My summary of the Git Contributors' Summit 2026, was Re: Git v3.0 timeline, was Re: What's cooking in git.git (Sep 2026, #08)

Hi Junio,

On Tue, 22 Sep 2026, Junio C Hamano wrote:

I originally wrote this summary of the Git Contributor Summit only for my own records, but since you hinted at wanting some meeting notes, figured it might be interesting to other people, too. So here goes my distillation of the breakout-sessions from this year's Contributors' Summit. Many of these ideas still need discussion on the list; proposals below are not project-wide decisions.

Security mailing list and process

The security list is seeing a large influx of outside reports, apparently often AI-assisted or generated. Duplicates and reports outside Git's security model consume triage time, while some patches have waited for months. Waiting for an empty queue is not a workable release criterion: we need to release the fixes we have, rather than hold them indefinitely while more reports arrive. No fixed release cadence or guaranteed response time was settled.

Documenting Git's security model would save repeated explanations of what does and does not constitute a vulnerability. (Personal note, not discussed at the Summit: Stolee had tried to start a conversation about Git's security boundary a long time ago, but nobody replied. Maybe the AI onslaught will provide enough motivation to get that discussion going.) Moving non-security bugs to the public list more readily was also encouraged. A timeout after which public discussion would be presumed OK was proposed, but not agreed.

More company staffing would help, without turning this into an obligation for volunteers. Paying for dedicated help was discussed, with onboarding costs a concern. Using AI for triage or fixes raises confidentiality and DCO questions of its own. (Personal note: There seemed to be some sentiment in the room that contradicted the earlier agreed-on finding on the mailing list that using AI for triaging and for investigating wasn't a copyright concern and should therefore be considered permissible.)

The release bottleneck is not really tag automation. Merging and backporting fixes, preparing advisories, and handling CVEs take work, and too much of that knowledge lives in people's heads. Peff volunteered to start a public discussion of the process and dig up existing resources. (Personal note: I have done that merging, backporting, etc plenty of times, and I don't think that it is the bottleneck, and I was rather surprised to hear that it is complicated. Sure, there are the expected merge conflicts when merging `maint-*` branches, and running -- and fixing! -- CI on all of the tags in a private repository should go without saying, but that's all craft of the trade. Rather, the indecision and lack of engagement on the git-security mailing list is what I see as the blocker. I'd happily volunteer to juggle those branch thickets if that was truly the make-or-break issue here.)

Microsoft's release process reportedly needs about seven weeks. There were no objections in the room to proceeding without waiting for that schedule. (Personal note: That release process was misrepresented, which is surprising, as I coordinated two or three Git for Windows security bugfix releases _on the git-security list_ since the most recent Git security bugfix release, it's always the same thing: release on a second Tuesday of the month, three weeks before that the patches need to have settled, everybody goes home with a dependable timeline. It's not really that big of a deal. Testing patches, constructive feedback, these are the things that are missing and therefore blocking the process. I sensed a lot of finger-pointing in this discussion. I mean, I don't blame anybody for avoiding security work: it is stressful and intense. The responsibility for getting things wrong is enormous. I know that because I've done my share of that, probably more than most in the Git project, and I will do even more in the future. But when I don't have the time, or the energy, I am aware that I, myself, am the bottleneck; I don't need to blame others.)

Git v3.0

The proposed target is spring 2027: v2.56 in September 2026, v2.98 in December, then v2.99 and v3.0 next spring. The jump in version numbers is intended to signal the approaching breaking changes. March and April were both mentioned; the precise timing is not settled.

There was agreement on releasing v2.99.1 and v3.0 together, differing only in the BREAKING_CHANGES (v3.0 switches them on). That keeps the transition separate from another round of feature development. Nobody in the room objected to Rust becoming mandatory in v3.0. (Personal note: probably because Randall wasn't there, to say that NonStop support would be a blocker and that Git please wait.) A v2.x LTS remains an open question, with Gentoo mentioned as an interested party. (Personal note: I am still advocating for an in-tree Long Term Support branch, and since Junio indicated that he's less than eager to take care of that, I would love for Patrick Steinhardt to be the "LTS lieutenant", I vaguely remember that he said he'd do it if asked, and I trust his judgement, so I'd ask.)

SHA-256 support across the ecosystem had been a blocker. GitHub reported experimental support, with general availability expected around November. GitLab already had public, non-experimental support, and libgit2 supports it, too. JGit remains a gap; Google was not planning to fund that work.

The SHA-1/SHA-256 interoperability work, including historical tags, was reported to be implemented but not yet sent to the list. (Personal note: I think that the room seriously "mis-underestimated" the real-world impact of this. The code is not even on the Git mailing list, and the ramifications of not having a robust plan how to deal with partial clones or submodules or even signed tags strikes me as a dealbreaker. I would not be surprised if the decision to enforce SHA-256 as default would have to be revisited before v3.0, and possibly overturned.)

