FORGE playbook
changelog

What changed.

All notable changes to FORGE CLI are documented here. The format follows Keep a Changelog. FORGE adheres to Semantic Versioning.
keepachangelog.com · semver.org

Unreleased

Nothing planned yet.

[0.1.0]2026-04-26Initial Release
Added
  • Immersive Ink terminal dashboard with real-time agent rendering
  • Responsive dashboard: compact and tiny fallback modes for narrow terminals
  • Interactive dashboard: help overlay, keybindings (q / ESC / h), and cancel support
  • Cycling bar spinner beside active agent tool calls
  • Cancellable runs via AbortController across all commands
  • Text renderer with --output flag — three modes: dashboard, text, json
  • Subagent text and tool history visible in all output renderers
  • Three-tier model resolution: --model flag › FORGE_MODEL env › ~/.claude/settings.json
  • Skills for generating ADR, knowledge base, plan, and spec artifacts
  • generate-knowledges command — produces AGENTS.md and knowledge-base/BASE.md from codebase analysis
  • --model flag on all CLI commands with inline alias expansion
  • Gate event handling with confidence score justification output
  • implement command — accepts multiple task files in a single run
  • plan, spec, and implement CLI commands
  • Forge Lead orchestrator coordinating all specialist subagents
Changed
  • Monolithic dashboard.tsx split into modular dashboard/ directory
  • generate-spec phases 3 and 4 now optional — specialists self-review in one pass
  • Phase 6 task file creation now dispatches parallel subagents instead of sequential writes
Fixed
  • Gate score parsing: added markdown fallback for when lead omits the canonical <forge:gate> XML tag
  • run.start event was being dropped — caused missing phase headers in dashboard
  • Agent tool not recognised by SDK translator — subagents were not appearing in dashboard
  • End-of-run failure messages now surface full lead transcript for debugging
  • Phase header now falls back to command name before the first skill phase fires
  • Forge Lead row now renders in AgentsBox before any Task event is dispatched
Performance
  • generate-spec streamlined from ~20 min to ~10 min — context preparation phase removed; task files now created in parallel by specialist subagents