daily-builder

An autonomous portfolio builder that turns idle compute into shipped GitHub projects.

Every session, start.sh picks a domain, generates a fresh project idea, dedupes it against history, hands the spec to Claude Code, builds the thing, evaluates the result, and updates a live mission-control dashboard. You wake up to another repo on your profile.

Why this exists

The mini-projects are a side-effect. The real product is the builder itself — the domain rotation, the dedupe pass, the wishlist with AI-curated alternates, the git-derived authoritative status, the token telemetry, the live dashboard, the finishing-pass workflow.

A few of the things it has shipped so far:

Project	Domain	Stack
`circuit-trace`	AI/ML interpretability	Python, PyTorch, Three.js, FastAPI
`heap-forge`	Systems / low-level	C11, Go (cgo), TypeScript, React
`gossip-mesh-visualizer`	Distributed systems	Go, WebSocket, D3.js, SWIM
`packet-replay-forge`	Networking	Python, Scapy, FastAPI, React
`logflow-anomaly-engine`	Data engineering	Python, DuckDB, sklearn, React
`dep-drift-analyzer`	Devtools	Python, FastAPI, React
`quorum-probe`	Distributed systems	Go, Raft

All built end-to-end by the same orchestration loop documented below.

The full workflow

1 · One command, zero ceremony

bash ~/daily-builder/start.sh

The dashboard boots in the background at http://localhost:8765 and the launcher checks for an unfinished project (via git activity, never state.md — state.md lies). If something's resumable, you get a one-key prompt:

  ◉ Unfinished project detected:

    heap-forge

  Status:   IN PROGRESS (working tree clean)
  Progress: 12 / 12 steps · 100%
  Last:     5b1c2af release: v1.0.0 — heap-forge complete

  Resume this project? (y/n):

If you say no — or there's nothing to resume — you get a mode prompt.

2 · Pick a mode

  Pick a mode:
    s — surprise   LRU domain + auto-generate idea
    g — guided     pick domain + give angle
    w — wishlist   pick from wishlist.md

Mode	What it does
surprise	Picks the least-recently-used domain (weighted by `config.json`), generates an idea inside it, dedupes against history.
guided	You choose the domain (1–6) and write a one-line angle. The angle gets injected into the idea prompt.
wishlist	Pulls from `wishlist.md` + an AI-generated alternate list. (Documented in section 3 below.)

After idea generation, every mode runs a TF-IDF cosine similarity check against project_history.md. Above the configured threshold? It regenerates with an explicit "this was too similar to X — go in a totally different direction" correction. Up to 3 attempts.

3 · Wishlist mode — pick, add, or curate

The wishlist is your idea backlog. The launcher always offers three things at once:

Numbered entries from your ## Unused list
AI-generated curated alternates (silently regenerated when the wishlist changes)
An a shortcut to add a new idea right then and there, even when the list isn't empty

  Your wishlist:
    1 — LLM-Automated Incident Response Sandbox. Ingests raw SOC alerts and...
    2 — gossip-cluster-paint: real-time SWIM convergence visualizer

  Curated suggestions (LLM-generated alternates):
    3 — trace-bottleneck-hunter: LLM-powered OpenTelemetry analyzer
    4 — inline-rule-transpiler: convert MITRE ATT&CK rules into Splunk/ELK
    5 — config-drift-fixer: Terraform/CloudFormation divergence detector
    6 — gossip-cluster-paint: SWIM membership convergence visualizer

  Actions:
    a — add a new idea to your wishlist
    s — switch to surprise mode instead
    q — quit

  Pick a number, or a/s/q:

Picking a numbered item → that idea becomes the build target. Your own entries get moved to the ## Used section so they don't get picked twice. Curated suggestions stay in curated.md and regenerate next time you change the wishlist.

The dashboard exposes the same workflow in a panel — same ideas, same counts, same add affordance:

4 · What happens when a project is built

When you exit the Claude session at the end of a build, the launcher does the following automatically:

Evaluator runs — 8 heuristic checks (README quality, git activity, tests, stub detection, secret scan, runnability, state.md health, plan progress) plus an optional LLM qualitative pass. Score gets written to evaluation.json inside the project, and appended to project_history.md.
Status flips — git mtime + final commit subject pattern (release:, complete, v1.0.0) flip the dashboard chip from Building to Shipped.
Wishlist housekeeping — if you started from a wishlist item, it was already moved to ## Used at session start so it can't be picked again. Curated alternates regenerate on the next wishlist change.
Dashboard updates live — SSE pushes the new state to all connected browsers; the calendar gets the day's commits, the velocity graph picks up the new bars, the Shipped count ticks up.
Finishing-pass suggestion — if the score is below the configured threshold (default 75), the launcher prints Suggest: start.sh --polish <repo> so you know to schedule a cleanup session.