Documentation

Julia's work highlights the gap between documentation written by people who know Git inside out and users who do not yet know what objects, the index, or upstream mean. We need approachable learning material as well as reference documentation. Both need work; keeping manpages concise does not mean they cannot have better explanations and examples. (Personal note: I am beyond excited that Julia, whose work I have always admired, got interested in improving Git's documentation, which is in dear need of being improved, mainly because it does not cater to the majority of Git users out there who are unlikely to wander onto the Git mailing list, ever. I just hope that old-timers who really do not need the documentation nor understand the need of those who do need it show enough appreciation for the fresh views and for Julia's understanding of the target audience.)

Discoverability matters, too. The website (https://git-scm.com/) needs clearer entry points for learning Git, and existing guides are harder to find than manpages. Missing subsection links are another improvement we could make incrementally. (Personal note: Judging by the history of that site, I do wonder whether the core Git contributors are interested in helping this effort at all. For example, there are a growing number of PRs suggesting to add new UIs to the growing list, but I gave up reviewing them because I was the only one doing so.)

There was support for replacing outdated material and for merging useful improvements, then iterating, rather than trying to perfect everything before it lands. Bringing user feedback to the list without flooding it remains a challenge.

The current funding covers only 100 hours split between two people. Additional project and company funding was encouraged; brian, Emily, and Mark offered to explore company support. (Personal note: I had tried, back when GitHub still funded my team, to start something like that, without any success. To the contrary, even Git for Windows and Git Credential Manager got defunded.)

On the tooling side, using only Asciidoctor instead of maintaining both AsciiDoc and Asciidoctor support was proposed as a possible Git v3.0 change. Distribution support and rendering differences need checking, with doc-diff suggested for comparing the outputs. Patrick filed an issue during the discussion. (Personal note: AFAIU the AsciiDoc spec is now maintained by Asciidoctor, and I am aware already of one change that was made to the spec without adapting AsciiDoc accordingly. So the entire discussion might be quite moot already.)

Other ideas included richer diagrams for HTML while retaining text versions for manpages, and privacy-respecting traffic measurements to help prioritize documentation work. No diagram format was chosen, and caching and AI scraping complicate getting useful traffic data. Mermaid was proposed, and even GraphViz. (Personal note: I added support for Mermaid diagrams to https://git-scm.com/, but it turned out to be too limited, so I added GraphViz support. The support code for this is a bit of a beast, having a wasm version of GraphViz for development, pre-rendering the diagrams as SVG and as PDF during deployment of the site; it was quite a bit of fun to implement all that.)

Outreachy sponsoring

The goal is to support three interns in the round starting in early December, at $10,000 each. The corporate sponsorship previously provided by GitLab and GitHub has dried up, leaving Git itself to pay. There are company contacts to follow up with; Emily offered to ask Google's OSPO, without high expectations. No new sponsorship commitments were made. (Personal note: I don't think that these internships provide enough publicity to give companies much of an incentive to fund this. Which I find a bit of a shame, Outreachy in particular does a lot of important, good work, and if I wasn't so constantly overworked, I would want to mentor again; I always found it rewarding, even if I hold myself to a quite high bar which is quite draining.)

A related point for Git Merge 2027: announcing the location early would help Outreachy and GSoC interns plan attendance. No location was selected.

Pluggable object database

Patrick's pluggable object database is working, but it is not complete: commit-graph and multi-pack-index integration are still outstanding, and a repository extension is planned. (Personal note: It might be interesting to see whether implementing a storage backend is easier in core Git or in another Git-compatible implementation. JGit should be a natural target, having originated within BigTable-sized constraints, i.e. a different storage system, but funding seems to have dried up, there's not even SHA-256 support, so JGit might not be as hackable as it once was.)

Content-defined chunking prompted an important distinction between changing how objects are stored and changing the logical object model. Starting at the storage layer would let us preserve existing blob OIDs rather than require ecosystem-wide changes. A new pack/index format could provide another representation of the same object, much as deltas do today. No particular representation was agreed. (Personal note: It is curious to me why nobody tought about inventing a "meta blob", i.e. an object much like a tree object, except that it stitches together a larger blob. This would allow for the content-defined chunking that `rsync` already championed, way before Git was born! It would have allowed a Git native large file support worth writing home about, and could have replaced Git LFS. Xet (https://huggingface.co/docs/hub/xet/index) would not have had to be invented, and it would have allowed game development to move to Git. I can only imagine that the time it would cost to get even the first patches of this into core Git would be seen as prohibitive by any company who may have considered the effort.)

