Agent Routing

Agent routing is MadoHub's core feature. When one agent completes a task, MadoHub can automatically extract the result, transform it into an actionable prompt, and deliver it to another agent.

How it works

Trigger — A connected source tile fires based on the configured trigger type (idle detection, completion, keyword match, or continuous).
Extract — MadoHub reads the agent's last response. For Claude Code, this uses JSONL session files for clean extraction. For other agents, it reads terminal output directly.
Transform — The extracted output is transformed into a prompt for the target agent. This can use AI-powered rewriting, template wrapping, or raw passthrough.
Deliver — The transformed prompt is typed into the target terminal with human-like delays (50–80ms per line), then submitted with the agent-appropriate Enter sequence.

JSONL session resolution

For Claude Code, MadoHub extracts clean responses by reading session log files:

Find the terminal's tmux pane, then walk the process tree (ps -eo pid=,ppid=,comm=) breadth-first to the first descendant whose command name contains claude/codex/amp/opencode — that PID is the agent process.
Read ~/.claude/sessions/{PID}.json (keyed by the agent's PID, not by session id) and extract its sessionId field.
Scan every subdirectory of ~/.claude/projects/ for a file named {sessionId}.jsonl — that's the transcript.
Parse the JSONL and extract the last assistant message.

This gives much cleaner content than raw terminal scraping, avoiding ANSI codes and formatting artifacts.

Separately, cwd_to_project_dir() maps a tile's working directory to a project directory by replacing / with -, falling back to a substring search when no exact match exists.

Trigger types

Type	Behavior
on-idle	Waits for a completion-keyword match, then 8 seconds of silence (`IDLE_DELAY_MS`). Default and most reliable.
on-complete	Handled identically to on-idle: waits for the same completion-keyword match, then the same 8-second silence delay before firing.
on-keyword	Fires when a custom regex pattern matches in output.
always	Streams all new terminal output continuously.

Transform types

Type	Description
ai-routing	Uses an AI model to generate a context-aware prompt for the target agent. Requires API key in settings. Supports custom templates.
prompt-wrap	Wraps output in a user-provided template. Variables: `{output}`, `{round}`, `{maxRounds}`, `{source}`.
full-output	Sends the last 30 lines of terminal output as-is.
raw	Sends the last 10 lines of terminal output.
summary	Sends an AI-generated summary of the output.

Routing dispatch defaults to the madoagent path (also selectable: legacy, shadow).

Content delivery

MadoHub types content into the target terminal line by line with 50–80ms delays to simulate human input. The Enter key behavior is agent-specific:

Claude Code, Codex, Amp, OpenCode: Double carriage return (\r\r) to submit. Codex has an 800ms pre-delay with 200ms between the two presses; the other agents use a 500ms pre-delay.
Shell: Single newline (\n) to submit.

Loop protection

Routing connections have built-in safeguards to prevent infinite loops:

Setting	Default	Range	Description
maxRounds	5	1–20	Maximum routing cycles before auto-deactivation.
cooldownMs	3s	1–30s	Minimum delay between triggers.
stopKeyword	—	Any string	Stops routing when detected (e.g., "LGTM").

Example workflow

Terminal 1 (Claude Code)          Terminal 2 (Claude Code)
┌──────────────────────┐    →    ┌──────────────────────┐
│ "Refactor auth module"│         │ Receives:            │
│                       │         │ "Review the refactor │
│ ✓ Done — 3 files     │         │  in src/auth/..."    │
└──────────────────────┘         └──────────────────────┘

Agent 1 refactors the code. On completion, MadoHub reads the JSONL response, uses AI routing to generate a review prompt with the specific files and changes, and delivers it to Agent 2.