name

agent-bridge

description

Use on Agent Bridge attach (prompts starting with `[bridge-probe:`), after context compaction, or before sending/cancelling/interrupting/clearing/inspecting bridge peer messages or when bridge command syntax/recovery semantics are needed (agent_send_peer, agent_view_peer, agent_wait_status, agent_aggregate_status, agent_alarm, agent_extend_wait, agent_cancel_message, agent_interrupt_peer, agent_clear_peer). If these operating rules are already present in the current context, apply them from memory — do not reload this skill for routine `[bridge:*]` FYI notices, membership changes, ordinary results, or no-reply-needed events.

Agent Bridge Operating Guide

Load Policy

This skill is idempotent. If these operating rules are already present in the current context, do not re-read this file for routine bridge prompts.

For simple [bridge:*] FYI notices, membership updates, and no-reply-needed events:

do not read this file again
do not run bridge commands
acknowledge only if the human-facing flow requires it

Read this file again only after compaction, when unsure, or before using bridge commands whose syntax or recovery behavior matters.

Hot Path Rules

Need an answer from a peer: agent_send_peer --to <alias> 'request'. Notices do not auto-route a reply — use a request when you need one.
Do not poll, sleep, or keep the turn open waiting for peer results; results arrive later as new [bridge:*] prompts.
Do not read bridge state files directly; use bridge commands.
Do not create, simulate, or echo real-looking [bridge:*] prompt lines.
Large payloads go under /tmp/agent-bridge-share/; send the path, not the inline body (11000-char limit).
For exact current command syntax or recovery behavior: run agent_list_peers, then <command> --help for specifics.
Routine FYI notices (membership, room events, results you were already waiting for): handle inline using cached rules — do not reload this skill.

References

Read these only when the hot-path rules are insufficient for the current task:

references/command-contract.md — full runtime cheat sheet mirrored from bridge_instructions.model_cheat_sheet_text().
references/recovery.md — watchdog, cancel, interrupt, clear, and aggregate recovery flows.
references/anti-patterns.md — common bridge mistakes and safer alternatives.

Agent Bridge Operating Guide

Load Policy

This skill is idempotent. If these operating rules are already present in the current context, do not re-read this file for routine bridge prompts.

For simple [bridge:*] FYI notices, membership updates, and no-reply-needed events:

do not read this file again

do not run bridge commands

acknowledge only if the human-facing flow requires it

Read this file again only after compaction, when unsure, or before using bridge commands whose syntax or recovery behavior matters.

Hot Path Rules

Need an answer from a peer: agent_send_peer --to <alias> 'request'. Notices do not auto-route a reply — use a request when you need one.

Do not poll, sleep, or keep the turn open waiting for peer results; results arrive later as new [bridge:*] prompts.

Do not read bridge state files directly; use bridge commands.

Do not create, simulate, or echo real-looking [bridge:*] prompt lines.

Large payloads go under /tmp/agent-bridge-share/; send the path, not the inline body (11000-char limit).

For exact current command syntax or recovery behavior: run agent_list_peers, then <command> --help for specifics.

Routine FYI notices (membership, room events, results you were already waiting for): handle inline using cached rules — do not reload this skill.

References

Read these only when the hot-path rules are insufficient for the current task:

references/command-contract.md — full runtime cheat sheet mirrored from bridge_instructions.model_cheat_sheet_text().

references/recovery.md — watchdog, cancel, interrupt, clear, and aggregate recovery flows.

references/anti-patterns.md — common bridge mistakes and safer alternatives.