ftl-sdlc-loop

name	ftl-sdlc-loop
description	Run autonomous multi-agent SDLC pipeline — plan, implement, review, test, and ship code changes
argument-hint	[task description] [options]
allowed-tools	Bash(ftl-sdlc-loop *), Bash(uv run ftl-sdlc-loop *), Bash(uv run supervisor.py *), Read

You are running the autonomous multi-agent SDLC pipeline using the ftl-sdlc-loop CLI tool. This tool orchestrates multiple Claude agents (planner, implementer, reviewer, tester, user) to collaboratively build software with real feedback loops.

Why Use This Tool

Autonomous development with real feedback. Instead of a single agent writing code and hoping it works, this pipeline has separate agents plan, implement, review, test, and actually run the code — catching real issues through structured iteration.

Claude is the user. The key insight: the User agent actually runs the generated code, discovers real errors, and provides structured feedback. This creates genuine feedback loops, not simulated ones.

Structured verdicts and exit gates. Each agent emits APPROVED/NEEDS_CHANGES/TESTS_PASSED/TESTS_FAILED/SATISFIED/NEEDS_IMPROVEMENT verdicts. The supervisor validates these — if an agent says APPROVED but lists open issues, it overrides the verdict.

How to Run

First run ftl-sdlc-loop --help to see all available options and examples.

Try these in order until one works:

1. ftl-sdlc-loop $ARGUMENTS (if installed via uv/pip)
2. uv run ftl-sdlc-loop $ARGUMENTS (if in the ftl-sdlc-loop repo)
3. uv run supervisor.py $ARGUMENTS (direct script invocation)

Common Workflows

Simple task

ftl-sdlc-loop "write a function to check if a number is prime"
ftl-sdlc-loop --effort minimal "solve two-sum"

Work on an existing codebase

ftl-sdlc-loop --workspace myproject --init-from /path/to/repo "add caching layer"
ftl-sdlc-loop --workspace myproject --init-from [e-mail]:user/repo.git "fix auth bug"

GitHub issue workflow — fix issue, create PR, review

ftl-sdlc-loop --workspace issue-42 --init-from ~/git/repo --github-issue 42 --github-repo owner/repo --github-pr --code-review --effort moderate --no-questions --clean

GitLab issue workflow — fix issue, create MR

ftl-sdlc-loop --workspace issue-285 --init-from ~/git/repo.git --gitlab-remote [e-mail]:org/repo.git --gitlab-issue 285 --effort minimal
ftl-sdlc-loop --workspace issue-285 --gitlab-mr --push

Push or create PR after completion

ftl-sdlc-loop --workspace myproject --push          # Push to remote
ftl-sdlc-loop --workspace myproject --pr             # Create GitHub PR
ftl-sdlc-loop --workspace myproject --push --clean   # Push, strip SDLC artifacts

Feature branch development

ftl-sdlc-loop --workspace feature-x --init-from /path/to/repo --branch feature-x "implement feature"
ftl-sdlc-loop --workspace feature-x --branch feature-x --push

Plan review workflow

ftl-sdlc-loop --workspace myproject --plan-only "build feature X"    # Generate plan, exit
# (human reviews PLAN.md)
ftl-sdlc-loop --workspace myproject --plan workspaces/myproject/PLAN.md "build feature X"

Continue a previous run

ftl-sdlc-loop --workspace myproject --continue "fix the bug from last run"

Continuous batch mode

ftl-sdlc-loop --continuous                    # Process queue.txt
ftl-sdlc-loop --continuous --queue tasks.txt  # Custom queue file

With environment variables (secrets)

ftl-sdlc-loop --workspace myproject --env ~/.secrets/myproject.env "build API integration"

Effort Levels

• minimal — Fast (~5-15 min): planner, implementer, tester. Skips review and user stages.
• moderate — Balanced (~30-60 min): full pipeline with review. Default.
• maximum — Production (~2-3 hours): comprehensive testing, multiple iterations.

Key Concepts

• Workspace: Named directory under workspaces/ containing the git repo and all artifacts. Reuse a workspace name to continue working on the same codebase.
• --init-from: Clones a repo (local path or git URL) into the workspace. Only needed the first time.
• --clean: Strips SDLC artifacts (PLAN.md, REVIEW.md, etc.) before push/PR. Use for clean deliverables.
• --no-questions: Fully autonomous — no interactive prompts. Required for unattended runs.
• --code-review: Runs multi-model code review after PR creation. Requires --github-pr.
• FINAL_REPORT.md: Human-readable summary generated at the end of each run.

Converting Natural Language to Commands

When the user describes what they want, map it to flags:

• "fix issue 42 on my repo" → --workspace issue-42 --init-from <repo> --github-issue 42
• "create a PR when done" → add --github-pr --push
• "run a code review too" → add --code-review
• "quick/fast" → --effort minimal
• "thorough/production" → --effort maximum
• "no prompts" → --no-questions
• "clean up artifacts" → --clean

After Any Command

• The pipeline runs autonomously and may take minutes to hours depending on effort level
• When complete, read FINAL_REPORT.md in the workspace for the summary
• If the user wants to see what happened, check workspaces/<name>/entries/ for full agent outputs