Skip to content

Releases

Adaptive Agentic SDD 1.1.0

Choose a tag to compare

@seok-jun seok-jun released this 07 Nov 11:18
4f96841

Small tasks keep Medium's safeguards with Trivial's paperwork: three records instead of seven.

This release includes PR #16 and PR #17.

git clone https://github.com/seok-jun/adaptive-agentic-sdd
python3 adaptive-agentic-sdd/try/setup.py        # Windows: python; --lang ko for Korean
  • Lighter Small route.
    • small-work.md names a fixed reading set; other common documents open only on listed exceptions.
    • A normal Trivial/Small task writes CLAIM, IMPLEMENTATION SUMMARY and CLOSED. CLOSED now holds merge, closure, product completion, cleanup and release as fields.
    • Grading, owning-module tests first, required checks, claims, merge authority and human-decision stops are unchanged.
    • Six bookkeeping rules now apply to Medium and above only.
  • Workflow docs match external adoption: project-owned W first, legacy install.py copies after it. Git 2.38+ is recommended for the local-docs merge check.
  • First-use fixes from newcomer runs:
    • A Windows command in the README, and P/W/M defined where they first appear.
    • START-HERE asks to allow local Git and claim-file deletion, and prints the resolver command for PowerShell and Bash.

Before/after (three Codex runs per version, one planted Small task): all six runs completed through merge and claim release.

  • Records fell from 7 to 3.
  • Command output returned to the agent fell by a third.
  • Input tokens fell 8%.
  • Wall time showed no difference.

Known limit: the fixed reading set is still about 23k words. Command output fell by a third without a change in wall time, so these runs do not show reading volume as the main driver of time or cost.

Existing project workflows are not updated automatically. Compare your W's small-work.md and tracker adapter with this tag, and merge what fits.

Validation: all 123 executable tests passed on Windows. Local links and all five manifests check clean. Two independent Codex reviews found eight issues, and all were fixed. Linux and macOS were not rerun for this release. The effect on outcomes, rework or cost in real projects remains unmeasured.

한국어: Small 작업의 안전장치는 Medium과 같고, 기록은 Trivial처럼 가벼워졌습니다. 정상 경로의 기록이 7개에서 3개로 줄었습니다. 같은 과제로 버전마다 Codex를 3회씩 실행했더니 Agent가 받은 명령 출력은 3분의 1 줄었고, 소요 시간은 차이가 없었습니다. 기존 프로젝트 워크플로는 자동으로 바뀌지 않으니, 필요하면 small-work.md와 tracker 어댑터를 비교해 반영하세요.

Full release record · Before/after runs and limits · Changes since v1.0.2

Adaptive Agentic SDD 1.0.2

Choose a tag to compare

@seok-jun seok-jun released this 07 Nov 04:34
6afd3f1

Try it in a minute: a one-command sandbox, a clearer README, and a demo from real runs.

This patch includes PR #11, #12, #13 and #14.

git clone https://github.com/seok-jun/adaptive-agentic-sdd
python3 adaptive-agentic-sdd/try/setup.py        # --lang ko for Korean
  • try/setup.py builds a disposable, offline sandbox: a small product with three planted tasks (Trivial, Small, Large), a project-owned workflow filled for it, local remotes and records. It confirms READY and writes nothing outside its folder.
  • README (EN/KO) leads with a one-line summary, a risk-grading diagram, the quick start and an animated demo shortened from two real Codex runs.
  • Eval runners keep Claude Code session records so headless runs can be counted locally.
  • Issue forms for try / first-use feedback and bugs, and a code of conduct.

The distribution is unchanged from v1.0.1; existing project workflows need no update.

Validation: 123 executable tests passed on Windows; both editions of the sandbox reached READY and all tests passed on macOS 12.7.6. In the sandbox, Codex completed the Small task end to end and stopped the Large task for human decisions without changing product code; Claude Code completed the Small task on macOS (its own report). One run per task. The Large stop follows a rule the sample product states. Linux and host sandbox enforcement were not tested; effect on outcomes, rework or cost remains unmeasured.

한국어: 명령 하나로 버려도 되는 체험 환경을 만들어 1분 안에 써 볼 수 있게 했습니다. README 첫 화면을 다시 쓰고 실제 실행으로 만든 데모를 넣었습니다. 배포본은 1.0.1과 같아서 기존 프로젝트는 할 일이 없습니다. 써 보다 막힌 곳은 이슈의 "Try / first-use feedback" 양식으로 알려 주세요(한국어 가능).

