fix: use current selected model

The CI port job's tsc --noEmit caught SessionEventAccumulator literals not
initializing the new sawMessage field (tests don't typecheck, so bun test
was green). Add a committed .githooks/pre-commit that mirrors the CI port
job — regenerate the port into a temp dir and tsc --noEmit it — so this
class of error fails at commit time, not in CI.

.gitea/workflows/publish.yml

@@ -0,0 +1,51 @@
+name: publish
+
+# Publish this package and its omp port to the npm registry on version tags.
+#
+# Prerequisites on npmjs.com / git.freno.me:
+#   - an npm automation token (publish scope) stored as the repo secret
+#     NPM_TOKEN (consumed via actions/setup-node registry auth)
+#   - the scoped package names claimed on npm (@mikefreno/<name> and
+#     @mikefreno/omp-<name>)
+#
+# Both packages ship TypeScript sources — nothing is built — so lifecycle
+# scripts (prepublishOnly tsc) are skipped: the port workflow already
+# typechecks the generated port against the pinned @oh-my-pi SDK, and the pi
+# package typechecks against the global pi SDK that CI does not install.
+#
+# Manual publish: Actions tab → Run workflow (workflow_dispatch).
+
+on:
+  push:
+    tags: ['v*']
+  workflow_dispatch:
+
+jobs:
+  publish:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Install bun
+        uses: oven-sh/setup-bun@v2
+
+      - name: Setup node (npm registry auth)
+        uses: actions/setup-node@v4
+        with:
+          node-version: '22'
+          registry-url: 'https://registry.npmjs.org'
+
+      - name: Publish pi package
+        env:
+          NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
+        run: npm publish --access public --ignore-scripts
+
+      - name: Port to omp
+        run: bun port-to-omp.mjs --out "$RUNNER_TEMP/omp-port"
+
+      - name: Publish omp package
+        working-directory: ${{ runner.temp }}/omp-port
+        env:
+          NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
+        run: npm publish --access public --ignore-scripts

.githooks/pre-commit

@@ -0,0 +1,37 @@
+#!/bin/sh
+# pre-commit — typecheck the generated omp port before committing.
+#
+# Mirrors the CI port job (.gitea/workflows/port-to-omp.yml). The source
+# repo's own tsconfig extends the host harness tsconfig and is not
+# self-contained, so the reliable typecheck target is the regenerated port:
+# it ships a self-contained tsconfig and the pinned @oh-my-pi SDK as a real
+# devDependency. Regenerating into a temp dir and running `tsc --noEmit`
+# there catches exactly what CI will fail on (e.g. an interface field added
+# without updating its object literals).
+#
+# Enable (per clone):  git config core.hooksPath .githooks
+set -euo pipefail
+
+ROOT="$(git rev-parse --show-toplevel)"
+cd "$ROOT"
+
+# Fast path: no TypeScript-adjacent change staged -> nothing to typecheck.
+if git diff --cached --quiet -- src/ agents/ package.json tsconfig.json port-to-omp.mjs; then
+  exit 0
+fi
+
+TMP="$(mktemp -d)"
+LOG="$(mktemp)"
+trap 'rm -rf "$TMP" "$LOG"' EXIT
+
+if ! bun port-to-omp.mjs --out "$TMP" >"$LOG" 2>&1; then
+  echo "pre-commit: port regeneration failed (CI would fail too) — output:" >&2
+  tail -20 "$LOG" >&2
+  exit 1
+fi
+
+if ! (cd "$TMP" && bun run typecheck) >"$LOG" 2>&1; then
+  echo "pre-commit: port typecheck failed (this is what CI runs) — output:" >&2
+  tail -30 "$LOG" >&2
+  exit 1
+fi

README.md

@@ -202,6 +202,21 @@ pygienium/
 └─ agents/                 ← scanner.md  fixer.md  deep-modules.md  defensive-guards.md
 ```
 
+## Development
+
+The omp port at `~/.omp/agent/extensions/pygienium` is regenerated **only by CI
+on push** (`.gitea/workflows/port-to-omp.yml`) — never by hand.
+
+The source repo's own `tsconfig.json` extends the host harness tsconfig, so the
+reliable typecheck target is the regenerated port (self-contained tsconfig +
+pinned `@oh-my-pi` SDK devDependency). A committed pre-commit hook runs exactly
+what CI runs — regenerate the port into a temp dir and `tsc --noEmit` it — and
+fails the commit on any error:
+
+```sh
+git config core.hooksPath .githooks
+```
+
 ## License
 
 MIT

agents/deep-modules.md

@@ -71,8 +71,9 @@ noise the user cannot act on.
 
 ## Skip (directory names — never descend into)
 
-`.cache`, `.git`, `.hg`, `.idea`, `.next`, `.nuxt`, `.pygienium`, `.ralpi`,
-`.svelte-kit`, `.svn`, `.turbo`, `.vscode`, `__pycache__`, `build`, `coverage`,
+`.cache`, `.git`, `.hg`, `.idea`, `.netlify`, `.next`, `.nuxt`, `.output`,
+`.pygienium`, `.ralpi`, `.svelte-kit`, `.svn`, `.turbo`, `.vercel`, `.vscode`,
+`__pycache__`, `build`, `coverage`,
 `dist`, `node_modules`, `out`, `vendor`, `venv` (and `.venv`)
 
 ## Skip (file patterns)

agents/defensive-guards.md

@@ -87,8 +87,9 @@ noise the user cannot act on.
 
 ## Skip (directory names — never descend into)
 
-`.cache`, `.git`, `.hg`, `.idea`, `.next`, `.nuxt`, `.pygienium`, `.ralpi`,
-`.svelte-kit`, `.svn`, `.turbo`, `.vscode`, `__pycache__`, `build`, `coverage`,
+`.cache`, `.git`, `.hg`, `.idea`, `.netlify`, `.next`, `.nuxt`, `.output`,
+`.pygienium`, `.ralpi`, `.svelte-kit`, `.svn`, `.turbo`, `.vercel`, `.vscode`,
+`__pycache__`, `build`, `coverage`,
 `dist`, `node_modules`, `out`, `vendor`, `venv` (and `.venv`)
 
 ## Skip (file patterns)

agents/scanner.md

@@ -36,8 +36,9 @@ noise the user cannot act on.
