mirror of
https://github.com/github/gh-stack.git
synced 2026-09-14 20:26:28 +08:00
Sameen Karim
754d190490
* Fork unmerged branches into a new stack when the base stack is fully merged
Once every PR that is officially part of a stack on GitHub has been merged
-- especially after the merged branches are deleted upstream -- you can no
longer add to that stack. A new PR on top would target the trunk directly
instead of chaining onto the merged PRs, so the remote stack's "each PR's
base ref is the previous PR's head ref" invariant no longer holds. On the
next `gh stack submit`, the stack update was rejected and surfaced as a
confusing, dead-end warning:
Failed to update stack on GitHub: Pull requests must form a stack,
where each PR's base ref is the previous PR's head ref
`submit` had no handling for this: `syncStack` always sent the full PR list
(including the merged-and-deleted ones), so the API rejected the broken
chain even though the new PRs had already been created with correct bases.
Fork the survivors into a fresh stack instead of failing. After syncing PR
state and before pushing, `runSubmit` now calls `maybeForkFromMergedBase`:
- It triggers only when every PR officially part of the tracked remote
stack (`s.ID`) has merged. Membership is read from the stacks API, so
open PRs that are not part of the remote stack do not count, and -- this
is the key guard -- a normal partial, bottom-up merge (where the remote
stack still lists an open PR) is left completely untouched. A cheap
pre-check (the local stack must have at least one merged branch) avoids
an extra ListStacks call on the common path.
- The local branches are partitioned: those still in the merged remote
stack stay behind; everything else (new branches, plus open PRs that were
never part of that remote stack) is lifted into a brand-new stack rooted
at the original trunk, with an empty remote ID. The bottom survivor is
re-based onto the trunk.
- `runSubmit` continues with the new stack, so the push loop, PR creation,
and `syncStack` all operate on it; the empty ID routes `syncStack` through
the adopt/create path and a fresh stack is created on GitHub.
- The original, fully merged stack is left untouched on GitHub. Locally it
is kept as a record only if at least one of its branches still exists in
the working copy; otherwise it is dropped. No data is lost -- those PRs
are already merged on GitHub.
To restructure the stack file safely, add `StackFile.IndexOfStack`, which
locates a stack by pointer identity so the fork can capture what it needs
before `AddStack`/`RemoveStack` reallocate the underlying slice.
Also soften the partial-merge case that does not fork: when an `UpdateStack`
call fails with the "must form a stack" 422 and the stack still contains
merged branches, report it as an informational note (the unmerged PRs were
pushed and re-based onto the trunk) rather than a scary failure warning.
Scope is limited to `submit`. `add` and `checkout` keep their existing
"refuse and suggest `gh stack init`" behavior on fully merged stacks.
Tests:
- cmd/submit_test.go: TestSubmit_ForksWhenRemoteStackFullyMerged covers both
disposition variants (the old stack is removed when its merged branches
are gone locally, kept when they still exist) and asserts that only the
new branches are pushed, the fork message is printed, a fresh stack is
created, and the local stack file is split into two stacks.
TestSubmit_NoForkWhenRemoteStackHasOpenPR verifies the everyday bottom-up
merge is not forked and that the broken-chain 422 is reported calmly.
TestUpdateStack_BrokenChainAfterMerge checks the calm-vs-warn branch.
- internal/stack/stack_test.go: TestIndexOfStack covers identity lookup and
the not-found case.
Docs: README, the CLI reference, the stacked-PRs guide, the FAQ, and the
agent SKILL.md note that submitting onto a fully merged stack starts a new
stack rooted at the trunk.
* Handle fully merged stacks gracefully in the view and modify TUIs
Merged branches (and their PRs) are not selectable, so once an entire stack
has landed there is nothing to act on -- yet the TUIs did not reflect that:
- `gh stack view` still drew a highlighted cursor on the top branch even
though it could not be selected. Navigation, checkout, and the per-branch
toggles all silently did nothing, with no indication of why.
- `gh stack modify` opened its full editor on a stack with nothing left to
restructure, instead of short-circuiting like `gh stack submit` does when
there is nothing to submit.
Reflect the "nothing actionable" state in both TUIs.
View (internal/tui/stackview/model.go):
- Hide the cursor when every branch is merged. `New` now starts the cursor
at -1 and only lands it on the current or first non-merged branch; when
none exists the cursor stays hidden, so no row is rendered as focused. The
existing `m.cursor >= 0` guards and merged-skipping `moveCursor` already
make every cursor action a no-op in that state, and mouse-wheel scrolling
still works for tall merged stacks.
- Dim the shortcuts that depend on the cursor. `buildHeaderConfig` marks
navigate, commits, files, open PR, and checkout as `Disabled` (rendered
gray via the existing ShortcutEntry.Disabled styling) when all branches
are merged, leaving only `q quit` active.
Modify (cmd/modify.go):
- Short-circuit before opening the TUI. After preconditions pass and PR
state is synced, `runModify` now returns early when the stack is fully
merged, printing "All branches in this stack have been merged" and
pointing at `gh stack init`, exiting cleanly (exit 0) like submit's
"nothing to submit" path. The linearity and merge-queue precondition
checks already skip merged branches, so they do not fire spuriously.
Tests:
- internal/tui/stackview/model_test.go: the cursor is hidden (-1) when all
branches are merged; up/down/enter do not move it or trigger a checkout;
View renders without panicking on a hidden cursor; buildHeaderConfig
disables every cursor-dependent shortcut (and only those) when all merged,
and leaves them all enabled when active branches remain.
- cmd/modify_test.go: runModify short-circuits on a fully merged stack,
printing the message and returning no error without launching the TUI.