CIAgent — Continuous Intelligence

Fully autonomous, git-native AI-driven software engineering harness.

Overview

CIAgent (Continuous Intelligence) is an autonomous-first software engineering harness that eliminates human-in-the-loop overhead while preserving the rigor of guided development. It receives a specification, resolves ambiguities through a single Clarify phase, then executes the full pipeline — research, plan, execute, verify — autonomously.

The git log IS the project memory. Every decision, escalation, lesson learned, and verification result is encoded in commit messages using structured ---ci--- YAML blocks. An agent's first impulse to gather context is git log, not file reads. Another agent with access to only commit messages (no code, no diffs) can reconstruct the project state completely.

Installation

From source (package not yet published to npm):

git clone https://git.cloudinit.dev/continuous-intelligence/ci.git
cd ciagent
npm install
npm run build
npm link

Quick Start

# Initialize from inline specification
ciagent init "Build a REST API for task management"

# Initialize from a specification file
ciagent init --spec ./specs/my-project.md

# Run the full autonomous pipeline
ciagent run --all

# Run a specific phase
ciagent run research
ciagent run plan
ciagent run execute
ciagent run verify

# Execute an ad-hoc task
ciagent quick "Add authentication middleware"

# Check project status (reads from git log + branches)
ciagent status

# Review autonomous decisions (extracted from git log ---ci--- blocks)
ciagent audit
ciagent audit --verbose

# Debug an issue
ciagent debug "Tests failing on CI"

# Rollback a phase
ciagent rollback 1

# Ship a phase (verify, security, commit, tag)
ciagent ship 1

Git-Native Architecture (v0.2.0)

The Commit Schema

Every CIAgent-generated commit contains a ---ci--- YAML block with structured metadata:

feat(P01-01-02): create user registration endpoint

---ci---
phase: 1
milestone: v1.0
plan: 01-01
task: 01-01-02
status: execute
decisions:
  - id: D-003
    decision: Use bcrypt with 12 rounds for password hashing
    rationale: Industry standard; argon2 not available in target env
    confidence: 0.88
    alternatives: [argon2, scrypt]
requirements:
  covered: [AUTH-01]
---/ci---

- POST /auth/register validates email and password
- Checks for duplicate users
- Returns JWT token on success

What Lives Where

Where	What	Why
.ciagent/config.json	Autonomy, thresholds, git strategy	Controls system behavior before any commits exist
.ciagent/PROJECT.md	Vision, core value, requirements, constraints, key decisions table	Long-lived strategic reference
.ciagent/ARCHITECTURE.md	System architecture, component boundaries, data flow	Long-lived technical reference
.ciagent/ROADMAP.md	Phase breakdown, milestone mapping, success criteria	Long-lived planning reference
.ciagent/REQUIREMENTS.md	v1/v2 requirements with REQ-IDs and traceability	Long-lived requirements reference
Git commit bodies	Decisions, escalations, lessons, compounds, verification results	Dynamic event stream — the audit trail
Git branches	Phase/milestone status	phase/NN-slug and milestone/vX.X-slug encode project structure

Branch Strategy

main
 └── milestone/v1.0-mvp
      ├── phase/01-authentication       # in progress if not merged
      ├── phase/02-task-management
      └── phase/03-realtime-notifications

• Branch exists + not merged = phase in progress
• Branch merged to milestone = phase complete
• Milestone branch merged to main = milestone complete

Context Reconstruction Protocol

An agent starting a session gathers context in this order:

1. git log --oneline -20 — recent activity
2. git branch -a — phase/milestone structure
3. git log -1 --format="%b" — latest ---ci--- block
4. .ciagent/config.json — autonomy + thresholds
5. .ciagent/PROJECT.md — vision + constraints (when needed)
6. .ciagent/ROADMAP.md — phase plan + success criteria (when needed)
7. .ciagent/REQUIREMENTS.md — REQ-IDs + traceability (when planning)
8. .ciagent/ARCHITECTURE.md — system structure (when researching)

Steps 1-3 take <1 second and provide 80% of the context needed.

The Reconstruction Test

An agent with access to only commit messages (no code, no diffs, no .ciagent/ files) can reconstruct:

Reconstructable	How
Project specification	Init commit body
Current phase	---ci---.phase field + branch status
Current milestone	Branch names + ---ci---.milestone field
All decisions with rationale	git log --grep="decisions:" --format="%b"
Decision confidence	Each decision has confidence: 0.XX
Alternatives considered	Each decision has alternatives: [...]
Requirements coverage	git log --grep="requirements:" --format="%b"
Lessons learned	git log --grep="lessons:" --format="%b"
Compounded solutions	git log --grep="compound:" --format="%b"
Escalation history	git log --grep="escalation:" --format="%b"

Commit Types

In addition to conventional commit types, CIAgent uses:

Type	When Used
decision	Autonomous decision logged (no code change)
compound	Compounded solution captured
escalation	Escalation raised or resolved
verify	Verification pass/fail
wip	Work-in-progress checkpoint

Autonomy Levels

Level	Behavior
full	No human interaction after Clarify. Escalate only irreversible decisions.
supervised	Escalate on every Escalation Gate plus verification failures.
guided	Escalate on every Decision Gate.

