specpowers v1.1.1 — skill slimming + 21 review fixes

Skill slimming — one authoritative definition per rule, loaded on demand.

v1.1.1 is a patch release: it restructures the specpowers skills from "every rule restated 2–4 times" into "one authoritative definition per rule + references loaded by step", and fixes 21 defects found by an independent whole-branch review. No judgment semantics changed — protocol numbers, subsection titles, the Gate chain, marker fields and enums are preserved verbatim.

Version note: the manifests carried an unreleased 1.2.0. Nothing was ever tagged or published at that version, so this ships as a patch on top of v1.1.0 and the manifests are now 1.1.1.

Size

File	Before	After
skills/specpowers/SKILL.md	386	341 (cap 345)
skills/specpowers-review/SKILL.md	657	480 (hard cap 480)
skills/specpowers-review/refs/	protocols.md, 777 lines	820 lines across 6 files (cap 820)
4 sub-skill SKILL.md (total)	628	619

refs/protocols.md is split by load timing: ledger.md (protocols 1–2), markers.md (3), dispatch.md (4–5), routing.md (6), ultrareview.md (8), convergence.md (7+9). Per-round cost now tracks the SKILL.md line count, and details are read only when the step that needs them is reached.

Fixes (21 of 34 findings: P0×2, P1×4, P2×15)

• P0 — expected_steps had three mutually contradictory definitions across the review recipes; the review protocol 9 subsection merge was ratified as a documented exception (the redline narrowed to "referenced subsection titles must not be renamed").
• P1 — the [拒绝异常] criteria were restated in 5 places; specpowers-design / specpowers-plan invoked the review skill without the args contract, silently losing Gate 0–2 tokens; three cross-file references pointed at the wrong file after the split.
• P2 — conservation-equation examples, the STEP_FINAL_READTHROUGH value set, tier-router 2b path fields, the empty-diff fallback in markers, a Windows note left inside a bash fence, and 10 more.

New guards

scripts/verify-dsh-provider.mjs gained two regression checks: every 本文件协议 N reference must point into the file that owns that protocol, and no refs/ file may point at a renamed or deleted refs file (75 references checked).

Breaking note

skills/specpowers-review/refs/protocols.md no longer exists. If anything outside this repo referenced it by path, point it at the file that now owns the protocol number (see the list above).

Install

# DeepSeek Harness — pinned
npx @deepseek-ai/dsh plugin --profile web add github:SAXEM1997/specpowers#v1.1.1

# track main
npx @deepseek-ai/dsh plugin --profile web add github:SAXEM1997/specpowers

Then restart the profile.

Verification

• CI on the release PR #10 (run 39596164327): Ubuntu, Windows and forced-CRLF legs all pass.
• CI on main after merge (run 39596293899): success.
• node scripts/verify-no-internal-refs.mjs: 0 hits across 93 tracked files, commit messages and author/committer identities.
• The final batch of 10 P2 fixes was re-reviewed by two independent verifiers (semantics ×4, pointer naming ×6) — all pass.
• evals/ was not run: neither skill-up nor claude is available on this machine. This release makes no claim that the eval suites pass.

Known gaps

13 findings from the same review remain open, all P3 (documentation wording, an environment checklist, historical docs/ references). They are logged in the review report and do not affect routing or gate behavior.

specpowers v1.1.0 — CRLF/BOM fix, warn-and-skip, provider CI

Windows/CRLF robustness + provider failure handling.

v1.1.0 is a minor release: fixes the CRLF/BOM frontmatter bug that broke every skill on Windows checkouts, adds warn-and-skip for malformed skill files, and puts the provider contract under CI on Linux, Windows and a forced-CRLF tree.

Fixes

• CRLF/BOM frontmatter (#2, PR #6): parseFrontmatter() normalizes CRLF and strips a leading UTF-8 BOM before parsing. Previously a Windows checkout produced name: "" for all six skills; DSH rejected the provider with invalid skill name "".
• Windows verifier import (PR #6): the self-check imports lib/index.js via pathToFileURL, fixing ERR_UNSUPPORTED_ESM_URL_SCHEME on D:\... paths found by the new Windows CI.

