[/improve] Self-improving agent UI — per-profile health + proposal/approval queue

[Interstellar-code]

Summary

Add an /improve page to Switch UI that monitors per-profile / per-skill agent health and drives the self-improving harness loop (Karpathy autoresearch ratchet) as a human-in-the-loop approval queue. This page is a thin consumer of the agent-side engine API (/api/improve/*) — all heavy machinery (metrics store, eval runner, meta-agent proposer, git ratchet, experiment state machine) lives in the hermes-agent plugin.

Companion / dependency: agent engine issue → Interstellar-code/hermes-agent#133. This UI issue is blocked on that plugin's /api/improve/* surface.

This is the UI half of docs/self-improving-agent-proposal.md (§4, §6c). Engine/agent work is out of scope here.

Architecture

• Consume /api/improve/* via the existing dashboard-proxy pattern (same as jobs/kanban).
• Capability-probe the endpoint — hide the /improve nav + page when the agent plugin isn't installed/enabled (mirror jobs/kanban probing).
• No business logic in the UI: it renders proposals, scores, diffs, history, and posts approve/reject/pause actions back to the API.

Page structure (/improve route)

1. Per-profile observability (P0 — zero-risk, immediate value)

Scorecard built from data the dashboard already fetches, scoped per profile:

• sessions/day, error+warn rate, cost trend, token efficiency, retries.
• Per-skill row (P4): invocations, error rate (logs↔sessions correlated), retries, avg tokens/cost, last-used, trend.
• Trends sourced from the agent metrics store via API (survives restarts).
• This section alone answers "is my harness degrading?" and ships before any loop.

2. Proposal queue (state: proposed)

Card per pending experiment:

• Side-by-side old vs new diff (SOUL.md / profile system_prompt) — exactly one atomic change, highlighted.
• Meta-agent rationale.
• Offline eval table: per-scenario pass/fail before vs after, aggregate delta, eval token cost.
• Actions: Approve / Reject / Edit-then-approve (reject is logged so the idea isn't re-proposed).

3. Observation window (state: live)

• Progress: "12 / 30 sessions observed" (window configurable per profile).
• Live metrics vs baseline: error/warn rate, completion, retries, token efficiency, periodic LLM-judge spot-scores.
• This is the second verification stage (production proof, beyond offline eval).

4. Verdict + History

• Verified → baseline vN+1; regressed → auto-revert notification ("experiment feat: Terminal Matrix revamp + Tasks kanban polish + sidebar unification #14 reverted: error rate +18%").
• History tab: results.tsv-equivalent — every experiment with diff, offline + live scores, verdict, cost.
• Baseline version curve per profile — score ratcheting up over time.

Controls

• Pause/resume per profile (pause stops new proposals; in-flight experiment finishes its window).
• One-experiment-in-flight-per-profile reflected in UI state.

Phasing (UI-side, tracks agent phases)

• P0 — observability page (consume metrics endpoints). Ship first, standalone value.
• P1 — surface scenario-suite runs + results (manual "run eval" trigger).
• P2 — proposal queue + approve/reject/edit + observation-window cards.
• P3 — memory-hygiene / USER.md staleness views (separate metrics).
• P4 — extend scorecards + queue to skills/plugins (reuses everything).

Out of scope (separate issue)

• The engine, metrics store, eval runner, meta-agent, git ratchet, and /api/improve/* implementation → Interstellar-code/hermes-agent#133.

Reference

docs/self-improving-agent-proposal.md — §4 (what Switch UI already has: profile file read/write, analytics/logs/sessions APIs), §6c (/improve experiment lifecycle spec). Never-upstream fork differentiator ("Skill Health / Agent Improvement" page).