feat: initialize project configuration and implementation plan for Step-Zero

.codex/config.toml

@@ -0,0 +1,3 @@
+[features]
+memories = true
+codex_git_commit = true

implemention_plan.md

@@ -0,0 +1,72 @@
+This is your **Final Execution Playbook**. All the VC fluff, fake science, and over-engineered bloat have been stripped out.
+
+What remains is a brutally effective, highly defensible 48-hour sprint to build **Step-Zero: The Cognitive Pacemaker**.
+
+The core thesis you are building (and pitching) is this: *“What if an AI refused to let you think about more than one thing at a time?”*
+
+Here is the step-by-step implementation plan to build the tech, hit the corporate bounties, and record the winning demo.
+
+---
+
+### The Bounty Matrix (Your Checklist)
+
+* **Track:** Backyard AI 🏡 (Target: Anyone paralyzed by task friction).
+* **NVIDIA ($10k) + Llama Champion:** `Nemotron-Mini-4B` running locally via `llama-cpp-python` to generate the strict logic steps.
+* **OpenBMB ($10k):** `MiniCPM-3-4B` running locally to translate the logic step into the user's chosen "Activation Style" (Calm, Structured, Direct).
+* **Modal ($20k) + Well-Tuned:** A fast LoRA fine-tune on Nemotron explicitly for *Syntactic Verb Enforcement* (forcing outputs under 8 words starting with an action verb).
+* **OpenAI ($10k) + Off-Brand:** Custom HTML/JS "Fog of War" frontend generated 100% by GitHub Copilot/Codex, served over Gradio `gr.Server`.
+* **Badges:** Tiny Titan (≤4B models), Off the Grid (100% offline inference), Sharing is Caring (Agent JSON trace export).
+
+---
+
+### Phase 1: The Modal Fine-Tuning Sprint (Friday Night)
+
+*You do this first so you can use the fine-tuned model all weekend.*
+
+1. **The Dataset (`verbs.jsonl`):** Create 200 rows of training data.
+* *Input:* "I need to clean my room but I am paralyzed."
+* *Output:* "Pick up one sock." *(Notice: No conversational fluff. Starts with a verb. Under 8 words).*
+
+
+2. **The Modal Script:** Use Modal's serverless GPU to run an Unsloth script. Fine-tune `Nemotron-Mini-4B` on this dataset to heavily bias it toward strict, atomic, imperative statements.
+3. **The Export:** Download the adapter weights, convert to `.gguf`, and save it to your local machine as `nemotron_atomic_v1.gguf`.
+
+### Phase 2: The Local Engine & Circuit Breaker (Saturday Morning)
+
+*You build the Python backend that hot-swaps the models to keep RAM under 6GB.*
+
+1. **The GBNF Grammar:** Write a `Llama.cpp` grammar file that forces Nemotron to *only* output a single string, stripping its ability to hallucinate markdown or lists.
+2. **The State Machine (`app.py`):** * Initialize a fastAPI/Gradio backend with a simple Python dictionary for state: `{"current_task": "", "too_hard_count": 0, "past_steps": []}`.
+3. **The `[TOO HARD]` Fallback Loop:** * If the user hits `[TOO HARD]`, Python increments the counter.
+* It passes the current task back to Nemotron: *"The user failed to do: [Task]. Provide a physically smaller, easier pre-requisite action."*
+
+
+4. **The Hard Circuit Breaker:** * If `too_hard_count == 2`, the backend bypasses the LLM entirely and hard-returns: *"You are out of activation energy. Step away from the screen for 3 minutes. Leave this window open."* (This proves to judges you care about UX, not just spamming AI outputs).
+
+### Phase 3: The OpenBMB Stylizer (Saturday Afternoon)
+
+1. **The Pipeline:** Once Nemotron outputs the raw atomic step (e.g., "Open the PDF"), Python pipes that string into OpenBMB.
+2. **The Persona Prompt:** OpenBMB receives: *"Rewrite this task: 'Open the PDF' in a [Calm/Direct/Encouraging] tone. Keep it under 2 sentences."*
+3. **WebSocket Broadcast:** The final stylized text is pushed to the frontend via Gradio's Async API or a raw WebSocket.
+
+### Phase 4: The Codex "Fog of War" UI (Saturday Night)
+
+*Do not write this yourself. You must use Codex/Copilot to win the OpenAI track.*
+
+1. Open VS Code and use the Copilot Chat panel.
+2. **Your exact prompt to Copilot:** > *"Generate a single-page HTML/JS/Tailwind interface. It must have a pure black background. At the top, put a very faint, blurred text area called 'Breadcrumbs'. In the dead center, place a massive text block (72px font, white) for a single instruction. Below it, place two buttons: A green [I DID THIS] button, and a muted gray [THIS IS TOO HARD] button. Write the WebSocket JS to connect to a local backend at `ws://localhost:7860`. When the user clicks [I DID THIS], the center text fades out, moves to the blurred 'Breadcrumbs' section, and the new text fades in. Do not use standard chatbot layouts."*
+3. **Commit the Code:** Make a git commit with the message: `feat: Generated proprietary Fog of War SPA interface using OpenAI Codex`.
+4. Serve this `index.html` file using `gr.Server(serve_static=["index.html"])`.
+
+### Phase 5: The Demo & The Pitch (Sunday)
+
+The demo video is where you separate yourselves from the "wrapper" projects.
+
+**The Video Script (Keep it tight and grounded):**
+
+1. **The Hook (0:00-0:15):** *"Standard AI assistants are built for conversation and expansion. But when a user is experiencing executive dysfunction or task paralysis, a conversational list is overwhelming. We built Step-Zero: a local-only Cognitive Pacemaker that enforces strict attentional constraints. It refuses to let you see the future."*
+2. **The Tech Flex (0:15-0:45):** *"Because this involves sensitive behavioral friction, it runs 100% off-the-grid. We used Modal to fine-tune Nemotron-4B to strictly output 8-word imperative verbs. Nemotron manages our hidden friction-state machine. Once a task is scaled to the user's capability, OpenBMB-4B dynamically translates the tone to match the user's required activation energy."*
+3. **The UI Flex (0:45-1:15):** *"Gradio's default chat interface couldn't handle our 'Fog of War' design, so we used OpenAI Codex to generate a completely custom, headless DOM overlay hosted via `gr.Server`."*
+4. **The Live Demo (1:15-2:00):** Show the app. Type *"I need to write my tax report."* Show the screen displaying only *"Open a new folder."* Hit `[TOO HARD]`. Show Nemotron dynamically downgrading the task to *"Click the finder icon."* Hit `[TOO HARD]` again. Show the Circuit Breaker activating: *"Walk away for 3 minutes."*
+
+**Final check:** Does this plan align with your team's capability to split the work? One person on the Llama.cpp backend (Python), one person wrangling the Copilot frontend (HTML/JS), and one person managing the Modal Fine-tune (Data/Unsloth)?