Dottie

Dottie is a closed-loop LLM factory run as one monorepo: data pipeline → train → eval gate → serve → agent → traces → retrain. The agent's only tool is scout, a CLI it can extend at runtime when a task needs a new capability.

Solo personal project, no connection to employer, built with public/free-tier only (R2/Workers/Supabase/HF ZeroGPU, ONNX WASM, public pip).

The Dottie site (apps/bluehenre, live at https://www.bhenre.com) is the operator's window into and steering wheel for this loop — one product with three faces: a Guide (Manus/OpenClaw/Hermes-style agentic assistant), a Hub (HuggingFace-style datasets/models/research registry), and a Monitor (Weights&Biases-style real-time view of the local development the org runs) — differentiated by provenance-honesty by construction. See apps/bluehenre/SPEC.md.

Monorepo layout

Path	Was	What it is
apps/ava-factory	ava-agi-factory-v6-4	Model factory: data pipeline, trainer, FastAPI serve engine, CPU-pilot chain, research engine
apps/scout-cli	scout-cli	The scout CLI (ex-BigBang) — 20+ capability-declared plugins plus the forge engine
apps/scout-rtx	scout-rtx	Windows RTX hill-climb runner (torch cu128 hard-pin; excluded from the uv workspace)
apps/dottie	—	Agent OS: engine, flywheel, policy guardrails, demand queue, task inbox
apps/bluehenre	—	The Dottie site: agentic-assistant Guide + provenance-honest artifact Hub + real-time dev Monitor (SPEC.md)
packages/ava-skills	ava-skills	Skill system (memory-router, memory-mint, code-bench, safety-scanner, …)
packages/ava-open-harness	ava-open-harness	Eval gate: J-Space tests, 11-category rubric, anti-mock guard
packages/personal-graphify	personal-graphify	Code knowledge-graph CLI/library (Ollama-first)

Root pyproject.toml is a virtual uv workspace over the light packages; apps/scout-rtx and apps/ava-factory are deliberately excluded (heavy, pinned deps).

uv sync                          # install workspace members editable
uv run pytest packages/ava-skills
uv run pytest apps/scout-cli/tests
scout --json forge list
scout --json system doctor

Single-CLI doctrine

The LLM and harness get exactly one tool: scout. Every capability is a plugin behind scout --json ..., and the agent adds capabilities itself through the forge engine:

scout --json forge new github --description "GitHub API wrapper" --domains api.github.com --network
scout --json forge cat github            # read the generated cli.py
scout --json forge edit github --code '<implementation>'
scout --json forge test github           # smoke test
scout skill install github --target dottie   # persist as a skill for the next session

# or generate from a spec
scout --json forge from-openapi --name linear --url https://api.linear.app/openapi.json
scout --json forge from-mcp --name notion --url https://mcp.notion.com/sse

Every plugin declares capabilities (network / filesystem / secrets, default deny) in a manifest and speaks --json. Forge engine: apps/scout-cli/bigbang/plugins/forge/cli.py (shipped in 4528f85; self-evolution loop verified end-to-end: new demo_tool → hello → test pass → rm).

Quickstart

git clone https://github.com/jcdavis131/dottie.git && cd dottie
export DOTTIE_ROOT=$(pwd)

uv sync
pip install -e apps/scout-cli

scout --help
scout --json ava status

# factory in smoke mode (CPU, nano preset)
cd apps/ava-factory
python scripts/dottie_continuous_loop.py --mode monitor
python scripts/dottie_continuous_loop.py --mode data --tokens 500000
python scripts/dottie_continuous_loop.py --mode train --preset nano

Full-scale training runs on a local RTX box; the nano preset (13.8M params) exists so the whole loop is exercisable on CPU.

Training

Model presets: nano 13.8M → mini 171M → base1b 1.4B, with a J-Space architecture (four workspaces at different update half-lives: S1 fast, S2 slow, Critic, Planner) and WSD schedule + YaRN context extension. The phase curriculum (P0 logic → P1 math → P2 foundation → P3 reasoning → P4 long-context → P5 anneal) with its token budgets is the design target, not a completed run — actual progress is published to the live console, and training telemetry (reports/dottie_telemetry.jsonl, dottie_live_status.json) is generated locally and gitignored, never committed.

Data pipeline: streaming manifests (fineweb-edu, proof-pile-2, synthetic logic, openwiki), six-stage curation (clean, dedup, decontamination with 13-gram overlap vs eval sets, tagging), packed with a 98/1/1 split.

Eval gate

uv run pytest packages/ava-open-harness -q

Checkpoints are only promoted to serving if they pass ava-open-harness: J-Space behavioral tests, the 11-category weighted rubric, safety evals, and test_no_mock.py — a guard that exists because an earlier version of this project fabricated eval scores, and every number must now come from a live forward pass or fail with a structured error.

Repo hygiene

• Generated telemetry and status files are gitignored and verified untracked by a CI job (Verify no telemetry tracked).
• reports/ keeps only curated reports; runs/cpu_pilot/ keeps text evidence (manifest, tokenizer, reports) — checkpoint binaries are regenerable via scripts/cpu_pilot_e2e.py and never committed.
• Research experiments under research-engine/experiments/<arxiv>/ are committed as curated notes; their run logs are not.
• CI (.github/workflows/ci.yml): uv sync, ruff, pytest for skills + harness + scout-cli, forge smoke test, quick eval gate, telemetry hygiene check.

Status

• Done: monorepo cutover (six repos subtree-merged, 2026-07-18), forge self-evolution engine (4528f85), telemetry hygiene + CI (2026-07-19).
• In progress: training the mini 171M on local GPU and promoting it through the eval gate to serve as the default brain; wiring scout ava as the single wrapper over the continuous-loop modes; closing the flywheel from site traces back into the anneal phase.

License

MIT — Solo personal project, no connection to employer, built with public/free-tier only. See LICENSE and per-package READMEs.