Large Update: Rate Limit Radar, Project Hooks, and Safer Shells

[zemaj]

Since September 16, 2025 we’ve poured a mountain of improvements into Code - and as of the latest main build (September 22) they’re all live.

---

Terminal

Code’s terminal overlay has grown into a true command cockpit. Type $ <command> to run it instantly in a dedicated shell window, or use $$ <task> when you want Code to draft the command first so you can approve, edit, or reject it. The overlay keeps native scrollback, mirrors every run into history, and still pipes through Code’s guardrails so you get human-level control with AI-level safety.

• Manual runs now show richer titles (“Shell: ”) and display truncated previews so you can spot the exact command before sending it.
• Guided sessions keep controller handles alive for reruns and cancellations, and every background status update (start, streaming chunks, completion) is logged so nothing goes missing.
• Auto-run commands suggested by Code are echoed back in the overlay via the new “surface update notice” plumbing, making it crystal clear what just executed when you let Code drive.

/limits

The /limits command evolved into a full dashboard that actively looks out for you. It renders synchronized hourly and weekly gauges, stores the official reset timestamps on disk so sessions resume accurately, and keeps the session footer in lockstep with reality. As you cross each threshold (50%, 75%, 90%), Code now raises in-chat alerts - “Hourly usage exceeded 75% of the limit. Run /limits for detailed usage.” - so you can throttle or swap models before the API responds with errors.

• Hourly windows now display the exact top-of-hour reset after the latest fixes, while weekly usage shows the precise day/time Code will reset the counter.
• The view offers both a compact summary row for quick glances and an expanded grid with detailed counts, deltas, and projected remaining capacity.
• Rate-limit snapshots stream in from RateLimitSnapshotEvent and refresh automatically in the background, so the gauges stay correct even during long-lived sessions.

Hooks

Project automation became dramatically more capable. In config.toml you can declare hooks for session.start, session.end, tool.before, tool.after, file.before_write, and file.after_write. Each hook runs in the same sandbox as your session, outputs to its own exec cell, and receives rich metadata via CODE_HOOK_EVENT, CODE_HOOK_PAYLOAD, CODE_SESSION_CWD, and more. Failures are surfaced but never block your main task.

• You can string multiple hooks together (e.g., lint after every apply_patch, post notifications when a session shuts down) and specify per-hook timeouts, custom working directories, or environment overrides.
• Project-scoped commands assigned to /cmd <name> now show up right in the slash menu, giving you one-keystroke access to smoke tests, migrations, bootstrap scripts, or deploy pipelines tailored to each repo.

Validation

The validation harness received a full control panel. The /validation command now exposes a polished interface where you can toggle the entire patch harness or specific tools - including the new cargo check validator - without editing config files. If Code detects that a required validator (say, shellcheck or actionlint) is missing, it launches the terminal overlay with the recommended install command prefilled, so you can accept it, watch progress, and retry immediately.

• Validation runs capture their outputs as structured exec cells, making it effortless to skim warnings, copy command lines, or rerun with tweaks.
• Background completions arrive automatically, and the harness now keeps defaults disabled until explicitly switched on, preventing accidental slowdowns on new installs.
• Compile blockers in the validation pipeline have been cleaned up, so enabling the new cargo check path works out of the box.

TUI & Performance

Rendering in the TUI feels dramatically faster and clearer. History cells cache layouts and parsed metadata, wait-tool chatter collapses into the final exec output, and every run stamps its duration so you can skim outcomes at speed. The status panel clears spinners once final answers land, bottom overlays clamp to the viewport, and the /update workflow now surfaces notices plus echoes the exact command that ran on your behalf.

• /limits has a refreshed layout with aligned color palettes, synchronized headings, and off-main-thread refresh logic to keep the UI smooth.
• The composer footer now reflects rate-limit snapshots in real time, while the history gutter keeps the right icons visible even after merges, validations, or agent handoffs.
• Update flows gained inline confirmations so you’re never left guessing whether an auto-upgrade or managed installer actually fired.

CLI & Automation

Command-line ergonomics leveled up. code resume once again restores paused sessions perfectly, shell completion scripts now reference the code shim everywhere, and scripts/wait-for-gh-run.sh can follow any GitHub Actions run - optionally dumping logs for failed jobs with a single flag. Logging defaults were tuned (WARN on file logs) to reduce noise without losing signal.

• A new code mcp subcommand exposes the MCP server directly, with code acp acting as a friendly alias for the same functionality.
• The npm package now ships the Rust code-mcp-server binary automatically, so installing @just-every/code gives you the full MCP/ACP toolchain straight away.
• MCP initialization tolerates numeric protocolVersion fields, making it easier for other clients to connect without fiddly patches.

ACP + Zed

ACP integration is officially live. The Rust MCP server advertises acp/new_session and acp/prompt alongside Code’s classic tools, so Zed (or any ACP-capable IDE) can drive Code without touching the terminal. The new CLI aliases (code acp, code mcp) make setup trivial, and the documentation now covers minimal configs for Code, Coder, and Zed.

• docs/integration-zed.md and the refreshed ACP docs walk through wiring Zed’s MCP transport into Code, including environment variables, permission flows, and launch commands.
• Code preserves confirm guards, sandbox policies, and validation harnesses while forwarding every event over ACP, so safety and automation stay exactly as you expect.

---

Ready to experience all of this?

npm install -g @just-every/code@latest

Whether you’re spinning up the terminal overlay, exploring /limits, or driving Code from Zed via ACP, we can’t wait to see what you build next!

[zemaj]

Oh, one thing we didn't mention - there's now an /update command. You can use it to enable auto-updates. Once it's battle tested auto-updates will be enabled by default.

[cbusillo]

In the current Validation harness, can we outsource the checks to a MCP? I have a custom MCP for Odoo that uses Jetbrains's Inspection interface. This is due to a few of the tools I have in PyCharm being able to work with Odoo's dynamic system.

BTW, great work, I am a fan.

[zemaj]

Hmmm interesting idea. Short answer is no, but should be relatively straightforward. We now have a hook system and an MCP system, so just would be a case of adding a hook when the validation would have been called and get it to talk to the MCP.

Will take a look next time I revisit this system.

[cbusillo]

Thank you for the answer :)