Full release record · Try-sandbox runs and limits · Changes since v1.0.1

Adaptive Agentic SDD 1.0.1

Choose a tag to compare

@seok-jun seok-jun released this 06 Nov 23:48

Small-task workflow improvements and explicit temporary-output boundaries.

This patch includes PR #10.

  • Adds English/Korean compact Trivial/Small routes, reuses unchanged inputs, and avoids repeating the full SUMMARY at closure.
  • Treats helper scripts, caches and compiler outputs as scoped writes; uses authorized task scratch and documents real host sandbox checks.
  • Preserves required verification, escalation, merge authority and claim ownership checks.
  • Adds three synthetic policy cases per edition (89 total per edition).

Validation: 117 executable tests passed on the final Windows release tree. Both 70-file bundle manifests and all three M manifests passed exact serialization/hash checks; publication-document local links passed. Five Python/JavaScript/TypeScript/Java/Go cases using requested Codex gpt-6.1-sol/high passed independent outcome checks after two preserved environment-blocked attempts were corrected. Completed runs showed no observed out-of-area agent writes and one SUMMARY each.

This is not a controlled speed comparison or a guarantee of agent compliance. Completed tasks still took about 4.5–5.7 minutes. The methodology does not install an OS sandbox; host enforcement must be configured separately.

한국어: 작은 작업의 중복 준비·기록을 줄이고 임시 파일·캐시·빌드 출력의 작업 경계를 명확히 했습니다. 필수 검증과 병합·claim 소유권 확인은 유지합니다. 기존 프로젝트의 W는 자동 갱신되지 않으므로 변경을 검토해 적용한 뒤 새 W commit을 선택하세요.

English guide · 한국어 적용 가이드

Full release record · Trial evidence and limitations · Changes since v1.0.0

Adaptive Agentic SDD 1.0.0

Choose a tag to compare

@seok-jun seok-jun released this 06 Nov 10:49

Short for small changes, deep for risky ones.

Adaptive Agentic SDD 1.0 is a risk-adaptive development workflow for Claude Code or Codex on existing codebases. It scales planning, review and verification across four change grades: Trivial, Small, Medium and Large.

Start here

Included

  • Project-owned, editable external workflows, explicitly selected per task. Continued use of the upstream methodology is optional.
  • A read-only task resolver and agent-assisted adoption procedure.
  • Four workflow skills for work-item management, parallel development, implementation and completion.
  • GitHub and local-document tracking adapters, English and Korean editions, and MIT licensing.
  • A principle for choosing improvements by cause, expected benefit and cost: reuse existing measures, use mechanical checks for decidable conditions, and guidance/review for judgments that remain.

The legacy installer remains available for explicit legacy use. Existing projects are not automatically migrated.

Verification and limits

The pre-publication review covered 223 files. Claude and independent GPT-6.1 Sol / xhigh reviewers completed the review and correction checks; no unresolved P1/P2 remained in scope. The final Windows test run passed all 117 tests. Earlier final-candidate validation passed eight legacy installation combinations (English/Korean × GitHub/local-docs × in-repository/external).

These are tool and review checks, not evidence of improved productivity, lower cost or agent compliance. This project provides guidance, not an agent runner or security boundary. Historical behavior evaluations include failures and have documented limitations. No new native macOS/Linux or live GitHub tracker lifecycle is claimed for 1.0.

Full release record · File review ledger · Changelog

한국어

작은 변경은 짧게, 위험한 변경은 깊게 검토하는 v1.0.0 정식 릴리스입니다. 프로젝트가 자체 워크플로를 소유하고 수정하며, 작업마다 명시적으로 선택해서 사용합니다. 도입은 에이전트 지원 절차이며 기존 프로젝트를 자동으로 이전하지 않습니다. 검증 결과와 한계는 위 릴리스 기록을 참고하세요.

1.0 preview.7

1.0 preview.7 Pre-release
Pre-release

Choose a tag to compare

@seok-jun seok-jun released this 31 Oct 14:45
51f62c4

Release commit: 51f62c49a524fd6fa8056ec55790388a56423e91 (merge of PR #7; tree identical to 8141d33a13e77f9b01288f03dd884c4524722d7d, which is the reviewed candidate 0a65db904cfe71f2e8f96a0b1b3fa7f7d181bb61 plus the review fix and validation lines). Follows preview.6 (9e69616). Earlier tags are unchanged.

