projectinator 0.1.5 → 0.3.0

package/README.md

@@ -6,8 +6,8 @@
 
 [![npm](https://img.shields.io/npm/v/projectinator.svg?color=e0a72d&label=npm)](https://www.npmjs.com/package/projectinator)
 ![license: MIT](https://img.shields.io/badge/license-MIT-blue.svg)
-![node](https://img.shields.io/badge/node-%E2%89%A520-brightgreen.svg)
-![tests: 140 passing](https://img.shields.io/badge/tests-140%20passing-brightgreen.svg)
+![node](https://img.shields.io/badge/node-%E2%89%A522.19-brightgreen.svg)
+![tests: 181 passing](https://img.shields.io/badge/tests-181%20passing-brightgreen.svg)
 ![TypeScript](https://img.shields.io/badge/TypeScript-strict-3178c6.svg)
 ![built on Pi](https://img.shields.io/badge/built%20on-Pi%20agent%20harness-e0a72d.svg)
 
@@ -22,7 +22,7 @@ watch it happen from a terminal cockpit: a live board, budget bar, and a standup
 
 Built on the [Pi](https://pi.dev) agent harness (Node/TypeScript). Bring your own API key.
 
-**Install & run** (Node ≥ 20):
+**Install & run** (Node ≥ 22.19):
 
 ```bash
 npx projectinator                 # run without installing
@@ -49,7 +49,8 @@ npm start
 </details>
 
 > **Optional:** `npx playwright install chromium` lets the tester actually run web apps in a
-> headless browser and enables live preview. Everything else works without it.
+> headless browser and enables live preview. Without it the tester can only read the code —
+> such passes show as **PASS\*** with a warning. Everything else works without it.
 
 > Projectinator spends **your** API money. Every screen shows the running cost; you set a
 > budget cap and it halts before crossing it. A tiny landing page is cents; a full app is
@@ -64,13 +65,15 @@ Type an idea → it plans → you approve → it builds, tests, and hands you wo
 - **Best model per role.** Roles bind to a *capability + tier*, never a model name. A
   swappable registry maps capabilities to models — new frontier model next month, edit one
   place, every route updates. Run a **bake-off** to pick empirically.
-- **A real pipeline.** PM decomposes → Designer specs → Developer writes files → Tester
-  **runs the app headless and catches real bugs** → feedback loop re-runs the dev on failure.
+- **A real pipeline.** PM decomposes → Designer specs → Developer writes files → **Reviewer**
+  (cheap, read-only) checks the wiring → Tester **runs the app headless and catches real
+  bugs** → feedback loop re-runs the dev on failure.
 - **Multi-file apps.** Vanilla HTML/CSS/JS or **React (CDN, no build)** — your choice.
 - **The cockpit.** A polished terminal UI: editable board, Kanban, standup, per-task cost,
   live budget bar, desktop notification when done.
 - **Honest cost.** Live spend tracking, per-project budget cap + an alert before the cap,
-  and predicted-vs-actual reporting that sharpens itself over real runs.
+  **per-task timeout and cost ceiling**, and predicted-vs-actual reporting that sharpens
+  itself over real runs.
