flowchart TD
A[Bounded implementation] --> B[Scope and test checks]
B --> C[Audit and independent review]
C --> D{Acceptance criteria met?}
D -->|No| A
D -->|Yes| E[Verified bundle for human approval]
Dev Autopilot
AI development work accepted against explicit review evidence
A persistent Python and SQLite orchestrator with reviews tied to exact code changes, bounded correction and human release authority.
v0.9.1 · CLI preview View repository
The engineering problem
An AI agent completing a task is not enough evidence that the code is ready. A scientific software workflow needs a record of what changed, which checks passed, which findings remain unresolved and who can approve the result.
Dev Autopilot coordinates bounded implementation, deterministic checks, adversarial audit, independent review and correction. SQLite preserves progress across restarts. A verified review bundle brings the code changes and their evidence together for human approval.
Reviews follow the exact code changes
Each finding records an identifier, severity, evidence and a hash of the code diff. Changing the diff invalidates earlier conclusions. Open critical or high-severity findings, and stale findings, block acceptance until the required review is completed.
Implementation and review are separate workflow roles. Local actor labels do not authenticate separate people or prove that reviewers are statistically independent. Deterministic checks cannot be overridden by a favourable agent response.
What the 0.9.1 preview adds
- Durable execution: state transitions, retries, leases and quota pauses survive a process restart.
- Reviewable evidence: findings and review conclusions are attached to the exact diff; verified bundles include a human-readable report.
- Bounded code retrieval: searches over clean Git snapshots with context limits and source provenance.
- Budget controls: durable reservations and shared project caps prevent the same reservation from being executed twice.
- Execution boundaries: explicit permissions, minimal subprocess environments and an optional Docker adapter using a digest-pinned image.
- Operational inspection: health commands, progress reporting and a deterministic offline demonstration.
The original charter, milestone and correction workflow remains. A blocked run records the required action and resume condition instead of silently proceeding.
What has been demonstrated
The current repository records the following evidence. Software tests and synthetic demonstrations do not substitute for measured live-agent performance.
| Area | Evidence | Scope or remaining work |
|---|---|---|
| Linux compatibility | 418 tests passed; Python 3.11–3.13 CI | One opt-in Docker test skipped in the suite |
| Docker runtime | Separate live isolation and timeout-cleanup smoke passed | Requires a configured Linux host and pinned image |
| Restart durability | SQLite recovery tests and about two observed hours of soak testing | A full eight-hour run is not yet demonstrated |
| Colab | Drive mount, installation and queue startup verified | No completed job or interruption/recovery demonstration |
| Provider quality and cost | Evaluation harness and reporting implemented | Matched live-provider comparisons and measured billing pending |
See the repository validation record and capability/evidence matrix.
A practical demonstration
The packaged offline demo shows deterministic agents, explicit checks, saved state and a verifiable HTML/JSON review bundle. It makes recovery and acceptance rules inspectable without requiring paid model calls. It is a demonstration of workflow behaviour, not a benchmark of model quality.
The 0.9.1 preview and installation instructions support CLI use on Linux or WSL2. There is no graphical interface. Published downloads include checksums; those detect corruption and tampering but do not authenticate an unsigned bundle’s origin.