fix(install): unblock upgrades from the deprecated @synsci/cli package (#15)

* chore: clean up agent prompt language, launcher install fallback, v1.2.3

- Reword the ml agent's 'flywheel' phrasing as 'model specialization' across the
  prompt and skills; only Atlas skills remain in the bundle.
- npx synsci: when global npm install fails (common on permissions), fall back
  to the standalone installer at openscience.sh/install instead of dead-ending.
- Bump CLI + launcher to 1.2.3.

* fix(managed): point default API base at app.syntheticsciences.ai

The managed/Atlas base defaulted to api.syntheticsciences.ai, which never
existed as a host — so managed mode failed with 'fetch failed' / DNS errors for
everyone. The real backend is app.syntheticsciences.ai (Vercel-fronted, /api/*
proxied to Fly). Repoint the default plus the hardcoded github/share/OIDC refs
(api.dev → app.dev for the preview path).

* fix(install): unblock upgrades from the deprecated @synsci/cli package

The package was renamed on npm from @synsci/cli to @synsci/openscience, and
both declare the same `openscience` bin. npm refuses to overwrite a bin file
owned by another package, so anyone with the old package still installed
globally hits `npm error code EEXIST` and dead-ends. Fix it in three layers:

- npx synsci: detect a global @synsci/cli up front and remove it (telling the
  user why) before resolving or installing the CLI; if the global install
  still fails with an EEXIST bin conflict, remove the stale package and retry
  once before falling back to the standalone installer. Also forward extra
  args through to `openscience web` and bump the launcher to 1.2.4.
- @synsci/openscience: ship a best-effort preinstall guard that clears a
  stale global @synsci/cli when npm_config_global is set; it always exits 0
  so it can never fail an install (script-blocking npm policies make this
  advisory only — the launcher and installer handle the conflict
  deterministically).
- openscience.sh/install: after installing the binary, warn when another
  `openscience` earlier on PATH would shadow it and point at
  `npm rm -g @synsci/cli`. Kept the root copy of the script in sync.

Author: aayambansal

backend/cli/package.json

@@ -1,12 +1,13 @@
 {
   "$schema": "https://json.schemastore.org/package.json",
-  "version": "1.2.2",
+  "version": "1.2.3",
   "name": "@synsci/openscience",
   "type": "module",
   "description": "AI-powered CLI for ML research and development workflows",
   "license": "Apache-2.0",
   "private": true,
   "scripts": {
+    "preinstall": "node ./script/preinstall.mjs || exit 0",
     "typecheck": "tsgo --noEmit",
     "test": "bun test",
     "build": "bun run script/build.ts",

backend/cli/script/preinstall.mjs

@@ -0,0 +1,22 @@
+#!/usr/bin/env node
+
+// Best-effort guard for GLOBAL installs only: the deprecated @synsci/cli
+// package links the same `openscience` bin, and npm refuses to overwrite a
+// bin file owned by another package (EEXIST) — which dead-ends upgrades.
+// Remove the stale package before ours is linked.
+//
+// This must NEVER fail the install: npm policies that block nested npm calls
+// (or script execution entirely) make this best-effort. The `npx synsci`
+// launcher and the openscience.sh/install script handle the conflict
+// deterministically.
+
+import { execSync } from "child_process"
+
+try {
+  if (process.env.npm_config_global === "true" || process.env.npm_config_global === "1") {
+    execSync("npm rm -g @synsci/cli", { stdio: "ignore", timeout: 60000 })
+  }
+} catch {
+  // swallow everything — the install must proceed
+}
+process.exit(0)

backend/cli/script/publish.ts

@@ -22,6 +22,7 @@ const version = Object.values(binaries)[0]
 
 await $`mkdir -p ./dist/${pkg.name}`
 await $`cp -r ./bin ./dist/${pkg.name}/bin`
+await $`cp ./script/preinstall.mjs ./dist/${pkg.name}/preinstall.mjs`
 await $`cp ./script/postinstall.mjs ./dist/${pkg.name}/postinstall.mjs`
 
 await Bun.file(`./dist/${pkg.name}/package.json`).write(
@@ -32,6 +33,9 @@ await Bun.file(`./dist/${pkg.name}/package.json`).write(
         openscience: `./bin/openscience`,
       },
       scripts: {
+        // best-effort: clears a stale global @synsci/cli whose `openscience`
+        // bin link would make npm refuse the install (EEXIST); never fails
+        preinstall: "node ./preinstall.mjs || exit 0",
         postinstall: "bun ./postinstall.mjs || node ./postinstall.mjs",
       },
       version: version,

backend/cli/skills/ml-training/model-economics/SKILL.md

@@ -13,7 +13,7 @@ dependencies: []
 
