Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.
Python agent-contextagent-memoryagent-pluginsagentic-ragcontext-databasedsh-pluginknowledge-basememorymemory-managementself-evolvingvector-database
/** MCP content adapted to pi's text/image tool results. */
export const MAX_RESULT_BYTES = 50 * 1024;
export const MAX_RESULT_LINES = 2000;
const isPlainObject = (value) => value !== null && typeof value === "object" && !Array.isArray(value);
export function joinText(blocks) {
return (Array.isArray(blocks) ? blocks : [])
.filter((block) => block && block.type === "text")
.map((block) => String(block.text ?? ""))
.join("\n");
}
/**
* An MCP `tools/call` result as pi content blocks.
*
* pi's content union is text or image; everything else becomes one readable
* text line so the model still learns what came back.
*/
export function mcpContentToPi(result) {
const blocks = Array.isArray(result?.content) ? result.content : [];
const out = [];
for (const block of blocks) {
if (!isPlainObject(block)) continue;
if (block.type === "text") {
out.push({ type: "text", text: String(block.text ?? "") });
} else if (block.type === "image") {
out.push({
type: "image",
data: String(block.data ?? ""),
mimeType: String(block.mimeType ?? block.mime_type ?? "image/png"),
});
} else if (block.type === "audio") {
out.push({ type: "text", text: `[audio content omitted (${String(block.mimeType ?? "unknown type")})]` });
} else if (block.type === "resource") {
const resource = isPlainObject(block.resource) ? block.resource : {};
const inline = typeof resource.text === "string" ? `\n${resource.text}` : "";
const mime = resource.mimeType ? ` (${resource.mimeType})` : "";
out.push({ type: "text", text: `[resource ${String(resource.uri ?? "?")}${mime}]${inline}` });
} else if (block.type === "resource_link") {
const label = block.name ? ` — ${block.name}` : "";
out.push({ type: "text", text: `[resource link ${String(block.uri ?? "?")}${label}]` });
} else {
out.push({ type: "text", text: `[unsupported MCP content block: ${String(block.type ?? "?")}]` });
}
}
// structuredContent is appended only when it says something the upstream text
// does not. FastMCP echoes a {"result": "<the same text>"} beside the text
// block for every tool annotated `-> str`, so the comparison must be against
// the UPSTREAM text — never against the notes synthesised above for
// image/audio/resource blocks, which would make every such echo look new.
const structured = result?.structuredContent;
if (isPlainObject(structured)) {
const text = joinText(blocks).trim();
const serialized = JSON.stringify(structured);
const duplicate =
(typeof structured.result === "string" && structured.result.trim() === text)
|| serialized.trim() === text;
if (!duplicate && serialized) out.push({ type: "text", text: serialized });
}
if (out.length === 0) out.push({ type: "text", text: "" });
return out;
}
/** All text blocks share one budget, including the truncation notice. */
export function toPiResult(tool, result) {
const content = mcpContentToPi(result);
const text = joinText(content);
const truncated = Buffer.byteLength(text) > MAX_RESULT_BYTES || text.split("\n").length > MAX_RESULT_LINES;
if (truncated) {
const hint = "\n[OpenViking] Output truncated. Request fewer items or use a narrower URI, offset or limit.";
const lines = text.split("\n").slice(0, MAX_RESULT_LINES - 1).join("\n");
const buffer = Buffer.from(lines);
let end = Math.min(buffer.length, MAX_RESULT_BYTES - Buffer.byteLength(hint));
while (end > 0 && (buffer[end] & 0xc0) === 0x80) end--;
// Consolidating text keeps the limit independent of MCP block boundaries.
const shortened = buffer.subarray(0, end).toString("utf8") + hint;
if (result.isError) throw new Error(shortened);
return {
content: [{ type: "text", text: shortened }, ...content.filter((block) => block.type === "image")],
details: { tool, truncated: true },
};
}
if (result.isError) throw new Error(text.trim() || "OpenViking " + tool + " failed");
return { content, details: { tool, truncated: false } };
}