Skip to main content

Documentation Index

Fetch the complete documentation index at: https://langwatch.ai/docs/llms.txt

Use this file to discover all available pages before exploring further.

Aider is a CLI pair-programmer that edits files in your repo via LLM round-trips. It supports any OpenAI-compatible endpoint via --openai-api-base, which is exactly what the LangWatch AI Gateway exposes.

​
Setup

export OPENAI_API_KEY="lw_vk_live_01HZX..."
export OPENAI_API_BASE="https://gateway.langwatch.ai/v1"

aider --model gpt-5-mini
For Anthropic Claude models, Aider uses litellm under the hood — it will route anthropic/claude-* model names through its Anthropic handler, which uses ANTHROPIC_BASE_URL: 
export ANTHROPIC_API_KEY="lw_vk_live_01HZX..."
export ANTHROPIC_BASE_URL="https://gateway.langwatch.ai"

aider --model anthropic/claude-sonnet-4-6
This hits the gateway’s /v1/messages endpoint, which is Anthropic-native and preserves cache_control blocks byte-for-byte (load-bearing for Aider’s repo-map caching — see below).

​
Cache-aware repo maps

Aider sends your repo map as a long system prompt (often 20–50k tokens). Anthropic caches the repo map with cache_control: {type: "ephemeral"}, giving a 90% discount on cache hits. The gateway preserves these cache markers on /v1/messages, so your Aider session pays full price once and discounted price on every follow-up in the 5-minute cache window. For this to work:
  1. Use /v1/messages (Anthropic shape), not /v1/chat/completions.
  2. Set VK cache.mode: respect (the default). Do NOT use cache.mode: disable with Aider — it 10×s your Claude bill.
See Caching Passthrough for the complete contract.

​
Trace propagation

Aider doesn’t natively emit a trace, but you can thread one manually: 
python -c "import uuid; print('00-{:032x}-{:016x}-01'.format(uuid.uuid4().int, uuid.uuid4().int >> 64))" > /tmp/aider-trace
export OPENAI_EXTRA_HEADERS="{\"traceparent\": \"$(cat /tmp/aider-trace)\", \"X-LangWatch-Thread-Id\": \"aider-$(date +%s)\"}"
OPENAI_EXTRA_HEADERS passthrough works with openai Python SDK ≥ v1.30. Aider versions pinned to older SDKs ignore it.
Every request from that Aider session shares a trace id; the gateway nests its span under it and no cost is double-attributed. See SDKs → trace propagation for the generator helper.

​
Governance recipes

​
Per-engineer Aider budget

  • Scope: principal, target: the engineer.
  • Window: day, limit: $25.
  • on_breach: block (Aider errors out cleanly — better than a $400 mistake).

​
Block destructive tools

Aider uses a tool-based interface for applying diffs. If you want to route Aider through a read-only VK (for demo purposes):
  • VK policy_rules.tools: ["shell", "run"].
Aider’s /run command returns 403 tool_not_allowed and the agent surfaces the error.

​
Fallback to a cheaper model on 429

Chain: openai-primary → openai-cheap-backup (same provider, different-region key with a lower quota). When OpenAI rate-limits your primary, Aider keeps working at slightly degraded quality instead of erroring. See Fallback Chains.

​
Troubleshooting

  • Aider complains “invalid_api_key” — verify you exported OPENAI_API_KEY with the VK value, not your real OpenAI key.
  • ANTHROPIC_BASE_URL ignored — Aider’s litellm handler needs litellm ≥ v1.40; earlier versions hardcode api.anthropic.com.
  • Cache misses on repo maps — confirm your model is anthropic/* (uses /v1/messages) and VK cache.mode is respect. disable defeats the Anthropic discount.
  • Gateway returns 413 payload_too_large — Aider can send 100k+ token prompts on large repos. Check the gateway’s LW_GATEWAY_MAX_BODY_BYTES setting (default 10MB).