← Blog · June 21, 2026 · Olaf Freund

Program update: week of June 15–21, 2026

SpecKit-informed interop and the concurrency core shipped; Job-native defaults are converging — the honest fleet-wide status across PFactory, AIFactory, TFactory and CFactory

This is a program-level weekly update across the whole Factory fleet. Where the per-service posts go deep on one product, this one steps back: what shipped this week across PFactory, AIFactory, TFactory and CFactory, what is genuinely live on the cluster, and — just as important — what is still on its way and should not be claimed as done. Links to each service’s own post for the week are at the bottom.

Shipped this week

RFC-0015 — Spec-driven interop, end to end

The fleet now speaks the same spec-driven language that tools like GitHub’s spec-kit validate the thesis for. This week RFC-0015 went from schema foundation to a working end-to-end proof, with the epic (#183) and all ten of its children closed.

What that means in practice:

• A constitution — a declarative set of house rules — is ingested by PFactory and carried through the contract so AIFactory and TFactory consume the same standards the planner was held to.
• .specify-style spec ingest and emit, so a SpecKit project and the Factory pipeline can hand work back and forth without lossy translation.
• An adversarial spec-review pass that argues against a spec before it is approved, surfacing gaps the happy-path review would miss.
• Requirement → test traceability, so every acceptance criterion can be traced to the test that exercises it (or flagged as unproven).
• A declarative extension registry that lets the schema grow without code changes scattered across four repos.

This is the connective-tissue work paying off: one contract, one set of rules, read the same way at plan, code and verify time.

RFC-0016 — the concurrency core, proven live

For most of the program the PARR pipeline ran one task at a time. RFC-0016 changes that, and the core landed and was verified live this week (epic #188). The pieces:

• A stateless control plane — the web servers hold no per-run state, so they can be replicated.
• Durable job-state in Postgres — the source of truth for every in-flight run, survivable across restarts and replicas.
• Admission / concurrency caps wired to the cost budget from RFC-0014, so fan-out is bounded by money as well as by cluster capacity.
• The Nix-per-task Kubernetes Job substrate as the default for build and verify gates — each task gets a reproducible, isolated environment.
• A shared MinIO / S3 object store for artifacts so Jobs on different nodes can exchange inputs and outputs.
• KEDA queue-depth autoscaling.

The proof is the part that matters: on the factory cluster we ran 8 concurrent Job-per-task executions, and watched KEDA scale a worker deployment 1→3 under queue pressure. That is the substrate the fleet needed to stop being a single-lane road.

RFC-0017 — Job-native execution mechanisms

RFC-0017 is the follow-through on RFC-0016: take the Job-native substrate and make it the default path for build and verify across replicas. This week the mechanisms shipped (epic #206):

• Job-native log streaming, so a build or verify running inside a Kubernetes Job streams its output back to the portal the same way an in-pod run does.
• A Redis-backed multi-replica run-multiplexer (rmux) — and multi-replica is now running live, so more than one control-plane replica can serve and observe runs.
• Workspace pack/unpack via object storage (#207), the mechanism that lets a workspace produced on one node be consumed on another.

Read the next section before treating RFC-0017 as finished — it is not.

AWS demo resources cleaned up

The App Runner and EKS demo stacks from the recent deploy-then-verify and three-tier posts were torn down. There is zero ongoing cloud spend from the demos.

In progress — honestly not done yet

The temptation each week is to round “mechanisms merged” up to “shipped.” Here is the part that is still converging.

RFC-0017 default flips — NOT live

The Job-native build+verify default flips (#671 / #466) are not yet live. The credential-injection work they depend on has merged (AIFactory #688, TFactory #480), but the flips themselves are still on safe in-pod defaults and are converging through bug rounds:

• The build default flip is blocked on a /work-has-no-.git problem — the Job’s workspace arrives without the git metadata the build step expects.
• The verify default flip is pending re-validation after that build path is sound.

Until those are green, the truthful statement is: the mechanisms exist and multi-replica is live, but build and verify still run in-pod by default. We will flip them when they are proven, not before.

Stage E — multi-node workspace consumption

Workspace pack/unpack exists, but the full multi-node consumption path — a workspace packed by one Job and unpacked by another across nodes as the normal flow — is Stage E and still has code to land. It is the last structural piece between “concurrency core proven” and “fully job-native by default.”

Where the RFC ladder stands

• RFC-0014 (cost-aware routing) — Proposed. Referenced as a dependency by 0015 and 0016; not yet implemented on its own.
• RFC-0015 (spec-driven interop) — Implemented, end-to-end proof landed.
• RFC-0016 (horizontal & concurrent execution) — Core implemented and verified live; remaining default flips and multi-node workspace consumption tracked under RFC-0017.
• RFC-0017 (full Job-native default execution & scale-out) — In progress; mechanisms shipped and multi-replica live, default flips converging.

The program roadmap and the RFC index carry the same status, kept deliberately in sync with this post.

Per-service posts for the week

Each product has its own write-up for the week of June 15–21:

• PFactory — olafkfreund.github.io/PFactory/blog/
• AIFactory — olafkfreund.github.io/AIFactory/blog/
• TFactory — olafkfreund.github.io/TFactory/blog/
• CFactory — olafkfreund.github.io/CFactory/blog/

The bet has not changed: lead with governance, verification and observability, and make the substrate underneath strong enough to run many of them at once. This week moved the substrate a long way — and was honest about the last mile that remains.

← All posts