Configuration

Configure loopflow via CLI flags, global config (~/.lf/config.yaml), or repo config (.lf/config.yaml). Precedence: CLI flags > repo config > global config.

Quick Reference

Context Assembly

Every step gets context assembled automatically. Run any command to see the breakdown:

Tokens: 12,847

files          6,892 █████
  STYLE.md     2,854 ██
  README.md      988 █
  scratch/     3,050 ██
diff_files     4,721 ███
  src/auth.py  2,847 ██
  tests/       1,874 █
clipboard      1,234 █

The token breakdown shows what’s included:

Defaults work well for most repos. Summaries require configuration.

Config Files

Global config (~/.lf/config.yaml) sets user-wide defaults. Repo config (.lf/config.yaml) overrides for that repo.

For most settings, repo overrides global. For additive settings (context, exclude, skill_sources, summaries), lists combine from both.

# ~/.lf/config.yaml (global)
agent_model: claude:opus
direction: product-engineer

# .lf/config.yaml (repo)
agent_model: codex        # overrides global
context:
  - docs/api.md           # combined with global context

Example repo config:

agent_model: claude:opus
push: true

direction: product-engineer

context:
  - src/schema.py
  - docs/api.md

exclude:
  - "*.test.ts"
  - node_modules

branch_names:
  schema: "{user}.{name}.{date}_{ts}"

Flows

Flows are stored as Python files in .lf/flows/:

def flow():
    return Flow(["implement", "rebase", "polish", "draft_commit"])

Options Reference

LF Docs (files)

roadmap/, scratch/, and root markdown files.

Includes README.md, STYLE.md, and similar guidance files. Also auto-includes scratch/ (current PR scratchpad) and roadmap/ (internal docs).

Branch Files (diff_files)

Full content of files modified on the current branch.

This is how the coding agent sees your changes. It gets complete files, not just diffs.

Summaries

Token-limited overviews of large directories. Configure paths and budget:

summary_tokens: 25000

summaries:
  - path: src
  - path: lib
    tokens: 5000

Generate with lf ops summarize. Summaries give the coding agent codebase context without consuming the full token budget. Only summarize directories you’re not including directly.

Clipboard

Paste content (errors, stack traces, context) into the prompt.

Use -c when debugging: copy an error, then lf debug -c.

Raw Diff

Include git diff main...HEAD output showing exact line changes.

Use when you want the agent to see precisely what changed. Can combine with --diff-files.

Context Files

Additional files always included in every step.

Config sets baseline files for all steps.

Exclude Patterns

Glob patterns to exclude from file listings.

Model

Default model for all steps.

Backends: claude, codex, gemini, opencode. Use backend:variant for specific models.

OpenCode passes model strings through to its own provider system:

agent_model: opencode                          # use opencode's default model
agent_model: opencode:anthropic/claude-sonnet  # explicit model
agent_model: opencode:openai/gpt-4o            # any provider opencode supports

Run Mode

Auto mode runs to completion. Interactive mode allows interruption and chat.

Set a step’s default mode in its frontmatter:

---
interactive: true
---
# Your step prompt here

CLI flags override the frontmatter default.

Direction

Directions shape judgment and intent—how the coding agent approaches work.

Direction files live in .lf/directions/ as markdown. Built-in directions: adapt, roadmap, ship, product-engineer, designer, infra-engineer, ceo.

Chrome

Enable browser automation for Claude Code.

Requires the Chrome extension and a paid Claude plan.

Push

Auto-push after commits in auto mode.

Yolo

Skip permissions. Codex uses the danger-full-access sandbox.

Claude: --dangerously-skip-permissions. Codex: --sandbox danger-full-access with approval_policy="never". Gemini: --yolo. OpenCode: permission: "allow" via OPENCODE_CONFIG_CONTENT.

Autoprune

Automatically remove worktrees when their PRs merge. Requires lfd daemon running.

With options:

autoprune:
  enabled: true
  poll_interval_seconds: 60  # default

IDE

Configure which tools open when launching sessions.

ide:
  warp: true           # open Warp terminal
  cursor: true         # open Cursor editor
  workspace: app.code-workspace  # optional workspace file

Branch Names

Customize branch naming for lf ops wt create.

branch_names:
  schema: "{user}.{name}.{date}_{ts}"

Available placeholders:

• {name} — short name passed to lf ops wt create
• {user} — git user.name (sanitized)
• {date} — YYYYMMDD
• {ts} — YYYYMMDD_HHMM

Summaries

Pre-generated codebase overviews for large repos.

summary_tokens: 25000

summaries:
  - path: src
  - path: lib
    tokens: 5000

Generate with lf ops summarize.

SkillRegistry

Enable the SkillRegistry directory as a remote skill source.

skill_registry:
  enabled: true
  prefix: sr
  cache_ttl_seconds: 86400

Run registry skills by prefix:

lf sr:gog

Registry skills are cached under ~/.lf/skills/skillregistry/ with a registry.json index.

Skill Sources

External skill libraries that extend loopflow with additional steps.

skill_sources:
  - name: superpowers
    prefix: sp
    path: ~/.superpowers

After configuring, run skills with their prefix:

lf sp:brainstorm              # run superpowers brainstorm skill
lf sp:write-plan -m codex     # with a different model

Auto-detection: If ~/.superpowers exists, it’s automatically registered with prefix sp even without explicit config.

See superpowers for an example skill library.