docs(ai): reword stale workers-ai-only comments across the review stack

.env.example

@@ -463,11 +463,13 @@ REDIS_URL=redis://redis:6379                # REQUIRED for the self-host review
 #                                            #   synthesis = both review, then ONE merged decision.
 # AI_ON_MERGE=either                         # synthesis merge rule: either (block if EITHER reviewer flags) |
 #                                            #   both (block only when both do). Ignored unless AI_COMBINE=synthesis.
-# AI_DAILY_NEURON_BUDGET=10000000            # daily spend cap (Cloudflare Workers AI "neurons") shared by AI
-#                                            #   summaries, the free consensus-defect reviewer pair, and the
-#                                            #   AI-slop scorer. Default 10,000,000/day; clamped to [0, 10000000].
-#                                            #   Only meters FREE Workers-AI calls — BYOK/provider calls above bill
-#                                            #   the maintainer's own account and are not counted against this.
+# AI_DAILY_NEURON_BUDGET=10000000            # daily spend cap (unit name is a Cloudflare Workers AI "neurons"
+#                                            #   holdover; applied as a provider-agnostic heuristic budget) shared
+#                                            #   by AI summaries, the free/default consensus-defect reviewer pair,
+#                                            #   and the AI-slop scorer. Default 10,000,000/day; clamped to
+#                                            #   [0, 10000000]. Only meters the free/default-reviewer path —
+#                                            #   BYOK/provider calls above bill the maintainer's own account and
+#                                            #   are not counted against this.
 # Ollama reviewer (AI_PROVIDER=ollama). Defaults: OLLAMA_AI_BASE_URL=http://localhost:11434/v1,
 # OLLAMA_AI_MODEL=llama3.1, no API key. Set the base URL to http://ollama:11434/v1 when using the compose
 # --profile ollama service.

.gittensory.yml

@@ -40,7 +40,7 @@ gate:
     minScore: 60
   # aiReview:                  # opt-in AI maintainer review (off by default; needs the AI flags enabled)
   #   mode: advisory           # block | advisory | off — block only blocks on a dual-model consensus defect
-  #   byok: false              # use a maintainer Anthropic/OpenAI key for the write-up; consensus stays free Workers AI
+  #   byok: false              # use a maintainer Anthropic/OpenAI key for the write-up; consensus stays on the free/default reviewer
   #   allAuthors: false        # true reviews every PR author with the selected self-host model(s)
   #   provider: anthropic      # anthropic | openai — which BYOK provider (the secret key is set via the dashboard, never here)
   #   model: claude-3-5-sonnet-latest   # optional model override for the BYOK write-up

src/config/gittensory-repo-focus-manifest.ts

@@ -44,7 +44,7 @@ gate:
     minScore: 60
   # aiReview:                  # opt-in AI maintainer review (off by default; needs the AI flags enabled)
   #   mode: advisory           # block | advisory | off — block only blocks on a dual-model consensus defect
-  #   byok: false              # use a maintainer Anthropic/OpenAI key for the write-up; consensus stays free Workers AI
+  #   byok: false              # use a maintainer Anthropic/OpenAI key for the write-up; consensus stays on the free/default reviewer
   #   allAuthors: false        # true reviews every PR author with the selected self-host model(s)
   #   provider: anthropic      # anthropic | openai — which BYOK provider (the secret key is set via the dashboard, never here)
   #   model: claude-3-5-sonnet-latest   # optional model override for the BYOK write-up

src/db/repositories.ts

@@ -962,8 +962,9 @@ async function recordAiKeyChange(
 
 /**
  * Decrypt a repo's BYOK key for an AI call. Returns null when no key is configured OR the encryption
- * secret is unavailable OR decryption fails — so the caller silently falls back to free Workers AI and
- * a misconfiguration never blocks the review. The plaintext key must be used immediately and never cached.
+ * secret is unavailable OR decryption fails — so the caller silently falls back to the free/default
+ * reviewer and a misconfiguration never blocks the review. The plaintext key must be used immediately
+ * and never cached.
  */
 export async function getDecryptedRepositoryAiKey(env: Env, fullName: string): Promise<DecryptedRepositoryAiKey | null> {
   const secret = env.TOKEN_ENCRYPTION_SECRET;

