fix(multi-source): plumb sourceId through performFullSync (closes PR #707 gap; closes #497, #540)

[jeremyknows]

Summary

Follow-up + complement to PR #707 (mdcruz88's source_id-routing fix). PR #707 correctly threads sourceId through sync's incremental loop (lines 581/641) but performFullSync (the path --full invokes) at sync.ts:892 calls runImport(engine, importArgs, { commit: headCommit }) without sourceId. Result: gbrain sync --source X --full updates sources.last_sync_at (looks like binding) but writes pages to default source.

This PR adds the missing thread + a regression test that exercises the full-sync path the existing 19-test suite at test/source-id-tx-regression.test.ts doesn't cover.

Relates to #497, #540. Branch rebases cleanly onto current master (commit dffb607, v0.30.1) and includes all of @mdcruz88's PR #707 commits as base.

Diff scope

3 files for the gap-fix (+18/-4) — preserves PR #707's design intent of gbrain import CLI being default-only by making sourceId a programmatic-only opt (no CLI flag added):

• src/commands/import.ts — runImport accepts opts.sourceId, threads to importFile + importImageFile
• src/commands/sync.ts:924 — performFullSync passes opts.sourceId to runImport
• src/core/import-file.ts — ImportImageOptions type accepts sourceId (TS-only fix; image-import body wiring deferred — out of scope, noted as a separate follow-up)

1 test file (+139) — regression coverage for the gap:

• test/performfullsync-source-id.test.ts — 2 PGLite tests:
  • performFullSync with --source routes pages to named source (not default)
  • performFullSync WITHOUT --source still targets default (back-compat preserved)

Verification

$ bun run typecheck
$ tsc --noEmit
[clean]

$ bun run verify
[8 checks, all OK]
- check-jsonb-pattern: OK (no JSON.stringify(x)::jsonb interpolation)
- check-progress-to-stdout: OK
- check-test-isolation: OK (271 non-serial unit files scanned)
- check-wasm-embedded: OK
- check-admin-build: OK (vite v6.4.2, 311ms)
- check-admin-scope-drift: OK
- check-cli-executable: OK
- tsc --noEmit: clean

$ bun test test/source-id-tx-regression.test.ts test/performfullsync-source-id.test.ts
21 pass / 0 fail / 65 expect() calls / 1.90s

The 19 existing tests (PR #707) + 2 new tests (this PR) all pass.

Real-behavior proof

• Behavior addressed: performFullSync (the --full sync path) does not thread sourceId to runImport. With PR fix(multi-source): thread source_id through per-page tx surface (closes Postgres 21000 mid-import) #707 alone, gbrain sync --source X --full updates sources.last_sync_at to look like binding worked, but pages still land in default. Verified empirically against rebased PR fix(multi-source): thread source_id through per-page tx surface (closes Postgres 21000 mid-import) #707 commit 46cd197 (i.e. PR fix(multi-source): thread source_id through per-page tx surface (closes Postgres 21000 mid-import) #707 without this gap-fix): the new regression test FAILS with expected counts['testsrc-pfs'] toBeGreaterThan(0) because pages went to default.

• Real environment tested: Bun 1.3.8 / macOS 25.4.0 / PGLite (no DATABASE_URL). Test fixture: temp git repo (mkdtempSync + git init + 2 markdown files) + performSync(engine, { repoPath, full: true, sourceId: 'testsrc-pfs', noPull: true, noEmbed: true }) + SELECT source_id, COUNT(*) FROM pages GROUP BY source_id.

• Exact command: bun test test/performfullsync-source-id.test.ts

• Evidence after fix (commit 694f9c9 + test 37dc598):

Found 2 markdown files
[import.files] start
[import.files] 2/2 (100%) imported=2 skipped=0 errors=0
[import.files] 2/2 (100%) done

Import complete (0.0s):
  2 pages imported
  0 pages skipped (0 unchanged, 0 errors)
  2 chunks created

(pass) performFullSync threads sourceId end-to-end >
       performFullSync with --source routes pages to named source (not default) [167.07ms]
(pass) performFullSync threads sourceId end-to-end >
       performFullSync WITHOUT --source still targets default (back-compat preserved) [147.11ms]

 2 pass
 0 fail
 8 expect() calls

• Observed result: With the fix applied, pages land at (source_id='testsrc-pfs', slug='topics/foo') and (source_id='testsrc-pfs', slug='topics/bar'). With no sourceId opt passed (back-compat path), pages land at (source_id='default', ...). Direct PGLite probe confirms.

• What was NOT tested:

  • Postgres engine path (test runs PGLite-only because no DATABASE_URL was available; the engine-layer fix in PR fix(multi-source): thread source_id through per-page tx surface (closes Postgres 21000 mid-import) #707 has Postgres parity per its own e2e tests, but performFullSync itself is engine-agnostic plumbing).
  • The Atlas-internal monorepo case (sync requires local_path to be a git root, so atlas's 9 agent-diary subdirs of ~/atlas still can't be source-bound via sync even with this fix; that's a separate concern filed at feat: sync should support subdir-of-git-repo as a source (atlas-style monorepo with N logical sources) #753).
  • Image-import body wiring — ImportImageOptions accepts sourceId here as a TS-only fix to keep runImport typecheck-clean, but withImportTransaction doesn't yet pass sourceId through to image tx.putPage. PR fix(multi-source): thread source_id through per-page tx surface (closes Postgres 21000 mid-import) #707's commit body explicitly enumerates this as a deferred follow-up; not in scope here.
  • Watch-loop interaction (--watch + --source) — not covered.

