v2.0 — Task Graph Kernel

octie

State-oriented task graph kernel for the agentic era

Loop engineering substrate. Context engineering infrastructure. State is derived, not set. The task layer that every agent framework can call.

View on GitHub →
Octie - Harness Your Agents

Up and running in seconds

No cloud. No database. No external services. Just a local CLI that runs anywhere Node.js runs.

Terminal

1 — Install

$ npm install -g octie-cli

2 — Initialize

$ mkdir my-project && cd my-project
$ octie init

3 — Create a task

$ octie create \
  --title "Implement login endpoint" \
  --description "POST /auth/login with JWT" \
  --success-criterion "Returns 200 with valid JWT" \
  --deliverable "src/auth/login.ts" \
  --priority top

4 — Visualize

$ octie list --graph
$ octie serve

State-oriented. Agent-native.

Not another todo list. Octie is the task substrate that bridges human intent and agent execution — where state is derived, not set.

01

Loop Engineering Substrate

Design the system that prompts your agents — not the prompts themselves. Octie provides discovery, verification, memory, and guardrails for agent loops.

02

Context Engineering Infrastructure

Structured, typed, dependency-aware context for agents. Not just text — tasks, criteria, deliverables, blockers. The context that lives between steps.

03

Agent-Native Task Layer

CLI-first, file-based, local-first. Designed for any shell agent — Claude Code, Codex, Gemini CLI, or any terminal-based AI. No cloud dependency.

04

State-Oriented Architecture

Status is derived, not set. Five states compute from item completion and blockers. State-oriented programming — the system recalculates itself, agents cannot drift out of sync.

05

Agentic Development Lifecycle

Tasks, criteria, deliverables, approval gates, audit trails. The full ADLC — from planning to implementation to review — tracked in a single graph.

06

Dual Output for Agents & Humans

Markdown for AI consumption. JSON for visualization. The same task graph serves both agent loops and human Kanban boards.

Five derived states

Status is never set manually. It derives from item completion, blockers, and approval gates. The system recalculates itself.

Set up auth middleware

top

Add JWT auth middleware to protect API routes, verify tokens on each request.

Build registration endpoint

top

Create POST /register with input validation, password hashing, and email verification.

Database migration scripts

second

Write migration scripts for new user profile fields. Include rollback scripts.

Add health check endpoint

later

Simple GET /health returning service status, DB connectivity, and uptime.

Deploy CI/CD pipeline

top

Set up GitHub Actions for automated testing and deployment. Blocked pending IAM role.

Auto-broadcast. Manual approve.

Two complementary mechanisms keep the task graph in sync: automatic state broadcast through the graph, and a single human approval gate.

Auto State Broadcast

Status is never set manually. It derives from item completion and blockers, then propagates through the graph via BFS. When a parent completes, children unblock automatically.

ready in_progress in_review completed
# Status auto-derives when criteria are met
octie update <id> --complete-criterion <id>

Manual Approval Gate

The only manual transition: in_review → completed. This is the human gate — the one moment where intent meets execution. Approval unblocks all dependent tasks.

ready in_progress in_review completed
# Only manual transition — the human gate
octie approve <id>

Structured workflows for any agent

Six skills that guide any shell agent through the complete Octie workflow — from installation to refinement. Works with Claude Code, Codex, Gemini CLI, or any terminal-based AI.

octie-install

Setup & Onboarding

Install the Octie CLI globally, verify the installation, initialize a project, and guide the user through basic usage.

Trigger: /octie-install
octie-research

Deep Investigation

Use web search tools, C7 MCP, and Tavily to research technology stack, architecture patterns, edge cases, and best practices.

Trigger: /octie-research
octie-plan

Task Creation

Create well-structured, atomic Octie tasks with proper dependencies, success criteria, and deliverables. Validate the graph.

Trigger: /octie-plan
octie-dev

Development Loop

Execute a development round — pick up tasks, implement features, verify criteria, update progress, create audit trails, and commit.

Trigger: /octie-dev
octie-fix

Bug Investigation

Analyze bug reports, investigate root cause, present pre-modify workflow, and create properly structured fix tasks.

Trigger: /octie-fix
octie-refine

Graph Optimization

Rearrange dependencies, merge small tasks, split oversized tasks, remove unnecessary blockers, and validate graph integrity.

Trigger: /octie-refine

Plan · Execute · Approve · Loop

The agentic loop. Tasks are discovered, executed, verified, and approved — then the next task is ready. The loop drives itself.

01

Plan — Define the Work

Break complex goals into atomic tasks with quantitative success criteria and concrete deliverables. Octie validates they're specific, measurable, and executable.

octie create --title "..." --success-criterion "..." --deliverable "..."
02

Execute — Let Agents Work

Agents discover ready tasks, implement changes, and mark criteria complete. Status auto-derives from item completion. No manual status updates — the system knows.

octie list --status ready --format md
03

Approve — Gate the Quality

When all items are complete, approve the task. This is the human gate — the one moment where intent meets execution. Approval unblocks the next work.

octie approve <id>
04

Loop — The Cycle Continues

Approved tasks unblock dependents. New tasks become ready. The agent loop discovers them, executes, and the cycle repeats — until the project is done.

octie list --graph

Dependencies made visible

Tasks form a directed acyclic graph. Status propagates through dependencies. When a parent completes, children unblock automatically.

┌─ Implement JWT login [ready]
├──→ ┌─ Add password hashing [in_progress]
│    │
│    └──→ ┌─ Password reset flow [blocked]
└──→ ┌─ Write auth tests [ready]
octie list --graph
O(1)
Graph lookups
6
Graph operations
5
Auto-derived states
16+
CLI commands

The loop drives itself

Agents discover ready work, execute, and mark progress. Status derives automatically. The loop continues until the project is done.

# Get ready tasks — markdown for AI consumption
$ octie list --status ready --format md

# Get task details before implementing
$ octie get <task-id> --format md

# Mark progress — status auto-derives
$ octie update <id> --complete-criterion <criterion-id>
$ octie update <id> --complete-deliverable <deliverable-id>

# Status: ready → in_progress → in_review (automatic)
# Approve to unblock dependents
$ octie approve <id>

State-oriented task graph kernel

Loop engineering substrate. Context engineering infrastructure.
State is derived, not set. Local-first. CLI-native. Agent-ready.

Get Started ↓ Contact

Get in Touch

Questions, feedback, or collaboration ideas?
I'd love to hear from you.

zhangyc0307@gmail.com