Kimi CLI — SKILL.md

---
name: "Kimi CLI"
description: "Fleet skill: Kimi CLI — software inventory and operations reference"
version: "1.0.0"
author: "skynet"
category: "fleet"
agents: ["claude-code", "codex", "gemini", "kimi"]
tags: ["software-kimi", "fleet", "fleet", "software"]
---

# Kimi CLI

---
name: software-kimi
description: Guidance for using Moonshot's Kimi CLI (K2.5 model) for autonomous coding, refactoring, and fleet-wide task execution across James's infrastructure.
metadata:
  author: skynet
  version: 1.0.0
---

# Kimi CLI (software-kimi)

Moonshot's high-performance coding CLI, optimized for large-context operations (256K) using the K2.5 model. This tool is a primary driver for complex refactoring and deep codebase analysis within the fleet.

## Fleet Installation Status

| Machine | IP | Status | Version | Method |
|---------|----|--------|---------|--------|
| **Dev Workstation** | 192.168.86.22 | Installed | 1.11.0 | Binary |
| **Mac Mini (Bots)** | 192.168.86.50 | Installed | 1.21.0 | Binary |
| **Mac Mini (Jarvis)** | 192.168.86.51 | Installed | 1.21.0 | Binary |
| **Mac Mini (Vault)** | 192.168.86.27 | Installed | Latest | `brew install kimi-cli` |
| **DGX Spark** | 192.168.86.48 | NOT INSTALLED | - | - |
| **Dev1** | - | NOT INSTALLED | - | - |
| **Server** | 192.168.86.33 | NOT INSTALLED | - | - |

> **Note:** For machines without Kimi CLI installed, use `ssh vault` or `ssh bots` to trigger remote jobs, or use the LiteLLM gateway on Spark (Port 8000) if raw API access is sufficient.

## Configuration & Authentication

- **Primary Auth Path:** `~/.kimi/credentials`
- **Config File:** `~/.kimi/config.json`
- **Model:** `k2.5` (Default)
- **Context Window:** 256,000 tokens

### Credentials Management
On the **Vault Mac**, credentials are managed via the secure keychain. For other machines, ensure the `credentials` file is synced from the Vault via the `infra-sync` script if tokens expire.

```bash
# Verify authentication status
kimi auth status

# Refresh token from Vault (from local workstation)
ssh vault "cat ~/.kimi/credentials" > ~/.kimi/credentials
```

## Key Commands

### General Usage
- `kimi chat`: Start an interactive session with K2.5.
- `kimi code [path]`: Open a codebase-aware session for the specified directory.
- `kimi exec "command"`: Run a specific prompt against the current context.
- `kimi config list`: View current runtime configuration.

### Advanced Operations
- `kimi scan`: Analyzes the current directory to build a temporary local index for better RAG performance.
- `kimi token-count [file]`: Calculate token usage before sending large files to the 256K context.

## Common Workflows

### 1. Deep Codebase Analysis
Kimi's 256K context is ideal for tasks that span multiple modules.
```bash
# Ingest the entire src directory for a logic audit
kimi code ./src --prompt "Analyze the dependency flow between the auth module and the database layer. Identify potential circularities."
```

### 2. Fleet-Wide Batch Refactoring
Use Kimi in conjunction with `fleet-build.sh` to propagate changes.
```bash
# Example: Updating all scripts in ~/dev/projects/skills/scripts
for file in ~/dev/projects/skills/scripts/*.py; do
  kimi exec "Update the logging pattern in $file to use the new Skynet standard" --write
done
```

### 3. Remote Execution via Vault
When working from a machine without Kimi installed (like Spark):
```bash
ssh vault "cd ~/dev/target-repo && kimi exec 'Generate unit tests for all functions in utils.py' --write"
```

## Fleet-Specific Patterns

### The "256K Dump" Strategy
Because Kimi handles 256K tokens, avoid aggressive `grep` filtering if the project is under 100MB. Instead, provide the whole context to allow the K2.5 model to see the "big picture."

### Model Overrides
The K2.5 model is the fleet default. If a task requires higher reasoning but lower context, check if `k2.5-pro` is available in `kimi config models`.

### Output Redirection
Kimi's CLI often outputs markdown. For automation, use the `--raw` flag when available to avoid stripping code blocks manually.

## Troubleshooting

### Token Limit Errors
Even with 256K, massive node_modules or build artifacts can cause overflows.
- **Fix:** Ensure `.kimiignore` or `.gitignore` is present. Kimi CLI respects these by default.
- **Manual Exclusion:** `kimi code . --exclude "dist/*,logs/*"`

### Authentication Failures (Error 401)
Usually caused by an expired session on the non-Vault machines.
- **Action:** Run `kimi login` or pull the latest credentials from the Vault Mac.

### Version Mismatch
The Workstation is on `1.11.0` while the Macs are on `1.21.0`.
- **Known Issue:** `1.11.0` lacks the `kimi scan` command.
- **Workaround:** For indexing-heavy tasks, delegate the execution to `ssh bots`.

## Documentation & Resources
- **Official Docs:** [https://github.com/anthropics/kimi](https://github.com/anthropics/kimi)
- **Internal Specs:** See `~/infra/docs/kimi-integration.md` on the Workstation.
- **Support:** Reach out to James (skynet) for API quota increases on Moonshot platform.

## Best Practices for Agents
1. **Always verify context:** Run `kimi token-count` if you suspect the file set is large.
2. **Prefer `--write`:** When instructed to modify files, use the `--write` flag to avoid manual copy-pasting of diffs.
3. **Atomic Commits:** After a `kimi exec --write` operation, immediately run `git diff` to validate changes before proceeding.
4. **Machine Selection:** Use **Bots** or **Jarvis** for long-running batch jobs to keep the **Workstation** responsive.