Short for small changes, deep for risky ones. Start with the README: paste the install prompt into Claude Code or Codex.

What changed

  • Model routing, c′ policy relaxation (rule meaning changed, owner-approved): routine Trivial/Small work on the current or manually chosen model no longer needs positive capability evidence unless the model is denied. Other current/manual routes need fresh host evidence or a human-confirmed scoped assessment citing observed results. Automatic switching, delegation and fallback still need PASS.
  • Portability (no rule change): generated in-repository commands and external POSIX checks use python3; external PowerShell checks keep python. Both editions say to retry with another Python 3.8+ interpreter before treating a check as failed, and both guides add POSIX copy commands.
  • Evaluation tooling and records (no rule change): behavior runner v2 removes answer keys from subject bundles; grader calibration, first-use and c′ records are published. Earlier records now say that their answer-key files were readable.
  • Release docs: the README opening describes the methodology without comparisons to other tools; the Codex sandbox note now covers the first task's worktree, commits and merge.

Evidence and limits

  • 89 tests passed on Windows (Python 3.12.2). On macOS, a rerun of the python3 fix passed all 89 tests without a python command (report). Both 66-file manifests matched, and an archive of the reviewed candidate passed init/render/check/check-evals (83 cases) for both editions × both trackers.
  • Independent read-only review (Codex gpt-6-astra, high): no blocker. The one P2 was fixed; two optional P3 wording suggestions were not applied.
  • A Codex first-use run with an empty capability table completed a Small task end to end. The c′ behavior check (5 PASS / 9 FAIL / 2 INVALID) is weak evidence: Codex subjects could not read files.
  • The macOS walkthrough was assisted, not a blind first use. No effect on outcomes, rework or cost has been measured. Live GitHub tracker operations were not run.

Full details: CHANGELOG.

1.0 preview.6

1.0 preview.6 Pre-release
Pre-release

Choose a tag to compare

@seok-jun seok-jun released this 31 Oct 05:59
9e69616

Release commit: 9e6961678a33c5e5da0e7b3e6c48ed96f6994110 (merge of PR #5; tree identical to dcf00418c058c4a20d679dca0320d7aa46da2c9c, which is the reviewed candidate 277cfc58a5523f4bfe02c06fe67864d852ae897e plus validation lines). Follows preview.5 (ee5c60a). Earlier tags are unchanged.

Short for small changes, deep for risky ones. Start with the README: paste the install prompt into Claude Code or Codex.

What changed

  • First experience (no rule change): both READMEs now show how depth scales by grade, two illustrative transcripts, a paste-in install prompt, the expected check/check-evals output, the added files, how to undo, and when the first task counts as done.
  • Tracker-neutral procedures: W defines Issue, comment, field values, assignee, PR and closing reference as tracker-neutral terms. GitHub-only mechanics moved to tracker/github.md; tracker/local-docs.md got the matching rows. Meaning for both existing adapters is unchanged; no new tracker is supported.
  • Skill and evaluation clarity: the II-S02 step applies the owning-module-first check to each verification-affecting changed file, matching verification.md. FI-S02 now supplies its failing command; new FI-S02b covers missing evidence.

Evidence and limits

  • 65 tests passed (Windows, Python 3.12.2); both 66-file manifests matched; an archive of the reviewed candidate passed init/render/check/check-evals (77 cases) for both editions × both trackers.
  • Independent read-only review (Codex gpt-6-astra, high): no blocker; all findings fixed and reconfirmed.
  • First-use runs from the README: a Claude Code agent installed, finished a Small task and undid the installation; a Codex run installed, but its sandbox blocked Git writes.
  • Behavior evidence is weak and published as graded: a paired diagnostic found no overall improvement, and a bounded tracker-neutral check scored 4 PASS / 6 FAIL with no failure traced to the change. No effect on outcomes, rework or cost has been measured. POSIX execution and live GitHub tracker operations were not run.

Full details: CHANGELOG.

1.0 preview.5

1.0 preview.5 Pre-release
Pre-release

Choose a tag to compare

@seok-jun seok-jun released this 30 Oct 08:20

Release commit: ee5c60a246a95218e0f429681036fa9bd4bd1e03

Packages the candidate from 110d943c6db638d72bbafbbdc2bad4037e66b45c on v1-preview, following preview.4 (5760ac66b2a4c943360a56c22892fe35e1a076db). This packaging step changes release documentation and records the evaluation limitations; both bundles and earlier tags are unchanged. methodology/ (English M) is unchanged.

