Skip to content

Kiln v2.0.0 — The Native-Workflow Rebuild

Choose a tag to compare

View all tags
@Fredasterehub Fredasterehub released this 30 May 01:24
v2.0.0
5c56a8b

"I am not an oven."

Kiln v2.0.0 is a ground-up rebuild. The old engine — persistent agent teams, ordered inter-agent messaging, and a wall of PreToolUse hooks — is gone. The autonomous work now runs as native Claude Code Dynamic Workflows: deterministic orchestration that keeps your session lean and never sits idle. Less machinery. More result.

One command still turns a conversation into working, tested software.

What changed

  • Native Dynamic Workflows replace the runtime. Research, architecture, build, validate, and report each run as a deterministic orchestration script — fan agents out, gather structured results, decide the next move in code. No daemon, no event bus, no idle stalls.
  • Brainstorm is an interactive teammate. The heavy creative dialogue lives in Da Vinci's context, not yours. You leave with a VISION.md.
  • Minds are files. Architecture, decisions, codebase state, and design tokens live as documents every worker reads and the conductor consults on demand — not as live processes.
  • Multi-model by design. Opus 4.8 reasons, plans, and reviews; GPT-5 (via Codex CLI) writes code; a Claude/Sonnet-only fallback runs when Codex isn't installed. Builder and reviewer are always different model families.
  • Essentially zero hooks. The guardrails that used to be runtime shell scripts are now properties of the workflow graph itself.

The build engine

The crown jewel of v2:

  • JIT vertical slices. A slicer scopes one independently-runnable behavior at a time from the live codebase — a multi-command CLI yields a slice per command (add, list, done, delete); a CRUD resource, a slice per operation.
  • Cross-family review on every cut. The builder and reviewer are never the same model family. The reviewer re-runs the tests itself, up to three fix cycles, before anything is accepted.
  • A milestone tribunal — when it's earned. Once a milestone's slices integrate, Ken (Opus) and Ryu (GPT-5) adversarially review the whole, independently; a deterministic reconcile dedupes and severity-ranks findings (a blocking gate the judge cannot soften); Judge Dredd delivers a binary verdict. For a single-slice milestone the tribunal is skipped — the cross-family slice review already is the milestone QA.
  • Validate tests what a person would do. Argus installs and runs the app, exercising every acceptance criterion the way a user would. Verdict: PASS / PARTIAL / FAILED, with up to three correction cycles back into Build.

Install

claude plugin marketplace add Fredasterehub/kiln
claude plugin install kiln

Then /kiln-doctor to pre-flight, and /kiln-fire to launch (or resume). Pass an intent to skip the cold open:

/kiln-fire build me a habit-tracker PWA

Requirements: Claude Code ≥ 2.1.154 with Dynamic Workflows; git, node. Optional: Codex CLI (the GPT-5 build path), Playwright MCP (browser validation). Kiln runs end-to-end on Claude alone.

Honest status

This is a freshly forged .0. It has been proven end to end — a full conductor run took an idea to a working, tested CLI app at VALIDATE_PASS, with the slicer producing per-behavior slices and the tribunal firing on the multi-slice milestone — but expect some edges still cooling. Functional and evolving; you may occasionally need to steer where it would normally carry you.

Perfection is achieved, not when there is nothing more to add, but when there is nothing left to take away. — Antoine de Saint-Exupéry

🤖 Generated with Claude Code

Assets 2
Loading