You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Hive capability: safely update Utah fork branches when rebases include workflow changes #1611
Packages #812 remains conflicting. Hive reports that its rebase push is rejected because inherited workflow changes require the GitHub App workflows permission, then refuses to re-home the unreviewed fork into privileged CI. These are real security boundaries, but the permanent outcome must not be repeated manual rebasing/adoption.
Utah #1049 separately records that Hive cannot implement the accepted shared download-pinning workflow because its App token cannot write workflows.
Owner decision / solution
Provide a least-privilege, auditable workflow-update capability for the exact Utah repositories, or a separate trusted writer that validates the intended diff/head and preserves original authorship/DCO. Installation-wide permission expansion across unrelated repositories is not implicitly authorized by this issue. Preserve /test, LGTM exact-head binding and existing review controls.
Acceptance
Document actual installed App grants and scoped execution identity; existing secret names are not proof of a grant.
Hive can safely reconcile a reviewed fork branch with main without force-overwriting concurrent contributor work.
Unreviewed fork code is never executed with the writer's credentials.
Workflow-only inheritance and intentional workflow edits are distinguished in the review evidence.
Summary: The automated contributor that keeps these images in sync can no longer safely update a small set of supporting repositories when an automatic rebase pulls in changes to their "build and test" (CI) files. The automation's token is not allowed to edit those CI files, and force-pushing unreviewed changes into the build system is a deliberate safety boundary, not a bug. This issue is waiting on an owner decision about whether and how to give the automation a narrow, auditable way to make exactly those edits — so nothing more can be finished here until a person decides.
What I need from you: Approve (or decline) a narrow, auditable way for the automation to edit CI files in only the affected supporting repos during a rebase.
What the issue asks (from its body)
Owner decision: provide a least-privilege, auditable workflow-update capability for the exact Utah fork repositories, or a separate trusted writer that validates the intended diff/head and preserves original authorship and license record (DCO). Installation-wide permission expansion across unrelated repositories is explicitly not authorized.
Acceptance items: document the real installed App grants and execution identity; safely reconcile a reviewed fork branch with main without force-overwriting concurrent work; never run unreviewed fork code with the writer's credentials; distinguish inherited workflow edits from intentional ones in the review evidence; let PR Add systemd unit to keep uupd functional with hibernation #812 re-enter real exact-head CI without a manual push.
Why it cannot be finished in this repository (projectbluefin/common)
The rebase/writer capability lives in the factory automation (projectbluefin/actions, e.g. reusable-promote-squash.yml), which runs with contents: write + pull-requests: write but notworkflows: write. This shared-config repo documents and consumes that contract; per AGENTS.md it does not own the implementation.
The actual blocker is a GitHub App permission grant (the workflows scope) and the security boundary of running Hive-managed rebases with privileged CI — both are org-admin / credential decisions, not code that lives here.
Per this repo's secrets-policy, a new bot or new secret is a Design gate requiring maintainer approval, and the issue itself is framed as "Owner decision / solution".
The park reason was written before these changes. Re-read it against the current state before deciding. This note does not un-park the issue: it stays with a maintainer, and Hive updates this note if more of the referenced work moves.
Closing as completed: what this asked for now works wherever GitHub allows it, and the two cases that started it are gone (Danathar, maintainer, 11-19).
Branches in the repo itself: the hive can push workflow changes. On 11-19 its logs showed its trusted-tier tokens are issued with workflows: write. The update-branch 403s seen that day (projectbluefin/utah-packages#1173, projectbluefin/actions#752) were both on fork heads, not same-repo branches. This comes from the hive's token logs; the installation's grants weren't read directly.
Contributor forks: GitHub never lets an App push workflow files to a fork, whatever the App has been granted. So no narrow writer can be added for that case. The hive handles it instead:
It records the refusal as fork-workflow-update and hands the PR to its author, or to a maintainer when "Allow edits by maintainers" is on, to run gh pr update-branch or merge the base in.
Or the hive adopts the PR: it copies it to its own branch, keeps the original commits and sign-off, and the copy goes through normal review and CI.
Either way, unreviewed fork code never runs with the hive's credentials.
Not changing: the App's permissions. On 2026-11-18 the owner decided against widening them.
@hanthor, this was yours: if you hit a case this doesn't cover, reopen it with the PR link and we'll look at it fresh.
Activity
kylerankin commentedon Nov 2, 2026
Summary: The automated contributor that keeps these images in sync can no longer safely update a small set of supporting repositories when an automatic rebase pulls in changes to their "build and test" (CI) files. The automation's token is not allowed to edit those CI files, and force-pushing unreviewed changes into the build system is a deliberate safety boundary, not a bug. This issue is waiting on an owner decision about whether and how to give the automation a narrow, auditable way to make exactly those edits — so nothing more can be finished here until a person decides.
What I need from you: Approve (or decline) a narrow, auditable way for the automation to edit CI files in only the affected supporting repos during a rebase.
What the issue asks (from its body)
Why it cannot be finished in this repository (projectbluefin/common)
projectbluefin/actions, e.g.reusable-promote-squash.yml), which runs withcontents: write+pull-requests: writebut notworkflows: write. This shared-config repo documents and consumes that contract; per AGENTS.md it does not own the implementation.workflowsscope) and the security boundary of running Hive-managed rebases with privileged CI — both are org-admin / credential decisions, not code that lives here.Evidence
hive/488-vm-tools, still open): its rebase push is rejected because inherited CI-file changes need theworkflowspermission; refusing to re-home the unreviewed fork into privileged CI is the correct boundary.utahreferences exist anywhere in projectbluefin/common.— hive: backend=pi model=lemonade/ornith-1.5-35b-a3b-gguf-q6_k
🐝 Hive Agent:
contributor| SHA:e3d4cbbfix(adoption): name the branch update a person must run when a fork's…
hivecommons-hive commentedon Nov 13, 2026
The reason this issue is parked may be out of date. It has been parked for a maintainer (
stage/045-needs-decision) since 2026-11-02, and since then:The park reason was written before these changes. Re-read it against the current state before deciding. This note does not un-park the issue: it stays with a maintainer, and Hive updates this note if more of the referenced work moves.
Danathar commentedon Nov 20, 2026
Closing as completed: what this asked for now works wherever GitHub allows it, and the two cases that started it are gone (Danathar, maintainer, 11-19).
The two cases that started it:
Branches in the repo itself: the hive can push workflow changes. On 11-19 its logs showed its trusted-tier tokens are issued with
workflows: write. The update-branch 403s seen that day (projectbluefin/utah-packages#1173, projectbluefin/actions#752) were both on fork heads, not same-repo branches. This comes from the hive's token logs; the installation's grants weren't read directly.Contributor forks: GitHub never lets an App push workflow files to a fork, whatever the App has been granted. So no narrow writer can be added for that case. The hive handles it instead:
fork-workflow-updateand hands the PR to its author, or to a maintainer when "Allow edits by maintainers" is on, to rungh pr update-branchor merge the base in.Not changing: the App's permissions. On 2026-11-18 the owner decided against widening them.
@hanthor, this was yours: if you hit a case this doesn't cover, reopen it with the PR link and we'll look at it fresh.