 
 ## Highlights
 
@@ -79,13 +82,16 @@ Type an idea → it plans → you approve → it builds, tests, and hands you wo
 | 🧠 **PM intake** | Vague request? The PM asks 2–4 clarifying questions (with pickable options) before planning. Specific requests skip straight through. |
 | 🏛 **Deep plan (council)** | Opt-in: architect + product + risk leads propose epics in parallel, a synthesizer merges them, you approve, then they expand into the backlog. |
 | 🆚 **Model bake-off** | Run one task across models, an LLM judge scores the outputs, compare cost/latency/quality — save the winner to the registry. |
-| 🧪 **Real test execution** | The tester loads the built app in headless Chromium and fails on JS/console errors — not just by reading the code. |
+| 🔍 **Reviewer** | A cheap read-only model checks each code task's wiring (missing files, unresolved `<script src>`, undefined functions, ES modules on a static site) before the Tester spends a browser run. Fails feed the same Developer fix loop. |
+| 🧪 **Real test execution** | The tester loads the built app in headless Chromium and fails on JS/console errors — not just by reading the code. Without Chromium, passes are marked **PASS\*** so you know the app was never run. |
 | 👁 **Live preview** | Local server + auto-reload; ES modules and fetch resolve like production. |
 | 🚀 **Deploy** | One click to Cloudflare Pages, Vercel, or Netlify (their CLI + your login). |
 | 📤 **Export** | Backlog → Markdown, CSV, **Jira** CSV, **Trello** CSV. |
 | 📜 **Git per build** | The workspace is a git repo; one commit per task. History view + **undo a task**. |
 | 📊 **Analytics** | Retro (with optional AI narrative), burndown, cost by epic/model, and estimate accuracy that self-calibrates from real runs. |
 | 💾 **Templates** | Save a project's brief as a reusable template; import/share as a file. |
+| ⛔ **Task limits** | Per-task timeout and cost ceiling (Settings → Preferences). A runaway task is aborted, billed for what it spent, and the build halts — resumable. |
+| ✎ **Task notes** | Annotate any task on the board (`n`). Yours only — never sent to a model; included in exports. |
 
 ## How it works
 
@@ -95,8 +101,7 @@ idea
  └─ intake?     PM asks clarifying questions if the request is vague
  └─ plan mode?  Quick (one PM) or Deep (planning council → approve epics)
  └─ decompose   → a routed, epic-tagged backlog with a cost estimate
- └─ approve     → auto-run, or gate the backlog / gate again before dev
- └─ build       toposort deps · design → code → test · Tester→Dev feedback loop
+ └─ build       toposort deps · design → code → review → test · Reviewer/Tester→Dev feedback loop
  └─ done        working files + retro + deploy/export/preview
 ```
 
@@ -120,8 +125,9 @@ npm start                                         # the cockpit (the normal way
 npm run build -- --live --mini                    # cheap end-to-end proof (~$0.10)
 npm run build -- --live --lock anthropic "idea"   # full pipeline on one provider
 npm run build -- --live --mini --resume           # resume a halted/finished build (skips done tasks)
+npm run build -- --live --mini --task-cap 0.5 --task-timeout 5   # per-task limits (USD / minutes)
 npm run bakeoff -- --capability design "Design a pricing page"   # model bake-off
-npm test                                          # 140 tests
+npm test                                          # 181 tests
 npm run typecheck
 ```
 
@@ -129,7 +135,8 @@ npm run typecheck
 
 - **Keys** are stored at `~/.projectinator/config.json` (chmod 0600) — never in the repo.
 - **Settings** (in the app): API keys, preferred provider, default workflow, default stack,
-  model assignments, budget cap + alert %, estimate accuracy.
+  model assignments (incl. the Reviewer), budget cap + alert %, per-task timeout + cost cap,
+  estimate accuracy.
 - **User data** lives under `~/.projectinator/` (config, calibration, templates, exports).
   Each build gets its own git-versioned workspace folder (one commit per finished task).
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "projectinator",
-  "version": "0.1.5",
+  "version": "0.3.0",
   "description": "Your AI build team in the terminal — hand it an app idea, a PM model plans a Scrum backlog, and the best model per role designs, codes, and tests it into working files. Bring your own API key.",
   "type": "module",
   "private": false,
@@ -36,7 +36,7 @@
     "src"
   ],
   "engines": {
-    "node": ">=20"
+    "node": ">=22.19.0"
   },
   "scripts": {
     "start": "tsx src/tui.tsx",
@@ -61,13 +61,13 @@
     "vitest": "^2.1.8"
   },
   "dependencies": {
-    "@earendil-works/pi-coding-agent": "0.80.7",
+    "@earendil-works/pi-coding-agent": "0.85.1",
     "@inkjs/ui": "2.0.0",
     "ink": "7.1.0",
     "ink-spinner": "5.0.0",
     "playwright": "1.61.1",
     "react": "19.2.7",
     "tsx": "4.23.1",
-    "typebox": "1.1.38"
+    "typebox": "1.3.7"
   }
 }

package/src/bakeoff.ts

@@ -7,15 +7,13 @@
 // scoring) is a later step.
