Delete README.md

README.md

@@ -1,89 +0,0 @@
-# agent-fleet
-
-A minimal, runnable scaffold for the pipeline we designed:
-
-```
-Scout (cheap, high volume) -> Generate (cheap) -> Score (mid, kill gate) -> Build spec (mid)
-```
-
-Each stage is a small class (`src/core/Stage.js`) with its own model, prompt,
-and output parser. `Pipeline` (`src/core/Pipeline.js`) chains them and stops
-early if a stage returns `{ kill: true }` — this is what keeps a bad idea
-from burning money on the more expensive Score/Build steps.
-
-## Setup
-
-```bash
-cd agent-fleet
-npm install
-cp .env.example .env
-# edit .env: add your OPENROUTER_API_KEY, adjust MAX_SPEND_USD
-```
-
-Get an OpenRouter key at https://openrouter.ai/keys — it gives you one API
-for Gemini Flash, Llama, Claude, GPT, etc, which is why everything routes
-through it here even though you mentioned wanting to call Gemini directly
-later (swap `src/providers/openrouter.js` for a Gemini-native client when
-you're ready to scale — same `callModel()` interface, just point it at
-Gemini's endpoint instead).
-
-## Run it
-
-```bash
-# Sanity check: run ONLY the scout stage, see real output + cost
-node index.js scout
-
-# Run the full 4-stage pipeline once, sequentially
-node index.js pipeline
-
-# Run multiple pipelines in parallel (worker_threads), one per seed idea
-node src/runFleet.js
-```
-
-Every run prints running spend (`$X.XXXX`) after each stage so you can see
-exactly where your OpenRouter balance is going before you scale up.
-
-## What this does NOT do yet
-
-This scaffold stops at producing a **build spec** (a structured JSON plan),
-not an actual game/app/video. That's intentional — build (generate the
-artifact) and publish (put it live) are separate concerns, and you don't
-want an LLM call auto-publishing to an app store or social account. Next
-steps, in order of leverage:
-
-1. **Feed real signal into Scout.** Right now `seed.rawSignal` is hand-written
-   in `index.js` / `src/runFleet.js`. Replace with a real scraper (App Store
-   review pulls, Reddit/trend APIs, CrazyGames charts) so Scout is reacting
-   to real demand, not your guesses.
-2. **Wire BuildStage's spec into an actual codegen step.** For browser
-   games/apps, the `build_steps` array is meant to be handed to a coding
-   agent (e.g. Claude Code) to actually produce the artifact. For video,
-   it's meant to be handed to a script-to-assets pipeline.
-3. **Add a PublishStage with a manual approval gate.** Keep a human in the
-   loop before anything goes to an app store or social account — those are
-   account-level risks, not post-level ones.
-4. **Swap cheap-model calls to Gemini directly** once volume actually
-   justifies leaving OpenRouter's per-call overhead behind.
-
-## Cost model
-
-`src/providers/openrouter.js` tracks a rough running spend estimate and
-throws before any call that would exceed `MAX_SPEND_USD` in your `.env`.
-The price table (`ROUGH_PRICE_PER_1K` in `src/config.js`) is a rough
-estimate, not live pricing — check OpenRouter's pricing page and update it
-if you're being strict about a small balance.
-
-## File map
-
-```
-src/
-  config.js              # model routing + spend limit, all env-driven
-  providers/openrouter.js# API client + spend guard
-  core/Stage.js           # base class every stage extends
-  core/Pipeline.js        # chains stages, honors early "kill" signals
-  prompts/                # the actual prompt engineering (scout/generate/score)
-  stages/                 # thin classes wiring prompts + models together
-  worker.js               # one pipeline run, for use inside a worker_thread
-  runFleet.js             # spawns N parallel pipeline threads
-index.js                  # single-thread entry point for testing
-```