stakk graph renders the change graph: --format=pretty (default) for a human,
--format=json and --format=json-full as a schema-versioned document for a machine.
It is offline — jj only — so it needs no network or credentials.
Nothing in it has been near a forge: no PR numbers, titles, review or CI state,
and nothing about what a push would do. stakk learns both during stakk submit's plan phase,
and --dry-run prints them.
--format=json is sparse: identifiers, titles and states — enough to drive stakk submit.
--format=json-full adds commit bodies, authors and files, which dominate the byte count.
Sparse is a strict subset of full (same names, types, values), and both report the same schema_version.
schema_version— currently3; bumped on breaking schema changesdefault_branchremotes[]—name,url,host(lowercased, keeping the port of anhttp(s)URL),repo(owner/repo);hostandrepoarenullwhen the URL is not<host>/<owner>/<repo>, such as a local path. The document does not say which forge a host runs — that isstakk submit's decision, which is why there is noforgefieldexcluded_bookmarks[]— bookmarks left out because their history contains a merge, which the stacking model does not represent; stakk cannot manage theseexcluded_heads[]— unbookmarked heads left out for the same reason, each aschange_idandshort_change_id; a head has no bookmark to name itstacks[]— one per leaf
One entry per leaf, segments trunk-to-leaf. Shared ancestor segments are repeated in full in every stack containing them, so each entry is self-contained.
The order is not part of the contract.
It follows commit recency, so stacks[0] changes whenever anyone commits, and it is not the TUI's leaf numbering.
Identify a stack by its content; committer_timestamp is in the sparse projection
so "the most recently modified stack" can be computed rather than assumed.
A segment is a run of commits ending at a PR boundary.
bookmarks[]— one{name}per bookmark on the boundary commit; empty for an unbookmarked head. An object rather than a bare name, so a per-bookmark field can be added without renaming the listcommits[]— oldest first, matching the--bookmark-commandpayload
Several bookmarks on one commit are one boundary: two --keeps naming them is stakk::selection::duplicate_mark.
In both projections:
change_id,short_change_id— either works in the selection flags. The short form is jj's shortest prefix unique right now, so storechange_id, or a later run may fail withstakk::selection::rev_unresolvabletitle— first line of the commit message; empty when there is no description ((no description set)belongs to the pretty renderer, never to the document)committer_timestamp— offset-aware, e.g.2026-02-19T19:47:54+01:00; the key stack order is derived from. Compare instants, not strings: that value sorts after2026-02-19T19:00:00Zas text while being twelve minutes earlieris_immutable— cannot take a new bookmarklocal_bookmark_names[]— unfiltered; includes bookmarks the bookmarks revset excluded, unlike the segment'sbookmarks[]is_boundary— the newest commit of its segmentis_leaf— the tip of its stack
--format=json-full adds commit_id, description (the full message, title line included), author
(name, email, timestamp) and files[].
They are absent from sparse, not null.
Stack order follows committer_timestamp, not author.timestamp, because a rebase updates the former.
Field names, types and the sparse/full subset relationship are covered by schema_version.
The order of stacks[] and the pretty format are not.
remotes[].githubis renamed toremotes[].repoand is filled for every<host>/<owner>/<repo>URL, not only GitHub ones.remotes[].hostis added.- There is no top-level
forgefield: the graph is offline and does not select a forge. bookmarks[].remote_stateis removed. It was derived from jj alone and could not know which remote a submission pushes to, sosyncedmeant "on some remote"; what a push would do isstakk submit --dry-run's answer.excluded_head_countis replaced byexcluded_heads[], listing each head'schange_idandshort_change_idinstead of counting them.