 
 import {
-  AuthStorage,
-  ModelRegistry,
   createAgentSession,
   defineTool,
   type AgentSession,
 } from "@earendil-works/pi-coding-agent";
 import { Type, type Static } from "typebox";
 import type { Capability, Difficulty, Provider, Task } from "./types.js";
-import { resolvePiModel } from "./executor.js";
+import { piRuntime, resolvePiModel } from "./executor.js"
 import { buildRolePrompt } from "./roles.js";
 import { estimateTokens } from "./estimate.js";
 import { addSessionCost } from "./session-cost.js";
@@ -72,13 +70,11 @@ const id = (c: Candidate) => `${c.provider}/${c.model}`;
 async function runCandidate(task: Task, cand: Candidate): Promise<BakeoffEntry> {
   const base: BakeoffEntry = { provider: cand.provider, model: cand.model, output: "", cost: 0, ms: 0, outputTokens: 0 };
   try {
-    const authStorage = AuthStorage.create();
-    const registry = ModelRegistry.create(authStorage);
-    const model = resolvePiModel(registry, cand.provider, cand.model);
+    const runtime = await piRuntime();
+    const model = resolvePiModel(runtime, cand.provider, cand.model);
     const { session } = await createAgentSession({
       model,
-      authStorage,
-      modelRegistry: registry,
+      modelRuntime: runtime,
       thinkingLevel: "medium",
       noTools: "all",
     });
@@ -144,14 +140,12 @@ async function judge(task: Task, entries: BakeoffEntry[], judgeCand: Candidate):
 
   const letters = scored.map((_, i) => String.fromCharCode(65 + i)); // A, B, C…
   const blocks = scored.map((e, i) => `### Option ${letters[i]}\n${e.output}`).join("\n\n");
-  const authStorage = AuthStorage.create();
-  const registry = ModelRegistry.create(authStorage);
-  const model = resolvePiModel(registry, judgeCand.provider, judgeCand.model);
+  const runtime = await piRuntime();
+  const model = resolvePiModel(runtime, judgeCand.provider, judgeCand.model);
   const { tool, get } = buildJudgeTool();
   const { session } = await createAgentSession({
     model,
-    authStorage,
-    modelRegistry: registry,
+    modelRuntime: runtime,
     thinkingLevel: "medium",
     noTools: "all",
     customTools: [tool],

package/src/build-state.ts

@@ -40,7 +40,8 @@ export function loadState(path: string): BuildState | undefined {
   }
 }
 
-/** Which task ids are already finished (last outcome wins). Used to skip on resume. */
+/** Which task ids are already finished (last outcome wins; a failed attempt does not
+ *  count). Used to skip on resume. */
 export function completedIds(state: BuildState): Set<string> {
-  return new Set(state.outcomes.map((o) => o.taskId));
+  return new Set(state.outcomes.filter((o) => !o.error).map((o) => o.taskId));
 }

package/src/cost.ts

@@ -22,7 +22,9 @@ function ratesFor(model: Model, inputTokens: number) {
 }
 
 /** Estimate USD cost for one task on a given model.
- *  A fraction of input can be served from cache at the cheaper cacheRead rate. */
+ *  A fraction of input can be served from cache at the cheaper cacheRead rate.