 
 ## Skip (directory names — never descend into)
 
-`.cache`, `.git`, `.hg`, `.idea`, `.next`, `.nuxt`, `.pygienium`, `.ralpi`,
-`.svelte-kit`, `.svn`, `.turbo`, `.vscode`, `__pycache__`, `build`, `coverage`,
+`.cache`, `.git`, `.hg`, `.idea`, `.netlify`, `.next`, `.nuxt`, `.output`,
+`.pygienium`, `.ralpi`, `.svelte-kit`, `.svn`, `.turbo`, `.vercel`, `.vscode`,
+`__pycache__`, `build`, `coverage`,
 `dist`, `node_modules`, `out`, `vendor`, `venv` (and `.venv`)
 
 ## Skip (file patterns)

agents/todos.md

@@ -100,8 +100,9 @@ noise the user cannot act on.
 
 ## Skip (directory names — never descend into)
 
-`.cache`, `.git`, `.hg`, `.idea`, `.next`, `.nuxt`, `.pygienium`, `.ralpi`,
-`.svelte-kit`, `.svn`, `.turbo`, `.vscode`, `__pycache__`, `build`, `coverage`,
+`.cache`, `.git`, `.hg`, `.idea`, `.netlify`, `.next`, `.nuxt`, `.output`,
+`.pygienium`, `.ralpi`, `.svelte-kit`, `.svn`, `.turbo`, `.vercel`, `.vscode`,
+`__pycache__`, `build`, `coverage`,
 `dist`, `node_modules`, `out`, `vendor`, `venv` (and `.venv`)
 
 ## Skip (file patterns)

package.json

@@ -1,6 +1,6 @@
 {
-  "name": "pygienium",
-  "version": "0.1.0",
+  "name": "@mikefreno/pygienium",
+  "version": "0.1.2",
   "description": "Code hygiene extension for pi — isolated sub-agent checks that scan a target, apply fixes, and emit a findings+changes report.",
   "keywords": [
     "pi-package",
@@ -51,5 +51,11 @@
   "scripts": {
     "typecheck": "tsc --noEmit",
     "test": "bun test"
-  }
+  },
+  "files": [
+    "src/",
+    "agents/",
+    "README.md",
+    "LICENSE"
+  ]
 }

port-to-omp.mjs

@@ -7,12 +7,12 @@
  * --out) is a generated artifact. Every op asserts its target and fails
  * loudly on base drift — never silently producing a stale port.
  *
- * Usage:
- *   bun port-to-omp.mjs                # write ~/.omp/agent/extensions/pygenium
- *   bun port-to-omp.mjs --out <dir>    # write elsewhere (CI: the omp repo clone)
+ * The port is updated ONLY by CI on push (see below) — never by hand; local
+ * edits to ~/.omp/agent/extensions/pygienium are overwritten by the next run.
  *
  * CI: .gitea/workflows/port-to-omp.yml clones the omp-pygienium repo and
- * runs this script into it, then commits + pushes when the port changed.
+ * runs this script into it (`bun port-to-omp.mjs --out <dir>`), then
+ * commits + pushes when the port changed.
  */
 
 import {
@@ -35,6 +35,7 @@ const OMP = join(HOME, ".omp", "agent", "extensions");
 
 const SKIP = new Set([
   "port-to-omp.mjs",
+  "release-tag.sh",
   ".gitea",
   ".github",
   "node_modules",
@@ -217,6 +218,11 @@ omp install @mikefreno/omp-pygienium
 This is the omp port of [Mike/pygienium](https://git.freno.me/Mike/pygienium), regenerated automatically from the source repo. See the source repo for full documentation.
 `;
 
+
+// publish workflow emitted into the port repo so the omp package can be
+// released independently (tag push or manual dispatch on the omp repo).
+const PORT_PUBLISH_WORKFLOW = "name: publish\n\n# Publish this omp port package to the npm registry on version tags.\n#\n# Prerequisites:\n#   - npm automation token (publish scope) stored as the repo secret NPM_TOKEN\n#   - package name claimed on npm (@mikefreno/omp-<name>)\n#\n# Manual publish: Actions tab → Run workflow (workflow_dispatch).\n\non:\n  push:\n    tags: ['v*']\n  workflow_dispatch:\n\njobs:\n  publish:\n    runs-on: ubuntu-latest\n    steps:\n      - name: Checkout\n        uses: actions/checkout@v4\n\n      - name: Setup node (npm registry auth)\n        uses: actions/setup-node@v4\n        with:\n          node-version: '22'\n          registry-url: 'https://registry.npmjs.org'\n\n      - name: Publish\n        env:\n          NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}\n        run: npm publish --access public --ignore-scripts\n";
+
 const FILE_RULES = {
     "src/index.ts": [
       { from: " * Read the pygienium chat style from pi's settings.json.", to: " * Read the pygienium chat style from omp's settings.json." },
@@ -250,9 +256,9 @@ const FILE_RULES = {
       { from: '\t\tagent.allowedTools ?? ["read", "bash", "grep", "find"];', to: '\t\tagent.allowedTools ?? ["read", "bash", "grep", "glob"];', label: "default tools" },
       {
         from:
-          "\tconst { session } = await createAgentSession({\n\t\tcwd: opts.cwd,\n\t\ttools,\n\t\tsessionManager: SessionManager.inMemory(opts.cwd),\n\t\tresourceLoader: loader,\n\t});",
+          "\tconst { session } = await createAgentSession({\n\t\tcwd: opts.cwd,\n\t\t// Pin the sub-agent to the model the user has selected in the invoking\n\t\t// session rather than the settings default. Omitted when the caller\n\t\t// has no live session model (print/RPC modes), which keeps the\n\t\t// settings-default fallback.\n\t\t...(opts.model ? { model: opts.model } : {}),\n\t\ttools,\n\t\tsessionManager: SessionManager.inMemory(opts.cwd),\n\t\tresourceLoader: loader,\n\t});",
         to:
-          "\tconst { session } = await createAgentSession({\n\t\tcwd: opts.cwd,\n\t\ttoolNames: tools,\n\t\t// `tools` is an allowlist, not a request list.\n\t\trestrictToolNames: true,\n\t\tsessionManager: SessionManager.inMemory(opts.cwd),\n\t\t// Replace the fully rendered default prompt with the agent body.\n\t\tsystemPrompt: agent.systemPrompt,\n\t\t// Keep the sub-agent isolated: no nested extensions/skills/prompts/etc.\n\t\tdisableExtensionDiscovery: true,\n\t\tskills: [],\n\t\tpromptTemplates: [],\n\t\trules: [],\n\t\tcontextFiles: [],\n\t\tenableMCP: false,\n\t\tenableLsp: false,\n\t\t// Private registry: the host session owns the process-global \"Main\"\n\t\t// identity, so a per-run registry keeps these in-process workers\n\t\t// disjoint from the main agent.\n\t\tagentRegistry: new AgentRegistry(),\n\t});",