  • Independent translation review (2026-10-30): gpt-6-astra (reasoning medium) compared the entire English edition against the Korean edition in five groups. It reported 14 findings (2 P1, 10 P2, 2 P3): 11 accepted, 2 partially accepted and 1 rejected. Partial acceptance: the README upgrade sentence was corrected, but the finding was treated as P2 because the sentence already existed in preview.4; the Methodology-only copy section was retained as existing content carried over from the preview.4 English guide, with its provenance stated and a pointer added in the Korean guide. The proposed cross-agent → cross-tool change was rejected to follow the source term Cross-agent behavior eval and the preview.4 English README. The finishing-issues cleanup exception was clarified in both editions; its ambiguity originated in the Korean plain-language rewording.

Execution profile (meaning change, opt-in; defaults keep preview.4 behavior):

  • review.large_mode: dual | single — single replaces the Large AS-IS/PLAN Review A + Blind Audit B with one independent review in a separate session that also covers Review B's questions, uses the same four severities and keeps the blocker-0 progress condition. Blind isolation, the human's B selection and reconciliation do not apply in single.
  • review.blind_reviewer_selection: required | optional — with optional, when no human has chosen, an agent may choose the Review B reviewer under the automatic-candidate rules (fresh role qualification PASS, strong + high floor, separate from the author and Review A, no frontier without escalation evidence) and records that it did. A human's choice is still never replaced automatically.
  • sdd.medium_artifacts: full | condensed — condensed writes a new Medium SDD as one README.md with ## AS-IS, ## PLAN, ## Implementation specification; the required content of each part is unchanged, and re-grading to Large splits it into the four documents before the AS-IS review.
  • None of the values reduces M's human-judgment requirements, automatic grade escalation, required verification, the CODE Review or human judgment on Large merges. Changes apply to newly started reviews and SDDs. Owners: design-review.md (new section "Large design review mode"), sdd-artifacts.md (new section "Medium condensed layout"), execution-routing.md; project template sdd-workflow.md gains an "Execution profile" table.
  • The installer rejects values outside the choices, non-string values and unknown keys in review and sdd (a misspelled key would otherwise silently keep the default), before writing anything.
  • design-review.md is a hash-pinned guard in the structure tests. The change is additive only; the test now proves that removing exactly the added section, index row and sentence reproduces the earlier pinned bytes, instead of re-pinning the file.

Readability (Korean edition):

  • New workflow/glossary.md (both editions), linked from the workflow README and each skill. It explains terms and adds no rules.
  • The four SKILL.md bodies were reworded in plain Korean. Headings, description, literal markers, stop strings, paths and the order and conditions of every step are unchanged; skill headings referenced by other documents keep their anchors.
  • One wording fix in issue-writing-guide.md (a pointer said "below" for a section above).

English edition:

  • New full bundle adaptive-sdd/ at the repository root with the same file set as ko/adaptive-sdd/; its methodology/ is byte-identical to the root methodology/. Workflow, templates, eval cases and both guides (docs/customization.md now covers the full installation, with a methodology-only section; docs/sdd-reading-audit.md) are translated. The English text is canonical.
  • The translation was produced by the coordinating agent and translation subagents with a fixed heading map and shared term list, then checked mechanically (no Korean text left, same heading structure, same template-token sequence, same case IDs and item counts, all installed links and anchors resolve). It then had one independent agent translation review (recorded above); it has not had an independent human review.
  • The installer is byte-identical in both bundles. Edition wording (skill entry descriptions, marker note) moved to each bundle's templates/strings.json, which the installer validates. Marked blocks written by either edition are recognized, so switching editions re-renders instead of failing; look-alike or duplicated markers are still rejected. Korean generated files are byte-identical to preview.4 except sdd-workflow.md.
  • tools/review_snapshot.py accepts the English tracker headings Meta / Records as well as the Korean ones; a document with headings of both editions is rejected.