Regression check against pre-fix

Verified by checking out 46cd197 (rebased PR #707 without this gap-fix) and running the same test:

$ git checkout 46cd197 -- src/commands/sync.ts src/commands/import.ts src/core/import-file.ts
$ bun test test/performfullsync-source-id.test.ts

Found 2 markdown files
[import.files] start
[import.files] 2/2 (100%) imported=2 skipped=0 errors=0
[import.files] 2/2 (100%) done

(fail) performFullSync threads sourceId end-to-end >
       performFullSync with --source routes pages to named source (not default) [166.21ms]
       Expected: > 0
       Received: 0
       at counts['testsrc-pfs']
(pass) performFullSync threads sourceId end-to-end >
       performFullSync WITHOUT --source still targets default (back-compat preserved) [137.66ms]

 1 pass / 1 fail / 7 expect() calls

Result: with PR #707 alone (no gap-fix), the --source path fails — pages went to default. This is the bug. The fix in commit 694f9c9 resolves it.

Test plan

• Run bun test test/source-id-tx-regression.test.ts test/performfullsync-source-id.test.ts — should be 21 pass.
• Run bun run typecheck — clean.
• Run bun run verify — all 8 guard checks pass.
• Optional: e2e tests via DATABASE_URL — performFullSync's source threading is identical for Postgres + PGLite (it's at the runImport CLI dispatch layer, not engine-layer), so PGLite coverage gives high confidence.

Maintainer-friendly notes

@mdcruz88 — earlier comment on PR #707 (#707 (comment)) offered this as either a separate PR or as commits to pull into PR #707. This PR is the standalone form. If you'd prefer to incorporate commits 694f9c9 (gap-fix) + 37dc598 (test) into PR #707 directly, happy to close this PR — just leave a comment.

@garrytan — appreciate the source isolation infrastructure in v0.18+. The combination of PR #707 (engine-layer source_id threading) + this PR (sync-command-layer threading) closes the open #497 + #540 paths fully. Filed alongside #753 (a related architectural follow-up about monorepo subdir-source support, not in this PR's scope).

Commits

1. 46cd197 — fix(multi-source): thread source_id through per-page tx surface (rebased from PR fix(multi-source): thread source_id through per-page tx surface (closes Postgres 21000 mid-import) #707)
2. 694f9c9 — fix(multi-source): plumb sourceId through performFullSync (PR fix(multi-source): thread source_id through per-page tx surface (closes Postgres 21000 mid-import) #707 gap)
3. 37dc598 — test: regression test for performFullSync sourceId threading

Branch: https://github.com/jeremyknows/gbrain/tree/fix-707-performfullsync-gap-rebase

---

^{Need help on this PR? Tag @codesmith with what you need.}

• Let Codesmith autofix CI failures and bot reviews

[mdcruz88]

Going to fold all three commits (46cd197 rebase + 694f9c97 gap-fix + 37dc5982 regression test) into PR #707 with Co-authored-by on each, then add two more on top:

1. Image-import body wiring — ImportImageOptions.sourceId is accepted in your fix as a TS-only patch, but withImportTransaction doesn't yet thread it through to image tx.putPage. Under GBRAIN_EMBEDDING_MULTIMODAL=true, performFullSync reaches importImageFile via collectMarkdownFiles and silently routes image pages/chunks/files to default. Same class of bug the gap-fix addresses for text — should ship in the same PR. Will add image regression test to mirror your text test.
2. writeSyncConfig?: boolean opt on runImport — runImport writes global sync.last_commit / sync.last_run / sync.repo_path keys (not source-scoped), so after a --source X --full run, a later bare gbrain sync can pick up X's repo as the default-source repo. performFullSync will pass writeSyncConfig: false and own its own anchors via writeSyncAnchor.

Once I force-push to PR #707 I'll ping here so you can close #757. Realistically tonight or tomorrow morning. Thanks again — the regression test was the missing piece I'd planned to add myself, glad to have it landed already.

(Heads-up: the test docstring references some Atlas-internal vocabulary — "PRISM Round 2", "Atlas Terminal agent", "PR-E" — I'll scrub those in a follow-up cleanup commit during the fold-in so the upstream test reads as a generic regression. Doesn't affect the test logic.)

[mdcruz88]

Done — PR #707 force-pushed to v0.30.1 base with all 3 of your commits (46cd197 rebase + 694f9c97 gap-fix + 37dc5982 test) preserved verbatim, plus 3 of mine on top:

• Scrub of internal vocabulary from your test's docstring (test logic unchanged).
• Image-import body sourceId wiring (withImportTransaction + importImageFile + image_of sibling-link source-scoping) + 2 new image source-routing tests.
• writeSyncConfig?: boolean opt on runImport so performFullSync(--source X) doesn't overwrite global sync.repo_path / sync.last_commit / sync.last_run keys + 2 new isolation tests.

Full stack + rationale on #707 (comment).

Feel free to close #757 — appreciate you driving the rebase + the empirical proof of the gap. Couldn't have shipped this in a single coherent PR without your work.

[jeremyknows]

Closing as requested by @mdcruz88 — our 3 commits (rebase + gap-fix + regression test) have been folded into PR #707 (force-pushed). Appreciate the integration and the additional sourceId wiring on the image path + isolation. The regression test does the job it was designed for either way.