AutoForge

Human intent. AI execution. One shared understanding.

AutoForge is a project-intelligence and orchestration framework for continuous AI-assisted software, data, and AI development. It turns brainstorming, research, planning, design, implementation, validation, and handoff into durable project state that humans and different coding agents can share.

AutoForge does not replace Codex, Claude Code, Cursor, Gemini, Grok, or another agent. The Core CLI gives every agent the same scoped work, decisions, governance, evidence, and next action instead of making each conversation start from scratch.

AutoForge owns deterministic project intelligence and orchestration—not model reasoning or autonomous agency. A separately named future CoJack Labs AI may consume AutoForge, but general conversation, personal memory, model routing, and prompt-to-product experiences are outside this framework's scope.

AutoForge v0.25.2 is the first supported registry release for the v0.25 architecture. Always begin with autoforge version and use explicit Core subcommands in automation.

What v0.25 Provides

| Layer | Package | Responsibility | | -------- | -------------------------------- | -------------------------------------------------------------------------------------------- | | Protocol | @cojacklabs/autoforge-protocol | Versioned contracts, capabilities, launch negotiation, and handoffs | | Core | @cojacklabs/autoforge-core | Deterministic, model-independent project intelligence | | SDK | @cojacklabs/autoforge-sdk | Supported programmatic facade with injected filesystem, Git, clock, and storage effects | | Core CLI | @cojacklabs/autoforge | Project lifecycle, memory, governance, orchestration, validation, and compatibility behavior | | Agent | @cojacklabs/autoforge-agent | Historical experimental runtime; not the framework's future product direction |

Core remains usable without an account, model provider, hosted service, or Agent installation. A first-party AI, consumer billing, personal memory, and prompt-to-product platform are not AutoForge framework milestones.

Install and Attach a Project

Use one global installation so projects share a workspace registry and update path:

npm install --global @cojacklabs/autoforge
# or: pnpm add --global @cojacklabs/autoforge
# or: yarn global add @cojacklabs/autoforge

autoforge version
autoforge attach "$PWD"
autoforge --project "$PWD" doctor
autoforge projects list

attach initializes local .autoforge/ state and registers the Git repository root globally. A nested path still attaches the whole repository. Use autoforge init only for an intentionally local-only environment such as a disposable CI container.

For a local dependency instead, prefix commands with npx:

npm install --save-dev @cojacklabs/autoforge
npx autoforge init
npx autoforge doctor

Upgrading from v0.24? Read the v0.25 migration guide.

The Continuous AutoForge Loop

AutoForge is most effective when it stays involved throughout the product lifecycle, not only when code is written.

| Stage | Human or agent activity | AutoForge capability | | ---------- | -------------------------------------------------------------------- | -------------------------------------------------- | | Brainstorm | Capture an idea without prematurely coding it | intent assess, intent register | | Research | Preserve findings and provenance | research register, knowledge | | Govern | Check standing rules and domain invariants | constitution, domain, doctrine | | Design | Validate screens, components, flows, states, and relationships | design validate, design import, design check | | Plan | Turn intent into features, phases, tasks, issues, or workflows | planning, workflow, add | | Prioritize | Explain why work is now, next, later, or backlog | strategy assess | | Execute | Open a scoped session and resolve the relevant context | start, context --explain, check | | Validate | Run retained quality/security checks and preserve evidence | gate check, evidence | | Remember | Record why a choice was made and what it changes | decide, why, changelog | | Handoff | Transfer structured state between agents without transcripts | recap, Protocol handoffs, orchestrate handoff | | Learn | Connect hypotheses, experiments, and observations to later decisions | learning | | Reorganize | Reassess dependencies, priority, impact, and project location | orchestrate, trace, twin, projects |

Start every session by asking what is already true:

autoforge --project "$PWD" doctor
autoforge --project "$PWD" recap
autoforge --project "$PWD" status --view next

