fix(git): never discard committed local work in rebase_onto_base (#683)

The shared rebase primitive (dev sync_branch verb, PM/CEO rebase path,
submit-freshen, merge-conflict resolver) opened with fetch -> checkout ->
unconditional reset --hard origin/<head_branch>. The dirty-tree gate
protects uncommitted edits only; the reset silently rewound past every
committed-but-unpushed commit — and the commit do-verb never pushes, so
mid-rework a dev routinely has exactly that. The force-with-lease push
then republished the truncated branch as authoritative (the lease
matched the freshly-fetched, never-moved origin ref).

rebase_onto_base now classifies local vs origin/<head> post-fetch:
- behind/equal: reset --hard origin as before (origin loses nothing)
- strictly ahead: reset skipped — the rebase runs from the local tip and
  the lease'd push publishes the previously-doomed commits
- diverged: a patch-equivalence probe (rev-list --right-only
  --cherry-pick) first rescues the self-inflicted residue of a prior
  rebase whose force-push failed (treated as ahead, self-heals on
  retry); only genuine two-sided divergence returns a new
  {status: diverged, local_only, origin_only} — no reset, no rebase,
  no push, neither side silently discarded
- an absent local ref is recovered from origin (branch + checkout,
  never reset)

Callers: the sync_branch verb maps diverged to an actionable envelope
steering to i_am_blocked (stash-preserved note included); the
submit-freshen hard-rejects it like conflicts; the merge-conflict
resolver already escalates any non-rebased/superseded status to the
CEO and degrades gracefully (pinned by test, no code change).

New real-git suite (bare origin + clone, no subprocess mocking)
asserts origin-side outcomes: ahead-publishes, behind-adopts,
diverged-refuses-untouched, absent-ref recovery, superseded,
conflicts, and wedge self-heal on retry via a rejecting pre-receive
hook.

Co-authored-by: Renn F <rennf93@users.noreply.github.com>
This commit is contained in:
Renzo F
2026-07-24 14:34:25 +02:00
committed by GitHub
co-authored by Renn F
parent 3516d925fe
commit 71f5426e40
9 changed files with 744 additions and 63 deletions
+67 -12
View File
@@ -2704,8 +2704,10 @@ class Choreographer:
branch onto its base is safe; ``sync_task_branch`` pushes only the branch onto its base is safe; ``sync_task_branch`` pushes only the
HEAD branch (master/main are never written). Fail-open on probe/sync HEAD branch (master/main are never written). Fail-open on probe/sync
errors the PR/merge layer keeps its own behind checks but a rebase errors the PR/merge layer keeps its own behind checks but a rebase
CONFLICT is a hard reject naming the files, so the PM routes a CONFLICT is a hard reject naming the files, and a DIVERGED branch (the
conflict-resolution revision instead of re-submitting blind. PM's own clone and origin each carry unique commits) is likewise a
hard reject, so the PM routes a conflict-resolution revision instead
of re-submitting blind.
""" """
if not getattr(t, "branch_name", None) or not base_branch: if not getattr(t, "branch_name", None) or not base_branch:
return None return None
@@ -2723,7 +2725,34 @@ class Choreographer:
"assembled_freshen_sync_failed", task_id=str(t.id), error=str(exc) "assembled_freshen_sync_failed", task_id=str(t.id), error=str(exc)
) )
return None return None
if result.get("status") == "conflicts": reject = self._freshen_rejection_for(
result, behind=behind, base_branch=base_branch, verb=verb
)
if reject is not None:
return reject
logger.info(
"assembled_branch_freshened",
task_id=str(t.id),
verb=verb,
base_branch=base_branch,
behind=behind,
status=result.get("status"),
)
return None
@staticmethod
def _freshen_rejection_for(
result: dict[str, Any], *, behind: int, base_branch: str, verb: str
) -> Envelope | None:
"""Hard-reject shapes for ``_freshen_assembled_branch``'s sync result.
Conflicts and divergence both mean the auto-sync could not safely
reconcile the branch on its own extracted so the caller's
return-statement count stays under the PLR0911 budget (mirrors
``_sync_branch_preflight_rejection``).
"""
status = result.get("status")
if status == "conflicts":
files = ", ".join(result.get("files") or []) or "unknown files" files = ", ".join(result.get("files") or []) or "unknown files"
return Envelope.invalid_state( return Envelope.invalid_state(
message=( message=(
@@ -2740,13 +2769,22 @@ class Choreographer:
), ),
context_briefing={}, context_briefing={},
) )
logger.info( if status == "diverged":
"assembled_branch_freshened", return Envelope.invalid_state(
task_id=str(t.id), message=(
verb=verb, f"{verb} refused: the assembled branch has DIVERGED from "
base_branch=base_branch, f"its origin copy ({result.get('local_only', '?')} "
behind=behind, f"local-only commit(s), {result.get('origin_only', '?')} "
status=result.get("status"), "origin-only commit(s)) — this workspace and origin each "
"carry work the other lacks"
),
remediate=(
"a human must reconcile the two histories by hand (fetch, "
"inspect both tips, merge or rebase deliberately) before "
"re-submitting — auto-sync refuses to guess which side to "
"keep"
),
context_briefing={},
) )
return None return None
@@ -4614,11 +4652,28 @@ class Choreographer:
def _sync_branch_next_hint(t: Any, result: dict[str, Any]) -> str: def _sync_branch_next_hint(t: Any, result: dict[str, Any]) -> str:
"""Compute the ``next`` hint for a completed ``sync_branch`` run. """Compute the ``next`` hint for a completed ``sync_branch`` run.
