Pre-configured repo for running skills fleet demos against a real JavaScript codebase.
Slides: https://def1-20-124-190-109.ngrok-free.app/ (temporary ngrok link)
- Upstream: qiao/PathFinding.js — comprehensive pathfinding library for grid-based games
- Fork point: commit
2904a9a(latest upstream as of 2026年04月15日) - Patch applied:
should.jsupgraded from4.3.xto^13.2.3for Node 25 compatibility
The codebase is untouched upstream code — no fleet-generated modifications included.
node --version # or install via nvm curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.3/install.sh | bash nvm install 22
Fleets run workers in tmux sessions. Required.
- Ubuntu/Debian:
sudo apt install -y tmux - Fedora/RHEL:
sudo dnf install -y tmux - macOS:
brew install tmux
Verify: tmux -V
npm install -g @anthropic-ai/claude-code
Verify: claude --version
From inside this repo:
npx skills add quickcall-dev/skills
When prompted:
- Skills: select all (or pick specific fleets you want)
- Agents: select Claude Code
- Scope: Project
- Method: Symlink
Verify skills loaded — start a new Claude Code session and check /dag-fleet appears in the skill list.
git clone <this-repo-url> && cd <repo-name> npm install
./dev.sh
Launches a tmux session (pathfinding-dev) with:
- server window: visual demo on
http://localhost:8080 - tests window: runs the test suite
Keep this running. Fleet workers need the dev server for visual testing.
Tests window should show all tests passing.
Four ready-to-launch fleets in docs/experiments/002-live-demo/fleets/:
| Fleet | Type | Workers | What it does |
|---|---|---|---|
fleet-01-test-blitz |
dag | 9 | Coverage audit → distribute → parallel test writing → validation |
fleet-02-scenario-builder |
iterative | 6 | Build interactive scenario builder with reviewer loop |
fleet-03-algorithm-race |
dag | 3 | Benchmark A* vs Dijkstra on identical maps, compile leaderboard |
fleet-04-dijkstra-optimize |
autoresearch | 1 | Autonomous loop: optimize Dijkstra to close gap with A* |
Open Claude Code in a separate terminal (not inside the dev tmux):
cd <repo-root> claude
Then pick a fleet:
# Easiest — 3 workers, fast
/dag-fleet launch docs/experiments/002-live-demo/fleets/fleet-03-algorithm-race
# Medium — 9 workers, DAG dependencies
/dag-fleet launch docs/experiments/002-live-demo/fleets/fleet-01-test-blitz
# Advanced — iterative with reviewer loop
/iterative-fleet launch docs/experiments/002-live-demo/fleets/fleet-02-scenario-builder
# Autonomous — runs until stopped or budget exhausted
/autoresearch-fleet launch docs/experiments/002-live-demo/fleets/fleet-04-dijkstra-optimize
/<fleet-type> status # worker states, durations /<fleet-type> view # live tmux pane view /<fleet-type> report # summary report
graph TD
subgraph "fleet-01: test-blitz (dag)"
A1[coverage-auditor] --> O1[orchestrator]
A2[visual-auditor] --> O1
O1 --> T1[test-writer-1]
O1 --> T2[test-writer-2]
O1 --> T3[test-writer-3]
O1 --> T4[test-writer-4]
O1 --> S1[scenario-builder]
T1 --> V1[validator]
T2 --> V1
T3 --> V1
T4 --> V1
S1 --> V1
end
graph TD
subgraph "fleet-02: scenario-builder (iterative)"
C[canvas-worker] --> R[reviewer]
CT[controls-worker] --> R
SC[scorer-worker] --> R
P[persistence-worker] --> R
I[integration-worker] --> R
R -->|iterate| C
R -->|lgtm| DONE[done]
end
graph TD
subgraph "fleet-03: algorithm-race (dag)"
RA[racer-astar] --> L[leaderboard]
RD[racer-dijkstra] --> L
end
graph TD
subgraph "fleet-04: dijkstra-optimize (autoresearch)"
LOOP[edit → eval → keep/discard → repeat]
LOOP -->|plateau| WEB[web search]
WEB --> LOOP
end
- fleet-03 (algorithm-race) — simplest, 3 workers, finishes fast
- fleet-01 (test-blitz) — shows DAG dependencies in action
- fleet-02 (scenario-builder) — shows iterative reviewer loop
- fleet-04 (dijkstra-optimize) — autonomous research, runs indefinitely
This repo comes with a pre-initialized docs/ structure managed by the /doc skill. As you run fleets, use /doc to track plans, findings, checkpoints, and learnings — all organized under numbered experiments.
docs/experiments/002-live-demo/
├── .meta.json # auto-managed metadata
├── fleets/ # fleet definitions (ready to launch)
├── plans/ # numbered plans (01, 02, ...)
├── findings/ # discoveries during work
├── checkpoints/ # progress snapshots
├── research/ # experiment-scoped research
└── review/ # structured reviews
# Inside Claude Code: /doc plan 2 "test-strategy" /doc ckpt 2 "tests-passing" /doc finding 2 "perf-regression" /doc learn 2 "testing" "mock-pitfalls" /doc status 2 /doc list /doc resume 2
Stateless and parallel-safe — multiple fleet workers can write to the same experiment without conflicts.
├── src/ # pathfinding library source
│ ├── core/ # Grid, Node, Heap, Util
│ └── finders/ # A*, Dijkstra, BFS, JumpPoint, etc.
├── test/ # mocha test suite
├── visual/ # browser demo app
├── dev.sh # tmux dev environment launcher
└── docs/
└── experiments/
└── 002-live-demo/
└── fleets/
├── fleet-01-test-blitz/
├── fleet-02-scenario-builder/
├── fleet-03-algorithm-race/
└── fleet-04-dijkstra-optimize/
npm install npx mocha --require should test/**/*.js # run tests npx http-server visual -p 8080 -c-1 # start visual demo
| Problem | Fix |
|---|---|
skills not found |
Restart Claude Code after installing skills |
tmux: command not found |
Install tmux (see prerequisites) |
port 8080 in use |
fuser -k 8080/tcp (Linux) or lsof -ti:8080 | xargs kill (macOS) |
| Workers stuck on "waiting" | Check DAG — upstream workers must complete first |
| Tests fail on install | Make sure you're on Node 18+ (node --version) |
npx skills hangs |
Check internet connection, skills installs from GitHub |
Based on prior runs with Claude models:
| Fleet | Approximate cost |
|---|---|
| fleet-01 (test-blitz) | ~6ドル |
| fleet-02 (scenario-builder) | ~2ドル |
| fleet-03 (algorithm-race) | ~3ドル |
| fleet-04 (dijkstra-optimize) | ~5ドル-50 (depends on iterations) |
Costs vary by model choice and worker budget caps. See fleet.json for per-worker max_budget_usd settings.
PathFinding.js provides 10 pathfinding algorithms for 2D grids: A*, Dijkstra, BFS, BestFirst, JumpPoint, IDA*, and their bi-directional variants. See the online demo.
MIT License — Xueqiao Xu