Autofab CLI Reference & Invocation Guide

Autofab is a literate, specification-driven software engineering framework for autonomous AI pair programming, double-blind verification, and deterministic DAG convergence.

All operations in the repository are invoked via uv run autofab <command>.

---

1. Quick Start

Autonomous Pair-Programming (clean)

autofab clean is the primary entry point for converging software components. It evaluates dirty nodes across your project's dependency DAG, commissions isolated double-blind workspaces, and coordinates autonomous subagents to converge every node clean.

# Clean an entire package scope across all roles in topological order
uv run autofab clean my_project

# Clean a package scope up to a specific role boundary (e.g., up to planning or lib)
uv run autofab clean my_project#planning
uv run autofab clean my_project#lib

# Clean a specific component target node or implementation file
uv run autofab clean //my_project/parts/agent:agent_config#lib
uv run autofab clean my_project/parts/agent/lib/agent_config.py

# Select an explicit model profile defined in model_configs.toml
uv run autofab clean my_project --model gemini-flash
uv run autofab clean my_project --model deepseek-flash

Marking Targets Dirty (mark-dirty)

When you want to invalidate a node, trigger downstream re-verification, or prompt agents to rework an implementation, use mark-dirty:

# Mark an entire scope dirty with a default reason
uv run autofab mark-dirty my_project

# Mark a specific target dirty with an actionable diagnostic reason
uv run autofab mark-dirty my_project/parts/agent/lib/agent_config.py "Type signature mismatch with low specification"
uv run autofab mark-dirty //my_project/parts/agent:agent_config#lib -m "Unit test regression on timeout handling"

---

2. Advanced & Internal Subagent Tooling

The commands below are primarily used by autonomous subagents and role workers inside isolated workspaces, or by advanced operators inspecting the DAG:

work-queue

Inspects the dependency DAG and outputs ready vs. blocked tasks for a scope or role.

# Inspect all ready and blocked nodes across a scope
uv run autofab work-queue my_project

# Inspect work queue for a specific role boundary
uv run autofab work-queue my_project#lib
uv run autofab work-queue my_project#qa

# Output JSON structured task metadata (used by agent coordinators)
uv run autofab work-queue my_project --json

mark-clean

Marks target nodes or entire directory scopes clean in metadata and DAG storage without invoking subagents.

# Mark all nodes in a scope clean
uv run autofab mark-clean my_project

# Mark a specific target clean
uv run autofab mark-clean my_project/parts/agent/lib/agent_config.py

commission & decommission

Manages isolated, double-blind role workspaces (autofab_role_workspaces/<role>_<scope>/).

# Initialize role workspaces bound to a model profile
uv run autofab commission --all my_project --model deepseek-flash

# Decommission and delete provisioned role workspaces
uv run autofab decommission my_project

refresh-sys

Synchronizes system tools, linters, methodology configs (autofab_roles.toml), and companion guides from canonical roots to provisioned role workspaces.

uv run autofab refresh-sys

get_work

Acquires the next assigned dirty target inside an isolated role workspace.

bin/get_work

check_files

Runs deterministic verification checks (linters, type checkers, test suites) against targets in a role workspace.

bin/check_files my_project/parts/agent/lib/agent_config.py

submit

Verifies and submits changes for a file with in-band metadata updates (LAST_CLEANED, LAST_CHANGED, CODE_HASH).

bin/submit my_project/parts/agent/lib/agent_config.py "Implemented timeout helper"

blame

Attributes an actionable defect critique or feedback to an upstream culprit contract or specification file, triggering upstream invalidation.

bin/blame my_project/parts/agent/low/agent_config.pyi "Contract missing parameter 'timeout'"

fail

Records an unresolvable failure or impasse for a target, logging diagnostic failure details and leaving the node dirty.

bin/fail my_project/parts/agent/lib/agent_config.py "Abstract method resolution error in pytest"

coverage

Evaluates statement test coverage between an implementation and its double-blind unit test suite.

uv run autofab coverage --impl my_project/parts/agent/lib/agent_config.py --test my_project/parts/agent/tests/agent_config_test.py -t 100.0

change & prompt

# Record an explicit manual change description on a target
uv run autofab change my_project/parts/agent/lib/agent_config.py "Refactored timeout helper"

# Print the task prompt and contract instructions for a target
uv run autofab prompt //my_project/parts/agent:agent_config#lib

configure

Scaffolds user global configuration directory ~/.config/autofab with starter model_configs.toml and credentials templates.

uv run autofab configure

install-antigravity-plugin

Installs or refreshes the Autofab Antigravity plugin and agent definitions into ~/.gemini/config/plugins/autofab. When running uv run autofab clean with an Antigravity model profile, this is performed automatically.

uv run autofab install-antigravity-plugin

---

3. Target Addressing Syntax

Autofab targets can be addressed using multiple canonical notations:

Format	Example	Description
Fully Qualified Target	//my_project/parts/agent:agent_config#lib	Exact unit and role address
File Path	my_project/parts/agent/lib/agent_config.py	Automatically resolved to target address
Role Partition	my_project/parts/agent/lib or my_project#lib	All units under the lib role
Part Directory	my_project/parts/agent	All roles and units under agent part
Scope Directory	my_project	Entire directory scope across all parts & roles

---

4. Model Configuration Schema (model_configs.toml)

Autofab resolves model profiles and rate cards from:

1. Local workspace: ./model_configs.toml
2. User global configuration: ~/.config/autofab/model_configs.toml (default)
3. Canonical starter template: autofab_python/templates/model_configs.toml

Run uv run autofab configure to scaffold ~/.config/autofab/model_configs.toml.

TOML Schema

# Top-Level Configuration
default = "deepseek-flash"               # Default model profile name
harness = "dsh"                          # Default runtime harness ("dsh", "antigravity")
antigravity_default = "gemini-flash"     # Default model profile for Antigravity harness
dsh_default = "deepseek-flash"           # Default model profile for DeepSeek Harness
max_unproductive_rounds = 3              # Global circuit-breaker round limit

# Model Profiles: [models.<profile-name>]
[models.deepseek-flash]
model = "deepseek-flash"                 # Model identifier passed to harness
provider = "deepseek"                    # Provider adapter ("deepseek", "google", "mlx", "anthropic")
harness = "dsh"                          # Runtime harness ("dsh", "antigravity")
tier = 2                                 # Capability tier (1=Local, 2=Budget, 3=Mid-Tier, 4=Frontier)
tier_label = "Budget API"                # Descriptive capability label
rate_uncached_in = 0.15                  # USD per 1,000,000 uncached input tokens
rate_cached_in = 0.003                   # USD per 1,000,000 cached input tokens (prompt cache hit)
rate_out = 0.60                          # USD per 1,000,000 output tokens
subagent_batch_size = 10                 # Maximum concurrent role subagent dispatches
timeout_seconds = 120.0                  # Request/process timeout in seconds

[models.gemini-flash]
model = "gemini-flash"
provider = "google"
harness = "antigravity"
tier = 2
tier_label = "Budget Frontier API"
rate_uncached_in = 0.10
rate_cached_in = 0.025
rate_out = 0.40
subagent_batch_size = 5
timeout_seconds = 120.0

[models.gemini-pro]
model = "gemini-pro"
provider = "google"
harness = "antigravity"
tier = 3
tier_label = "Mid-Tier Frontier API"
rate_uncached_in = 1.25
rate_cached_in = 0.3125
rate_out = 5.00
subagent_batch_size = 5
timeout_seconds = 180.0