egm-review-loop

1# Goldfish Review: Loop (Orchestration Driver)
2 
3When running in Codex or ChatGPT Work, first read [Codex integration](../../docs/codex.md).
4In Claude Code or Cowork, read `${CLAUDE_PLUGIN_ROOT}/docs/claude.md` first.
5 
6## Overview
7 
8This skill acts as the central orchestration engine driving the multi-agent
9review loop. It coordinates two distinct agent memory models:
10 
11*   **The Goldfish (Adversarial Auditors)**: Subagents spawned in isolated,
12    fresh context windows to catch unstated assumptions without historical bias.
13*   **The Elephant (Persistent Resolver)**: The primary agent context holding
14    institutional knowledge to evaluate findings and apply approved repairs.
15 
16The driver manages argument intake, resolves paths for reviewers and standards,
17unconditionally re-parses target line ranges before every iteration, spawns
18concurrent specialized Goldfish auditors based on configuration, and coordinates
19with the Elephant context to resolve findings.
20 
21## Delegation
22 
23A Goldfish runs in a fresh context with no conversation history. Spawn each
24subagent with your harness's subagent tool (Antigravity: `invoke_subagent`; Claude Code: `Agent`; Codex: `spawn_agent`). Each subagent is defined in `agents/<file>.md`, and its
25frontmatter `name` (for example `egm-zoom-out`) is the subagent type to request
26where your harness registers named subagents. Otherwise spawn a general-purpose
27subagent and tell it to read that persona file first. Pass the prompt verbatim
28…