no-mistakes Lite (nml)

nml is a lightweight local workflow driver inspired by no-mistakes. It inspects a git repository, records original user intent, prepares committed changes for isolated validation, and is being built toward review, test, lint, PR, CI, optional automerge, and deploy steps.

Inspiration and key differences

no-mistakes gates changes by putting a local git proxy in front of your real remote: push to the gate, let the AI validation pipeline run in a disposable worktree, then forward only clean work to the configured push target and PR flow.

nml explores a smaller command-first take on the same idea:

• No git proxy remote required - run nml run, nml status, nml respond, and nml resume directly from the repository.
• Agent-friendly by default - headless commands emit compact TOON on stdout, progress on stderr, and avoid prompts unless an explicit interactive command is used.
• Lightweight local state - run state lives under .git/nml/runs and is mirrored to ~/.nml/runs for resume and logs.
• Explicit worktree isolation - Treehouse leases provide isolated validation worktrees, with current worktree reuse when already inside a Treehouse-managed checkout.
• Incremental scope - nml focuses on a small binary, simple config, direct GitHub PR/CI automation, and hooks or skills that can be installed separately.

Use no-mistakes when you want the full git-push gate workflow. Use nml when you want a minimal CLI that agents and scripts can call directly.

Current commands

nml                                  # print compact workspace status in TOON
nml init --yes --agent <name>         # create ~/.config/nml/config.yaml
nml init --interactive                # guided setup wizard for humans
nml doctor                            # check local tools, auth, config, and repo state
nml run --message "fix: handle empty input"
nml run --test-command "go test ./..."
nml run --skip-review --test-command "go test ./..."
nml status --format toon
nml status --run <id> --full
nml findings --format toon
nml findings --run <id> --full
nml runs --resumable
nml resume --run <id>
nml respond --action fix --run <id>      # fix all latest review findings
nml respond --action approve
nml config --format toon
nml config --interactive
nml config --scope global --set agent.name=codex
nml config --scope project --set review.yolo=true --set ci.timeout=15m
nml config --scope project --set review.auto_approve_after_rounds=true
nml config --scope global --set auto_merge.enabled=true
nml config --scope project --set conflict_resolution.mode=rebase
nml config --scope project --set cleanup.auto=false
nml hooks install --apps claude,codex,opencode
nml tui                                 # timeline, or interactive review gate

Headless commands keep stdout structured and compact. Running nml with no arguments is content-first and never prompts. Progress goes to stderr. Exit codes are 0 for success or no-op, 1 for runtime errors, and 2 for usage errors.

Agent integrations

Install live session context hooks when you want agents to see nml workspace status at startup:

nml hooks install --apps claude,codex,opencode
nml hooks install --scope project --apps claude,codex

The repository also carries an Agent Skill at skills/no-mistakes-lite/SKILL.md. scripts/install.sh installs only the nml binary and does not install the skill automatically. If you want the skill, install or copy it explicitly. The hook and skill are complementary: use the hook for ambient live state, and the skill for on-demand workflow guidance.

Implemented first

• Go module and nml binary skeleton.
• Global and per-repo config loading from ~/.config/nml, including persistent run defaults such as yolo review fixing, auto-approve after review rounds, PR conflict repair mode, auto-merge, auto-cleanup, CI timeout, and validation commands.
• Git preflight classification and no-op exits.
• Non-interactive first-run setup with --yes, plus an explicit --interactive wizard for humans.
• Doctor command with TOON output.
• Dirty worktree staging, agent-generated commit metadata, and commit hook fix retries.
• Agent or fallback intent generation and JSON run state under .git/nml/runs, mirrored to ~/.nml/runs with event and log files for resume.
• Treehouse worktree leasing with completed-run auto-cleanup that preserves a worktree when the current terminal started inside it, current Treehouse worktree reuse, exact source commit reuse when already based on the current base, and cherry-pick or format-patch fallback.
• First review loop using configured agents, exact LGTM, Markdown finding parsing, review gate output, configured auto-approval after review rounds, --skip-review, and --yes or --yolo automatic fix attempts.
• Test command execution only when configured per repo or supplied per run, plus lint command execution with skip reasons and agent fix retries.
• Docs evaluation and optional agent documentation updates.
• GitHub remote detection, safe push of tool-owned nml/* branches, and GitHub PR create/update through gh when available.
• Bounded CI watch with delayed check registration handling, repeated PR merge-conflict detection with configurable merge or rebase repair before or after CI watches, failed-log collection, persisted CI logs, stricter no-checks handling for yolo, skipped-review, or unresolved-finding auto-approved runs, agent CI fix retries, optional per-run auto-merge with GitHub CLI prompts disabled, and optional deploy command retries.
• Resumable failed or interrupted runs via nml resume, review finding parser, secret redaction, PR body generation, Bubble Tea run timeline and interactive review gate response picker, AXI session hook installation for Claude Code, Codex, and OpenCode, installable Agent Skill, binary-only install script, usage docs, and unit tests.

See docs/usage.md for workflow details. Future work is mostly provider expansion.