fix(desktop): honest sidebar load-more — source-scoped lists, matching totals, pin-aware counts

[OmarB97]

Why

The operator reported "Load more" appearing when nothing visibly overflows, and count labels that never settle (the morning recording showed Sessions 17/18 with everything loaded). Investigation found the affordance was structurally dishonest, for three independent reasons:

1. The recents universe was polluted. The unified session fetch had no source filter, so cron and messaging-platform sessions rode the recents list and its totals — the primary profile today carries 129 cron + 92 telegram sessions. Meanwhile the cron/messaging sidebar slices documented in store/session.ts had no writers at all (setCronSessions/setMessagingSessions are referenced only in the store): the per-platform messaging sections never rendered, and a pinned cron run resolved only while it happened to sit in the recents page. The source-scoping machinery upstream carries (source=/exclude_sources= on both session endpoints + the controller slice fetches) was dropped in a fork merge.
2. Counts and pages could disagree. Without exclude_sources on session_count/surfaced_session_count, any filtered page was paired with an unfiltered total — "Load more" then advertises pages that can never arrive.
3. Pinned rows broke the label math. The Sessions label counted unpinned rows against a pin-inclusive total (17/18 forever after pinning one row), and the messaging sections (post drag-to-pin) treated pinned rows as not-loaded, over-reporting their pagers.

What changed

Commit 1 — fix(desktop): restore source-scoped session lists… (port of upstream's design, fork-only):

• hermes_state.py — session_count + surfaced_session_count gain exclude_sources (mirroring list_sessions_rich, which already had it), so counts can apply the exact filter their paired page uses.
• hermes_cli/web_server.py — /api/sessions and /api/profiles/sessions accept source= / exclude_sources= (comma-separated) and thread them into BOTH the page query and the count, per profile on the unified endpoint.
• apps/desktop/src/hermes.ts — SessionSourceFilter + the filter argument on listAllProfileSessions.
• apps/desktop/src/app/desktop-controller.tsx — recents fetch excludes cron + messaging sources; refreshCronSessions (source=cron, pin resolution), refreshMessagingSessions (inverse exclusion, split per platform, truncation flag), and loadMoreMessagingForPlatform (per-source pager resolving exact totals) now actually populate the store slices; recents are scoped to the active profile with a scope-change refetch effect (identity-stable variant of upstream's reactive dep — refreshSessions flows into useGatewayBoot/usePromptActions, so its identity is kept stable and the effect owns the refetch); onLoadMoreMessaging is finally passed to the sidebar.
• tests/hermes_cli/test_web_server.py — wiring-capture test asserting both params reach page AND count on /api/sessions (the exact "param silently dropped in a merge" failure mode), plus a seeded-DB test pinning total == listable rows for all three slices (recents-excluded, cron, telegram) on /api/profiles/sessions.

Commit 2 — fix(desktop): make sidebar count labels and load-more honest about pinned rows (upstream-relevant):

• apps/desktop/src/app/chat/sidebar/index.tsx — pinned rows are dropped from BOTH sides of the Sessions label (they're always loaded via the refresh keep-set, so the unpinned remainder is identical in both universes and the label now agrees with the footer); messaging sections count pinned rows as loaded (loadedCount) for their label + pager while still housing the rows in Pinned, take section recency from every loaded row (no reshuffle when the newest thread gets pinned), and only drop a section when it has nothing to show AND nothing more on disk.

How to review

1. hermes_state.py — the two count functions; the added clause is byte-identical to list_sessions_rich's.
2. web_server.py — both handlers thread source/exclude_list to page + count; nothing else moved.
3. desktop-controller.tsx — SIDEBAR_EXCLUDED_SOURCES/MESSAGING_EXCLUDED_SOURCES constants and the three new callbacks; note the comment explaining the deliberate deviation from upstream (scope read at call time + separate effect, instead of a reactive dep).
4. sidebar/index.tsx — pinnedFromRecentsCount/agentKnownTotal for the label; loadedCount/latestActivity in messagingGroups and their two render-site uses.
5. Tests last — they encode the invariant: a section's total counts only rows that section can ever list.

