MobaDex

A React Native app that runs on iOS, Android and the web from one codebase, built and verified by a team of AI agents.

Screens

• Home — searchable roster in a two-column grid
• Explore — same roster filtered by primary attribute and role
• Superstar — hero of the day with stats, hype quote and lore
• Hero detail — attributes, combat stats and abilities with icons
• Settings — data sources, cache controls and app info

Screenshots

Captured on the iPhone 18 Pro simulator (iOS 27) and framed by scripts/screenshots.sh (booted simulator + yarn ios); the raw captures are in screenshots/ and the process is documented in docs/screenshots.md.

Built by an agent team

Changes to this app are made by a team of AI agents and validated by deterministic checks before a human sees them. A message to a Telegram bot ("make the hero name gold") comes back as verified screenshots of the running app on an isolated branch, ready for /pr.

flowchart LR
  A[Telegram / CLI] --> B[harness/run.py]
  B --> C[planner] --> D[implementer] --> E[test-writer]
  E --> F{gate: tsc · eslint · jest<br/>architecture · secrets}
  F --> G[reviewer] --> H[visual-qa<br/>simulator + OCR + vision]
  F -. fail .-> D
  G -. blocking .-> D
  H -. fail .-> D
  H --> I{harness re-verifies<br/>independently}
  I --> J[branch + Telegram report] --> K[PR → CI gate]

Results

Measured with the repo's own evals (docs/benchmarks.md):

	pass	cost	time
Agents on Claude Opus/Sonnet	16/16	$1.72	42 s/case
Agents on Claude Sonnet/Haiku	16/16	$0.43	35 s/case
Agents fully local (qwen3 8B + gemma4, M4 Pro)	9/16	$0	164 s/case
End to end, Claude	4/4 tasks	$3.71	2.4 min/task
End to end, hybrid (Claude judges, local builds, escalation)	4/4 tasks	$0.49	23 min/task

Watch it live

The agents, their routes and every run live in Neo4j; harness/graph.py serve redraws the graph while the agents work. Here a real hybrid run (51 min, replayed): the reviewer blocks the first change, the local implementer times out, the gate goes red, the fix escalates to Claude Haiku and the run passes (docs/graph.md).

The same run in Neo4j Browser: each call linked to the next (THEN) and to the model it ran on (ON); the implementer's route to qwen3-14b and its escalation to Claude Haiku:

And you can ask it in plain language: .mcp.json plugs Neo4j's official MCP server (read-only) into Claude Code, which writes the Cypher itself:

Which agents in the hybrid backend run on a model that scored below 100% in the evals, and which of them have an escalation route?

Two: implementer (qwen3-14b, 66.7%, escalates to Haiku) and visual-qa (gemma4, 66.7%, no escalation). Planner, reviewer and test-writer score 100%.

MATCH (a:Agent)-[r:ROUTED_TO {backend:'hybrid'}]->(m:Model) OPTIONAL MATCH (a)-[s:SCORED]->(m) OPTIONAL MATCH (a)-[e:ESCALATES_TO {backend:'hybrid'}]->(em:Model) RETURN … · Sonnet, 12 s

How it works

• 5 single-purpose agents (.claude/agents/): judges are read-only and builders can't edit the harness that judges them. The gate enforces both.
• Code decides "done", not the model. harness/gate.sh (types, lint, tests, Clean Architecture layers, secret scan) plus OCR on real simulator captures, re-run by the harness after the agent says it's finished.
• Plans as contracts. The planner declares the screens, the exact on-screen text and how much to verify. The harness checks all three.
• Self-correcting loop. The raw failure goes back to the agent, with capped retries and escalation to a stronger model when a cheap one gets stuck.
• Measured per agent and as a system. Deterministic graders including mutation testing, a roster fingerprint on every result, and compare showing what an added or changed agent bought and cost (docs/evals.md).
• Three backends, one definition. Claude Code, fully local (opencode + Ollama) or hybrid, with a route per agent picked by harness/eval.py matrix (docs/hybrid.md, docs/local-backend.md).

Guardrails

• Credentials never enter the repo or an agent's context: they're loaded by scripts/telegram.sh from outside the repo, and .claude/hooks/guard.py blocks any tool call that reads them or calls the Telegram API by hand.
• No destructive git, no skipping git hooks, and no git at all in unattended runs: the harness commits on an isolated agent/<run-id> branch.
• Per-agent least privilege, rendered for both backends and tested (harness/local/guardrails_test.py).
• The bot only answers its owner's private chat, drops the backlog on start-up and shell-quotes anything an agent wrote. Captions are sent with --form-string, so agent text can't make curl read a local file.
• CI runs the same gate with a read-only token and no AI in the loop: no repository secrets, no paid API. Agents run on Claude Code or on local models, never through a pay-per-call API key.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