Configuration

CIAgent uses .ciagent/config.json for project configuration:

{
  "autonomy": {
    "level": "full",
    "escalation_hooks": ["deploy", "delete_data", "merge_to_main"],
    "clarify_budget": 10,
    "decision_confidence_threshold": 0.6,
    "max_revision_iterations": 3,
    "max_verification_retries": 2,
    "escalation_timeout_ms": 300000
  },
  "model_profile": "quality",
  "parallelization": {
    "enabled": true,
    "max_concurrent_agents": 5,
    "min_plans_for_parallel": 2
  },
  "verification": {
    "automated_only": true,
    "escalate_visual": true,
    "escalate_external_integration": true,
    "test_first": false
  },
  "security": {
    "auto_accept_low_severity": true,
    "auto_mitigate_medium_severity": true,
    "escalate_high_severity": true
  },
  "git": {
    "branching_strategy": "phase",
    "auto_commit": true,
    "auto_push": false
  }
}

Architecture

Pipeline

SPECIFY → CLARIFY → RESEARCH → PLAN → EXECUTE → TEST → VERIFY → COMPLETE
                 ↕               ↕         ↕     ↕      ↕
            (questions)    (auto-decide) (auto-run) (auto-test) (auto-verify)

Git-Native Core Modules

Module	Purpose
commit-parser	---ci--- YAML block extraction and parsing from commit messages
commit-builder	Structured commit message generation for all commit types
git-context	Project state reconstruction from git log + git branch
git-branch	Phase/milestone branch lifecycle management
ciagent-files	.ciagent/ long-lived reference file management with update discipline

Decision Engine

Every autonomous decision is classified by confidence:

• High (>0.85): Auto-decide, commit as ---ci--- block
• Medium (0.60-0.85): Auto-decide with assumption logging, flag for review
• Low (<0.60): Escalate to human

Decisions are committed to git as decision type commits. The audit trail is git log --grep="decisions:".

19 Agents

Agent	Role	CIAgent Modification
orchestrator	Pipeline controller	Git-first context loading, ---ci--- commit generation
planner	Plan creation	Never sets autonomous: false
executor	Task execution	Never pauses for checkpoints
verifier	Output verification	Generates automated tests, not human UAT
researcher	Domain research	Logs assumptions, never flags for human
tester	Integration/e2e tests	Detects and runs existing test files, never writes tests
challenger	Plan stress-testing	Binding verdicts, only escalates <0.60
security-auditor	Security audit	Auto-dispositions threats
debugger	Bug fixing	Auto-fixes when confidence > threshold
Others	Various	Delegates to active intelligence backend

Verification Layers

1. Structural: File existence, import/export wiring, no stubs
2. Behavioral: Test infrastructure and requirement traceability (partially implemented — static analysis, no test generation yet)
3. Security: Regex-based threat pattern scanning with auto-disposition (partially implemented — no STRIDE analysis yet)
4. Code Quality: Regex-based code quality checks (partially implemented — no multi-persona review yet)

Specification Format

# Project: My Project

## Objective
Build a REST API for task management.

## Requirements
- User authentication (JWT-based)
- CRUD operations for tasks
- Real-time notifications

## Constraints
- Must use Node.js
- Must be production-ready

## Out of Scope
- Admin dashboard
- Mobile apps

Escalation Protocol

When CIAgent cannot proceed autonomously:

1. Irreversible Action: Deploy, delete, merge to protected branch
2. Verification Failure: Tests pass but functional verification fails
3. Low Confidence Decision: Critical decision below threshold
4. Security Escalation: High-severity threat detected
5. Specification Ambiguity: Multiple valid interpretations

Each escalation is committed as an escalation type commit. Resolved escalations produce a follow-up commit with the resolution. The full escalation history is available via git log --grep="escalation:".

Current Limitations

• Agent implementations: 5 core agents have intrinsic logic (planner, executor, verifier, researcher, tester); 13 agents delegate to backends. Full LLM-powered agent behavior requires an intelligence backend.
• Package not published to npm: Install from source only until a publishing pipeline is configured.
• Behavioral/Security/Quality verification layers: Partially implemented — structural verification is complete; behavioral does static analysis; security does regex-based threat scanning; quality does regex-based code quality checks.

Differences from Learnship

Dimension	Learnship	CIAgent
Project memory	.planning/ directory files (legacy)	Git log + ---ci--- commit blocks
Audit trail	.ciagent/audit/*.json files (legacy)	git log --grep="decisions:"
State management	STATE.md + STATE.md.json (legacy)	Reconstructed from git on demand
Phase discovery	Read .planning/phases/ directory (legacy)	git branch -a | grep phase/
Human Interactions	19+/lifecycle	1-2/lifecycle
Decision Making	Human decides, agent implements	Agent decides, human reviews post-hoc
Verification	Human UAT	Automated tests + escalation
Specification	Multi-round conversation	Single spec file

Repository

git.cloudinit.dev/continuous-intelligence/ci

License

MIT