Nothing about "shipped" lives in state.md. Status is always derived from the git working tree, the commit log, and the eval score — three sources you can't fake.

The dashboard

Five sections, all clickable, all live-updating over Server-Sent Events.

Overview · live mission status

Shows what's currently building (or "no project running"), the streak counter, the shipped count, and total commits. Each card opens to a detail view.

Click the Shipped stat → modal listing every project that reached COMPLETE, with eval scores and direct GitHub links:

Activity · cadence + token burn

Token usage tile (Max plan input/output/cache breakdown by week, top projects), 365-day commit calendar, language distribution donut, 30-day velocity bars, hour×day heatmap, recent commits feed.

Click any green calendar day → the day's commits drill-down with hour bars and per-project stats:

Projects · the portfolio

All projects with authoritative status (Building / Shipped / Stalled based on days_since_last_commit), eval-score chip, line counts, commits, and time-since- last-activity. Filter by status, sort by recency / commits / lines / name.

Click a project card → drawer with five tabs (Overview, Commits, Files, Languages, README) and four actions:

Action	What it does
▶ Resume	Spawns `Terminal.app` running `start.sh --resume <repo>`.
◆ Polish	Spawns Terminal running the dedicated finishing-pass session.
✓ Evaluate	Runs the evaluator in-place; the card re-renders with the new score.
⊘ Archive	Confirms, then moves the repo to `_archive/`.

Wishlist · the backlog

Same data the launcher uses — Unused, Built, AI-Curated counts, with quick previews. Click anywhere to open the side panel and add/manage ideas.

Log · session telemetry

Append-only feed of every start.sh session — mode, project name, timestamp.

Install

git clone https://github.com/rayancheca/daily-builder ~/daily-builder
cd ~/daily-builder

# Python venv (only needed for pytest; main scripts are stdlib only)
python3 -m venv venv
venv/bin/pip install pytest

# Verify
venv/bin/python -m pytest tests/ -q

Requires:

macOS or Linux with bash, Python 3.11+, git
Claude Code CLI authenticated
GitHub CLI for repo creation
Optional: osascript (macOS) for the dashboard's Resume-in-Terminal buttons

Other launcher modes

bash start.sh --polish <repo>     # dedicated finishing-pass session
bash start.sh --evaluate <repo>   # just score it
bash start.sh --resume <repo>     # resume a specific project
bash start.sh --mode guided       # skip the mode prompt

Architecture

~/daily-builder/
├── start.sh                    # the only script you run
├── config.json                 # all thresholds: stalled days, dedupe cutoff, etc.
├── wishlist.md                 # your idea backlog (Unused / Used)
├── curated.md                  # auto-generated AI alternates
├── project_history.md          # append-only log + eval scores
├── lib/
│   ├── paths.py                # typed Config loader
│   ├── project_state.py        # git-derived authoritative status
│   ├── pick_domain.py          # LRU + weighted domain picker
│   ├── dedupe.py               # stdlib TF-IDF similarity
│   ├── wishlist.py             # parser + add/mark_built + curated regen
│   ├── telemetry.py            # Claude Code transcript parser
│   └── evaluate.py             # heuristic + LLM evaluator
├── prompts/
│   ├── generate_idea.md        # domain-aware idea prompt
│   ├── evaluate.md             # LLM evaluation rubric
│   ├── finishing_pass.md       # polish-only session brief
│   ├── new_project.md          # full build spec
│   ├── dashboard_polish.md     # separate-session dashboard audit brief
│   └── rules/                  # modular agent rules
├── dashboard/
│   ├── server.py               # Python stdlib HTTP server + SSE
│   ├── index.html
│   ├── app.js                  # Chart.js + vanilla JS
│   └── style.css               # OKLCH tokens, no framework
├── docs/
│   └── screenshots/            # README assets (regenerated via Playwright)
├── portfolio-finish/           # workflow for finishing/shipping existing repos
└── tests/                      # pytest, all stdlib targets

Data sources — single source of truth

Question	Source
Is a project shipped?	git commit history + `state.md` cross-check
How much progress?	git feat/fix commit count vs. numbered plan in `CLAUDE.md`
Is a project stalled?	`days_since_last_commit >= stalled_days` (config)
Token usage?	`~/.claude/projects/*/.jsonl` message-usage fields
Evaluation score?	8 heuristic checks + LLM qualitative pass
Idea too similar?	TF-IDF cosine against `project_history.md` entries

state.md is trusted for display labels only, never for counts or gating.