Enhancements

• Warn-and-skip malformed skill files (#4, PR #7): parseSkillFile() now warns with the offending path and skips a SKILL.md with an unparseable frontmatter or missing/empty name; the remaining skills keep registering, and get() returns undefined for the bad file.
• Provider CI (#5, PR #8): .github/workflows/provider.yml runs the self-check on Ubuntu, on Windows with skills converted to CRLF, and on a Linux CRLF simulation. It caught a real Windows bug on its first run.
• Line-ending policy: .gitattributes is now * text=auto eol=lf, so Windows checkouts stay LF; the CRLF regression is still exercised explicitly in CI.

Install

# DeepSeek Harness — pinned
npx @deepseek-ai/dsh plugin --profile web add github:SAXEM1997/specpowers#v1.1.0

# track main
npx @deepseek-ai/dsh plugin --profile web add github:SAXEM1997/specpowers

Then restart the profile. The installed profile may still pin v1.0.0; reinstall with #v1.1.0 to upgrade.

Verification

• Main CI run 39572919588: verify (ubuntu-latest), verify (windows-latest), verify (CRLF skills) all pass.
• Merged main self-check: node scripts/verify-dsh-provider.mjs → ALL PASS.
• Anonymous shallow clone at v1.1.0: same self-check → ALL PASS.

Compatibility

No breaking changes. LF checkouts behave as before; malformed skill files now warn and are skipped instead of surfacing as empty-name candidates.

---

specpowers v1.1.0（中文）

Windows/CRLF 健壮性 + provider 失败处理。

v1.1.0 是次版本：修复 Windows 检出下 CRLF/BOM 导致全部技能空名的 bug，新增解析失败告警并跳过，并把 provider 契约纳入 Linux / Windows / 强制 CRLF 的 CI。

修复

• CRLF/BOM frontmatter（#2，PR #6）：parseFrontmatter() 先归一化 CRLF 并去掉 UTF-8 BOM；此前 Windows 检出会让 6 个技能全部 name: ""，DSH 报 invalid skill name ""。
• Windows verifier import（PR #6）：自检用 pathToFileURL 导入 lib/index.js，修复新 Windows CI 抓到的 ERR_UNSUPPORTED_ESM_URL_SCHEME。

增强

• 解析失败告警并跳过（#4，PR #7）：parseSkillFile() 对不可解析 frontmatter 或缺失/空 name 输出带路径的告警并跳过；其余技能继续注册，坏文件的 get() 返回 undefined。
• Provider CI（#5，PR #8）：Ubuntu、Windows+CRLF、Linux+CRLF 三路自检；第一次运行就抓到真实的 Windows bug。
• 行尾策略：.gitattributes 改为 * text=auto eol=lf，Windows 检出保持 LF；CRLF 回归仍在 CI 中显式覆盖。

安装

npx @deepseek-ai/dsh plugin --profile web add github:SAXEM1997/specpowers#v1.1.0

装完重启 profile。已安装的 profile 可能仍钉在 v1.0.0，需要用 #v1.1.0 重装升级。

验证

• main CI 39572919588：三个 job 全绿。
• 合并后的 main 自检：ALL PASS。
• 匿名浅克隆 v1.1.0：同一自检 ALL PASS。

兼容性

无 breaking change。LF 检出行为不变；坏技能文件从“空名候选”变为“告警并跳过”。

specpowers v1.0.0 — SDD+TDD skill group for DeepSeek Harness & Claude Code

SDD + TDD engineering methodology skill group — one repository, one shared skills/ tree, installable as a DeepSeek Harness plugin and as a Claude Code plugin. It fuses OpenSpec (spec-driven development) with Superpowers (test-driven discipline) into a Phase 0→4 workflow.

Highlights

• Phase 0→4 workflow — requirements clarification → OpenSpec format conversion → implementation plan → subagent-driven TDD → verification & archiving. Every phase exits through a review Gate.
• DeepSeek Harness plugin — package.json + cordis.patch.yml + lib/index.js register the skills on the host skill registry. Zero runtime dependencies.
• Bare-name skill addressing — skills are referenced by bare name; the Claude Code plugin-namespace form is kept as a documented fallback in each skill's platform block.
• Multi-platform reference — skills/specpowers/refs/platform-tools.md maps tools and degradation paths for DSH and Claude Code; Codex is reserved.
• Two zero-dependency self-checks — scripts/verify-dsh-provider.mjs (packaging, provider contract, frontmatter parsing) and scripts/verify-no-internal-refs.mjs (leak guard).
• Bilingual README (中文 / English), MIT, upstream attribution in NOTICE.

