@ssk_dev/pi-docs-slim

See my full setup for Pi

简体中文

A Pi extension that keeps Pi's built-in documentation guidance out of the system prompt. When you actually have a question about Pi itself, ask it with /pi and the guidance comes back for that one run.

This is a fork of robzolkos/pi-slim by Rob Zolkos. The idea and the /pi command come from his extension. The fork exists mainly to make it work on newer Pi versions.

Install

pi install npm:@ssk_dev/pi-docs-slim

If you have the original pi-slim installed, remove it first. Both register a /pi command.

Requires Pi 0.87.1 or newer (tested on 0.87.1, 1.0.0, 1.3.0 and 1.4.0).

Usage

• Normal requests: Pi's documentation guidance is removed from what the model sees.
• /pi <question about Pi>: this run reads the guidance at the head of its own message instead of in the system prompt, and keeps it for its tool-call turns. Once the guidance has gone into a request, the transcript shows a one-line 📖 pi-docs marker between your message and the reply, so you can see which run carried it. The next normal request goes back to removing it.
• Your own <docs> blocks, project instructions, and code examples are left alone. Saved user messages, assistant replies, and tool output are never changed (/pi only adds the guidance to the request it sends).

After installing or updating, run /reload, send a normal message, and check /context.

Why fork instead of the original

The original 0.2.1 removes the docs by replacing a string in the system prompt. Starting with Pi 0.87.1, the system prompt is built from structured sections and saved as system messages, so that replacement no longer takes effect: the docs still reach the model (see pi-slim#3).

What this fork changes:

• Removes the docs from the new structured prompt. It blocks Pi from filling the docs section back in, and strips the docs from every request before the provider sees them. The saved system message keeps a short <!-- pi-slim:omit --> marker in that section: Pi derives each request's sections from the transcript, so a cleared section makes it append an empty system message to every request.
• Cleans old history too. Docs left over from an earlier /pi run are stripped from later normal requests.
• Works with /context in any load order. pi-context-view never shows Pi's docs and never mislabels the prompt as forced, whether it loads before or after this extension (see Limits for the marker).
• Makes /pi reliable. The "keep docs" permission is tied to the exact request /pi sends, so it can't leak into your next message if that request gets intercepted.
• /pi no longer breaks the prompt cache. Most providers get the system prompt at the very start of the request, so putting the guidance back into it made the provider re-read the whole session: one /pi run cost two full cache misses (the run itself, plus the normal request that switched the section back off). Now the system prompt is byte-identical for every request and the cached prefix before /pi is kept. Only the /pi run itself is new: about 500 tokens of guidance, plus re-reading that run once on the next normal request, which no longer carries the guidance.
• Doesn't freeze the prompt. Tools and rules added later by other extensions still show up.
• Adds offline integration tests. They run real Pi sessions with an offline model provider, so no model service is called.

No new runtime dependencies.

Tests

npm test

The tests need Pi 0.87.1+ and pi-context-view 0.6.0 or 0.8.0 installed. They look for the global Pi and ~/.pi/agent/npm/node_modules/pi-context-view by default. For other locations, point to the package roots:

PI_CODING_AGENT_DIR=/path/to/pi-coding-agent PI_CONTEXT_VIEW_DIR=/path/to/pi-context-view npm test

Set PI_SLIM_ENTRY to run the suite against another implementation.

Limits

• Old session files aren't rewritten. /context's Usage view is clean after the next normal request. The Initial view is a snapshot from the first request; reload to refresh it. If that first request was /pi, it will include the docs, which is expected.
• During a /pi run, /context shows the guidance as content of that message (about 500 tokens). It is added to the request only: the saved session keeps your question as you typed it, and later requests don't carry the guidance.
• The 📖 pi-docs marker is a session CustomEntry: Pi keeps custom entries out of the model's context, so it adds no prompt tokens, /context does not count it, and it stays in the transcript across reloads. It is written after the run's message is saved, so forking or /tree from that message leaves it behind, and compaction picks the same cut as without it. Its token count is a rough estimate, and an extension that edits the request after this one can still change what is sent.
• Saved sessions keep the ~20-character omission marker in their docs section so Pi's per-request section diff converges. The marker is removed from every request and never reaches the model. /context may still count it (about 5 tokens): pi-context-view 0.6.0 reads the saved transcript, and 0.8.0 captures each request before this extension cleans it when it loads first. That is pi-context-view's documented behavior for handlers that run after it (see its capture notes); you don't need to change your load order.
• If the model reads Pi's docs with a tool, that output is kept.
• The cleanup relies on Pi 0.87.1's event order. Rerun the tests after upgrading Pi. If another extension re-injects the docs after the final context step, they won't be removed.

License

MIT. Original work © pi-slim contributors (Rob Zolkos); see LICENSE.