elanous v0.2.37

Behavior changes

• Release resume validates writable traffic configuration, dependency directories, and saved release artifact ownership before changing run state.

Plus 170 internal changes.

elanous v0.2.36

Behavior changes

• Continue bounded remaining work in the TUI, yielding when new input or a human decision is needed. (#30657)
• Route larger goal work through subagents when delegation is appropriate. (#30617)
• Record doctor visits for long-running work and transitions that need owner attention. (#30403, #30468, #30608, #30618)
• Improve first-run onboarding for Korean and English users and guide new users toward a first useful result. (#30536, #30646)
• Connect Software Wizard requests to the actual run path. (#30668)
• Stop the release workflow when npm publishing fails. (#30677)
• Verify accepted gate-radius receipts and fail closed when a receipt does not match. (#30678)

Plus 204 internal changes.

Known issues

The following issues are known in this release.

• Publishing Markdown files for internal use may not work reliably.

elanous v0.2.35

Behavior changes

• 기존 TUI를 기본으로 유지하고, --tui next로 새 TUI를 선택할 수 있습니다. (#29674, #29708, #29998)
• 새 TUI에서 로컬 대화를 스트리밍하고 도구·하니스 명령을 실행할 수 있습니다. (#29683, #29703, #29841, #29823, #29852)
• 새 대화, 작업 진행 상태, 제공자 선택과 테마 적용 상태를 화면에서 확인할 수 있습니다. (#30015, #29743, #29751, #29724, #29770)
• 새 TUI에 턴 전 압축과 휠 로그 스크롤을 더했습니다. (#30030, #29969)
• PR·커밋을 지정해 격리된 Pod 명령을 실행할 수 있습니다. (#29974)
• 큰 위임 작업을 나눌 때 단계가 부모 계약과 완료 조건을 이어받고, 완료 경계와 분해 결과를 확인합니다. (#29993, #30026, #30052, #30041, #29753)
• 원격 작업에서 저장된 인증과 선택한 모델 설정을 유지합니다. (#29746, #30043)
• 릴리스 자료 수집·레지스트리 오류 처리를 보강하고, Pod·게이트 실패 때 이미지·시험 조각·기준선·종료 근거를 보존합니다. (#29632, #29650, #29646, #29662, #29676, #29757, #29853, #29868, #29900, #29941, #29956, #29962, #30009, #30033, #30053, #30117)
• 같은 작업의 PR 재사용과 연결을 개선하고 중복 생성·잘못된 연결을 줄였습니다. (#29735, #29869, #29912, #29915, #29948)
• RFC 선행 안내와 리뷰 판정을 보강해 차단 사유와 긍정 근거를 더 분명히 합니다. (#29822, #29762, #30007, #30023, #30038)
• 보류·모호한 분류, TUI 선택기와 유휴 클립보드·Git 확인 동작을 다듬었습니다. (#29889, #29978, #30108, #30044, #30051, #30060)

Plus 142 internal changes.

elanous v0.2.34

Behavior changes

• Task Agent graph journeys can generate and verify documents, recover from landing refusals, and enqueue a bounded landing to the train.
• The guardian reports when the selected Grok subscription is within four hours of expiring, without returning its token or account ID.
• The TUI shows concise progress summaries for GPT-6 turns and flushes them when a turn ends with an error.
• Task decomposition now preserves each child task's completion boundaries, including code, live-run, and evidence requirements.
• Task Agent repair follows graph next moves, binds branch updates to the source card, and caps repeated review and landing attempts.
• Release gate fixes cover six introduced 0.2.32 regressions across scenario and draft tests, and gate-triage classifies graph gate failures as regressions.
• Pod checkout skips AppleDouble sidecar files so macOS metadata does not disrupt remote work.

Plus 204 internal changes.

