오데덕 (odeduck) v0.21.2

Fixed

• 활용신청 확인 후 폼·대상 데이터·상세기능·활용목적·분류·동의·제출 경로가 바뀌면 제출을 차단합니다.
  확인 상태는 브라우저 안에 보관하고, 재검증과 1회 제출을 같은 실행에서 처리합니다.
• 제출 후 고정 4초 대기를 제거하고 포털 완료·검증 오류 응답을 최대 15초 기다립니다.
  응답이 없으면 신청내역을 확인하고, 결과가 불확실하면 자동 재제출하지 않고 확인 필요 오류를 반환합니다.

Validation and limits

• 로컬 Chrome의 기존 REST·ODCloud 폼, 확인 후 필드·DOM·URL 변경, 1회 제출, 취소와 응답 대기를 검증했습니다.
• 실제 정부 계정에 새 활용신청을 제출하는 검증은 수행하지 않았습니다. 포털의 장기 가용성이나 속도 개선율을 보장하지 않습니다.

오데덕 (odeduck) v0.21.1

현재 연결된 AI로 목표를 실행하고, 반복해서 전달하던 지침과 상태를 줄였습니다.
v0.21.0 배포는 Windows CI의 줄바꿈 문제로 공개 전에 중단했습니다. 상세 topic은 LF·CRLF 모두 지원합니다.

Changed

• MCP 기본 경로는 연결된 host가 계획·해석하며 다른 agent CLI 설치가 필요하지 않습니다.
  별도 모델 검토는 선택 사항입니다. 단독 CLI goal은 --review-with 없이 계획 모델을 사용할 수 있고,
  검토 미설정 상태에서 계산 결과를 얻으면 추가 계획 호출 없이 artifact와 미검증 항목을 반환합니다.
• 기본 계획 지침은 약 54 KB에서 5.9 KB로 줄이고 read_guide로 필요한 상세 topic을 읽습니다.
  원천·연산·권한의 상세 계약과 결정적 검증은 유지합니다.
• MCP goal은 최초 snapshot 뒤 상태 변경분을 반환합니다. RFC 6902 changes를 이전 state에
  적용하며 baseRevision을 확인합니다. 맥락 복구·기존 전체 상태 소비자는 fullState:true를 사용합니다.

Added

• modelUsage: 실제 자식 모델 호출별 바이트·시간과 provider 보고 입력·출력·캐시 토큰을 기록합니다.
  현재 Codex·Claude 보고 형식을 지원하며 미보고는 별도로 표시합니다. MCP host 추론·임베딩은 포함하지 않습니다.

Validation and limits

• 동일한 계산 fixture에서 전체 응답 대비 변경분 전달은 약 45% 작았습니다. 지침은 89% 작아졌습니다.
  토큰·과금액이 같은 비율로 줄었다는 의미는 아닙니다. 검증 기록.
• 기본 경로의 review_required를 독립 검토된 output_ready로 변경하지 않습니다. CLI는 미검토
  결과가 있더라도 목표 미완료 종료 코드를 유지하므로 JSON의 artifact와 evaluation도 확인해야 합니다.
• 단독 CLI에는 계획 모델이 필요합니다. CLI의 행동별 모델 호출, 프로세스 재개, INTENT 전체 완주는 남아 있습니다.

오데덕 (odeduck) v0.20.0

목표를 말하면 필요한 공공데이터를 발견하고 결과와 근거까지 만드는 흐름을 goal로 통일했습니다.

Changed

• CLI는 odeduck goal, MCP tool은 goal을 사용합니다. 데이터셋을 지정하지 않은 발견·조사·보고서·분석
  요청은 목표 실행으로 시작하며, 검색은 필요한 자료를 찾는 단계입니다. CLI의 solve는 기존 스크립트용으로 남습니다.
• --review-with codex|claude|gemini로 고정된 계획·검토 수신자와 선택 근거 전송을 명시합니다.
  보고·분석·요청 범위 검토에 쓰이던 여러 옵션을 묶었으며, 기존 옵션은 호환용으로 유지합니다.
  CLI goal은 이 설정이 빠졌다면 모델 호출 전에 설정 방법을 안내합니다.
• 넓은 발견 목표는 서로 다른 구체적 가설을 세우고 실제 후보를 검사하며 수정하도록 공통 계획 지침을 개선했습니다.
  소재·데이터셋·결합 정답을 지침에 고정하지 않습니다.

