1) Do you not feel self-conscious or weird about calling this "EvanFlow"? Seems like a lot of people these days are naming their AI tools/skills/whatever after themselves which seems self-absorbed. Either that or they hope that if their thing takes off like OpenClaw did then they'll grab the fame that comes along with it.

2) Why does your TDD flow miss the refactor step of TDD?

On jtfrench's unanswered question about dumb zone evasion: context length is what drives the drift. Agents go off-track when a loop runs long enough that early design context falls out. Resetting at each RED-GREEN-REFACTOR boundary keeps cycles short enough to avoid it. The hard cap of 5 iterate rounds is the same instinct applied at the macro level.

We ran into the parallel integration seam problem building tonone, a 23-agent Claude Code plugin where each domain agent works in its own worktree and integration tests are the merge contract.

https://github.com/tonone-ai/tonone if curious.

> Several rules come from 2025-2026 industry research on agentic coding failure modes

What are some of the papers you read?

How are these separate steps?

TDD is how you execute, not something you tack on afterwards.

EvanFlow is a single TDD-driven loop. Say "let's evanflow this" and it walks brainstorm → plan → execute → tdd → iterate → STOP. Real checkpoints at design and plan approval. Never auto-commits, never auto-stages, never proposes integration - every git op is your call.

The three things that actually changed how I work:

1. Vertical-slice TDD. One failing test → minimal impl → next test. Watch each test fail before writing the impl that passes it. (Sounds obvious. Almost no agent does it by default. ~62% of LLM-generated test assertions are wrong per HumanEval research, so testing TDD discipline matters more than the impl discipline.)

2. Embedded grilling at decision points. Before locking a plan: what breaks if a user does X? What's the rollback? What's explicitly out of scope? Catches design flaws while they're still cheap.

3. Iterate-until-clean (hard cap of 5 rounds). Re-read the diff against dead code, naming, the deletion test, assertion correctness, and a Five Failure Modes pass (hallucinated actions, scope creep, cascading errors, context loss, tool misuse). For UI: screenshot via headless Chromium.

For bigger plans with 3+ independent units sharing types, it forks into a parallel coder/overseer orchestration. Integration tests at touchpoints ARE the cohesion contract.

Three install paths: Claude Code plugin marketplace, npx skills add, manual copy. MIT.