+ *  Kept to 4 decimals, not cents: a backlog of sub-cent tasks would otherwise
+ *  estimate to $0.00 and under-reserve budget in the parallel scheduler. */
 export function estimateCost(est: TokenEstimate, model: Model): number {
   const cachedFraction = clamp01(est.cachedInputFraction ?? 0);
   const rate = ratesFor(model, est.input);
@@ -33,13 +35,9 @@ export function estimateCost(est: TokenEstimate, model: Model): number {
   const inputCost = (freshInput / PER_MILLION) * rate.input + (cachedInput / PER_MILLION) * rate.cacheRead;
   const outputCost = (est.output / PER_MILLION) * rate.output;
 
-  return round2(inputCost + outputCost);
+  return Math.round((inputCost + outputCost) * 10_000) / 10_000;
 }
 
 function clamp01(n: number): number {
   return Math.max(0, Math.min(1, n));
 }
-
-function round2(n: number): number {
-  return Math.round(n * 100) / 100;
-}

package/src/council.ts

@@ -4,8 +4,6 @@
 // via the normal decomposer seeded with these epics.
 
 import {
-  AuthStorage,
-  ModelRegistry,
   createAgentSession,
   defineTool,
   type AgentSession,
@@ -13,7 +11,7 @@ import {
 import { Type, type Static } from "typebox";
 import type { Backend, Provider } from "./types.js";
 import { findEntry } from "./registry.js";
-import { resolvePiModel } from "./executor.js";
+import { piRuntime, resolvePiModel } from "./executor.js"
 import { addSessionCost } from "./session-cost.js";
 
 export interface Epic {
@@ -67,17 +65,15 @@ interface Ctx {
 }
 
 async function runEpicAgent(idea: string, system: string, toolName: string, ctx: Ctx): Promise<Epic[]> {
-  const authStorage = AuthStorage.create();
-  const registry = ModelRegistry.create(authStorage);
+  const runtime = await piRuntime();
   const { entry } = findEntry("plan", "mid");
   const pick = ctx.modelOverride ?? entry.byBackend[ctx.backend];
   try {
-    const model = resolvePiModel(registry, pick.provider, pick.model);
+    const model = resolvePiModel(runtime, pick.provider, pick.model);
     const { tool, get } = buildEpicsTool(toolName);
     const { session } = await createAgentSession({
       model,
-      authStorage,
-      modelRegistry: registry,
+      modelRuntime: runtime,
       thinkingLevel: "low",
       noTools: "all",
       customTools: [tool],
@@ -145,15 +141,14 @@ export async function councilEpics(idea: string, ctx: Ctx): Promise<CouncilResul
   if (!proposals.length) return { epics: [], proposals: [] };
 
   // Synthesize.
-  const authStorage = AuthStorage.create();
-  const registry = ModelRegistry.create(authStorage);
+  const runtime = await piRuntime();
   const { entry } = findEntry("plan", "mid");
   const pick = ctx.modelOverride ?? entry.byBackend[ctx.backend];
   try {
-    const model = resolvePiModel(registry, pick.provider, pick.model);
+    const model = resolvePiModel(runtime, pick.provider, pick.model);
     const { tool, get } = buildEpicsTool("submit_epics");
     const { session } = await createAgentSession({
-      model, authStorage, modelRegistry: registry, thinkingLevel: "low",
+      model, modelRuntime: runtime, thinkingLevel: "low",
       noTools: "all", customTools: [tool], tools: ["submit_epics"],
     });
     try {

package/src/estimate.ts

@@ -37,6 +37,13 @@ const BUCKETS: Record<Capability, Record<Difficulty, Bucket>> = {
     medium: { input: 25_000, output: 7_000 },
     high: { input: 60_000, output: 14_000 },
   },
+  // Review reads the same tree the tester does but never runs it — shorter output.
+  review: {
+    trivial: { input: 15_000, output: 800 },
+    low: { input: 25_000, output: 1_500 },
+    medium: { input: 50_000, output: 3_000 },
+    high: { input: 80_000, output: 5_000 },
+  },
   test: {
     trivial: { input: 15_000, output: 1_200 },
     low: { input: 25_000, output: 2_500 },

package/src/executor.ts

@@ -6,8 +6,7 @@
 // hits the provider API and spends money — that path is guarded by the caller.