+          "\tconst { session } = await createAgentSession({\n\t\tcwd: opts.cwd,\n\t\t// Pin the sub-agent to the model the user has selected in the invoking\n\t\t// session rather than the settings default.\n\t\t...(opts.model ? { model: opts.model } : {}),\n\t\ttoolNames: tools,\n\t\t// `tools` is an allowlist, not a request list.\n\t\trestrictToolNames: true,\n\t\tsessionManager: SessionManager.inMemory(opts.cwd),\n\t\t// Replace the fully rendered default prompt with the agent body.\n\t\tsystemPrompt: agent.systemPrompt,\n\t\t// Keep the sub-agent isolated: no nested extensions/skills/prompts/etc.\n\t\tdisableExtensionDiscovery: true,\n\t\tskills: [],\n\t\tpromptTemplates: [],\n\t\trules: [],\n\t\tcontextFiles: [],\n\t\tenableMCP: false,\n\t\tenableLsp: false,\n\t\t// Private registry: the host session owns the process-global \"Main\"\n\t\t// identity, so a per-run registry keeps these in-process workers\n\t\t// disjoint from the main agent.\n\t\tagentRegistry: new AgentRegistry(),\n\t});",
         label: "createAgentSession",
       },
     ],
@@ -321,6 +327,10 @@ function portExtension() {
     }
   }
   writeFileSync(join(dstDir, "tsconfig.json"), PYGENIUM_TSCONFIG);
+  const wfDir = join(dstDir, ".gitea", "workflows");
+  mkdirSync(wfDir, { recursive: true });
+  writeFileSync(join(wfDir, "publish.yml"), PORT_PUBLISH_WORKFLOW);
+
   console.log("== bun install (regenerates bun.lock + node_modules)");
   execSync("bun install", { cwd: dstDir, stdio: "inherit" });
   console.log("== done");

scripts/release-tag.sh

@@ -0,0 +1,196 @@
+#!/bin/bash
+
+# release-tag.sh — version bump, commit, tag, and push for this pi extension.
+#
+# Mirrors the release flow from PodTui's scripts/release-tag.sh, adapted for
+# the pi extension repos: the version lives in package.json, and the omp port
+# derives its version from it during regeneration.
+#
+# Usage:
+#   scripts/release-tag.sh            interactive release
+#   scripts/release-tag.sh --dry-run  plan the bump/tag/pushes without doing
+#
+# After this script runs:
+#   - pushing master triggers the port-to-omp workflow, which regenerates and
+#     pushes the omp port (picking up the new version)
+#   - pushing the v* tag triggers the publish workflow, which npm-publishes
+#     this package and its omp port together
+
+set -e
+
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+BLUE='\033[0;34m'
+CYAN='\033[0;36m'
+NC='\033[0m' # No Color
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PROJECT_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
+cd "$PROJECT_ROOT"
+
+DRY_RUN=0
+for arg in "$@"; do
+	case "$arg" in
+	--dry-run) DRY_RUN=1 ;;
+	-h | --help)
+		echo "usage: $0 [--dry-run]"
+		exit 0
+		;;
+	esac
+done
+
+if [ ! -d .git ] && [ ! -f .git ]; then
+	echo -e "${RED}✗ not a git repository: $PROJECT_ROOT${NC}"
+	exit 1
+fi
+
+if ! git diff-index --quiet HEAD --; then
+	echo -e "${RED}✗ working tree is dirty — commit or stash before releasing${NC}"
+	git status --short
+	exit 1
+fi
+
+NAME=$(python3 -c "import json;print(json.load(open('package.json'))['name'])")
+OMP_NAME=$(python3 -c "import json,re;n=json.load(open('package.json'))['name'];print(re.sub(r'^@mikefreno/(.+)$', r'@mikefreno/omp-\1', n))")
+CURRENT_VERSION=$(python3 -c "import json;print(json.load(open('package.json'))['version'])")
+
+echo -e "${CYAN}Package:${NC} ${GREEN}${NAME}${NC}"
+echo -e "${CYAN}Current version:${NC} ${GREEN}v${CURRENT_VERSION}${NC}"
+echo ""
+
+IFS='.' read -r MAJOR MINOR PATCH <<<"$CURRENT_VERSION"
+MAJOR=$(echo "$MAJOR" | sed 's/[^0-9].*//')
+MINOR=$(echo "$MINOR" | sed 's/[^0-9].*//')
+PATCH=$(echo "$PATCH" | sed 's/[^0-9].*//')
+
+echo -e "${CYAN}Select version bump type:${NC}"
+echo "  1) Major (breaking changes)     ${MAJOR}.${MINOR}.${PATCH} → $((MAJOR + 1)).0.0"
+echo "  2) Minor (new features)         ${MAJOR}.${MINOR}.${PATCH} → ${MAJOR}.$((MINOR + 1)).0"
+echo "  3) Patch (bug fixes)            ${MAJOR}.${MINOR}.${PATCH} → ${MAJOR}.${MINOR}.$((PATCH + 1))"
+echo "  4) Custom version"
+echo "  5) Cancel"
+echo ""
+read -p "Enter choice (1-5): " -n 1 -r CHOICE
+echo ""
+echo ""
+
+case $CHOICE in
+1) NEW_VERSION="$((MAJOR + 1)).0.0" ;;
+2) NEW_VERSION="${MAJOR}.$((MINOR + 1)).0" ;;
+3) NEW_VERSION="${MAJOR}.${MINOR}.$((PATCH + 1))" ;;
+4)
+	read -p "Enter new version (e.g. 0.6.1): " -r NEW_VERSION
+	echo ""
+	;;
+5)
+	echo -e "${YELLOW}Cancelled.${NC}"
+	exit 0
+	;;
+*)
+	echo -e "${RED}✗ invalid choice${NC}"
+	exit 1
+	;;
+esac
+
+if ! echo "$NEW_VERSION" | grep -Eq '^[0-9]+\.[0-9]+\.[0-9]+$'; then
+	echo -e "${RED}✗ version must be MAJOR.MINOR.PATCH (got: ${NEW_VERSION})${NC}"
+	exit 1
+fi
+
+if git rev-parse -q --verify "refs/tags/v${NEW_VERSION}" >/dev/null; then
+	echo -e "${RED}✗ tag v${NEW_VERSION} already exists${NC}"
+	exit 1
+fi
+
+echo -e "${CYAN}New version:${NC} ${GREEN}v${NEW_VERSION}${NC}"
+echo ""
+echo -e "${YELLOW}This will:${NC}"
+echo "  1. Set package.json → \"version\": \"${NEW_VERSION}\""
+echo "  2. Commit the bump"
+echo "  3. Create annotated tag v${NEW_VERSION}"
+echo "  4. Push master and the tag to every remote"
+REMOTES=$(git remote)
+for r in $REMOTES; do
+	echo "       → $r"
+done
+echo ""
+echo -e "${YELLOW}Note: master push triggers the port-to-omp workflow; the v* tag${NC}"
+echo -e "push triggers the npm publish workflow (this package + omp port).${NC}"
+echo ""
+read -p "Proceed? (y/n) " -n 1 -r
+echo ""
+if [[ ! $REPLY =~ ^[Yy]$ ]]; then
+	echo -e "${YELLOW}Aborted.${NC}"
+	exit 0
+fi
+
+if [ "$DRY_RUN" -eq 1 ]; then
+	echo -e "${BLUE}── dry run: no changes made ──${NC}"
+	exit 0
+fi
+
+# ── Apply the bump ───────────────────────────────────────────────────────────
+echo ""
+echo -e "${CYAN}[1/4]${NC} Updating package.json..."