Three shapes: a rebase conflict (files listed, stash noted if one was Four shapes: a genuine divergence (neither side touched, a human
taken), a clean rebase whose stash pop then conflicted (stash must reconcile), a rebase conflict (files listed, stash noted if one
was taken), a clean rebase whose stash pop then conflicted (stash
preserved, never dropped), or the plain spec default. preserved, never dropped), or the plain spec default.
""" """
status = str(result.get("status", "unknown")) status = str(result.get("status", "unknown"))
if status == "diverged":
hint = (
"sync_branch refused: your branch and its origin copy have "
f"DIVERGED ({result.get('local_only', '?')} commit(s) only "
f"in your workspace, {result.get('origin_only', '?')} only "
"on origin) — neither side was touched. escalate via "
"i_am_blocked(reason='...') so a human can reconcile the two "
"histories by hand"
)
if result.get("stash_pop_conflict"):
hint += (
" — your stashed changes were also popped back into a "
"conflict; the stash is preserved (not dropped), resolve "
"it by hand once the divergence is reconciled"
)
return hint
if status == "conflicts": if status == "conflicts":
hint = ( hint = (
f"sync_branch hit conflicts on {result.get('files', [])};" f"sync_branch hit conflicts on {result.get('files', [])};"
+136 -17
View File
@@ -5180,6 +5180,18 @@ class GitService(BaseService):
can now merge cleanly. can now merge cleanly.
- ``{"status": "conflicts", "files": [...]}`` the rebase hit - ``{"status": "conflicts", "files": [...]}`` the rebase hit
conflicts and was aborted; a developer must resolve by hand. conflicts and was aborted; a developer must resolve by hand.
- ``{"status": "diverged", "local_only": int, "origin_only": int}``
the local ``head_branch`` and ``origin/<head_branch>`` each
carry commits the other lacks with no patch-equivalent on the
other side. Most often this is the residue of a PRIOR call to
this same method that rebased locally but whose force-push then
failed (network blip, a flow-verb timeout kill, a container
reap between rebase and push) that case is recognized by
patch-equivalence and self-heals as "ahead" instead (see
:meth:`_reset_head_or_diverged`). What's left is a genuine
divergence, e.g. the task bounced to a different agent's clone
that pushed meanwhile. Refused outright: neither side is
touched and nothing is pushed a human must reconcile.
Any of the above may carry ``"stash_pop_conflict": True`` when Any of the above may carry ``"stash_pop_conflict": True`` when
``stash`` popped into a conflict (see below). ``stash`` popped into a conflict (see below).
@@ -5188,32 +5200,47 @@ class GitService(BaseService):
legitimate when it is the head's true merge target; the choreographer legitimate when it is the head's true merge target; the choreographer
refuses only a mis-resolved one. refuses only a mis-resolved one.
Safety gate (mirrors :meth:`pull`): refuses on a dirty worktree so the Safety gate (mirrors :meth:`pull`): refuses on a dirty worktree so
``git reset --hard`` below can't discard uncommitted agent edits — uncommitted agent edits are never discarded UNLESS ``stash=True``,
UNLESS ``stash=True``, in which case the dirty worktree (tracked + in which case the dirty worktree (tracked + untracked, ``-u``) is
untracked, ``-u``) is stashed first and popped back after the rebase stashed first and popped back after the rebase instead of refusing
instead of refusing outright (the dev-facing dead end this closes: outright (the dev-facing dead end this closes: DIRTY_WORKSPACE had no
DIRTY_WORKSPACE had no in-gate remedy other than a raw ``git`` the in-gate remedy other than a raw ``git`` the agent is denied). A pop
agent is denied). A pop conflict is never auto-resolved the stash conflict is never auto-resolved the stash is left in place (never
is left in place (never dropped) and the result gets dropped) and the result gets ``stash_pop_conflict: True`` so the
``stash_pop_conflict: True`` so the caller returns an actionable caller returns an actionable envelope; the agent's uncommitted work
envelope; the agent's uncommitted work is never lost. is never lost.
Beyond uncommitted edits, a COMMITTED local tip is never discarded
either. The ``commit`` do-verb never pushes, so mid-rework a dev
routinely has committed-but-unpushed commits on ``head_branch`` the
old unconditional ``reset --hard origin/<head_branch>`` right after
checkout silently rewound past them before the force-with-lease push
republished the truncated branch as authoritative. This now
classifies local vs ``origin/<head_branch>`` (post-fetch) first: an
absent local ref is recovered from origin (checkout, never reset
nothing local to discard); local behind-or-equal resets to origin as
before (origin has nothing to lose); local strictly ahead skips the
reset and rebases from the local tip instead (a superset the push
below publishes); a genuine divergence refuses via ``"diverged"``
rather than guessing which side to keep.
