Plan in the open
The goal becomes a visible task graph with risk, dependencies, and expected verification before edits start.
Autonomous development, kept accountable
Phonton gives coding agents a local control plane: plan the work, route focused repo context, verify the output, and hand back evidence a developer can trust.
$ phonton plan "add rate limiting to auth"
scope middleware, config, tests
context 14 source handles
provider gemini / budget locked
verify cargo test -p auth
$ phonton review latest
status verified
diff ready for approval
Product loop
Phonton is shaped around the moments that make AI coding trustworthy: scope, context, execution, verification, and review.
The goal becomes a visible task graph with risk, dependencies, and expected verification before edits start.
Source handles, repo memory, and semantic context keep model calls focused on the actual change.
Work routes through the developer machine with provider choice, sandbox boundaries, and inspectable state.
Checks and review payloads decide whether the output is ready, failed, or needs human direction.
Trust boundary
Phonton keeps the agent loop inspectable by making memory, provider routing, verification, and review artifacts explicit instead of hiding them behind a hosted proxy.
Configuration, memory, task history, and review payloads stay on disk.
Use OpenAI, Anthropic, Gemini, OpenRouter, Ollama, or compatible endpoints.
Plans, retries, approvals, rejections, and rollback context remain explicit.
Benchmarks report provider, commit, checks, spend, and review burden.
CLI first
Start with `phonton doctor`, confirm the local environment, preview a plan, then let verification decide what is ready for review.
Audience
Keep the terminal workflow and gain an agent loop that produces inspectable plans and diffs.
Fit autonomous coding into existing trust, CI, secrets, and repository policy.
Let agents help while public stewardship still depends on readable evidence.
Next step