255 lines
9.3 KiB
TypeScript
255 lines
9.3 KiB
TypeScript
import type {
|
|
ExtensionAPI,
|
|
ExtensionContext,
|
|
BeforeAgentStartEvent,
|
|
BeforeAgentStartEventResult,
|
|
ExtensionCommandContext,
|
|
} from "@earendil-works/pi-coding-agent";
|
|
import { compress } from "headroom-ai";
|
|
|
|
// Global state for manual model lock
|
|
let isLocked = false;
|
|
let lockedModel: any = null;
|
|
|
|
// Tags that trigger Headroom context compression.
|
|
// read / discuss / search skip compression entirely.
|
|
const COMPRESS_TAGS = new Set([
|
|
"devops-low",
|
|
"devops-high",
|
|
"code-analysis-low",
|
|
"code-analysis-high",
|
|
"codewrite-low",
|
|
"codewrite-high",
|
|
]);
|
|
|
|
// Minimum message size (in chars) before compression activates.
|
|
// ~5K tokens ≈ 20K characters (rough 4:1 ratio).
|
|
const COMPRESS_MIN_CHARS = 20_000;
|
|
|
|
// Current routing tag for the active turn (used by context handler)
|
|
let currentTag: string | null = null;
|
|
|
|
// Model ID mappings for routing
|
|
// Cost-conscious: flash/economy for most tasks, Pro only for genuinely complex work
|
|
const MODELS: Record<string, { provider: string; id: string }> = {
|
|
"free-core": { provider: "openrouter", id: "free" },
|
|
"router-eval": { provider: "openrouter", id: "free" },
|
|
"economy-devops": { provider: "opencode-go", id: "deepseek/deepseek-v4-flash" },
|
|
"economy-code": { provider: "opencode-go", id: "deepseek/deepseek-v4-flash" },
|
|
"precision-devops": { provider: "opencode-go", id: "deepseek/deepseek-v4-flash" },
|
|
"precision-code-high":{ provider: "opencode-go", id: "deepseek/deepseek-v4-pro" },
|
|
"precision-react": { provider: "opencode-go", id: "deepseek/deepseek-v4-pro" },
|
|
"context-heavy": { provider: "openrouter", id: "free" },
|
|
};
|
|
|
|
// Thinking level mappings — only set for tasks that truly benefit.
|
|
// Everything else keeps the default (off) for speed and cost.
|
|
const THINKING: Record<string, string> = {
|
|
"precision-code-high":"high",
|
|
"precision-react": "high",
|
|
};
|
|
|
|
function getModel(ctx: ExtensionContext, key: string): any {
|
|
const mapping = MODELS[key];
|
|
if (!mapping) return undefined;
|
|
return ctx.modelRegistry.find(mapping.provider, mapping.id);
|
|
}
|
|
|
|
function modelLabel(key: string): string {
|
|
const m = MODELS[key];
|
|
if (!m) return key;
|
|
// Return just the model name for short display (e.g. "v4-flash", "v4-pro", "free")
|
|
const id = m.id.includes("/") ? m.id : `${m.provider}/${m.id}`;
|
|
const parts = id.split("/");
|
|
return parts[parts.length - 1];
|
|
}
|
|
|
|
export default function (pi: ExtensionAPI) {
|
|
|
|
// 1. Register /lock-model command
|
|
pi.registerCommand("lock-model", {
|
|
description: "Lock to a specific model. Disables dynamic routing until /unlock-model.",
|
|
handler: async (args: string, ctx: ExtensionCommandContext) => {
|
|
const modelId = (args || "").trim();
|
|
if (!modelId) {
|
|
ctx.ui.notify(
|
|
"Usage: /lock-model <provider/model-id>\nExample: /lock-model openrouter/anthropic/claude-3.5-sonnet",
|
|
"error"
|
|
);
|
|
return;
|
|
}
|
|