""" """
stashed = await self._stash_if_dirty(workspace, stash=stash) stashed = await self._stash_if_dirty(workspace, stash=stash)
await self._run_git(workspace, ["fetch", "origin"], token=git_token) await self._run_git(workspace, ["fetch", "origin"], token=git_token)
await self._ensure_local_head_ref(workspace, head_branch)
await self._run_git(workspace, ["checkout", head_branch]) await self._run_git(workspace, ["checkout", head_branch])
await self._run_git(workspace, ["reset", "--hard", f"origin/{head_branch}"]) diverged = await self._reset_head_or_diverged(workspace, head_branch)
if diverged is not None:
if stashed:
await self._pop_stash_into(workspace, diverged)
return diverged
rebase = await self._run_git( rebase = await self._run_git(
workspace, ["rebase", f"origin/{base_branch}"], check=False workspace, ["rebase", f"origin/{base_branch}"], check=False
) )
if rebase.returncode != 0: if rebase.returncode != 0:
return await self._abort_rebase_conflict(workspace, stashed=stashed) return await self._abort_rebase_conflict(workspace, stashed=stashed)
count = await self._run_git( unique = await self._rev_list_count(workspace, f"origin/{base_branch}..HEAD")
workspace,
["rev-list", "--count", f"origin/{base_branch}..HEAD"],
)
unique = int(count.stdout.strip() or "0")
if unique == 0: if unique == 0:
result: dict[str, Any] = {"status": "superseded"} result: dict[str, Any] = {"status": "superseded"}
else: else:
@@ -5227,6 +5254,97 @@ class GitService(BaseService):
await self._pop_stash_into(workspace, result) await self._pop_stash_into(workspace, result)
return result return result
async def _rev_list_count(self, workspace: Path, range_spec: str) -> int:
"""``git rev-list --count <range_spec>`` as an int (empty stdout → 0)."""
result = await self._run_git(workspace, ["rev-list", "--count", range_spec])
return int(result.stdout.strip() or "0")
async def _ensure_local_head_ref(self, workspace: Path, head_branch: str) -> None:
"""Recover an absent local ``head_branch`` ref from origin before checkout.
Mirrors :meth:`_assert_on_task_branch`'s recovery: worktree flows
normally guarantee the local ref already exists, but a caller running
against a bare clone root (or a re-provisioned workspace) may only
have the branch on ``origin`` (this runs post-fetch) create the
local ref (never reset one that already exists) so the unconditional
checkout right after this never fails on a missing branch.
"""
exists = await self._run_git(
workspace,
["rev-parse", "--verify", "--quiet", f"refs/heads/{head_branch}"],
check=False,
)
if exists.returncode != 0:
await self._run_git(
workspace, ["branch", head_branch, f"origin/{head_branch}"]
)
async def _reset_head_or_diverged(
self, workspace: Path, head_branch: str
) -> dict[str, Any] | None:
"""Classify checked-out ``head_branch`` against ``origin/<head_branch>``.
Resets local to origin when local carries nothing origin lacks
(behind or equal origin is authoritative, today's behavior).
Leaves the local tip untouched when it's strictly ahead
(committed-but-unpushed work a superset of origin the
force-with-lease push below will publish along with the rebase).
BOTH sides carrying unique commits by raw SHA isn't proof of a
genuine divergence: a prior run of this same method can rebase
locally and then have its force-push fail after, leaving local and
origin both non-empty forever on retry even though origin's tip is
just local's old history under new SHAs. :meth:`_origin_rewritten_locally`
tells the two apart by patch-equivalence and this self-heals as
"ahead" instead. Only a real two-sided divergence e.g. the task
bounced to a different agent's clone that pushed meanwhile — still
returns a ``{"status": "diverged", ...}`` dict; neither side is
silently discarded.
"""
local_only = await self._rev_list_count(
workspace, f"origin/{head_branch}..HEAD"
)
origin_only = await self._rev_list_count(
workspace, f"HEAD..origin/{head_branch}"
)
if local_only > 0 and origin_only > 0:
if await self._origin_rewritten_locally(workspace, head_branch):
return None
return {
"status": "diverged",
"local_only": local_only,
"origin_only": origin_only,
}
if local_only == 0:
await self._run_git(workspace, ["reset", "--hard", f"origin/{head_branch}"])
return None
async def _origin_rewritten_locally(
self, workspace: Path, head_branch: str
) -> bool:
"""True when every origin-only commit has a patch-equivalent local one.
Rescues the self-inflicted wedge from a rebase whose force-push
failed afterwards: local already carries origin's commits under new
SHAs, so a raw SHA rev-list count sees both sides positive forever.
``rev-list --cherry-pick --right-only`` drops any origin-only commit
whose patch (context-adjusted patch-id, same mechanism as
:meth:`unmerged_child_commits`'s ``git cherry``) matches one of
local's exclusive commits; whatever survives is truly exclusive to
origin, so a non-zero count still refuses as a real divergence.
"""
result = await self._run_git(
workspace,
[
"rev-list",
"--count",
"--right-only",
"--cherry-pick",
f"HEAD...origin/{head_branch}",
],
)
return int(result.stdout.strip() or "0") == 0
async def _stash_if_dirty(self, workspace: Path, *, stash: bool) -> bool: async def _stash_if_dirty(self, workspace: Path, *, stash: bool) -> bool:
"""Clean-tree gate for :meth:`rebase_onto_base`. """Clean-tree gate for :meth:`rebase_onto_base`.
@@ -5348,7 +5466,8 @@ class GitService(BaseService):
rebase through the dev ``sync_branch`` verb instead of the CEO/PM-only rebase through the dev ``sync_branch`` verb instead of the CEO/PM-only
``/rebase`` HTTP route. Mirrors ``rebase_pr_for_task``'s workspace/token ``/rebase`` HTTP route. Mirrors ``rebase_pr_for_task``'s workspace/token
resolution and delegates to :meth:`rebase_onto_base`, returning the same resolution and delegates to :meth:`rebase_onto_base`, returning the same
classification dict (``rebased`` / ``superseded`` / ``conflicts``). classification dict (``rebased`` / ``superseded`` / ``conflicts`` /
``diverged``).
