feat(selfhost): per-repo feature config + all-authors AI review + RAG embed stack

migrations/0075_ai_review_all_authors.sql

@@ -0,0 +1,5 @@
+-- AI review for all authors (per-repo opt-in). The AI maintainer review is confirmed-contributor-gated by
+-- default (an AI-spend guard, see runAiReviewForAdvisory). `ai_review_all_authors` lets a self-host operator
+-- run the review for EVERY PR's author — intended for an operator who wants real reviews on all PRs (incl. their
+-- own) and pays for the AI themselves. Default 0 (off) — additive, existing repos are byte-identical.
+ALTER TABLE repository_settings ADD COLUMN ai_review_all_authors INTEGER NOT NULL DEFAULT 0;

src/api/routes.ts

@@ -233,6 +233,7 @@ import { buildRepoOutcomeCalibration } from "../services/outcome-calibration";
 import { loadGatePrecisionReport } from "../services/gate-precision";
 import { computeOpsStats, isOpsEnabled } from "../review/ops-wire";
 import { computeParityReadiness, isParityAuditEnabled } from "../review/parity-wire";
+import { isRagEnabled } from "../review/rag-wire";
 import { getPublicStats, isPublicStatsEnabled } from "../review/public-stats";
 import { buildMaintainerQualityDashboard, isMaintainerQualityDataStale } from "../services/maintainer-quality-dashboard";
 import { MAX_LOCAL_SCORER_WARNING_CHARS, MAX_LOCAL_SCORER_WARNING_COUNT } from "../signals/local-scorer-diagnostics";
@@ -619,6 +620,7 @@ const repositorySettingsSchema = z.object({
   aiReviewByok: z.boolean().default(false),
   aiReviewProvider: z.enum(["anthropic", "openai"]).nullable().optional(),
   aiReviewModel: z.string().trim().min(1).max(120).nullable().optional(),
+  aiReviewAllAuthors: z.boolean().default(false),
   autoLabelEnabled: z.boolean().default(true),
   gittensorLabel: z.string().trim().min(1).max(50).default("gittensor"),
   blacklistLabel: z.string().trim().min(1).max(50).default("slop"),
@@ -710,6 +712,7 @@ const repositoryAiReviewSchema = z.object({
   byok: z.boolean().default(false),
   provider: z.enum(["anthropic", "openai"]).nullable().optional(),
   model: z.string().trim().min(1).max(120).nullable().optional(),
+  allAuthors: z.boolean().default(false),
 });
 
 const contributorIssueDraftGenerateSchema = z.object({
@@ -2246,13 +2249,15 @@ export function createApp() {
       aiReviewByok: parsed.data.byok,
       aiReviewProvider: parsed.data.provider,
       aiReviewModel: parsed.data.model,
+      aiReviewAllAuthors: parsed.data.allAuthors,
     });
     // getRepositorySettings normalizes these to a concrete value or null (never undefined).
     return c.json({
       aiReviewMode: updated.aiReviewMode,
       aiReviewByok: updated.aiReviewByok,
       aiReviewProvider: updated.aiReviewProvider ?? null,
       aiReviewModel: updated.aiReviewModel ?? null,
+      aiReviewAllAuthors: updated.aiReviewAllAuthors,
     });
   });
 
@@ -3056,6 +3061,21 @@ export function createApp() {
     return c.json({ ok: true, status: "queued" }, 202);
   });
 
+  // Operator-facing RAG (re)index trigger for a self-host maintainer. Bearer-gated by the `/v1/internal/*`
+  // middleware (INTERNAL_JOB_TOKEN). With NO body it enqueues the fan-out (re-indexes every RAG-active configured +
+  // registered repo); with `{ "repoFullName": "owner/repo" }` it indexes just that repo. Either way the job is
+  // gated downstream by convergedFeatureActive, so a repo where RAG is off is a no-op. 404 when RAG is globally off
+  // so the endpoint doesn't exist on a deploy that isn't running RAG. This is how an operator adds/indexes a new
+  // repo on demand instead of waiting for the 6-hourly cron.
+  app.post("/v1/internal/jobs/rag-index", async (c) => {
+    if (!isRagEnabled(c.env)) return c.json({ error: "not_found" }, 404);
+    const body = (await c.req.json().catch(() => ({}))) as { repoFullName?: unknown };
+    const repoFullName = typeof body?.repoFullName === "string" && body.repoFullName.trim().length > 0 ? body.repoFullName.trim() : undefined;
+    const message: JobMessage = { type: "rag-index-repo", requestedBy: "api", ...(repoFullName ? { repoFullName } : {}) };
+    await c.env.JOBS.send(message);
+    return c.json({ ok: true, status: "queued", scope: repoFullName ?? "all-configured-repos" }, 202);
+  });
+
   app.post("/v1/internal/jobs/refresh-registry/run", async (c) => {
     return c.json(await refreshRegistry(c.env));
   });
@@ -3365,6 +3385,7 @@ export function createApp() {
         aiReviewByok: parsed.data.aiReviewByok,
         aiReviewProvider: parsed.data.aiReviewProvider,
         aiReviewModel: parsed.data.aiReviewModel,
+        aiReviewAllAuthors: parsed.data.aiReviewAllAuthors,
         autoLabelEnabled: parsed.data.autoLabelEnabled,
         gittensorLabel: parsed.data.gittensorLabel,
         blacklistLabel: parsed.data.blacklistLabel,