When work is active, resolve the bounded execution packet before editing:

autoforge --project "$PWD" context --explain

From an Idea to Validated Work

Before creating structured JSON, inspect the installed schema rather than guessing its shape:

autoforge schemas list
autoforge intent assess --schema

Then assess the idea, create scoped work, and complete it with durable rationale:

autoforge intent assess product-idea.json --kind planning

autoforge add feature --name "Customer onboarding" --description "Guide new customers to first value"
autoforge add phase --feature <feature-id> --name "First-run flow" --description "Design and implement onboarding"
autoforge add task --phase <phase-id> --name "Build first-run flow" --description "Implement the approved design" --include "src/**" --include "test/**"

autoforge start task <task-id>
autoforge context --explain
autoforge check --path src/onboarding.ts
autoforge gate check
autoforge decide --statement "..." --reasoning "..." --consequence "..." --scope onboarding --keyword first-run --work <task-id> --kind feature-note
autoforge done

done requires a decision linked with --work so completed work leaves an explanation, not merely a changed Git tree. For genuinely trivial work, use autoforge done --no-decision "<reason>"; the bypass remains auditable.

For a new application idea that needs formal discovery, architecture, design, data, and security approval, follow the bootstrap pipeline.

Give This Prompt to Any Coding Agent

Use this at project onboarding, after an AutoForge update, or when switching between Claude, Codex, Cursor, Gemini, Grok, and another repository-aware agent:

Use the globally installed AutoForge CLI for this repository. Review `README.md` and follow the current documentation walkthrough in `docs/README.md` to use AutoForge continuously throughout brainstorming, planning, documentation, design, bootstrapping, development, validation, decision-making, and handoff; follow every applicable `AGENTS.md` and summarize the project's current AutoForge state before making changes.

The complete procedure is in the Agent Setup Guide, and behavioral guidance is in the Agentic AI Guide.

Humans and Agents Share One Source of Truth

Durable project truth includes work, decisions, doctrines, governance, domain knowledge, specifications, traceability, approved evidence, and structured handoffs. Reproducible context packets, leases, provider caches, logs, credentials, and raw transcripts are operational state and should not become portable project memory.

Useful continuity commands:

autoforge recap
autoforge why --query "checkout"
autoforge evidence summary
autoforge twin generate
autoforge twin query --type decision

See Cross-Agent Handoffs and Governance and Memory.

Parallel Agent Work

Do not run multiple writing agents against the same checkout. Build a dependency-aware orchestration plan, then let AutoForge assign isolated worktrees and reject overlapping write scopes:

autoforge orchestrate plan orchestration-plan.json
autoforge orchestrate ready
autoforge orchestrate claim <work-id> --agent codex --role backend
autoforge orchestrate explain <work-id>
autoforge orchestrate status
autoforge orchestrate handoff <assignment-id> handoff.json
autoforge orchestrate release <assignment-id>

If orchestrate explain reports stale or unavailable context, stop, release the claim, and reclaim it before editing.

Core CLI and Agent Compatibility

Explicit subcommands are deterministic. Automation should use commands such as autoforge status --json, autoforge recap, and autoforge gate check --json.

The released CLI retains compatibility code for the historical experimental Agent. That behavior does not establish a first-party AutoForge Agent roadmap. Automation should always use explicit deterministic subcommands. A later compatibility issue will decide how bare invocation and Agent-launch shims are retired without breaking supported installations.

Provider credentials remain in the operating-system credential store:

autoforge credentials set openai
autoforge credentials status openai
autoforge credentials delete openai

See Local Provider Credentials.

Documentation Map

Start with docs/README.md. It identifies current operational guides, v0.25 release material, specialist references, and historical documents. The authoritative product boundary is the AutoForge Framework North-Star. The installed command remains the final authority for syntax:

autoforge help
autoforge schemas list

Development and Release Validation

This repository uses pnpm 11 and Turborepo. From a clean checkout:

pnpm install --frozen-lockfile
pnpm workspace:check
pnpm exec turbo run build typecheck test format:check
pnpm release:status

The completed release sequence, validation evidence, known exclusions, and follow-up work are documented in v0.25 Release Readiness.

Released under the MIT License.