``stash`` forwards to :meth:`rebase_onto_base` auto-stash a dirty ``stash`` forwards to :meth:`rebase_onto_base` auto-stash a dirty
worktree instead of refusing DIRTY_WORKSPACE. worktree instead of refusing DIRTY_WORKSPACE.
@@ -85,6 +85,27 @@ async def test_freshen_conflicts_reject_with_files() -> None:
assert "stats.json" in body["message"] assert "stats.json" in body["message"]
@pytest.mark.asyncio
async def test_freshen_diverged_rejects() -> None:
"""A diverged branch is refused just like a conflict — never guessed at."""
git = AsyncMock()
git.is_behind_base.return_value = (2, 3)
git.sync_task_branch.return_value = {
"status": "diverged",
"local_only": 1,
"origin_only": 2,
}
c = Choreographer(_make_deps(git=git))
env = await c._freshen_assembled_branch(
_cell_task(), base_branch="feature/main_pm/root", verb="submit_up"
)
assert env is not None
body = env.as_dict()
assert body["error"] == "invalid_state"
assert "DIVERGED" in body["message"]
assert "reconcile" in body["remediate"]
@pytest.mark.asyncio @pytest.mark.asyncio
async def test_freshen_fails_open_on_probe_error() -> None: async def test_freshen_fails_open_on_probe_error() -> None:
git = AsyncMock() git = AsyncMock()
@@ -148,6 +148,38 @@ async def test_genuine_conflict_escalates_to_ceo_and_does_not_loop(
assert env.error is None assert env.error is None
@pytest.mark.asyncio
async def test_diverged_rebase_outcome_escalates_rather_than_completing(
monkeypatch: pytest.MonkeyPatch,
) -> None:
"""A diverged rebase (local and origin each carry unique commits) must
escalate like any other non-rebased/non-superseded outcome — never
silently complete or re-merge a branch that could still be missing one
side's work."""
git = AsyncMock()
git.rebase_pr_for_task = AsyncMock(
return_value={"status": "diverged", "local_only": 1, "origin_only": 2}
)
git.close_pull_request = AsyncMock()
git.pr_merge = AsyncMock()
task = AsyncMock()
task.admin_set_status = AsyncMock()
task.get = AsyncMock(return_value=MagicMock(status="awaiting_ceo_approval"))
choreo = _choreo(task, git, monkeypatch)
t = MagicMock(
pr_number=161, project_id=uuid4(), parent_task_id=None, team="backend"
)
await choreo._resolve_merge_conflict_on_complete(
uuid4(), uuid4(), t, "feature/backend/root--cell", "notes", _EXC
)
task.admin_set_status.assert_awaited_once()
task.cell_pm_complete.assert_not_awaited()
git.close_pull_request.assert_not_awaited()
git.pr_merge.assert_not_awaited()
@pytest.mark.asyncio @pytest.mark.asyncio
async def test_unknown_rebase_outcome_escalates_rather_than_completing( async def test_unknown_rebase_outcome_escalates_rather_than_completing(
monkeypatch: pytest.MonkeyPatch, monkeypatch: pytest.MonkeyPatch,
+67
View File
@@ -131,6 +131,73 @@ async def test_sync_branch_conflicts_aborts_and_steers_to_resolve() -> None:
assert "sync_branch again" in env.next assert "sync_branch again" in env.next
@pytest.mark.asyncio
async def test_sync_branch_diverged_steers_to_i_am_blocked() -> None:
"""A diverged branch is not an error — it's a clean ok-envelope hint to
escalate, mirroring how a conflict is surfaced (git-only op, no DB
transition either way)."""
aid = uuid4()
tid = uuid4()
t = _task(tid=tid, aid=aid)
task_svc = AsyncMock()
task_svc.get.return_value = t
task_svc.agent_for.return_value = MagicMock(role="developer", team="backend")
git_svc = AsyncMock()
git_svc.sync_task_branch.return_value = {
"status": "diverged",
"local_only": 2,
"origin_only": 1,
}
deps = _make_deps(task=task_svc, git=git_svc)
c = Choreographer(deps)
with patch(
"roboco.services.gateway.choreographer._impl.resolve_parent_branch",
new=AsyncMock(return_value=_BASE),
):
env = await c.sync_branch(aid, tid)
assert env.error is None
assert env.next is not None
assert "DIVERGED" in env.next
assert "i_am_blocked" in env.next
@pytest.mark.asyncio
async def test_sync_branch_diverged_with_stash_pop_conflict_notes_preserved_stash() -> (
None
):
"""A diverged result can ALSO carry a stash-pop conflict (the stash was
taken, the rebase never touched either side, but popping it back still
collided) — the hint must not drop that warning."""
aid = uuid4()
tid = uuid4()
t = _task(tid=tid, aid=aid)
task_svc = AsyncMock()
task_svc.get.return_value = t
task_svc.agent_for.return_value = MagicMock(role="developer", team="backend")
git_svc = AsyncMock()
git_svc.sync_task_branch.return_value = {
"status": "diverged",
"local_only": 2,
"origin_only": 1,
"stash_pop_conflict": True,
}
deps = _make_deps(task=task_svc, git=git_svc)
c = Choreographer(deps)
with patch(
"roboco.services.gateway.choreographer._impl.resolve_parent_branch",
new=AsyncMock(return_value=_BASE),
):
env = await c.sync_branch(aid, tid, stash=True)
assert env.error is None
assert env.next is not None
assert "DIVERGED" in env.next
assert "preserved" in env.next.lower()
@pytest.mark.asyncio @pytest.mark.asyncio
async def test_sync_branch_not_found_for_unknown_task() -> None: async def test_sync_branch_not_found_for_unknown_task() -> None:
aid = uuid4() aid = uuid4()
+8 -2
View File
@@ -1463,7 +1463,10 @@ async def test_rebase_onto_base_stash_true_auto_stashes_and_pops() -> None:
if args[:2] == ["status", "--porcelain"]: if args[:2] == ["status", "--porcelain"]:
res.stdout = " M dirty.py\n" res.stdout = " M dirty.py\n"
elif args[:2] == ["rev-list", "--count"]: elif args[:2] == ["rev-list", "--count"]:
