SF

The evolution of Singularity Forge — now a real coding agent.

The original SF went viral as a prompt framework for Claude Code. It worked, but it was fighting the tool — injecting prompts through slash commands, hoping the LLM would follow instructions, with no actual control over context windows, sessions, or execution.

This version is different. SF is now a standalone CLI built on the Pi SDK, which gives it direct TypeScript access to the agent harness itself. That means SF can actually do what v1 could only ask the LLM to do: clear context between tasks, inject exactly the right files at dispatch time, manage git branches, track cost and tokens, detect stuck loops, recover from crashes, and auto-advance through an entire milestone without human intervention.

One command. Walk away. Come back to a built project with clean git history.

npm install -g sf-run@latest

SF now provisions a managed RTK binary on supported macOS, Linux, and Windows installs to compress shell-command output in bash, async_bash, bg_shell, and verification flows. SF forces RTK_TELEMETRY_DISABLED=1 for all managed invocations. Set SF_RTK_DISABLED=1 to disable the integration.

📋 NOTICE: New to Node on Mac? If you installed Node.js via Homebrew, you may be running a development release instead of LTS. Read this guide to pin Node 24 LTS and avoid compatibility issues.

---

What's New in v2.71

MCP Secure Env Collect

• Secure credential collection over MCP — the new secure_env_collect tool uses MCP form elicitation to collect secrets (API keys, tokens) from external clients without exposing values in tool output. Masks input in interactive mode.
• Hardened elicitation schema — MCP elicitation schema handling is stricter, with proper validation and fallback for providers that don't support forms.

MCP Reliability

• Stream ordering preserved — MCP tool output now renders in the correct order, fixing interleaved output in Claude Code and other MCP clients.
• isError flag propagation — workflow tool execution failures now correctly return isError: true, so MCP clients can distinguish success from failure.
• Multi-round discuss questions — new-project discuss phase supports multi-round questioning with structured question gates.

Model Selection Hardening

• Unconfigured models blocked — models without a configured provider are filtered from selection surfaces, preventing dispatch failures.
• Provider readiness required — saved default model selection now verifies the provider is ready before accepting it.
• Session override honored — /sf model selection persists as a session override across all dispatch phases.
• Minimal context guard — model override logic is skipped in minimal command contexts where it doesn't apply.

Auto-Mode Resilience

• Credential cooldown recovery — auto-mode survives transient 429 rate-limit responses with structured cooldown errors and a bounded retry budget.
• Fire-and-forget auto start — auto start is detached from active turns to prevent blocking.
• Scoped forensics — stuck-loop forensics are now scoped to auto sessions only, preventing false positives in interactive use.

TUI Improvements

• Overlay subscription fix — resolved overlay subscription lifecycle and Ctrl+Shift+P shortcut conflict.
• Improved overlays and shortcuts — SF overlays, keyboard shortcuts, and notification flows redesigned for consistency.
• Pinned output restored — pinned output bar displays above the editor during tool execution again.
• Turn completion cleanup — pinned latest output is cleared on turn completion, preventing stale output from persisting.
• Secure input masking — extension input values are masked in interactive mode when collecting secrets.

Provider Fixes

• Full OAuth login URLs — OAuth login URLs are now displayed in full instead of being truncated.
• MiniMax bearer auth — MiniMax Anthropic API requests use proper bearer authentication.
• Case-insensitive tool rendering — renderable tool matching is now case-insensitive, fixing missed tool output.
• Headless idle timeout — idle timeout is kept off during interactive tool execution in headless mode.

Reliability & Internals

• TOCTOU file locking — race conditions in event log and custom workflow graph file locking are fixed with proper atomic lock acquisition.
• State derive refactor — deriveStateFromDb god function extracted into composable, testable helpers.
• Windows portability — hardened cross-platform portability across runtime, tooling, and CI.
• Model routing transparency — dynamic routing is skipped for interactive dispatches; model changes are always shown in the banner.
• Capability-aware routing (ADR-004) — full implementation of capability scoring, before_model_select hook, and task metadata extraction.
• Multi-model provider strategy (ADR-005) — infrastructure for multi-provider model selection wired into live paths.
• Anti-fabrication guardrails — discuss prompts enforce turn-taking to prevent fabricated user responses.
• Milestone worktree cleanup — merged worktree cleanup uses the milestone branch instead of generic lookups.
• Tool cache control — cache_control breakpoints added to tool definitions for improved prompt caching.

See the full Changelog for details on every release.

