@whamp/pi-inline-identifier

Use inline $skill, &agent, and /template references in one user request. Each recognized definition is added once, in first-mention order, and the original request is appended at the tail.

$poteto-mode $teach explain why this extension only inserts one skill

With the pstack extension loaded first (as in this repository's root manifest), $poteto-mode enables sticky Poteto Mode from a complete token anywhere in interactive or RPC user input. Extension-generated input is ignored. /poteto-mode off still disables the mode.

Install and configuration

Install the published package with pi install npm:@whamp/pi-inline-identifier. The root pi package also registers this extension as inline-identifier. Do not load the upstream @pi-kaush/pi-inline-identifier package or its skill/agent compatibility packages in the same Pi session; they also handle these tokens and could transform the request twice.

Identifier behavior

• $skill resolves only a loaded Pi skill. A single distinct skill keeps Pi's native /skill:name expansion. Multiple distinct references render each skill as a Pi-format <skill name="…" location="…"> block.
• &agent discovers file-backed user agents under ~/.pi/agent/agents and trusted project agents under the nearest .pi/agents. Project agents are listed only when ctx.isProjectTrusted() is true and the detected subagent tool accepts the supported direct-execute shape. The current pi-subagents catalog uses { action: "execute", input: { agent, task } }; project delegation adds input.agentScope: "both".
• /template resolves loaded Pi prompt templates inline, including one template by itself. Native leading slash commands (for example /model or /skill:teach) bypass inline processing. Unknown tokens remain ordinary request text.
• Repeated references to the same category and name resolve once. Different categories with the same name remain distinct. Blocks follow the first source offset; generated content is not scanned again. Attachments remain on the transformed input.
• Extension-generated inputs are not transformed. If any recognized file-backed resource cannot be read, Pi receives the unchanged request and gets a warning instead of a partial transformation.

Agent references add instructions; this extension never calls or schedules the subagent tool. Its agent inventory is intentionally limited to on-disk user and trusted project definitions. Built-in or package-provided agents that are not exposed through those files are not discoverable here. Tool schemas that expose neither an explicit nested input agent field nor the supported compact { action, input } execute contract do not enable agent identifiers. A top-level agent field alone is not compatible with the nested call rendered here.

Template placeholders

The complete request is treated as one argument. $1, $@, $ARGUMENTS, defaults such as ${2:-fallback}, and slices such as ${@:1} retain Pi's literal placeholder substitution. Each template is expanded in full for every composed request; context-reuse reminders are not used for flat blocks.

The original request is always appended once after all instruction blocks. A placeholder that inserts the request therefore creates another literal copy inside that template block. This is intentional template-authored duplication. Repeating a placeholder repeats the request as often as the template asks; a template without a consuming placeholder gets no extra copy inside its body.

Editor support and attribution

The extension retains upstream autocomplete, prompt slash-trigger support, theme-based token coloring, and guarded Editor/CustomEditor render wrappers. Inline / completion and coloring depend on Pi's editor internals; if those internals change, input routing is unaffected but the TUI affordances may need an update.

Automated checks cover the coordinator, pstack's sticky token handling, editor adapter, and the root manifest. A real Pi print-mode input run confirmed two-skill ordering, literal template expansion, the request tail, and sticky mode. A second run captured the ordered skill blocks and request tail at before_agent_start and received OK from the model. Interactive Pi showed skill completion and theme coloring. Agent discovery with installed subagent and project trust remains environment-dependent.

Ported from @pi-kaush/pi-inline-identifier, upstream revision d59aab65b94328071b8bcd936b7cd87584723715 (MIT, Copyright (c) 2026 Kaushik Gopal). This package retains the upstream license and attribution. The main behavior change is ordered flat composition with one request tail; the upstream whole-message prompt reuse reminder and its compaction fallback are not used.