res.stdout = "1" # Only the post-rebase unique-vs-base count is non-zero; the
# pre-rebase local-vs-origin(HEAD) classification reads as
# "nothing unique on either side" (behind/equal, not diverged).
res.stdout = "1" if args[2] == "origin/master..HEAD" else "0"
else: else:
res.stdout = "" res.stdout = ""
return res return res
@@ -1499,7 +1502,10 @@ async def test_rebase_onto_base_stash_pop_conflict_preserves_stash() -> None:
if args[:2] == ["status", "--porcelain"]: if args[:2] == ["status", "--porcelain"]:
res.stdout = " M dirty.py\n" res.stdout = " M dirty.py\n"
elif args[:2] == ["rev-list", "--count"]: elif args[:2] == ["rev-list", "--count"]:
res.stdout = "1" # Only the post-rebase unique-vs-base count is non-zero; the
# pre-rebase local-vs-origin(HEAD) classification reads as
# "nothing unique on either side" (behind/equal, not diverged).
res.stdout = "1" if args[2] == "origin/master..HEAD" else "0"
elif args == ["stash", "pop"]: elif args == ["stash", "pop"]:
res.returncode = 1 # pop conflicted — stash is NOT dropped by git res.returncode = 1 # pop conflicted — stash is NOT dropped by git
else: else:
+99 -30
View File
@@ -94,21 +94,27 @@ async def test_success_path_returns_rebased_and_does_not_call_abort(
Call sequence for the success path (rebase OK, 2 unique commits): Call sequence for the success path (rebase OK, 2 unique commits):
[0] status --porcelain clean (the H8 dirty-tree gate) [0] status --porcelain clean (the H8 dirty-tree gate)
[1] fetch origin [1] fetch origin
[2] checkout HEAD branch [2] rev-parse --verify --quiet local HEAD ref exists
[3] reset --hard origin/HEAD [3] checkout HEAD branch
[4] rebase origin/BASE exits 0 [4] rev-list --count origin/H..HEAD local_only=0
[5] rev-list --count returns "2" [5] rev-list --count HEAD..origin/H origin_only=0 (not diverged)
[6] push --force-with-lease pushes the rebased branch [6] reset --hard origin/HEAD local has nothing unique
[7] rebase origin/BASE exits 0
[8] rev-list --count returns "2"
[9] push --force-with-lease pushes the rebased branch
""" """
run = AsyncMock( run = AsyncMock(
side_effect=[ side_effect=[
_result(stdout=""), # [0] status --porcelain → clean _result(stdout=""), # [0] status --porcelain → clean
_result(), # [1] fetch _result(), # [1] fetch
_result(), # [2] checkout _result(returncode=0), # [2] rev-parse --verify (local ref exists)
_result(), # [3] reset _result(), # [3] checkout
_result(), # [4] rebase ← success _result(stdout="0\n"), # [4] rev-list origin/H..HEAD → local_only=0
_result(stdout="2\n"), # [5] rev-list _result(stdout="0\n"), # [5] rev-list HEAD..origin/H → origin_only=0
_result(), # [6] push _result(), # [6] reset (local not ahead → reset)
_result(), # [7] rebase ← success
_result(stdout="2\n"), # [8] rev-list origin/BASE..HEAD
_result(), # [9] push
] ]
) )
monkeypatch.setattr(GitService, "_run_git", run) monkeypatch.setattr(GitService, "_run_git", run)
@@ -148,21 +154,27 @@ async def test_conflict_path_calls_diff_then_abort_and_returns_conflict_files(
Call sequence: Call sequence:
[0] status --porcelain clean (the H8 dirty-tree gate) [0] status --porcelain clean (the H8 dirty-tree gate)
[1] fetch origin [1] fetch origin
[2] checkout HEAD branch [2] rev-parse --verify --quiet local HEAD ref exists
[3] reset --hard origin/HEAD [3] checkout HEAD branch
[4] rebase origin/BASE exits 1 (conflict) [4] rev-list --count origin/H..HEAD local_only=0
[5] diff --name-only lists conflicted files [5] rev-list --count HEAD..origin/H origin_only=0 (not diverged)
[6] rebase --abort exits 0 [6] reset --hard origin/HEAD local has nothing unique
[7] rebase origin/BASE exits 1 (conflict)
[8] diff --name-only lists conflicted files
[9] rebase --abort exits 0
""" """
run = AsyncMock( run = AsyncMock(
side_effect=[ side_effect=[
_result(stdout=""), # [0] status --porcelain → clean _result(stdout=""), # [0] status --porcelain → clean
_result(), # [1] fetch _result(), # [1] fetch
_result(), # [2] checkout _result(returncode=0), # [2] rev-parse --verify (local ref exists)
_result(), # [3] reset _result(), # [3] checkout
_result(returncode=1), # [4] rebase ← conflict _result(stdout="0\n"), # [4] rev-list origin/H..HEAD → local_only=0
_result(stdout="src/a.py\nsrc/b.py\n"), # [5] diff _result(stdout="0\n"), # [5] rev-list HEAD..origin/H → origin_only=0
_result(), # [6] rebase --abort _result(), # [6] reset (local not ahead → reset)
_result(returncode=1), # [7] rebase ← conflict
_result(stdout="src/a.py\nsrc/b.py\n"), # [8] diff
_result(), # [9] rebase --abort