 
 import {
-  AuthStorage,
-  ModelRegistry,
+  ModelRuntime,
   createAgentSession,
   type AgentSession,
   type AgentSessionEvent,
@@ -20,17 +19,23 @@ import { getModel } from "./models.js";
 import { addSessionCost } from "./session-cost.js";
 
 /** Pi's own Model type, derived so we don't depend on a deep sub-path import. */
-type PiModel = NonNullable<ReturnType<ModelRegistry["find"]>>;
+export type PiModel = NonNullable<ReturnType<ModelRuntime["getModel"]>>;
+
+/** Pi's model/auth runtime: built-in catalog + env keys / ~/.pi/agent/auth.json.
+ *  Created per call on purpose — a key added in Settings must apply to the next session. */
+export function piRuntime(): Promise<ModelRuntime> {
+  return ModelRuntime.create();
+}
 
 /** Resolve a Projectinator (provider, modelId) to Pi's executable Model.
  *  Offline + free — reads Pi's built-in registry. Throws with a clear message
  *  if the id isn't one Pi knows (our ids are kept identical to Pi's on purpose). */
 export function resolvePiModel(
-  registry: ModelRegistry,
+  runtime: ModelRuntime,
   provider: Provider,
   modelId: string,
 ): PiModel {
-  const m = registry.find(provider, modelId);
+  const m = runtime.getModel(provider, modelId);
   if (!m) {
     throw new Error(
       `Pi has no model "${provider}/${modelId}". ` +
@@ -47,8 +52,6 @@ export interface ExecuteOptions {
   thinkingLevel?: "off" | "low" | "medium" | "high";
   /** Optional progress hook — receives raw Pi session events. */
   onEvent?: (event: AgentSessionEvent) => void;
-  /** Override auth (tests). Default resolves env keys / ~/.pi/agent/auth.json. */
-  authStorage?: AuthStorage;
   /** Tools the agent may use. Default: the coding set. */
   tools?: string[];
 }
@@ -95,15 +98,13 @@ export async function executeTask(
   decision: RouteDecision,
   opts: ExecuteOptions,
 ): Promise<ExecuteResult> {
-  const authStorage = opts.authStorage ?? AuthStorage.create();
-  const registry = ModelRegistry.create(authStorage);
-  const model = resolvePiModel(registry, decision.provider, decision.model.id);
+  const runtime = await piRuntime();
+  const model = resolvePiModel(runtime, decision.provider, decision.model.id);
 
   const { session } = await createAgentSession({
     model,
     cwd: opts.workspace,
-    authStorage,
-    modelRegistry: registry,
+    modelRuntime: runtime,
     thinkingLevel: opts.thinkingLevel ?? "medium",
     tools: opts.tools ?? ["read", "write", "edit", "bash", "ls", "grep", "find"],
   });

package/src/intake.ts

@@ -6,15 +6,13 @@
 // schema, coerced/validated in code.
 
 import {
-  AuthStorage,
-  ModelRegistry,
   createAgentSession,
   defineTool,
 } from "@earendil-works/pi-coding-agent";
 import { Type, type Static } from "typebox";
 import type { Backend, Provider } from "./types.js";
 import { findEntry } from "./registry.js";
-import { resolvePiModel } from "./executor.js";
+import { piRuntime, resolvePiModel } from "./executor.js"
 import { addSessionCost } from "./session-cost.js";
 
 const IntakeSchema = Type.Object(
@@ -78,18 +76,16 @@ export interface AssessOptions {
 /** Ask the PM whether the request needs clarification; returns up to 4 questions
  *  (empty = clear enough to plan directly). Never throws — returns [] on trouble. */
 export async function assessIntake(idea: string, opts: AssessOptions): Promise<IntakeQuestion[]> {
-  const authStorage = AuthStorage.create();
-  const registry = ModelRegistry.create(authStorage);
+  const runtime = await piRuntime();
   const { entry } = findEntry("plan", "mid");
   const pick = opts.modelOverride ?? entry.byBackend[opts.backend];
 
   try {
-    const model = resolvePiModel(registry, pick.provider, pick.model);
+    const model = resolvePiModel(runtime, pick.provider, pick.model);
     const { tool, get } = buildIntakeTool();
     const { session } = await createAgentSession({
       model,
-      authStorage,
-      modelRegistry: registry,
+      modelRuntime: runtime,
       thinkingLevel: "low",
       noTools: "all",
       customTools: [tool],

package/src/models.ts

@@ -1,36 +1,42 @@
-// Model pricing table. Rates USD per 1,000,000 tokens (verified July 2026).
