fix(import): sync imported checkpoints once logged in by sdshah09 · Pull Request #1929 · entireio/cli

sdshah09 · 2026-08-08T17:06:10Z

Fixes #1773 (fix 2 — sync after login). @karthik-rameshkumar handed this over on the issue after landing fix 1 (#1774).

The problem

entire import writes read-only checkpoints and creates no commit, so the pre-push git hook — the only sync trigger — never fires for it. Because idempotency is keyed on local existence, history imported while logged out is stranded permanently: the post-login re-run reports 0 turn(s) (N already imported) and makes no attempt to sync.

The insight that keeps this small

The git-refs store already enqueues every checkpoint ref as it writes it (gitRefsStore.setRef → enqueueForPush, checkpoint/refs_store.go:209), and nothing drains that queue but a push. So turns imported while logged out are still queued afterwards — the data is packed and addressed, nobody mails it.

Draining the queue after a logged-in import is therefore enough to recover them, with no change to the import idempotency logic and no new push machinery.

What this adds

syncImportedCheckpoints (new import_sync.go, ~70 lines), called from the two existing warnIfImportNotSynced sites — import_cmd.go and setup_import.go. It reuses strategy.PushQueuedCheckpointRefs, whose only current caller is doctor migrate-checkpoints' opt-in "push now"; this is the same foreground-drain shape.

Deliberately not gated on TurnsImported > 0. That is the fix: a logged-in re-run that imports nothing new still pushes, which is what makes it the recovery path.

Best-effort by design — the import already succeeded locally, so a push failure prints a Note: and returns rather than failing the command; context.Canceled stays silent.

Scope: git-refs only

Since #1900 made git-refs the sole entire enable default, this covers every repo newly hitting the bug. git-branch has no push queue — its checkpoints ride the entire/checkpoints/v1 branch the pre-push hook already ships — so it is skipped rather than paying for a checkpoint-policy sync to drain a queue it never fills. Those repos keep syncing exactly as they do today; no regression, just not fixed here. Happy to extend if you'd prefer one PR covering both.

I looked at ManualCommitStrategy.PrePush first (it covers both backends and has no production callers — only PrePushFromGitHook is wired, at hooks_git_cmd.go:329). I went with PushQueuedCheckpointRefs instead because it returns (pushed, pushDisabled, err) so the user can be told accurately what happened, PrePush passes protectFirstUserBranch: false and so skips the empty-remote guard, and it forces non-interactive SSH — right for a hook, wrong for a foreground command.

Remote resolution

resolveMigratePushRemote → resolveCheckpointPushRemote, now shared with the migration command: both are non-hook drain paths and must resolve the elected checkpoint sync remote the same fail-closed way (a misconfigured checkpoint_push_remote is an error, never a fallback to origin). entire import gains --remote to override, matching doctor migrate-checkpoints.

Verification

End-to-end against a real bare remote, identical sequence on both binaries:

	0.9.0	this branch
import logged out	local only ✓	local only ✓
login + re-run output	`0 turn(s) (1 already imported)`	same
refs on the remote	0 ❌	1 ✅
push queue file	still present	drained
user told	nothing	`Synced 1 checkpoint(s) to origin.`

Tests in import_sync_test.go cover: logged out never pushes; git-branch never pushes; success/disabled/empty-queue/failure/cancelled messaging; --remote override; remote-resolution failure reported not fatal; dry-run never pushes. The regression test is TestImportSyncsAlreadyImportedTurns — imports twice and asserts the second run still syncs despite importing nothing new.

mise run check green: lint 0 issues, unit + integration pass, e2e canary 60/60 (vogon) and 4/4 (roger-roger).

Open question for a maintainer

The issue's own closing note is still unanswered, and it affects this PR:

confirm with the backend team that a connected-repo git push of the checkpoint refs is sufficient for dashboard ingestion (GitHub App), and whether the CLI should nudge the "connect repo / install GitHub App" step when enabling while logged out.