Previous highlights (v2.70 and earlier)

• Full workflow over MCP (v2.68) — slice replanning, milestone management, slice completion, task completion, and core planning tools exposed over MCP
• Transport-gated MCP (v2.68) — workflow tool availability adapts to provider transport capabilities automatically
• Contextual tips system (v2.68) — TUI and web terminal surface contextual tips based on workflow state
• Ask user questions over MCP (v2.70) — interactive questions exposed via elicitation for external integrations
• Tiered Context Injection (M005) — relevance-scoped context with 65%+ token reduction
• Resilient transient error recovery — defers to Core RetryHandler and fixes cmdCtx race conditions
• Anthropic subscription routing — auto-routed through Claude Code CLI provider with proper display names
• 5-wave state machine hardening — critical data integrity fixes across atomic writes, event log reconciliation, session recovery
• Discussion gate enforcement — mechanical enforcement with fail-closed behavior
• Slice-level parallelism — dependency-aware parallel dispatch within a milestone
• Persistent notification panel — TUI overlay, widget, and web API for real-time notifications
• MCP server — 6 read-only project state tools for external integrations, auto-wrapup guard, and question dedup
• Ollama extension — first-class local LLM support via Ollama, with dynamic routing enabled by default
• Discord bot & daemon — dedicated daemon package, Discord bot, and headless text mode with tool calls
• Capability-aware model routing (ADR-004) — capability scoring, before_model_select hook, and task metadata extraction
• VS Code sidebar redesign — SCM provider, checkpoints, diagnostics panel, activity feed, workflow controls, session forking
• /sf parallel watch — native TUI overlay for real-time worker monitoring
• Codebase map — automatic codebase map injection for fresh agent contexts
• --resume flag — resume previous sessions from the CLI
• Concurrent invocation guard — prevents overlapping auto-mode runs
• VS Code integration — status bar, file decorations, bash terminal, session tree, conversation history, and code lens
• Skills overhaul — 30+ skill packs covering major frameworks, databases, and cloud platforms
• Single-writer state engine — disciplined state transitions with machine guards and TOCTOU hardening
• DB-backed planning tools — atomic SQLite tool calls for state transitions
• Declarative workflow engine — YAML workflows through auto-loop
• Doctor: worktree lifecycle checks — validates worktree health, detects orphans, consolidates cleanup

---

Documentation

Full documentation is in the docs/ directory:

User Guides

• Getting Started — install, first run, basic usage
• Autonomous Mode — autonomous execution deep-dive
• Configuration — all preferences, models, git, and hooks
• Custom Models — add custom providers (Ollama, vLLM, LM Studio, proxies)
• Token Optimization — profiles, context compression, complexity routing
• Cost Management — budgets, tracking, projections
• Git Strategy — worktree isolation, branching, merge behavior
• Parallel Orchestration — run multiple milestones simultaneously
• Working in Teams — unique IDs, shared artifacts
• Skills — bundled skills, discovery, custom authoring
• Commands Reference — all commands and keyboard shortcuts
• Troubleshooting — common issues, doctor, forensics, recovery
• Visualizer — workflow visualizer with stats and discussion status
• Remote Questions — route decisions to Slack or Discord when human input is needed
• Dynamic Model Routing — complexity-based model selection and budget pressure
• Web Interface — browser-based project management and real-time progress
• Migration from v1 — .planning → .sf migration
• Docker Sandbox — run SF autonomous mode in an isolated Docker container

Developer Docs

• Architecture — system design and dispatch pipeline
• CI/CD Pipeline — three-stage promotion pipeline (Dev → Test → Prod)
• Pipeline Simplification (ADR-003) — merged research into planning, mechanical completion
• VS Code Extension — chat participant, sidebar dashboard, RPC integration

---

What Changed From v1

The original SF was a collection of markdown prompts installed into ~/.claude/commands/. It relied entirely on the LLM reading those prompts and doing the right thing. That worked surprisingly well — but it had hard limits:

• No context control. The LLM accumulated garbage over a long session. Quality degraded.
• No real automation. "Auto mode" was the LLM calling itself in a loop, burning context on orchestration overhead.
• No crash recovery. If the session died mid-task, you started over.
• No observability. No cost tracking, no progress dashboard, no stuck detection.

