Skip to content

Systems and records

Cross-repository engineering handoff: 1,038 patches, six states

By SourceX Editorial · Updated

10 sections

Short answer

A cross-repository engineering handoff record should say where a change came from, where it was sent, what state it is in, who owns the next step and what proves the outcome. OpenEmbedded-Core's 1,038 carried patches show the pattern: every one states its upstream status and 86.5 percent of submitted fixes link the submission, yet 72 of 135 still-pending patches have waited since at least 2020.

System export

Cross repository engineering

This page covers cross-repository engineering handoff: 1,038 patches, six states.Illustrative

Key takeaways

  • A handoff record needs six parts: origin, destination, state, a link that still resolves, an owner role and dated outcomes.
  • All 1,038 patches in OpenEmbedded-Core carry a valid Upstream-Status line, a field the build system has rejected as missing or malformed for core-layer patches since the yocto-5.0 release.
  • Links are strong where a handoff is claimed: 225 of 260 Submitted patches link the submission and 116 of 167 Backports link the exact upstream commit, while 129 of 135 Pending patches say nothing beyond the status word.
  • Open handoffs linger: 72 Pending and 46 Submitted patches kept the same status and place from April 2020 to October 2026.
  • The median patch touches one file, so a handoff can be licensed as a header, a diff and its linked records, without either repository.

What is a cross-repository engineering handoff, and what record must follow it?

A cross-repository engineering handoff happens when a change made in one codebase has to be accepted, finished or maintained in another that a different team owns. The record that follows it should name six things: where the change came from, where it was sent, what state the handoff is in, a link to the other side that still resolves, who owns the next step, and dated evidence of the outcome.

Inside a company this is routine. A product team patches a shared library it does not own and asks the platform team to take the fix; a vendored fork carries local changes that ought to go back to the vendor; a service moves into a new repository and its open issues follow it, or do not. Each crossing is a place where context drops: the reason stays in one tracker, the code lands in another, and a year later nobody can say whether the local patch is still needed.

For an AI team, these records can teach an agent to route a fix to the right owner, to notice when a carried patch has become redundant and to keep the reason attached as work crosses a boundary. This page measures how one large open-source project records such handoffs, then turns the result into a trace and an outcome rubric a buyer can put in a data request.

Why OpenEmbedded's patch headers are a fair public test

Because the project makes every handoff state explicit and enforces it. OpenEmbedded-Core, which its README describes as the layer containing the core metadata for current versions of OpenEmbedded, carries 1,038 patches across 254 recipe folders, each against an upstream project it does not control, and each patch must state in an Upstream-Status line where its handoff stands.

The Yocto Project's recipe style guide defines the values. Pending means 'No determination has been made yet, or patch has not yet been submitted to upstream.' Submitted means sent upstream and waiting for approval, optionally with where it went. Backport covers a change accepted upstream or pulled from a newer upstream version, with version information such as a commit ID. Denied means 'Not accepted by upstream, include reason in patch.' Inactive-Upstream marks a project with no activity for a long time, 'measured in years', and Inappropriate marks a patch that will not be sent upstream, with a reason in brackets.

The build system checks the line. In oe-core's QA code, a patch without a line matching one of those six values raises a patch-status error, reported as missing or malformed with a pointer to the guide. The check is absent from the yocto-3.1 and yocto-4.0 releases, turns into an error for core-layer patches at yocto-5.0, and now sits among the tests a layer must pass for Yocto Project Compatible status. Debian's DEP-3 patch guidelines define a comparable Forwarded field but call the format highly recommended and say that, until Debian policy endorses it, 'it will not be required.'

Where 1,038 handoffs stand, status by status

Every one of the 1,038 patches under meta/ at commit fa83da7, dated 26 October 2026, carries a valid status, and four in ten stay local by design. The rest record handoffs at different stages, and how much context each stage keeps varies sharply.

The 419 Inappropriate patches never leave: 216 cite something specific to OpenEmbedded or Yocto, 79 a build-environment or configuration need, and 42 give no reason. Submitted and Pending are open handoffs, Backport records work pulled from upstream, Denied a handoff upstream refused and Inactive-Upstream one with nobody left to receive it. Across all six, 233 status lines, 22.4 percent, carry nothing beyond the status word.

