cj-ab (agent bridge)

agent-bridge runs one headless prompt against another agent CLI and streams that agent's answer to stdout:

agent-bridge [-m light|medium|heavy|<model>] <claude|codex|copilot> "<prompt>"|- [file_path]

If agent-bridge is not on PATH, call it by path from this skill folder.

Writing the prompt

The target agent starts cold: it sees only the prompt and the file, never your conversation. Write the prompt as a complete brief, with the goal, the constraints that matter, and the output format you want back ("output only the code", "list findings as bullets", "reply with a verdict and three reasons"). Pass a single file_path and let the bridge inline it. For several files, name their paths in the prompt; the target runs in your working directory and can read them itself.

Pass the prompt as - and feed it through a heredoc with a quoted delimiter, so the shell leaves quotes, $() and backticks in it alone:

agent-bridge codex - src/auth.py <<'EOF'
Convert this Python file to idiomatic Rust. Output only the code.
EOF

Any target is valid, including your own kind. agent-bridge claude from Claude gives a fresh-context second opinion.

Choosing the model

If the user named a model, pass it with -m exactly as given. Otherwise pick the tier from the task's complexity:

• medium: the usual choice, including hard work: a focused or cross-cutting review, a hard bug, a multi-file change, an alternative draft, a translation with real logic in it.
• light: only for trivially mechanical work with one right answer: reformatting, a quick lookup, a yes/no sanity check.
• heavy: expensive (Fable on claude), so reach for it only when the user asks for it, or when the problem is extraordinarily difficult and medium has already fallen short on it.

Torn between light and medium, take medium; torn between medium and heavy, stay on medium. Omitting -m runs the CLI's configured default model. The tier names map to concrete models in the script's TIERS table.

Running it

• Timeout. A call takes 1 to 5 minutes, longer on heavy. Set the shell tool's timeout to 10 minutes (600000 ms), or use the shell tool's own background mode, which keeps the output and reports completion. With a plain &, redirect stdout and stderr to a file and wait on the process before reading it.
• Full access. The target runs with no sandbox and no approval prompts: it can edit files, run commands and reach the network in your working directory. Say in the prompt whether you want edits made or only an answer back, and review git diff after a call that edits.
• One hop. The bridge refuses a call made from inside a bridged agent. The guard is an inherited env var, so it stops accidental ping-pong only.
• Cancellation. Killing the bridge (timeout, Ctrl-C) takes the target's whole process tree down with it.
• Failures. Errors start with [agent-bridge] on stderr: bad usage, an unreadable file, or a CLI missing from PATH. The bridge passes the target's own errors and exit code through as-is.

Using the answer

The answer is input to your judgement. Check it, and any edits it made, against the code before building on it or relaying it, and tell the user which agent and model it came from. When you are cross-checking, report where the two answers agree and where they differ.