+sed -i.bak "s/\"version\": \"[^\"]*\"/\"version\": \"${NEW_VERSION}\"/" package.json
+rm -f package.json.bak
+echo -e "${GREEN}✓ package.json updated${NC}"
+
+if git diff --quiet -- package.json; then
+	echo -e "${RED}✗ package.json did not change${NC}"
+	exit 1
+fi
+
+echo -e "${CYAN}[2/4]${NC} Committing..."
+git add package.json
+git commit -m "chore: bump version to v${NEW_VERSION}"
+echo -e "${GREEN}✓ committed${NC}"
+
+echo -e "${CYAN}[3/4]${NC} Tagging..."
+git tag -a "v${NEW_VERSION}" -m "${NAME} v${NEW_VERSION}"
+echo -e "${GREEN}✓ tagged v${NEW_VERSION}${NC}"
+echo ""
+
+echo -e "${CYAN}[4/4]${NC} Pushing..."
+FAILED=""
+for r in $REMOTES; do
+	echo -e "  → ${BLUE}${r}${NC} master..."
+	if git push "$r" master >/dev/null 2>&1; then
+		echo -e "    ${GREEN}✓ master pushed${NC}"
+	else
+		echo -e "    ${RED}✗ master push failed${NC}"
+		FAILED="$FAILED $r"
+		continue
+	fi
+	echo -e "  → ${BLUE}${r}${NC} v${NEW_VERSION}..."
+	if git push "$r" "v${NEW_VERSION}" >/dev/null 2>&1; then
+		echo -e "    ${GREEN}✓ tag pushed${NC}"
+	else
+		echo -e "    ${RED}✗ tag push failed${NC}"
+		FAILED="$FAILED $r"
+	fi
+done
+
+echo ""
+if [ -n "$FAILED" ]; then
+	echo -e "${RED}✗ push failed for remote(s):${FAILED}${NC}"
+	echo -e "${YELLOW}Re-run: git push <remote> master && git push <remote> v${NEW_VERSION}${NC}"
+	exit 1
+fi
+
+echo -e "${GREEN}═══════════════════════════════════════${NC}"
+echo -e "${GREEN}✓ ${NAME} v${NEW_VERSION} released${NC}"
+echo -e "${GREEN}═══════════════════════════════════════${NC}"
+echo ""
+echo -e "${CYAN}Version:${NC} ${CURRENT_VERSION} → ${GREEN}${NEW_VERSION}${NC}"
+echo -e "${CYAN}Tag:${NC} v${NEW_VERSION}"
+echo ""
+echo -e "${BLUE}Next steps (automatic, nothing to do):${NC}"
+echo "  1. Gitea Actions port-to-omp regenerates and pushes the omp port"
+echo "     (which now carries version ${NEW_VERSION})"
+echo "  2. Gitea Actions publish npm-publishes ${NAME} and"
+echo "     ${OMP_NAME} to the npm registry"
+echo ""
+echo -e "${CYAN}Local port (optional):${NC} run 'bun port-to-omp.mjs' in this repo, or"
+echo "pull in ~/.omp once CI has pushed."

src/agent-runner.ts

@@ -20,6 +20,7 @@ import { dirname, isAbsolute, join } from "node:path";
 import type {
 	AgentSession,
 	AgentSessionEvent,
+	ExtensionCommandContext,
 } from "@earendil-works/pi-coding-agent";
 import { loadAgents, extensionRoot, type AgentDef } from "./agents.js";
 
@@ -52,6 +53,11 @@ export interface AgentTaskOptions {
 	task: string;
 	/** Optional tool allowlist override (else uses the agent's `allowedTools`). */
 	allowedTools?: string[];
+	/**
+	 * The currently selected model from the invoking session. When omitted,
+	 * `createAgentSession` falls back to the settings default model.
+	 */
+	model?: ExtensionCommandContext["model"];
 	/** Optional explicit agent definition (skips `loadAgents`). */
 	agent?: AgentDef;
 	/**
@@ -208,6 +214,11 @@ export async function defaultAgentRunner(
 
 	const { session } = await createAgentSession({
 		cwd: opts.cwd,
+		// Pin the sub-agent to the model the user has selected in the invoking
+		// session rather than the settings default. Omitted when the caller
+		// has no live session model (print/RPC modes), which keeps the
+		// settings-default fallback.
+		...(opts.model ? { model: opts.model } : {}),
 		tools,
 		sessionManager: SessionManager.inMemory(opts.cwd),
 		resourceLoader: loader,
@@ -243,7 +254,7 @@ async function runSessionToCompletion(
 	session: AgentSession,
 	opts: AgentTaskOptions,
 ): Promise<AgentRunResult> {
-	const acc: SessionEventAccumulator = { text: "" };
+	const acc: SessionEventAccumulator = { text: "", sawMessage: false };
 	try {
 		const unsubscribe = session.subscribe((event: AgentSessionEvent) => {
 			applySessionEvent(acc, event, opts.onEvent);
@@ -256,16 +267,15 @@ async function runSessionToCompletion(
 		unsubscribe();
 		// Surface session errors that didn't throw but left no useful output.