There is also an API question: does a backend seeing only object content have enough context to make good storage and delta choices, or should it receive richer information? More searchable tree storage and a Git "commit cloud" were other possibilities raised, not committed plans. (Personal note: At a previous GitMerge, Facebook presented their work, see e.g. https://github.com/facebook/sapling/blob/main/eden/mononoke/blobstore/packblob/README.md, which includes separating actual storage from transport. That is, already at push time, derived metadata is computed in async jobs which provide several potential deltas ready-to-go when a client clones or fetches. They reported clones with regular Git clients that are twice as fast, just because the server doesn't need to spend much compute on the data it sends. So there is a lot to be learned out there already.)

AI

The current SubmittingPatches policy is rooted in DCO certification and advice from SFC lawyers. The unresolved question is whether, and to what extent, contributors can certify AI-generated code. There was substantial disagreement about acceptable use, provenance and legal risks, community trust, review burden, and whether the current caution excludes useful tools. No policy change was agreed. (Personal note: You'd think that the opinion of lawyers is taken at face value, but no, it seems that some core Git contributors seem to disagree with the lawyers in favor of their own opinion...)

Several participants found language and proofreading assistance useful. A particular concern was submissions where the human does little more than relay agent output, leaving reviewers to deal with the consequences. The influx of poor GSoC contributions was one example. Attribution such as Assisted-by was suggested to make tool use clearer, but attribution alone does not answer the quality or DCO questions. (Personal note: my precedent of "Assisted-by" was called out as helpful, and I do think it is. I make a difference between AI-generated and AI-assisted. I'm not a fast typer, so I benefit a lot from being able to tell an LLM to please refactor out these four lines with the appropriate signature. There's not much creativity in there. I also like to let AI present me the call graphs for certain code locations, because due to the choice of C, which thanks to the C preprocessor is not easy to analyze statically, there are no competent tools other than LLMs that I can use for the task. I was highly surprised, though, to see how much enmity against AI in general was voiced, not by many, but many, many times, and how that contrasts with the Linux project which I hitherto had not considered to be as particularly open to modern practices.)

brian and Taylor agreed to put differing policy proposals on the list. The suggested process is to have alternatives examined by SFC counsel, make the risks clear, and then consider a vote. Emily volunteered to organize the voting procedure. Eligibility and the details remain open; Junio's authority as maintainer remains central. (Personal note: Since Taylor works for OpenAI now, I was not surprised by his stance, but brian works at GitHub, home of GitHub Copilot, and I am not sure how favorable their employer would look at their semi-public utterings about AI...)

Protocol v2 for pushes

There are concrete use cases now: repositories with millions of refs, including a reported 896 MB ref advertisement. Reftable improves ref update throughput, but does not by itself solve the advertisement problem. Nobody objected to push protocol v2, and the fetch-v2 infrastructure already provides much of the foundation.

The discussion covered advertising fewer refs, letting clients identify useful branches, and replacing large advertisements with a few rounds of push negotiation. Negotiation results could also help optimize the server's connectivity checks. Some improvements might be possible in the existing protocol before introducing v2.

We need to measure the tradeoffs rather than assume fewer bytes means faster pushes. One example involved a shallow push taking 35 seconds instead of two because of work to minimize the transfer. Shallow boundaries and unrelated histories complicate the proposed heuristics.

SHA-256 interoperability is another reason to want push v2: the current push protocol requires using the server's primary hash algorithm. Negotiation could make that more flexible.

Forge replication to thousands of mirrors would benefit from finding out cheaply whether refs have changed, rather than downloading full advertisements from every target. Checksums, ETag-like values, and reftable generation numbers were discussed, with concurrent updates complicating the picture.

Compact or compressed advertisements are also worth exploring and benchmarking, possibly reusing reftable's format. That does not mean sending the server's actual reftable, including hidden refs. There are several promising directions here, but no final design yet.

Some breakout sessions were planned, but apparently had to be cut.

Personal notes: I wasn't present for all of the sessions, in the afternoon I had other commitments; Therefore these notes (which AI assisted me in distilling) came partially from what I dictated and partially from the shared Google Document in which a few volunteers gracefully wrote notes. I found it challenging to connect as a remote participant. The link to the Google Meet, as well as control over the lobby thereof, seems to have been restricted to at most a few people, which might have contributed to the long waiting time before I was allowed in, and it definitely contributed to my comments not reaching the discussion in time to have an impact. I would have loved for Junio or the other two brave souls who also participated remotedly to have had more "air time". I am still a fan of the idea to have more frequent, smaller, virtual Contributor Summits, organized by a rotating cast. (Maybe I can get Emily to host the next one.) I was very happy that Junio was participating, as he _is_ the project lead, and in past Contributor Summits decisions were taken without him, which I found odd. Timing was really challenging for him, though, it was way past midnight for him. I'm all the more grateful that he did participate.

Final remark: This summary is obviously biased. I lightly edited it to separate better between my personal views and a hopefully unbiased account of what was discussed, and how, and by who. Nevertheless, I am but human. As a consequence, I would be delighted if other participants would share their summaries, so that my bias can be balanced out.

Ciao, Johannes