跳到正文
DEV Community· Charan Nihaal R·· 2 小時前AI 評分70

Entire 從 CLI 走向完整程式碼託管平台:內部體驗者回顧演進歷程

from a cli to a full forge: my ride as an Entire insider

AI 導讀

Entire 今日正式公開推出,從把智能體工作階段連結至 Git commit 的 CLI,擴展成包含 Trails、原生 Git 託管、記憶外掛及 CI 整合的平台。

正文

Entire went public today. i've been an insider since the beta dropped, and i've tried pretty much everything that shipped along the way. so instead of a quick "congrats" post, here's the whole ride, release by release, the way i lived it.

if you just want the launch stuff, skip to what's live today. but honestly the journey is the fun part.

the problem it started with

every time i code with an agent, the most useful part of the work is the conversation. why we picked this approach, what we tried first, what broke. and then i commit, and all of that just... disappears. git keeps the what, never the why.

Entire's first idea was simple: save the agent session right next to the commit. that's it. that one idea is what hooked me.

february: just a cli and some hooks

the early builds (0.4.x, starting around feb 10) were all about getting the basics right.

  • entire enable hooked into my repo and started capturing agent sessions as checkpoints tied to commits
  • entire explain showed me the session behind a commit
  • rewind and resume let me jump back to an earlier point in a session
  • secret detection got layered on top (gitleaks patterns + entropy scanning) so transcripts don't leak keys
  • git hooks started chaining instead of overwriting, so it played nice with whatever hooks you already had, and it even warns you if you use Husky or Lefthook
  • OpenCode support landed in 0.4.6, Cursor checkpoints in 0.4.8, and Factory AI Droid in 0.4.9

this was the "wait, how does this work" phase for me. running entire status and seeing session cards show up for the first time was a small but very satisfying moment.

march: plugins, codex and windows

  • 0.5.0 (mar 6) brought the external agent plugin system, so other agents could plug in without being built into the cli
  • 0.5.1 (mar 19) added entire trace for figuring out slow hooks, and opt-in PII redaction
  • 0.5.2 (mar 30) was a big one: Codex CLI integration, proper Windows support, and the first work on checkpoints v2

this is where it stopped feeling like a claude-only toy. i could use different agents on the same repo and everything still landed in one history.

april: searching your own history

  • 0.5.3 gave us entire sessions (list, info, stop) and entire attach for linking sessions it missed
  • 0.5.4 made entire search available, plus active agents in entire status
  • 0.5.6 (apr 24) added entire activity and the first version of entire dispatch

entire search changed how i used it. instead of scrolling through old chats, i could just search "why did we do x" across my sessions.

may: review agents show up

  • 0.6.0 brought a nicer search palette and protected-branch push warnings
  • 0.6.1 (may 7) introduced entire labs, entire labs review (run review skills inside an agent session, with a live dashboard) and entire recap
  • 0.6.2 made the Pi agent integration internal, added custom redaction rule packs, and live streaming for entire review

looking back, entire review was the first hint of what became Trails. agents reviewing agent code, with a score you can actually read.

june & july: the forge starts to appear

  • the 0.7.x releases kept polishing everything through june
  • 0.8.0 (jul 6) was huge: a new git-refs checkpoint store, entire agent-help (docs written for agents), entire experts, entire api, and the first org / project / repo commands with git-remote-entire
  • 0.8.42 promoted org, project, repo and grant out of labs into the main cli, and started cross-region code search
  • 0.9.0 (jul 27) made git-refs the default, added local agent picking for entire dispatch, and multi-repo semantic search

this is the moment it clicked for me that Entire wasn't "a cli that saves sessions" anymore. it was becoming a place where your repo lives.

august & september: native repos, mirrors and runners

  • 0.10.0 (aug 12) added remote plugin install and plugin dependencies
  • 0.10.1 – 0.10.3 made login smoother and made subagent work survive into checkpoints for Claude Code, Codex, Cursor, Copilot CLI and Droid
  • 0.10.4 (sep 2): entire repo clone /et/<project>/<repo> to clone an Entire-native repo, and region-scoped dispatch
  • 0.10.5 made mirror creation async with live progress
  • 0.10.6 (sep 7) brought a Windows PowerShell installer, reworked entire runner setup, and trails working on native repos
  • 0.11.0 (sep 22) cleaned up the whole command surface (repo view, auth switch, etc.)

cloning a repo with an /et/ path for the first time felt weirdly special. like, this is not github, this is its own thing.

october: trails grow up

  • 0.11.4 (oct 6): entire repo view became one view for both forges, showing every region holding a copy
  • 0.11.5 (oct 14): no more shadow branches, checkpoints only get written when you commit, and entire checkpoint delete
  • 0.11.6 (oct 20): per-reviewer agent configs in review profiles, and entire trail merge got smarter about branches
  • 0.12.0 (oct 21): project trails became the default, and links moved to entire.com
  • 0.12.1 / 0.12.2 (oct 24 – 26): entire checkpoint list --outgoing shows what your next push uploads, a privacy notice in entire enable, IP redaction, and per-call token usage in transcripts

and then, today.

what's live today

Trails instead of PRs. a trail can start from a bug report, a feature request, or straight from your agent. every push gets scored by review agents (security, slop, human-review-required, or your own) running in cloud sandboxes. you see the last few runs for each eval, findings stay pinned until they're fixed or dismissed, and gates stop a merge until evals, human approvals and CI are happy. it reads like a pilot's dashboard, not a wall of comments.

native git hosting. you can create and host repos directly on Entire, replicate them across regions (US, EU, Australia and India at launch), or mirror from GitHub and keep it as the source of truth. their benchmarks say pushes are up to 89x faster than GitHub.

continuous memory. two open source, local-first plugins:

  • entire-graph maps your codebase (definitions, callers, routes, change impact) so agents don't burn tokens grepping around. up to 71% fewer tokens in their testing.
  • entire-brain ties past agent sessions to that map, so your agent remembers the why, and reviewers see those memories right on the trail.

CI with Buildkite and Depot. connect either one at the project level, and builds show up as checks on your trail. there's even bk pipeline convert and depot ci migrate to bring GitHub Actions workflows over.

and marvin 💜

⬛⬛⬛⬛⬛
⬛⚪⬛⚪⬛
⬛⬛⬛⬛⬛

this is Marvin, Entire's mascot. a little black block with two eyes. i don't know when it happened, but i'm fully attached to him now. no notes.

thank you

being an insider meant watching a product get built in real time. huge thanks to Haimantika Mitra, Rizel Scarlett and Lizzie Siegle for keeping us insiders in the loop the whole way, answering questions and making us feel like part of it.

and congrats to Thomas Dohmke and the whole Entire team. going from a cli i was poking at in february to a full forge going public in october is wild.

try it 👉 entire.com
meet the team 👉 entire.com/about

welcome to the rebellion 🚀

來源:DEV Community · dev.to