 		// A session ending with stopReason "error" and no text means the model
-		// call failed silently — treat that as a failed run, not ok:true.
-		if (acc.errorMessage) {
-			return { ok: false, text: acc.text, error: acc.errorMessage };
-		}
-		if (!acc.text.trim() && acc.stopReason === "error") {
-			return {
-				ok: false,
-				text: acc.text,
-				error: "sub-agent session ended in error with no output.",
-			};
+		// call failed silently — treat that as a failed run, not ok:true. A
+		// session that settled with NO text and NO observed events means the
+		// model never produced anything at all (dead provider stream, failed
+		// start) — reporting that as a successful analysis would let an empty
+		// scan masquerade as a clean one, and the check's verify hook would
+		// fail only later with a confusing "artifact missing" error.
+		const sessionError = emptySessionError(acc);
+		if (sessionError) {
+			return { ok: false, text: acc.text, error: sessionError };
 		}
 		return { ok: true, text: acc.text };
 	} catch (err) {
@@ -287,12 +297,34 @@ async function runSessionToCompletion(
 export interface SessionEventAccumulator {
 	/** Joined assistant text seen so far (text_delta stream). */
 	text: string;
+	/** Whether any assistant-message or tool event was observed at all. */
+	sawMessage: boolean;
 	/** stopReason of the final assistant message, when reported. */
 	stopReason?: string;
 	/** errorMessage of the final assistant message, when reported. */
 	errorMessage?: string;
 }
 
+/**
+ * Classify a settled session's capture: `undefined` when the result is a
+ * legitimate (possibly empty-text) outcome, else the error that should fail
+ * the run. Kept pure so the decision is unit-testable without a session.
+ */
+export function emptySessionError(
+	acc: SessionEventAccumulator,
+): string | undefined {
+	if (acc.errorMessage) return acc.errorMessage;
+	if (!acc.text.trim()) {
+		if (acc.stopReason === "error") {
+			return "sub-agent session ended in error with no output.";
+		}
+		if (!acc.sawMessage) {
+			return "sub-agent session settled with no output — no assistant message or tool activity was observed. Check model/provider connectivity, then resume.";
+		}
+	}
+	return undefined;
+}
+
 /**
  * Interpret one session event into the running accumulator and forward the
  * stream-driving events to the chat.
@@ -313,6 +345,16 @@ export function applySessionEvent(
 ): void {
 	if (!event) return;
 	try {
+		// Any message or tool event means the session actually ran — a settled
+		// capture with none of these is a dead session, not an empty scan.
+		if (
+			event.type === "message_update" ||
+			event.type === "message_end" ||
+			event.type === "tool_execution_start" ||
+			event.type === "tool_execution_end"
+		) {
+			acc.sawMessage = true;
+		}
 		if (
 			event.type === "message_update" &&
 			event.assistantMessageEvent?.type === "text_delta"

src/checks/deep-modules.ts

@@ -21,14 +21,13 @@
  * @module pygienium/checks/deep-modules
  */
 
-import { readdirSync } from "node:fs";
 import { join } from "node:path";
 import {
 	registerCheck,
 	type CheckDefinition,
 	type CheckScope,
 } from "./registry.js";
-import { isScopeSource, scopeRulesMarkdown } from "./scope.js";
+import { hasScopeSources, scopeRulesMarkdown } from "./scope.js";
 
 /** Output directory for this check's persistent reports. */
 export function deepModulesOutputDir(cwd: string): string {
@@ -50,23 +49,13 @@ export function changesPath(cwd: string): string {
  * with zero source files gives the scanner nothing to classify.
  */
 function deepModulesGate(cwd: string): string | undefined {
-	let found = false;
 	try {
-		const entries = readdirSync(cwd);
-		for (const entry of entries) {
-			if (isScopeSource(entry)) {
-				found = true;
-				break;
-			}
-		}
+		if (hasScopeSources(cwd)) return undefined;
 	} catch {
 		// unreadable cwd → let the agent decide; don't block.
 		return undefined;
 	}
-	if (!found) {
-		return "no source files found to inspect";
-	}
-	return undefined;
+	return "no source files found to inspect";
 }
 
 /**

src/checks/defensive-guards.ts

@@ -41,14 +41,13 @@
  * @module pygienium/checks/defensive-guards
  */
 
-import { readdirSync } from "node:fs";
 import { join } from "node:path";
 import {
 	registerCheck,
 	type CheckDefinition,
 	type CheckScope,
 } from "./registry.js";
-import { isScopeSource, scopeRulesMarkdown } from "./scope.js";
+import { hasScopeSources, scopeRulesMarkdown } from "./scope.js";
 
 /** Output directory for this check's persistent reports. */
 export function defensiveGuardsOutputDir(cwd: string): string {
@@ -70,23 +69,13 @@ export function changesPath(cwd: string): string {
  * with zero source files gives the scanner nothing to analyse.
  */
 function defensiveGuardsGate(cwd: string): string | undefined {
-	let found = false;
 	try {
-		const entries = readdirSync(cwd);
-		for (const entry of entries) {
-			if (isScopeSource(entry)) {
-				found = true;
-				break;
-			}
-		}
+		if (hasScopeSources(cwd)) return undefined;
 	} catch {
 		// unreadable cwd → let the agent decide; don't block.
 		return undefined;
 	}
-	if (!found) {
-		return "no source files found to inspect";
-	}
-	return undefined;
+	return "no source files found to inspect";
 }
 
 /**

src/checks/scope.ts

@@ -9,6 +9,10 @@
  * @module pygienium/checks/scope
  */
 
+import { readdirSync, statSync } from "node:fs";
+import type { Dirent } from "node:fs";
+import { join } from "node:path";
+
 /**
  * Implementation-code file extensions pygienium inspects.
  *
@@ -56,6 +60,13 @@ export const SCOPE_EXCLUDE_DIRS: ReadonlySet<string> = new Set([
 	".nuxt",
 	".turbo",
 	".svelte-kit",
+	// Framework/deploy build output: Nitro, Vercel, Netlify artifact dirs.
