-06-07)

Fixed

  • Agent variant global consistency alignment: drafter→high, builder→high, spec-writer→high
  • Auditor routing table: fixed from L2 Think High → L3 Pro Think Max (two tables)
  • README agent config examples added variant column
  • REQUIREMENTS agent allocation table added superpowers-implementer/spec-writer rows
  • Cost distribution updated: ~75% L2, ~25% L3, saving 85% vs all-Pro

v1.1.0 (2026-06-07)

Added

  • Thinking intensity variants (off / high / max), Desktop UI dropdown + Ctrl+T
  • Model routing optimization: Planner and Auditor use Pro Think Max, rest use Flash
  • Provider switching: dual opencode-go and bifrost providers
  • Route table documentation: complete stage routing in agent definitions
  • Global consistency sync: opencode.json, agent .md, spec docs three-way alignment

Improved

  • Edit discipline validation: three-level matching (exact → normalized → fuzzy) + post-edit verification
  • Session persistence recovery: auto-detect and prompt resume on interruption
  • Learning extraction: auto-record learnings after each Builder task
  • Upward feedback loop: sub-agents signal architecture contradictions upstream

CLI

  • bin/forcoding.mjs unified CLI: evaluate / health / install / uninstall / init

v1.0.0 (2026-06-04)

Initial Release

  • 6 sub-agents (forcoding/scout/drafter/planner/builder/auditor)
  • 7 skills (core/intent/reliable-edits/clean-comments/flash-optimization/autopilot/parallel)
  • Two-stage intent refinement (Flash extract → classification mapping)
  • Four-stage workflow (Scout → Drafter → Planner → Builder)
  • Autonomous mode (gogogo)
  • Three-tier reasoning model routing (L1/L2/L3)
  • Self-test framework (5 tasks × 4 configs)
  • .opencode/plugins/forcoding.js plugin entry
  • templates/AGENTS.md + PROMPT-TEMPLATES.md