const slashIdx = modelId.indexOf("/");
|
|
if (slashIdx < 0) {
|
|
ctx.ui.notify("Invalid model ID format. Use: provider/model-id", "error");
|
|
return;
|
|
}
|
|
const provider = modelId.substring(0, slashIdx);
|
|
const id = modelId.substring(slashIdx + 1);
|
|
let model = ctx.modelRegistry.find(provider, id);
|
|
if (!model) {
|
|
// OpenRouter model IDs include provider prefix (e.g. "openrouter/owl-alpha")
|
|
// so try with the full provider/model-id format
|
|
model = ctx.modelRegistry.find(provider, `${provider}/${id}`);
|
|
}
|
|
if (!model) {
|
|
ctx.ui.notify(`Model not found: ${modelId}`, "error");
|
|
return;
|
|
}
|
|
isLocked = true;
|
|
lockedModel = model;
|
|
pi.setModel(model);
|
|
ctx.ui.notify(`🔒 Router disabled. Model locked to: ${modelId}`, "info");
|
|
},
|
|
});
|
|
|
|
// 2. Register /unlock-model command
|
|
pi.registerCommand("unlock-model", {
|
|
description: "Re-enable dynamic prompt routing.",
|
|
handler: async (_args: string, ctx: ExtensionCommandContext) => {
|
|
isLocked = false;
|
|
lockedModel = null;
|
|
ctx.ui.notify("🔓 Router enabled. Prompts will be dynamically analyzed and routed.", "info");
|
|
},
|
|
});
|
|
|
|
// 3. Intercept prompts before agent starts
|
|
pi.on("before_agent_start", async (event: BeforeAgentStartEvent, ctx: ExtensionContext): Promise<BeforeAgentStartEventResult | undefined> => {
|
|
const prompt = event.prompt;
|
|
|
|
// Pass-through if user locked the model manually
|
|
if (isLocked && lockedModel) {
|
|
pi.setModel(lockedModel);
|
|
ctx.ui.setStatus("router", `🔒 ${lockedModel.id || "unknown"}`);
|
|
return;
|
|
}
|
|
|
|
// Skip routing for short prompts / simple greetings
|
|
if (prompt.trim().length < 15) {
|
|
const model = getModel(ctx, "router-eval");
|
|
if (model) pi.setModel(model);
|
|
ctx.ui.setStatus("router", `⚡ ${modelLabel("router-eval")}`);
|
|
return;
|
|
}
|
|
|
|
// Show analyzing indicator in status bar
|
|
ctx.ui.setStatus("router", "🤔 Analyzing...");
|
|
|
|
try {
|
|
const OPENROUTER_API_KEY = process.env["OPENROUTER_API_KEY"] || "";
|
|
|
|
const response = await fetch("https://openrouter.ai/api/v1/chat/completions", {
|
|
method: "POST",
|
|
headers: {
|
|
"Authorization": `Bearer ${OPENROUTER_API_KEY}`,
|
|
"Content-Type": "application/json",
|
|
},
|
|
body: JSON.stringify({
|
|
model: "openrouter/owl-alpha",
|
|
messages: [
|
|
{
|
|
role: "system",
|
|
content:
|
|
'You are a minimal command interceptor router. Analyze the prompt and output ONLY a minified JSON object containing the tag and language. No markdown, no conversation.\n\nValid Tags:\n- "read" (reading/ingesting logs, docs, read-only code)\n- "discuss" (general chat, architectural concepts)\n- "search" (requires web search/external lookups)\n- "devops-low" (editing yaml, single dockerfile, basic bash, env adjustments)\n- "devops-high" (complex multi-container network layers, server crashes)\n- "code-analysis-low" (finding a bug in a short file)\n- "code-analysis-high" (refactoring multiple complex files, tracking syntax errors)\n- "codewrite-low" (generating simple syntax boilerplate or standard functions)\n- "codewrite-high" (complex features, heavy React state, complex PHP architectures)\n\nRespond exactly like this: {"tag":"tag-name","lang":"react|php|python|rust|bash|none"}',
