v0.9.0: the enterprise foundation
A merge gate, a no-AIFactory front door, multi-tenant groundwork, and the start of real polyglot — without diluting the one thing that matters: proving a generated test is worth keeping.
An LLM will write you a thousand tests. That number means nothing. The question that decides whether a test suite is an asset or a liability is the one almost no tool answers: is this generated test actually worth keeping — or is it a tautology that passes today and lies tomorrow?
That question is TFactory’s whole job, and it stays the centre of v0.9.0. What this release adds is everything around it that an enterprise needs before the answer can matter: a place to enforce it, a way in that doesn’t assume our sister tool, the start of tenancy, and a second real language.
The verdict, where merges happen
The moat was never generation — it’s the verdict pipeline: a generated test has to clear coverage-delta, 3× stability, a mutation-kill probe, and CI-parity before it counts.
Planner → Gen-Functional → Executor → Evaluator → Triager
v0.9.0 puts that verdict where decisions are actually made. The Triager can now
publish a red/green TFactory / tests status on the pull request, graded
against a quality_gate policy you set in .tfactory.yml — minimum accept-rate,
no surviving mutants, no tests that pass only by mocking out the thing they test.
It’s opt-in and dry-run-first; turn it on and “did the tests hold up” becomes a
merge gate in the workflow your team already runs.
A front door that doesn’t need AIFactory
TFactory grew up taking finished features from AIFactory. It no longer has to.
A new task_create_from_spec tool and a POST /api/specs/ingest endpoint take a
raw acceptance-criteria spec — markdown, Gherkin, or EARS — and turn it into
a native test-generation task. No branch, no handover.
The same seam fixed an inherited wart: creating a task from a GitHub issue in the portal used to mint an AIFactory coding task. Now it creates a native TFactory test task, through the ingest endpoint, where it belongs.
Multi-tenant, before it’s a rewrite
Enterprise means tenancy, and retrofitting it late is how you earn a quarter of
pain. So v0.9.0 lays the groundwork now: a migration of legacy project metadata
into the database, a ProjectStore abstraction (JSON by default, org-scoped DB
behind APP_PROJECTS_BACKEND), and a request→org resolver. We’ll be honest —
the cutover that actually enforces org isolation on every route is staged, not
flipped. But the seam is in, so “both deployment models” isn’t a future rewrite.
Polyglot, for real this time
Python and TypeScript were genuinely end-to-end. Java was a wedge. v0.9.0 wires JaCoCo into the Evaluator’s coverage signal with format-aware parsing, so the Java lane’s coverage-delta actually computes instead of erroring on the wrong parser. C# and Go are on the roadmap, not in the box — and we’ll say so.
Plumbing that unblocks adoption
Two smaller things that matter for real deployments: per-user acw_ API
tokens with an api:full scope, so the handover skill and CLI authenticate as
you instead of sharing one host-wide secret; and verified Ollama Cloud
support via the OpenAI-compatible path, for teams that want a capable model with
their own egress posture.
A forward bet: WebMCP
One thing we’re planning, not shipping. WebMCP
is an early W3C Community-Group draft — navigator.modelContext.registerTool() —
that lets a web page expose typed, callable tools to an in-browser agent. It’s
Chrome-Canary-only today, so treat this as a forward bet. But a page’s agent
tools are a new public surface nobody verifies, and that is squarely our lane:
bring the verdict pipeline to it, and expose TFactory’s own actions as WebMCP
tools so an agent can drive the portal from the open tab. Epic’s scoped,
default-off, experimental.
The honest ledger
Shipped (0.9.0, with a 0.9.1 hotfix and 0.9.2 in flight): the PR gate, generic ingestion, issue→native-task, tenant groundwork, Java coverage, per-user tokens, Ollama Cloud. Planned or experimental: the org-scoping cutover, C#/Go, WebMCP. We’d rather tell you which is which than blur the line.
Next: how TFactory decides a test is worth keeping.