Cut worker isolate baseline memory: stub just-bash, lazy-load cheerio (#340)

Production OOM triage: every 'Worker exceeded memory limit' burst hits
unrelated cheap routes right after a deploy — the 128MB limit is per
isolate, and the main worker's eagerly-evaluated module graph is what
crowds it, not any single request.

- Alias just-bash to a throwing stub (worker never uses it): removes
  just-bash + turndown + @mixmark-io/domino from the bundle. The chain
  was pulled in eagerly by @cloudflare/think via the SamChatAgent
  re-export in src/server.ts; SAM only uses its own MCP tools.
- Disable Think's workspace bash tool on SamChatAgent so the stub is
  unreachable at runtime.
- Dynamic-import page-analyzer (cheerio) in the site-audit crawl step
  so it evaluates only when an audit runs, not in every isolate.
- Drop the turndown CJS alias workaround (#339): turndown is no longer
  in the graph at all.

Main eager server chunk: 14,434 kB -> 11,712 kB (-19%); cheerio's
503 kB now a lazy chunk.

src/server/features/sam/SamChatAgent.ts

@@ -78,6 +78,12 @@ type SamContext = {
  * the app DB, so every session in a project shares them.
  */
 export class SamChatAgent extends Think {
+  // SAM's toolset is the MCP tools from beforeTurn; it has no use for Think's
+  // workspace bash tool, whose just-bash dependency is stubbed out of the
+  // bundle anyway (see vite.config.ts) to keep ~30 MB of eagerly-evaluated
+  // source out of every isolate's baseline heap.
+  override workspaceBash = false;
+
   // Session row + project, resolved once per DO lifetime (the binding is
   // immutable). Null until a turn/provider needs it — and left null when the
   // registry row is gone, which beforeTurn turns into a polite refusal.

src/server/lib/just-bash-stub.ts

@@ -0,0 +1,33 @@
+/**
+ * Build-time stand-in for `just-bash`, wired up via the vite.config.ts alias.
+ *
+ * `@cloudflare/think` eagerly imports just-bash (~21 MB of source, plus
+ * turndown and the 8.6 MB @mixmark-io/domino DOM implementation) at module
+ * init, and SamChatAgent is re-exported from src/server.ts, so the whole chain
+ * lands in the main worker's startup module graph — raising every isolate's
+ * baseline heap toward the 128 MB limit (production OOM bursts on unrelated
+ * routes after each deploy). SAM only uses its own MCP toolset and disables
+ * Think's workspace bash tool (`workspaceBash = false`), so the real library
+ * is unreachable; this stub keeps it out of the bundle entirely.
+ *
+ * Remove once https://github.com/cloudflare/agents/issues/1673 lands and
+ * @cloudflare/think loads just-bash lazily.
+ */
+const STUBBED_MESSAGE =
+  "just-bash is stubbed out of the worker bundle (see vite.config.ts); " +
+  "the workspace bash tool is disabled for this deployment";
+
+// Covers every named import in the dependency graph: `Bash` (Think's
+// workspace bash tool) and `defineCommand` (agents/skills bash scripts).
+// Both are only reached when a model invokes bash, which SAM never exposes.
+// Must stay a class: call sites construct it with `new Bash({...})`.
+// oxlint-disable-next-line typescript-eslint/no-extraneous-class
+export class Bash {
+  constructor() {
+    throw new Error(STUBBED_MESSAGE);
+  }
+}
+
+export function defineCommand(): never {
+  throw new Error(STUBBED_MESSAGE);
+}

src/server/workflows/site-audit-workflow-helpers.ts

@@ -1,4 +1,3 @@
-import { analyzeHtml } from "@/server/lib/audit/page-analyzer";
 import type { StepPageResult } from "@/server/lib/audit/types";
 import { isSameOrigin, normalizeUrl } from "@/server/lib/audit/url-utils";
 
@@ -31,6 +30,11 @@ export async function crawlPage(
     }
 
     const html = await response.text();
+    // Dynamic import keeps cheerio (page-analyzer's HTML parser) out of the
+    // worker's startup module graph: SiteAuditWorkflow is re-exported from
+    // src/server.ts, so a static import would evaluate cheerio in every
+    // isolate's baseline heap, not just when an audit actually crawls.
+    const { analyzeHtml } = await import("@/server/lib/audit/page-analyzer");
     const analysis = analyzeHtml(
       html,
       finalUrl,

vite.config.ts

@@ -1,3 +1,4 @@
+import { fileURLToPath } from "node:url";
 import { tanstackStart } from "@tanstack/react-start/plugin/vite";
 import { defineConfig, loadEnv } from "vite";
 import tsConfigPaths from "vite-tsconfig-paths";
@@ -23,17 +24,19 @@ export default defineConfig(({ mode }) => {
   return {
     resolve: {
       alias: {
-        // TODO: Remove this workaround once fixed upstream — either turndown
+        // TODO: Remove this workaround once @cloudflare/think stops eagerly
-        // drops the bare require from its ESM build, or @cloudflare/think
+        // importing just-bash at module init
-        // stops eagerly importing just-bash/turndown at module init
         // (https://github.com/cloudflare/agents/issues/1673).
         //
-        // turndown's ESM build (pulled in via just-bash's html-to-markdown
+        // just-bash (plus its turndown → @mixmark-io/domino chain, ~30 MB of
-        // command) contains a bare CommonJS `require("@mixmark-io/domino")`
+        // source) is only used by Think's workspace bash tool, which SAM
-        // that the Cloudflare Workers runtime rejects at deploy time (error
+        // disables — but the eager import drags it into the main worker's
-        // 10021). Its CJS build goes through Vite's CommonJS transform, which
+        // startup module graph, inflating every isolate's baseline heap
-        // rewrites that require into a bundled import.
+        // toward the 128 MB limit (production OOM bursts on unrelated
-        turndown: "turndown/lib/turndown.cjs.js",
+        // routes). Alias it to a throwing stub so it never ships.
+        "just-bash": fileURLToPath(
+          new URL("./src/server/lib/just-bash-stub.ts", import.meta.url),
+        ),
       },
     },
     envPrefix: [