0.2.15

번역 준비 중

이 쪽은 아직 한국어로 옮기지 않았습니다. 아래는 영어 원문입니다.

Behavior changes

• the harness no longer deletes files that were added to main after a run started when they are outside the run's target paths.
• elanous release auto-start previews the scheduled release cut within a window and, with --apply, starts the guarded release run (freeze or readiness changes defer instead of failing).
• elanous hq fence-audit lists cron jobs that write outside the HQ fence and proposes the wrapped line (no crontab changes).
• the public-demo "now" view no longer crashes on installs without brand rules and never returns a developer error page.
• when the last review round fails, the harness child opens a review-blocked draft PR with its work and must-fix list instead of exiting without a result.
• every elanous-related cron line now shows up in the loop registry as a loop or an unregistered warning; none is silently dropped.
• a recorded heal event now starts the heal loop once.
• The author-depth shadow tick checks selected release-cell titles and text, holding missing-text cells with ledger receipts, without enqueueing harness work.
• brand-dependent generators now report missing rules as unmeasured rather than passing an unchecked brand assessment.
• decision cards are no longer re-sent every 20 seconds when no device is subscribed.
• a seat loop that falls back to shadow now says why.
• hq move-plan lists what a headquarters host owns before moving it.
• Each traffic tick records author-depth shadow receipts and reports held and queued-for-author counts plus incomplete ledger writes in JSON without enqueueing harness work.
• Lecture-note graph collection now OCRs photos into draft inputs, recording unreadable images with a reason.
• Lesson handbook pages show hash-free dated IDs and readable, 200-character cause and remedy summaries.
• a failed Pod run now reports why it failed instead of «no error diagnostic».
• The scheduled harness queue tick now records author-depth shadow receipts at most once every 30 minutes without changing dispatch decisions.
• Generate a capability map from maturity grades, green release cells, and the marketing feature map, with evidence gaps and patent-held concepts marked.
• the Telegram intake digest now leads with what matters to us, at most three items in a short situation-reason-action form.
• tests no longer start the real heal loop (and its paid external grounding) when they record a failure event.
• Feature maps now use marketing-map maturity, publication status, and PR citations when a row has no release checklist cell.
• seat requests can be listed and closed from the CLI.
• a waiting harness-queue item now names which seat cap is holding it back.
• seat launch caps follow each seat's remaining work and lend
• at most two heal loops run at once; further failures wait in the inbox.
• intake check now flags paid dependencies a new tool could replace, and asks for a hands-on trial first.
• agent-env capture records a secret-free profile of the local Claude, Codex and Grok setup.
• Rhythm daily dry runs preserve review files, deduplicate PR landings, and bound unlanded runs; weekly reviews use scheduled cuts and recent daily news.
• Returned Pod child logs now populate the host logs database once per job, preserving event identity and source attribution.
• A new public guide shows how to attach a Claude Code or Codex session to an elanous seat using a project-local marker and context checks.
• links saved from Telegram are absorbed right away instead of waiting for the daily run.
• the HQ arbiter asks the operations seat to promote the standby after the holder goes silent; automatic promotion stays opt-in and a failed runbook is never reported as a promotion.
• the draft sweep no longer treats two drafts as the same goal just because they share a title; it needs a matching goal id, or the same title on the same harness branch stem, and never an auto-generated file-name title.
• Idle seat checks now request the next evidence-free yellow release cell after 30 minutes, with shadow mode by default.
• test processes no longer run the real author-depth shadow on every queue tick (it timed out the seat-loop tests).
• doctor node-modules reports missing or mismatched installed packages in a working tree.
• a finished run that stalls while catching up with main is kept as harvestable instead of being dropped.
• The public loop agents guide now explains the orchestrator's handoff flow, shadow and live modes, human decision boundary, and status commands.
• the changed-file type check fails instead of passing when it cannot see TypeScript changes that are in the diff.
• a review that failed to run is no longer shown as PASS in PR comments or the PR body, and it is retried once with another provider.
• AX briefs now include the measured, deduplicated self-implement merge count and report source read failures explicitly.
• the child model provider chosen for a run is now used in Pods and local runs, or the launch is refused.
• the release checklist gate fails on a carry-over ID collision instead of silently leaving the item behind.
• several release checklist commands opening a fresh ledger at the same time no longer fail with "duplicate column name".
• one seat at its cap no longer blocks other seats' goals in the harness queue.
• draft sweeps record a daily count of old harness drafts and say why a sweep failed.
• pod run always saves the job's log to the artifact folder it announces.
• the run supervisor no longer stops a run as «no progress» when its children finished and committed on a non-main base.
• the draft cleanup and steward loops no longer fail outside a git checkout, and a failing graph step keeps the real error line.
• Pod launches now pick the pool member with the most real free capacity instead of always the first one.
• Pods no longer receive a child effort flag that self implement rejects, which would stop the Pod at startup.
• Pods launched at the same moment share one GitHub App token instead of each minting one, and a Pod whose GitHub login fails is retried once after a short backoff.
• every harness run records which door it came through and whether it went via the queue; eln harness doors shows the daily table.

Security

• agent-env capture redacts key-shaped values (sk-, ghp_, AKIA… and long digit-heavy strings) even in plain identifier fields.

Plus 103 internal changes.

Known issues

The following issues are known in this release.

• Known issues have been reported for this release.