] ]
) )
monkeypatch.setattr(GitService, "_run_git", run) monkeypatch.setattr(GitService, "_run_git", run)
@@ -214,21 +226,27 @@ async def test_resilience_when_both_rebase_and_abort_fail_returns_conflict_no_ex
Call sequence: Call sequence:
[0] status --porcelain clean (the H8 dirty-tree gate) [0] status --porcelain clean (the H8 dirty-tree gate)
[1] fetch origin [1] fetch origin
[2] checkout HEAD branch [2] rev-parse --verify --quiet local HEAD ref exists
[3] reset --hard origin/HEAD [3] checkout HEAD branch
[4] rebase origin/BASE exits 1 (conflict) [4] rev-list --count origin/H..HEAD local_only=0
[5] diff --name-only lists conflicted files [5] rev-list --count HEAD..origin/H origin_only=0 (not diverged)
[6] rebase --abort exits 1 (abort also fails) [6] reset --hard origin/HEAD local has nothing unique
[7] rebase origin/BASE exits 1 (conflict)
[8] diff --name-only lists conflicted files
[9] rebase --abort exits 1 (abort also fails)
""" """
run = AsyncMock( run = AsyncMock(
side_effect=[ side_effect=[
_result(stdout=""), # [0] status --porcelain → clean _result(stdout=""), # [0] status --porcelain → clean
_result(), # [1] fetch _result(), # [1] fetch
_result(), # [2] checkout _result(returncode=0), # [2] rev-parse --verify (local ref exists)
_result(), # [3] reset _result(), # [3] checkout
_result(returncode=1), # [4] rebase ← conflict _result(stdout="0\n"), # [4] rev-list origin/H..HEAD → local_only=0
_result(stdout="src/conflict.py\n"), # [5] diff _result(stdout="0\n"), # [5] rev-list HEAD..origin/H → origin_only=0
_result(returncode=1), # [6] rebase --abort ← also fails _result(), # [6] reset (local not ahead → reset)
_result(returncode=1), # [7] rebase ← conflict
_result(stdout="src/conflict.py\n"), # [8] diff
_result(returncode=1), # [9] rebase --abort ← also fails
] ]
) )
monkeypatch.setattr(GitService, "_run_git", run) monkeypatch.setattr(GitService, "_run_git", run)
@@ -246,6 +264,57 @@ async def test_resilience_when_both_rebase_and_abort_fail_returns_conflict_no_ex
assert result == {"status": "conflicts", "files": ["src/conflict.py"]} assert result == {"status": "conflicts", "files": ["src/conflict.py"]}
# ---------------------------------------------------------------------------
# Test 4 — diverged: both sides carry unique commits, neither is touched
# ---------------------------------------------------------------------------
@pytest.mark.asyncio
async def test_diverged_path_returns_diverged_and_touches_nothing(
monkeypatch: pytest.MonkeyPatch,
) -> None:
"""When local HEAD and origin/HEAD each carry commits the other lacks
AND the patch-equivalence probe finds a real origin-only commit (no
local match), the method refuses immediately no reset, no rebase, no
push.
Call sequence:
[0] status --porcelain clean (the H8 dirty-tree gate)
[1] fetch origin
[2] rev-parse --verify --quiet local HEAD ref exists
[3] checkout HEAD branch
[4] rev-list --count origin/H..HEAD local_only=2
[5] rev-list --count HEAD..origin/H origin_only=1
[6] rev-list --count --right-only --cherry-pick HEAD...origin/H
1 origin-only commit survives patch-equivalence DIVERGED
"""
run = AsyncMock(
side_effect=[
_result(stdout=""), # [0] status --porcelain → clean
_result(), # [1] fetch
_result(returncode=0), # [2] rev-parse --verify (local ref exists)
_result(), # [3] checkout
_result(stdout="2\n"), # [4] rev-list origin/H..HEAD → local_only=2
_result(stdout="1\n"), # [5] rev-list HEAD..origin/H → origin_only=1
_result(stdout="1\n"), # [6] cherry-pick probe → 1 genuinely unmatched
]
)
monkeypatch.setattr(GitService, "_run_git", run)
svc = _git_service()
result = await svc.rebase_onto_base(
_WORKSPACE,
head_branch=_HEAD,
base_branch=_BASE,
git_token=_TOKEN,
)
assert result == {"status": "diverged", "local_only": 2, "origin_only": 1}
# Nothing past the classification ran: no reset, no rebase, no push.
subcommands = {c.args[1][0] for c in run.call_args_list}
assert subcommands.isdisjoint({"reset", "rebase", "push"})
# --------------------------------------------------------------------------- # ---------------------------------------------------------------------------
# Safety gate tests for rebase() — protected-branch guard # Safety gate tests for rebase() — protected-branch guard
# --------------------------------------------------------------------------- # ---------------------------------------------------------------------------
@@ -0,0 +1,305 @@
"""Real-git regression tests for the committed-work-preserving reconciliation
in ``rebase_onto_base``.
