Summary

Two cross-tab navigation entry-points into the Sources panel set one axis (namespace filter or source path) but did not align the panel's vendor sub-tab axis (user / claude / openai). Because the panel only renders the active vendor's tree (#570), the user-visible chip / active-source state suggested the link succeeded while the rendered content was an unrelated vendor — reported as "Sources 화면으로 넘어가긴 하는데 직접 링크가 안 된다."

Audit (3 parallel Explore agents) found _navigateToSource is the same shape of bug; both fixes share the helper and live next to each other in app.js, so they ship together as one focused change.

Changes

navigateToSourcesByNs (Settings → Namespaces → Sources button)

packages/memtomem/src/memtomem/web/static/app.js — _renderMemorySourceTree:

• filterActive now also returns true when STATE.sourcesNsFilter is set, not just when the path-text input is non-empty. Empty memory_dir headers (28 of 29 in the repro for a claude:… namespace) are now hidden under the existing visibleCatsRaw guard.
• After computing per-vendor totals, if the current vendor has zero matches and another vendor has any, the active sub-tab shifts to the vendor with the most. Not persisted via _writeActiveSourcesVendor — this is a navigation hint, not the user's stored preference, so the next manual click still wins.

_navigateToSource (Home recent source, command palette)

Same root cause, narrower entry-point. Resolves the target vendor from the source's memory_dir → provider (looked up via STATE.allSources + STATE.memoryStatusByPath) before activateTab('sources'), so a Claude/OpenAI source clicked from Home doesn't land on whichever sub-tab the user last viewed and silently miss the .source-item[title=path] scroll-into-view that follows. Falls through to current behaviour when STATE.allSources isn't populated yet (cold load) — no regression.

Cache bust

index.html app.js?v=82 → v=84 so the JS body change is actually picked up by browsers (per feedback_static_asset_cache_bust.md, unbumped query strings get served stale from disk cache).

Verification

• uv run ruff check packages/memtomem/src — clean.
• uv run pytest -m "not ollama" -k "namespace or sources or web" — 986 passed.
• Playwright (against mm web):
  • localStorage forced to user vendor → click claude:-Users-pdstudio-Work-agent-harness-memtomem namespace's Sources button → vendor tab shifts to claude, exactly 1 dir group with 284 file items, chip shows the namespace.
  • Clear chip → 29 dir groups + 356 files reappear under Claude.
  • Click default namespace from user vendor → vendor stays user, 2 dirs + 22 files (auto-switch correctly inert when current vendor has matches).
  • _navigateToSource(<claude path>) from user vendor → vendor switches to claude, target source becomes .source-item.active and gets scrollIntoView.
  • _navigateToSource(<user path>) from user vendor → no vendor switch, target source becomes active (regression check).

Out-of-scope sibling

_searchByTag(tag) (app.js:3631-3640) populates both search-input and tag-filter with the tag string — same axis-mismatch shape but v0.1.0-since behaviour, not a recent regression. Filed as a separate issue (#672) for an intent-vs-design-flaw decision; user-visible behaviour change so it needs its own CHANGELOG entry.

Test plan

• ruff check clean
• Pytest namespace/sources/web suite green
• Playwright: cross-vendor namespace card → auto-switch to matching vendor sub-tab
• Playwright: chip-clear restores full vendor tree
• Playwright: same-vendor namespace card preserves current vendor
• Playwright: _navigateToSource cross-vendor + same-vendor regression
• Manual: refresh in real browser to confirm ?v=84 bust takes effect

🤖 Generated with Claude Code