pi-sap-aicore 0.3.1 → 0.3.3

package/CHANGELOG.md

@@ -7,6 +7,32 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.3.3] - 2026-07-01
+
+### Added
+
+- Added `npm run validate:foundation`, a live SAP AI Core validation matrix for
+  direct foundation executables. It verifies text generation, real tool execution
+  side effects, and image input across Azure OpenAI, AWS Bedrock, and GCP Vertex
+  AI foundation routes.
+
+### Changed
+
+- Raised SAP AI SDK dependency floors to `^2.12.0` so fresh installs prefer the
+  patched transitive `axios` / `form-data` tree that resolves the npm audit
+  advisory for `form-data <4.0.6`.
+
+## [0.3.2] - 2026-07-01
+
+### Fixed
+
+- Direct AWS Bedrock and GCP Vertex AI foundation routes now advertise pi tools
+  using provider-native tool/function declaration schemas, so Claude and Gemini
+  foundation models can perform real coding-agent tool calls instead of acting as
+  text-only generators.
+- Vertex/Gemini foundation tool-call replay now preserves Gemini thought signatures
+  and normalizes provider-prefixed function names before emitting pi tool calls.
+
 ## [0.3.1] - 2026-07-01
 
 ### Added
@@ -135,7 +161,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   `reasoning_effort` for OpenAI).
 - MIT license and npm packaging.
 
-[Unreleased]: https://github.com/ttiimmaahh/pi-sap-aicore/compare/v0.3.1...HEAD
+[Unreleased]: https://github.com/ttiimmaahh/pi-sap-aicore/compare/v0.3.3...HEAD
+[0.3.3]: https://github.com/ttiimmaahh/pi-sap-aicore/compare/v0.3.2...v0.3.3
+[0.3.2]: https://github.com/ttiimmaahh/pi-sap-aicore/compare/v0.3.1...v0.3.2
 [0.3.1]: https://github.com/ttiimmaahh/pi-sap-aicore/compare/v0.3.0...v0.3.1
 [0.3.0]: https://github.com/ttiimmaahh/pi-sap-aicore/compare/v0.2.2...v0.3.0
 [0.2.2]: https://github.com/ttiimmaahh/pi-sap-aicore/compare/v0.2.1...v0.2.2

package/README.md

@@ -132,8 +132,9 @@ The extension registers **two providers**, both backed by the same service key:
 |---|---|---|
 | SAP deployment | one orchestration deployment fronts **every** model | one foundation deployment **per model** |
 | Models | Claude, GPT-5*, Gemini | GPT/OpenAI (`azure-openai`), Anthropic/Claude (`aws-bedrock`), and Gemini (`gcp-vertexai`) |
+| Tool use | yes | yes — tools are translated to OpenAI, Bedrock Converse, or Vertex/Gemini function declarations by executable |
 | Streaming | subject to orchestration's per-model allow-list — new models can 400 `Streaming is not supported` (we fall back to non-streaming) | Azure OpenAI streams natively; AWS Bedrock and Vertex AI currently use non-streaming endpoints and replay responses into pi stream events |
-| Reasoning effort | tunable (`reasoning_effort` / `thinking`) | model **default** only for Azure; Bedrock/Anthropic and Vertex/Gemini thinking controls are not fully wired yet |
+| Reasoning effort | tunable (`reasoning_effort` / `thinking`) | model **default** only for Azure; Bedrock/Anthropic and Vertex/Gemini effort controls are intentionally conservative |
 | Content filter / grounding / templating | yes | no — raw model access |
 | SDK / endpoint | `@sap-ai-sdk/orchestration` | `AzureOpenAiChatClient` for `azure-openai`; SAP `/inference/deployments/{id}/converse` for `aws-bedrock`; SAP `/inference/deployments/{id}/models/{model}:generateContent` for `gcp-vertexai` |
 
@@ -159,6 +160,26 @@ there. An id with no matching deployment 404s at call time. Run
 `/sap-models discover` in pi (or `node scripts/list-sap-models.mjs` from this
 repo) to see what your tenant actually deploys.
 
