CodeForge v0.2.0 — The contract around the model

CodeForge v0.2.0 — The contract around the model

Release date: 2026-10-22
Download: Windows installer · Repo

---

The AHA moment

Most “local AI IDEs” stall when the model is small: they over-read, under-edit, or declare victory without verifying.

CodeForge fixes the agent loop itself — read → edit → test → finish — and adapts the harness to each model.

On an RTX 3060 12GB, with Qwen3.6-35B (IQ2_S) via llama.cpp:

Task	Mean (3 reps)	Stuck
T1 — bugfix	100	0
T2 — feature (coupon)	97	0
T3 — multi-file	100	0
T4 — rename refactor	100	0

Zero stuck runs. Zero bad tool-args. A ~Q2 35B completing real agentic coding tasks end-to-end on a gaming GPU.

The model is not the product — the contract around the model is.

---

What's new

Agentic loop that finishes

• Unlimited exploration within the turn budget (Cursor / Claude Code style)
• Continue when the step budget ends — no soft “explore cap” that refuses new reads
• Honest outcomes: stuck / truncated / oracle-red are failed, not “done”
• Spec self-check and oracle before finish with writes
• Diagnose-only finishes refused until the agent mutates or says blocked:

Adaptive harness

• Per-model profiles (~/.codeforge/ai/model-profiles/)
• Calibrate suite packaged in the extension (T2 / T4 / CS2 / CS3)
• Harness revise with paired bootstrap — keep only when the gain is real
• Optional toolTransport for XML-leaky locals
• Online hysteresis without quietly contaminating harness.json

You own the freedoms

• Sandbox level, Allow all, outside-workspace writes, MCP, Full agent freedom
• All user settings, revocable — the model cannot grant itself permissions
• API keys in SecretStorage (not plain settings JSON)
• DOMPurify + CSP nonce on chat / settings webviews

Measurable & robust

• Protocol telemetry: native vs content tool calls, bad args, length exits
• MCP / hooks / sessions hardened for daily use
• Windows process kill and shell prep fixes
• LinkedIn release announce workflow (optional)

---

Install

1. Download CodeForge-Setup-0.2.0-win32-x64.exe
2. Run the installer (per-user under %LOCALAPPDATA%\Programs\CodeForge)
3. Open Settings → AI → point at your OpenAI-compatible endpoint (llama.cpp, Ollama, vLLM, cloud APIs…)
4. Pick a model → Agent mode → ship

No Docker. No vendor lock-in. Your models. Your machine.

---

Not in this release

• Intent crutches that rewrite user messages based on phrase matchers
• One global “weak mode” that punishes strong models
• Claiming 100% forever — residual failures stay visible as model limits

What's Changed

• CodeForge v0.2.0: adaptive harness and agent contract by @Kortexio in #1

New Contributors

• @Kortexio made their first contribution in #1

Full Changelog: v0.1.3...v0.2.0