SF v2 solves all of these because it's not a prompt framework anymore — it's a TypeScript application that controls the agent session.

	v1 (Prompt Framework)	v2 (Agent Application)
Runtime	Claude Code slash commands	Standalone CLI via Pi SDK
Context management	Hope the LLM doesn't fill up	Fresh session per task, programmatic
Auto mode	LLM self-loop	State machine reading .sf/ files
Crash recovery	None	Lock files + session forensics
Git strategy	LLM writes git commands	Worktree isolation, sequential commits, squash merge
Cost tracking	None	Per-unit token/cost ledger with dashboard
Stuck detection	None	Retry once, then stop with diagnostics
Timeout supervision	None	Soft/idle/hard timeouts with recovery steering
Context injection	"Read this file"	Pre-inlined into dispatch prompt
Roadmap reassessment	Manual	Automatic after each slice completes
Skill discovery	None	Auto-detect and install relevant skills during research
Verification	Manual	Automated verification commands with auto-fix retries
Reporting	None	Self-contained HTML reports with metrics and dep graphs
Parallel execution	None	Multi-worker parallel milestone orchestration

Migrating from v1

Note: Migration works best with a ROADMAP.md file for milestone structure. Without one, milestones are inferred from the phases/ directory.

If you have projects with .planning directories from the original Singularity Forge, you can migrate them to SF's .sf format:

# From within the project directory
/sf migrate

# Or specify a path
/sf migrate ~/projects/my-old-project

The migration tool:

• Parses your old PROJECT.md, ROADMAP.md, REQUIREMENTS.md, phase directories, plans, summaries, and research
• Maps phases → slices, plans → tasks, milestones → milestones
• Preserves completion state ([x] phases stay done, summaries carry over)
• Consolidates research files into the new structure
• Shows a preview before writing anything
• Optionally runs an agent-driven review of the output for quality assurance

Supports format variations including milestone-sectioned roadmaps with <details> blocks, bold phase entries, bullet-format requirements, decimal phase numbering, and duplicate phase numbers across milestones.

---

How It Works

SF structures work into a hierarchy:

Milestone  →  a shippable version (4-10 slices)
  Slice    →  one demoable vertical capability (1-7 tasks)
    Task   →  one context-window-sized unit of work

The iron rule: a task must fit in one context window. If it can't, it's two tasks.

The Loop

Each slice flows through phases automatically:

Plan (with integrated research) → Execute (per task) → Complete → Reassess Roadmap → Next Slice
                                                                                      ↓ (all slices done)
                                                                              Validate Milestone → Complete Milestone

Plan scouts the codebase, researches relevant docs, and decomposes the slice into tasks with must-haves (mechanically verifiable outcomes). Execute runs each task in a fresh context window with only the relevant files pre-loaded — then runs configured verification commands (lint, test, etc.) with auto-fix retries. Complete writes the summary, UAT script, marks the roadmap, and commits with meaningful messages derived from task summaries. Reassess checks if the roadmap still makes sense given what was learned. Validate Milestone runs a reconciliation gate after all slices complete — comparing roadmap success criteria against actual results before sealing the milestone.

/sf autonomous — The Main Event

This is what makes SF different. Run it, walk away, come back to built software.

/sf autonomous

Autonomous mode is a state machine driven by files on disk. It reads .sf/STATE.md, determines the next unit of work, creates a fresh agent session, injects a focused prompt with all relevant context pre-inlined, and lets the LLM execute. When the LLM finishes, autonomous mode reads disk state again and dispatches the next unit. /sf auto remains supported as a short alias.

What happens under the hood:

1. Fresh session per unit — Every task, every research phase, every planning step gets a clean 200k-token context window. No accumulated garbage. No "I'll be more concise now."

2. Context pre-loading — The dispatch prompt includes inlined task plans, slice plans, prior task summaries, dependency summaries, roadmap excerpts, and decisions register. The LLM starts with everything it needs instead of spending tool calls reading files.

3. Git isolation — When git.isolation is set to worktree or branch, each milestone runs on its own milestone/<MID> branch (in a worktree or in-place). All slice work commits sequentially — no branch switching, no merge conflicts. When the milestone completes, it's squash-merged to main as one clean commit. The default is none (work on the current branch), configurable via preferences.

4. Crash recovery — A lock file tracks the current unit. If the session dies, the next /sf autonomous reads the surviving session file, synthesizes a recovery briefing from every tool call that made it to disk, and resumes with full context. Parallel orchestrator state is persisted to disk with PID liveness detection, so multi-worker sessions survive crashes too. In headless mode, crashes trigger automatic restart with exponential backoff (default 3 attempts).

5. Provider error recovery — Transient provider errors (rate limits, 500/503 server errors, overloaded) auto-resume after a delay. Permanent errors (auth, billing) pause for manual review. The model fallback chain retries transient network errors before switching models.

6. Stuck detection — A sliding-window detector identifies repeated dispatch patterns (including multi-unit cycles). On detection, it retries once with a deep diagnostic. If it fails again, auto mode stops with the exact file it expected.

