Status

Where the build stands

Milestones 0 to 3 build a working code host for people, one that GitHub clients treat as GitHub Enterprise Server 3.22. AM1 adds what that host needs to carry hundreds of AI coding agents safely on one repository. The rest makes it operable, recoverable and installable anywhere.

  1. M0
  2. M1
  3. M2
  4. M3
  5. AM1
  6. M4
  7. M5
  8. AM2
  9. AM3
  10. M6
  11. AM4
  12. M7

4 closed3 under way5 to comeIn build order, as of 15 November 2026

  1. M0ClosedClosed 31 October

    Measure the platform and the protocol

    Measured Cloudflare and the replication protocol on prototype code before any product code. Every number the design depends on was tested here, and no test hit its fail line.

    Rerun since on the production Rust host

    • Commit at 100 commands/s: 6 of 6 runs under 600 ms, p99 244 to 377 ms B0-01
    • Group commit at 300 and 330 commands/s: 16 of 16 runs under 600 ms B0-02
    • A gray site during another site's deploy: 235 of 235 groups moved within 20 s, 0 failed writes B0-07
  2. M1ClosedClosed 4 November

    Groups and git

    Replicated groups across three sites, and git push, clone and fetch over HTTPS with stock git. Packs are stored in R2 with a clone cache in front, plus basic accounts and tokens.

    Gate

    • 348 of 348 client checks on git 2.39.5, 2.51.0 and 2.56.0, with a site failing in half the runs BM-1
    • A 24-hour chaos run lost 0 of 904,294 acknowledged writes BM-1
  3. M2ClosedClosed 7 November

    Issues, pull requests and merge

    Issues, pull requests, reviews, check runs, merge and the merge queue, LFS objects and release assets, with REST and GraphQL that match GHES 3.22.

    Gate

    • gh 2.102.1, git and git-lfs 3.8.0 pass unmodified against the hosted edge
    • A merge on a newly elected leader: p99 1.36 s in its region and 1.83 s from another continent BM-2
  4. M3ClosedClosed 9 November

    Identity and authorization

    Fine-grained tokens, OAuth and device flows, GitHub Apps with installation tokens, single sign-on with OIDC, collaborators and invitations, repository transfers, and read replicas (learners) for busy repositories.

    Gate

    • 200 of 200 revoked tokens refused on the next push, over HTTPS BM-3
    • Terraform's own acceptance test for collaborators, 6 of 6 runs BF-34

    After Milestone 3, git, gh, Terraform and Octokit work against Graftyard unchanged, at human scale.

  5. AM1Under wayFeatures built, gate open

    Agent-scale safety and load

    What the host needs to carry hundreds of AI coding agents safely on one repository. The features are built and their functional tests pass. What is left is scale.

    New features

    • Agent identity. The host knows which commits and requests come from an agent, which vendor and kind of agent it is, and records provenance on each push.
    • Agent policy. Protected paths an agent can't touch, a merge gate for agent changes, and per-org switches to turn agents on or off.
    • Tasks. An API under /api/graftyard/v1 to create and track agent tasks.
    • Leases. An agent claims files or refs before it edits them, so two agents don't collide. Pushes outside the lease are refused.
    • Activity feed. Agents long-poll for repository events instead of hammering the API.
    • Trust view. An opt-in field that flags hidden characters and agent authors in issue and pull request text. Default GHES responses stay byte-identical.

    Load handling for hundreds of agents on one repository

    • Admission lanes. Human and merge-queue commands always go first. Agents get a share, and past it they are refused cleanly with "retry after" instead of slowing people down.
    • Inline small packs and log archiving, so one hot repository's log doesn't grow without bound.
    • Fan-out to more git cache shards under heavy push load, and budgeted test merges.

    Passing at 500 agents on one repository

    • Merges: p99 1,208 to 1,571 ms in three ten-minute runs, against a 2 s line B0-29
    • API reads: p99 1,339 to 1,982 ms with 100.00% of reads served and no leader change, three runs in a row, against a 2 s line B0-29
    • Leases, refusals for six kinds of actor, the activity feed, the trust view on eight resources, and a byte-identical GHES default L0-11, L0-13, L0-16, LF-4, LF-12, LF-13

    Still to pass

    • Commit p99 under 600 ms with 500 agents pushing B0-29
    • A 30-day storage run on a repository with a 4 GB metadata tier, which needs metadata child groups, now being built B0-30
    • 1 MiB agent pushes with every bucket failing, a burst of 500 pushes in one second, and 5,000 agents with human commits still under 1 s B0-31 to B0-33
    • Git CPU for one 5 GB repository under load, with clones starting within 1 s B0-34
    • The activity feed under a leader kill every 60 s L0-12
    • Test merges with Actions dispatch, which waits on Milestone 4 B0-35

    Each piece passes on its own. The gate asks for all of it at once on one repository: 500 agents, human commits fast, no leader churn, memory under the isolate limit, and 30 days of storage.

