Cerebretron's AI coding assistant inside VS Code: an authenticated, streaming
chat sidebar powered by the Cerebretron /v1 API.
Status: 0.3 — agent + rich UI. A workspace-aware coding agent with a
polished chat: markdown + syntax-styled code blocks (Copy / Apply to
editor), clean tool cards, and a model selector (Code / Reason / Chat /
Flash / Innovate). It reads your files to answer (read-only tools auto-run),
edits files / runs commands behind VS Code approval dialogs, and taps
Cerebretron's cloud tools (engines, real CalculiX FEA, math). Falls back to
plain chat when no folder is open.
0.4 — inline diff review: agent edits open as a native VS Code diff
(current ↔ proposed) you Accept/Reject; applied via WorkspaceEdit (undoable).
0.5 — deep context: multi-turn memory (it remembers the conversation),
auto-included diagnostics (it sees the file's errors/warnings to fix them),
and @file mentions (reference any file by @path to pull it into context).
0.6 — inline completions: Copilot-style ghost-text suggestions as you
type (Tab to accept), using the fast tier with a fill-in-the-middle prompt.
Toggle via Cerebretron: Toggle Inline Completions.
0.7 — editor power-tools: Cmd/Ctrl+K Cmd/Ctrl+I inline edit (select
code → describe the change → review the diff); "Fix with Cerebretron"
quick-fix on any error/warning + "Edit with Cerebretron" on a selection
(lightbulb / right-click); a status-bar entry to open the chat; and chat
slash commands (/explain, /fix, /tests, /doc, /review).
0.8 — codebase intelligence: a codebase_search tool lets the agent find
where anything is defined/used across the whole workspace (ripgrep, with a
cross-platform JS fallback) before reading specific files.
0.9 — terminal + git: a run_in_terminal tool runs commands (dev servers,
watchers, interactive) in the visible integrated terminal (gated); the
agent uses run_shell for git status/diff/commit + builds/tests where it
needs the captured output.
0.10 — @-picker + history: type @ in the chat for a file
autocomplete dropdown (arrow keys / Enter / click) that pulls a file into
context, and the conversation now persists across reloads (per workspace;
"New" clears it).
0.11 — hardening pass: an adversarial review pass. Diff-apply now re-checks
the file on disk and aborts if it changed since you reviewed (no silent
clobber); the apply prompt lists Reject first so a reflexive Enter never
auto-applies. Cancelled/errored turns no longer leave a dangling history entry,
persisted chat history is validated on load, the @-picker cache invalidates
when files are added/removed, and the markdown renderer is NUL-hardened.
0.12 — control + safety core: Permission modes (status-bar +
Cmd/Ctrl+K Cmd/Ctrl+M): Ask (review each edit), Auto-Edit (apply
without a diff prompt), Plan (read-only — the agent investigates then
drafts a plan you Approve / Edit / Discard before anything runs), and
YOLO (auto-apply + auto-run, confirmation-gated). Checkpoints + Rewind:
every file the agent (or Cmd+K) changes is snapshotted first — hit ⟲ on any
step (or Cerebretron: Rewind Checkpoint…) to revert all file changes from
that step on, while the conversation is kept. (Rewind restores tracked file
edits only — shell/network side effects are not undone.) Live task list:
the agent maintains a pinned todo checklist that updates as it works.
0.13 — onboarding + awareness: Browser sign-in (Sign In → Sign in with
browser) — authorize in your browser, no key to copy/paste (key-paste still
available). A status line in chat shows context %, tokens, daily message
quota, and today's cost (from /v1/usage); the model dropdown is now driven
by your account's available models. A real slash-command framework: type
/ for an autocomplete of built-ins — /explain /fix /tests /doc /review,
/plan, /init (generate a CEREBRETRON.md), /model, /usage, /context,
/compact (summarize to reclaim context), /clear, /help — plus your own
commands from .cerebretron/commands/*.md (filename = name, $ARGS = the rest
of the line; trusted workspaces only).
0.14 — deep context: CEREBRETRON.md project rules — drop a
CEREBRETRON.md in your repo (or .cerebretron/rules/*.md) and the agent
follows it every turn (walks up parent dirs, nearest wins); /init generates
one for you. @-mentions beyond files: @problems (workspace diagnostics),
@git (uncommitted diff), @folder:path (file listing), @symbol:Name
(pull a symbol + its code into context) — alongside @file. Code-
intelligence tools: the agent can now use the language server —
go_to_definition, find_references, document_symbols, workspace_symbols
— for precise navigation instead of guessing from text search.
0.15 — editing power: Parallel tools — the agent can fire several
read-only/search/cloud calls at once (file edits + shell still run one-at-a-
time, in order, each gated), so investigation is much faster. Subagents —
it can spawn_agent to hand a focused subtask to a nested agent with its own
context and get back just the summary (depth-capped; its steps show indented).
Changes panel — every file edited in a run is listed in a Changes card
with per-file Diff / Revert and Revert all (great with Auto-Edit
mode); reverts use the checkpoint snapshots.
0.16 — sessions + extensibility: Multiple sessions — keep several
conversations per workspace; the Chat ▾ button (or Sessions / History…)
switches, renames, deletes, and starts new ones (old single-conversation
history is migrated automatically). MCP client — point Cerebretron at your
own Model Context Protocol servers via the cerebretron.mcpServers setting
(http or stdio); their tools become agent tools (mcp__server__tool).
Full SSRF/DNS-rebind hardening on http (https/no-creds/no-private-IPs/no-
redirects); stdio servers (local processes) load only in a trusted workspace.
Manage with /mcp.
0.17 — rich output + multimodal: Inline rendering — diagram/visual
cloud tools (graphviz, mermaid, latex, image generation, knowledge) are now in
the tool set, and their output renders inline in chat (SVG diagrams,
generated images); SVG is injected under the strict nonce-CSP that blocks any
script. Image input — paste a screenshot into the prompt (or ＋ to
attach a file) and the agent answers about it on a vision turn. Reasoning
— when a reasoning-tier model streams its thinking, it shows in a collapsible
Thinking… block above the answer. (Image input and reasoning light up once
the matching server build is deployed; the client degrades gracefully until
then.)
0.18 — Claude-grade chat surface: tool calls now render as clean
collapsible cards — a ✓ tool detail header you click to expand the
command/result output (read, search, shell, cloud, MCP) — instead of one-line
labels; a turn that's only tool calls no longer leaves an empty bubble. While
a file is being written the chat shows a tidy editing index.html…
indicator (never raw tool JSON), and reasoning collapses to "Thought for Ns"
with the elapsed time.
0.19 — open in the editor (like Claude Code): the chat can now open as a
full editor tab, not just the narrow sidebar — click the ⤢ button in
the chat header, the Cerebretron status-bar item, or run Cerebretron: Open
Chat in Editor. The editor tab and sidebar stay in sync (same session). In the
wide editor the conversation is centered in a readable column.
0.20 — Claude-level visual polish: assistant replies render as clean prose
(no heavy box), user turns as a subtle bubble, with roomier spacing + a gentle
fade-in. A unified rounded composer with a focus ring, an icon attach
button, and a pill Send. Refined code-block chrome (language tag, ghost
Copy/Apply), styled scrollbar, rounded @// picker, and an empty-state
welcome with one-click example prompt chips.
0.21 — MAX mode: a ⚡ Max toggle in the chat bar. When on, the Code
tier runs two top coder models together server-side and a boss synthesizes
the single best answer (slower, higher quality). Opt-in + resilient — if a
model is unavailable it degrades to the single best draft, so Max never breaks
a turn. Other tiers are unaffected.
Each message runs a client-side tool loop (in the extension host): the model
emits one tool call at a time, the extension executes it, feeds the result
back, and loops until the model gives a final answer.
The agent is primed with your active file, selection, and the workspace layout.
All file paths are confined to the workspace (symlink-escape-proof). Brand-
stealth holds — only Cerebretron / genesis-* labels are ever shown.
