CLI Builder

CLI Builder

Build production-quality CLI tools for any module or application, in any language.

Reference files (read each one on demand, not upfront, to keep the agent's context budget small):

• references/cli-libraries.md — read during Step 2 (Design) to recommend libraries and during Step 4 (Execute) for starter scaffolds
• references/testing-patterns.md — read during Step 4 (Execute) when writing tests
• references/final-report.md — read during Step 5 (Summarize) to write the final report

Repo Sync Before Edits (mandatory)

Run this section once, before Step 1, so the analysis reads the current code. If git rev-parse --git-dir fails, the directory is not a git repository: skip this section, the Branch-First Safety Rule, and the Step 4 commits, and list no git repository: no branch or commits under Uncertainty: in the final report.

In a git repository, sync the current branch with remote:

branch="$(git rev-parse --abbrev-ref HEAD)"
git fetch origin
git pull --rebase origin "$branch"

If the working tree is not clean, stash first, sync, then restore:

git stash push -u -m "pre-sync"
branch="$(git rev-parse --abbrev-ref HEAD)"
git fetch origin && git pull --rebase origin "$branch"
git stash pop

If origin is missing, pull is unavailable, or rebase/stash conflicts occur, stop and ask the user before continuing.

Branch-First Safety Rule

Run this rule at the start of Step 4, before the first file write, with CLI_NAME set to the tool name approved in Step 2. Only create a new branch if on main or master — otherwise continue on the existing branch (the user likely set it up already or is resuming work):

current_branch="$(git rev-parse --abbrev-ref HEAD)"
if [ "$current_branch" = "main" ] || [ "$current_branch" = "master" ]; then
  slug="$(echo "${CLI_NAME:-cli}" | tr '[:upper:] ' '[:lower:]-' | tr -cd 'a-z0-9-')"
  ts="$(date +%Y%m%d-%H%M%S)"
  git checkout -b "feat/cli-${slug}-${ts}"
fi

Mandatory 5-Step Workflow (approval-gated)

Explicit approval is a user reply that accepts the presented item as it stands, such as "approved" or "looks good, go ahead". A question, a change request, or no reply is not approval. Steps 1-3 write no files.

Step 1: Analyze

Understand the project before proposing anything.

Auto-detect language by checking for manifest files:

• package.json / tsconfig.json -> JavaScript/TypeScript
• pyproject.toml / setup.py / setup.cfg / requirements.txt -> Python
• go.mod -> Go
• Cargo.toml -> Rust
• pom.xml / build.gradle / build.gradle.kts -> Java/Kotlin
• Gemfile / *.gemspec -> Ruby

Identify existing CLI/entry points: check for bin fields, __main__.py, main.go, fn main(), existing arg parsing code, or scripts in package.json.

Record the module structure: list the public functions or classes the CLI will expose, their inputs and outputs, the core data types, and runtime dependencies.

Ask clarifying questions (only what cannot be inferred):

• Primary use case (automation, developer tool, data processing, admin)
• Target audience (developers, ops, end users)
• Single command or multi-command (subcommand tree)
• Output formats needed (text, JSON, table, CSV)
• Distribution method (pip/npm/go install, standalone binary, source)

Present the findings: detected language, entry points, the list of exposed functions, and the open questions. Wait for explicit approval before Step 2.

Step 2: Design

Present a structured CLI design document:

• Tool name and binary/entry point name
• Command tree (visual hierarchy for multi-command tools)
• Arguments and options per command (name, type, required/optional, default, help text)
• Global options (verbose, quiet, output format, config file, no-color)
• I/O behavior (stdin support, stdout/stderr separation, piping)
• Config strategy (CLI args > env vars > config file > defaults)
• Example invocations (at least 3 realistic examples showing common use cases)

Approval loop:

1. Present the design document.
2. Ask for feedback.
3. If the reply is not explicit approval, revise the design and return to 1.

No implementation before design approval. If the user ends the run without approving the design, stop; the final status is BLOCKED.

Step 3: Plan

Break implementation into three phases, each with granular tasks.

Phase 1 — Foundation (get a working CLI skeleton):

• Entry point and arg parsing setup
• One core command (the most important one)
• Help text and version flag
• Basic tests (help output, version, one command)

Phase 2 — Complete (full feature set):

• All remaining commands
• Input validation and error handling
• Output formatting (text, JSON, table as designed)
• Comprehensive tests

Phase 3 — Polish (optional; include it only when the user confirms it):

• Config file support
• Environment variable overrides
• Shell completions (bash, zsh, fish)
• Distribution/packaging setup (setup.py, package.json bin, goreleaser, etc.)

Each task includes:

• Goal: one sentence
• Files: create or modify
• Expected behavior: what the user can do after this task
• Test: how to verify
• Effort: S / M / L

Use the same approval loop as Step 2 for the plan.

No execution before plan approval. If the user ends the run without approving the plan, stop; the final status is BLOCKED.

Step 4: Execute

Before the first file write, apply the Branch-First Safety Rule. Then, for each task in the approved plan:

1. Implement the task.
2. Run the project's test command, taken from the manifest (for example pytest, npm test, go test ./..., cargo test).
3. If a test fails, fix the code and re-run the test command. After 3 failed fix attempts on the same task, stop Step 4, show the failing output, and ask the user how to proceed. Do not start the next task while a test fails.
4. If the test command cannot run (missing toolchain or test runner), tell the user. Continue only after explicit approval, and record that task's tests as not run.

At the end of each phase:

5. Run the demo (--help plus at least one approved example invocation) and show the output. A demo that exits non-zero is a failing test (item 3).
6. In a git repository, stage only the files this phase created or modified, then commit them with a descriptive message.

If a task needs a change to the approved design (a command, option, or output format differs), stop Step 4. Present the proposed change and wait for explicit approval before you continue.

Step 5: Summarize

Print the final report once, after the last Step Completion Report, as plain text in the chat. Do not write a report file unless the user asks for one. Read references/final-report.md for each part's contents, the status rules, and two examples. The four parts, in this order:

1. Result: — the status first (COMPLETE, PARTIAL — <reason>, or BLOCKED — <reason>), then the tool name.
2. Evidence: — only checks that ran: files, test counts, demo exit codes.
3. Uncertainty: — untested behavior and assumptions, labeled apart from verified facts.
4. Decision: — the approval needed, or No approval needed., then each remaining user action.

After the four parts, print the usage quick-start (install command and 3-5 example invocations) and the next steps (suggested improvements, missing features, distribution TODO).

Status rules — apply the first rule that matches:

1. BLOCKED — the run stopped before the first file write, for example because the design or plan was not approved, there was no module to wrap and the user gave no answer, or Repo Sync hit a conflict or a missing origin the user did not approve working around.
2. PARTIAL — at least one file was written, and then any of these happened: a test or demo still fails, a test command could not run, the user stopped Step 4 early, an approved task is not done, or a design change awaits approval.
3. COMPLETE — every task in the approved plan is done, and every test command and demo ran and passed. A Phase 3 the user declined does not make the run partial.