|
|
},
|
|
{ role: "user", content: prompt },
|
|
],
|
|
max_tokens: 35,
|
|
temperature: 0,
|
|
}),
|
|
});
|
|
|
|
const payload = await response.json();
|
|
const content = payload?.choices?.[0]?.message?.content?.trim();
|
|
if (!content) {
|
|
const fallback = getModel(ctx, "free-core");
|
|
if (fallback) pi.setModel(fallback);
|
|
ctx.ui.setStatus("router", `⚠️ ${modelLabel("free-core")}`);
|
|
return;
|
|
}
|
|
|
|
const decision = JSON.parse(content);
|
|
const tag: string = decision.tag || "read";
|
|
const lang: string = decision.lang || "none";
|
|
|
|
// Deterministic routing tree
|
|
let modelKey: string;
|
|
switch (tag) {
|
|
case "read":
|
|
case "discuss":
|
|
case "search":
|
|
modelKey = "free-core";
|
|
break;
|
|
case "devops-low":
|
|
modelKey = "economy-devops";
|
|
break;
|
|
case "devops-high":
|
|
modelKey = "precision-devops";
|
|
break;
|
|
case "code-analysis-low":
|
|
modelKey = "free-core";
|
|
break;
|
|
case "code-analysis-high":
|
|
modelKey = "context-heavy";
|
|
break;
|
|
case "codewrite-low":
|
|
modelKey = "economy-code";
|
|
break;
|
|
case "codewrite-high":
|
|
modelKey = lang === "react" ? "precision-react" : "precision-code-high";
|
|
break;
|
|
default:
|
|
modelKey = "free-core";
|
|
}
|
|
|
|
const model = getModel(ctx, modelKey);
|
|
if (model) {
|
|
pi.setModel(model);
|
|
}
|
|
|
|
// Set thinking level based on task complexity
|
|
const thinkingLevel = THINKING[modelKey];
|
|
if (thinkingLevel) {
|
|
pi.setThinkingLevel(thinkingLevel as any);
|
|
}
|
|
|
|
// Store tag for compression check in context event
|
|
currentTag = tag;
|
|
|
|
// Show routing decision in status bar (shortened to avoid truncation)
|
|
ctx.ui.setStatus("router", `${modelLabel(modelKey)}`);
|
|
|
|
} catch (_error) {
|
|
// Fallback gracefully on network drops
|
|
const fallback = getModel(ctx, "free-core");
|
|
if (fallback) pi.setModel(fallback);
|
|
ctx.ui.setStatus("router", `⚠️ fallback ${modelLabel("free-core")}`);
|
|
}
|
|
});
|
|
|
|
// 4. Compress large contexts before LLM turns (Headroom)
|
|
pi.on("context", async (event, ctx) => {
|
|
// Only compress for analysis/coding/devops tags
|
|
if (!currentTag || !COMPRESS_TAGS.has(currentTag)) return;
|
|
|
|
// Quick size check before calling the proxy
|
|
const totalChars = JSON.stringify(event.messages).length;
|
|
if (totalChars < COMPRESS_MIN_CHARS) return;
|
|
|
|
try {
|
|
const result = await compress(event.messages, {
|
|
baseUrl: "http://192.168.20.13:8787",
|
|
fallback: true,
|
|
timeout: 15_000,
|
|
});
|
|
if (result.messages && result.messages.length > 0) {
|
|
const saved = ((result.tokensBefore - result.tokensAfter) / result.tokensBefore * 100).toFixed(0);
|
|
ctx.ui.setStatus("compression", `📦${saved}%`);
|
|
return { messages: result.messages };
|
|
}
|
|
} catch {
|
|
// Proxy down — pass through (fallback: true already handles transport errors)
|
|
ctx.ui.setStatus("compression", "⚠️");
|
|
}
|
|
});
|
|
}
|