-// Shape mirrors Pi's models.json `cost` block so this ports to Pi later.
-// cacheRead ~= 0.1x input, cacheWrite ~= 1.25x input (Anthropic-style; approximations
-// for models whose exact cache rates we haven't pinned — refine against provider docs).
+// Model pricing table. Rates USD per 1,000,000 tokens — copied verbatim from Pi's
+// bundled catalog (pi-coding-agent 0.85.1, Sept 2026); test/executor.test.ts pins every
+// entry against it. Shape mirrors Pi's models.json `cost` block.
 
 import type { Model } from "./types.js";
 import { findOpenRouterModel } from "./openrouter.js";
 
 export const MODELS: Record<string, Model> = {
-  // ---- OpenAI: GPT-5.6 family ----
+  // ---- OpenAI: GPT-5.6 family (prices cut Sept 2026; past 272k input costs 2x) ----
   "gpt-5.6-sol": {
     id: "gpt-5.6-sol",
     provider: "openai",
     name: "GPT-5.6 Sol",
     contextWindow: 272_000,
-    cost: { input: 5, output: 30, cacheRead: 0.5, cacheWrite: 6.25 },
+    cost: { input: 4, output: 20, cacheRead: 0.4, cacheWrite: 5, tiers: [{ inputTokensAbove: 272_000, input: 8, output: 30, cacheRead: 0.8, cacheWrite: 10 }] },
   },
   "gpt-5.6-terra": {
     id: "gpt-5.6-terra",
     provider: "openai",
     name: "GPT-5.6 Terra",
     contextWindow: 272_000,
-    cost: { input: 2.5, output: 15, cacheRead: 0.25, cacheWrite: 3.125 },
+    cost: { input: 2, output: 12, cacheRead: 0.2, cacheWrite: 2.5, tiers: [{ inputTokensAbove: 272_000, input: 4, output: 18, cacheRead: 0.4, cacheWrite: 5 }] },
   },
   "gpt-5.6-luna": {
     id: "gpt-5.6-luna",
     provider: "openai",
     name: "GPT-5.6 Luna",
     contextWindow: 272_000,
-    cost: { input: 1, output: 6, cacheRead: 0.1, cacheWrite: 1.25 },
+    cost: { input: 0.2, output: 1.2, cacheRead: 0.02, cacheWrite: 0.25, tiers: [{ inputTokensAbove: 272_000, input: 0.4, output: 1.8, cacheRead: 0.04, cacheWrite: 0.5 }] },
   },
 
   // ---- Anthropic: Claude ----
+  "claude-fable-5-1": {
+    id: "claude-fable-5-1",
+    provider: "anthropic",
+    name: "Claude Fable 5.1",
+    contextWindow: 1_000_000,
+    cost: { input: 10, output: 50, cacheRead: 0.25, cacheWrite: 12.5 },
+  },
   "claude-fable-5": {
     id: "claude-fable-5",
     provider: "anthropic",
@@ -38,6 +44,13 @@ export const MODELS: Record<string, Model> = {
     contextWindow: 1_000_000,
     cost: { input: 10, output: 50, cacheRead: 1, cacheWrite: 12.5 },
   },
+  "claude-opus-5": {
+    id: "claude-opus-5",
+    provider: "anthropic",
+    name: "Claude Opus 5",
+    contextWindow: 1_000_000,
+    cost: { input: 5, output: 25, cacheRead: 0.5, cacheWrite: 6.25 },
+  },
   "claude-opus-4-8": {
     id: "claude-opus-4-8",
     provider: "anthropic",
@@ -45,6 +58,13 @@ export const MODELS: Record<string, Model> = {
     contextWindow: 1_000_000,
     cost: { input: 5, output: 25, cacheRead: 0.5, cacheWrite: 6.25 },
