본문으로 건너뛰기

0.2.11

번역 준비 중

이 쪽은 아직 한국어로 옮기지 않았습니다. 아래는 영어 원문입니다.

Behavior changes​

  • Add a bounded context journal for Claude Code hooks, Codex agent missions, and harness child lifecycle events.
  • Workflow runs now support single-node execution, resuming from prior runs, and pin-required test mode.
  • Decision-card answers now resume the waiting harness run, while deadlines retain safe fallback behavior.
  • two programs writing to the same conversation at once no longer lose messages from its count or undo its project assignment; assigning a conversation to a project that does not exist now answers «not found» (404), and the dashboard suggests a project from the session folder whenever it changes.
  • The INSIDE2 live-event SSE endpoint (GET /v1/inside/events) streams graph, harness, seat, steward, and PTY decision updates with secret values masked.
  • Steward live mode launches eligible goals only after pre-launch duplicate, budget and location checks, while money, publishing and security decisions still require human approval.
  • Add a rate-limited diagnostics intake endpoint that stores doctor bundles in private S3 and returns a support code.
  • The elanous-basics plugin now includes a Google Workspace skill for your own Gmail and Calendar, plus Drive and Sheets reads.
  • A new elanous-essentials pack bundles five everyday skills — YouTube digests, web search and crawling, link digests, diagrams and lecture-note digitizing — each with a free path and a bring-your-own-key path.
  • elanous harness stop <run> now finds and stops a run that was launched from another worktree: the run is bound to its process in a per-user host registry, and the stop still checks the process start time before any signal.
  • Pod pool members can use separate registry ports, and lease checks exclude members whose in-Pod DNS lookup fails.
  • Setup automatically enables skills for signed-in gh, gws, op, and himalaya CLIs, skips unavailable or signed-out CLIs, and keeps public GitHub repository reads available without signing in.
  • Plugin make now researches the web and Aside local memory before drafting a connector, skill, and graph, and continues when research is unavailable.
  • workflows can run independent nodes in parallel (concurrency, default 1) and route a failed node to an on_error handler instead of stopping.
  • Condense a day of context events and decisions into source-linked, project-and-seat memory with correction, conflict and retirement signals.
  • Install, list and remove YAML team sets of titled personas with validated handoffs.
  • Pods have a lite tier (2 GiB) you opt into: elanous pod run --lite runs network-skill commands (omni-crawl, omni-digest) on a small image, and --pod-memory lite or a Pod 메모리: lite goal line gives a harness goal 2 GiB; a lite run that runs out of memory retries once on the high tier, and pod lease status shows lite slots.
  • elanous graph variant <template> --plan <json> --dry-run|--run validates a variant, saves its graph beside the run ledger, and executes the validated wiring.
  • The harness gate runs up to 30 reverse-import and HTTP-route consumer tests, including PWA tests for removed routes, and warns with the full overflow list.
  • elanous make "<one-line request>" classifies the request and routes it to the existing plugin maker or workflow builder, asking which to use when classification fails.
  • Isolate provider availability tests from host credentials and verify Codex-first selection with the default fallback chain.
  • Plugin make can prepare a leak-checked official-market bundle after local installation, with an optional configured signature and no publishing.
  • elanous skills install essential --code <CODE> --email <address> --agree installs the essential skill pack from a gift code — nothing is sent without --agree, existing skill folders are never overwritten, and an oversized or mismatched download is refused.
  • Asking how X reacts to something now searches X itself, so the answer cites x.com posts.
  • Delegated work now picks a default output for each task — a report, slides, a research doc or a post draft — instead of ending in a chat answer, and what was produced is recorded in one outputs list.
  • Cache successful per-member DNS checks for ten minutes under a host lock and exclude labeled DNS probe Pods from goal counts.
  • the harness launch gate no longer blocks every launch when the launch universe's config has no model chain: it judges with the built-in default chain and prints a warning, and an explicit --child-llm-provider counts as the chain.
  • classification calls no longer fail when no global LLM provider is configured: a role that names its own provider resolves without one, and an injected classifier transport still runs (marked unresolved).

Plus 206 internal changes.