7. Timeout supervision — Soft timeout warns the LLM to wrap up. Idle watchdog detects stalls. Hard timeout pauses autonomous mode. Recovery steering nudges the LLM to finish durable output before giving up.

8. Cost tracking — Every unit's token usage and cost is captured, broken down by phase, slice, and model. The dashboard shows running totals and projections. Budget ceilings can pause autonomous mode before overspending.

9. Adaptive replanning — After each slice completes, the roadmap is reassessed. If the work revealed new information that changes the plan, slices are reordered, added, or removed before continuing.

10. Verification enforcement — Configure shell commands (npm run lint, npm run test, etc.) that run automatically after task execution. Failures trigger auto-fix retries before advancing. Auto-discovered checks from package.json run in advisory mode — they log warnings but don't block on pre-existing errors. Configurable via verification_commands, verification_auto_fix, and verification_max_retries preferences.

11. Milestone validation — After all slices complete, a validate-milestone gate compares roadmap success criteria against actual results before sealing the milestone.

12. Escape hatch — Press Escape to pause. The conversation is preserved. Interact with the agent, inspect what happened, or just /sf autonomous to resume from disk state.

/sf and /sf next — Step Mode

By default, /sf runs in step mode: the same state machine as autonomous mode, but it pauses between units with a wizard showing what completed and what's next. You advance one step at a time, review the output, and continue when ready.

• No .sf/ directory → Start a new project. Discussion flow captures your vision, constraints, and preferences.
• Milestone exists, no roadmap → Discuss or research the milestone.
• Roadmap exists, slices pending → Plan the next slice, execute one task, or switch to auto.
• Mid-task → Resume from where you left off.

/sf next is an explicit alias for step mode. You can switch from step → auto mid-session via the wizard.

Step mode is the on-ramp. Auto mode is the highway.

---

Getting Started

Install

npm install -g sf-run

Log in to a provider

First, choose your LLM provider:

sf
/login

Select from 20+ providers — Anthropic, OpenAI, Google, OpenRouter, GitHub Copilot, and more. If you have a Claude Max or Copilot subscription, the OAuth flow handles everything. Otherwise, paste your API key when prompted.

SF auto-selects a default model after login. To switch models later:

/model

Use it

Open a terminal in your project and run:

sf

SF opens an interactive agent session. From there, you have two ways to work:

/sf — step mode. Type /sf and SF executes one unit of work at a time, pausing between each with a wizard showing what completed and what's next. Same state machine as autonomous mode, but you stay in the loop. No project yet? It starts the discussion flow. Roadmap exists? It plans or executes the next step.

/sf autonomous — autonomous mode. Type /sf autonomous and walk away. SF researches, plans, executes, verifies, commits, and advances through every slice until the milestone is complete. Fresh context window per task. No babysitting. /sf auto is an alias.

Two terminals, one project

The real workflow: run autonomous mode in one terminal, steer from another.

Terminal 1 — let it build

sf
/sf autonomous

Terminal 2 — steer while it works

sf
/sf discuss    # talk through architecture decisions
/sf status     # check progress
/sf queue      # queue the next milestone

Both terminals read and write the same .sf/ files on disk. Your decisions in terminal 2 are picked up automatically at the next phase boundary — no need to stop autonomous mode.

Headless mode — CI and scripts

sf headless runs any /sf command without a TUI. Designed for CI pipelines, cron jobs, and scripted automation.

# Run autonomous mode in CI
sf headless --timeout 600000

# Create and execute a milestone end-to-end
sf headless new-milestone --context spec.md --auto

# One unit at a time (cron-friendly)
sf headless next

# Instant JSON snapshot (no LLM, ~50ms)
sf headless query

# Force a specific pipeline phase
sf headless dispatch plan

Headless auto-responds to interactive prompts, detects completion, and exits with structured codes: 0 complete, 1 error/timeout, 2 blocked. Auto-restarts on crash with exponential backoff. Use sf headless query for instant, machine-readable state inspection — returns phase, next dispatch preview, and parallel worker costs as a single JSON object without spawning an LLM session. Pair with remote questions to route decisions to Slack or Discord when human input is needed.

Multi-session orchestration — headless mode supports file-based IPC in .sf/parallel/ for coordinating multiple SF workers across milestones. Build orchestrators that spawn, monitor, and budget-cap a fleet of SF workers.

