Updates

Product

Forge 0.4.2 released

Forge 0.4.2 is available, with a signed build attached. The editor opens .gz files, decompressed by the DEFLATE implementation written for the PNG decoder in 0.4.0 — the same payload under a different header, so it added a header parse and no dependency, which is the kind of payoff the earlier decision to shrink dependencies was made for. A saved agent conversation no longer stores whole tool results: one had reached 11.4 MB because of a single 7.4 MB result, and the transcript is rewritten in full on every message, so that was eleven megabytes re-serialised each time. Capped at 64 KB keeping both ends, the largest conversation drops to 3.4 MB. The composer grows with what is typed, and the send button now says the agent is working and stops it rather than sitting inert for the whole turn. See the Forge project page.

Product

Forge 0.4.1 released

Forge 0.4.1 is available, with a signed build attached to the release. A patch release carrying three editor fixes, each one a cost that had been invisible until it was measured. Undo history stored a copy of the entire buffer per step, which meant roughly 136 MB for one large file; steps now record only the lines that changed, measured at 25 KB where whole copies cost 215 MB. Saved terminal scrollback stored a colour value beside every character, taking real sessions from 12.3 MB to 0.5 MB once colours were described as runs instead — losslessly, and still reading sessions written by older builds. And a restarted window kept its identity rather than coming back as a new one while its saved state stayed stranded on disk. See the Forge project page.

Product

Forge 0.4.0 released

Forge 0.4.0 is available. The IDE decodes PNG and GIF itself now, in about a thousand lines of decode-only code validated byte for byte against the crate it replaces — six crates left the dependency graph as a result, continuing the stated direction of shrinking rather than growing dependencies. The agent gained a scratch working area under the system temporary directory, so throwaway scripts stop landing in real repositories; writes there skip approval, and the containment check resolves paths rather than comparing strings. A conversation pushed far past the context limit recovers instead of wedging. And the terminal client patches only the transcript lines that changed rather than reprinting the live block on every spinner tick, which had been sustaining 11.2 KB/s of escape sequences for a whole run. See the Forge project page for the project log entry.

Product

Forge 0.3.0 released — Forge IDE ships alongside the CLI

Forge 0.3.0 is publicly available at github.com/Vulkgryph/Forge/releases/tag/v0.3.0. The Forge IDE — the native code editor built on the same Rust agent core as the terminal client — is now public as a second surface; both surfaces ship together and share one version number. The repository is a Cargo workspace carrying forge-agent, forge-tui-rs, and forge-ide. macOS on Apple Silicon is the verified platform; Linux x86-64 builds and passes tests in CI, and Linux ARM64, Windows, and Intel Macs are intended but not yet end-to-end smoke-tested — the release notes carry the per-component platform table. See the Forge project page and the full forge-ide CHANGELOG, forge-agent CHANGELOG, and forge-tui-rs CHANGELOG.

Studio

Neural Substrate retired from the portfolio

Neural Substrate has been removed as a project. Active development had already stopped, and on review the two papers did not stand on equal ground: A different kind of learning system was a research-direction note written before the methodology bar the site has since committed to (via NN-Revival), and it does not meet that bar — single unreplicated anecdote, no frozen artifacts, no pre-registered predictions. It has been retired rather than retrofitted. Provenance over behavior, by contrast, is a methodology note whose lesson generalizes beyond neural systems, and it has been relocated to Research as a standalone note. Old NS URLs 404; the methodology note's old URL redirects to the new one.

Product

Forge repository temporarily private during IDE release prep

The Forge repository has been set to private while the Forge IDE — the graphical companion surface built on the same Rust agent core — is prepared for release and undergoes internal review. The CLI is not changing in substance; both surfaces will re-emerge together as a monorepo at the same URL (github.com/Vulkgryph/Forge) when the IDE release is ready. Existing Forge CLI installs continue to work; new clones and forge-update pause during this window. See the Forge project page for details, and reach [email protected] if you need something in the meantime.

Research

NN-Revival #3 published: Cascade-Correlation on Modern Compute

The third entry in the Forgotten Architectures series is live at Cascade-Correlation on Modern Compute. Fahlman & Lebiere's 1990 constructive learner re-run faithfully in its native domain (two-spirals + a low-dim regression surface) with a labeled MNIST arm. The mechanism reproduces cleanly, but the founding legend does not survive modern training: a two-layer MLP with Adam solves two-spirals at 99.65% held-out vs CasCor's ~79%. Honest verdict: (mostly) earned retirement — a road-not-taken overtaken by three decades of cumulative backprop tooling, not a flaw exposed. The scorecard keeps N1 as a MISS, and a post-freeze exploratory pass documents a leakage-clean rerun that produced a firm null. Two partial-revivals (#1, #2) and one retirement (#3) is the series doing its sorting job. Code + frozen results at github.com/Vulkgryph/cascor-study.

Studio

Focus consolidating on Forge and Bastion

A portfolio-level focus update. Neural Substrate is being rebuilt from the ground up: the path toward a public open source release, as it had been framed on this site, is no longer active, and the current implementation is being set aside. The work returns to fully internal research with no schedule attached to any future public artifact. The two methodology papers already published under NN-Revival (weightless neural networks and CMAC) are unaffected and remain live.

