Relay Context

简体中文

Relay Context is a lightweight personal Codex skill that turns the latest relevant topic segment into a concise, self-contained prompt for another agent or conversation.

It relays a focused task—not the entire conversation. The recipient does not need this skill or access to the source chat.

Why use it

• Move a focused issue from a sidebar to the main conversation.
• Ask another agent for independent analysis or review.
• Continue a specific task in a new conversation without copying the full chat.
• Convert a discussion into an implementation-ready prompt.

How scope is selected

Relay Context uses semantic relevance instead of a fixed number of turns:

1. Use the topic or range explicitly specified by the user.
2. Otherwise find the newest coherent unresolved topic segment.
3. Scan backward for decisions, constraints, evidence, and background required by that topic.
4. Stop at unrelated, resolved, superseded, repetitive, or independently reconstructable material.
5. Include older details when the current topic depends on them.

The goal is sufficient context for the next step, not a lossless transcript.

Usage

Normally, invoke the skill without arguments:

$relay-context

Codex-specific shorthands are optional:

$relay-context main
$relay-context side

Natural-language requests work as well:

Turn the discussion above into a prompt for the implementation agent.
Package this issue for a read-only review in another conversation.
Relay the migration topic from the decision point onward.

Output behavior

The skill:

• Produces an implementation, analysis, or review prompt based on the user's intent.
• Separates user-confirmed decisions from unconfirmed assistant suggestions.
• Assumes the recipient cannot see the source conversation.
• Carries intent, decisions, constraints, evidence, and other context the recipient cannot reconstruct.
• References known workspace artifacts instead of copying content the recipient can recover there.
• Preserves failed approaches only when needed to prevent a significant repeated mistake.
• Redacts secrets, credentials, tokens, and unnecessary personal information.
• Outputs exactly one copyable fenced code block.
• Never reads transcripts, inspects files or repositories, performs the task, calls tools, operates the clipboard, or creates a handoff file.

Relay Context vs. Handoff

Use Relay Context to transfer one focused topic as a prompt. Use a full-session handoff workflow when a fresh agent must resume the overall state of a long-running conversation, including progress across multiple topics.

Installation

Clone or download the repository, then copy only the runtime files into your personal Codex skills directory:

target="${CODEX_HOME:-$HOME/.codex}/skills/relay-context"
mkdir -p "$target/agents"
cp SKILL.md "$target/SKILL.md"
cp agents/openai.yaml "$target/agents/openai.yaml"

Restart Codex or begin a new conversation if the skill does not appear immediately.

Runtime files

relay-context/
├── SKILL.md
└── agents/
    └── openai.yaml

Repository documentation is not required in the installed skill directory.