Worktree sandbox
Each task was its own git repo; a run branched into an isolated worktree, worked to completion autonomously, and never mutated the task folder until you approved.
← Forge Pro
Archive · 1st generation
Forge Lite
The first-generation agentic coder — built inside GaudeauAI. A natural-language command drove a real tool-use loop (read, write, edit, run, search, patch) against two 256K local coding models, controllable from an iPhone or Mac, with no cloud coder anywhere. Its defining trick was safety by construction: every run executed in its own throwaway git worktree, and nothing touched the real files until you reviewed the diff and tapped Apply.
Now
Forge is now a standalone native Apple app — Forge Pro — with its own server, provider choice, and mobile clients.
Go to Forge Pro →
Subsystem · Forge CLI
An agentic coder that ran on your box — and that you approved.
Approve-before-apply diff
You got a per-file unified diff with include/exclude toggles — Apply All, apply only the files you picked, or discard the whole worktree. Roadmap auto-apply existed, but only as an explicit option; review-first was the default.
8 built-in tools
read / write / edit / undo / applyPatch / listFiles / searchCode / runCommand — plus any registered skill exposed as a tool.
Pause, resume & recover
Pause was cooperative — it held at a clean feature boundary instead of cutting off live model or tool work. Disconnects didn't lose server-side work or event history, and persisted sessions, roadmap progress and Forge state restored after a backend restart.
Sandboxed HTML preview
Generated artifacts — including JS canvas games — ran live in the app inside a CSP-locked WebView that couldn't phone home or navigate away.
Admin-gated + audited
Because it held shell + filesystem power it was the highest rung of the injection threat model — owner-only, home-dir-scoped, with a full JSONL command audit trail.
Run controls
Pick Qwen3-Coder-Next 80B or Qwen3.6 27B MTP, then set reasoning depth, temperature, tool-iteration limit, output-token ceiling, compaction threshold and continuation limits before the run started.
Live context telemetry
The run surface exposed model and phase, output-token activity, per-step and cumulative prompt/eval tokens, context occupancy and compaction state while the agent worked.
Autonomous roadmaps
Speed built then ran a deterministic artifact gate; Balanced planned first; Quality assigned Qwen3.6 to planning/review and Coder-Next to implementation, with up to two build repairs and three review rounds before stopping as Needs Attention.
Two selectable coding models, both at 256K: Qwen3-Coder-Next 80B was the default for coding and interactive Forge; Qwen3.6 27B MTP handled Reason Before Acting and Quality-roadmap planning/review. Sessions persisted 24 hours in SQLite and followed you between iOS and Mac; LSP diagnostics fed back after edits, and fuzzy smart-replace landed changes even when whitespace differed.
This page is an archive of the in-GaudeauAI prototype. The ideas here — worktree isolation, review-first diffs,
local models, run controls — carried forward into Forge Pro,
now a standalone native Apple app.