Status and roadmap

Where things stand

Priompt is working, pre-release software, currently v0.7, and all 22 end-to-end suites pass.

Version	Shipped
v0.1	gRPC server, Python adapter, validation
v0.2	L1/L2 caching (TTL), keyed on version_hash
v0.3	promptctl: commit/diff/log/promote/push/pull on go-git
v0.4	pub/sub over embedded NATS, TTL sync, subscriber model
v0.5	PostgreSQL, Redis L2, org-scoped multi-token auth
v0.6	token expiry/rotation, mTLS, Prometheus + audit log, per-org rate limits, backup/restore, migrations
v0.7	server-side versioning: commit DAG, branches, merges, history, commit diffs, pinning, rollback, verdict-carrying notifications

Before the first public release

• Final registry names: Docker image, PyPI distribution, npm package
• Public Go module paths (replace priomptproto, priomptdb, priomptauth, and drop the replace directives)
• License
• Prebuilt binaries for Linux, macOS, and Windows

Once the names are picked, tag v* and the existing release workflows publish to GHCR and PyPI.

Known gaps

• Merges take the source branch's content. Three-way auto-merge isn't implemented.
• There's no ListBranches RPC, so the web UI's branch, compare, search, and settings pages are still on mock data.
• Key rotation for encryption at rest isn't automated (re-put a prompt to re-encrypt it).
• Authorization works at org-prefix granularity. Per-prompt rules aren't modeled.

Proposed: the eval gate

Priompt already does CD. This proposal adds the CI.

The validation gate proves a prompt is well-formed, and the semantic diff estimates how much it changed. Neither answers "did it get worse?" The proposal:

• The server never calls a model. An external runner you own (promptfoo, deepeval, custom) evaluates the candidate branch and posts scores. The server stores them on the commit graph and gates the merge to main.
• Scores are compared with statistics, not a bare threshold. A candidate is compared against a baseline with an explicit noise band, and each verdict carries a reason code: ABOVE_BASELINE, WITHIN_NOISE, REGRESSION_EXCEEDS_TOLERANCE, AMBIGUOUS, MODEL_CHANGED, and so on. A changed model, dataset, or grader is flagged instead of being averaged into a pass.
• The semantic diff schedules the evals. A minor edit runs a cheap tier and a structural change runs the full suite.
• It reuses what Priompt already has. Invisible branches give zero-exposure candidates. Content-addressed commits guarantee that what was measured is exactly what ships. NATS triggers the runner.

A server with no eval policy configured behaves exactly as it does today.