/advisor mode: Open-source Python coding agent that pairs a cheap worker model with an expensive reviewer at decision points (no need to pay Opus rates for the whole session)

Most agent CLIs make you pick one model — Opus is great but burns money, Haiku is cheap but misses the architectural calls. This Claude Code feature is wired in an /advisor mode that pairs both in an open source project called ClawCodex. You can search it in github or see the discussion thread after this post for the link. How it works: a cheap worker (e.g. haiku-4-5, or deepseek-v4-pro) does the grinding — file reads, edits, test runs. At decision points (before committing to an interpretation, before declaring done, when stuck) the worker pauses and consults a stronger reviewer (e.g. opus-4-7). The reviewer sees the entire conversation — every tool call, every result — and returns short Gaps / Risks / Do-next advice. Then the worker continues. Net cost on typical sessions is several-fold lower than running Opus end-to-end, without losing the architectural judgment on the calls that matter. Two execution modes under the hood: \- Server-side (Anthropic 1P): advisor beta header — one roundtrip, prompt-cache friendly. Worker + advisor both on Anthropic. \- Client-side (any provider): worker emits a regular tool\_use, the agent intercepts and makes a separate call to the configured advisor model. Two roundtrips, but you can mix providers — e.g. DeepSeek worker + Claude Opus advisor, or Gemini worker + GLM advisor. Config is one line in the REPL: /advisor anthropic:claude-opus-4-7 /advisor deepseek:deepseek-v4-pro Status bar shows worker tokens, advisor tokens, and USD cost separately so you can see where the spend is going. It's part of a Python port of Claude Code with native support for Anthropic, OpenAI, Gemini, DeepSeek, GLM, Minimax, OpenRouter. On SWE-bench Verified the agent scores 58.2% on Gemini 2.5 Pro vs openclaude's 53% under the same harness. The actually-hard part was getting the advisor prompt to STOP restating the worker's plan back at it — early versions burned the worker's context on echoes. The fix was a hard "no first-person voice, no echoes" rule plus a Gaps / Risks / Do-next template. Happy to dig into the prompt design if anyone's curious. Source link in a comment below.