A tier add aligns every organization holding a live provision of the product inside one transaction that holds the exclusive rendezvous, so its hold grew with the product's holders, and it locked their pools in the order their provisions were created, the one catalog path that did not lock in pool id order (#194). It checked the product outside its transaction, answered "Product not found." for any failed read, showed the database's text when a pool already held a rung, and answered a lock wait with "see server logs" (#216). AddTier now runs on runPlanAct. Inside the exclusive transaction it: - reads the product and refuses one that is missing ("Product not found.") or not published; any other failed read is a failed add; - counts the holders and refuses above planActHolderCap, naming the count and the limit, and, when more than half hold the product through an org type's default on a ladder, the way through: a default change with Migrate; - locks the holders' pools in ascending id, reads their provisions again, records the act (tier_added, with names read inside) and inserts the tier; - reads ListTierAddCollisions for the locked pools and refuses the first pool, in pool id order, that already holds a rung on the ladder or holds two live provisions of the product, naming the organization and the product, before any alignment; - aligns each pool and materializes it once, with a reason that names the ladder rather than its id. planActHolderCap is 200. A timed run on PostgreSQL 18.1, one run per holder count from 50 to 3,200, measured a removal with migrate at 2.2 to 2.6 ms per holder and an add at 0.7 to 1.0 ms. The removal, which takes the same cap in a later change, binds: 200 holders hold the rendezvous for about half a second. The timing test lifts the cap so it can still measure past it. CreatePlanLadderTier only parses, calls and renders. A lock wait or a deadlock at any step answers "Another change collided with this one; try again." A collision answers under the product field. Every other failure, including a CF003 the collision read should have prevented, answers "Failed to add tier. Reference: {request id}." and logs the step and the driver text through the request's logger; the driver text is never shown. Goldens: add-aligns-a-live-provision and tier-change-during-conferral/ a-tier-add gain the act row, the ascending pool locks and the ladder's name in the reason; add-begin-fails answers with the reference and logs the step; add-collides-with-a-held-rung is refused by the collision read with no alignment statement; contention's add answers the collision sentence. Placeholders renumber in three of them because a request id or a provision id moved. New: add-above-the-limit, add-above-the-limit-default-sourced, add-product-unread and add-product-held-twice. The full suite passes.
Server Package
This package implements the HTTP server for the member-console application, including HTMX partial handlers.
Overview
The server uses Go's standard net/http package with:
- Bootstrap 5 for styling
- HTMX 2.0 for dynamic content updates
- gorilla/csrf for CSRF protection
- scs (
alexedwards/scs) for server-side session management (Valkey/Redis-backed)
Architecture
The HTTP server handles several types of requests:
- Static Files (
/static/*) - CSS, JS, images served from embedded assets - Page Handlers (
/) - Full HTML page rendering - Partial Handlers (
/partials/*) - HTMX partial HTML responses - Auth Handlers (
/login,/logout,/callback) - OAuth2/OIDC authentication
Partial handlers interact with Temporal workflows for async operations like site provisioning.
Files
server.go- Main server setup, middleware chain, route registrationmember_plans.go,member_plan_moves.go,member_entitlements.go,member_addons.go,member_invoices.go,workspace_partials.go,operator*.go- HTMX partial handlers for this package's own featuresprice_text.go- price text shared by the member and operator pages
Integration-owned handlers (e.g. FedWiki's member-facing API and partials)
live in their own tree under internal/integrations/<key>/web instead of
here — see docs/building-an-integration.md and
internal/integrations/fedwiki/web for the worked example. The patterns
below apply equally to handlers in either location.
HTMX Integration
Partial Handler Pattern
Partial handlers return HTML fragments (not full pages) that HTMX swaps into the DOM:
// Handler returns HTML fragment
func (h *Handler) GetItems(w http.ResponseWriter, r *http.Request) {
// ... fetch data ...
h.Templates.Render(w, "items.html", data)
}
<!-- Template returns partial HTML -->
<div class="table-responsive">
<table class="table">
{{ range .Sites }}
<tr>...</tr>
{{ end }}
</table>
</div>
HTMX Triggers
Use HX-Trigger header to trigger events that refresh other parts of the page:
// After successful operation, trigger refresh
w.Header().Set("HX-Trigger", `{"refreshSites": true}`)
<!-- Element listens for the trigger -->
<div hx-get="/partials/items"
hx-trigger="load, refreshItems from:body">
</div>
Cross-origin protection with HTMX
There is no CSRF token. internal/middleware/csrf.go wraps net/http's
CrossOriginProtection, which refuses a non-safe request whose
Sec-Fetch-Site (or, failing that, Origin) says it came from another
site. An htmx request from one of the console's own pages is same-origin
and passes with nothing added; a refusal is a 403 whose body names the
CSRF check, which error-handler.js reads to keep the person's input.
Workflow Execution
Handlers that start a Temporal workflow for a member's request answer when the start is confirmed, not when the workflow finishes. An integration records the request first (FedWiki's site lifecycle records a row in core.lifecycle_requests), starts the workflow under the identifier stored on that row, and returns a fragment or an HX-Trigger refresh; the workflow records its own outcome, and the list that shows the request re-renders until the outcome lands. A handler never waits for a workflow to finish: one that reads a result (the Stripe environment check) waits on WorkflowRun.Get under a context of a few seconds and answers that the check started when it runs out, because a workflow can outlast the server's write timeout. FedWiki starts a recorded request's workflow detached from the member's request, so a closed connection cannot leave the request with no workflow.
Route Organization
Routes are organized by feature using handler structs:
// ItemPartialsHandler groups related routes
type ItemPartialsHandler struct { ... }
func (h *ItemPartialsHandler) RegisterRoutes(mux *http.ServeMux) {
mux.HandleFunc("GET /partials/items", h.GetItems)
mux.HandleFunc("POST /partials/items", h.CreateItem)
mux.HandleFunc("DELETE /partials/items/{id}", h.DeleteItem)
// ...
}
Error Rendering
Errors are rendered back into the same partial/modal that initiated the request:
func (h *Handler) renderCreateFormError(w http.ResponseWriter, userID int64, errMsg string) {
// Re-fetch form data
data := h.getCreateFormData(userID)
data.Error = errMsg
// Re-render the form partial with error displayed
h.Templates.Render(w, "create_form.html", data)
}
This keeps the user in context - they see the error in the same modal/form they were using.
Middleware Stack
handler = middleware.Logging(logger)(handler) // Request logging
handler = middleware.Recover(logger)(handler) // Panic recovery
handler = middleware.Compress()(handler) // Gzip compression
handler = middleware.SecurityHeaders()(handler) // Security headers
handler = middleware.RequestID()(handler) // Request ID generation
handler = csrf.Protect(key, opts...)(handler) // CSRF protection