pi-file-reference

A pi extension that resolves @filepath references in AGENTS.md and injects the referenced file contents into the system prompt.

How it works

1. On session_start, the extension resets its cache for the new session
2. On the first before_agent_start, parses @filepath references from Pi's loaded context files (AGENTS.md, CLAUDE.md, custom context files), filters to .md/.mdc only, ignores dot-files and files over 100KB
3. Reads the referenced files and caches their contents (survives reload)
4. On every before_agent_start, injects the file contents into the system prompt as <project_references> blocks inside Pi's <project_context> section

@filepath syntax

@path/to/file.md
@./path/to/file.md
@~/path/to/file.md
@"path with spaces.md"
@/absolute/path/to/file.md

The @ must be at the start of a line or preceded by whitespace.

Only .md and .mdc file references are resolved. References to other file types (e.g., @notes.txt) are silently ignored.

Directory references

When @path resolves to a directory, only .md and .mdc files at depth 1 are injected. Dot-files (e.g., .hidden.md) are skipped. Files larger than 100KB are also skipped.

@./docs

Injects @docs/style-guide.md, @docs/patterns.md, etc. (sorted alphabetically). Subdirectories are skipped. Trailing / is stripped (@./docs/ works the same).

Installation

# Install from npm (once published)
pi install npm:@josephyoung/pi-file-reference

# Or clone and link locally
git clone https://github.com/your-org/pi-file-reference.git
cd pi-file-reference
npm install
pi install .

Usage

In your AGENTS.md:

# Project guidelines
Always follow the patterns in @./docs/style-guide.md

# Architecture
@~/.project-arch/backend-overview.md

The referenced files' content is automatically injected into the system prompt at the start of each agent turn.

Caching & LLM Prompt Cache

File contents are read once per session (first before_agent_start) and reused for every subsequent turn. This guarantees the injected system prompt bytes are identical within a session, making the extension LLM prompt-cache friendly.

Decision	Rationale
Cache at session level	New session = fresh content. No stale data across sessions.
Cache file contents, not full prompt	lastIndexOf + string concat is O(n) and memcpy-cheap. Caching the full prompt + hashing the raw prompt for hit/miss adds complexity with no observable gain.
Inject before </project_context>	References are project context material. Fallback to appending if the tag is absent.

The only cache-invalidation risk is other extensions producing non-deterministic system prompt output, which is rare and outside our control.

See docs/lifecycle-diagram.md for a visual flow and docs/system-prompt-injection.md for full design analysis.

License

MIT

See CHANGELOG.md for version history.