+### Live foundation validation
+
+To validate normal coding-agent behavior across the three direct foundation
+executables, run the live smoke-test matrix from this repo:
+
+```bash
+npm run validate:foundation
+```
+
+The script uses `pi --no-extensions -e ./index.ts` so it tests the local checkout,
+not a globally installed package. It makes real SAP model calls and validates:
+
+- text generation
+- real tool execution via side-effect files
+- image input smoke tests
+
+Default models are `gpt-5.5`, `anthropic--claude-4.8-opus`, and
+`gemini-3.5-flash`. Override them with `GPT_MODEL`, `BEDROCK_MODEL`, and
+`VERTEX_MODEL`; set `SKIP_IMAGE=1` to skip vision tests.
+
 ## Models
 
 The model list is composed of three sources, merged at startup:
@@ -337,7 +358,8 @@ the direct Azure OpenAI SDK pinned to API version `2024-10-21`, which has no
 thinking-level cycle is a no-op there. Anthropic/Claude models use SAP's AWS
 Bedrock `/converse` endpoint; Gemini models use SAP's Vertex AI `generateContent`
 endpoint with `thinkingBudget: 0` by default so small pi output budgets produce
-visible text instead of only hidden thoughts. Use the orchestration route if you
+visible text instead of only hidden thoughts. Tool/function calling is supported
+on all three direct foundation executables. Use the orchestration route if you
 need explicit effort control.
 
 To override budgets per model, edit `thinkingLevelMap` on the relevant entry in
@@ -448,9 +470,10 @@ npmjs.com:
 │   └── publish.yml           # tag-driven npm publish via OIDC trusted publishing
 ├── index.ts                  # ExtensionAPI factory + registerProvider calls (both providers)
 ├── scripts/
-│   ├── update-models.mjs     # maintainer script: fetches models.dev, writes models-snapshot.json
-│   ├── list-sap-models.mjs   # lists models your tenant actually deploys (diff vs snapshot)
-│   └── diagnose-streaming.mjs # probes orchestration streaming support per model
+│   ├── update-models.mjs                    # maintainer script: fetches models.dev, writes models-snapshot.json
+│   ├── list-sap-models.mjs                  # lists models your tenant actually deploys (diff vs snapshot)
+│   ├── diagnose-streaming.mjs               # probes orchestration streaming support per model
+│   └── validate-foundation-executables.mjs  # live text/tool/image smoke tests for direct foundation executables
 └── src/
     ├── auth.ts                  # service-key validation + pi oauth registration
     ├── model-catalog.ts         # loads snapshot/cache/overlay and adapts models.dev metadata

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-sap-aicore",
-  "version": "0.3.1",
+  "version": "0.3.3",
   "description": "SAP AI Core (orchestration + foundation) provider for the pi coding agent",
   "license": "MIT",
   "author": "Tim Pearson (https://github.com/ttiimmaahh)",
@@ -31,12 +31,13 @@
   },
   "scripts": {
     "update-models": "node scripts/update-models.mjs",
+    "validate:foundation": "node scripts/validate-foundation-executables.mjs",
     "prepublishOnly": "tsc --noEmit"
   },
   "dependencies": {
-    "@sap-ai-sdk/ai-api": "^2.10.0",
-    "@sap-ai-sdk/foundation-models": "^2.10.0",
-    "@sap-ai-sdk/orchestration": "^2.10.0"
+    "@sap-ai-sdk/ai-api": "^2.12.0",
+    "@sap-ai-sdk/foundation-models": "^2.12.0",
+    "@sap-ai-sdk/orchestration": "^2.12.0"
   },
   "peerDependencies": {
     "@earendil-works/pi-ai": "*",