+	// Generated bundles dominate a naive marker/stub scan (freno-dev alone
+	// flagged 119 of 124 candidates inside `.output/`/`.vercel/` minified
+	// bundles) and must never feed the pre-scan or the agent's file walk.
+	".output",
+	".vercel",
+	".netlify",
 	"__pycache__",
 	".venv",
 	"venv",
@@ -99,6 +110,38 @@ export function isScopeSource(path: string): boolean {
 	return SCOPE_EXTENSIONS.has(lower.slice(dot));
 }
 
+/**
+ * True when the tree rooted at `root` contains at least one in-scope source
+ * file. Walks recursively (honoring {@link SCOPE_EXCLUDE_DIRS}) — a top-level
+ * entry scan alone would skip any repo whose source lives in subdirectories,
+ * e.g. `game/` or `src/`, even though recon's git inventory finds hundreds of
+ * files. A file root is judged by {@link isScopeSource} directly.
+ */
+export function hasScopeSources(root: string): boolean {
+	const st = statSync(root, { throwIfNoEntry: false });
+	if (!st) return false;
+	if (st.isFile()) return isScopeSource(root);
+	const stack = [root];
+	while (stack.length > 0) {
+		const dir = stack.pop() as string;
+		let entries: Dirent[];
+		try {
+			entries = readdirSync(dir, { withFileTypes: true });
+		} catch {
+			continue;
+		}
+		for (const entry of entries) {
+			if (entry.isDirectory()) {
+				if (SCOPE_EXCLUDE_DIRS.has(entry.name)) continue;
+				stack.push(join(dir, entry.name));
+			} else if (entry.isFile() && isScopeSource(entry.name)) {
+				return true;
+			}
+		}
+	}
+	return false;
+}
+
 /**
  * Markdown section injected into every scan task string so the sub-agent knows
  * exactly what to inspect and what to skip — stated once here, not copy-pasted

src/checks/todos.ts

@@ -36,7 +36,6 @@
  * @module pygienium/checks/todos
  */
 
-import { readdirSync } from "node:fs";
 import { readFile, readdir, stat } from "node:fs/promises";
 import { join, relative } from "node:path";
 import { loadRunState } from "../run-state.js";
@@ -46,6 +45,7 @@ import {
 	type CheckScope,
 } from "./registry.js";
 import {
+	hasScopeSources,
 	isScopeSource,
 	SCOPE_EXCLUDE_DIRS,
 	scopeRulesMarkdown,
@@ -124,6 +124,21 @@ const MAX_CANDIDATES = 500;
 const FALLBACK_CAP = 16;
 /** Candidate list embedded in the live prompt is truncated at this many. */
 const PROMPT_CAP = 40;
+/**
+ * Max characters of a candidate's code line embedded in the task prompt.
+ * Generated/bundled single lines can be hundreds of KB (e.g. minified assets
+ * sneaking past scope); embedding them wholesale balloons the task to
+ * megabytes and chokes the sub-agent. `path:line` plus a truncated prefix is
+ * enough to classify — the agent can read the file for full context.
+ */
+const CODE_DISPLAY_CAP = 160;
+
+/** Truncate a candidate's code line for prompt embedding. */
+function displayCode(code: string): string {
+	return code.length > CODE_DISPLAY_CAP
+		? `${code.slice(0, CODE_DISPLAY_CAP)}…`
+		: code;
+}
 
 /** Extract the declared function name from a header line, when present. */
 function headerName(line: string): string | undefined {
@@ -369,7 +384,7 @@ function renderFindings(
 		parts.push(`## ${title}`);
 		list.slice(0, FALLBACK_CAP).forEach((c, i) => {
 			parts.push(
-				`### ${i + 1}. ${relative(cwd, c.path)}:${c.line} — ${c.code} | snippet: ${c.snippet} | context: ${c.context || relative(cwd, c.path)}`,
+				`### ${i + 1}. ${relative(cwd, c.path)}:${c.line} — ${displayCode(c.code)} | snippet: ${c.snippet} | context: ${c.context || relative(cwd, c.path)}`,
 			);
 		});
 		if (list.length > FALLBACK_CAP) {
@@ -390,23 +405,13 @@ function renderFindings(
  * with zero source files gives the scanner nothing to analyse.
  */
 function todosGate(cwd: string): string | undefined {
-	let found = false;
 	try {
-		const entries = readdirSync(cwd);
-		for (const entry of entries) {
-			if (isScopeSource(entry)) {
-				found = true;
-				break;
-			}
-		}
+		if (hasScopeSources(cwd)) return undefined;
 	} catch {
 		// unreadable cwd → let the agent decide; don't block.
