Admin Models And Maintenance Agent Planning Notes

Status: mixed implementation/planning page, 2026-05-22

This page tracks a grouped set of user-facing configuration and agent-governance requests that need design before implementation. It is intentionally a page, not a structured memory record, because the items describe desired future behavior rather than existing system state.

Immediate UI Corrections

• The Tag Manager should remain a main navigation item for Admin users and should not appear as a tab inside /admin.
• The /admin provider-management tab should be labeled OAuth so it is clear that this surface controls external sign-in providers.
• Editable settings should show the friendly setting name and info icon on the first line, then the configuration key on the next line.
• The Proposals page should remain usable at narrow desktop widths: the proposal list and selected proposal should stay visible, and approve/respond/reject controls should fit in one horizontal row where practical.
• Starting a maintenance run from the Proposals page should immediately show that work is active instead of appearing idle until a refresh.

Model Registry

Implemented baseline: /admin now includes a Models tab backed by ChatModelProfileService. Admins can define chat model profiles with name, provider, provider-discovered model dropdown, optional context-window tokens, enabled state, default chat selection, role allowlist, description, and assignments for maintenance runs, proposal reviews, and Admin Maintenance chat. The dropdown is populated through IChatProvider.ListModelsAsync, so Ollama uses its api/tags results and GitHub/Copilot uses the provider model list path. /chat now selects from enabled model profiles allowed for the current user role and disables send when no enabled default profile is available. Existing installs continue to get an implicit legacy default derived from the older Chat provider/model settings until explicit profiles are configured.

Remaining design work: model profiles still need provider-safe chat settings beyond context-window metadata. Future profile fields should include at least:

Maintenance Agent Tasks And Logs

Implemented baseline: maintenance runs now publish process-local active run state through MaintenanceActiveRunStore, append compact JSONL summaries to MemorySmith:MaintenanceAgent:Storage:ActivityLogPath, and the Proposals page shows active trigger/task/start time plus a Recent task activity panel with recent completed/skipped task runs, timestamps, finding/proposal counts, proposal-id drilldown buttons, and the first warning. LastRunPath remains the single latest-run state file for scheduler/operational checks. /maintenance is now a first-class Admin page with run controls, active-run state, task trace history, proposal action history, non-mutating maintenance-agent chat, transcript search, and proposal links. /admin still includes the Maintenance tab surface for admin workflow continuity. In production DI, Admin Maintenance chat delegates to MemoryChatAgent, so it can use the same read-only wiki tool pipeline as Chat while remaining non-mutating. Transcript persistence trims older entries according to TranscriptRetentionEntries and redacts common token/secret/password patterns when TranscriptRedactionEnabled is true.

Remaining design work: admins still need deeper per-step inspection beyond the current active snapshot, activity rows, proposal action rows, and transcript entries. A later implementation should consider:

• deeper run inspection and richer task drilldown beyond the current active-run snapshot and completed activity rows.

Proposal Review Agent

Implemented baseline: the Proposals page now exposes a Request Agent Review action for actionable proposals. The button calls MaintenanceAgentService.ReviewProposalAsync, records an agent_review_requested history event plus the optional human comment, runs the configured maintenance-agent LLM provider when enabled, records agent_review_completed feedback, and can save a validated revised proposal through SubmitAgentRevisionAsync while preserving the original proposal and diff.

Implemented usability update: proposal details now show a quick human-readable summary before metadata, and the proposal comment box is a full row above the Request Agent Review, Approve, Respond, and Reject button row. Maintenance LLM review parsing accepts fenced or wrapped JSON by extracting the JSON object before deserialization, so fenced JSON responses no longer produce the backtick start-value parse error.

Remaining design work: proposal review still needs richer review verdict metadata, detailed transcript logs, and optional fresh wiki-context retrieval. Users can currently disagree with the review by keeping the original proposal path open and ignoring or rejecting the revised proposal.

Open design questions:

• Should review verdicts become filterable proposal metadata?
• How should conflicting human and agent comments be represented in proposal history?
• Should review agents be allowed to inspect only proposal evidence, or also pull fresh wiki context?

Chat Agent Writes Through Proposals

Implemented current state: standard chat-agent memory/page write approval now submits a maintenance proposal instead of applying writes directly in production DI. The chat turn stores the submitted proposal id in its References drawer, and /proposals remains the review/approval surface that applies file changes after before-text validation. The legacy direct-apply path remains available only when a host constructs MemoryChatAgent without MaintenanceProposalWorkflow.

Open design questions:

• Should viewers/editors be able to draft proposals without approve permission?
• How should proposal history connect back to the chat turn that created it?
• Should chat-submitted proposal links deep-link into a selected proposal once /proposals supports query-string selection?

Chat Compaction Mode

Long-running chat sessions may need a compacting mode. The design should decide whether compaction is automatic, user-triggered, or model-profile controlled. It should preserve auditability: compacted summaries should not erase raw history needed to understand proposal decisions, agent writes, or admin actions.

Incomplete Note

The note We need an admin page that is just the ... is incomplete and needs clarification before it can become an implementation task.