First launch

On first run, SF launches a branded setup wizard that walks you through LLM provider selection (OAuth or API key), then optional tool API keys (Brave Search, Context7, Jina, Slack, Discord). Every step is skippable — press Enter to skip any. If you have an existing Pi installation, your provider credentials (LLM and tool keys) are imported automatically. Run sf config anytime to re-run the wizard.

Commands

Command	What it does
/sf	Step mode — executes one unit at a time, pauses between each
/sf next	Explicit step mode (same as bare /sf)
/sf autonomous	Autonomous mode — researches, plans, executes, commits, repeats
/sf auto	Alias for /sf autonomous
/sf quick	Execute a quick task with SF guarantees, skip planning overhead
/sf stop	Stop autonomous mode gracefully
/sf steer	Hard-steer plan documents during execution
/sf discuss	Discuss architecture and decisions (works alongside autonomous mode)
/sf rethink	Conversational project reorganization
/sf mcp	MCP server status and connectivity
/sf status	Progress dashboard
/sf queue	Queue future milestones (safe during autonomous mode)
/sf prefs	Model selection, timeouts, budget ceiling
/sf migrate	Migrate a v1 .planning directory to .sf format
/sf help	Categorized command reference for all SF subcommands
/sf mode	Switch workflow mode (solo/team) with coordinated defaults
/sf forensics	Full-access SF debugger for auto-mode failure investigation
/sf cleanup	Archive phase directories from completed milestones
/sf doctor	Runtime health checks — issues surface across widget, visualizer, and reports
/sf keys	API key manager — list, add, remove, test, rotate, doctor
/sf logs	Browse activity, debug, and metrics logs
/sf export --html	Generate HTML report for current or completed milestone
/worktree (/wt)	Git worktree lifecycle — create, switch, merge, remove
/voice	Toggle real-time speech-to-text (macOS, Linux)
/exit	Graceful shutdown — saves session state before exiting
/kill	Kill SF process immediately
/clear	Start a new session (alias for /new)
Ctrl+Alt+G	Toggle dashboard overlay
Ctrl+Alt+V	Toggle voice transcription
Ctrl+Alt+B	Show background shell processes
Alt+V	Paste clipboard image (macOS)
sf config	Re-run the setup wizard (LLM provider + tool keys)
sf update	Update SF to the latest version
sf headless [cmd]	Run /sf commands without TUI (CI, cron, scripts)
sf headless query	Instant JSON snapshot — state, next dispatch, costs (no LLM)
sf --continue (-c)	Resume the most recent session for the current directory
sf --worktree (-w)	Launch an isolated worktree session for the active milestone
sf sessions	Interactive session picker — browse and resume any saved session

---

What SF Manages For You

Context Engineering

Every dispatch is carefully constructed. The LLM never wastes tool calls on orientation.

Artifact	Purpose
PROJECT.md	Living doc — what the project is right now
DECISIONS.md	Append-only register of architectural decisions
KNOWLEDGE.md	Cross-session rules, patterns, and lessons learned
RUNTIME.md	Runtime context — API endpoints, env vars, services (v2.39)
STATE.md	Quick-glance dashboard — always read first
M001-ROADMAP.md	Milestone plan with slice checkboxes, risk levels, dependencies
M001-CONTEXT.md	User decisions from the discuss phase
M001-RESEARCH.md	Codebase and ecosystem research
S01-PLAN.md	Slice task decomposition with must-haves
T01-PLAN.md	Individual task plan with verification criteria
T01-SUMMARY.md	What happened — YAML frontmatter + narrative
S01-UAT.md	Human test script derived from slice outcomes

Git Strategy

Branch-per-slice with squash merge. Fully automated.

main:
  docs(M001/S04): workflow documentation and examples
  fix(M001/S03): bug fixes and doc corrections
  feat(M001/S02): API endpoints and middleware
  feat(M001/S01): data model and type system

sf/M001/S01 (deleted after merge):
  feat(S01/T03): file writer with round-trip fidelity
  feat(S01/T02): markdown parser for plan files
  feat(S01/T01): core types and interfaces

One squash commit per milestone on main (or whichever branch you started from). The worktree is torn down after merge. Git bisect works. Individual milestones are revertable. Commit messages are generated from task summaries — no more generic "complete task" messages.

Verification

Every task has must-haves — mechanically checkable outcomes:

• Truths — Observable behaviors ("User can sign up with email")
• Artifacts — Files that must exist with real implementation, not stubs
• Key Links — Imports and wiring between artifacts

The verification ladder: static checks → command execution → behavioral testing → human review (only when the agent genuinely can't verify itself).