src/env.d.ts

@@ -47,7 +47,13 @@ declare global {
     PUBLIC_SITE_ORIGIN?: string;
     AI_SUMMARIES_ENABLED?: string;
     AI_PUBLIC_COMMENTS_ENABLED?: string;
+    /** Model id for a genuine Cloudflare Workers AI binding only — no live deployment (hosted or self-host)
+     *  binds `env.AI` to Workers AI today (see CONVERGENCE_RUNBOOK.md), and self-host discards any
+     *  `@cf/`-prefixed value here. Self-host operators should use the provider-specific `*_AI_MODEL` vars below. */
     WORKERS_AI_SUMMARY_MODEL?: string;
+    /** Daily spend cap in Cloudflare Workers AI "neurons" for the free/default-reviewer path (shared across
+     *  ai-review/ai-slop/ai-summaries/planner). The unit name is a Workers-AI holdover; it's applied as a
+     *  provider-agnostic heuristic budget regardless of which configured provider actually serves the request. */
     AI_DAILY_NEURON_BUDGET?: string;
     /** Per-repository/day cap for maintainer-paid BYOK AI review provider calls. */
     AI_BYOK_DAILY_REPO_LIMIT?: string;

src/review/adapters.ts

@@ -52,8 +52,9 @@ export function reviewVectorAdapter(vectorize: Vectorize): VectorAdapter {
   };
 }
 
-// ── Inference (Workers AI → InferenceAdapter). Feature-gated. Mirrors `ai.run(model, options)`; the cast
-//    bridges the overloaded Workers-AI `run` to the portable single-signature shape. ──
+// ── Inference (the Ai-shaped adapter → InferenceAdapter). Feature-gated. Mirrors `ai.run(model, options)`;
+//    the cast bridges the overloaded `run` signature to the portable single-signature shape. `ai` is
+//    Workers AI historically, and on self-host is the generic provider router (src/selfhost/ai.ts). ──
 export function reviewInferenceAdapter(ai: Ai): InferenceAdapter {
   return { run: (model, options) => (ai as unknown as { run(m: string, o: Record<string, unknown>): Promise<unknown> }).run(model, options) };
 }

src/review/rag.ts

@@ -22,7 +22,7 @@
 // additive module. The host injects concrete adapters at the call site.
 
 // ── Injected infra interfaces (inlined from reviewbot src/platform/types.ts) ──────────────────────
-// These mirror the platform-adapter shapes so the host can pass its Vectorize/Workers-AI/D1-backed
+// These mirror the platform-adapter shapes so the host can pass its Vectorize/self-host-AI/D1-backed
 // implementations unchanged; nothing here depends on env bindings.
 
 /** Vector search surface (Vectorize → Qdrant / pgvector / sqlite-vec). */
@@ -43,7 +43,8 @@ export interface VectorAdapter {
   deleteByIds(ids: string[]): Promise<void>;
 }
 
-/** Inference (Workers AI → Ollama / OpenAI-compatible). Mirrors `ai.run(model, options)`. */
+/** Inference (the configured AI provider — self-host Codex/Claude Code/Ollama/OpenAI-compatible, or the
+ *  legacy Workers-AI binding). Mirrors `ai.run(model, options)`. */
 export interface InferenceAdapter {
   run(model: string, options: Record<string, unknown>): Promise<unknown>;
 }
@@ -87,7 +88,10 @@ export type RagRetrievalResult = {
 };
 
