Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Latest commit

 

History

History
58 lines (44 loc) · 2.13 KB

File metadata and controls

58 lines (44 loc) · 2.13 KB

Agents — Istok

This directory defines the multi-agent setup for the Istok project. Two agents collaborate with distinct responsibilities:

Agent Tool Primary role
Claude Code claude CLI Design, architecture, planning, RFCs, code review
Codex codex CLI Feature implementation, mechanical code changes

Role division

Claude Code (roles/claude.md):

  • Owns milestone scoping and definition-of-done criteria
  • Authors RFCs before large refactors
  • Reviews Codex output for correctness and architectural fit
  • Makes decisions about crate boundaries and public API shape

Codex (roles/codex.md):

  • Implements features as specified in milestones and RFCs
  • Follows all skills in skills/ unconditionally
  • Does not make architectural decisions or modify docs/milestones.md
  • Hands off to Claude Code when a task requires design judgment

Skills

Skills in skills/ apply to both agents unless the frontmatter says otherwise. Each skill is a folder containing a SKILL.md with YAML frontmatter and Markdown body.

skills/
  rust_style/         — Rust coding conventions for all crates
  protocol_testing/   — Deterministic harness and codec test patterns
  no_std_rules/       — no_std compatibility requirements for istok-core
  milestone_playbook/ — How to scope, implement, and close out a milestone

Task handoff workflow

Work is handed from Claude Code to Codex via .agents/TASK.md (gitignored, local only).

Claude Code writes TASK.md
  → user runs Codex: "read .agents/TASK.md and implement it"
    → Codex implements, checks off acceptance criteria, sets status to ready_for_review
      → user reviews code
        → user asks Claude Code to review
          → Claude Code approves or writes a new TASK.md with follow-up

TASK.md is overwritten for each new task. History lives in git commits, not the file.

Reading order for agents

  1. Read your own role file in roles/
  2. Read all applicable skills in skills/
  3. Cross-check against CLAUDE.md in the repo root (authoritative on hard rules)
  4. Codex only: read .agents/TASK.md for the current task