A terminal menu that patches your Kimi Code install — reasoning effort, the tool catalogue, transcript window, system prompts, hooks and defaults, plus the way it looks — and can always put it back.
Requirements · Install · What it changes · Undo · How it works · FAQ · Internals
Kimi Code ships as a single binary with its UI, prompts and defaults compiled
in. There is no settings file for most of it. kimi-code-mods opens that
binary, applies the changes you picked in a menu, and puts it back together —
starting from a pristine copy every time, so nothing accumulates and one
command undoes all of it.
The header line is the state check: which Kimi is installed, whether this tool
patched it, and how it is signed. Everything above the divider is a setting —
What it changes goes through them. Below it are the
actions: Apply is the only row that reaches your Kimi at all,
Restore Original Kimi Code puts it back, and the three Open rows drop
you into Kimi's own config.toml, the env-profile.conf the launcher exports
its variables from, or the extracted bundle.js when you want to search for an
anchor yourself.
At a glance
| Verified against | Kimi Code 0.38.0 |
| Patches | 17 JavaScript patches over 25 settings |
| System prompts | 132, extracted as Markdown |
| Tests | 29 checks, about 950 assertions |
| Patching | macOS only — the menu itself runs everywhere |
| Undo | one command, from a frozen pristine baseline |
| Patching (Apply) | Menu | ||
|---|---|---|---|
| macOS | ✅ | ✅ | Reads the Mach-O header of Kimi's binary and re-signs with codesign. |
| Linux / WSL | ❌ | ✅ | Apply refuses — the two steps above are macOS only. |
| Windows | ❌ | via WSL | The PowerShell installer sets things up inside WSL, with the same limit. |
Also needed: python3 (menu), node (patch runner), git, and on macOS the
Xcode command line tools for codesign.
macOS, Linux, WSL
curl -fsSL https://raw.githubusercontent.com/Kirchlive/kimi-code-mods/main/install.sh | bashWindows (PowerShell)
iwr -useb https://raw.githubusercontent.com/Kirchlive/kimi-code-mods/main/install.ps1 | iexInspect before you run it — recommended, this tool rewrites a binary
Piping a script into a shell means running code you have not read, and this one goes on to modify an application you use. Download it first if that matters to you:
curl -fsSL -o install.sh https://raw.githubusercontent.com/Kirchlive/kimi-code-mods/main/install.sh
less install.sh
bash install.shThe installer itself touches no Kimi installation. It checks prerequisites,
puts the repository in ~/.kimi-code-mods, and links kimi-code-mods into
~/.local/bin. Patching happens later, when you choose Apply in the menu.
From a checkout
git clone https://github.com/Kirchlive/kimi-code-mods.git ~/.kimi-code-mods
~/.kimi-code-mods/kimi-code-mods.shThen open the menu:
kimi-code-modsArrow keys move, enter opens, ‹› change a value, esc goes back, q
quits. Nothing is written to Kimi until you pick Apply.
Every switch is on Kimi's own value until you change it. The menu writes to
patch-settings.conf and ~/.kimi-code/config.toml; Apply is what reaches
the binary.
| Lever | What it does |
|---|---|
| Transcript window | How much history goes back with every turn. The one lever that moves running cost directly. |
| Tool setup | Every tool description ships in every request, so an unused tool is a tax on each turn. Turn tools off individually and see what the catalogue weighs. |
| Reasoning | How hard the model thinks, and whether thinking is re-sent with each turn. |
| Group | What it does |
|---|---|
| Complexity effort router | off · pin · free — sets reasoning effort per turn from the prompt itself; pin only ever raises it, never lowers. |
| AGENTS.md alternative names | off · claude · all — have Kimi also read CLAUDE.md and friends. AGENTS.md keeps priority, one file per directory. |
| Toolsets | Named lists of disabled tools, applied in one keystroke, so a "minimal" and a "full" setup are one row apart. |
| Subagent models | Which model each subagent runs on. Needs Kimi's secondary-model flag. |
| Skill setup | Kimi's own product skills, and which directories skills are read from. |
| Hook setup | Run a shell command on any of Kimi's 20 events. |
| Loop control | Attempts per step, and how much context is held back for the answer. |
| Patches | The JavaScript patches themselves: what each one is, whether it applied, and what it did to the bundle. |
Misc — the ten switches belonging to no group of their own
| Setting | Values | What it does |
|---|---|---|
| Command suggestions | default half full |
Height of the slash-command list: Kimi's five, half the window, or nearly full. |
| Working directory /wd | off on |
Adds a /wd command that starts a session in another directory. |
| Click to position cursor | off on |
Place the cursor in the composer with a mouse click. Fullscreen only. |
| Line numbers in Read | on off |
Off saves tokens on every read, and costs the model the ability to cite a line. |
| Expanded by default | off thinking tools both |
Show thinking blocks and tool output unfolded. Costs screen, not tokens. |
| Read limits | default moderate large |
How much one Read returns. Higher trades round trips for context. |
| Auto-accept plans | off on |
Skip the plan approval prompt. A multi-option plan then has no option chosen. |
| Composer border | default off single double bold |
The frame around the input box. |
| Fullscreen renderer | off on |
The alternate screen buffer, patched into the binary so it holds however Kimi is started — not only through a launcher that exports an environment variable. |
| Welcome banner | off on |
Greet with Welcome to Kimi Code Mods! and give the logo its horns. |
Themes — a palette editor over Kimi's 19 colour tokens with a live preview
of a session, and two presets that cannot be deleted: Kimi-Code-Mods (Kimi's
dark palette with the project red) and Default Kimi.
Two spinners, settable apart — the one turning while Kimi thinks, and the one turning while it waits on the model or a tool. Kimi keeps two alphabets and used to change both at once; here they are separate.