Configuration

All knobs in config.json:

{
  "abandonment": { "stalled_days": 3, "dead_days": 14 },
  "dedupe":      { "similarity_threshold": 0.55, "max_regens": 3 },
  "evaluator":   { "score_threshold": 75, "use_llm": true },
  "domain": {
    "weights": {
      "cybersecurity": 1.0, "systems": 1.0, "data": 1.0,
      "ai_ml": 1.0, "networking": 1.0, "devtools": 1.0
    }
  },
  "telemetry": {
    "transcripts_dir": "~/.claude/projects",
    "max_weekly_tokens": 10000000000,
    "max_session_tokens": 1000000000
  },
  "finishing_pass": { "auto_suggest_below_score": 75 }
}

Set a domain weight to 0 to exclude it. Raise another to bias toward it.

Running the dashboard remotely

The dashboard is a local-only HTTP server that reads ~/dev/daily-projects/ and ~/.claude/projects/ directly off your filesystem. It cannot be deployed to Vercel, Netlify, or any serverless platform without a full rewrite.

To access it from another device, use a secure tunnel — the dashboard keeps running on your desktop, the tunnel exposes it over a temporary public URL.

Cloudflare Tunnel (recommended — free, persistent)

brew install cloudflared
cloudflared tunnel --url http://localhost:8765

Prints a https://something-random.trycloudflare.com URL. Works anywhere, TLS-terminated by Cloudflare. Kill with Ctrl+C.

ngrok (quick one-off)

brew install ngrok
ngrok http 8765

Tailscale (cleanest for personal use)

brew install --cask tailscale
# Install on every device, then:
# http://<your-mac-name>:8765

No tunnel latency, no random URLs, devices reach each other over a private mesh.

Testing

venv/bin/python -m pytest tests/ -v

27 tests across 5 modules — project_state, pick_domain, dedupe, evaluate, telemetry. All stdlib-only assertions, no mocks beyond tmp git repos and synthetic transcripts.

Regenerating the screenshots

The README is documented as a workflow demo, so the screenshots are part of the repo. Regenerate them whenever the dashboard changes:

# 1. Boot the dashboard
python3 ~/daily-builder/dashboard/server.py &

# 2. Run the capture script (uses Playwright)
cd /tmp && npm i playwright
node ~/daily-builder/scripts/capture-screenshots.mjs

Outputs land in docs/screenshots/.

License

MIT.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

daily-builder

Why this exists

The full workflow

1 · One command, zero ceremony

2 · Pick a mode

3 · Wishlist mode — pick, add, or curate

4 · What happens when a project is built

The dashboard

Overview · live mission status

Activity · cadence + token burn

Projects · the portfolio

Wishlist · the backlog

Log · session telemetry

Install

Other launcher modes

Architecture

Data sources — single source of truth

Configuration

Running the dashboard remotely

Cloudflare Tunnel (recommended — free, persistent)

ngrok (quick one-off)

Tailscale (cleanest for personal use)

Testing

Regenerating the screenshots

License

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Name		Name	Last commit message	Last commit date
Latest commit History 19 Commits
dashboard		dashboard
docs/screenshots		docs/screenshots
lib		lib
portfolio-finish		portfolio-finish
prompts		prompts
scripts		scripts
tests		tests
.gitignore		.gitignore
README.md		README.md
config.json		config.json
curated.md		curated.md
dashboard.sh		dashboard.sh
project_history.md		project_history.md
reference.md		reference.md
resume_prompt.md		resume_prompt.md
start.sh		start.sh
wishlist.md		wishlist.md

Folders and files

Latest commit

History

Repository files navigation

daily-builder

Why this exists

The full workflow

1 · One command, zero ceremony

2 · Pick a mode

3 · Wishlist mode — pick, add, or curate

4 · What happens when a project is built

The dashboard

Overview · live mission status

Activity · cadence + token burn

Projects · the portfolio

Wishlist · the backlog

Log · session telemetry

Install

Other launcher modes

Architecture

Data sources — single source of truth

Configuration

Running the dashboard remotely

Cloudflare Tunnel (recommended — free, persistent)

ngrok (quick one-off)

Tailscale (cleanest for personal use)

Testing

Regenerating the screenshots

License

About

Topics

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages