- --turns N: control how many conversation turns to include (default 25)
- --include all|conversation|git|todos: pick what context to send
- Reduced per-turn size to prevent Codex prompt overflow
- UTF-8 safe truncation throughout
- Interactive Codex launch with inherited stdio
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Instead of codex exec (non-interactive), Relay now launches codex
with inherited stdin/stdout/stderr so it opens the interactive TUI.
The user can keep talking to Codex with the full handoff context
already loaded as the initial prompt.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Relay now reads the actual Claude Code JSONL session transcript and
extracts the FULL conversation: user messages, assistant reasoning,
every tool call (Bash, Read, Write, Edit), and every tool result
including errors. This is the real context, not just git state.
- 80 conversation turns captured from live Claude session
- Tool calls summarized: [Write] path, [Bash] command, [Edit] path
- Tool results with errors preserved
- UTF-8 safe truncation
- Status command shows last 15 turns with emoji prefixes
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When Claude Code hits its rate limit, Relay captures the full session
state (task, todos, git diff, decisions, errors) and hands it off to
Codex, Gemini, Ollama, or GPT-4 — so your work never stops.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