elanous v0.2.32

Behavior changes

• The TUI can wake for a completed child while idle, show background job counts, and schedule its own next turn. PRs: #28892, #28743, #28742.
• Reusable warm Pods are available as an opt-in transport, with request cleanup between uses. PRs: #28292, #28727.
• External /stop and /esc requests reach the active turn's escape gate immediately. PR #28769.
• Task Agent repair launches accept an explicit dispatch tier and select models using subscription availability. PR #27786.
• The TUI shows when a fallback model answered and keeps the session boot card current after history reset. PRs: #28734, #28746.
• PWA live run graphs show the current execution graph. PR #28879.
• TUI click-to-fold controls can be toggled back open in place. PR #28825.
• The TUI detects a doom loop and changes approach without asking, and avoids false “NO FINAL SYNTHESIS” after a goal loop has ended. PRs: #28795, #28772.
• Resizing the terminal preserves the open slash picker border; folding panes preserves their position. PRs: #28766, #28825.
• Long RunShell sleeps explain how to schedule a wakeup, and CLI legacy loop warnings no longer pollute standard error. PRs: #28763, #28432.
• Background input wakeups, PTY state reporting, and session tool traces now reflect the active turn more reliably. PRs: #28750, #28740, #28741, #28876, #28882.
• Task Agent closes stale drafts after a code PR opens and retries completed no-change children once. PRs: #28829, #28824, #28813.
• The PR tool warns on overlapping changed lines while allowing a new PR to open. PR #28881.
• Webclone reads ruler versions portably on Linux. PR #27855.
• Internal marker text is escaped from the public test export. PR #28754.

Plus 103 internal changes.

elanous v0.2.31

Behavior changes

• External /stop and /esc requests now bypass queued input and reach the active turn's escape gate. PR: #28724.
• Task Agent repair launches can carry an explicit dispatch tier and use subscription availability for model selection; older task cards remain compatible. PR: #27786.
• /session new refreshes the boot card before asynchronous ACP renewal, keeping the card in sync when conversation history is cleared. PR: #28746.
• Legacy loop warnings no longer appear on CLI standard error. PR: #28432.

Plus 203 internal changes.

Known issues

The following issues are known in this release.

• Release verification may take longer than expected after small changes because some checks can run again instead of reusing earlier results.

elanous v0.2.29

Behavior changes

• Keep multiline tool commands within one TUI log row so streaming input and status remain visible.
• Do not apply the fact-question tool budget on goal-loop turns, so status questions in the TUI no longer stall with «Tool budget reached».
• Fold consecutive RunShell results into one expandable TUI row in task-unit mode.
• Open an inspectable shell or Codex session from the TUI with /term new [command] or the terminal shortcut.
• Restore the in-memory TUI history prefix after a failed Codex 400 repair retry when it has not changed, while retaining later messages.
• Text injected from outside with pty text no longer merges into the line you are typing or the mid-turn typeahead; it waits as a separate queued message (marked in the TUI) and is sent after the turn once the input is empty, and pty text --when-idle waits for an idle screen first.

Plus 123 internal changes.

elanous v0.2.28

Behavior changes

• Failed reviews can launch a bounded commit-only repair on the same PR branch with exact-head protection and owner notification.
• Seat messages can stream from a separate authenticated SSE post office, with durable replay and polling fallback.
• eln pr land --cwd now reads TypeScript gate inputs from the selected worktree and finds an installed compiler before running; missing compiler dependencies fail with an install hint.
• Seat messages can link replies, wait for the first answer with an explicit sender, and display ordered threads.
• Explicit child rework escalation can use an optional same-provider model map, with env precedence and an observed target decision.
• Telegram briefings now collect item button and reply feedback, order measured topics by weekly action rate, and keep incomplete split items eligible for retry; delivery remains at least once.

Plus 103 internal changes.

elanous v0.2.27

Behavior changes

• Seat boot now shows a read-only local handoff card from current seat sources.
• Self utterances now exports paginated, redacted human and agent speech by date, surface, and origin.
• eln tasks hand --checklist now uses the checklist title when no request text is supplied and rejects request text without a target path before launching a worker.
• eln msg watch --since-ack now explicitly reads a seat's unacknowledged messages without acknowledging them.
• The inside architecture view now follows the current product hierarchy of loop agents, graph engineering, wizard, and supporting runtime.
• eln tasks hand --checklist now finds cells across open release versions, with --version to select a specific release.
• The expose rubric CLI now checks public marketing blog drafts passed through --files for reproducibility, confidentiality, and brand fit before review.
• eln tasks cover now counts authored goals only after their goal run completes and merges, and marks unreadable ledgers as unknown.
• PR landing now accepts review result filenames with a 10 to 40 character commit prefix when the review content confirms the full PR head.
• eln msg post --to ALL now stores one broadcast message; each recipient reads it alongside direct messages with their existing acknowledgement cursor.
• Channel posts now mirror addressed messages into the durable inbox, and seat boot shows each seat's unread message count.

Plus 92 internal changes.

elanous v0.2.25

Behavior changes

• Retro facts now includes run exit incidents in its read-only output.

Plus 133 internal changes.