Coasts

Multiple Harnesses

If one repository is used from more than one harness, one way to consolidate the Coasts setup is to keep the shared /coasts workflow in one place and keep the harness-specific always-on rules in the files for each harness.

AGENTS.md
CLAUDE.md
.cursor/rules/coast.md           # optional Cursor-native always-on rules
.agents/skills/coasts/SKILL.md
.agents/skills/coasts/agents/openai.yaml
.claude/skills/coasts -> ../../.agents/skills/coasts
.cursor/commands/coasts.md       # optional, thin, harness-specific
.claude/commands/coasts.md   # optional, thin, harness-specific

Use this layout like this:

  • AGENTS.md — short, always-on rules for working with Coasts in Codex and T3 Code
  • .cursor/rules/coast.md — optional Cursor-native always-on rules
  • CLAUDE.md — short, always-on rules for working with Coasts in Claude Code and Conductor
  • .agents/skills/coasts/SKILL.md — canonical reusable /coasts workflow
  • .agents/skills/coasts/agents/openai.yaml — optional Codex/OpenAI metadata
  • .claude/skills/coasts — Claude-facing mirror or symlink when Claude Code also needs the same skill
  • .cursor/commands/coasts.md — optional Cursor command file; one simple option is to have it reuse the same skill
  • .claude/commands/coasts.md — optional explicit command file; one simple option is to have it reuse the same skill

Step-by-step

  1. Put the Coast Runtime rules in the always-on instruction files.
    • AGENTS.md, CLAUDE.md, or .cursor/rules/coast.md should answer the "every task" rules: run coast lookup first, use coast exec, read logs with coast logs, ask before coast assign or coast run when there is no match.
  2. Create one canonical skill for Coasts.
    • Put the reusable /coasts workflow in .agents/skills/coasts/SKILL.md.
    • Use the Coast CLI directly inside that skill: coast lookup, coast ls, coast run, coast assign, coast unassign, coast checkout, and coast ui.
  3. Expose that skill only where a harness needs a different path.
    • Codex, T3 Code, and Cursor can all use .agents/skills/ directly.
    • Claude Code needs .claude/skills/, so mirror or symlink the canonical skill into that location.
  4. Add a command file only if you want an explicit /coasts entrypoint.
    • If you create .claude/commands/coasts.md or .cursor/commands/coasts.md, one simple option is to have the command reuse the same skill.
    • If you give the command its own separate instructions, you are taking on a second copy of the workflow to maintain.
  5. Keep Conductor-specific setup in Conductor, not in the skill.
    • Use Conductor Repository Settings scripts for bootstrap or run behavior that belongs to Conductor itself.
    • Keep Coasts policy and use of the coast CLI in CLAUDE.md and the shared skill.

Concrete /coasts example

A good shared coasts skill should do three jobs:

  1. Use Existing Coast
    • run coast lookup
    • if a match exists, use coast exec, coast ps, and coast logs
  2. Manage Assignment
    • run coast ls
    • offer coast run, coast assign, coast unassign, or coast checkout
    • ask before reusing or disrupting an existing slot
  3. Open UI
    • run coast ui

That is the right place for the /coasts workflow. The always-on files should only hold the short rules that must apply even when the skill is never invoked.

If you want Claude Code to reuse the same skill as Codex, T3 Code, or Cursor, one option is a symlink:

mkdir -p .claude/skills
ln -s ../../.agents/skills/coasts .claude/skills/coasts

A checked-in mirror is also fine if your team prefers not to use symlinks. The main goal is just to avoid unnecessary drift between copies.

Harness-specific cautions

  • Claude Code: project skills and optional project commands are both valid, but keep the logic in the skill.
  • Cursor: use AGENTS.md or .cursor/rules/coast.md for the short Coast Runtime rules, use a skill for the reusable workflow, and keep .cursor/commands optional.
  • Conductor: treat it as CLAUDE.md plus Conductor scripts and settings first. If you add a command and it does not appear, fully close and reopen the app before checking again.
  • T3 Code: this is the thinnest harness surface here. Use the Codex-style AGENTS.md plus .agents/skills pattern, and do not invent a separate T3-specific command layout for docs about Coasts.
  • Codex: keep AGENTS.md short and put the reusable workflow in .agents/skills.