Install

# DeepSeek Harness — pinned (recommended, reproducible)
npx @deepseek-ai/dsh plugin --profile web add github:SAXEM1997/specpowers#v1.0.0

# DeepSeek Harness — track main
npx @deepseek-ai/dsh plugin --profile web add github:SAXEM1997/specpowers

Then restart the profile and refresh the browser; verify with dsh --profile web --dump-config (a specpowers line must appear).

Claude Code:

/plugin marketplace add https://github.com/SAXEM1997/specpowers.git
/plugin install specpowers@specpowers-marketplace

Dependencies

• superpowers-dsh (DSH) or Superpowers (Claude Code) — 8 upstream skills are hard dependencies: brainstorming, writing-plans, subagent-driven-development, test-driven-development, systematic-debugging, requesting-code-review, verification-before-completion, finishing-a-development-branch.
• OpenSpec CLI — openspec init generates .claude/skills/openspec-* and the /opsx:* commands.

Known limitations

• The DSH install path was verified by executing the same provider code in-process plus byte-level manifest parity with a known-installable package — not by an end-to-end dsh plugin add run in CI.
• The eval suite (skills/*/evals/, 15 cases) is executed by skill-up; DSH is not one of its built-in engines (built-ins: claude_code, codex, qodercli, qwen_code).
• Codex support is reserved, not implemented — no Codex environment was available to verify it.

---

specpowers v1.0.0（中文）

SDD + TDD 工程化开发方法论技能组 —— 同一仓库、同一份 skills/ 真源，既可装为 DeepSeek Harness 插件，也可装为 Claude Code 插件。融合 OpenSpec（规范驱动开发）与 Superpowers（测试驱动纪律），形成 Phase 0→4 完整工作流。

亮点

• Phase 0→4 工作流 —— 需求澄清 → OpenSpec 格式转换 → 实施计划 → subagent-driven TDD → 验证归档；每个阶段出口都经审查 Gate。
• DeepSeek Harness 插件 —— package.json + cordis.patch.yml + lib/index.js 把技能注册进 host 技能注册表，零运行时依赖。
• 技能调用统一裸名 —— 调用一律用裸名；Claude Code 的插件命名空间形式作为回退记录在每个技能头部的「平台适配」块。
• 三平台对照 —— refs/platform-tools.md 给出 DSH 与 Claude Code 的工具映射与降级路径；Codex 为预留。
• 两个零依赖自检 —— verify-dsh-provider.mjs（打包、provider 契约、frontmatter 解析）与 verify-no-internal-refs.mjs（泄漏闸门）。
• 中英双语 README，MIT 许可，上游归属见 NOTICE。

安装

# 钉版本（推荐，可复现）
npx @deepseek-ai/dsh plugin --profile web add github:SAXEM1997/specpowers#v1.0.0

# 跟随 main
npx @deepseek-ai/dsh plugin --profile web add github:SAXEM1997/specpowers

依赖

• superpowers-dsh（DSH）或 Superpowers（Claude Code）—— 8 个上游技能为硬依赖：brainstorming、writing-plans、subagent-driven-development、test-driven-development、systematic-debugging、requesting-code-review、verification-before-completion、finishing-a-development-branch。
• OpenSpec CLI —— openspec init 生成 .claude/skills/openspec-* 与 /opsx:* 命令。

已知限制

• DSH 安装路径通过同一份 provider 代码的进程内执行 + 与已知可装包的清单逐字节一致性验证，未在 CI 中做端到端 dsh plugin add。
• 评测套件（skills/*/evals/，15 个用例）由 skill-up 执行；DSH 不是其内置引擎（内置为 claude_code/codex/qodercli/qwen_code）。
• Codex 支持为预留，未实现 —— 无 Codex 环境可验证。