 		return undefined;
 	}
-	if (!found) {
-		return "no source files found to inspect";
-	}
-	return undefined;
+	return "no source files found to inspect";
 }
 
 /**
@@ -434,7 +439,7 @@ export async function buildTodosScanTask(
 
 	const candidateList = candidates
 		.slice(0, PROMPT_CAP)
-		.map((c) => ` - ${relative(cwd, c.path)}:${c.line} [${c.kind}] ${c.code}`);
+		.map((c) => ` - ${relative(cwd, c.path)}:${c.line} [${c.kind}] ${displayCode(c.code)}`);
 	if (candidates.length > PROMPT_CAP) {
 		candidateList.push(
 			` - … and ${candidates.length - PROMPT_CAP} more (truncated for brevity)`,

src/commands.ts

@@ -48,7 +48,7 @@ import type { SendChatMessage } from "./phases.js";
 /** Narrow context slice handlers need (a subset of `ExtensionCommandContext`). */
 export type PygieniumCtx = Pick<
 	ExtensionCommandContext,
-	"cwd" | "mode" | "hasUI" | "ui"
+	"cwd" | "mode" | "hasUI" | "ui" | "model"
 > & {
 	/** Optional callback to post messages to the chat window. */
 	sendChatMessage?: SendChatMessage;
@@ -176,6 +176,7 @@ export async function handleCheckCommand(
 		existingState: existing,
 		ui: ctx.ui,
 		hasUI: ctx.hasUI,
+		model: ctx.model,
 		sendChatMessage: ctx.sendChatMessage,
 		onAgentEvent: ctx.onAgentEvent,
 		sendPhaseLine: ctx.sendPhaseLine,
@@ -215,6 +216,7 @@ export async function handleAllCommand(
 		only: parsed.only,
 		ui: ctx.ui,
 		hasUI: ctx.hasUI,
+		model: ctx.model,
 		sendChatMessage: ctx.sendChatMessage,
 		onAgentEvent: ctx.onAgentEvent,
 		sendPhaseLine: ctx.sendPhaseLine,
@@ -303,6 +305,7 @@ export async function handleResumeCommand(
 			scope: { cwd, target: cwd, fix: entry.fix, rest: [] },
 			ui: ctx.ui,
 			hasUI: ctx.hasUI,
+			model: ctx.model,
 			existingState: state,
 			sendChatMessage: ctx.sendChatMessage,
 			onAgentEvent: ctx.onAgentEvent,

src/index.ts

@@ -410,6 +410,7 @@ export default async function pygieniumExtension(
 					mode: ctx.mode,
 					hasUI: ctx.hasUI,
 					ui: ctx.ui,
+					model: ctx.model,
 					sendChatMessage,
 					onAgentEvent,
 					sendPhaseLine,

src/modes/all.ts

@@ -22,10 +22,13 @@
 
 import { mkdir, writeFile } from "node:fs/promises";
 import { dirname, join, resolve } from "node:path";
-import type { ExtensionUIContext } from "@earendil-works/pi-coding-agent";
+import type {
+	AgentSessionEvent,
+	ExtensionCommandContext,
+	ExtensionUIContext,
+} from "@earendil-works/pi-coding-agent";
 import { getAllChecks, type CheckDefinition } from "../checks/registry.js";
 import { runCheck } from "./check-runner.js";
-import type { AgentSessionEvent } from "@earendil-works/pi-coding-agent";
 import { createPhaseStrip, type SendChatMessage } from "../phases.js";
 import { createPipelineFooter, type ItemStatus } from "../footer.js";
 import { runRecon } from "../recon.js";
@@ -74,6 +77,12 @@ export interface AllRunOptions {
 	ui?: ExtensionUIContext;
 	/** Whether dialog-capable UI is available. */
 	hasUI?: boolean;
+	/**
+	 * The currently selected model from the invoking session; forwarded to
+	 * each check's sub-agents so scans run on the model the user picked, not
+	 * the settings default.
+	 */
+	model?: ExtensionCommandContext["model"];
 	/** Optional callback to post completion messages into the chat. */
 	sendChatMessage?: SendChatMessage;
 	/** Optional callback forwarding raw sub-agent events to the chat stream. */
@@ -369,6 +378,7 @@ export async function runAllChecks(
 			scope: { cwd, target, fix, rest: [] },
 			ui: opts.ui,
 			hasUI,
+			model: opts.model,
 			existingState: state,
 			sendChatMessage: opts.sendChatMessage,
 			onAgentEvent: opts.onAgentEvent,

src/modes/check-runner.ts

@@ -17,7 +17,10 @@
 
 import { rm } from "node:fs/promises";
 import { resolve, join } from "node:path";
-import type { ExtensionUIContext } from "@earendil-works/pi-coding-agent";
+import type {
+	ExtensionCommandContext,
+	ExtensionUIContext,
+} from "@earendil-works/pi-coding-agent";
 import type { CheckDefinition, CheckScope } from "../checks/registry.js";
 import { runAgentTask } from "../agent-runner.js";
 import { runRecon } from "../recon.js";
@@ -84,6 +87,12 @@ export interface RunCheckOptions {
 	ui?: ExtensionUIContext;
 	/** Whether dialog-capable UI is available. */
 	hasUI?: boolean;
+	/**
+	 * The currently selected model from the invoking session; forwarded to
+	 * each sub-agent so scans run on the model the user picked, not the
+	 * settings default.
+	 */
+	model?: ExtensionCommandContext["model"];
 	/** Pre-existing run state to update (for `/pygienium-all` and resume). */
 	existingState?: RunState;
 	/** Optional callback to post completion messages into the chat. */
@@ -278,6 +287,7 @@ async function runCheckImplInner(
 			cwd: scope.target,
 			agentName: check.agentName,
 			task: scanTask,
+			model: opts.model,
 			onEvent: forward(PHASE_ANALYSIS),
 		});
 		findings = scanResult.text;
@@ -310,6 +320,7 @@ async function runCheckImplInner(
 				cwd: scope.target,
 				agentName: check.fixAgentName ?? "fixer",
 				task: fixTask,
+				model: opts.model,
 				onEvent: forward(PHASE_FIX),
 			});
 			changes = fixResult.text;

tests/agent-runner.test.ts

@@ -10,12 +10,13 @@
 import { describe, expect, it } from "bun:test";
 import {
 	applySessionEvent,
+	emptySessionError,
 	type SessionEventAccumulator,
 } from "../src/agent-runner.js";
 import type { AgentSessionEvent } from "@earendil-works/pi-coding-agent";
 
 function fresh(): SessionEventAccumulator {
-	return { text: "" };
+	return { text: "", sawMessage: false };
 }
 
 /** Build a typed-as-unknown event so malformed shapes compile in tests. */
@@ -134,4 +135,54 @@ describe("applySessionEvent", () => {
 		expect(forwarded).toEqual([]);
 		expect(acc.text).toBe("x");
 	});
+
+	it("marks sawMessage when any message or tool event is observed", () => {
+		const fromUpdate = fresh();
+		applySessionEvent(fromUpdate, event({ type: "message_update" }));
+		expect(fromUpdate.sawMessage).toBe(true);
+
+		const fromEnd = fresh();
+		applySessionEvent(fromEnd, event({ type: "message_end", message: null }));
+		expect(fromEnd.sawMessage).toBe(true);
+
+		const fromTool = fresh();
+		applySessionEvent(fromTool, event({ type: "tool_execution_start" }));
+		expect(fromTool.sawMessage).toBe(true);
+	});
+});
+
+describe("emptySessionError", () => {
+	it("fails a session that settled with no text and no observed events", () => {
+		expect(emptySessionError({ text: "", sawMessage: false })).toContain(
+			"no output",
+		);
+	});
+
+	it("fails an empty session whose final message reported stopReason error", () => {
+		expect(
+			emptySessionError({ text: "", sawMessage: true, stopReason: "error" }),
+		).toBe("sub-agent session ended in error with no output.");
+	});
+
+	it("surfaces a recorded errorMessage regardless of text", () => {
+		expect(
+			emptySessionError({
+				text: "partial output",
+				sawMessage: true,
+				errorMessage: "upstream 529",
+			}),
+		).toBe("upstream 529");
+	});
+
+	it("accepts an empty-text session that demonstrably ran (tool activity)", () => {
+		expect(
+			emptySessionError({ text: "", sawMessage: true, stopReason: "end_turn" }),
+		).toBeUndefined();
+	});
+
+	it("accepts any session with text", () => {
+		expect(
+			emptySessionError({ text: "report", sawMessage: false }),
+		).toBeUndefined();
+	});
 });

tests/check-runner.test.ts

@@ -110,6 +110,28 @@ describe("check-runner integration", () => {
 		);
 	});
 
+	it("forwards the selected model to every sub-agent", async () => {
+		const check = smokeCheck();
+		const selectedModel = {
+			provider: "test-provider",
+			id: "test-model",
+		} as unknown as PygieniumCtx["model"];
+		const seen: unknown[] = [];
+		setAgentRunner(async (opts) => {
+			seen.push(opts.model);
+			return fakeAgentRunner(opts);
+		});
+		await handleCheckCommand(
+			check,
+			"--fix",
+			{ ...stubCtx(cwd), model: selectedModel } as PygieniumCtx,
+		);
+
+		// Analysis + fix phases each dispatch one sub-agent.
