dsh-llm-key-rotation

Seamless API-key rotation for DeepSeek Harness.
Hit a quota limit or 429 Rate Limit? The plugin swaps the dead key with a fresh spare and retries instantly — no restarts, no context loss, and zero interruptions.

---

⚡️ Highlights

• 🔄 Invisible to Agent & User: Request fails with quota/rate limit → plugin hot-swaps the key → request retries and succeeds seamlessly.
• 🎛 Native Web UI: Add, manage, and toggle spare keys right inside Settings.
• 🧠 Smart Anti-Spin: Cooldown discipline prevents infinite loops if all your keys are exhausted.
• 🧼 Zero Core Patches: Plugs cleanly into the Harness recovery waterfall — safe to install, safe to remove.

---

🛠 How It Works

┌─────────────────┐          429 / QUOTA / AUTH          ┌──────────────────────┐
│  Model Request  │ ───────────────────────────────────► │  llm-key-rotation    │
└─────────────────┘                                      └──────────┬───────────┘
         ▲                                                          │
         │                  { kind: 'retry' }                       │ 1. Pick next spare key
         │             (Adapter re-resolves key)                    │ 2. Hot-swap env ref
         └──────────────────────────────────────────────────────────┘

1. Your primary key works as usual.
2. If a request fails (QUOTA, RATE_LIMIT, or AUTH), the plugin loads the next spare key from your chain.
3. The turn retries with the new key, and future requests continue with the working key.

---

🚀 Quick Start

1. Install the plugin

dsh plugin --profile web add @m1khal3v/dsh-llm-key-rotation

2. Configure via Web UI

Navigate to Settings → Plugins → Key Rotation:

1. Toggle on rotation for your target provider.
2. Select your trigger codes (default: QUOTA, AUTH).
3. Click Add key, paste your spare keys, and hit Save.

(Keys are safely saved to the Harness credential store).

---

⚙️ Behavior & Good to Know

• Smart Rotation Window (300s): During consecutive failures, the plugin walks forward through your spare keys. If no failures occur for >5 minutes, the chain resets to start from the top again.
• Interplay with dsh-llm-retry:
  • QUOTA and AUTH rotate immediately.
  • RATE_LIMIT is handled by dsh-llm-retry first (with backoff). To rotate instantly on 429 errors instead, simply remove RATE_LIMIT from your provider's retryableCodes.
• Live Terminal Logs: Watch rotation events in real time right in your dsh terminal:

  [llm-key-rotation] rotated provider="opencode-go" chain[0]→"OPENCODE_GO_API_KEY" (QUOTA)
  

  (Secret values are never logged — only provider names, indices, and error codes).

---

🧑‍💻 Development

pnpm install
pnpm run verify    # typecheck + test
pnpm run build

📄 License

MIT