The mix has shifted over six years. At the yocto-3.1 release of April 2020, Pending was the largest status, 38.5 percent of 1,365 patches, and 34 lines were malformed; today Pending is 13.0 percent, Submitted has risen from 12.2 to 25.0 percent, and with the check in place none is malformed. Each status, with how much of the record survives in the line itself:

Where 1,038 handoffs stand, status by status
StatusPatchesShare of 1,038Give any detailLink in the status line
Inappropriate (stays local)41940.4%37716
Submitted (sent, awaiting answer)26025.0%242225 (86.5%)
Backport (pulled from upstream)16716.1%141125, of which 116 to an exact commit
Pending (not yet sent)13513.0%62
Inactive-Upstream (no receiver)393.8%261
Denied (refused upstream)181.7%139

Which handoffs keep a cross-reference a reader can follow?

The ones that claim to have crossed. Of 260 Submitted patches, 225 link the submission: 108 to a pull request, merge request or patch-review page, 59 to a mailing-list archive and 28 to a bug tracker. Of 167 Backports, 116 link the exact upstream commit and 10 more name a commit hash without a link. Pending patches, which have not crossed, almost never explain themselves: 6 of 135 give any detail.

Some cross-references point at a person rather than a record. Eleven Submitted patches give only an email address, which says whom to ask but not where the work went, and puts personal data in a field a dataset would have to strip. Eighteen more give no detail at all.

Links also depend on systems the project does not control. The 378 status lines with links (379 links) point to 56 hosts, and each link is only as durable as its host. A clang patch titled 'AsmMatcherEmitter: sort ClassInfo lists by name as well', written in November 2020 to make builds reproducible, has been marked Submitted with a link to review D97477 on reviews.llvm.org since at least the yocto-4.0 release, when copies sat under both the llvm and rust-llvm recipes. That site now says it 'hosts a static archive of the differential pages for historical purposes', and LLVM has moved its review and contribution system to GitHub. The handoff record still points at a review system its upstream retired.

How long do open handoffs stay open?

Often for years. Of the 135 patches Pending at commit fa83da7, 72 held the same file name and status in the same recipe folder at the yocto-3.1 release of April 2020 and at every release since; 46 of the 260 Submitted patches did the same.

Following handoffs forward tells the same story. Of 341 patches Pending at the yocto-4.0 release in April 2022, 208 had left the tree by October 2026, 91 were still Pending and 42 had changed status, 14 of them to Submitted. Of 213 Submitted patches, 141 had gone, 66 were still Submitted and only 2 had become Backports, the status that records upstream acceptance.

Removal is the most common ending and the least informative. A patch disappears when a version upgrade absorbs the fix, when upstream solves the problem another way or when the recipe is dropped, and the header cannot say which. The ages below are floors, because a patch that moved to another recipe or was renumbered, like the clang patch, counts as new under the matching rule:

How long do open handoffs stay open?
Same status sincePending at HEADSubmitted at HEAD
yocto-3.1, April 2020 or earlier7246
yocto-4.0, April 20221920
yocto-5.0, April 20241540
After yocto-5.029154
A flow diagram from the 341 Pending and 213 Submitted OpenEmbedded-Core patches of April 2022 to their state in October 2026: 349 removed with no recorded cause, 157 unchanged and 48 relabeled.
Where April 2022's open handoffs in OpenEmbedded-Core stood in October 2026. Counted from git history at yocto-4.0 and fa83da7 on 2026-10-27.

The handoff trace: fields a usable record needs

A usable record lets a reader who was never on either team see where the work came from, where it went, what state it is in, who owns it and how it ended. OpenEmbedded's headers supply some of those fields every time and others never.