package/scripts/validate-foundation-executables.mjs

@@ -0,0 +1,178 @@
+#!/usr/bin/env node
+// Live validation for SAP AI Core direct foundation executables.
+//
+// This script makes real, billed model calls through pi using the local extension.
+// It validates the three high-value foundation executable families:
+//   - azure-openai  → GPT/OpenAI models
+//   - aws-bedrock   → Anthropic/Claude models
+//   - gcp-vertexai  → Gemini models
+//
+// Usage from repo root:
+//   node scripts/validate-foundation-executables.mjs
+//
+// Optional model overrides:
+//   GPT_MODEL=gpt-5.5 \
+//   BEDROCK_MODEL=anthropic--claude-4.8-opus \
+//   VERTEX_MODEL=gemini-3.5-flash \
+//   node scripts/validate-foundation-executables.mjs
+//
+// Optional: SKIP_IMAGE=1 to skip vision tests.
+
+import { mkdtempSync, readFileSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { spawnSync } from "node:child_process";
+
+const ROOT = new URL("..", import.meta.url).pathname;
+const PI = process.env.PI_BIN ?? "pi";
+const SKIP_IMAGE = process.env.SKIP_IMAGE === "1";
+
+const SCENARIOS = [
+	{
+		name: "azure-openai/gpt",
+		model: process.env.GPT_MODEL ?? "gpt-5.5",
+		marker: "GPT_TOOL_OK",
+	},
+	{
+		name: "aws-bedrock/anthropic",
+		model: process.env.BEDROCK_MODEL ?? "anthropic--claude-4.8-opus",
+		marker: "BEDROCK_TOOL_OK",
+	},
+	{
+		name: "gcp-vertexai/gemini",
+		model: process.env.VERTEX_MODEL ?? "gemini-3.5-flash",
+		marker: "VERTEX_TOOL_OK",
+	},
+];
+
+const tinyPng = Buffer.from(
+	"iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mP8/x8AAwMCAO+/p9sAAAAASUVORK5CYII=",
+	"base64",
+);
+
+function runPi(args, options = {}) {
+	const result = spawnSync(PI, args, {
+		cwd: ROOT,
+		encoding: "utf8",
+		maxBuffer: 10 * 1024 * 1024,
+		...options,
+	});
+	return {
+		status: result.status ?? 1,
+		stdout: result.stdout ?? "",
+		stderr: result.stderr ?? "",
+		combined: `${result.stdout ?? ""}${result.stderr ?? ""}`,
+	};
+}
+
+function baseArgs(model) {
+	return [
+		"--no-extensions",
+		"-e",
+		"./index.ts",
+		"--model",
+		`sap-aicore-foundation/${model}`,
+		"--no-context-files",
+		"--no-skills",
+		"-p",
+	];
+}
+
+function assert(condition, message, detail) {
+	if (condition) return;
+	console.error(`\n❌ ${message}`);
+	if (detail) console.error(detail);
+	process.exitCode = 1;
+	throw new Error(message);
+}
+
+function truncate(s, max = 2000) {
+	return s.length > max ? `${s.slice(0, max)}…[+${s.length - max} chars]` : s;
+}
+
+function validateText(scenario) {
+	const prompt = "Reply with exactly the single token: TEXT_OK";
+	const result = runPi([...baseArgs(scenario.model), prompt]);
+	assert(
+		result.status === 0,
+		`${scenario.name} text scenario exited ${result.status}`,
+		truncate(result.combined),
+	);
+	assert(
+		/TEXT_OK/.test(result.combined),
+		`${scenario.name} text scenario did not contain TEXT_OK`,
+		truncate(result.combined),
+	);
+	console.log(`  ✓ text generation`);
+}
+
+function validateToolUse(scenario, workDir) {
+	const file = join(workDir, `${scenario.name.replaceAll("/", "-")}.txt`);
+	const prompt =
+		`Use the bash tool to run exactly: printf ${scenario.marker} > ${file}. ` +
+		"Then say done.";
+	const result = runPi([...baseArgs(scenario.model), prompt]);
+	assert(
+		result.status === 0,
+		`${scenario.name} tool scenario exited ${result.status}`,
+		truncate(result.combined),
+	);
+	let actual = "";
+	try {
+		actual = readFileSync(file, "utf8");
+	} catch {
+		// handled below
+	}
+	assert(
+		actual === scenario.marker,
+		`${scenario.name} tool scenario did not create expected side-effect`,
+		`expected file ${file} to contain ${scenario.marker}\nmodel output:\n${truncate(result.combined)}`,
+	);
+	console.log(`  ✓ tool execution side effect`);
+}
+
+function validateImage(scenario, workDir) {
+	const imagePath = join(workDir, "tiny.png");
+	writeFileSync(imagePath, tinyPng);
+	const prompt =
+		"Look at @" +
+		imagePath +
+		" and reply with exactly IMAGE_OK if you can inspect the attached image.";
+	const result = runPi([...baseArgs(scenario.model), prompt]);
+	assert(
+		result.status === 0,
+		`${scenario.name} image scenario exited ${result.status}`,
+		truncate(result.combined),
+	);
+	assert(
+		/IMAGE_OK/.test(result.combined),
+		`${scenario.name} image scenario did not contain IMAGE_OK`,
+		truncate(result.combined),
+	);
+	console.log(`  ✓ image input smoke test`);
+}
+
+const workDir = mkdtempSync(join(tmpdir(), "pi-sap-aicore-foundation-"));
+console.log(`Live SAP AI Core foundation validation`);
+console.log(`repo: ${ROOT}`);
+console.log(`workDir: ${workDir}`);
+console.log(`skip image: ${SKIP_IMAGE ? "yes" : "no"}\n`);
+
+try {
+	for (const scenario of SCENARIOS) {
+		console.log(`${scenario.name}  model=${scenario.model}`);
+		validateText(scenario);
+		validateToolUse(scenario, workDir);
+		if (!SKIP_IMAGE) validateImage(scenario, workDir);
+		console.log("");
+	}
+	console.log("✅ all foundation executable scenarios passed");
+} catch {
+	// assert() already printed details and set exitCode.
+} finally {
+	if (process.env.KEEP_VALIDATION_ARTIFACTS !== "1") {
+		rmSync(workDir, { recursive: true, force: true });
+	} else {
+		console.log(`kept artifacts: ${workDir}`);
+	}
+}