  6. M4Under way3 of 6 work packages done

    Webhooks, notifications, search, Actions and the web app

    Everything that reads the committed log: webhooks, notifications and search, the Actions engine and its runners, and the web app.

    Passing on a hosted test install

    • Webhooks: 12,325 of 12,325 deliveries arrived after a site was killed, none under a new GUID, and healthy hooks kept p99 1.07 s beside one timing out BF-57, BF-58
    • Notifications and search: an inbox rebuilt with read state kept, 10,000 stars listed, and EU documents kept in EU shards BF-5, BF-67, BF-68
    • Actions with an unmodified actions/runner: a killed runner's job fails at its lease end, required checks conclude after the commit path is killed, and secrets survive two key rotations while a fork pull request sees none BF-4, BF-56, BF-59
    • Merge queue workflows: a merge_group run on each queue ref built, and each ref deleted 10 s after its pull request landed or was ejected BF-37
    • A site killed during a 256-job matrix: runners polled the new leader's site 11.5 s later, and all 256 jobs ran exactly once BF-55
    • The web app under Playwright: every page renders with its first data in the HTML, and the live job log streams within 2 s and falls back when its relay dies BF-38, BF-61

    Still to pass

    • 2,500 runners polling one broker, past its cap BF-54
    • Hosted runners on Cloudflare Sandboxes, under a per-owner cap BF-60
    • The gate: an unmodified actions/runner runs a matrix while a site fails, with no job state lost BM-4
  7. M5Under way2 of 7 work packages done

    Release control and site days

    How a new version reaches an install: releases the customer signs, applied one site a day with the site drained first, and a pre-signed way back for every step.

    Passed

    • The release group, its genesis and its keys: a genesis signed by another key is refused by the members themselves (13 November) BF-41
    • The front door and pin-only pointers: 200,416 of 200,416 requests stayed pinned through a 601 s blackhole (14 November) BF-8

    Still to come

    • The graft-ops executor, site days with automatic drain, the change feed and export endpoints, and the revocation and rotation runbooks
    • The gate: a bad release commits nothing, and a vendor token revoked mid-roll returns every site to the last good version BM-5
  8. AM2PlannedAfter Milestone 5

    graft-mcp, the remote MCP server

    An MCP server at mcp.<host> with its own sign-in and a consent page per client. Its tools cover coordination and provenance, and every text field comes through the trust view.

    Gate

    • The MCP row passes, and a bad graft-mcp at 1% swaps back with the Cloudflare API blocked LM-2
  9. AM3PlannedAfter Milestone 5

    The controller and Graftyard-run agents

    Graftyard's own agents, off until an org turns them on. A controller in a separate account turns an assigned issue into a task, runs it in a Sandbox workspace whose network is closed by default, and keeps a session log for every run.

    Gate

    • An issue to a merged pull request end to end, in a fork and in the repository; a Sandbox killed mid-task resumes; a bad controller at 1% swaps back LM-3
  10. M6PlannedPlanned

    Region loss, SSH and the lifeboat

    Replacing a lost region's members with every account API token revoked, opt-in SSH through Cloudflare Tunnel, and the lifeboat: a read-only copy on the customer's own hosts that serves clones and reads when Cloudflare is down.

    Gate

    • One site fails for a day and its members are replaced with no new script version, every acknowledged write present, and gh pr list working against the lifeboat BM-6
  11. AM4PlannedAfter Milestone 6

    Insights, the predictor and Code Mode

    Conflict prediction between agents' branches before a pull request exists, insights behind an agents page, and Code Mode behind an org switch.

    Gate

    • Every textual conflict on current tips found, with at most 5% false conflicts LM-4
  12. M7PlannedPlanned, one package started early

    Breadth and modes

    Self-hosted installs and installs in the customer's own Cloudflare account, managed cell accounts, metadata child groups for very large repositories, a single-site evaluation mode, import from GHES, and whole-install restore.

    Started early

    • Metadata child groups, which AM1's 30-day storage run needs M7-WP4

    Gate

    • A self-hosted install with no vendor key passes the gates of Milestones 1 to 6, and two restores into new accounts lose no entry up to their cutoff BM-7

From build/BUILD-PLAN.md and the results files it cites, in the Graftyard repository. Each milestone closes only when every row of its gate meets its pass line, on a hosted install, unchanged. The labels after each line are the plan's row ids. All results lists every row with its numbers.

See the evidence