Delphin v0.1.0 — talk to your AI agent while it thinks

First release of Delphin — a duplex companion for AI agent CLIs. Keep typing while the agent works; prompts queue, an arbiter decides interrupt-vs-wait, and the whole conversation is remembered locally.

Features

• Duplex wrapper — runs your agent CLI in a PTY; you can keep sending input while it's busy.
• Pluggable arbiter (--arbiter):
  • heuristic (default) — urgency words ("stop", "wait", "no", …) interrupt; everything else queues.
  • question — questions also interrupt; plain instructions queue.
• Smarter idle detection — --ready 'you> ' flips to idle the moment the agent shows a prompt (beyond plain output-silence).
• Local memory — every prompt, verdict, and reply is logged to a local SQLite file. Query it with delphin recall [query].
• Config file — ./.delphin.toml or <config-dir>/delphin/config.toml; CLI flags override.
• Agent skills — ships a Claude Code plugin (/plugin marketplace add wuisabel-gif/Delphin) and a Codex prompt.

Ecosystem

Sibling of MemoryWhale (the memory layer). Point Delphin at MemoryWhale's DB with --db to feed conversations into it. See ECOSYSTEM.md.

Install

• Binary: download delphin-v0.1.0-macos-arm64 below, chmod +x, run.
• From source: cargo build --release → target/release/delphin.

Quality

CI (fmt + clippy -D warnings + test) on every PR; 15 unit tests + an e2e test.

delphin --interrupt ctrl-c -- bash examples/mock-agent.sh   # try it (no real model)
delphin -- claude                                            # wrap a real agent