Fixed

• 이전 대화를 참조하는 목표는 시작 시 context(CLI --context)로 필요한 맥락을 계획과 별도 검토에
  함께 전달할 수 있습니다. 원천 근거·권한·승인과 분리하고 크기·인증정보 검사를 적용합니다.
• 목표 실행 전에 같은 카탈로그와 의미 검색 경로를 점검하여 준비되지 않은 환경에서 계획 호출을 낭비하지 않습니다.
• 의미 검색 환경을 복구한 뒤 MCP의 같은 목표에서 retry_search로 진행할 수 있습니다. 원래 검색어·역할·
  목표 계약·실패 이력·만료·예산을 보존하며, 일반 검색으로 조용히 낮추지 않습니다.
• 실행 중인 바이너리·빌드·공통 지침의 해시와 허용 행동을 runtime에 표시합니다. 파일이 업데이트돼도
  기존 MCP 프로세스를 새 버전으로 오인하지 않습니다.

Migration and limits

• MCP의 advance_goal은 goal로 바뀌었습니다. 서버를 다시 연결하고 새 tool schema를 읽어야 합니다.
• 새 CLI 예: odeduck goal "사람들이 잘 모르는 흥미로운 사실을 발견하고 출처를 확인해주세요" --review-with codex
• MCP 검토 설정: odeduck mcp --review-with codex. 연결된 host와 지정한 검토 제공자에 선택 근거가 전달됩니다.
• 인증키 보호·근거 크기 제한·별도 검토·완료 기준은 유지합니다. 독립 CLI의 자동 신청, 프로세스 간 재개,
  INTENT.md 전체 목표의 범용 자율 완주가 완료됐다는 의미는 아닙니다.

오데덕 (odeduck) v0.19.1

로그인 상태 확인과 자동 활용신청 흐름을 개선하고, 사용자용 Agent Skill과 확장 가능한 개발 지침을 정리했습니다.

Fixed

• login은 저장된 세션을 먼저 검증하고, 유효하면 브라우저를 다시 열지 않습니다.
  활용신청용 headless 브라우저도 저장된 세션을 확인한 뒤 시작합니다.
• status는 세션 없음·만료 시 실패 상태를 반환합니다. 포털 통신 장애와 페이지 형식 변경을
  로그인 실패로 바꾸지 않으며, applications도 원래 오류를 보존합니다.
• MCP 가이드와 사용자용 Skill은 신청 전에 로그인·기존 승인 상태를 확인하도록 안내합니다.
  오류 응답에 동반된 빈 신청 목록을 미신청 상태로 해석하지 않고, 로그인 후 원래 작업을 이어갑니다.

Added

• 설치 가능한 odeduck Agent Skill을 제공합니다. 보고·비교·계산은 solve / advance_goal로,
  개별 데이터 조회는 검색·검사·신청·호출로 연결합니다. 실행 조건은 설치된 CLI/MCP에서 확인합니다.
• catalog validate-release --snapshot으로 배포 카탈로그를 로컬 설치 상태 변경 없이 검사합니다.
  별도 Catalog refresh 작업은 검증된 카탈로그만 독립적으로 발행합니다.

