@alexleekt/pi-bump

Invisible continuation for the Pi agent.

Resume the agentic loop without sending the LLM a single token.

Usage

Double-tap Enter (quickest)

1. Make sure the chat editor is empty
2. Press Enter twice within ~300 ms
3. The agent continues — the LLM sees nothing

/continue command

Command	What it does
/continue	Resume the loop invisibly (or visibly if escalated)
/continue status	Show if the agent is idle, busy, or escalated
/continue help	Show available commands

Both methods only fire when the agent is idle and no messages are pending.

How it works

Default (invisible): A hidden customType message is sent with display: false. The context handler replaces it with "Continue" before the LLM sees it. The LLM receives a clean semantic nudge and the chat stays uncluttered.

Escalated (visible): When loop detection notices the last two assistant responses had identical tool calls (or exact text duplicates), the next continue sends a real visible user message like "What's next?" or "Keep going" — stronger signal, same goal.

A context event handler proactively removes any leaked invisible markers as insurance. Real user input always resets escalation state.

Installation

npm install -g @alexleekt/pi-bump

Pi auto-discovers globally installed pi-package extensions.

Or copy index.ts into .pi/extensions/ (project) or ~/.pi/agent/extensions/ (global).

Debug mode

BUMP_DEBUG=1 pi

Then toggle per-session debugging with /bump-debug-keypresses:

• Enter — triggers invisible continue + shows timing
• Backspace, Delete — shows timing only
• Ctrl+Enter, Alt+Enter — also monitored

Debug mode resets when the session ends. Only available when BUMP_DEBUG=1 is set.

Note: Debug mode monitors 5 keys (enter, backspace, delete, ctrl+enter, alt+enter). In normal mode, only Enter is monitored for performance. The other keys are only checked when debug mode is active for that session.

Performance

pi-bump is designed to stay out of the way:

• Minimal keystroke overhead — In normal mode, only Enter is matched against incoming keystrokes. No matchesKey() calls for backspace, delete, or modifier keys unless debug mode is active.
• Zero-allocation context guard — The context handler scans for invisible markers with .some() before allocating. In typical conversations (no invisible messages), no array is created.
• Synchronous state reset — The input handler resets escalation state immediately (no microtask deferral), preventing false loop detection.

Hybrid Escalation Strategy (v0.3.1+)

pi-bump uses a two-tier strategy to keep the agent moving without polluting the chat:

1. Invisible tier (default) — sends a hidden customType message (display: false). The context handler replaces it with "Continue" for the LLM. Chat stays clean.
2. Visible tier (escalation) — when loop detection sees identical tool calls (or exact text duplicates) across the last two assistant responses, the next continue sends a real visible user message with a randomized nudge like "Continue", "Go deeper", or "Show me where this leads".
3. Auto-reset — a non-loop assistant response or real user input resets back to the invisible tier.

The visible tier only triggers when the invisible tier isn't breaking the loop — so most continues stay silent and seamless.

Acknowledgments

The invisible continuation technique was adapted from pi-invisible-continue by Tom X Nguyen.

License

MIT