flowchart LR
A[Planning advice] --> B[Implementation]
B --> C[Deterministic gates]
C --> D[Adversarial audit]
D --> E[Independent review]
E --> F[Bounded correction]
F --> C
E --> G[Ready for human review]
Dev Autopilot
Persistent, auditable orchestration for scientific software development
A fail-closed Python and SQLite workflow coordinating implementation, adversarial audit, independent review, and final human approval.
v0.1.1 · Alpha View repository
The problem
AI coding agents can produce useful work, but a scientific software project also needs durable state, explicit permissions, reproducible tests, independent review and a clear boundary around human authority.
Dev Autopilot turns those requirements into a workflow instead of relying on one long prompt or one agent grading its own work.
What persists
SQLite records:
- every legal state transition;
- append-only events and stop reasons;
- retry counts for each agent role;
- quota pauses and exact resume states;
- idempotency hashes and diff-keyed review caches;
- expiring exclusive run locks.
This allows a workflow to recover after process restarts without guessing where it left off.
Safety design
Version 0.1.1 includes:
- validated YAML jobs with explicit file, tree and glob permissions;
- scope, test, context-size and Git-metadata gates;
- fail-closed transitions;
- deterministic jitter and quota-reset support;
- bounded review/correction rounds;
- fake adapters for complete offline integration tests;
- a generic file-based bridge for real subscription agents.
Real execution requires an explicit network opt-in. Shell syntax inside test commands also requires an explicit opt-in because it expands the trust boundary.
Why it belongs in a scientific portfolio
The project is not a claim that automation removes scientific judgment. It demonstrates the opposite: agentic workflows become more useful when scientific scope, mechanical correctness, adversarial review and human authority are represented as different responsibilities.
Current boundary
Dev Autopilot is alpha software. It coordinates and audits work, but never stages, commits, pushes, tags or merges product code. A human must inspect the final diff and evidence.
Review the state machine, job schema, offline smoke run and TargetIntel example profile.