Relevant because reportRepoEnabled (setup.go:976) is silently skipped when logged out, so a repo enabled before login is never reported to the backend. If pushing alone is not sufficient, this PR would print "Synced" while the dashboard stays empty, and the notice wording should probably also point at connecting the repo. Happy to follow up either way.

First contribution here — glad to adjust any of the above.

Import writes read-only checkpoints and creates no commit, so the pre-push git hook -- the only sync trigger -- never fires for it. Combined with idempotency keyed on local existence, history imported while logged out was stranded permanently: a post-login re-run reported "0 turn(s) (N already imported)" and made no attempt to sync, so the sessions never reached the dashboard. The git-refs store already enqueues every checkpoint ref as it writes it (gitRefsStore.setRef -> enqueueForPush) and nothing drains that queue but a push, so those turns are still queued afterwards. Draining the queue after a logged-in import is therefore enough to recover them, with no change to the import idempotency logic. syncImportedCheckpoints runs on every logged-in import, deliberately NOT gated on this run having imported anything new -- that is what makes a post-login re-run the recovery path. It is best-effort: the import already succeeded locally, so a push failure is reported and swallowed rather than failing the command, and Ctrl-C stays silent. Scoped to the git-refs backend, the sole `entire enable` default since entireio#1900. git-branch has no push queue -- its checkpoints ride the entire/checkpoints/v1 branch that the pre-push hook already ships -- so it is skipped rather than paying for a checkpoint-policy sync to drain a queue it never fills. Those repos keep syncing exactly as they do today. resolveMigratePushRemote is renamed resolveCheckpointPushRemote and shared with the migration command: both are non-hook drain paths and must resolve the elected checkpoint sync remote the same fail-closed way. `entire import` gains --remote to override it, matching `entire doctor migrate-checkpoints`. Verified against a real remote: on 0.9.0 the checkpoint stays local after login + re-import (0 refs on the remote); with this change the same sequence pushes it and prints "Synced 1 checkpoint(s) to origin." Refs entireio#1773 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Entire-Checkpoint: 01KZH4HXC95ET9V176H4SS6NTZ

Soph

Thanks for this! The diagnosis is right, the queue insight is the right one, and I appreciate that the PR body pre-empts the "why not PrePush" question. I verified the load-bearing claim independently: enqueueForPush has no auth gating, so the refs really are still queued. Two things I'd want fixed before merge.

Blockers

--remote can route checkpoint data off the elected sync remote. PushQueuedCheckpointRefs has no checkpointSyncAllowedForRemote gate and the flag isn't validated, so import --remote <anything> pushes transcripts there. CLAUDE.md states the opposite invariant, so this is a contract break.
The reported count is the whole repo queue, not the import's refs — an import that finds nothing can print Synced 12 checkpoint(s).

One reframe, not a blocker

For git-refs the next ordinary git push already drains the queue regardless of login state, so imported history is pending rather than stranded. That doesn't undercut the PR - syncing promptly instead of at the next push is worth having - but the docstring and the notice copy should say that, and it's why (2) matters: if prompt sync is the value, the count has to be about the import.

Relatedly, warnIfImportNotSynced still says "Log in with 'entire login' before importing", which reads as "too late for this history". Worth fixing in the same PR.

Also

The four existing TestRunSelectedImports_* tests weren't stubbed, so they now run the real sync, which reads ENTIRE_TOKEN — not covered by the testdirs net. Add stubImportSync (or force importLoggedIn false) to them.
No docs changes: ref-checkpoint-backend.md still names PrePush as the only drain, and CLAUDE.md needs the --remote behavior reflected once it's gated.

On your question regarding the push. The enable call is one step in the onboarding flow, if actual checkpoints are pushed then it does not matter if the enable call came before.

Again thanks for opening this!

Soph · 2026-08-20T20:35:23Z

+//
+// Import writes read-only checkpoints and creates no commit, so the pre-push git
+// hook — the only sync trigger — never fires for it. Before this, history
+// imported while logged out stayed local-only forever: import's idempotency is


For git-refs — the only backend this touches — the next ordinary git push already drains the whole queue regardless of login state (prePushCheckpointRefs), which is what doctor migrate-checkpoints tells users: "Refs are queued; they push on the next git push." So it isn't stranded forever, it's pending until the next push. Worth correcting here, since as written this will convince the next maintainer the queue has no other drain trigger.

That makes the value of this PR "imported history syncs promptly" rather than "recovers otherwise-unrecoverable history" — still worth having, just a different claim.

Soph · 2026-08-20T20:35:23Z

+//
+// Best-effort by design: the import itself already succeeded locally, so a push
+// failure is reported and swallowed rather than failing the command.
+func syncImportedCheckpoints(ctx context.Context, w io.Writer, repo *git.Repository, explicitRemote string) {


No WithNonInteractiveSSH and no timeout on the batch push, so a passphrase-protected key prompts mid-enable and a hung remote blocks it. The only other caller (doctor migrate-checkpoints) gates its drain behind CanPromptInteractively() + a confirm; this has neither.

Soph · 2026-08-20T20:35:23Z

+	// entire/checkpoints/v1 branch, which the pre-push hook ships alongside the
+	// user's own push. Draining a queue it never fills would be a no-op that
+	// still paid for a checkpoint-policy sync, so skip the backend entirely.
+	cpCfg, cfgErr := settings.LoadCheckpointsConfig(ctx)


cfgErr is dropped with no log and no output, so an invalid checkpoints block silently disables the whole fix — the same silent stranding #1773 is about. Both other error paths here log.

Soph · 2026-08-20T20:35:23Z

+	remote, err := importResolvePushRemote(ctx, explicitRemote)
+	if err != nil {
+		logging.Debug(ctx, "import: skipping checkpoint sync", "error", err.Error())
+		fmt.Fprintf(w, "Note: imported history is not synced yet: %v\n", err)


Surfaces resolveCheckpointPushRemote's "pass --remote explicitly" — a flag entire enable doesn't have — and fires on the ordinary git init && entire enable no-remote-yet case, not only on errors.

Soph · 2026-08-20T20:35:23Z

+		return
+	}
+
+	pushed, pushDisabled, err := importPushQueuedRefs(ctx, repo, remote)


Blocker: PushQueuedCheckpointRefs has no checkpointSyncAllowedForRemote gate (only prePush does, manual_commit_push.go:84), and --remote is returned verbatim without checking it is a configured remote. So import --remote fork (or a pasted URL) pushes refs/entire/checkpoints/*, transcripts included, there. CLAUDE.md: "Pushes to any other remote or to a raw URL never carry checkpoint data."

Separately: the policy fetch inside runs before the queue-empty check, so every logged-in import pays a remote round-trip with nothing to push — queue.Peek() (pushqueue.go:117) makes that cheap to skip.

Soph · 2026-08-20T20:35:24Z

+	case pushed == 0:
+		// Nothing queued — everything this import produced was already synced.
+	default:
+		fmt.Fprintf(w, "Synced %d checkpoint(s) to %s.\n", pushed, remote)


Blocker: pushed is the whole repo-wide queue, not this import's refs, so 12 unpushed session checkpoints give Imported 0 turn(s)… Synced 12 checkpoint(s). Matters more if prompt sync is the point of the PR. And under checkpoint_remote the refs go to that URL, not to remote — stderr and stdout then name different destinations.

Soph · 2026-08-20T20:35:24Z

 	warnIfImportNotSynced(w, importedLocalHistory)
+	// When it runs logged in, push what was just imported: enable creates no
+	// commit either, so nothing else would trigger the pre-push hook (#1773).
+	syncImportedCheckpoints(ctx, w, repo, "")


Ctrl-C breaks the loop above and lands here with a cancelled ctx: git config fails instantly and the swallow at checkpoint_sync_remote.go:120 prints no git remotes configured in a repo that has them (or a false checkpoint_push_remote misconfiguration). import_cmd.go:98 guards this; this path doesn't.