package/src/stream-foundation-vertexai.ts

@@ -29,6 +29,7 @@ type VertexGenerateContentResponse = {
 			role?: string;
 			parts?: Array<{
 				text?: string;
+				thoughtSignature?: string;
 				functionCall?: { name?: string; args?: unknown };
 			}>;
 		};
@@ -192,14 +193,18 @@ function replayVertexGenerateContentResponse(
 
 		if (part.functionCall) {
 			const contentIndex = output.content.length;
+			const providerName = part.functionCall.name ?? "";
 			const toolCall = {
 				type: "toolCall" as const,
 				id: randomUUID(),
-				name: part.functionCall.name ?? "",
+				name: normalizeVertexFunctionName(providerName),
 				arguments:
 					part.functionCall.args && typeof part.functionCall.args === "object"
 						? (part.functionCall.args as Record<string, unknown>)
 						: {},
+				...(part.thoughtSignature
+					? { thoughtSignature: part.thoughtSignature }
+					: {}),
 			};
 			output.content.push(toolCall);
 			stream.push({ type: "toolcall_start", contentIndex, partial: output });
@@ -212,3 +217,8 @@ function replayVertexGenerateContentResponse(
 		}
 	}
 }
+
+function normalizeVertexFunctionName(name: string): string {
+	const index = name.indexOf(":");
+	return index >= 0 ? name.slice(index + 1) : name;
+}

package/src/translate-foundation-bedrock.ts

@@ -3,6 +3,7 @@ import type {
 	Context,
 	Message,
 	TextContent,
+	Tool,
 	ToolResultMessage,
 	UserMessage,
 } from "@earendil-works/pi-ai";
@@ -35,9 +36,20 @@ export type BedrockConverseMessage = {
 	content: BedrockConverseContentBlock[];
 };
 
+export type BedrockToolConfig = {
+	tools: Array<{
+		toolSpec: {
+			name: string;
+			description: string;
+			inputSchema: { json: Record<string, unknown> };
+		};
+	}>;
+};
+
 export function piContextToBedrockConverse(context: Context): {
 	system?: Array<{ text: string }>;
 	messages: BedrockConverseMessage[];
+	toolConfig?: BedrockToolConfig;
 } {
 	const messages: BedrockConverseMessage[] = [];
 
