Skip to content

Triage memory proposals with the assistant and prepare batch decisions #1144

Open

Description

@zlareb1

Company or project name

ClickHouse

Component

Dashboard

Use case

A growing proposal queue is not useful if every administrator must reconstruct the evidence and review each item through a separate interaction. Users want the assistant to help settle the work, with automation where Loom's policy permits it.

This is a focused Review queue workflow under the existing assistant direction in #160. It adds proposal triage and batch decision preparation; it does not replace that issue's general inspection, tuning or assistant implementation.

Describe the solution you'd like

Add “Help review” to the Proposals tab. Pass the current authorized scope and selected proposal revisions to the existing assistant workflow. The assistant groups related proposals, identifies possible duplicates or contradictions, gathers visible source and run evidence, and drafts accept/reject/request-evidence recommendations with reasons and uncertainty.

A reviewer can inspect a group and its exact selected claim revisions, adjust the selection and confirm a bounded batch once. Reuse ordinary proposal decision operations and audit records. This reduces evidence gathering and repeated confirmations without turning model confidence into a new approval authority.

Current memory.proposal_decide requires an administrator, and acceptance requires an independent authenticated human with the displayed content hashes. Preserve that contract. The assistant retains process attribution; it must not run a background acceptance job with a person's credentials or bypass proposal state through copying, moving, rewriting or importing.

Acceptance criteria

  • Start triage from Review with scope and selected rows preserved. Show queue age, why an item needs review, recommendation reasons, relevant evidence, contradictions and missing/unavailable sources. The assistant cannot read or enumerate a broader scope than the caller.
  • Run success, popularity and model confidence are labeled as evidence with limits, not proof of factual truth. Uncertain or conflicting items can be left undecided or sent back for evidence.
  • Show the exact batch before confirmation, including the selected keys, content hashes and proposed decisions. Reauthorize and recheck revisions at Apply. Edited, erased, ambiguous or otherwise ineligible items remain undecided with a per-item reason; retries do not replay successful decisions.
  • Final acceptance uses the deliberate human review action and existing independent-review checks. Self-acceptance, service-account approval and stale-hash acceptance remain refused. A recommendation is never displayed as an already applied decision.
  • Expose and explain existing owner-set proposal_policy automation, including rejection after graded failures. Automatic acceptance remains refused. Any future change to that admission contract is a separate policy decision, not a hidden assistant shortcut.
  • Provide bulk rejection/deferral only where the current API and authorized policy support it. Do not manufacture a decision operation for an unsupported state.
  • The manual queue works when the assistant is disabled. Assistant work is on demand, bounded and cancellable, with its model cost visible; it adds no model call to memory writes or recall.
  • Check the workflow on a labeled proposal set for grounding, unauthorized-source refusal, stale plans, independent reviewer identity, partial batches and decision corrections. Report human review effort and errors rather than claiming that fewer clicks prove better memory quality.

Describe alternatives you've considered

Automatic acceptance based on LLM confidence defeats the current independent-human contract and can promote unsupported claims. Manual review of every item leaves the reported scaling problem. Existing automatic rejection can reduce some work, but it does not assemble evidence or prepare a reviewed batch.

Where the need comes from

The need is mine, an AI tool helped shape the proposal

Additional context

User-requested feature, not a report of a reproduced assistant failure. No model evaluation or mutation of live proposals was run.

Parent direction: #160
Related: #745
Related implementation: #995

This deliverable can reuse the current assistant foundation; it does not depend on completing every feature in #160 and must work on ClickHouse-only deployments.

Related product work: #1141

Activity

  1. zlareb1 commented on Nov 22, 2026

    @zlareb1
    MemberAuthor

    Additional acceptance criteria: channel coverage

    “Help review” is a UI shortcut for a capability that must also work headlessly.

    • Expose proposal selection, evidence gathering, triage, batch-plan preview, per-item status and permitted decision operations through API, CLI and supported SDKs, with corresponding MCP/agent tools. Do not require chat-UI state to run triage or inspect its plan.

    • A properly authenticated independent human can review and confirm the exact batch through CLI/API/SDK as well as the UI. Retain the same reviewer identity, selected content hashes, revision checks and audit receipts on every route.

    • Service-account/agent callers cannot gain human approval authority by changing interface. They may prepare recommendations or use existing authorized rejection policy; automatic acceptance remains refused under the current contract.

    • 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

  2. added
    up-for-grabsContributor tasks marked help wanted; sizes and skill requirements vary.
    on Nov 22, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requesthelp wantedExtra attention is neededup-for-grabsContributor tasks marked help wanted; sizes and skill requirements vary.

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions