--- status: current version: 0.4.0 canonical: true audience: [ai, human] last_updated: 2026-07-01 --- # Origin This file records the problem EveGlyph Editor addresses and the idea it is built around. It is grounded only in the project's own README and PROGRESS notes — no founding narrative beyond what those documents state. ## The problem: an agent that edits your files, and a human who wants to stay in control AI coding and writing agents increasingly do not just suggest text — they reach onto disk and rewrite files directly. That capability is useful, but it creates a control problem for a human working in a plain Markdown workspace: - An **autocomplete-style** AI writing tool keeps the human in the loop by construction — it proposes the next few words or a paragraph, inline, and the human accepts or ignores each suggestion as they type. Control is fine-grained, but the tool cannot do much unsupervised work: it does not restructure a document, touch other files, or carry out a multi-step edit across a workspace. - A **local CLI agent** (Claude Code, Codex, Gemini) can do that larger-scope work — read a workspace, plan an edit, and write it — but the moment it is allowed to touch files on disk without a human approving each individual change, the human's sense of what actually happened to their documents gets weaker exactly as the agent's capability gets stronger. EveGlyph Editor exists in the space between those two: a human writing clean Markdown wants an agent that can do real, multi-file, disk-level work, without losing the ability to see — and undo — exactly what changed. ## The answer: the workspace ↔ agent ↔ diff-review ↔ human loop EveGlyph Editor's own framing of its north star is a single closed loop: > humans write clean Markdown, local CLI agents edit files on disk, every > change surfaces as a reviewable git diff. Concretely: before a local agent runs, the workspace is git-snapshotted (a repo is initialized if one does not exist). The agent then runs with auto-approve inside that one workspace folder — it can read, create, edit, and delete files there without asking per file. Afterward, EveGlyph Editor does not just say "done" — it shows a real `git diff`, grouped into per-file cards with +/− counts, and the human makes the final call: **Accept** commits the agent's changes (`agent: `), **Reject** runs `git reset --hard HEAD` plus `git clean -fd`, discarding all of the agent's edits and any untracked files. This is a different shape of control than autocomplete. The human is not approving every keystroke; the human is reviewing every *change set*, after the fact, with the full diff in front of them, and the decision is binary and reversible up to that point. The README states this trade-off plainly rather than hiding it: local-agent mode runs a CLI with auto-approve, and the project's answer to the risk that creates is the diff-review step, not a promise that the agent will never touch a file without asking first. The front stage stays deliberately minimal — an editor, a file tree, a settings panel — while the capability (the agent, the git snapshot/diff machinery, the workspace memory that steers the agent) lives backstage. That front-stage-minimal, back-stage-strong shape is itself part of the answer to the problem: a Markdown workspace should not look or feel more complicated just because a capable agent is standing behind it. ## Why this needed its own format, not just an editor The same review loop needs the documents themselves to carry a small amount of machine-legible structure — what kind of document this is, its status, its tags — so an agent's context and a human's filter/search both have something honest to key off. That is what EveGlyph-MD (`EG-MD-2026`) is: a lightweight `type` / `status` / `tags` frontmatter classification layer, and EveGlyph Editor is described in its own README as "the editor half" of that format. ## See also - [current.md](./current.md) — what v0.4.0 actually ships today. - [design-history.md](./design-history.md) — the v0.1 → v0.4 version lineage. - [public-summary.md](./public-summary.md) — a short citable summary.