feature-loop

feature-loop

Autonomous build → verify → fix loop for any repo, on top of the agent-skills plugin. Run /spec and /plan, then one command takes the plan to a green feature branch with no approval prompts — building, testing, running your CI gates, security-auditing, and simplifying, looping fixes back into the build until everything passes.

The engine is generic; everything project-specific lives in a small .featureloop file in the target repo.

What it does

/spec → SPEC.md     /plan → tasks/plan.md      (interactive, first)

feature-loop PROJ-123 add-rollup-metric
 ├ 1. worktree off origin/<base>  → feature/PROJ-123-add-rollup-metric
 ├ 2. /build         (the ONLY writer)  ← plan + tasks/failures/*.md
 ├ 3-5. parallel READ-ONLY gates:
 │      • /test                  → tasks/failures/test.md      (if gaps)
 │      • $FL_GATES (your CI)    → tasks/failures/pipeline.md  (if red)
 │      • security audit         → tasks/failures/security.md  (if any)
 │      any failures? ─► back to /build  (≤ FL_MAX_ITERS)
 ├ 6. /code-simplify → re-verify gates
 └ live state in <worktree>/tasks/STATUS.md

Gates 3–5 are read-only; /build is the only writer. That single-writer invariant is what makes the parallelism safe.

Install

As a Claude plugin (interactive /auto-feature):

claude plugin marketplace add ecukalla/feature-loop
claude plugin install feature-loop@ecukalla-plugins
# then, in any repo: /auto-feature PROJ-123 add-rollup-metric

Migrating from 0.1.x

The marketplace was renamed from feature-loop to ecukalla-plugins in 0.2.0, so the install grammar moved from feature-loop@feature-loop to feature-loop@ecukalla-plugins. An existing 0.1.x install keeps working, but to pick up the new marketplace name, remove the old one and re-add it:

claude plugin uninstall feature-loop
claude plugin marketplace remove feature-loop
claude plugin marketplace add ecukalla/feature-loop
claude plugin install feature-loop@ecukalla-plugins

Headless / CLI:

git clone https://github.com/ecukalla/feature-loop ~/tools/feature-loop
export PATH="$HOME/tools/feature-loop/bin:$PATH"
# then, in any repo (bring your own image):
ANTHROPIC_API_KEY=… feature-loop-docker PROJ-123 add-rollup-metric
ANTHROPIC_API_KEY=… feature-loop-docker --image python:3.14-slim PROJ-123 add-rollup-metric

# Use your Claude.ai / Max subscription instead of an API key:
feature-loop-docker --auth oauth PROJ-123 add-rollup-metric

Billing — changes 2026-06-15. feature-loop runs entirely on headless claude -p, which Anthropic counts as programmatic usage. From 2026-06-15, --auth oauth runs draw from your plan's separate monthly Agent SDK credit and then bill at full API rates (the ANTHROPIC_API_KEY path is unchanged). See docs/billing-and-auth.md.

Bring your own Docker. You name the base (an image, a Dockerfile, or --image); feature-loop wraps it with a cached overlay that injects only the Claude CLI, the agent-skills plugin, and the engine. There is no dependency on any image I publish — see docs/featureloop-config.md.

Update: claude plugin update feature-loop · git -C ~/tools/feature-loop pull.

Configure a repo

Drop a .featureloop in the repo root (see examples/.featureloop and docs/featureloop-config.md). Minimum:

FL_BASE_BRANCH=main
FL_GATES='make test'        # your authoritative CI gate

Projects needing extra toolchain point FL_DOCKERFILE (or FL_IMAGE) at a pure toolchain image — install tools to /usr/local so the non-root runtime user can use them. feature-loop adds Claude/plugin/engine on top automatically.

On an interactive terminal the engine shows colored ==> headers, a spinner, and a live in-place display of the parallel gates; piped/headless runs stay plain (zero escape codes). Set NO_COLOR to drop color or FL_NO_SPINNER=1 to drop the animation (see docs/featureloop-config.md).

Layout