pi-spark 0.8.0 → 0.9.1

package/README.md

@@ -9,6 +9,7 @@ A small, opinionated collection of [pi](https://pi.dev/) extensions.
 - **Editor:** replaces the default editor with a compact working indicator (inspired by [Amp](https://ampcode.com/)) and current model info.
 - **Footer:** shows session information, extension statuses, cost, and context usage on one line.
 - **Fullscreen:** clears the screen and scrollback on session start, pins the editor and footer to the bottom for a full-screen session, and clears again on exit.
+- **Models:** exposes a `model` tool so the agent can list the models the user can use currently and inspect the provider, model, and thinking level in use.
 - **Name:** exposes a `name` tool so the agent can give the current session a concise, recognizable name in the session selector.
 - **Presets:** switches named model presets with `/preset`, `--preset`, and quick cycle shortcuts.
 - **Recap:** generates a short idle-session recap and exposes a `/recap` command for manual generation, inspired by [Claude Code's session recap](https://code.claude.com/docs/en/interactive-mode#session-recap).
@@ -76,6 +77,12 @@ Example:
 
 - pi-spark clears the screen and scrollback at session start and exit, pins the editor and footer to the bottom, and enables pi's `clearOnShrink` behavior programmatically so pinned UI stays aligned after taller components close.
 
+### Models
+
+- The agent can call the `model` tool with two actions:
+  - `list`: gets all currently usable models with their metadata, with optional `provider` and `model` substring filters and `offset`/`limit` paging.
+  - `current`: gets the active provider, model, and thinking level.
+
 ### Name
 
 - The agent can set or refresh the current session's name and optionally give a reason.

package/extensions/models/config.ts

@@ -0,0 +1,3 @@
+import * as z from "zod";
+
+export const modelsConfigSchema = z.object({});

