diff --git a/docs/arch/README.md b/docs/arch/README.md
index ecce68d9..31743fd5 100644
--- a/docs/arch/README.md
+++ b/docs/arch/README.md
@@ -24,7 +24,7 @@ Cross-layer reads go via HTTP (`$fetch('/api/...')`), never shared server utilit
 
 ## Server-side architecture
 
-- **Platform**: request-scoped object on `event.context.platform` carrying every infrastructure binding a handler needs: `db` (D1), `ai` (Workers AI), `github` (resolved client), `requestId`. Mounted by `server/plugins/platform.ts`. Handlers must read bindings from here, never directly from `event.context.cloudflare.env`. `db` is one D1 session per request: a safe method starts on any read replica, a mutating method starts on the primary, and a write's bookmark rides a `d1-bookmark` cookie so the next request reads it. Cron tasks and queue consumers use the raw binding, which always reaches the primary.
+- **Platform**: request-scoped object on `event.context.platform` carrying every infrastructure binding a handler needs: `db` (D1), `ai` (Workers AI), `github` (resolved client), `requestId`. Mounted by `server/plugins/platform.ts`. Handlers must read bindings from here, never directly from `event.context.cloudflare.env`. `db` is one D1 session per request. Safe methods use an incoming bookmark or `first-unconstrained`; mutating methods start with `first-primary`. A write's bookmark rides a `d1-bookmark` cookie so the next request reads it. With replication disabled, all session queries use the primary. Keep replication disabled for `skilld-db`, following the [D1 reliability runbook](../runbooks/d1-reliability.md). Cron tasks and queue consumers use the raw binding, which always reaches the primary.
 - **defineApiHandler**: the shape of every internal API route: `{ schema, policy, handler, presenter }`. Schema is a zod input, policy is an array of atomic predicates AND-ed, handler receives `{ body, platform, user, event }`, presenter shapes the response. Defined in `shared/server/handler.ts`.
 - **Policy**: a `(ctx) => boolean | Promise<boolean>` atom in `layers/<layer>/server/policies/`. Composes by array.
 - **Presenter**: a `(row) => dto` in `layers/<layer>/server/presenters/`. Response shape lives here, never inline in handlers.
diff --git a/docs/runbooks/d1-reliability.md b/docs/runbooks/d1-reliability.md
new file mode 100644
index 00000000..270b6e31
--- /dev/null
+++ b/docs/runbooks/d1-reliability.md
@@ -0,0 +1,69 @@
+# D1 reliability
+
+Keep read replication disabled for `skilld-db`.
+Use this runbook to remove restoring replicas from production Skill reads.
+Codex prepared this runbook for Harlan Wilton.
+
+Cloudflare [case 02370111](https://www.support.cloudflare.com/s/case/500Nv00000mk4JZ) identified a replica restoration loop.
+Replicas restore a snapshot while the primary continues writing.
+A replica that cannot catch up disconnects and starts restoring again.
+Reads waiting on that replica fail during this cycle.
+
+Cloudflare attributed the October 5 and 7 internal errors to reads waiting on a restoring European replica.
+Individual error references could not be mapped one-to-one to retained logs.
+Engineering considers the October 9 storage timeout a separate event that coincided with primary restarts.
+
+## Disable replication
+
+Use `cf` with access to the configured Cloudflare account and `D1:Edit` permission.
+Run these commands from the repository root.
+Check the database name and account in the dry run before applying the change.
+
+```sh
+cf d1 get a5e53f35-f5e5-4987-8c67-c0175addc7cc
+cf d1 edit a5e53f35-f5e5-4987-8c67-c0175addc7cc --read-replication-mode disabled --dry-run
+cf d1 edit a5e53f35-f5e5-4987-8c67-c0175addc7cc --read-replication-mode disabled
+cf d1 get a5e53f35-f5e5-4987-8c67-c0175addc7cc
+```
+
+The final response must name `skilld-db` and return `read_replication.mode: disabled`.
+This changes database configuration without a Worker deployment or a schema migration.
+Cloudflare support advises allowing up to 24 hours for existing replicas to stop processing requests.
+A successful request during that interval does not prove replica retirement.
+
+Without replication, D1 routes reads and writes to the primary.
+Requests from Europe and the Americas may take longer because the primary is in Oceania.
+See [Cloudflare's replication model](https://developers.cloudflare.com/d1/best-practices/read-replication/#primary-database-instance-vs-read-replicas).
+
+## Check production reads
+
+Use the existing page, redirect, and asset checks:
+
+```sh
+pnpm production:smoke
+cf d1 query a5e53f35-f5e5-4987-8c67-c0175addc7cc --sql "SELECT owner, repo, name FROM skills WHERE owner = 'anthropics' AND repo = 'skills' AND name = 'skill-creator' LIMIT 1"
+```
+
+The smoke result must return `_tag: passed`.
+The query must return the Skill with `success: true` and `meta.served_by_primary: true`.
+The REST query checks the primary directly; it does not prove where Worker session reads ran.
+
+After the retirement interval, inspect Sentry for new internal errors and replica-disconnection errors.
+Keep primary storage timeouts separate when comparing recurrences.
+If errors continue, give Cloudflare the UTC timestamps, deployed release, and full error cause.
+
+## Preserve session consistency
+
+Keep the request sessions and bookmark cookie in [`server/utils/db.ts`](../../server/utils/db.ts).
+Disabling replication keeps those constraints valid while D1 serves queries from the primary.
+`first-primary` directs only the first query to the primary; later session queries can use replicas when replication is enabled.
+See [D1 session constraints](https://developers.cloudflare.com/d1/worker-api/d1-database/#withsession).
+
+D1 automatically retries eligible read-only queries up to twice.
+Cloudflare support specifies a 45-second retry budget and excludes overloaded failures from those automatic retries.
+The reported lifecycle failures lasted approximately 47 to 93 seconds, exceeding that budget.
+See [D1 automatic retries](https://developers.cloudflare.com/d1/observability/debug-d1/#automatic-retries).
+
+Application retries must preserve bookmarks and the request deadline, with bounded exponential backoff and jitter.
+Retry only genuinely read-only queries, or writes whose idempotency has been established.
+Keep replication disabled unless Cloudflare confirms the restoration loop is fixed and regional reads pass checks.