Evaluation:

  • First executed cross-agent behavior run of 12 pre-selected Korean cases × Claude Code and Codex × 3 runs, cross-graded; runner tests/run_behavior_evals.py; record in docs/eval-results/2026-10-30/. 71 valid runs: Claude Code 21 PASS / 14 FAIL / 1 INVALID, Codex 22 PASS / 14 FAIL. A usage limit invalidated 65 first-attempt runs; each was retried once (31 reruns, 34 regrades of saved answers) as skill-evals.md allows. This is a baseline, not evidence of effect.
  • It found one case defect: PD-S06 required BLOCKED: origin/main fetch failed, while the skill says BLOCKED: declared base fetch failed; all six FAILs of that case came from the stale case. The recorded results are unchanged; the case is corrected in both editions.
  • Known limitations: II-S02 missed the explicit first-targeted-verification requirement in all three Claude runs. FI-S02 requires naming a failing command but supplies neither its name nor a failure log; this is an evaluation-input limitation, not established evidence of a skill defect. Raw answers are unavailable for a grading audit. Skill changes and case repair are deferred to a later version; existing results remain unchanged. See the source comparison in the evaluation record.
  • Six execution-profile policy-regression cases (II-K01–K06 and their English counterparts) were added; they are expected-behavior specifications and were not executed.

Release-commit validation (2026-10-30, Windows, Python 3.14.7): all 53 executable tests passed, both bundle manifests (66 files each) matched, and both editions passed init, render, check and check-evals (76 cases) in disposable projects. The initial sandbox run hit Windows temporary-directory locks (WinError 32) and a Git/MSYS access denial; the full suite was rerun outside the sandbox. Actual POSIX execution, live remote/host integration and an independent review of this change remain undone.

Bundle MANIFEST.json SHA-256:

  • English: ff001a1b8f07650687836e18f348a440af3a9c3e483f2393fe9eff6c81bdce26
  • Korean: b56018812a1b8e4e9c90fbff5c4058336b50c3b5d6019dc0157796bf5ecb5058

1.0 preview.4

1.0 preview.4 Pre-release
Pre-release

Choose a tag to compare

@seok-jun seok-jun released this 30 Oct 04:32

Adaptive Agentic SDD 1.0 preview.4

Includes the M/W/P correction discussed after preview.3. M owns common principles, W owns execution procedures, and P owns product facts, settings and authority. External installation is solely the optional --harness-root storage location; S is no longer a required workflow or Skill role.

  • Generated AGENTS, skill entry files and project documents declare actual installation paths. Fixed workflow references retain M/W/P ownership.
  • Existing CLI options, connection schema, checkout binding, integrity checks and hand-edit protection remain supported.
  • Existing consumers adopt the new pinned bundle explicitly, then render and check. The source tag and commit must both match preview.4; prior preview tags remain unchanged.

Installation

Validation

The correction payload at 6d37203 passed 36 Windows tests and an independent read-only review, with no actionable findings. Synthetic upgrades from preview.3 passed for both in-repository and external installations, preserving project settings and replacing the source pin. This release changes only documentation relative to that payload.

A ZIP built from the exact preview.4 tag passed init/render/check/check-evals for in-repository and external installation, with matching tag/commit pins and M/W/P entry paths. Both methodology manifests and the 64-file bundle manifest matched. The 70 cases across four skills are structural checks, not executed model behavior evaluations.

Actual POSIX execution, real consumer deployment, model behavior and efficacy remain unverified. The full workflow/templates/tools are still available only in Korean; the English methodology edition remains supported.

한국어

이번 .4에는 M·W·P/S 보정이 포함되어 있습니다. M·W·P 소유 구조를 유지하고 외부 설치는 경로 옵션으로 정리했습니다. 기존 설치는 새 사본으로 명시적으로 갱신한 뒤 render/check를 실행하세요. .3 및 이전 태그는 그대로 보존합니다.

Release commit: 5760ac66b2a4c943360a56c22892fe35e1a076db

1.0 preview.3

1.0 preview.3 Pre-release
Pre-release

Choose a tag to compare

@seok-jun seok-jun released this 30 Oct 02:43

Adaptive Agentic SDD 1.0 preview.3

This pre-release packages the changes since preview.2 on v1-preview. The v0.2 line remains on main; previous preview tags are unchanged.

Changes

  • Korean workflow: Epic is an optional management grouping, separate from the four execution grades. Shared-change verification, human decisions and overall completion evidence remain required.
  • Splits SDD rules by phase and separates human-readable AS-IS views from detailed evidence, with optional templates and migration guidance.
  • Adds external harness installation, source-commit/tag upgrade safeguards, operating and review settings, and exact-commit review snapshot tooling.
  • Fixes GitHub tracker ID guidance. Clarifies reuse of valid evidence for unchanged candidates and verification inputs while retaining required checks.
  • Aligns the English and Korean installation guides with preview.3 and its verified commit, and corrects references to relocated policy documents.