package/extensions/models/index.ts

@@ -0,0 +1,188 @@
+import { clampThinkingLevel, getSupportedThinkingLevels, StringEnum } from "@earendil-works/pi-ai";
+import { DEFAULT_MAX_BYTES, DEFAULT_MAX_LINES, truncateHead } from "@earendil-works/pi-coding-agent";
+import { Container, Spacer, Text } from "@earendil-works/pi-tui";
+import { Type } from "typebox";
+
+import { loadConfig } from "../shared/config";
+import { formatModel } from "../shared/format";
+
+import type { Api, Model, ModelThinkingLevel } from "@earendil-works/pi-ai";
+import type { ExtensionAPI, TruncationResult } from "@earendil-works/pi-coding-agent";
+
+/** Pi's built-in default thinking level, clamped per model. Not exported by pi's public API. */
+const DEFAULT_THINKING_LEVEL: ModelThinkingLevel = "medium";
+
+type ModelMetadata = Omit<Model<Api>, "headers" | "compat"> & {
+  thinkingLevels: ModelThinkingLevel[];
+  defaultThinkingLevel: ModelThinkingLevel;
+};
+
+type ModelToolDetails =
+  | { action: "current"; current: { model: ModelMetadata; thinkingLevel: ModelThinkingLevel } }
+  | { action: "list"; models: ModelMetadata[]; total: number; truncation?: TruncationResult }
+
+function toMetadata(model: Model<Api>): ModelMetadata {
+  const { headers: _headers, compat: _compat, ...metadata } = model;
+
+  return {
+    ...metadata,
+    thinkingLevels: getSupportedThinkingLevels(model),
+    defaultThinkingLevel: clampThinkingLevel(model, DEFAULT_THINKING_LEVEL),
+  };
+}
+
+export default function (pi: ExtensionAPI) {
+  pi.on("session_start", (_event, ctx) => {
+    const config = loadConfig(ctx, "models");
+    if (!config) return;
+
+    pi.registerTool({
+      name: "model",
+      label: "model",
+      description:
+        "Inspect pi's model state. The \"current\" action returns the active model with metadata " +
+        "and thinking level. The \"list\" action returns all models the user can use " +
+        "currently (auth configured), including metadata such as provider, id, name, API type, " +
+        "reasoning support, input modalities, cost, context window, and max output tokens. " +
+        "Lists can be filtered with provider/model queries and paged with offset/limit. " +
+        `List output is one JSON object per line, truncated to ${DEFAULT_MAX_LINES} models or ` +
+        `${DEFAULT_MAX_BYTES / 1024}KB (whichever is hit first).`,
+      promptSnippet: "List available models or show the current model in use",
+      promptGuidelines: [
+        "Use model when you need to know available pi models, the active provider or model, or the current thinking level.",
+      ],
+      parameters: Type.Object({
+        action: StringEnum(["list", "current"] as const, {
+          description:
+            "\"list\" returns all currently usable models with metadata; " +
+            "\"current\" returns the active model with metadata and thinking level.",
+        }),
+        provider: Type.Optional(Type.String({
+          description: "For list: filter by provider, case-insensitive substring (e.g., \"vercel\", \"moonshot\")",
+        })),
+        model: Type.Optional(Type.String({
+          description: "For list: filter by model id or display name, case-insensitive substring (e.g., \"claude\", \"deepseek\")",
+        })),
+        offset: Type.Optional(Type.Number({
+          description: "For list: model number to start from (1-indexed)"
+        })),
+        limit: Type.Optional(Type.Number({
+          description: "For list: maximum number of models to return"
+        })),
+      }),
+      renderCall(args, theme) {
+        let text = `${theme.bold(theme.fg("toolTitle", "model"))} ${theme.fg("accent", args.action)}`;
+
+        if (args.provider) text += theme.fg("muted", ` provider:${args.provider}`);
+        if (args.model) text += theme.fg("muted", ` model:${args.model}`);
+        if (args.offset !== undefined || args.limit !== undefined) text += theme.fg("warning", ` from:${args.offset ?? 1}`);
+        if (args.limit !== undefined) text += theme.fg("warning", ` to:${(args.offset ?? 1) + args.limit - 1}`);
+
+        return new Text(text, 0, 0);
+      },
+      renderResult(result, { expanded }, theme, context) {
+        const details = result.details as ModelToolDetails | undefined;
+
+        const container = new Container();
+        container.addChild(new Spacer(1));
+
+        if (context.isError || !details) {
+          const output = result.content
+            .filter((content) => content.type === "text")
+            .map((content) => content.text)
+            .join("\n");
+          container.addChild(new Text(theme.fg("error", output), 0, 0));
+
+          return container;
+        }
+
+        if (details.action === "current") {
+          const { model, thinkingLevel } = details.current;
+          container.addChild(new Text(theme.fg("muted", formatModel(model.provider, model.id, thinkingLevel)), 0, 0));
+
+          return container;
+        }
+
+        const filtered = context.args.provider !== undefined || context.args.model !== undefined;
+        const models = details.models ?? [];
+        const total = details.total ?? models.length;
+
+        if (expanded) {
+          models.forEach((model) => {
+            container.addChild(new Text(theme.fg("muted", formatModel(model.provider, model.id)), 0, 0));
+          });
+
+          container.addChild(new Spacer(1));
+        }
+
+        const summary = `${models.length !== total ? `${models.length} of ` : ""}${total} ${filtered ? "matched" : "available"} model${total === 1 ? "" : "s"} listed`;
+        container.addChild(new Text(theme.fg("muted", summary) + (details.truncation?.truncated ? theme.fg("warning", " (truncated)") : ""), 0, 0));
+
+        return container;
+      },
+      async execute(_toolCallId, params, _signal, _onUpdate, ctx) {
+        if (params.action === "current") {
+          const model = ctx.model;
+          if (!model) throw new Error("No model is currently selected.");
+
+          const thinkingLevel = pi.getThinkingLevel();
+          const current = { model: toMetadata(model), thinkingLevel };
+
+          return {
+            content: [{ type: "text", text: JSON.stringify(current) }],
+            details: { action: "current", current } satisfies ModelToolDetails,
+          };
+        }
+
+        const providerQuery = params.provider?.trim().toLowerCase();
+        const modelQuery = params.model?.trim().toLowerCase();
+        const filtered = providerQuery !== undefined || modelQuery !== undefined;
+
+        const matched = ctx.modelRegistry.getAvailable().filter((model) => {
+          if (providerQuery && !model.provider.toLowerCase().includes(providerQuery)) return false;
+          if (modelQuery && !model.id.toLowerCase().includes(modelQuery) && !model.name.toLowerCase().includes(modelQuery)) return false;
+          return true;
+        }).map(toMetadata);
+        const total = matched.length;
+
+        if (total === 0) {
+          return {
+            content: [{ type: "text", text: `0 models ${filtered ? "matched" : "available"}.` }],
+            details: { action: "list", models: [], total } satisfies ModelToolDetails,
+          };
+        }
+
+        // Convert from 1-indexed offset to 0-indexed array access.
+        const startIndex = params.offset ? Math.max(0, params.offset - 1) : 0;
+        if (startIndex >= total) {
+          throw new Error(`Offset ${params.offset} is beyond end of list (${total} models total).`);
+        }
+
+        const endIndex = params.limit !== undefined ? Math.min(startIndex + params.limit, total) : total;
+        const selected = matched.slice(startIndex, endIndex);
+
+        // JSONL: one compact object per line, so truncation cuts at record boundaries.
+        const truncation = truncateHead(selected.map((model) => JSON.stringify(model)).join("\n"), {
+          maxLines: DEFAULT_MAX_LINES,
+          maxBytes: DEFAULT_MAX_BYTES,
+        });
+
+        const delivered = truncation.truncated ? selected.slice(0, truncation.outputLines) : selected;
+        const endDisplay = startIndex + delivered.length;
+        const nextOffset = endDisplay + 1;
+
+        let text = truncation.content;
+        if (truncation.truncated) {
+          text += `\n\n[Truncated: showing models ${startIndex + 1}-${endDisplay} of ${total}; use offset=${nextOffset} to continue]`;
+        } else if (endDisplay < total) {
+          text += `\n\n[${total - endDisplay} more models in list; use offset=${nextOffset} to continue]`;
+        }
+
+        return {
+          content: [{ type: "text", text }],
+          details: { action: "list", models: delivered, total, truncation } satisfies ModelToolDetails,
+        };
+      },
+    });
+  });
+}

package/extensions/shared/config/schema.ts

@@ -3,6 +3,7 @@ import * as z from "zod";
 import { editorConfigSchema } from "../../editor/config";
 import { footerConfigSchema } from "../../footer/config";
 import { fullscreenConfigSchema } from "../../fullscreen/config";
+import { modelsConfigSchema } from "../../models/config";
 import { nameConfigSchema } from "../../name/config";
 import { presetsConfigSchema } from "../../presets/config";
 import { recapConfigSchema } from "../../recap/config";
@@ -11,6 +12,7 @@ export const configSchemas = {
   editor: editorConfigSchema,
   footer: footerConfigSchema,
   fullscreen: fullscreenConfigSchema,
+  models: modelsConfigSchema,
   name: nameConfigSchema,
   presets: presetsConfigSchema,
   recap: recapConfigSchema,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-spark",
-  "version": "0.8.0",
+  "version": "0.9.1",
   "description": "A small, opinionated collection of pi extensions",
   "keywords": [
     "pi-coding-agent",