 ## When to Use This Skill
 
-Use this skill as the **first step** of any flywheel assessment to determine:
+Use this skill as the **first step** of any model-specialization assessment to determine:
 - Is it worth training a specialized model?
 - What will the total investment be?
 - When will the specialized model break even vs frontier APIs?

backend/cli/skills/research/initialize-atlas-graph/SKILL.md

@@ -18,7 +18,7 @@ clone at a different path resolves to the **same** graph.
 Run this skill when:
 - The canvas shows "no graph for this project" / the folder isn't linked yet.
 - The user asks to "initialize", "set up Atlas", or "start tracking" research.
-- You are about to begin research/flywheel work that should be recorded.
+- You are about to begin research work that should be recorded.
 
 ## Steps
 

backend/cli/src/agent/prompt/ml.txt

@@ -7,7 +7,7 @@ evidence: scope the problem, survey prior art, build the data pipeline, design a
 models, evaluate rigorously against real baselines, ablate what matters, and report honestly.
 You cover the full spectrum — classical ML, deep learning, LLMs, RL, and multimodal — and the
 complete experimental lifecycle. When the goal is to replace a frontier API dependency with a
-cheaper specialized model, you run the model flywheel (see the Model Specialization track).
+cheaper specialized model, you run the specialization loop (see the Model Specialization track).
 
 You are NOT a chatbot. You are an autonomous ML agent. Given a task, dataset, or research
 question, you independently produce a complete, reproducible analysis and result.
@@ -141,7 +141,7 @@ Before reporting, run the decoupled critique + reviewer gate. This is a blocking
 
 ---
 
-## Model Specialization Track (the flywheel)
+## Model Specialization Track
 When the goal is to REPLACE a frontier API dependency with a cheaper specialized model, run
 this loop instead of the generic pipeline: **production data → train specialized model →
 deploy → collect more data → retrain**. This is collaborative — present options with

backend/cli/src/cli/cmd/github.ts

@@ -355,7 +355,7 @@ export const GithubInstallCommand = cmd({
 
             async function getInstallation() {
               return await fetch(
-                `https://api.syntheticsciences.ai/get_github_app_installation?owner=${app.owner}&repo=${app.repo}`,
+                `https://app.syntheticsciences.ai/get_github_app_installation?owner=${app.owner}&repo=${app.repo}`,
               )
                 .then((res) => res.json())
                 .then((data) => data.installation)
@@ -664,7 +664,7 @@ export const GithubRunCommand = cmd({
 
       function normalizeOidcBaseUrl(): string {
         const value = process.env["OIDC_BASE_URL"]
-        if (!value) return "https://api.syntheticsciences.ai"
+        if (!value) return "https://app.syntheticsciences.ai"
         return value.replace(/\/+$/, "")
       }
 

backend/cli/src/endpoints.ts

@@ -17,7 +17,7 @@
  */
 
 /** Neutral public default. Contains no internal codename. */
-export const DEFAULT_MANAGED_API_BASE = "https://api.syntheticsciences.ai"
+export const DEFAULT_MANAGED_API_BASE = "https://app.syntheticsciences.ai"
 
 function stripTrailingSlashes(url: string): string {
   return url.replace(/\/+$/, "")

backend/cli/src/openscience/index.ts

@@ -185,8 +185,8 @@ export class InsufficientCreditsError extends Error {
 
 // ── Bundled atlas CLI resolution ─────────────────────────────────────────
 // The @synsci/atlas package ships as a dependency; its `atlas` binary lives in
-// node_modules. The agent shells out to native `atlas` commands (the research /
-// flywheel prompts drive the map + managed-compute path through it), so the
+// node_modules. The agent shells out to native `atlas` commands (the research
+// prompts drive the map + managed-compute path through it), so the
 // binary must be on the subprocess PATH without requiring a separate global
 // install. We resolve the package, prefer the npm-generated `.bin/atlas` shim,
 // and otherwise synthesize a tiny launcher in the openscience data dir that runs the

backend/cli/src/provider/provider.ts

@@ -109,7 +109,7 @@ export namespace Provider {
    * Blocker (c) — the inverse of requireAtlasProxyForManagedKey.
    *
    * A prior managed sync may have injected an Atlas proxy `*_BASE_URL` (e.g.
-   * ANTHROPIC_BASE_URL → api.syntheticsciences.ai). If the resolved key is the
+   * ANTHROPIC_BASE_URL → app.syntheticsciences.ai). If the resolved key is the
    * user's OWN key (BYOK), that key must NEVER be sent to the Atlas proxy — it
    * would leak the credential and mis-bill. Pin the base URL back to the
    * provider's public endpoint and drop the managed routing.