src/db/repositories.ts

@@ -431,6 +431,7 @@ export async function getRepositorySettings(env: Env, fullName: string): Promise
       aiReviewByok: false,
       aiReviewProvider: null,
       aiReviewModel: null,
+      aiReviewAllAuthors: false,
       autoLabelEnabled: true,
       gittensorLabel: "gittensor",
       blacklistLabel: "slop",
@@ -473,6 +474,7 @@ export async function getRepositorySettings(env: Env, fullName: string): Promise
     aiReviewByok: row.aiReviewByok,
     aiReviewProvider: normalizeAiReviewProvider(row.aiReviewProvider),
     aiReviewModel: row.aiReviewModel ?? null,
+    aiReviewAllAuthors: row.aiReviewAllAuthors,
     autoLabelEnabled: row.autoLabelEnabled,
     gittensorLabel: row.gittensorLabel,
     blacklistLabel: row.blacklistLabel,
@@ -519,6 +521,7 @@ export async function upsertRepositorySettings(env: Env, settings: Partial<Repos
     aiReviewByok: settings.aiReviewByok ?? false,
     aiReviewProvider: normalizeAiReviewProvider(settings.aiReviewProvider),
     aiReviewModel: typeof settings.aiReviewModel === "string" && settings.aiReviewModel.trim() ? settings.aiReviewModel.trim() : null,
+    aiReviewAllAuthors: settings.aiReviewAllAuthors ?? false,
     autoLabelEnabled: settings.autoLabelEnabled ?? true,
     gittensorLabel: settings.gittensorLabel ?? "gittensor",
     blacklistLabel: settings.blacklistLabel ?? "slop",
@@ -563,6 +566,7 @@ export async function upsertRepositorySettings(env: Env, settings: Partial<Repos
       aiReviewByok: resolved.aiReviewByok,
       aiReviewProvider: resolved.aiReviewProvider,
       aiReviewModel: resolved.aiReviewModel,
+      aiReviewAllAuthors: resolved.aiReviewAllAuthors,
       autoLabelEnabled: resolved.autoLabelEnabled,
       gittensorLabel: resolved.gittensorLabel,
       blacklistLabel: resolved.blacklistLabel,
@@ -608,6 +612,7 @@ export async function upsertRepositorySettings(env: Env, settings: Partial<Repos
         aiReviewByok: resolved.aiReviewByok,
         aiReviewProvider: resolved.aiReviewProvider,
         aiReviewModel: resolved.aiReviewModel,
+        aiReviewAllAuthors: resolved.aiReviewAllAuthors,
         autoLabelEnabled: resolved.autoLabelEnabled,
         gittensorLabel: resolved.gittensorLabel,
         blacklistLabel: resolved.blacklistLabel,

src/db/schema.ts

@@ -66,6 +66,7 @@ export const repositorySettings = sqliteTable("repository_settings", {
   aiReviewByok: integer("ai_review_byok", { mode: "boolean" }).notNull().default(false),
   aiReviewProvider: text("ai_review_provider"),
   aiReviewModel: text("ai_review_model"),
+  aiReviewAllAuthors: integer("ai_review_all_authors", { mode: "boolean" }).notNull().default(false),
   autoLabelEnabled: integer("auto_label_enabled", { mode: "boolean" }).notNull().default(true),
   gittensorLabel: text("gittensor_label").notNull().default("gittensor"),
   // Label applied to a blacklisted contributor's PR/issue (#1425); configurable so the disposition works

src/env.d.ts

@@ -4,6 +4,11 @@ declare global {
     JOBS: Queue;
     RATE_LIMITER?: DurableObjectNamespace;
     AI?: Ai;
+    /** Self-host (RAG): a DEDICATED embedding provider, kept SEPARATE from the review chat chain so the reviewer
+     *  stays frontier-only (claude-code/codex) while embeddings — which those CLIs cannot produce — route to a
+     *  local/openai-compatible endpoint (ollama). Built at boot from AI_EMBED_BASE_URL/AI_EMBED_MODEL. Absent ⇒
+     *  `createReviewAdapters` falls back to `env.AI` (byte-identical to before). */
+    AI_EMBED?: Ai;
     /** Convergence (infra): Vectorize index for codebase RAG retrieval (Layer C). Optional — the review is
      *  fully fail-safe without it (absent ⇒ no RAG, review proceeds with no retrieved context). The index is
      *  created with bge-m3's 1024 dimensions. Unused until the per-module RAG wiring chunk; an unbound deploy