Repository navigation
Description
Company or project name
ClickHouse
Component
Dashboard
Describe what's wrong
Pending memory proposals do not contribute to the sidebar's Review queue visibility or count. A namespace can have proposals waiting for a person while its Review queue entry is absent. When the entry is present for another reason, its badge counts conflicts rather than the proposals waiting inside it.
The user also reports difficulty finding Review queue and its Proposals tab. Review is a recurring memory workflow and needs a stable entry point for people who can access it.
Does it reproduce on the latest release or main?
Not checked
The browser interaction has not been independently reproduced on current main. Static inspection of main at f6e14b5c9e5c12e60c79adcc69a2e739567ff438 confirms the missing proposal-count path: adminItems only considers conflicts and autoResolved; the badge reads only conflicts. The queue itself has a Proposals tab and starts on Conflicts.
How to reproduce
On a private test installation with authentication enabled:
- Use a namespace where the reviewer has the required access. Create one pending fact with
memory.set(proposed=true)and a distinct key and value. Leave no open conflicts or recent auto-resolutions in that namespace. - Open the dashboard as the reviewer, select that namespace, and wait for the normal badge refresh.
- Inspect the sidebar and the mobile More navigation. Compare their Review entry and count with the pending rows from
memory.proposalsor the Proposals tab's pending rows. - Repeat with both conflicts and proposals present. Check whether the label and count describe the pending review work shown by the page.
The latest applied database migration and browser version were not captured. This report contains no claim that a private runtime fixture was created or tested.
Expected behavior
- Review remains reachable for a caller authorized to view review work, including when its queue is empty or a count request fails. Place it in the ordinary memory workflow rather than relying on an automatically appearing admin-drawer entry. Preserve backend roles and show the permission needed for restricted actions.
- Pending proposals contribute to a clearly labeled review count, with per-kind counts inside Review. Define the count as review items, distinguish it from the number of memories, and exclude accepted/rejected proposals and auto-resolved history.
- A proposal-only queue is visible and offers a direct route to its pending proposals. Empty Conflicts must not look like an empty overall queue.
- Counts match the complete authorized queue, including beyond one results page. Fetch bounded server-side summaries rather than memory bodies or a capped client-side sample.
- Counts refresh after a decision and on scope changes. A failed or partial count is unknown or partial, never a false zero. Cached counts must not cross organization, namespace, principal or permission boundaries.
- Desktop, mobile and keyboard navigation expose the same entry, accessible count label and selected review category.
Environment
Observed installation reports Loom 1.3.0, build c0c890ab4df0bb8e35ad620d32dcd7180c5157bc, with ClickHouse storage authority. Current-main source was checked separately at f6e14b5c9e5c12e60c79adcc69a2e739567ff438. ClickHouse server version and browser version were not captured; SDK and model provider are not relevant to this navigation report.
Related issues and pull requests
How this was found
An AI tool found it, and I have not reproduced it myself
The original observation is user-reported. An AI agent confirmed the current-main source behavior; it did not perform an independent browser reproduction.
Additional context
Treat the missing proposal count as the first fix, with stable Review navigation in the same workflow. Keep proposal admission policy separate from this navigation change.
- Sidebar visibility and badge calculation
- Sidebar count reads omit proposals
- Review queue categories and default category
Redaction
- I removed tokens, keys, passwords and personal or customer data from everything above.
Activity
zlareb1 commentedon Nov 22, 2026
Additional acceptance criteria: channel coverage
The sidebar itself is a UI control. Its review data and decisions must also be available to headless agents and users.
Expose authorized review summaries, pending counts by kind, filters, pagination and decision state through the API, CLI and supported SDKs, with corresponding MCP/agent operations where permitted. The UI consumes those same summaries.
Compare client results on the same fixed scope and documented freshness window. Restricted items must not leak through aggregate counts; failed counts remain unknown. CLI users need list/count/status output, not a graphical-sidebar equivalent.
Use one authoritative backend contract across HTTP API, CLI, every supported SDK, UI and appropriate MCP/agent tools. Preserve the same permissions, scope, errors, revision checks and idempotency; publishing a feature must not leave a UI-only mutation path.
Document the supported operations and capability matrix, including CLI JSON output and typed SDK request/response examples. Cover Python and TypeScript and any other SDK supported when the feature ships; this requirement does not silently add new SDK languages.
Verify the underlying workflow with the dashboard absent. Optional services and human-only actions keep their existing authorization requirements; interface parity does not mean granting every caller every action.
Related headless deployment requirement: #1146
DeviousCardi commentedon Nov 22, 2026
Fixed by #1153
Pending proposals now surface the Review queue and its sidebar count:
memory.proposalsacceptscount_only— one boundedcount()over the same scope (no value column, no scan cap), so counts match the complete authorized queue beyond one page. Onlystatus = 'proposed'rows count; accepted/rejected proposals and auto-resolved history are excluded.conflicts.listand shows the review-item total, labeled per kind for screen readers ("5 items to review: 2 conflicts, 3 pending proposals"). A failed/partial read is shown as unknown ("at least …"), never a false zero, and remembered answers are cached per org / namespace / principal / role.?tab=proposals.