Feedback Loop

The feedback loop lets users tell and show whether a persona-driven UI change is helping, then feeds that signal back into variant selection. It never changes the active UI mid-session — feedback shapes the next decision in the same context.

SDK learning loop (canonical)

The canonical loop lives in @dionysys/server and runs in MCP mode (no beta flags required). After a weak-signal blend decision is applied, feedback updates a per-context Thompson-sampling bandit:

• Explicit keep/revert → bandit.keepReward / bandit.revertReward applied to the (stateId, modality) arm (FeedbackService.submit / evaluate).
• Passive session reward on POST /api/dionysys/sessions/:sessionId/complete → applied at bandit.passiveRewardWeight.

At decision time DecisionService reads the arms for the current context and blends them with the connector via wBandit = n / (n + banditEvidenceK) (n derived from the arm's alpha + beta minus priors). The context key is stateId = "<modality>:<expertise>" (the deterministic guess); the arms are the applied modality. Cold arms reduce the blend to the model's free choice. Tune everything via mcp.gate / mcp.bandit — see Configuration.

Decay (discounted Thompson sampling). By default each reward update first discounts the arm toward its prior (mcp.bandit.decay, effective window 200) so stale evidence fades as new evidence arrives, and a quiet arm is further discounted on read by γ^(days idle) — in memory only. Set decay.enabled: false to keep the original forever-accumulating behaviour. Inspect and maintain the learned arms (posterior mean, credible interval, P(best), snapshot/reset/decay) from the admin console's Bandit tab.

How the demo collects feedback (beta)

The Excalidraw demo adds an optional feedback UI on top of the canonical loop, gated behind beta flags (off by default):

ADAPTIVE_FEEDBACK_BETA_ENABLED=true npm run dev --workspace=@dionysys-demo/excalidraw-backend
VITE_ADAPTIVE_FEEDBACK_BETA_ENABLED=true npm run dev --workspace=@dionysys-demo/excalidraw-frontend

After a decision is applied (a dionysys.decision_applied event starts the metrics window), two signals are collected:

• Explicit — a thumbs prompt ("This helped" / "Got in my way"), gated on engagement rather than the instant the UI changes: at least promptDelayMs (default 30s) since the decision and at least minProductiveActions (default 3) productive actions; it auto-dismisses after autoDismissMs (default 15s). These gates live in useFeedbackTrigger.
• Passive — behavioral metrics over the post-decision window (productive actions, creations/text, deletions, hidden-tool friction), evaluated on its own schedule.

Both reach the server through @dionysys/client over the standard SDK routes:

Signal	Route
Explicit keep/revert	POST /api/dionysys/feedback
Passive evaluation	POST /api/dionysys/feedback/evaluate
Per-session records	GET /api/dionysys/feedback/overview?sessionId=…
Session end (reward)	POST /api/dionysys/sessions/:sessionId/complete

There is one feedback path now — the SDK routes. The demo no longer has its own /api/reward or /api/adaptive-feedback endpoints.

Where it lives

Concern	File
Bandit math (sampleBeta, blend, reward→increments)	packages/core/src/bandit/ThompsonBandit.ts
Posterior stats (mean, credible interval, P(best))	packages/core/src/bandit/posteriorStats.ts
Decay math (discountTowardPrior, effectiveWindowToGamma)	packages/core/src/bandit/decay.ts
Decision blend (reads arms, read-time decay)	packages/server/src/services/DecisionService.ts
Feedback → bandit update (on-update decay)	packages/server/src/services/FeedbackService.ts
Bandit inspector + maintenance	packages/server/src/services/AdminConfigService.ts
Feedback routes	packages/server/src/routes/feedback.ts
Feedback widget	packages/react/src/feedback/AdaptiveFeedback.tsx
Feedback submission hook	packages/react/src/feedback/useFeedback.ts
Prompt timing	packages/react/src/feedback/useFeedbackTrigger.ts
Demo beta wiring	demos/excalidraw/backend/src/services/FeedbackBetaService.ts

Not yet implemented

The pre-extraction demo had extra cross-session machinery that was removed during the SDK split and is not yet reimplemented in @dionysys/server:

• Cross-session browser-prior warm-start — DionysysBrowserPrior exists in the storage contract but is not wired, so each context cold-starts on the model until the bandit warms.
• Graph/LLM recommendation workflow — feedback maps directly to keep/revert/observe; there is no separate recommendation graph.

The bandit learning itself (the canonical section) is fully wired; the items above are known gaps.

Boundaries

• The demo feedback UI is disabled unless both beta flags are true; the SDK bandit loop runs in MCP mode without any flags.
• Feedback applies to the next decision in the same context — never a mid-session auto-revert.

For the broader productization roadmap, see the SDK upgrade spec in the repository.