   },
+  "claude-sonnet-5": {
+    id: "claude-sonnet-5",
+    provider: "anthropic",
+    name: "Claude Sonnet 5",
+    contextWindow: 1_000_000,
+    cost: { input: 2, output: 10, cacheRead: 0.2, cacheWrite: 2.5 },
+  },
   "claude-sonnet-4-6": {
     id: "claude-sonnet-4-6",
     provider: "anthropic",
@@ -77,6 +97,13 @@ export const MODELS: Record<string, Model> = {
       tiers: [{ inputTokensAbove: 200_000, input: 4, output: 18, cacheRead: 0.4, cacheWrite: 5 }],
     },
   },
+  "gemini-3.8-flash": {
+    id: "gemini-3.8-flash",
+    provider: "google",
+    name: "Gemini 3.8 Flash",
+    contextWindow: 1_000_000,
+    cost: { input: 0.75, output: 3.75, cacheRead: 0.075 },
+  },
   "gemini-3-flash-preview": {
     id: "gemini-3-flash-preview",
     provider: "google",
@@ -88,6 +115,13 @@ export const MODELS: Record<string, Model> = {
   // ---- OpenRouter (one key → frontier models). ids are Pi's OpenRouter-catalog
   // slugs (vendor/model). Pricing mirrors the underlying model (OpenRouter passes
   // it through, ~small margin); ACTUAL cost still comes from Pi per run.
+  "anthropic/claude-opus-5": {
+    id: "anthropic/claude-opus-5",
+    provider: "openrouter",
+    name: "Claude Opus 5 (OpenRouter)",
+    contextWindow: 1_000_000,
+    cost: { input: 5, output: 25, cacheRead: 0.5, cacheWrite: 6.25 },
+  },
   "anthropic/claude-opus-4.8": {
     id: "anthropic/claude-opus-4.8",
     provider: "openrouter",
@@ -95,6 +129,13 @@ export const MODELS: Record<string, Model> = {
     contextWindow: 1_000_000,
     cost: { input: 5, output: 25, cacheRead: 0.5, cacheWrite: 6.25 },
   },
+  "anthropic/claude-sonnet-5": {
+    id: "anthropic/claude-sonnet-5",
+    provider: "openrouter",
+    name: "Claude Sonnet 5 (OpenRouter)",
+    contextWindow: 1_000_000,
+    cost: { input: 2, output: 10, cacheRead: 0.2, cacheWrite: 2.5 },
+  },
   "anthropic/claude-sonnet-4.6": {
     id: "anthropic/claude-sonnet-4.6",
     provider: "openrouter",
@@ -102,12 +143,19 @@ export const MODELS: Record<string, Model> = {
     contextWindow: 1_000_000,
     cost: { input: 3, output: 15, cacheRead: 0.3, cacheWrite: 3.75 },
   },
+  "google/gemini-3.8-flash": {
+    id: "google/gemini-3.8-flash",
+    provider: "openrouter",
+    name: "Gemini 3.8 Flash (OpenRouter)",
+    contextWindow: 1_000_000,
+    cost: { input: 0.75, output: 3.75, cacheRead: 0.075 },
+  },
   "openai/gpt-5.6-luna": {
     id: "openai/gpt-5.6-luna",
     provider: "openrouter",
     name: "GPT-5.6 Luna (OpenRouter)",
-    contextWindow: 272_000,
-    cost: { input: 1, output: 6, cacheRead: 0.1, cacheWrite: 1.25 },
+    contextWindow: 1_050_000,
+    cost: { input: 0.2, output: 1.2, cacheRead: 0.02, cacheWrite: 0.25 },
   },
 };
 

package/src/narrate.ts

@@ -3,15 +3,13 @@
 // (costs a small call) and cached on the build state by the caller.
