fableforge-anvil-agent 0.3.0

The open-source, self-verified coding agent. Generate → Execute → Verify → Recover.

Anvil — The Self-Verified Coding Agent

Generate → Execute → Verify → Recover

Every other open agent generates and hopes. Anvil generates, runs, checks, and fixes — because it was trained on 210,000 examples of real agents doing exactly that.

This isn't prompt engineering. This is behavior engineering.

---

Why Anvil?

Other Agents	Anvil
Generate code and hope it works	Generate code, then verify it works
No error recovery	Self-healing with 3 retry attempts
One-shot output	Iterative Plan→Execute→Verify→Recover loop
No cost awareness	Token tracking + model routing for cost optimization
Black box	Full session tracking, verify reports, telemetry
Requires expensive API	Runs fully local with ShellWhisperer (1.5B)

The Verification Loop

   ┌──────┐     ┌──────┐     ┌──────┐     ┌──────┐
   │ PLAN │────▶│ EXEC │────▶│VERIFY│────▶│ DONE │
   └──────┘     └──────┘     └──┬───┘     └──────┘
                                 │ Fail
                                 ▼
                            ┌──────┐
                            │RECOVR│────▶ back to EXEC
                            └──────┘

Anvil doesn't just write code. It verifies every change:

1. Syntax check — Does the code parse?
2. Test run — Do the tests pass?
3. Lint check — Is the code clean?
4. Import check — Are dependencies valid?

If verification fails, Anvil diagnoses the error, generates a fix, and re-verifies. Up to 3 retry cycles. This isn't optional — it's the core loop.

Quick Start

pip install anvil-agent

# Run with local model (ollama)
anvil run "Add error handling to main.py"

# Run with API model
anvil run -m gpt-4o "Refactor the auth module"

# Interactive chat with verification
anvil chat

# Verify existing code
anvil verify src/

# Start as persistent daemon
anvil daemon --port 8765

# List past sessions
anvil sessions

The Name

Anvil — where code gets forged, hammered, and tested until it holds.

Every blacksmith knows: you don't just shape metal on the anvil. You test it. You strike it, check it, and if it's not right, you heat it again and hammer it until it is. That's what this agent does with code.

Other agents shape and ship. Anvil shapes, verifies, and only then ships.

Architecture

anvil/
├── core/
│   ├── engine.py          # Plan→Execute→Verify→Recover loop
│   ├── config.py          # 7-layer configuration system
│   └── session.py          # Full session tracking + persistence
├── tools/
│   └── executor.py         # Bash, Read, Write, Edit, Grep, Glob, LS
├── verify/
│   └── pipeline.py         # Syntax, test, lint, import verification
├── models/
│   └── registry.py          # Local (ollama), OpenAI, Anthropic + cost tracking
├── integrations/
│   ├── verifyloop.py        # VerifyLoop framework integration
│   ├── error_recovery.py   # ErrorRecovery engine integration
│   ├── agent_swarm.py      # AgentSwarm coordination integration
│   └── cost_optimizer.py   # CostOptimizer routing integration
├── daemon/
│   └── server.py            # Persistent HTTP daemon mode
├── tui/
│   └── dashboard.py         # Rich terminal dashboard
└── cli.py                  # run, chat, verify, daemon, sessions, models

The FableForge Ecosystem

Anvil is the flagship product of the FableForge ecosystem — 21 open-source projects built from 210K real agent traces:

Project	What It Does
Anvil	Self-verified coding agent (this one)
VerifyLoop	Plan→Execute→Verify→Recover framework
ErrorRecovery	Self-healing middleware (3,725 error examples)
FableForge-14B	The fine-tuned model (4-stage training)
ShellWhisperer	1.5B edge agent (phone/RPi, 50ms)
ReasonCritic	Verification model (130 benchmark tasks)
TraceCompiler	Compile traces → LoRA skills
AgentRuntime	Persistent agent daemon (systemd for AI)
AgentSwarm	Multi-agent from real trace transitions
AgentTelemetry	Datadog for agents (token tracking, costs)
BenchAgent	HumanEval for tool-use (107 tasks)
AgentDev	VSCode extension with verification
TraceViz	Trace replay visualizer (Next.js)
AgentSkills.org	npm for agent behaviors
AgentCurriculum	5-stage progressive training
AgentFuzzer	Adversarial testing for agents
AgentConstitution	Safety guardrails from traces
CostOptimizer	Token cost reduction (50-80%)
AgentProfiler	Behavioral fingerprinting
TrajectoryDistiller	Trace→training data pipeline
Fable5-Dataset	HuggingFace dataset release

Configuration

Create .anvil.json in your project root:

{
  "model": {
    "model": "local",
    "temperature": 0.2,
    "max_tokens": 4096
  },
  "verify": {
    "enabled": true,
    "auto_recover": true,
    "max_retries": 3,
    "check_syntax": true,
    "check_tests": true,
    "check_lint": true
  },
  "tools": {
    "allow_shell": true,
    "sandbox": false
  },
  "safety": {
    "constitution_enabled": true,
    "blocked_commands": ["rm -rf /", "mkfs"],
    "require_confirmation_for": ["git push", "DROP TABLE"]
  },
  "cost": {
    "max_cost_per_session_usd": 5.0,
    "route_by_complexity": true,
    "simple_model": "local",
    "complex_model": "gpt-4o"
  }
}

Daemon Mode

Run Anvil as a persistent server:

anvil daemon --port 8765

curl -X POST http://localhost:8765/run \
  -H "Content-Type: application/json" \
  -d '{"task": "Add input validation to all API endpoints"}'

Model Backends

Model	Type	Input $/1M	Output $/1M
local (fableforge-14b)	Local	Free	Free
gpt-4o	API	$2.50	$10.00
gpt-4o-mini	API	$0.15	$0.60
o3-mini	API	$1.10	$4.40
claude-3.5-sonnet	API	$3.00	$15.00
claude-3.5-haiku	API	$0.80	$4.00

How It's Different

Trained on Real Behavior

The FableForge model was trained on 210K examples from real agent traces:

• 87.7% planning rate — agents plan before they act
• 39.5% error recovery rate — agents that hit errors and recover
• 1,311-step trace — the Boeing 747 trace proves agents need persistent runtime
• 31 tools mapped — transition matrices drive swarm coordination

Verification Is Not Optional

Other agents: "Here's the code, hope it works."

Anvil: "Here's the code. I ran it. Tests pass. Lint is clean. Imports resolve. Here's the proof."

Self-Healing

When verification fails, Anvil doesn't just report the error. It reads the error, generates a fix, applies it, and re-verifies. This is the ErrorRecovery engine with 3,725 real error examples baked in.

Ecosystem Integration

Anvil doesn't work alone. It's wired into the full FableForge stack:

• VerifyLoop → Sophisticated multi-step verification
• ErrorRecovery → Pattern-matched error resolution from real traces
• AgentSwarm → Multi-agent coordination via transition matrices
• CostOptimizer → Automatic model routing based on task complexity
• AgentConstitution → Safety guardrails from analysis of real traces

License

MIT

Built With

• 210,000+ real agent traces from the Fable-5 dataset collection
• 87.7% planning rate behavioral signal
• 39.5% error recovery success rate
• 303 tool calls in a single session (Boeing 747 trace)
• 5 specialized micro-models (ShellWhisperer, ReasonCritic, etc.)

---

Anvil: Forge your code. Verify it holds.