Evidence

• New endpoint tests: 3/3 pass; full tests/hermes_cli/test_web_server.py: 248 passed; tests/test_hermes_state.py -k "count or list_sessions": 21 passed.
• Live-data check (TestClient against a copy of the operator's real state.db, auth header included): old vs new recents paths return consistent rows == total with zero rows lost; source=cron / source=telegram slices return exactly their own rows with matching totals. (The live DB currently lists few unarchived roots — the operator ran the freshly-restored Archive All earlier today — so the synthetic-DB test carries the cardinality assertions.)
• Renderer harness (real Chromium, seeded stores): 5 local loaded / total 5 / 1 pinned → header reads Sessions 4 with no Load more (previously 4/5 + phantom footer); bumping the total to 9 → header 4/8 + footer "Load 4 more" (label and footer agree in the unpinned universe); Discord section with both conversations loaded and one pinned → header Discord 2, one row rendered, no phantom per-platform Load more.
• npm run type-check, npx eslint (all touched files), production vite build: pass. Full renderer vitest suite: failure set byte-identical to the feat(desktop): drag-to-pin follow-ups — positional drop, messaging pins render, drag hint #146 baseline run (diff-verified).

Verification

• …/.venv/bin/python -m pytest tests/hermes_cli/test_web_server.py -q → 248 passed.
• npx vitest run --environment jsdom → no new failures vs the recorded baseline set (diff clean).
• npm run type-check / npx eslint … / npm run build → pass.
• Re-runnable live check: HERMES_HOME=<copy> python -c "...TestClient..." with X-Hermes-Session-Token header — /api/profiles/sessions?exclude_sources=cron,telegram,… returns total == rows for the operator's data.

Risks / gaps

• First refresh after deploy may briefly keep stale messaging/cron rows inside $sessions (the merge keep-set preserves pinned ones indefinitely); they don't render in recents (pin-filtered) and only feed pin resolution — accepted, self-corrects for unpinned rows on the next page fetch.
• exclude_sources uses SQL NOT IN, which drops NULL-source rows; the operator's DBs have zero NULL sources (verified: cli/tui/cron/telegram/api_server/unknown) and the schema declares source TEXT NOT NULL — non-actionable here, and matching upstream's exact semantics keeps the port faithful.
• Messaging platforms whose threads all have message_count=0 (the operator's 92 telegram rows today) stay invisible under the min_messages=1 policy shared with recents — pre-existing policy, out of scope; noted on MeshBoard task hermes-desktop-sidebar-load-more-honesty.
• The sidebar-honesty commit applies to the upstream port PR (feat(desktop): refine sidebar session drag reordering NousResearch/hermes-agent#43661) and is cherry-picked there; the source-scoping commit is fork-only (upstream already ships it).

Collaborators

• @OmarB97 (operator — reported the phantom Load more)
• Claude Fable 5 (ko-mac.claude — investigation, port, tests, verification)

[github-actions]

🔎 Lint report: fix/sidebar-load-more-honesty vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 10597 on HEAD, 10583 on base (🆕 +14)

🆕 New issues (7):

Rule	Count
invalid-argument-type	6
invalid-parameter-default	1

First entries

hermes_cli/web_server.py:1814: [invalid-argument-type] invalid-argument-type: Argument to bound method `SessionDB.list_sessions_rich` is incorrect: Expected `list[str]`, found `(list[str] & ~AlwaysFalsy) | None`
hermes_cli/web_server.py:1750: [invalid-parameter-default] invalid-parameter-default: Default value of type `None` is not assignable to annotated parameter type `str`
hermes_cli/web_server.py:1717: [invalid-argument-type] invalid-argument-type: Argument to bound method `SessionDB.surfaced_session_count` is incorrect: Expected `list[str]`, found `(list[str] & ~AlwaysFalsy) | None`
hermes_cli/web_server.py:1813: [invalid-argument-type] invalid-argument-type: Argument to bound method `SessionDB.list_sessions_rich` is incorrect: Expected `str`, found `(str & ~AlwaysFalsy) | None`
hermes_cli/web_server.py:1824: [invalid-argument-type] invalid-argument-type: Argument to bound method `SessionDB.session_count` is incorrect: Expected `list[str]`, found `(list[str] & ~AlwaysFalsy) | None`
hermes_cli/web_server.py:1823: [invalid-argument-type] invalid-argument-type: Argument to bound method `SessionDB.session_count` is incorrect: Expected `str`, found `(str & ~AlwaysFalsy) | None`
hermes_cli/web_server.py:1716: [invalid-argument-type] invalid-argument-type: Argument to bound method `SessionDB.surfaced_session_count` is incorrect: Expected `str`, found `(str & ~AlwaysFalsy) | None`

✅ Fixed issues: none

Unchanged: 5544 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.