Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions changelog.d/security-no-fork-rehome-into-base-repo.md
Original file line number Diff line number Diff line change
@@ -0,0 +1 @@
- **Security:** contributor adoption no longer moves a fork pull request's commits into a `hive/adopt-<n>` branch of the base repository (supersede, DCO re-home, or in-place-stalled re-home). The re-homed PR was a same-repository PR, so the fork's unreviewed code ran with the repository's trusted CI and skipped its fork-only guards; one such run published and signed a canary image. A fork PR the hive cannot finish on the fork's own branch now parks at `stage/085-needs-decision` (`adoption-stalled:fork-rehome-refused`) with a maintainer ask. PRs from branches of the base repository are re-homed as before.
9 changes: 8 additions & 1 deletion src/docs/hive-merge.md
Original file line number Diff line number Diff line change
Expand Up @@ -598,7 +598,14 @@ sweep's own red-PR route: the superseding App PR it opens from `hive/adopt-<n>`
is red with the contributor's failures, lands in `ci-failing.json` like any
hive-authored red PR, and the scanner's FIX-BEFORE-NEW kick rebases and fixes
it on that branch (never a replacement PR), after which it follows the normal
`050 → 100` path. When adoption is in-place on the contributor PR's own
`050 → 100` path. A **fork** PR (or one whose head repository is unknown)
is never superseded or re-homed into `hive/adopt-<n>`: a branch in the base
repository makes a same-repository PR, which runs with the repository's
secrets and write tokens and skips its fork-only guards, so unreviewed fork
code would run as trusted. The hive records the refusal and the next pass
parks the PR at `085` as `adoption-stalled:fork-rehome-refused` with a
maintainer ask; in-place adoption on the fork's own branch (when the
contributor allows maintainer edits) is unchanged. When adoption is in-place on the contributor PR's own
branch, a changed head stays adopted only with hive-controlled evidence
(GitHub-linked App commit metadata; the stage store only remembers SHAs that
already passed that check);
Expand Down
2 changes: 1 addition & 1 deletion src/docs/labels-and-control-signals.md
Original file line number Diff line number Diff line change
Expand Up @@ -224,7 +224,7 @@ referenced it) a lone `stage/050-review-pending`
| `stage/070-fixing` | PR | hive-owned repair: App/taken-over PR red, changes requested (by the swarm, or by a trusted human's **changes requested** review / `/hive decision <text>` on the current head — reason `human-changes-requested`, which also kicks the author agent with the text verbatim, [#13041](https://github.com/hivecommons/hive/issues/13041)) or needs-rebase, **or** a non-App PR whose red head was caused by a hive-pushed commit that must be removed/re-homed; a hive agent is reworking it | 2h |
| `stage/071-needs-author` | PR | human-authored PR (`author/human`): CI red / changes requested / conflicting; the hive posted its diagnosis, the author must act. No nudge, no adoption — the person pushes or comments | 24h |
| `stage/072-returned-to-contributor` | PR | clanker PR (`author/clanker`, relay or external hive): CI red / changes requested / conflicting; the hive posted its diagnosis, the ball is in the contributor hive's court. (Briefly spelled `stage/072-needs-author` by [#12713](https://github.com/hivecommons/hive/pull/12713); the engine renames that label back in place) | 24h |
| `stage/074-adopted-by-hive` | PR | the author (a clanker) was unavailable — relay offline, never registered, or silent past the grace (see below); the hive took the PR over: in place on the PR's own branch when it may push there (the scanner's red-PR fix kick rebases and fixes it), otherwise via `hive/adopt-<n>` and a superseding App PR that keeps the author's commits (the original closes automatically when the superseder merges). Adoption is an **ownership** state, not a CI state: the PR reads 074 for as long as the hive owns it — through its fixes going pending, green, reviewed or red again (074 outranks 050/060/070/071/072) — and leaves 074 only at the merge gate (080/085/090/100), when the adoption is released, or to 085 `adoption-stalled:<reason>` when the head sat unchanged past the dwell with CI idle and a real blocker the hive cannot push past — a red optional check alone never counts, and the escalation releases back to 074 as soon as the reason clears (an in-place adoption with a blocking review finding re-homes to `hive/adopt-<n>` first instead; one with a squash-unsafe hive commit elects a merge commit on a `role/trusted` PR — `merge-method/merge` label, lane continues — and otherwise goes to 085 `adoption-stalled:hive-signoff-attribution` at once, without a supersede, naming why the election was refused). For in-place adoption, a changed head must be linked to the App bot by GitHub commit metadata (the stage store only remembers heads that already passed that check); otherwise the hive treats it as a contributor push and hands the PR back (072 if red, else the normal path). Forgeable git author/committer names and emails never keep 074. A PR whose author holds **admin** or **maintain** on the repository is never adopted or re-homed (one cached permission read per repo and author per hour; a failed read does not adopt), and an in-place adoption of such an author is released on the next pass. | 3h |