Dashboard

Ctrl+Alt+G or /sf status opens a real-time overlay showing:

• Current milestone, slice, and task progress
• Auto mode elapsed time and phase
• Per-unit cost and token breakdown by phase, slice, and model
• Cost projections based on completed work
• Completed and in-progress units

HTML Reports

After a milestone completes, SF auto-generates a self-contained HTML report in .sf/reports/. Each report includes project summary, progress tree, slice dependency graph (SVG DAG), cost/token metrics with bar charts, execution timeline, changelog, and knowledge base sections. No external dependencies — all CSS and JS are inlined, printable to PDF from any browser.

An auto-generated index.html shows all reports with progression metrics across milestones.

• Automatic — generated after milestone completion (configurable via auto_report preference)
• Manual — run /sf export --html anytime

---

Configuration

Preferences

SF preferences live in ~/.sf/PREFERENCES.md (global) or .sf/PREFERENCES.md (project). Manage with /sf prefs.

---
version: 1
models:
  research: claude-sonnet-4-6
  planning:
    model: claude-opus-4-6
    fallbacks:
      - openrouter/z-ai/glm-5
      - openrouter/minimax/minimax-m2.5
  execution: claude-sonnet-4-6
  completion: claude-sonnet-4-6
skill_discovery: suggest
auto_supervisor:
  soft_timeout_minutes: 20
  idle_timeout_minutes: 10
  hard_timeout_minutes: 30
budget_ceiling: 50.00
unique_milestone_ids: true
verification_commands:
  - npm run lint
  - npm run test
auto_report: true
---

Key settings:

Setting	What it controls
models.*	Per-phase model selection — string for a single model, or {model, fallbacks} for automatic failover
skill_discovery	auto / suggest / off — how SF finds and applies skills
auto_supervisor.*	Timeout thresholds for autonomous mode supervision
budget_ceiling	USD ceiling — autonomous mode pauses when reached
uat_dispatch	Enable automatic UAT runs after slice completion
always_use_skills	Skills to always load when relevant
skill_rules	Situational rules for skill routing
skill_staleness_days	Skills unused for N days get deprioritized (default: 60, 0 = disabled)
unique_milestone_ids	Uses unique milestone names to avoid clashes when working in teams of people
git.isolation	none (default), worktree, or branch — enable worktree or branch isolation for milestone work
git.manage_gitignore	Set false to prevent SF from modifying .gitignore
verification_commands	Array of shell commands to run after task execution (e.g., ["npm run lint", "npm run test"])
verification_auto_fix	Auto-retry on verification failures (default: true)
verification_max_retries	Max retries for verification failures (default: 2)
phases.require_slice_discussion	Pause auto-mode before each slice for human discussion review
auto_report	Auto-generate HTML reports after milestone completion (default: true)

Agent Instructions

Place an AGENTS.md file in any directory to provide persistent behavioral guidance for that scope. Pi core loads AGENTS.md automatically (with CLAUDE.md as a fallback) at both user and project levels. Use these files for coding standards, architectural decisions, domain terminology, or workflow preferences.

Note: The legacy agent-instructions.md format (~/.sf/agent-instructions.md and .sf/agent-instructions.md) is deprecated and no longer loaded. Migrate any existing instructions to AGENTS.md or CLAUDE.md.

Debug Mode

Start SF with sf --debug to enable structured JSONL diagnostic logging. Debug logs capture dispatch decisions, state transitions, and timing data for troubleshooting auto-mode issues.

Token Optimization

SF includes a coordinated token optimization system that reduces usage by 40-60% on cost-sensitive workloads. Set a single preference to coordinate model selection, phase skipping, and context compression:

token_profile: budget      # or balanced (default), quality

Profile	Savings	What It Does
budget	40-60%	Cheap models, skip research/reassess, minimal context inlining
balanced	10-20%	Default models, skip slice research, standard context
quality	0%	All phases, all context, full model power

Complexity-based routing automatically classifies tasks as simple/standard/complex and routes to appropriate models. Simple docs tasks get Haiku; complex architectural work gets Opus. The classification is heuristic (sub-millisecond, no LLM calls) and learns from outcomes via a persistent routing history.

Budget pressure graduates model downgrading as you approach your budget ceiling — 50%, 75%, and 90% thresholds progressively shift work to cheaper tiers.

See the full Token Optimization Guide for details.