Spinners, verbs and message display — every value
Thinking style / Working style
| Setting | Values |
|---|---|
| Spinner shape | default braille dots moon blocks wave glow colors arc star custom |
| Spinner speed | default, or 20–2000 ms per frame |
| Your own frames | any characters, space separated |
| Reverse-mirror run | off on — run the frames there and back, so it swings instead of jumping |
Working style takes the same presets plus follow, which keeps it equal to the
thinking spinner.
Thinking verbs — rotate a word beside the spinner instead of always saying "working": on or off, your own word list, and the format they are drawn in.
User message display
| Setting | Values |
|---|---|
| Your message marker | any prefix; Kimi's own is a sparkle |
| Your message border | off round single double bold topbottom |
| Your message style | default plain italic dim underline strikethrough |
132 system prompts are extracted out of the binary into system-prompts/. Edit
one as plain Markdown and it replaces the original on the next run. A prompt
whose anchor moved in a Kimi release is reported and skipped, never guessed at.
View System Prompts reads them, prices them, and migrates your edits onto
the new text after a Kimi update.
Cost report — what every prompt weighs in tokens, so you can see what you are paying for before you trim it.
There are two, and neither depends on the other.
kimi-code-mods # menu → Restore Original Kimi Code
~/.kimi-code-mods/kimi-patch.sh --restoreThe first thing a patch run ever does is freeze an untouched copy of your Kimi
binary under baseline/. Every later run starts from that copy, not from the
patched one — so patches never stack, and restoring is a file copy rather than
an attempt to undo edits. Your settings survive a restore; only the binary goes
back.
To remove the tool itself:
rm ~/.local/bin/kimi-code-mods
rm -rf ~/.kimi-code-mods # settings, baseline and patches live hereImportant
Re-signing is ad-hoc: the hardened runtime and Apple's notarisation are gone from the patched binary. That is inherent to modifying a signed application, not a shortcut taken here. If your setup requires a notarised Kimi, restore the baseline.
Kimi Code is a Node single-executable application: a JavaScript bundle packed into a Mach-O binary. A run is five steps, and stops at the first one that fails without touching what is installed.
- Freeze an untouched copy under
baseline/, once per Kimi version. - Extract the ~22 MB bundle out of the baseline.
- Apply your prompt overrides, then the patches in
patches/. Each patch locates its anchor and refuses to act if the anchor is missing or ambiguous. - Repack the bundle and re-sign ad-hoc with
codesign. - Verify the result runs and reports the expected version — and roll back to the baseline if it does not.
A run ends with a summary of what took, what had nothing to do, and what was left behind:
Apply summary
patches 12 applied, 5 no-op, 0 failed
prompts 0 applied, 132 unchanged, 0 drifted, 0 anchor missing, 0 rejected
binary repacked, re-signed ad-hoc, verified 0.38.0
result OK — Kimi Code 0.38.0 is patched and installed
Five of the seventeen patches are no-ops on this build: the setting they own is still on Kimi's own value, so there is nothing to splice in. Turn one on and it moves to applied.
Writing your own patch, the anchor rules, and how prompt extraction works are in docs/internals.md.
Verified against Kimi Code 0.38.0: all 17 patches and all 132 prompt overrides find their anchor in this build.
Patches anchor on text inside a minified bundle, and that text moves between
releases. When an anchor is gone the patch reports it and is skipped — the run
carries on with the rest, and nothing half-applied is installed. Prompt
overrides work the same way: a prompt whose text Kimi rewrote is reported as
anchor missing on every run and left alone, rather than forced over the new
wording. kimi-patch.sh --extract-prompts pulls the new text out of the
updated binary, and --migrate three-way-merges your edits onto it.
Kimi replaces its own binary when it auto-updates, which removes every patch. Your settings are untouched — run Apply again and they come back. To stop Kimi updating itself:
export KIMI_CLI_NO_AUTO_UPDATE=1The menu notices when the installed binary is no longer the one it patched and says so in its header.
Kimi dies at startup with no message (exit 137).
macOS sends SIGKILL to a signed binary whose signature no longer matches its
contents. Every run re-signs ad-hoc for exactly this reason; if the binary was
edited by other means, restore the baseline and start again.
The menu runs but Apply refuses.
You are on Linux or WSL. Patching reads a Mach-O header and re-signs with
codesign — both macOS only. Everything else in the menu works.
Kimi updated itself and everything is stock again.
Expected: the update replaced the binary. Your settings survive — run Apply
again, or set KIMI_CLI_NO_AUTO_UPDATE=1 to stop the updates.
A prompt reports "anchor missing".
Kimi moved that text in a release, so the override is skipped rather than
forced over the new wording. Nothing is lost. Re-extract with
kimi-patch.sh --extract-prompts and carry your change across with
--migrate, which does the three-way merge for you. See
docs/internals.md.
Is it safe to run on the Kimi I use every day?
The patched build is verified before it is installed and rolled back if it does not run, and the pristine binary is kept per version and never deleted. What you do give up is notarisation — see What this costs you.
Can I write my own patch?
Yes — drop a .js file into patches/. The anchor rules, the runner contract
and the test harness are in docs/internals.md.
./test.sh # unit checks, a few seconds
./test.sh --full # adds the real binary: Mach-O resize, signing, round-trip29 checks: the patch runner's contract, the auto-repatch guard, the launcher, and the self-checks of every menu module (about 950 assertions in total), including navigation driven through a real pseudo-terminal. The suite works on copies and verifies at the end that it wrote to no real settings file.
tweakcc does this for Claude Code and is where the idea came from — the marker, the bold-cyan selected row and the preview-beside-the-list layout are all lifted from it. This is the same idea pointed at a different binary, with a different container to open.
MIT — see LICENSE.