Changed

• 사업 검증·데이터 설계용 Skill은 짧은 진입 문서와 작업별 상세 문서로 나눴습니다.
  새 기능의 지침·실행 계약·평가 자료를 추가할 위치를 명시하고, 특정 제품 버전에 고정하지 않습니다.
• README를 한국어 사용자 중심으로 정리하고, 오데덕을 공공데이터 탐색부터 실행까지 연결하는
  오픈소스 플랫폼으로 소개합니다. 실데이터로 확인한 공개 파일·활용신청 사례를 홍보 문서에 담았습니다.
• 코드 릴리스는 검증된 기존 카탈로그를 재사용합니다. 체크섬·구조·제공형 범위·대표 검색 질의를
  다시 검사하고, 수집 날짜와 압축 파일은 보존하며 출처 태그·해시를 catalog-source.json에 동봉합니다.
  카탈로그 재사용은 최신 데이터 수집을 의미하지 않습니다.

목표 실행의 기존 지원 범위와 승인 조건은 유지됩니다. solve 내부 자동 활용신청이나 범용 목표 완주가
완료됐다는 의미는 아닙니다.

오데덕 (odeduck) v0.19.0

Start with a goal, inspect the actual sources, execute bounded analyses and review
selected supporting evidence through the same CLI/MCP engine. Goal execution and
model review are experimental: this release does not establish general autonomous
goal completion, canonical entity identity, causal claims or field verification.

Added

• Source-cited goal explanations through the shared CLI/MCP engine distinguish
  required source dates/definitions/limits from engine-written execution notes.
  Exact authorized packet-cell citations and a separate finding per explanation
  accompany the result; missing or unsupported explanations cannot complete a goal.
  Corrections retain execution/review history and expiry discards the draft prose.

• Goal analysis review reuses already-disclosed original XLSX cells across verified
  overlapping observations, without another disclosure or broader authority.
  Analysis v3 carries one packet collection with context references and cell
  attribution; conflicting provenance/values fail, and repackaging cannot buy
  another review. Comparison context carries its recipe once and references
  already-included original metadata, preserving complete provenance within the
  unchanged review limit. Source-report v1 and historical review archives stay unchanged.

• Local sample.compare through the shared CLI/MCP goal engine compares explicit
  keys and numeric measures from two retained original revisions. Exact decimal
  checks preserve duplicates, discrepancies and both unmatched sides. Selected
  computed summaries can support interpretation review without replacing source
  records, changing result roles or approving identity, applicability or coverage.

• Registered monthly FILE export through the shared CLI/MCP goal sample path,
  with typed period/registration/province/age choices, credentialless no-redirect
  form POSTs and complete CSV scan provenance. Publisher rows and selected-age
  columns stay original; source applicability and goal completion remain separate.

• Registered supporting HTML documents in FILE inspection and goal sample
  (delivery:document), with credentialless exact-origin acquisition, selected
  original DOM addresses and page/text hashes. Explicit evidence disclosure and
  composition.support carry document context into review; documents cannot
  become computational data or certify applicability or population coverage.

• Additional opt-in full-scope goal review (solve --review-full-scope, MCP startup
  --review-goal-full-scope), requiring existing analysis/disclosure authorization.
  Complete CSV retention or exact XLSX rectangle acquisition enables review, not
  population approval. Every original source needs a separately cited scope finding;
  missing evidence, truncated selections and unsupported acquisition types remain incomplete.

• Optional composition.support links already-disclosed evidence from another
  original observation to proposed interpretation targets in analysis review.
  Exact packets and acquisition provenance are preserved; support does not change
  calculations, satisfy required roles or certify applicability or population scope.

• Bounded portal FILE history discovery and exact edition selection through CLI,
  MCP and goal inspection. Historical requests retain edition-specific metadata
  and acquisition provenance, reject latest-file fallback and preserve old
  observations across reinspection. Listing versions grants no ledger verification.

• Optional composition.reportUnmatched projects selected excluded source fields
  into a separate goal result table, retaining stage-local record addresses and
  missing/null distinctions. It shares result limits, does not change calculations,
  and requires selected evidence before analysis review.

• Already-disclosed same-file header/context evidence in opt-in goal analysis
  review, pinned to source content and contract revisions. Context keeps its
  selected cells, original addresses and acquisition request separate from
  computational sources; file association is not proof of meaning or identity.

