v6.7.0

v6.7.0 (2026-10-14)

Changes merged to main after v6.6.0, other than this version bump. The product commits are only PR #60 (256c535, merge 7ec3304).

Skills

• Record questions and replies as GitHub issues. New skill recording-qa-to-issues (skills/recording-qa-to-issues/SKILL.md, Codex metadata in agents/openai.yaml with allow_implicit_invocation: true). When the human partner asks a question, reports a problem, requests a change, or follows up in Cursor, Claude Code, Codex, Kimi, OpenCode, or Pi — including 提问, 咨询, bug reports, feature requests, and UI or docs discussions — the session facing that person records the turn on one GitHub issue in the current repo. One trackable question is one issue; a follow-up is a comment on that issue; a greeting with nothing to track, a delegated subagent, an explicit request not to record, or a missing repo / unavailable gh is not recorded (and a missing repo is not reported as recorded). The title is <type>(<scope>): <description>, or <type>: <description> when there is no scope. The body is 来源 (the tool name), 提问 (original text, with secret values replaced by [REDACTED]), and 回复要点 (conclusions, not the full chain of thought). One primary type sets the title type; other types are extra labels. The skill's table is bug/fix, feature/feat, enhancement/feat, documentation/docs, ui/fix or feat, refactor/refactor, performance/perf, test/test, security/fix, chore/chore, question/question, and discussion/discussion. A missing label is created with gh label create and no custom color. Search open and closed issues first; the same topic is a comment, not a second issue. The skill does not close the issue and does not touch a Project board. (#60)

Repository

• Contributor guide points at the skill instead of copying its type table. CLAUDE.md (canonical; AGENTS.md is a symlink) tells agents to record the human partner's questions and replies with skills/recording-qa-to-issues/SKILL.md, and to take type, labels, and title from that skill. The rest of the issue section stays this repository's contributor rules. (#60)
• READMEs list the new skill. README.md, README.zh-CN.md, README.zh-TW.md, README.ja.md, and README.ko.md now count 27 top-level skills (17 general development skills and 10 Product Design skills) and name recording-qa-to-issues under Collaboration. (#60)

Full Changelog: v6.6.0...v6.7.0

v6.6.0

v6.6.0 (2026-10-14)

Changes merged to main after v6.5.0, other than this version bump. The product commits are only PR #55 (1b3542e, 840317c, merge 3b81a61).

Skills

• Grilling HANDOFF writes a derived technical spec. After the human partner confirms the Shared-Understanding Record, skills/grilling/SKILL.md writes a separate technical specification and waits for review of that file. The record stays the interview product. Confirmation does not make the record the spec, and it does not authorize implementation or writing-plans. The spec header is the work title (not "Shared-Understanding Record"), Status: Confirmed, Date, Source: grilling. The body is six sections filled only from confirmed content: Problem and objective; Architecture and component boundaries; Interfaces, data, and behavior; Constraints, compatibility, and non-goals; Errors, operations, and other confirmed concerns that affect implementation; Acceptance criteria and testing. The eight Completion Gate headings are not the outline, and the record is not pasted as the file. Closed branches become non-goals or constraints; authorized defaults become specified defaults; non-blocking open items stay labeled open. Inline self-review only (no spec-reviewer subagent). Commit only that spec file, then stop for spec-file review. writing-plans loads only after explicit approval of the written file (按这个写计划 and the other go-aheads already listed). Early stop still emits a partial record and does not write a spec. The default path is unchanged: docs/wukong-code/specs/YYYY-MM-DD-HHmm-<topic>-design.md, overridable by a spec-location preference. tests/skills/grilling-scenarios.md M4–M6 and U1 now require that derived file. Cursor subagent probes (not a Drill or tmux cohort) are in docs/wukong-code/evals/2026-10-13-grilling-derived-spec-handoff.md: RED wrote the eight-section record as the spec; GREEN wrote the derived outline, asked for spec-file review, and loaded writing-plans only after 按这个写计划. (#55)

Repository

• Contributor rules no longer override the grilling HANDOFF. CLAUDE.md (canonical; AGENTS.md is a symlink) now says issue and GitHub Project tracking is this repository's habit, not a grilling skill capability. Plugin users follow skills/grilling/SKILL.md. The section no longer forbids a grilling spec file and no longer tells agents on this repo to ignore the skill's write-and-commit step. Consensus is still recorded on a GitHub issue. If docs/prd.md already exists, it stays the product snapshot and is not replaced by a spec. (#55)

Full Changelog: v6.5.0...v6.6.0

v6.5.0

v6.5.0 (2026-10-13)

Changes merged to main after v6.4.0.

