███████╗ ██████╗ ██████╗ ██████╗ ███████╗
██╔════╝██╔═══██╗██╔══██╗██╔════╝ ██╔════╝
█████╗ ██║ ██║██████╔╝██║ ███╗█████╗
██╔══╝ ██║ ██║██╔══██╗██║ ██║██╔══╝
██║ ╚██████╔╝██║ ██║╚██████╔╝███████╗
╚═╝ ╚═════╝ ╚═╝ ╚═╝ ╚═════╝ ╚══════╝
Phase-locked development workflow for Claude Code
Evidence gates. Test-first discipline. No shortcuts.
What Forge Does
Forge treats software construction as a series of earned progressions. Each phase produces demonstrable evidence before the next phase unlocks. The agent cannot skip steps, cut corners, or claim completion without proof.
Rules are suggestions. Gates are enforcement. You can write "always test first" in CLAUDE.md, and the agent will agree, then skip it when it feels confident. Forge intercepts tool use: the phase-gate hook blocks code edits during pre-execution phases, the commit-guardian requires test evidence before commits, and discipline skills resist rationalization under pressure. The difference between a rule and Forge is the difference between a suggestion and a wall.
Workflow
stateDiagram-v2
direction LR
[*] --> Discovery
Discovery --> Design : direction approved
Design --> Planning : spec approved
Planning --> Execution : plan approved
Execution --> Verification : all tasks complete
Verification --> Integration : evidence collected
Integration --> [*] : code landed
state Execution {
direction TB
[*] --> ProveFail: write test
ProveFail --> Implement: RED confirmed
Implement --> ProvePass: run test
ProvePass --> Review: GREEN confirmed
Review --> [*]: approved
Review --> ProveFail: issues found
}
state Verification {
direction TB
[*] --> Tests
Tests --> Build
Build --> Security
Security --> SpecAudit
SpecAudit --> [*]: all pass
}
Each transition requires evidence. No phase advances on assertions alone.
| Phase | Skill | Gate |
|---|---|---|
| Discovery | discover-intent | User approves direction |
| Design | shape-design | Spec reviewed and approved |
| Planning | chart-tasks | All tasks have verification criteria |
| Execution | drive-execution + prove-first | All tasks pass test-first + review |
| Verification | inspect-work + confirm-complete | Tests, build, security, spec coverage |
| Integration | land-changes | User chooses merge/PR/keep/discard |
Skill Selection
Forge uses a three-tier routing system instead of blanket invocation rules.
Tier 1 -- Unconditional. These apply to all implementation and debugging work. No exceptions.
| Skill | Trigger |
|---|---|
prove-first | Any new production code or bugfix |
trace-fault | Any bug, test failure, or unexpected behavior |
confirm-complete | Any claim that work is done |
Tier 2 -- Intent-matched. Activate when their description matches the task.
discover-intent | shape-design | chart-tasks | drive-execution | inspect-work | land-changes | distill-lessons | receive-feedback
Tier 3 -- User-invoked. Available via slash command, never auto-triggered.
| Skill | Command |
|---|---|
isolate-work | /forge:isolate-work |
Install
# Add the marketplace and install
claude plugins marketplace add caseyrtalbot/claude-plugins
claude plugins install forge@caseyrtalbot
Or from a local clone:
git clone https://github.com/caseyrtalbot/forge.git
claude plugins install ./forge
Requirements: Claude Code v1.0.33+, Node.js 18+, Git
Quick Start
/forge:start "Add user authentication" # Begin a workflow
/forge:status # Check current phase
/forge:advance # Advance when gates pass
/forge:audit # Run full quality audit
After /forge:start, the agent enters Discovery: reads project context, asks questions one at a time, refines what to build. Once you approve a direction, it moves through Design (spec), Planning (tasks), Execution (test-first implementation with per-task review), Verification (tests, build, security scan), and Integration (merge/PR). The entire workflow persists across sessions in .forge/forge-state.json.