 /** bge-m3: large context window → a whole file/function embeds as one coherent chunk (fewer vectors
- *  than 512-token models, which helps both quality and the free-tier vector budget). */
+ *  than 512-token models, which helps both quality and the free-tier vector budget). This is a Workers-AI
+ *  model id; the self-host embed path (`createOpenAiCompatibleAi` in src/selfhost/ai.ts) discards any
+ *  `@cf/`-prefixed id and substitutes its own configured/default embed model (`AI_EMBED_MODEL`), so this
+ *  constant only matters for a genuine Cloudflare Workers AI inference binding. */
 export const EMBED_MODEL = "@cf/baai/bge-m3";
 /** Default bge-m3 output dimension. Self-host can override this when QDRANT_DIM selects another model width. */
 export const RAG_DIMENSIONS = 1024;
@@ -96,7 +100,8 @@ const CHUNK_CHARS = 16000; // per-file chunk budget; only files larger than this
 const CHUNK_OVERLAP = 1500;
 /** Hard per-repo stored-vector cap — the free-tier guard. Source is prioritized so it survives the cap. */
 export const MAX_CHUNKS_PER_REPO = 1500;
-const EMBED_BATCH = 96; // Workers AI caps embedding input at 100 items/call
+const EMBED_BATCH = 96; // Workers AI caps embedding input at 100 items/call; kept as a conservative general
+// bound — other embed providers (Ollama/vLLM/etc via the self-host adapter) may not share this exact cap.
 const MAX_CONTEXT_CHARS = 14000; // bound the injected block (mirrors diff/knowledge budgets)
 export const MAX_FILE_BYTES = 1_000_000; // skip files larger than ~1MB
 

src/types.ts

@@ -677,7 +677,8 @@ export type RepositorySettings = {
   /** Slop-risk threshold (0-100) at/above which `slopGateMode: block` blocks. Default 60 (the `high` band). */
   slopGateMinScore?: number | null | undefined;
   /** AI-assisted slop advisory (the `slopAiAdvisory` capability). When true AND `slopGateMode != off`, a
-   *  free Workers-AI pass adds an ADVISORY-only `ai_slop_advisory` finding for semantic slop the
+   *  free/default-reviewer pass (the configured self-host provider, or the legacy Workers-AI pair when
+   *  none is configured) adds an ADVISORY-only `ai_slop_advisory` finding for semantic slop the
    *  deterministic detector cannot quantify. It NEVER feeds slopRisk or the gate (only the deterministic
    *  core blocks). Default false — opt-in via `.gittensory.yml gate.slop.aiAdvisory`. */
   slopAiAdvisory: boolean;
@@ -686,13 +687,14 @@ export type RepositorySettings = {
    *  like every other blocker). Default `off` — AI is opt-in. */
   aiReviewMode: GateRuleMode;
   /** Bring-your-own-key: when true and a provider key is configured for the repo, the advisory AI review
-   *  is generated by the maintainer's frontier model (Anthropic/OpenAI) instead of free Workers AI. The
-   *  consensus blocker always uses the free Workers-AI model pair regardless, so BYOK never changes who
+   *  is generated by the maintainer's frontier model (Anthropic/OpenAI) instead of the free/default
+   *  reviewer. The consensus blocker always uses the free/default reviewer pair regardless (the configured
+   *  self-host provider, or the legacy Workers-AI pair when none is configured), so BYOK never changes who
    *  can be blocked. Default false. */
   aiReviewByok: boolean;
   /** Config-as-code BYOK provider for the advisory write-up. `null` = use the configured key's own
-   *  provider. When set, it must match the stored key's provider or BYOK is skipped (Workers-AI fallback).
-   *  The secret key itself is never here — only via the encrypted key store. */
+   *  provider. When set, it must match the stored key's provider or BYOK is skipped (falls back to the
+   *  free/default reviewer). The secret key itself is never here — only via the encrypted key store. */
   aiReviewProvider?: "anthropic" | "openai" | null | undefined;
   /** Config-as-code model override for the BYOK advisory write-up (e.g. "claude-3-5-sonnet-latest").
    *  `null` = use the key record's model, else a conservative per-provider default. */