Skip to content

Releases: DjinnFoundry/forge-loop

v0.6.0 — Forge-first surface cleanup

20 Mar 16:06

Choose a tag to compare

[0.6.0] - 2026-03-20

Added

  • /forge-cancel as the primary Claude Code stop command, with /cancel-ralph kept as a legacy alias

Changed

  • Forge now uses FORGE_COMPLETE and FORGE_PAUSE as the primary control markers while still accepting legacy Ralph markers
  • Claude driver docs and install flow now use Forge-first naming and forge-loop.*.local.md as the primary loop-state file name
  • README positioning now keeps Ralph and other influences in an explicit lineage section instead of mixing them into the main product surface

v0.5.1 — Claude hook install fix

20 Mar 16:07

Choose a tag to compare

[0.5.1] - 2026-03-20

Fixed

  • Claude stop hook now resolves its real source path before loading shared helpers, so symlinked installs under ~/.claude/hooks/stop-hook.sh no longer fail with missing helper-library errors
  • Added regression coverage for running the stop hook through a symlinked install path

v0.5.0 — Task-driven success contracts

20 Mar 16:07

Choose a tag to compare

[0.5.0] - 2026-03-20

Added

  • Optional explicit --done-when "TEXT" success contract for Forge task scopes

Changed

  • Forge now treats open-text task completion as the primary objective and KPI targets as guardrails
  • Codex driver status and prompt rendering now surface success mode and done_when text
  • README and driver docs now describe task-derived completion checks when no explicit success override is provided

v0.1.0 — The Forge

20 Mar 11:56

Choose a tag to compare

The Forge — Autoregressive Codebase Improvement for Claude Code

A structured, KPI-driven, self-correcting loop that tracks metrics (coverage, speed, quality), evaluates with fresh-context subagents, rotates strategies when stagnating, and knows when it's done.

What's included

  • The Forge Protocol — eight-phase iteration cycle (Orient, Measure, Evaluate, Decide, Execute, Verify, Record, Complete)
  • 7 named strategies with automatic selection based on normalized KPI gaps
  • Stagnation detection and automatic strategy rotation
  • Fresh-context evaluation via subagents every 3rd iteration (prevents anchoring bias)
  • Autoregressive state file that persists KPIs, strategies, and lessons across iterations
  • Stop hook for the iteration engine (compatible with Ralph Wiggum loops)
  • Multi-language support — Elixir, Python, JavaScript, Ruby, Go

Installation

git clone https://github.com/DjinnFoundry/forge-loop.git
cd forge-loop
./install.sh

Usage

/forge "API controllers" --coverage 90 --speed -30%

See README for full docs.

Standing on the shoulders of

  • Geoff Huntley's Ralph Wiggum — autonomous AI development loops
  • Andrej Karpathy — autoregressive mindset
  • SICA (ICLR 2025 SSI-FM Workshop) — compounding iterations