@madgagarin/pi-agentrouter

Pi Coding Agent extension for AgentRouter. Adds support for DeepSeek V4 Flash, Claude Opus 5, GPT-6 Astra, and GLM 5.3 with in-flight WAF auto-recovery, automatic retry orchestration, prompt caching optimization, and live pricing.

AgentRouter provides free trial credits on registration at agentrouter.org.

Features

• Multi-model access with one key: Use DeepSeek V4 Flash, Claude Opus 5/4.8, GPT-6 Astra, and GPT-5.6 Sol through a single endpoint. No need to manage separate billing accounts across OpenAI, Anthropic, and DeepSeek. Switch models in Pi with Ctrl+P.
• Automatic error recovery (no crashed runs): Intercepts transient gateway errors (rate limits, 500/503, thinking-mode 400/422 validation issues) and retries up to 10 times with exponential backoff and unpinned session headers. Long autonomous tasks continue running instead of aborting midway.
• Prompt caching optimization (~90% cost reduction): Enforces strict prefix stability across requests so DeepSeek and Claude prompt caches hit ~95% in multi-turn sessions, significantly reducing input token costs.
• Context compaction cleanup: Automatically strips raw tool output dumps and internal thinking traces when running /compact, shrinking the payload by 80–90% without losing instructions.
• Live batch quota probe (/agentrouter check): Premium models (Claude Opus, GPT-6) use periodic batch quotas on AgentRouter. Instead of guessing or getting 402 errors mid-run, run /agentrouter check to test live availability across all models in parallel (🟢 Ready vs ⏳ Exhausted) and inspect your total monthly spend in USD.
• In-flight WAF auto-recovery & false-positive bypass: Transparently catches upstream content filter triggers (400 content-blocked, 500 sensitive words detected, 405 challenge), proactively sanitizes raw binary dumps (ELF executables, null-byte sequences), performs non-destructive tool result redaction while strictly preserving tool call arguments and assistant history, and frames prompts with language preambles without losing user instructions or breaking context.
• Spend tracking & pricing sync: Queries current rates from the gateway on startup, updates settings automatically, and keeps model catalogs in sync.
• Granular provider rate pacing: Coordinates requests across concurrent subagents via local lock files (~/.pi/agent/.agentrouter-pacing), isolated per provider so that non-AgentRouter models (Anthropic, Gemini, Ollama) are never throttled.
• Isolated credentials: Stores keys in agentrouter-* namespaces in auth.json without modifying default provider keys.

Quick Start

1. Get an API key

Create an account at agentrouter.org and copy your sk-... key from the dashboard.

2. Install the extension

pi install npm:@madgagarin/pi-agentrouter

3. Set your key in Pi chat

/agentrouter key sk-your-agentrouter-key

Or export it in your shell profile:

export AGENTROUTER_API_KEY="sk-your-agentrouter-key"

Models & Pricing

Rates are pulled dynamically from the gateway API ($2.00 / 1M tokens base unit):

Claude and GPT quotas are released in batches throughout the day. When a batch is consumed (HTTP 402), check availability with /agentrouter check or switch to unlimited models (deepseek-v4-flash / glm-5.3).

In-Chat Commands

Keybindings

Recommended settings.json

Add this to ~/.pi/agent/settings.json for quick model switching:

{
  "defaultProvider": "agentrouter-openai",
  "defaultModel": "deepseek-v4-flash",
  "defaultThinkingLevel": "low",
  "enabledModels": [
    "agentrouter-openai/deepseek-v4-flash",
    "agentrouter-openai/gpt-6-astra",
    "agentrouter-openai/gpt-5.6-sol",
    "agentrouter-clode/claude-opus-5",
    "agentrouter-clode/claude-opus-4-8"
  ]
}

FAQ

How does error recovery work?

The extension intercepts requests at the transport level. If the gateway returns a temporary error (such as a 429 rate limit, 405 WAF challenge, 500/503 service issue, thinking validation mismatch, or WAF content-blocked error), it strips sticky routing headers to failing nodes, proactively strips binary dumps, performs non-destructive in-flight poison redaction on tool results while preserving all tool arguments and assistant turns, waits with exponential backoff, and retries up to 10 times. Diagnostic events are logged to ~/.pi/agent/.agentrouter-debug.log.

Why does prompt caching matter?

DeepSeek charges ~$0.014 per 1M cached input tokens versus ~$0.14 for uncached ones. Because this extension prevents dynamic prompt rewrites and ensures prefix determinism, multi-turn coding sessions routinely achieve 95%+ cache hit rates, saving substantial costs over long runs.

How do batch quotas and /agentrouter check work?

AgentRouter releases daily quotas for premium models (Claude Opus 5, GPT-6 Astra) in periodic batch drops. When a batch is fully consumed, the API returns HTTP 402. Running /agentrouter check sends lightweight live health probes to all models in parallel, immediately displaying which models are Ready (🟢 200 OK) or Exhausted (⏳ 402), along with your current monthly spend in USD. If a premium batch is empty, switch to deepseek-v4-flash or glm-5.3 which have unlimited capacity.

Does pacing slow down other providers?

No. Request pacing is isolated per provider via dedicated lock files and applied exclusively to requests routed to agentrouter.org. Direct provider connections (Google Gemini, Anthropic, local Ollama) run without pacing delays.

Using custom subagents (pi-subagents)

AgentRouter validates requests against the pi-code signature. If you define custom subagents in ~/.pi/agent/agents/*.md, set systemPromptMode: append in their frontmatter.

License

MIT © madgagarin