| `stage/074-adopted-by-hive` | PR | the author (a clanker) was unavailable — relay offline, never registered, or silent past the grace (see below); the hive took the PR over: in place on the PR's own branch when it may push there (the scanner's red-PR fix kick rebases and fixes it), otherwise via `hive/adopt-<n>` and a superseding App PR that keeps the author's commits (the original closes automatically when the superseder merges). Adoption is an **ownership** state, not a CI state: the PR reads 074 for as long as the hive owns it — through its fixes going pending, green, reviewed or red again (074 outranks 050/060/070/071/072) — and leaves 074 only at the merge gate (080/085/090/100), when the adoption is released, or to 085 `adoption-stalled:<reason>` when the head sat unchanged past the dwell with CI idle and a real blocker the hive cannot push past — a red optional check alone never counts, and the escalation releases back to 074 as soon as the reason clears (an in-place adoption with a blocking review finding re-homes to `hive/adopt-<n>` first instead; one with a squash-unsafe hive commit elects a merge commit on a `role/trusted` PR — `merge-method/merge` label, lane continues — and otherwise goes to 085 `adoption-stalled:hive-signoff-attribution` at once, without a supersede, naming why the election was refused). For in-place adoption, a changed head must be linked to the App bot by GitHub commit metadata (the stage store only remembers heads that already passed that check); otherwise the hive treats it as a contributor push and hands the PR back (072 if red, else the normal path). Forgeable git author/committer names and emails never keep 074. A PR whose author holds **admin** or **maintain** on the repository is never adopted or re-homed (one cached permission read per repo and author per hour; a failed read does not adopt), and an in-place adoption of such an author is released on the next pass. A fork PR is never superseded or re-homed into the base repository (that would run unreviewed fork code with the repository's trusted CI); it parks at `085` as `adoption-stalled:fork-rehome-refused` instead. | 3h |
| `stage/080-awaiting-merge-admission` | PR | nothing left for the author: CI green (no red check on the head, required or not), verdict approve, no modifiers — and the automerge lane has not admitted it yet (prefilter not evaluated for this head, labels not synced, lane tripped). A stall here means a lane problem, so the hive also posts one `<!-- hive:stage-stalled -->` comment with the lane's last skip reason | 30m |
| `stage/085-needs-decision` | PR | **PR-only** since the owner decision of 2026-10-17 (issues park at `stage/045-needs-decision`; before that this stage carried both, [#12683](https://github.com/hivecommons/hive/issues/12683)). A person owes a **decision**, not a merge: a `requires_human` swarm verdict (reason `reviewer-verdict:requires_human`), `sentinel-alert`, `hold` / `do-not-merge*` / `needs-decision` modifier, no reviewer configured, GitHub workflow approval is required for a trusted same-repo bot/App/clanker PR (`workflow-approval-required:<ctx>`); Hive has posted `/ok-to-test` and is waiting for the repo-local `hive-ok-to-test` workflow or a maintainer, or a human-only lane rule unrelated to trust (workflow-path edits, tests removed, human-merge path, reporter-trust hold), or an open issue the PR closes is held or parked (`linked-issue:hold:#N:<label>` / `linked-issue:needs-decision:#N:<label>`), or GitHub refused to update the branch from its base because the base changed `.github/workflows/` and the App lacks the workflows permission (`workflow-permission`: a person merges or rebases the base in and pushes, or grants the App Workflows: write). A trusted human's normal GitHub **Approve** review on the exact current head clears a `requires_human` verdict for any author kind/lane and records `reviewer-verdict:requires_human:human-approved@<login>`; a new head invalidates it. The same approval also clears the merge-lane review gate over any non-approved swarm outcome, a perspective's `changes_requested` included (the `review_evaluated` audit line records `human_approver` and `overrode`); without it, `changes_requested` / `reject` still require fixes or manual merge. Under `review.self_authored_decision_policy: resolve`, a Hive-authored PR's reviewer decision question parks here only when it matches a named risk class (`review.DecisionRiskClasses`: permissions/credentials, security, sensitive paths, data loss/irreversible, merge/trust policy, external side effects, licensing/legal, unverifiable) or carries `<!-- hive-decision-risk: high -->`; any other question is resolved by policy and recorded as an advisory reason on the approving aggregate. Hive-authored non-risk incomplete-fix reviews with concrete remediation are normalized to `changes_requested` at verdict load/aggregation time and by the stage-label detail selector, even when the finding was marked high severity; duplicate summary/changelog intent-alignment cleanup with a concrete trim/reword fix is included. Those PRs go to `stage/070-fixing`, not here. On a hive-authored PR the trusted human can also answer with a **changes requested** review on the head, or `/hive decision <what to change>`, which moves the PR to `stage/070-fixing` and kicks the author agent with the text verbatim ([#13041](https://github.com/hivecommons/hive/issues/13041)); `/hive decision` on a PR the hive did not open gets one reply naming the review-side equivalents. One `<!-- hive:stage-needs-decision -->` comment on entry names what is owed and, per lane, what clears it ([#13087](https://github.com/hivecommons/hive/issues/13087)): for a hive-authored `requires_human` park the two actions — approve the head (review or `/hive approve` reply after the verdict) to merge as-is, or request changes and the hive pushes the fix; for any other `requires_human` park a trusted merger's approval by review or `/hive approve` reply; for a hold / sentinel park the label a person must remove; for a linked-issue park the issue and the label to remove from it; and for a manual-merge-only lane rule that a maintainer must merge by hand, which no approval overrides | none |
| `stage/090-ready-for-human-merge` | PR | the hive would merge this itself but trust/role blocks it: reviewer verdict approve at the **current** head, required checks green (tide excluded), DCO yes, no sentinel alert, no conflicts, soak elapsed, no hold — and the only remaining lane reason is the author's tier/role. A maintainer's `/lgtm` `/approve` is the only missing step. One `<!-- hive:stage-ready-for-human-merge -->` comment on entry carries the evidence. Always read with the `role/*` and `author/*` labels. (Was `stage/090-needs-human` before [#12528](https://github.com/hivecommons/hive/issues/12528); the engine renames the GitHub label in place) | none |
Expand Down
30 changes: 30 additions & 0 deletions src/pkg/github/automerge/automerge_contributor_adopt.go
Original file line number Diff line number Diff line change
Expand Up @@ -469,6 +469,9 @@ func adoptionBlockerText(obs stageObservation, rec stagelabels.Record) string {
msg += ": " + q
}
return msg + "."
case rec.ForkRehomeRefused(obs.HeadSHA):
return fmt.Sprintf("this PR comes from a fork (`%s`) and the hive cannot finish it on the fork's own branch. Moving its commits to `%s%d` in this repository would run code no maintainer has reviewed with the repository's trusted CI (secrets, write tokens, no fork guards), so the hive did not. "+
"**Ask (maintainer):** review the change; if it should land, bring it into the repository yourself or follow the repository's process for fork pull requests, otherwise close it.", nonEmpty(obs.HeadRepo, "unknown"), adoptBranchPrefix, obs.Number)
case rec.AdoptionAttemptedNothingOpened(obs.HeadSHA):
return fmt.Sprintf("the hive tried to re-home this PR onto `%s%d` (`%s`) and could not open the superseding PR.", adoptBranchPrefix, obs.Number, rec.AdoptReason)
}
Expand Down Expand Up @@ -624,6 +627,14 @@ func (c *Engine) supersedeAdoptedPR(ctx context.Context, store *stagelabels.Stor
c.info("stage labels: not superseding a repository maintainer's PR; adoption released", "repo", obs.Display, "pr", obs.Number, "author", obs.Author, "reason", reason)
return 0, nil
}
if adoptionFromFork(obs) {
// Every supersede path (grace, offline, dco-rehome,
// in-place-stalled) ends here. The escalation pass parks the PR
// at 085 with adoptionBlockerText's maintainer ask.
store.MarkForkRehomeRefused(key, obs.HeadSHA)
c.info("stage labels: not re-homing a fork PR into the base repository", "repo", obs.Display, "pr", obs.Number, "head_repo", obs.HeadRepo, "head", obs.HeadSHA, "reason", reason)
return 0, nil
}
owner, repo, number := obs.Owner, obs.Repo, obs.Number
rec.AdoptReason = reason
if stored, ok := store.Get(key); ok && stored.AdoptDiffFence != "" {
Expand Down Expand Up @@ -815,6 +826,22 @@ func (c *Engine) authorMaintainsRepo(ctx context.Context, obs stageObservation,
return maintainer
}

// adoptStallForkRehomeRefused is the adoption-stalled suffix for a fork PR
// the hive would not move into the base repository.
const adoptStallForkRehomeRefused = "fork-rehome-refused"

// adoptionFromFork reports whether the PR's head is outside the base
// repository, or unknown (a deleted fork), which is treated the same.
// Superseding such a PR pushes its commits to a hive branch inside the base
// repository, where a pull_request run is a same-repository run: secrets,
// write tokens, and none of the repository's fork-only guards. Unreviewed
// fork code then runs as trusted (projectbluefin/utah-packages#609, re-homed
// as #626, published and signed a canary image), so the hive never does it;
// an in-place adoption on the fork's own branch stays a fork run.
func adoptionFromFork(obs stageObservation) bool {
return !strings.EqualFold(strings.TrimSpace(obs.HeadRepo), obs.Owner+"/"+obs.Repo)
}

// canPushToPRBranch reports whether the App can push to the PR's head: a
// head in the base repo always; a fork head only with "allow edits by
// maintainers", which the list payload omits, so one GET resolves it.
Expand Down Expand Up @@ -1086,6 +1113,9 @@ func adoptionStalledReason(obs stageObservation, rec stagelabels.Record, policy
if rec.AdoptedBy > 0 || obs.HeadSHA == "" {
return ""
}
if rec.ForkRehomeRefused(obs.HeadSHA) {
return stagelabels.LaneReasonAdoptionStalledPrefix + adoptStallForkRehomeRefused
}
if rec.AdoptionAttemptedNothingOpened(obs.HeadSHA) {
return stagelabels.LaneReasonAdoptionStalledPrefix + "supersede-failed:" + nonEmpty(rec.AdoptReason, obs.Reason)
}
Expand Down
Loading
Loading