Language Guidance

• Codex router understands Chinese prompts. hooks/language_router.py adds Chinese cues for every phase in the same precedence order as using-wukong-code (debugging → review → verification → testing → profile → implementation), so a Chinese prompt such as 「帮我修复 handler.ts 里的 bug」 routes to the TypeScript pack instead of abstaining. Word boundaries use explicit ASCII lookarounds because Python's \b treats CJK as word characters; gaps between cue words are clause-bounded (，。；,;) and ambiguous nouns (验证, 审核, 加载, 编辑器, 支持) only match in phrase shapes, so 「验证码」 or 「编辑器」 do not select a phase. English routing is unchanged: a 53-prompt differential against the previous router shows zero differences. 27 Chinese cases and 8 false-positive controls are pinned in tests/hooks/test-language-router.sh and tests/hooks/language-router-contract.py. (#41)
• The prompt hook leaves no hooks/__pycache__ behind. hooks/user-prompt-submit runs python3 -B, so the first Codex prompt no longer dirties a git-cloned plugin checkout with an untracked, un-ignored hooks/__pycache__/ that sync-to-codex-plugin.sh would otherwise rsync into the Codex fork; the router test copies the plugin root and asserts the directory stays clean. (#44)

Repository & Packaging

• Rust eval session captures moved to git history. The 48 raw JSONL captures under docs/wukong-code/evals/raw/2026-07-29-rust-language-guidance/candidate/jsonl/ are removed from the working tree; candidate.md and INTEGRITY.md keep the SHA-256 manifest and show how to retrieve each file with git show v6.4.0:<path>. Tracked size drops from 13.54 MB to 5.40 MB (shallow clone 17 MB → 8.2 MB). tests/skills/test-language-guidance.sh fails if a .jsonl or any file over 256 KiB is added back under docs/wukong-code/evals/raw/. (#42)
• .gitignore anchors evals/ to the repo root. The unanchored rule also matched docs/wukong-code/evals/, so new eval reports there were silently ignored by git add; the rule is now /evals/ and the language-guidance test checks both sides. (#43)
• Maintainer identity and provenance metadata unified. Manifests, hooks, and docs name the current maintainer; FUNDING.yml is removed; LICENSE keeps the 2025 Jesse Vincent line and adds a 2026 line; THIRD_PARTY_NOTICES.md gains a Superpowers section; upstream issue references are qualified as obra/superpowers#…. (#40)

Testing & CI

• Host-tool preflight. scripts/test.sh checks for every host tool the suites shell out to (git node python3 rg rsync jq zip unzip tar gzip shasum, plus npm for extended) and exits 3 with the full missing list before running anything, so a missing tool shows up once by name instead of as unrelated assertion failures. docs/testing.md documents the table. (#38)
• ShellCheck is a CI gate. A separate lint job in .github/workflows/test.yml runs bash scripts/lint-shell.sh --all on every pull request and push; the 11 warning lines that had accumulated in tests/claude-code/ (including a $?-after-echo that always reported exit code 0) are cleared. ubuntu-latest ships shellcheck, so the job adds no dependency. (#45)
• evals-static re-pinned to a reachable harness commit. The weekly manifest-validation workflow clones the harness branch explicitly, hoists repo/branch/commit into env, and fails with a ::error:: naming the unreachable pin; tests/test-automation/test-test-runner.sh fails if docs/evals-setup.md and the workflow disagree on the pin. (#37)

Documentation

• Issue and GitHub Projects collaboration rules. CLAUDE.md (canonical; AGENTS.md is a symlink) now requires agents working on this repository to search then create or update GitHub issues for 需求、Bug、and 咨询/一般讨论, and to record grilling consensus on the issue instead of writing a new spec under docs/wukong-code/specs/. The English PR bar, identity disclosure, and skill content are unchanged. (#50, tracking #49)
• Simplified Chinese product-requirements snapshot. docs/prd.md states current purpose, audience, in/out of scope, capabilities, and how success is judged. It is a snapshot of what the repo already ships, not a roadmap. (#51)

Full Changelog: v6.4.0...v6.5.0

v6.3.0

What's Changed

• feat: add experimental Go language guidance by @wukongnotnull in #11
• Release: promote dev to main (experimental Go language guidance) by @wukongnotnull in #12
• feat: add programming grilling skill by @wukongnotnull in #13
• Dev by @wukongnotnull in #14
• feat: add experimental Swift language guidance by @wukongnotnull in #15
• Release: promote dev to main (experimental Swift language guidance) by @wukongnotnull in #16
• Minute-precision timestamps for plans/specs filenames by @wukongnotnull in #17
• Release: promote dev to main (minute-precision plans/specs filenames) by @wukongnotnull in #18
• feat: add Rust language-guidance pack by @wukongnotnull in #19
• Release: promote dev to main (experimental Rust language guidance) by @wukongnotnull in #20
• feat: add experimental Java language guidance by @wukongnotnull in #21
• Release: promote dev to main (experimental Java language guidance) by @wukongnotnull in #22

Full Changelog: v6.2.1...v6.3.0

v6.2.1

v6.2.1 (2026-07-19)

Lower Resident Skill Cost

Mechanical slim of high-frequency skills (gates stay resident; depth on demand):

• writing-skills (~689 → ~243)
• subagent-driven-development (~345 → ~277)
• test-driven-development (~261 → ~175)
• receiving-code-review (~213 → ~151)

Gate test: tests/skills/test-skill-slim-gates.sh

Full notes: see RELEASE-NOTES.md

v6.2.0

v6.2.0 (2026-07-19)

Lower Ritual, Same Gates

• Primary workflow + progress budget in using-wukong-code
• Condensed brainstorming (Full path + HARD-GATE unchanged for creative work)
• Same-task verification reuse
• Intent-first finishing
• Focused systematic debugging

OpenCode / Bootstrap

• Dedup scans all messages for EXTREMELY_IMPORTANT
• Tests for marker-elsewhere and stripped frontmatter

Full notes: see RELEASE-NOTES.md in the repo.