Bundled Tools

SF ships with 24 extensions, all loaded automatically:

Extension	What it provides
SF	Core workflow engine, autonomous mode, commands, dashboard
Browser Tools	Playwright-based browser with form intelligence, intent-ranked element finding, semantic actions, PDF export, session state persistence, network mocking, device emulation, structured extraction, visual diffing, region zoom, test code generation, and prompt injection detection
Search the Web	Brave Search, Tavily, or Jina page extraction
Google Search	Gemini-powered web search with AI-synthesized answers
Context7	Up-to-date library/framework documentation
Background Shell	Long-running process management with readiness detection
Async Jobs	Background bash commands with job tracking and cancellation
Subagent	Delegated tasks with isolated context windows
GitHub	Full-suite GitHub issues and PR management via /gh command
Mac Tools	macOS native app automation via Accessibility APIs
MCP Client	Native MCP server integration via @modelcontextprotocol/sdk
Voice	Real-time speech-to-text transcription (macOS, Linux — Ubuntu 22.04+)
Slash Commands	Custom command creation
Ask User Questions	Structured user input with single/multi-select
Secure Env Collect	Masked secret collection without manual .env editing
Remote Questions	Route decisions to Slack/Discord when human input is needed in headless/CI mode
Universal Config	Discover and import MCP servers and rules from other AI coding tools
AWS Auth	Automatic Bedrock credential refresh for AWS-hosted models
Ollama	First-class local LLM support via Ollama
Claude Code CLI	External provider extension for Claude Code CLI
cmux	Claude multiplexer integration — desktop notifications, sidebar metadata, visual subagent splits
GitHub Sync	Auto-sync milestones to GitHub Issues, PRs, and Milestones
LSP	Language Server Protocol — diagnostics, definitions, references, hover, rename
TTSR	Tool-triggered system rules — conditional context injection based on tool usage

Bundled Agents

Five specialized subagents for delegated work:

Agent	Role
Scout	Fast codebase recon — returns compressed context for handoff
Researcher	Web research — finds and synthesizes current information
Worker	General-purpose execution in an isolated context window
JavaScript Pro	JavaScript-specialized execution and debugging
TypeScript Pro	TypeScript-specialized execution and debugging

---

Working in teams

The best practice for working in teams is to ensure unique milestone names across all branches (by using unique_milestone_ids) and checking in the right .sf/ artifacts to share valuable context between teammates.

Suggested .gitignore setup

# ── SF: Runtime / Ephemeral (per-developer, per-session) ──────────────────
# Crash detection sentinel — PID lock, written per auto-mode session
.sf/auto.lock
# Auto-mode dispatch tracker — prevents re-running completed units (includes archived per-milestone files)
.sf/completed-units*.json
# State manifest — workflow state for recovery
.sf/state-manifest.json
# Derived state cache — regenerated from plan/roadmap files on disk
.sf/STATE.md
# Per-developer token/cost accumulator
.sf/metrics.json
# Raw JSONL session dumps — crash recovery forensics, auto-pruned
.sf/activity/
# Unit execution records — dispatch phase, timeouts, recovery tracking
.sf/runtime/
# Git worktree working copies
.sf/worktrees/
# Parallel orchestration IPC and worker status
.sf/parallel/
# SQLite database and WAL sidecars — checkpoint state, forensics data
.sf/sf.db*
# Daily-rotated event journal — structured event log for forensics
.sf/journal/
# Doctor run history — diagnostic check results
.sf/doctor-history.jsonl
# Workflow event log — structured event stream
.sf/event-log.jsonl
# Generated HTML reports (regenerable via /sf export --html)
.sf/reports/
# Session-specific interrupted-work markers
.sf/milestones/**/continue.md
.sf/milestones/**/*-CONTINUE.md

Unique Milestone Names

Create or amend your .sf/PREFERENCES.md file within the repo to include unique_milestone_ids: true e.g.

---
version: 1
unique_milestone_ids: true
---

With the above .gitignore set up, the .sf/PREFERENCES.md file is checked into the repo ensuring all teammates use unique milestone names to avoid collisions.

Milestone names will now be generated with a 6 char random string appended e.g. instead of M001 you'll get something like M001-ush8s3

Migrating an existing git ignored .sf/ folder

1. Ensure you are not in the middle of any milestones (clean state)
2. Update the .sf/ related entries in your .gitignore to follow the Suggested .gitignore setup section under Working in teams (ensure you are no longer blanket ignoring the whole .sf/ directory)
3. Update your .sf/PREFERENCES.md file within the repo as per section Unique Milestone Names
4. If you want to update all your existing milestones use this prompt in SF: I have turned on unique milestone ids, please update all old milestone ids to use this new format e.g. M001-abc123 where abc123 is a random 6 char lowercase alpha numeric string. Update all references in all .sf file contents, file names and directory names. Validate your work once done to ensure referential integrity.
5. Commit to git

