A drop-in harness that makes Claude Code cheaper, sharper, and forgetful-proof: persistent memory, token savings, anti-drift, a code-map RAG, disciplined cheap-model coding.
Agent
The AUDITOR (codex voice). THE PROCEDURE step 6 (VERIFY), routed to Codex per Ro's directive 2026-08-02 ("queremos usar Codex más porque nos dan más créditos"). Reads a diff or finished unit against its spec, reports findings, and stops. It NEVER edits and NEVER spawns anything — this agent has no Write/Edit tools at…
Agent
The BUILDER. THE PROCEDURE step 5 (BUILD) — dispatches ONE unit to the real synchronous Codex CLI (codex exec), which edits files on disk, and returns the diff. This agent has NO file-editing tools itself; it cannot write code in Claude even by accident. This is NOT codex:codex-rescue — that plugin agent is a…
Agent
Delegates work to REAL Gemini 2.5 (Pro/Flash) via the gemini shim → OpenCode. A genuine independent model, so valid as the non-Claude voice in a council. Gemini gets its OWN Read/Bash/Edit tools inside OpenCode, so it can scrape, run commands, read repos, and analyse images — not just answer from context. Use for…
Agent
Runs after a unit of work is built and verified, to make every code map tell the truth about the repo again. THE PROCEDURE step 7 (PERSIST), before the commit. It regenerates the CACHED maps (.codemap, graphify) and VERIFIES the LIVE ones (L0/L1/skeleton) — those are computed on every call and have nothing to…
Agent
The AUDITOR. THE PROCEDURE step 6 (VERIFY) — a non-builder reviews the builder's work before it is accepted. Reads a diff (or a finished unit) against its spec, reports findings, and stops. It NEVER edits and NEVER spawns anything. Use for: auditing a codex diff; checking a unit against its CONTEXT/CHANGE/GOAL/VERIFY…