The historical bug (live, struck repeatedly): the preamble ran an
unconditional ``git reset --hard origin/<head_branch>`` right after checkout,
which rewound the local branch to the last-PUSHED tip silently discarding
every committed-but-unpushed commit, since the ``commit`` do-verb never
pushes. The subsequent ``push --force-with-lease`` then succeeded (the lease
matches the freshly-fetched origin ref, which never moved) and republished
the truncated branch as authoritative.
These run against a REAL bare origin + real clones (no mocked ``_run_git``)
the bug lived exactly in real-git ref semantics, and a mock can't
reproduce a fast-forward/divergence classification bug like this one.
"""
from __future__ import annotations
import subprocess
from typing import TYPE_CHECKING
from unittest.mock import MagicMock
import pytest
from roboco.exceptions import GitCommandError
from roboco.services.git import GitService
if TYPE_CHECKING:
from pathlib import Path
_HEAD = "feature/backend/task"
_BASE = "master"
def _git(repo: Path, *args: str) -> subprocess.CompletedProcess[str]:
return subprocess.run(
["git", *args], cwd=repo, check=True, capture_output=True, text=True
)
def _git_ok(repo: Path, *args: str) -> subprocess.CompletedProcess[str]:
"""Non-raising variant for existence probes."""
return subprocess.run(
["git", *args], cwd=repo, check=False, capture_output=True, text=True
)
def _init_bare(path: Path) -> None:
path.mkdir(parents=True, exist_ok=True)
subprocess.run(
["git", "init", "--bare", "--initial-branch=master", str(path)],
check=True,
capture_output=True,
)
def _configure(repo: Path) -> None:
_git(repo, "config", "user.email", "t@example.com")
_git(repo, "config", "user.name", "T")
_git(repo, "config", "commit.gpgsign", "false")
def _clone(origin: Path, dest: Path) -> None:
subprocess.run(
["git", "clone", str(origin), str(dest)], check=True, capture_output=True
)
_configure(dest)
def _commit(repo: Path, name: str, content: str) -> None:
(repo / name).write_text(content)
_git(repo, "add", name)
_git(repo, "commit", "-m", f"add {name}")
def _log_subjects(repo: Path, ref: str) -> list[str]:
out = _git(repo, "log", "--format=%s", ref)
return out.stdout.splitlines()
def _rev(repo: Path, ref: str) -> str:
return _git(repo, "rev-parse", ref).stdout.strip()
def _service() -> GitService:
svc = GitService.__new__(GitService)
svc.log = MagicMock()
svc.session = MagicMock()
return svc
@pytest.fixture
def repo_pair(tmp_path: Path) -> tuple[Path, Path]:
"""A bare origin plus a working clone, both carrying a pushed task branch
(``master`` with one root commit, ``_HEAD`` branched off it with one more)."""
origin = tmp_path / "origin.git"
_init_bare(origin)
seed = tmp_path / "seed"
_clone(origin, seed)
_commit(seed, "README.md", "root\n")
_git(seed, "push", "origin", "master")
_git(seed, "checkout", "-b", _HEAD)
_commit(seed, "feature.py", "v1\n")
_git(seed, "push", "origin", _HEAD)
work = tmp_path / "work"
_clone(origin, work)
_git(work, "checkout", _HEAD)
return origin, work
@pytest.mark.asyncio
async def test_local_ahead_survives_sync_and_reaches_origin(
repo_pair: tuple[Path, Path],
) -> None:
"""(a) Committed-but-unpushed local work is a superset of origin — the
reset is skipped and the rebase + force-push PUBLISHES it instead of
discarding it (the historical data-loss bug)."""
origin, work = repo_pair
_commit(work, "fix.py", "unpushed fix\n") # committed locally, never pushed
svc = _service()
result = await svc.rebase_onto_base(
work, head_branch=_HEAD, base_branch=_BASE, git_token=""
)
assert result["status"] == "rebased"
assert "add fix.py" in _log_subjects(origin, _HEAD)
@pytest.mark.asyncio
async def test_local_behind_adopts_origin_unchanged(
repo_pair: tuple[Path, Path],
) -> None:
"""(b) Local has nothing origin lacks (a sibling pushed while this clone
sat idle) resets to origin exactly as before the fix."""
origin, work = repo_pair
other = origin.parent / "other"
_clone(origin, other)
_git(other, "checkout", _HEAD)
_commit(other, "sibling.py", "from another clone\n")
_git(other, "push", "origin", _HEAD)
svc = _service()
result = await svc.rebase_onto_base(
work, head_branch=_HEAD, base_branch=_BASE, git_token=""
)
assert result["status"] == "rebased"
assert "add sibling.py" in _log_subjects(work, _HEAD)
assert _rev(work, _HEAD) == _rev(work, f"origin/{_HEAD}")
@pytest.mark.asyncio
async def test_diverged_refuses_and_touches_neither_side(
repo_pair: tuple[Path, Path],
) -> None:
"""(c) Both local and origin carry unique commits — refuse outright;
neither branch nor origin is touched, nothing is pushed."""
origin, work = repo_pair
other = origin.parent / "other"
_clone(origin, other)
_git(other, "checkout", _HEAD)
_commit(other, "sibling.py", "from another clone\n")
_git(other, "push", "origin", _HEAD)
_commit(work, "local.py", "local-only work\n") # unpushed local commit
local_before = _rev(work, _HEAD)
origin_before = _rev(origin, _HEAD)
svc = _service()
result = await svc.rebase_onto_base(
work, head_branch=_HEAD, base_branch=_BASE, git_token=""
)
assert result == {"status": "diverged", "local_only": 1, "origin_only": 1}
assert _rev(work, _HEAD) == local_before
assert _rev(origin, _HEAD) == origin_before
@pytest.mark.asyncio
async def test_local_ref_absent_recovers_from_origin(
repo_pair: tuple[Path, Path],
) -> None:
"""A workspace whose local ref for ``head_branch`` doesn't exist yet
(only fetched/tracked from origin, e.g. a bare clone-root caller) is
recovered checkout, never reset, since there's nothing local to lose."""