Sentinel is stepping back off the near-term work list; the development laptop is in place for when it returns. Focus is being consolidated on Forge (CLI released; IDE approaching release) and the Bastion OS. The roadmap has been updated to reflect these changes.

Research

NN-Revival #2 published: CMAC on Modern Compute

The second entry in the Forgotten Architectures series is live at CMAC on Modern Compute. Direct follow-through on #1's predicted fix — does the sparse tiled addressing Albus proposed in 1975 avoid the interference saturation vanilla WiSARD hit? Across the swept range, adding capacity never degrades CMAC's quality; the one sharp failure is coverage, not interference, and recovers fully with more data. The scorecard keeps its honest misses in place, including a PARTIAL MISS on N2 (the sample-efficiency prediction). Code and frozen results at github.com/Vulkgryph/cmac-study; all 5 SHA-256 hashes in the paper verifiable against the repo. Next in the backlog: Cascade-Correlation (#3).

Research

NN-Revival launched with first paper

A new research series is live at /research/nn-revival/: Forgotten Architectures, an empirical revisit of pre-backprop neural and learning architectures under modern experimental controls — pre-registered predictions, multi-seed runs, and frozen result artifacts with SHA-256 hashes. The first paper, Weightless Neural Networks on Modern Compute, re-runs classic WiSARD and Sparse Distributed Memory on binarized MNIST and includes a dual-mode continual-learning probe. Deliberately toy-scale and not a SOTA claim; the methodology is the point.

Studio

Bastion and Sentinel unblocked

Development hardware for Bastion (dedicated x86-64 workstation) and Sentinel (development laptop) is now in place. Both projects were previously flagged as hardware-blocked on the roadmap; active work resumes on both, with Bastion taking priority. Neither will match Forge's release cadence — the systems are longer-horizon by design — but the standstill is over.

Research

Neural Substrate moving toward open source release

Neural Substrate — the biophysical spiking-network tissue model and its embodied 3D navigation environment — is transitioning from internal research to preparation for open source release. Work is currently focused on readying the codebase, documentation, and release infrastructure. Release timing is not yet fixed. Weekly project logs continue at the project page.

Product

Forge 0.2.1 released

A reliability release focused on local-model edge cases. The edit_file tool no longer dead-ends on a failed match — it returns bounded recovery hints (closest match, whitespace diffs, or an occurrence list) so weaker local models can self-correct instead of looping.

Token usage tracking is restored on streaming OpenAI-compatible endpoints (mlx_lm, vLLM, llama.cpp, LM Studio, OpenAI). Without it, /usage and the context footer had been stuck at zero and auto-compaction was silently disabled on long local sessions.

Tool calls that some servers leak as raw text (notably mlx_lm at high context, including Qwen3-Coder's XML dialect) are now recovered into real calls, so the agent stops stalling or returning empty turns in that state. Installer preflight now checks for curl, ripgrep, and bun up front. See the CHANGELOG.

Product

Forge 0.2.0 released

0.2.0 ships streaming reasoning display for chain-of-thought models — a compact "Thinking…" line that settles into a "Thought for Xs" summary when the answer arrives, expandable with Ctrl+T. Works with any OpenAI-compatible endpoint that streams reasoning in a separate field (LM Studio, Ollama, vLLM, mlx_lm, and most cloud reasoning models).

Fixed: compatibility with strict OpenAI-compatible servers like mlx_lm that require the system message to come first — Forge now keeps the leading system prompt and relocates later ones. Ctrl-key shortcuts (Ctrl+F, Ctrl+T) no longer leak their letter into the input field.

Formally removed: the Claude subscription (Pro/Max) OAuth code path. The behavior was first patched out within 0.1.0 on June 22; 0.2.0 is the tagged release that completes the removal — client id, user-agent impersonation, token store, and weekly version self-check all gone. Anthropic is still fully supported via API key. See the CHANGELOG.

Product

Claude subscription login removed from Forge

Forge no longer supports Claude (Pro / Max) subscription OAuth login. Anthropic's terms restrict subscription OAuth credentials to its own applications, so the Claude OAuth code path has been removed entirely from Forge. ChatGPT Codex subscription OAuth is unaffected and remains supported. To connect to Anthropic models, use a direct Anthropic API key — see the CHANGELOG for details. This is a patch to 0.1.0; no version bump was issued because nothing else changed.

Product

Forge 0.1.0 released

Forge is now publicly available. The terminal-based autonomous AI coding agent runs locally against any OpenAI-compatible endpoint, or against Claude Pro and ChatGPT Plus subscriptions via OAuth. Source on GitHub at github.com/Vulkgryph/Forge. Installation and configuration notes live in the repository README. Inquiries: [email protected].

Site

Vulkgryph site published

The Vulkgryph public site is live at vulkgryph.com. It serves as the canonical public index for current projects, research tracks, and development state. Forge remains pre-launch, with public availability targeted for late June 2026.

Site

Vulkgryph public site draft

Initial site structure added: home, research, roadmap, and updates. The public site is structured as a systems index and engineering statement.

Product

Forge listed as first product surface

Forge is listed as local-first AI coding tooling and runtime systems. Public documentation remains pending.

Research

Research tracks separated from products

Neural Substrate and Bastion are presented as research and long-term systems work, separate from the near-term project list.