• Stage-local unmatched record addresses in goal join metrics, including empty
  joins, for bounded evidence reading and replanning. Relational review now also
  requires disclosed comparison fields for excluded direct records and replays
  exclusion metrics; unrelated excluded values remain private.

• Bounded exact value-set selection (sample.whereIn) for full direct CSV scans,
  preserving source record positions and separate scanned/matched/retained coverage.
  Selection order cannot bypass acquisition replay limits, and both review kinds
  require selected evidence for predicate fields. Development calibration can
  now acquire inspected sources and check unchanged independent records before review.

• Additional opt-in review of typed relational calculations via CLI
  --review-analyses or MCP startup --review-goal-analyses. Local replay and
  original record participation bind already-disclosed result values; separate
  relation, period, measurement and coverage findings accompany goal/output checks.
  Source-report authority does not enable this scope. Review budgets and canonical
  evidence deduplication are shared; spatial, causal and hypothesis approval remain unsupported.

• Source-grain aggregation of retained observations before cross-source joins via
  goal sample.reduce. Exact sums, contributing-record groups, source revisions,
  original-period checks and shared budgets survive CLI/MCP execution. Computed
  groups remain distinct from publisher records and do not acquire source-report approval.

• Opt-in, separately invoked model review for bounded source-field reports via
  CLI --review-source-reports or MCP startup --review-goals-with. Original-goal
  fit and every output's source support are assessed separately against already
  authorized evidence; qualified completion carries reviewer/method/input/execution
  provenance. This is not field verification or approval of joins, calculations,
  population claims or hypotheses. Default-off disclosure and three-review limits apply.

• Native STD inspection through CLI/MCP, plus bounded CSV, ZIP-member and exact
  XLSX-rectangle readers with source hashes, original positions and explicit limits.

• Optional, recipient-bound selected evidence for experimental goal planning:
  CLI --share-evidence requires one explicit agent; MCP disclosure is fixed by
  server startup --share-goal-evidence. Original record addresses, exact values,
  missing/null states, cumulative disclosure budgets and session expiry are retained.
  Default planning remains value-free; evidence access is not semantic approval.

• Experimental shared goal-based discovery and bounded composition via CLI solve
  and MCP advance_goal, with required-role/output checks, source-record lineage,
  typed FILE/STD acquisition, explicit retries and versioned scope-label comparison.
  Candidate artifacts remain distinct from goal completion; general semantic
  approval and durable goal resumption are not implemented.

Fixed

• Keep frozen vocabulary hashes identical across platform checkouts.
• Discard retained goal values before returning acquisitions that finish after
  session expiry, including source failures and cancellation.
• Reject malformed UTF-8 before decoding goal-planner decisions, and return
  bounded errors for oversized comparison numbers and result tables without panics.
• Continue experimental goal planning after review_required within the original
  contract, disclosure policy, budgets and expiry. Pin evaluations to execution
  revisions and clear stale current results before a replacement execution.
  Restarting the planning loop no longer refills its one replay correction.
• Share semantic-search policy across CLI and MCP, keeping degraded retrieval
  visible and preserving the connection limit through progressive discovery.
• Strengthen credential isolation at the shared REST and LINK response boundaries;
  withhold unsafe responses without creating replacement source observations.
• Preserve empty CSV strings instead of inventing nulls; keep STD reads bound to
  the first-party response and reject ambiguous JSON object members.
• Allow ordinary single-source projection and aggregation through the shared goal
  executor; remove the spatial-only zero-join restriction and unexecuted join explanations.
• Rename the experimental goal artifact state from sample_joined to sample_executed
  for all current operations. Historical diagnostics retain their original state;
  this does not change semantic approval or CLI completion requirements.
• Reject lossy Unicode decoding in native STD schema/row JSON before retaining
  source identifiers. Preserve valid replacement characters and surrogate pairs.
• Describe only executed key/measurement operations and preserve the conditional
  source coverage of nearest-record calculations in goal evidence explanations.
• Preserve original CSV empty strings in nearest-derived evidence and artifacts.