---

Architecture

SF is a TypeScript application that embeds the Pi coding agent SDK.

sf (CLI binary)
  └─ loader.ts          Sets PI_PACKAGE_DIR, SF env vars, dynamic-imports cli.ts
      └─ cli.ts         Wires SDK managers, loads extensions, starts InteractiveMode
          ├─ headless.ts     Headless orchestrator (spawns RPC child, auto-responds, detects completion)
          ├─ onboarding.ts   First-run setup wizard (LLM provider + tool keys)
          ├─ wizard.ts       Env hydration from stored auth.json credentials
          ├─ app-paths.ts    ~/.sf/agent/, ~/.sf/sessions/, auth.json
          ├─ resource-loader.ts  Syncs bundled extensions + agents to ~/.sf/agent/
          └─ src/resources/
              ├─ extensions/sf/    Core SF extension (auto, state, commands, ...)
              ├─ extensions/...     21 supporting extensions
              ├─ agents/            scout, researcher, worker, javascript-pro, typescript-pro
              └─ SF-WORKFLOW.md    Manual bootstrap protocol

Key design decisions:

• pkg/ shim directory — PI_PACKAGE_DIR points here (not project root) to avoid Pi's theme resolution collision with our src/ directory. Contains only piConfig and theme assets.
• Two-file loader pattern — loader.ts sets all env vars with zero SDK imports, then dynamic-imports cli.ts which does static SDK imports. This ensures PI_PACKAGE_DIR is set before any SDK code evaluates.
• Always-overwrite sync — npm update -g takes effect immediately. Bundled extensions and agents are synced to ~/.sf/agent/ on every launch, not just first run.
• State lives on disk — .sf/ is the source of truth. Auto mode reads it, writes it, and advances based on what it finds. No in-memory state survives across sessions.

---

Requirements

• Node.js ≥ 24.0.0 (24 LTS recommended)
• An LLM provider — any of the 20+ supported providers (see Use Any Model)
• Git — initialized automatically if missing

Optional:

• Brave Search API key (web research)
• Tavily API key (web research — alternative to Brave)
• Google Gemini API key (web research via Gemini Search grounding)
• Context7 API key (library docs)
• Jina API key (page extraction)

---

Use Any Model

SF isn't locked to one provider. It runs on the Pi SDK, which supports 20+ model providers out of the box. Use different models for different phases — Opus for planning, Sonnet for execution, a fast model for research.

Built-in Providers

Anthropic, Anthropic (Vertex AI), OpenAI, Google (Gemini), OpenRouter, GitHub Copilot, Amazon Bedrock, Azure OpenAI, Google Vertex, Groq, Cerebras, Mistral, xAI, HuggingFace, Vercel AI Gateway, and more.

OAuth / Max Plans

If you have a Claude Max, Codex, or GitHub Copilot subscription, you can use those directly — Pi handles the OAuth flow. No API key needed.

⚠️ Important: Using OAuth tokens from subscription plans outside their native applications may violate the provider's Terms of Service. In particular:

• Google Gemini — Using Gemini CLI or Antigravity OAuth tokens in third-party tools has resulted in Google account suspensions. This affects your entire Google account, not just the Gemini service. Use a Gemini API key instead.
• Claude Max — Anthropic's ToS may not explicitly permit OAuth use outside Claude's own applications.
• GitHub Copilot — Usage outside GitHub's own tools may be restricted by your subscription terms.

SF supports API key authentication for all providers as the safe alternative. We strongly recommend using API keys over OAuth for Google Gemini.

OpenRouter

OpenRouter gives you access to hundreds of models through a single API key. Use it to run SF with Llama, DeepSeek, Qwen, or anything else OpenRouter supports.

Per-Phase Model Selection

In your preferences (/sf prefs), assign different models to different phases:

models:
  research: openrouter/deepseek/deepseek-r1
  planning:
    model: claude-opus-4-6
    fallbacks:
      - openrouter/z-ai/glm-5
  execution: claude-sonnet-4-6
  completion: claude-sonnet-4-6

Use expensive models where quality matters (planning, complex execution) and cheaper/faster models where speed matters (research, simple completions). Each phase accepts a simple model string or an object with model and fallbacks — if the primary model fails (provider outage, rate limit, credit exhaustion), SF automatically tries the next fallback. SF tracks cost per-model so you can see exactly where your budget goes.

---

Ecosystem

Project	Description
SF2 Config Utility	Standalone configuration tool for managing SF preferences, providers, and API keys

---

Star History

---

License

MIT License

---

The original SF showed what was possible. This version delivers it.

npm install -g sf-run && sf