@pi-kaush/pi-inline-identifier

Use inline references for loaded Pi skills, named subagents, and prompt templates without replacing Pi's editor or command handling.

Install

pi install npm:@pi-kaush/pi-inline-identifier

Restart Pi or run /reload.

| Identifier | Example | Color | | --------------- | ----------------- | ------ | | Skill | $review | Purple | | Agent | &reviewer | Blue | | Prompt template | /publish-pi-ext | Green |

Do not load this package together with @pi-kaush/pi-inline-skill-identifier or @pi-kaush/pi-inline-agent-identifier. Those compatibility packages handle the same input independently and can transform it twice.

Local development

pi \
  -e ./extensions/pi-inline-identifier/src/skill.ts \
  -e ./extensions/pi-inline-identifier/src/agent.ts \
  -e ./extensions/pi-inline-identifier/src/prompt.ts

Selective loading

The package exposes skill, agent, and prompt handling as separate Pi extension entrypoints. Use pi config to disable any entrypoint you do not want. Disabled entrypoints are not imported, and the enabled entrypoints share one coordinator, autocomplete provider, and editor render patch.

Native slash commands stay native

If the first non-whitespace character on the first line is /, the package does nothing: no custom completion, coloring, or input transformation.

/model                             # untouched
/publish-pi-ext pi-inline-identifier # untouched; Pi expands it natively
Use /publish-pi-ext for this change # handled as an inline identifier

Every loaded prompt template is recognized inline. First-line slash commands still use Pi's native completion and expansion.

Request preservation

The package routes a submitted request at most once:

• Repeating the same identifier still counts as one reference.
• More than one distinct known identifier, including identifiers from different categories, leaves the input unchanged rather than guessing a composition order.
• Extension-generated input is left unchanged.

Skills

Exactly one known $skill reference is rewritten to Pi's native /skill:name command. Pi inserts the skill block followed by the original request once.

Agents

Exactly one known &agent reference adds one delegation instruction followed by the original request once. A compatible subagent tool with an agent parameter is required. Trusted project agents are included only when that tool supports agentScope.

Prompt templates

For an inline prompt reference, the package reads the already-discovered template on first use and caches its body and revision for the current session runtime. /reload, /new, /resume, /fork, or restarting Pi creates a fresh runtime that can pick up template edits. It expands Pi's positional, default, all-argument, and slicing placeholders while treating the complete surrounding request as one argument.

The first full expansion starts with an exact, revision-aware marker such as Inline prompt template "/publish-pi-ext" (revision 0123456789ab):. If that marker remains in a user message in Pi's active, compaction-aware context, a later reference sends a short instruction to reuse it as a fresh invocation instead of inserting the full body again. A summary that merely mentions the prompt does not count. If automatic compaction removes the marker after submission, a context guard restores the full expansion before the model call. When the original message leaves context or the template revision changes, the full body is inserted again.

If the template inserts the request through $1, $@, $ARGUMENTS, or a matching slice, the package does not append another copy. If the template has no placeholder that consumes the request, the expanded template is followed by one Original request section. Repeating a request placeholder inside the template still repeats it intentionally; the extension adds no extra copy.

Compatibility

The package does not replace Pi's editor. It installs guarded render and autocomplete-trigger wrappers so / can open completion after ordinary text. It covers both the extension-resolved Editor and Pi's host CustomEditor when those classes come from separate pi-tui module instances, so render-only custom editor factories such as pi-content-layout keep identifier coloring in either load order. Pi otherwise reserves / for leading command completion and exposes no public inline-slash trigger API. If those editor internals change, inline slash autocomplete may require an update; input routing remains independent.

Performance

• No prompt files, session entries, or agent directories are read while the extension loads.
• Prompt context is scanned backward only when a known inline prompt is submitted, and the scan stops at the first exact revision marker.
• The model-call context guard returns immediately unless a reuse reminder is active.
• Agent discovery is lazy and cached for the session after the first relevant & interaction.
• Disabled entrypoints are not evaluated.
• There are no runtime dependencies or background tasks.
• TUI-only autocomplete and coloring are skipped outside interactive mode; input routing remains available in print, JSON, and RPC modes.