Changed

• Share one engine-owned goal planning guide across CLI prompts and the MCP
  resource; remove duplicate operator instructions and obsolete join requirements.

오데덕 (odeduck) v0.18.0

Added

• Record evidence for cross-domain connections in a local, append-only ledger,
  including the datasets, identifier fields, provenance, sample overlap and join
  expansion used to support a decision.
• Verify sample claims against recent call_api profiles in the same MCP session,
  with delivery-, operation- and request-bound receipts that never persist raw
  response values.

Changed

• Require explicit expected-key mappings, consistent count aggregates, safe
  data.go.kr provenance URLs and delivery-compatible evidence records.
• Keep read-only empty-ledger queries side-effect free and bound receipt memory;
  preserve ambiguous or damaged JSONL tails instead of deleting evidence.

Fixed

• Prevent fabricated sample metrics, credential-bearing provenance and
  whitespace or ordering variants from bypassing validation or idempotency.
• Make ledger locking, concurrent append, cancellation, cross-platform builds
  and profile evidence hashing deterministic and test-covered.

오데덕 (odeduck) v0.17.2

Changed

• Adopted the stable black-glasses Odeduck logo and regenerated the social and
  launch artwork from the same source.
• Shortened the README around the product problem, a concrete cross-domain
  example, quick start, measured proof and safety boundaries.
• Kept the original no-glasses cabinet animation as the README hero and moved
  experimental video variants out of the public assets.

Fixed

• Updated the README, launch kit and installer examples to point to v0.17.2.

오데덕 (odeduck) v0.17.1

Fixed

• Align the deterministic lexical release gate with its documented contract:
  retain exact and cross-domain term checks, while leaving broad intent such as
  “cheap investment property” to agent-planned or semantic discovery instead of
  assuming an undocumented synonym mapping.

oddsock v0.16.1

Changed

• Reframed the README around the everyday problem of testing a small-business idea
  without knowing official public-data terms, while preserving the doorway character
  and existing brand mark.
• Clarified where existing public-data MCPs still hand work back to people and showed
  the real REST, FILE and LINK paths without implying that oddsock makes the final
  business decision.
• Added a reproducible, no-login first-run example plus GitHub and LinkedIn launch
  artwork that preserves the existing character mark.
• Made the macOS/Linux and Windows installers fall back to public GitHub downloads
  when the GitHub CLI is absent or unauthenticated, while retaining private-preview
  support for authenticated contributors.
• Added a current architecture map and a manually dispatched anonymous launch smoke
  test that verifies the public repository, release assets, installer, catalogue and
  first no-login search before promotional links are distributed.
• Replaced the stale pre-rename video brief with the current rainy-season cafe demo
  and marked its old Higgsfield job IDs and render as non-publishable records.

Security

• Added a redacted full-history Gitleaks check to CI, with exact-fingerprint ignores
  for reviewed test fixtures and protocol examples so new findings still fail.
• Ship checksummed installer scripts as versioned release assets, with hermetic
  fallback and checksum-rejection tests on Linux and Windows.

Community

• Added contribution, security and conduct policies, structured issue forms, a
  pull-request checklist, funding metadata and monthly dependency updates.
• Removed paused local-tracker scratch documents from the public tree; their history
  remains recoverable in Git while GitHub Issues stays the canonical tracker.

oddsock v0.16.0

Changed

• Reframed the README around oddsock's quiet doorway character and a single-line
  natural-language question, with concise proof points for discovery, inspection,
  application and first-call workflows.
• Replaced the meerkat mark with a transparent doorway-character SVG and matching
  2048px PNG, including a light keyline that stays legible on dark GitHub themes.
• Restored authenticated private-repository installation, source-install catalogue
  bootstrap and legacy Homebrew/Windows upgrade guidance.

Security

• Documented when standalone discovery sends a goal to a logged-in Codex, Claude,
  Gemini or Cursor process, including sensitive-input and untrusted-metadata bounds.