Endless Dev Kit

01 / The closed loop

Endless Dev Kit

Loop engineered. Wax sealed. Human approved.

  • AI can write the fix. That part's solved.
  • The real question: is it real — or is the AI grading its own homework?
  • Proof you can read. Proof the AI can't fake.
loop engineeringthe closed loop
Scroll to follow the work ↓

02 / The Workbench

The Workbench

Nothing starts until you're clear to work.

  • Green light first — identity, repo, tracker, tools, all live.
  • Shape the work: file a bug, pull one in, or flesh out a hunch.
  • Checked against the real code until the plan holds.
  • Leaves as a Fix Plan (something broke) or Build Plan (something new).
  • Built for Linear. Jira, Shortcut, GitHub connect too.
preflight · intake · plan · confirmbuilt for Linear

03 / The Mailman

The Mailman

Ready work gets packaged — not approved.

  • The brain between planning and building.
  • Reads the board. Takes only what's ready.
  • Bundles related work into one clean package.
  • He packages — he does not seal. (Remember that word.)
  • Drops it in the mailbox and moves on.
composepackaging

04 / The Mailbox

The Mailbox

Packages wait to be picked up.

  • A queue, on purpose — nothing pushes itself at your code.
  • Work waits a minute or a week. Never expires, never drifts.
  • Your reviewers set the pace — not the AI's speed.
ready queue

05 / The Mayor

The Mayor

The AI never gets ahead of the human.

  • Checks the mailbox on a pulse — or the second you say "now."
  • Opens several lanes at once, each its own crew of agents.
  • Lanes are capped to what you can actually review and merge.
  • Three lanes, three reviewers. One finishes, the next moves up.
orchestrateWIP cap

06 / The Floor

The Floor

Here's the question every engineer asks.

  • The shop floor — where the code actually gets written.
  • Safe work runs in fast batches. Risky work gets watched.
  • You approved the scope. That's all you signed.
  • So what stops the AI from editing the test till it passes?
  • If it writes the fix and owns the scoreboard, "all green" means nothing.
work-batchwork-issue

07 / The Wax Seal

The Wax Seal

We don't ask the AI not to cheat. We make cheating impossible.

  • Before any fix, we write a test that fails — proof the bug's real.
  • Then we press a wax seal on that test: a frozen fingerprint.
  • Fix comes back green? We check the seal.
  • Seal intact → the only way to green was the real fix.
  • Seal broken → the test was tampered with. Fix tossed. No override.
pinned failing testfrozen fingerprint

08 / The Evidence

The Evidence

One proof record per branch — then a second opinion you can't buy.

  • Every branch builds one evidence bundle before a PR can open.
  • Red on old code → green with the fix → seal intact → suite passes → lint, build clean.
  • Anything short of all-green — no PR opens.
  • A fresh AI on a different model reads the change and the proof.
  • It can block. It can't merge, and it can't touch a line.
evidence bundleindependent review gate

09 / The Human Decision

The Human Decision

There is no merge button for the machine.

  • Confirmed, fixed, proven, reviewed — and still can't ship.
  • The engine can't merge. Not a setting — it was never built in.
  • A person reads the proof and makes one merge call, per PR.
human merge

10 / Refine & Learn

Refine & Learn

The loop improves the product — and itself.

  • Visual-only tweaks ship on their own — same proof, same human merge.
  • A short retro looks back — honest about the run.
  • Real, repeatable failures route through skill-forge to sharpen the engine.
  • Then it starts again. Endless, by design.
polishretroskill-forge

The loop