Soph · 2026-08-20T20:35:24Z

 	cmd.Flags().StringVar(&pathFlag, "path", "", "Override the transcript directory to import from")
 	cmd.Flags().BoolVar(&dryRun, "dry-run", false, "Report what would be imported without writing")
 	cmd.Flags().StringSliceVar(&sessions, "session", nil, "Import only these session IDs (repeatable)")
+	cmd.Flags().StringVar(&remoteFlag, "remote", "", "Git remote to sync imported checkpoints to (default: the elected checkpoint sync remote)")


Worth validating this against the elected sync remote before use — see the gate note on import_sync.go:65.

Soph · 2026-08-20T20:35:24Z

+	require.NoError(t, os.WriteFile(filepath.Join(claudeDir, "s.jsonl"),
+		[]byte(`{"type":"user","uuid":"u1","message":{"role":"user","content":"hi"}}`+"\n"), 0o644))
+
+	calls := stubImportSync(t, true, func(context.Context, *git.Repository, string) (int, bool, error) {


Doesn't reproduce #1773: both runs are loggedIn=true, so the logged-out precondition never exists, and the push is stubbed — so the claim the PR rests on (refs written while logged out are still queued afterwards) is never exercised. Breaking enqueue would leave this green.

Soph · 2026-08-20T20:35:24Z

 	t.Chdir(tmpDir)

-	got, err := resolveMigratePushRemote(context.Background(), "explicit-remote")
+	got, err := resolveCheckpointPushRemote(context.Background(), "explicit-remote")


Rename the five TestResolveMigratePushRemote_* functions too — they now point at a symbol that no longer exists.

sdshah09 requested a review from a team as a code owner August 8, 2026 17:06

sdshah09 mentioned this pull request Aug 9, 2026

entire enable/import while logged out imports locally with no way to sync after login #1773

Open

Merge branch 'main' into fix/1773-sync-imported-checkpoints

1cd64f0

This was referenced Aug 17, 2026

fix(opencode): spawn hooks via node:child_process for Desktop #2018

Merged

OpenCode hooks never fire in the OpenCode Desktop app: plugin uses Bun globals, desktop runs the server on Node #2014

Closed

sdshah09 added 2 commits August 17, 2026 15:59

Merge branch 'main' into fix/1773-sync-imported-checkpoints

2259feb

Merge branch 'main' into fix/1773-sync-imported-checkpoints

9212599

Soph reviewed Aug 20, 2026

View reviewed changes

Navigation Menu

Repository navigation

fix(import): sync imported checkpoints once logged in - #1929

fix(import): sync imported checkpoints once logged in#1929
sdshah09 wants to merge 4 commits into
entireio:mainfrom
sdshah09:fix/1773-sync-imported-checkpoints

sdshah09 commented Aug 8, 2026 •

edited

Loading

Uh oh!

Soph left a comment

Uh oh!

Soph Aug 20, 2026

Uh oh!

Soph Aug 20, 2026

Uh oh!

Soph Aug 20, 2026

Uh oh!

Soph Aug 20, 2026

Uh oh!

Soph Aug 20, 2026

Uh oh!

Soph Aug 20, 2026

Uh oh!

Soph Aug 20, 2026

Uh oh!

Soph Aug 20, 2026

Uh oh!

Soph Aug 20, 2026

Uh oh!

Soph Aug 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

2 participants

Conversation

sdshah09 commented Aug 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

The problem

The insight that keeps this small

What this adds

Scope: git-refs only

Remote resolution

Verification

Open question for a maintainer

Uh oh!

Soph left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

2 participants

sdshah09 commented Aug 8, 2026 •

edited

Loading