The state field is complete because a tool enforces it; the outcome field is missing because nothing records why a patch left. The trace, with what the public record supplies and the check a buyer can run:

The handoff trace: fields a usable record needs
FieldWhat it holdsIn OpenEmbedded-CoreAcceptance check
OriginThe repository, team and change where the work startedThe patch file and recipe pathResolves to a commit in the origin repository
DestinationThe repository or team expected to take the workImplied by the recipe; named in 225 Submitted linksNamed explicitly, never inferred
StatePending, sent, accepted, refused, local by design or receiver goneValid on all 1,038 patches, enforced by a QA checkA value from a fixed list, checked by a tool
Cross-referenceLink to the review, issue or commit on the other side225 of 260 Submitted; 116 of 167 Backports to an exact commitLink resolves; host recorded; archive copy kept
ReasonWhy the work stays local, was refused or still waits377 of 419 Inappropriate give one; 6 of 135 Pending doRequired for local, refused and long-waiting states
OwnerWho carries the next step, as a roleA Signed-off-by line on 980 of 1,038 patchesRole kept; names and addresses removed
Dated historyWhen each state changedOnly in git history across releasesEvery state change carries a date
OutcomeWhat ended the handoffRemoval is visible; its cause is notRemoval tied to the upgrade or commit that ended it

Outcome rubric: seven labels for how a handoff ended

Seven labels cover the cases these records produce, and only one of them means the other side took the work. A case without the evidence a label requires takes a weaker label, down to unknown.

The labels, with the evidence each needs and how many OpenEmbedded patches earn it:

Outcome rubric: seven labels for how a handoff ended
LabelEvidence requiredIn OpenEmbedded-CoreUse for a buyer
Accepted upstream, linkedBackport status with a commit link116 of 167 BackportsCompleted handoffs to learn from
Sent, awaiting answerSubmitted with a resolving link225 of 260 SubmittedTracking tasks where age matters
Refused, with reasonDenied with a link or a stated reason13 of 18 DeniedWhy a receiver said no
Receiver goneInactive-Upstream with evidence of inactivity26 of 39 give any detailWhen to stop trying to hand off
Local by designInappropriate with a reason377 of 419What should never be sent upstream
StalledPending or Submitted unchanged for years72 Pending and 46 Submitted since April 2020 or earlierHandoffs that lost an owner
Ended, cause unknownPatch removed with no linked reason208 of 341 patches Pending in April 2022Unknown until the ending change is found

The strongest objection: a closed handoff may not close the problem

Correct, and the status line cannot settle it. A Backport says upstream accepted a change, not that the change fixed the problem that prompted it, and a removal says the local copy ended, not that users stopped hitting the bug. Status lines are also self-reported: nothing checks that a Submitted link still shows an open review or that a Pending patch was ever looked at.

Security fixes show why the link carries so much weight. Seventy-three patches carry a CVE tag, 64 of them are Backports and 48 link the exact upstream commit, which lets anyone confirm that the carried code matches what upstream released. Where the link is missing, a reader has to trust the label.

Three checks reduce the risk. Confirm acceptance from the upstream side, by commit, before calling a handoff complete. Tie every removal to the upgrade or change that ended it before assigning an outcome. And keep long-unchanged states under their own label, because a patch pending for six years is evidence about ownership, not about the fix.

How buyers would use handoff records without taking unrelated code

As routing and maintenance tasks with known answers. Give an agent a local patch and its upstream, ask whether to send it, keep it local or drop it after an upgrade, and grade the answer against the recorded status and later history. Stalled cases test whether an agent notices a handoff that has lost its owner, and Denied cases with links teach why a receiver says no.

The unit is small. The median OpenEmbedded patch touches one file, and 1,010 of 1,038 touch five or fewer. A handoff can therefore be licensed as its header, its diff and the linked records, without the rest of either repository, which matters when the origin side holds unrelated proprietary code or third-party code the supplier has no right to license.

Company histories hold the same structure under other names: issue links between repositories, issues transferred after a split, patches carried on vendored forks and internal packages waiting on a platform team. They also hold what open projects rarely link, the customer ticket or incident that started the work. Those records can contain customer names, internal hostnames and other teams' code, so rights review and de-identification come before anything is shared.

On its request page, SourceX says it follows up about potential sources once a buyer shares scope, volume, language, format, timing and licensing, because datasets are sourced on request rather than held in stock. For handoff records, add the states you need, the trace fields above and an acceptance test: for example, every accepted handoff resolves to an upstream commit, every long-waiting case is labeled stalled rather than open, and no email addresses survive in status fields. Whether a company's history can meet that test is decided by reviewing it.