The English edition still provides the common methodology only. The full workflow, skills, templates and tools are currently available in Korean; their English edition remains pending.

Installation

Use the archive and guide from this tag and record the verified full commit SHA with the manifest hash. Do not use the preview.2 guide for the new installer features.

Validation and limits

  • The integration payload at 01ca9a6 passed all 35 executable tests on Windows with Python 3.14.7. The release commit changes only five release documentation files; the validated methodology, installer and workflow bytes are unchanged.
  • A ZIP built from the exact preview.3 tag passed both methodology manifests and the 64-file bundle manifest checks. Disposable local-docs and GitHub-adapter installations passed init, render, check and check-evals, including the release tag/commit pin. The 70 cases across four skills were checked structurally, not run as model behavior evaluations.
  • Independent reviews and reconfirmations have the limited scopes recorded in the changelog. They are not a new independent review of the entire cumulative workflow.
  • Actual POSIX execution, live tracker/host integration, model behavior, and effects on outcomes, rework or cost remain unverified.

한국어 요약

preview.2 이후의 Epic 관리 분리, SDD 단계별 읽기, 외부 설치 및 tracker 수정을 preview.3으로 묶었습니다. 설치 시 위 한국어 안내와 같은 태그의 사본을 사용하고 전체 commit SHA·manifest 해시를 고정하세요. 실행 절차·Skill·설치 도구의 영문판과 모델 행동·효과 검증은 아직 완료되지 않았습니다.

Release commit: bd3dd97e0a91b63ba14e6a44c0fba5ba3534f131

Korean bundle manifest SHA-256: 49ac6d8781fef753fac91b9d9956cfe2856379bb92a5078e402fc083337f19ab

1.0 preview.2

1.0 preview.2 Pre-release
Pre-release

Choose a tag to compare

@seok-jun seok-jun released this 16 Oct 19:26

Adds executable procedures, skills, an eval structure, and an installer to the Korean edition, and fixes three human-judgment gaps in the common methodology for both editions.

  • methodology/core.md (both editions): human-judgment requirements cannot be waived by host or project rules; lists the minimum questions that need human judgment; a delegation counts as a human decision only when a human records it with its scope, date, and confirmer, and the Agent never creates, extends, or finalizes one; when a human cannot be asked, the dependent work stops and the question is recorded as unresolved; the adapter gains a slot for the work-record reference location and how to find earlier records.
  • methodology/adaptation.md (both editions): adds awaiting human confirmation to the re-evaluation dispositions.
  • methodology/supports/ordered-changes.md (both editions): progress that depends on human judgment stops and asks a human under core.md.
  • docs/customization.md: adds the new adapter slot and the delegation rule, and points to preview.2.
  • Korean edition: ko/methodology/ is replaced by the full pinned copy ko/adaptive-sdd/:
    • workflow/ — workflow procedures and four skills (organizing and creating work items, isolating work, design and implementation, verification and closing), extracted from procedures used in one product with product-specific parts moved out; tracker mappings for local documents and GitHub.
    • workflow/process/skill-evals.md and workflow/skills/*/evals/ — eval kinds, grading, result records, 30 fictional seed cases, turning recorded incidents into diagnostic cases, and validity conditions for comparing and adopting a change.
    • templates/ and tools/install.py — generate project documents, skill entry files for Claude Code and Codex, and marked blocks in AGENTS.md, CLAUDE.md, and .gitattributes from project.json; check the copy and generated files against their hashes.
    • MANIFEST.json — sha256 of every file in the copy.
  • Known limitations: the local-documents tracker mapping replaces GitHub issues and pull requests only; git still needs a code remote named origin. Squash and rebase merges are confirmed from a recorded merge entry rather than from git ancestry. The guidance for re-evaluating observations is condensed until its skill ships, and no deterministic routing validator is included. Some workflow references point to project-document sections that the generated documents do not carry yet.
  • Pending: the English edition of workflow/, templates/, tools/, and the installer guide. Until then the Korean edition is the only source for them. A skill for re-evaluating observations is also deferred.

Checked: manifest hashes, the installer's unit tests, one install rehearsal on Windows into a fictional project using a locally built archive, and one independent review whose should-fix findings were addressed before publication. Effect on outcomes has not been measured.