LoopCode — Autonomous Software Orchestrator

Autonomous software engineering

A headless CLI orchestrator built on OpenCode that decomposes natural language goals into a task DAG, executing branches concurrently in isolated Git worktrees with strict 5-layer verification.

curl -fsSL https://raw.githubusercontent.com/arjun-vegeta/loopCode/main/install.sh | bash

LoopCode v2.0.0 · loopcode · main · claude-3-5-sonnet

simulation running

Task Execution Graph esc / click to close

1. Index codebase workspace ✓ completed

2. Write profile route configs ✓ completed

3. Write profile controller implementation ✓ completed

4. Write profile endpoint integrations ✓ completed

5. Validate code & review merge ✓ completed

Verification Detail (5-Layer Gauntlet) esc / click to close

✓ Layer 1: Compilation (compile)

✓ Layer 2: Formatting & Linters (lint)

✓ Layer 3: Integration test execution (tests)

✓ Layer 4: Static security analysis (semgrep)

✓ Layer 5: Peer review assessment (ai-review)

Quota Spend Control esc / click to close

Goal cost audit

spend: $0.18 / limit: $10.00

[████░░░░░░░░░░░░░░░░░░░░░] 1.8%

Monthly cost audit

spend: $32.40 / limit: $100.00

[████████░░░░░░░░░░░░░░░░░] 32.4%

Keyboard Shortcuts esc / click to close

t / Tabopen task list

vopen verification details

bopen spend limits

ctrl+cinterrupt task execution

⠋ idle · waiting for input

spent: $0.00 / $10.00

> "Add a new endpoint for user profile retrieval"

[t] tasks

[v] verify

[b] budget

[h] help

[space] pause

[r] restart

engine mechanics

shared sqlite memory

Central structured persistence decouples agents. Communication happens asynchronously via data contracts, avoiding direct process dependencies.

topological schedules

Sorts task DAGs based on resource allowlists. Spawns independent subtasks in parallel, executing changes in isolated Git worktrees.

5-layer verification

Integrates compilation audits, linter scans, Bun unit testing checks, Semgrep code security patterns, and AI review loops before merging.

cli runtime parameters

loopcode "<goal>"

Accepts a natural language instruction to plan, construct worktrees, and verify code blocks.

-d, --db <path>

Specifies the path to the SQLite persistence database (defaults to loopcode.db).

-r, --resume <id>

Resumes execution from a previously interrupted Session ID or Task ID.

--login

Launches the in-terminal interactive authentication workflow for platform providers.

--headless

Forces non-interactive execution suite mode (ideal for CI/CD environments).

controlled configuration

Configure parameters inside ~/.loopcode/config.toml to restrict model routing definitions, set maximum spend ceilings, and establish validation protocols.

Compatible with all 100+ OpenCode model routers. Route subtasks dynamically to the optimal model based on prompt complexity, cost thresholds, and cached instructions.

[model]
default      = "anthropic/claude-sonnet-4-6"
planning     = "anthropic/claude-opus-4-6"

[budget]
maxMonthlyCostUsd = 100.0
maxGoalCostUsd    = 10.0
maxTaskCostUsd    = 2.0

[safety]
permissionMode           = "acceptEdits"
allowDestructiveRollback = false
maxParallelAgents        = 5