Methods note: repository, snapshots, rules and limits

Repository: openembedded/openembedded-core, cloned from its GitHub mirror on 27 October 2026 at commit fa83da7 (26 October 2026), with the release tags yocto-3.1 (7 April 2020), yocto-4.0 (16 April 2022) and yocto-5.0 (18 April 2024) fetched at depth one. A patch is any file under meta/ ending .patch or .diff; the meta-selftest and meta-skeleton layers, 14 files at HEAD, are excluded. The status is the first line matching oe-core's own case-sensitive pattern; files with only a loose match count as malformed, and the obsolete value Accepted is reported separately. A link is any http or https run in the status detail, and a commit link contains a token of 7 to 40 hexadecimal characters with at least one letter and one digit.

Snapshots are matched on the first three path parts plus the file name, so a patch that moves between version-named subfolders of one recipe keeps its identity, but one that changes recipe or number does not; ages are therefore floors. Reason groups and link kinds come from keyword rules printed by the script. Shares come from integers, rounded once, half up. Not measured: whether links still resolve, acceptance confirmed from the upstream side, why removed patches were removed, and anything inside a company. The saved file holds paths, statuses, link hosts and flags, with no names or addresses.

Frequently asked questions

What is a cross-repository engineering handoff?

It is the point where a change made in one codebase has to be accepted, finished or maintained in another, usually owned by a different team. Examples include a fix sent to a library's maintainers, a patch carried on a vendored fork and issues moved when a repository is split.

What should a handoff record contain?

Six things: where the change came from, where it was sent, its current state, a link to the other side that still resolves, the owner of the next step as a role, and dated evidence of the outcome. Without the link and the dates, a reader cannot tell a finished handoff from a forgotten one.

How long do handoffs between projects stay open?

In OpenEmbedded-Core, often for years: 72 of the 135 patches marked Pending in October 2026 had the same status and place at the April 2020 release, and 46 of 260 Submitted patches did too. Both figures are floors, because moved or renamed patches count as new.

Can handoff records be licensed without the whole codebase?

Often, yes. The median OpenEmbedded patch touches one file, so a handoff can travel as its header, its diff and the linked records. Third-party code inside a diff and personal data in headers, such as email addresses, still need review before anything is licensed.

Does removing a carried patch mean upstream accepted the fix?

Not necessarily. A patch disappears when an upgrade absorbs the fix, when upstream solves the problem another way or when the recipe is dropped. Of 213 patches Submitted in April 2022, 141 had gone by October 2026, but only 2 had been relabeled as accepted Backports.

Sources

  • OpenEmbedded-Core carries 1,038 patches under meta/ at commit fa83da7, each with a valid Upstream-Status; 225 of 260 Submitted link the submission and 116 of 167 Backports an exact commit. Source
  • OpenEmbedded-Core describes itself as the layer containing the core metadata for current versions of OpenEmbedded. Source
  • oe-core's QA code requires an Upstream-Status line with one of six values and reports patches without one as missing or malformed. Source
  • The patch-status check is among the tests a layer must pass for Yocto Project Compatible status. Source
  • At the yocto-5.0 release, patch-status became an error for core-layer patches; the check is absent from yocto-3.1 and yocto-4.0. Source
  • Of 341 patches Pending and 213 Submitted at yocto-4.0 (April 2022), 208 and 141 had left the tree by October 2026, and 2 Submitted patches had become Backports. Source
  • At yocto-3.1 (April 2020), Pending was 38.5 percent of 1,365 patches and 34 status lines were malformed; 72 of today's Pending patches already held that status. Source
  • A clang reproducibility patch is marked Submitted with a link to review D97477 on reviews.llvm.org, unchanged since at least yocto-4.0. Source
  • The Yocto Project's style guide defines the Pending, Submitted, Backport, Denied, Inactive-Upstream and Inappropriate statuses. Source
  • Debian's DEP-3 patch format is highly recommended but not required until Debian policy endorses it. Source
  • LLVM's former Phabricator review site now hosts a static archive, and LLVM moved its review system to GitHub. Source
  • SourceX sources datasets on request rather than holding them in stock, and its request form asks for scope, volume, language, format, timing and licensing. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies