CORAL CLAUDE.md

A project guide for CORAL, a system that runs coding agents in separate Git workspaces, scores their changes, and repeats the process. It explains the project layout, tools, and evaluation loop.

In plain words
What is it for?
Use it when developing or operating CORAL, configuring agent workspaces, running evaluations, checking scores, or understanding its technology and directory structure.
Why use it?
It gives agents the shared context they need to work consistently in an unfamiliar repository. It also explains how changes are submitted and graded.

Instructions file

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add instructions/human-agent-society/coral/claude-md
Clone the repo
git clone --depth 1 https://github.com/Human-Agent-Society/CORAL
Per session 4,578 This file is loaded in full into every session.
When invoked 4,578 The same file — it is already loaded in full.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.04578 $0.04578
Opus 5 $0.02289 $0.02289
Sonnet 5 $0.00916 $0.00916
Haiku 4.5 $0.00458 $0.00458

Measured yesterday against content hash 588d2cc7b2cf, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

CORAL CLAUDE.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

CLAUDE.md · 220 lines

How it starts

The opening of the file, as written. The whole thing — 220 lines — stays where its author put it; the contents beside it link to each section on GitHub.

CORAL

An orchestration system for autonomous coding agents — agents follow a CORAL.md guide, run experiments, share knowledge, and loop forever.

Project Overview

Core pattern: Spawn agents → agents read CORAL.md → commit changes → grader daemon scores them → repeat

Key concepts:

  • Agents are the optimizers — Claude Code (or Codex / Cursor / Kiro / OpenCode) subprocesses, each in its own git worktree.
  • Shared state via .coral/ — split into public/ (visible to agents through a runtime-specific symlink like .claude/, .codex/, .opencode/) and private/ (grader venv, hidden inputs — denied to agents). The grader's own source is surfaced read-only to agents as <shared_dir>/grader/ (a symlink to the real grader/ package) so they can read how they're scored.
  • Async eval loopcoral eval -m "..." stages+commits and writes a pending attempt; a long-running grader daemon picks it up, grades it inside a detached worktree, and writes the final score back. Default behavior blocks until the score lands; --no-wait returns immediately.
  • CLI orchestration — 18 commands (see Commands below), grouped under coral start / status / eval / log / ....

Directory Structure

Directory Purpose
coral/types.py Core types: Task, Score, ScoreBundle, Attempt
coral/config.py OmegaConf-backed YAML configuration (CoralConfig, GraderConfig, AgentConfig, GatewayConfig, WarmStartConfig, HeartbeatActionConfig, ...)
coral/task/ Frontend-independent task validation with structured reports, progress events, and baseline grading
coral/agent/ Agent lifecycle: manager.py (multi-agent supervisor), runtime.py (abstract), state.py, heartbeat.py, exit_classifier.py, warmstart.py, process.py, registry.py
coral/sandbox/ Pluggable agent sandboxing (agents.sandbox): protocol.py (SandboxProvider + spec/context types), registry.py (name or module:Class entrypoint resolution), srt.py (built-in srt provider: OS-level FS/network enforcement + allow-all proxy)
coral/agent/builtin/ Concrete runtimes: claude_code, codex, cursor_agent, kiro, opencode
coral/grader/ Grader stack: protocol.py, base.py, task_grader.py, loader.py, subprocess_grader.py, daemon.py (long-running grader), builtin/function_grader.py
coral/hub/ Shared state: attempts.py, notes.py, skills.py, checkpoint.py (git-tracked snapshots of .coral/public/), heartbeat.py, prompts/ (built-in heartbeat prompts)
coral/hooks/ post_commit.py — implements submit_eval (called by coral eval)
coral/workspace/ Run layout: project.py (run dir setup), worktree.py (per-agent git worktrees + symlinks), repo.py (clone/init), grader_env.py (.coral/private/grader_venv/)
coral/template/ coral_md.py + coral.md.template / coral_single.md.template; bundled agents/ (deep-researcher, librarian) and skills/ (deep-research, organize-files, skill-creator) seeded into every run
coral/gateway/ Optional LiteLLM gateway (server.py, middleware.py, config.py) for intercepting agent model traffic
coral/web/ Starlette web dashboard (app.py, api.py, events.py, logs.py, static/)
coral/cli/ CLI package: start.py, query.py, eval.py, heartbeat.py, agents.py (user-level bindings), ui.py, author.py, validation.py, _helpers.py
coral/user_agents.py User-level agent bindings: load/save ~/.config/coral/agents.yaml, expanded into concrete agent fields by config._expand_bindings
examples/ Task configs (circle_packing, swebench-verified, kernel_engineering, mnist, ...) — each is a task.yaml + seed/ + packaged grader (grader/ referenced by grader.entrypoint); hidden data is declared in grader.private (copied into .coral/private/) and must live outside the agent-visible grader/ package
plugin/ Skills-first, multi-harness plugin for driving coral from another harness (Superpowers-style: one shared skills/, per-harness .claude-plugin/ + .codex-plugin/ manifests, per-harness hooks/ with a SessionStart install check). Per-harness marketplace manifests at the repo root — .claude-plugin/marketplace.json (Claude) and .agents/plugins/marketplace.json (Codex git-backed, git-subdir./plugin) — for owner/repo discovery. No MCP. See plugin/README.md.
tests/ Pytest suite (config, grader, hooks, hub, manager reliability, daemon, workspace, ...)

Read the full file on GitHub · 220 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. yesterday First seen · 220 lines · 4,578 tokens per session scan A 588d2cc7b2cf

Subscribe to this mod's changes

CORAL CLAUDE.md is an instructions file published in the GitHub repository Human-Agent-Society/CORAL (935 stars, last pushed yesterday), licensed Apache-2.0. It adds 4,578 tokens to every session, about $0.0229 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.