+		expect(seen.length).toBe(2);
+		for (const m of seen) expect(m).toBe(selectedModel);
+	});
+
 	it("persists run-state.json at the expected path", async () => {
 		const check = smokeCheck();
 		await handleCheckCommand(check, "", stubCtx(cwd));

tests/deep-modules.test.ts

@@ -168,4 +168,17 @@ describe("deep-modules check", () => {
 			await rm(empty, { recursive: true, force: true }).catch(() => {});
 		}
 	});
+
+	it("does not skip when sources live in subdirectories (gate scans recursively)", async () => {
+		// Top level holds only a directory — the old shallow gate skipped any
+		// repo whose source lives under `game/`/`src/`-style subdirs.
+		await mkdir(join(cwd, "game"), { recursive: true });
+		await writeFile(join(cwd, "game", "main.lua"), "return {}\n", "utf8");
+		const check = getCheck("deep-modules")!;
+
+		await handleCheckCommand(check, "", stubCtx(cwd));
+
+		const state = await loadRunState(cwd);
+		expect(state?.checks["deep-modules"]?.status).toBe("complete");
+	});
 });

tests/defensive-guards.test.ts

@@ -229,4 +229,17 @@ describe("defensive-guards check", () => {
 		const state = await loadRunState(cwd);
 		expect(state?.checks["defensive-guards"]?.status).toBe("skipped");
 	});
+
+	it("does not skip when sources live in subdirectories (gate scans recursively)", async () => {
+		// Top level holds only a directory — the old shallow gate skipped any
+		// repo whose source lives under `game/`/`src/`-style subdirs.
+		await mkdir(join(cwd, "game"), { recursive: true });
+		await writeFile(join(cwd, "game", "main.lua"), "return {}\n", "utf8");
+		const check = getCheck("defensive-guards")!;
+
+		await handleCheckCommand(check, "", stubCtx(cwd));
+
+		const state = await loadRunState(cwd);
+		expect(state?.checks["defensive-guards"]?.status).toBe("complete");
+	});
 });

tests/todos.test.ts

@@ -205,6 +205,30 @@ describe("detectTodoStubs", () => {
 			loud.some((h) => h.path.endsWith("fetch.rs") && h.snippet === "todo!("),
 		).toBe(true);
 	});
+
+	it("never descends into build/deploy output directories", async () => {
+		// Generated bundles under framework build dirs must not feed the
+		// pre-scan: they dominate candidate counts with minified noise (the
+		// freno-dev failure flagged 119 of 124 candidates inside
+		// `.output`/`.vercel` bundles, ballooning the scan task to 2.5 MB).
+		for (const rel of [
+			join(".output", "public", "bundle.js"),
+			join(".vercel", "output", "static", "app.js"),
+			join(".netlify", "functions", "bundle.js"),
+		]) {
+			const full = join(dir, rel);
+			await mkdir(join(full, ".."), { recursive: true });
+			await writeFile(
+				full,
+				"// TODO: bundle placeholder\nfunction f(){ return 0; }\nthrow new Error('not implemented');\n",
+				"utf8",
+			);
+		}
+		const hits = await detectTodoStubs(dir);
+		expect(
+			hits.filter((h) => /(?:\.output|\.vercel|\.netlify)[/\\]/.test(h.path)),
+		).toHaveLength(0);
+	});
 });
 
 describe("todos check", () => {
@@ -311,6 +335,19 @@ describe("todos check", () => {
 		expect(state?.checks["todos"]?.status).toBe("skipped");
 	});
 
+	it("does not skip when sources live in subdirectories (gate scans recursively)", async () => {
+		// Top level holds only a directory — the old shallow gate skipped any
+		// repo whose source lives under `game/`/`src/`-style subdirs.
+		await mkdir(join(cwd, "game"), { recursive: true });
+		await writeFile(join(cwd, "game", "main.lua"), "return {}\n", "utf8");
+		const check = getCheck("todos")!;
+
+		await handleCheckCommand(check, "", stubCtx(cwd));
+
+		const state = await loadRunState(cwd);
+		expect(state?.checks["todos"]?.status).toBe("complete");
+	});
+
 	it("reports a new/resolved delta against the previous run's counts", async () => {
 		await seedStubs(cwd);
 		const check = getCheck("todos")!;
@@ -334,4 +371,25 @@ describe("todos check", () => {
 		);
 		expect(task).toContain("| new: 0 | resolved: 4 |");
 	});
+
+	it("truncates giant single-line candidates so the task prompt stays bounded", async () => {
+		// A minified/generated single line can be hundreds of KB; embedding it
+		// wholesale ballooned the freno-dev task to 2.5 MB and choked the
+		// analysis agent. The task must carry a truncated prefix, never the
+		// full line.
+		const long = `// TODO: ${"x".repeat(400)}`;
+		await writeFile(
+			join(cwd, "huge.ts"),
+			`${long}\nexport function f() { return 0; }\n`,
+			"utf8",
+		);
+		const task = await buildTodosScanTask(cwd, {
+			cwd,
+			target: cwd,
+			fix: false,
+			rest: [],
+		});
+		expect(task).not.toContain("x".repeat(400));
+		expect(task).toContain("…");
+	});
 });