@@ -46,9 +58,11 @@ export function piContextToBedrockConverse(context: Context): {
 		if (translated) messages.push(translated);
 	}
 
+	const tools = (context.tools ?? []).map(piToolToBedrockToolSpec);
 	return {
 		...(context.systemPrompt ? { system: [{ text: context.systemPrompt }] } : {}),
 		messages: coalesceAdjacentMessages(messages),
+		...(tools.length > 0 ? { toolConfig: { tools } } : {}),
 	};
 }
 
@@ -159,6 +173,18 @@ function imageFormatFromMimeType(mimeType: string): string {
 	return "png";
 }
 
+function piToolToBedrockToolSpec(tool: Tool): BedrockToolConfig["tools"][number] {
+	return {
+		toolSpec: {
+			name: tool.name,
+			description: tool.description,
+			inputSchema: {
+				json: tool.parameters as unknown as Record<string, unknown>,
+			},
+		},
+	};
+}
+
 function coalesceAdjacentMessages(
 	messages: BedrockConverseMessage[],
 ): BedrockConverseMessage[] {

package/src/translate-foundation-vertexai.ts

@@ -3,6 +3,7 @@ import type {
 	Context,
 	Message,
 	TextContent,
+	Tool,
 	ToolResultMessage,
 	UserMessage,
 } from "@earendil-works/pi-ai";
@@ -10,7 +11,7 @@ import type {
 export type VertexPart =
 	| { text: string }
 	| { inlineData: { mimeType: string; data: string } }
-	| { functionCall: { name: string; args: unknown } }
+	| { functionCall: { name: string; args: unknown }; thoughtSignature?: string }
 	| { functionResponse: { name: string; response: Record<string, unknown> } };
 
 export type VertexContent = {
@@ -18,9 +19,18 @@ export type VertexContent = {
 	parts: VertexPart[];
 };
 
+export type VertexTool = {
+	functionDeclarations: Array<{
+		name: string;
+		description: string;
+		parameters: Record<string, unknown>;
+	}>;
+};
+
 export function piContextToVertexGenerateContent(context: Context): {
 	systemInstruction?: { parts: Array<{ text: string }> };
 	contents: VertexContent[];
+	tools?: VertexTool[];
 } {
 	const contents: VertexContent[] = [];
 
@@ -29,11 +39,17 @@ export function piContextToVertexGenerateContent(context: Context): {
 		if (translated) contents.push(translated);
 	}
 
+	const functionDeclarations = (context.tools ?? []).map(
+		piToolToVertexFunctionDeclaration,
+	);
 	return {
 		...(context.systemPrompt
 			? { systemInstruction: { parts: [{ text: context.systemPrompt }] } }
 			: {}),
 		contents: coalesceAdjacentContents(contents),
+		...(functionDeclarations.length > 0
+			? { tools: [{ functionDeclarations }] }
+			: {}),
 	};
 }
 
@@ -68,7 +84,13 @@ function piAssistantToVertexContent(msg: AssistantMessage): VertexContent {
 			parts.push({ text: block.text });
 		} else if (block.type === "toolCall") {
 			parts.push({
-				functionCall: { name: block.name, args: block.arguments },
+				functionCall: {
+					name: block.name,
+					args: block.arguments,
+				},
+				...(block.thoughtSignature
+					? { thoughtSignature: block.thoughtSignature }
+					: {}),
 			});
 		}
 	}
@@ -111,6 +133,16 @@ function toolResultImagesAsUserParts(msg: ToolResultMessage): VertexPart[] {
 		}));
 }
 
+function piToolToVertexFunctionDeclaration(
+	tool: Tool,
+): VertexTool["functionDeclarations"][number] {
+	return {
+		name: tool.name,
+		description: tool.description,
+		parameters: tool.parameters as unknown as Record<string, unknown>,
+	};
+}
+
 function coalesceAdjacentContents(contents: VertexContent[]): VertexContent[] {
 	const result: VertexContent[] = [];
 	for (const content of contents) {