 
 import {
-  AuthStorage,
-  ModelRegistry,
   createAgentSession,
   type AgentSession,
 } from "@earendil-works/pi-coding-agent";
 import type { Backend, Provider } from "./types.js";
 import type { RetroReport } from "./retro.js";
 import { findEntry } from "./registry.js";
-import { resolvePiModel } from "./executor.js";
+import { piRuntime, resolvePiModel } from "./executor.js"
 import { addSessionCost } from "./session-cost.js";
 
 function lastAssistantText(session: AgentSession): string {
@@ -66,16 +64,14 @@ export interface NarrateOptions {
 
 /** Generate the narrative. Throws on failure (caller shows the error). */
 export async function narrateRetro(report: RetroReport, opts: NarrateOptions): Promise<string> {
-  const authStorage = AuthStorage.create();
-  const registry = ModelRegistry.create(authStorage);
+  const runtime = await piRuntime();
   const { entry } = findEntry("plan", "mid");
   const pick = opts.modelOverride ?? entry.byBackend[opts.backend];
-  const model = resolvePiModel(registry, pick.provider, pick.model);
+  const model = resolvePiModel(runtime, pick.provider, pick.model);
 
   const { session } = await createAgentSession({
     model,
-    authStorage,
-    modelRegistry: registry,
+    modelRuntime: runtime,
     thinkingLevel: "low",
     noTools: "all",
   });

package/src/openrouter.ts

@@ -6,7 +6,7 @@
 //  - openRouterModels() / findOpenRouterModel(): SYNC reads (disk cache, else Pi's
 //    built-in list) so cost estimation (getModel) can price any picked model.
 
-import { AuthStorage, ModelRegistry } from "@earendil-works/pi-coding-agent";
+import { piRuntime } from "./executor.js";
 import { readFileSync, writeFileSync, existsSync, mkdirSync } from "node:fs";
 import { homedir } from "node:os";
 import { join } from "node:path";
@@ -20,19 +20,25 @@ const CACHE = join(homedir(), ".projectinator", "openrouter-models.json");
 let builtinMemo: ORModel[] | null = null;
 let diskMemo: ORModel[] | null | undefined; // undefined = not read yet, null = no cache
 
-/** Pi's built-in OpenRouter catalog — offline, always available, names + pricing. */
+/** Pi's built-in OpenRouter catalog — offline, names + pricing. Sync read of the memo;
+ *  call warmBuiltinOpenRouterModels() once at startup to fill it (Pi's runtime is async). */
 export function builtinOpenRouterModels(): ORModel[] {
+  return builtinMemo ?? [];
+}
+
+/** Load Pi's built-in OpenRouter list into the memo. Safe to call repeatedly. */
+export async function warmBuiltinOpenRouterModels(): Promise<ORModel[]> {
   if (builtinMemo) return builtinMemo;
   try {
-    const reg = ModelRegistry.create(AuthStorage.create());
-    const all = (reg.getAll() as unknown as Array<{ id: string; provider: string; name?: string; contextWindow?: number; cost?: ModelCost }>);
-    builtinMemo = all
-      .filter((m) => m.provider === "openrouter" && m.cost)
+    const runtime = await piRuntime();
+    builtinMemo = runtime
+      .getModels("openrouter")
+      .filter((m) => !!m.cost)
       .map((m) => ({
         id: m.id,
         name: m.name ?? m.id,
         contextWindow: m.contextWindow ?? 200_000,
-        cost: m.cost as ModelCost,
+        cost: m.cost,
       }))
       .sort((a, b) => a.name.localeCompare(b.name));
   } catch {