origin, _work = repo_pair
fresh = origin.parent / "fresh"
_clone(origin, fresh) # only master is checked out; _HEAD is origin-only
assert _git_ok(fresh, "rev-parse", "--verify", "--quiet", _HEAD).returncode != 0
svc = _service()
result = await svc.rebase_onto_base(
fresh, head_branch=_HEAD, base_branch=_BASE, git_token=""
)
assert result["status"] == "rebased"
assert _git_ok(fresh, "rev-parse", "--verify", "--quiet", _HEAD).returncode == 0
@pytest.mark.asyncio
async def test_superseded_still_returns_superseded(
repo_pair: tuple[Path, Path],
) -> None:
"""(d) A head whose work is already in base is still 'superseded'
byte-for-byte unchanged by the reconciliation step."""
origin, work = repo_pair
# Fast-forward base past the branch's own tip so the rebase leaves it
# with zero unique commits over base.
other = origin.parent / "other-base"
_clone(origin, other)
_git(other, "merge", "--no-ff", "-m", "merge feature", f"origin/{_HEAD}")
_git(other, "push", "origin", "master")
svc = _service()
result = await svc.rebase_onto_base(
work, head_branch=_HEAD, base_branch=_BASE, git_token=""
)
assert result == {"status": "superseded"}
def _install_rejecting_hook(bare: Path) -> None:
"""Make every push to ``bare`` fail, simulating a push-side failure
(network blip, flow-verb timeout kill, container reap) after a rebase
already succeeded locally."""
hook = bare / "hooks" / "pre-receive"
hook.write_text("#!/bin/sh\nexit 1\n")
hook.chmod(0o755)
def _remove_rejecting_hook(bare: Path) -> None:
(bare / "hooks" / "pre-receive").unlink()
@pytest.mark.asyncio
async def test_self_inflicted_wedge_self_heals_on_retry(
repo_pair: tuple[Path, Path],
) -> None:
"""(e) A prior rebase that succeeded locally but whose force-push then
failed leaves local=rebased-history, origin=old-history both rev-list
counts positive by raw SHA. The patch-equivalence probe recognizes
origin's tip as fully rewritten into local's exclusive commits, so the
retry self-heals as 'rebased' instead of refusing as 'diverged' forever."""
origin, work = repo_pair
# Advance master so the rebase actually rewrites HEAD's commits' SHAs.
other = origin.parent / "other-base"
_clone(origin, other)
_commit(other, "base2.py", "advance master\n")
_git(other, "push", "origin", "master")
_commit(work, "fix.py", "unpushed fix\n") # local commit to be rebased
svc = _service()
_install_rejecting_hook(origin)
with pytest.raises(GitCommandError):
await svc.rebase_onto_base(
work, head_branch=_HEAD, base_branch=_BASE, git_token=""
)
_remove_rejecting_hook(origin)
result = await svc.rebase_onto_base(
work, head_branch=_HEAD, base_branch=_BASE, git_token=""
)
assert result["status"] == "rebased"
assert "add fix.py" in _log_subjects(origin, _HEAD)
@pytest.mark.asyncio
async def test_conflicts_still_refuses_and_reports_files(
repo_pair: tuple[Path, Path],
) -> None:
"""(d) A genuine rebase conflict is still reported exactly as before —
aborted, files listed, nothing pushed."""
origin, _work = repo_pair
# Branch off the ORIGINAL master and edit README.md.
conflict_head = "feature/backend/conflict"
conflicting = origin.parent / "conflicting"
_clone(origin, conflicting)
_git(conflicting, "checkout", "-b", conflict_head)
(conflicting / "README.md").write_text("branch change\n")
_git(conflicting, "add", "README.md")
_git(conflicting, "commit", "-m", "branch edits README")
_git(conflicting, "push", "origin", conflict_head)
# Move master with a COMPETING edit to the same line, so rebasing the
# branch onto the new master tip collides.
other = origin.parent / "other-base"
_clone(origin, other)
(other / "README.md").write_text("master change\n")
_git(other, "add", "README.md")
_git(other, "commit", "-m", "master edits README")
_git(other, "push", "origin", "master")
svc = _service()
result = await svc.rebase_onto_base(
conflicting, head_branch=conflict_head, base_branch=_BASE, git_token=""
)
assert result["status"] == "conflicts"
assert result["files"] == ["README.md"]
@@ -64,7 +64,14 @@ async def test_rebase_rebased_force_pushes_when_unique_commits() -> None:
pushed.append(args) pushed.append(args)
return _result() return _result()
if args[:2] == ["rev-list", "--count"]: if args[:2] == ["rev-list", "--count"]:
# Only the post-rebase unique-vs-base count is non-zero; the
# pre-rebase local-vs-origin(HEAD) classification must read as
# "nothing unique on either side" or this would misclassify as
# diverged before the rebase ever runs.
range_spec = next(iter(args[2:]), "")
if range_spec == f"origin/{_BASE}..HEAD":
return _result(stdout="3\n") return _result(stdout="3\n")
return _result(stdout="0\n")
return _result() return _result()
with patch.object(svc, "_run_git", new=fake_run): with patch.object(svc, "_run_git", new=fake_run):