github-router 0.3.11 → 0.3.13

package/dist/main.js

@@ -6,6 +6,8 @@ import os from "node:os";
 import path from "node:path";
 import { randomBytes, randomUUID } from "node:crypto";
 import process$1 from "node:process";
+import fs$1 from "node:fs";
+import { Writable } from "node:stream";
 import { execFileSync, spawn } from "node:child_process";
 import { serve } from "srvx";
 import { getProxyForUrl } from "proxy-from-env";
@@ -19,9 +21,11 @@ import clipboard from "clipboardy";
 //#region src/lib/paths.ts
 const APP_DIR = path.join(os.homedir(), ".local", "share", "github-router");
 const GITHUB_TOKEN_PATH = path.join(APP_DIR, "github_token");
+const ERROR_LOG_PATH = path.join(APP_DIR, "error.log");
 const PATHS = {
 	APP_DIR,
-	GITHUB_TOKEN_PATH
+	GITHUB_TOKEN_PATH,
+	ERROR_LOG_PATH
 };
 async function ensurePaths() {
 	await fs.mkdir(PATHS.APP_DIR, { recursive: true });
@@ -223,19 +227,68 @@ function filterBetaHeader(value) {
 	return value.split(",").map((v) => v.trim()).filter((v) => v && ALLOWED_BETA_PREFIXES.some((prefix) => v.startsWith(prefix))).join(",") || void 0;
 }
 /**
+* Normalize a model ID for fuzzy comparison: lowercase, replace dots with
+* dashes, insert dash at letter→digit boundaries, and collapse repeated
+* dashes. E.g. "gpt5.3-codex" → "gpt-5-3-codex", "GPT-5.3-Codex" → "gpt-5-3-codex".
+*/
+function normalizeModelId(id) {
+	return id.toLowerCase().replace(/\./g, "-").replace(/([a-z])(\d)/g, "$1-$2").replace(/-{2,}/g, "-");
+}
+/**
 * Resolve a model name to the best available variant in the Copilot model list.
-* Prefers the 1M context variant for opus models.
+*
+* Resolution cascade:
+* 1. Exact match
+* 2. Case-insensitive match
+* 3. Family preference (opus→1m, codex→highest version)
+* 4. Normalized match (dots→dashes, letter-digit boundaries)
+* 5. Return as-is with a warning
 */
 function resolveModel(modelId) {
 	const models = state.models?.data;
 	if (!models) return modelId;
 	if (models.some((m) => m.id === modelId)) return modelId;
-	if (modelId.toLowerCase().includes("opus")) {
+	const lower = modelId.toLowerCase();
+	const ciMatch = models.find((m) => m.id.toLowerCase() === lower);
+	if (ciMatch) return ciMatch.id;
+	if (lower.includes("opus")) {
 		const oneM = models.find((m) => m.id.includes("opus") && m.id.endsWith("-1m"));
 		if (oneM) return oneM.id;
 	}
+	if (lower.includes("codex")) {
+		const codexModels = models.filter((m) => m.id.includes("codex") && !m.id.includes("mini"));
+		if (codexModels.length > 0) {
+			codexModels.sort((a, b) => b.id.localeCompare(a.id));
+			return codexModels[0].id;
+		}
+	}
+	const normalized = normalizeModelId(modelId);
+	const normMatch = models.find((m) => normalizeModelId(m.id) === normalized);
+	if (normMatch) return normMatch.id;
+	consola.warn(`Model "${modelId}" not found in Copilot model list. Available: ${models.map((m) => m.id).join(", ")}`);
 	return modelId;
 }
+/**
+* Resolve a codex model ID, falling back to the best available codex model.
+* Used by the codex subcommand for model selection.
+*/
+function resolveCodexModel(modelId) {
+	const resolved = resolveModel(modelId);
+	const models = state.models?.data;
+	if (!models) return resolved;
+	if (models.some((m) => m.id === resolved)) return resolved;
+	const codexModels = models.filter((m) => {
+		const endpoints = m.supported_endpoints ?? [];
+		return m.id.includes("codex") && !m.id.includes("mini") && (endpoints.length === 0 || endpoints.includes("/responses"));
+	});
+	if (codexModels.length > 0) {
+		codexModels.sort((a, b) => b.id.localeCompare(a.id));
+		const best = codexModels[0].id;
+		consola.warn(`Model "${modelId}" not available, using "${best}" instead`);
+		return best;
+	}
+	return resolved;
+}
 async function cacheModels() {
 	state.models = await getModels();
 }
@@ -413,10 +466,104 @@ const checkUsage = defineCommand({
 	}
 });
 
+//#endregion
+//#region src/lib/file-log-reporter.ts
+const MAX_LOG_BYTES = 1024 * 1024;
+const DEDUP_MAX = 1e3;
+const ARG_MAX_LEN = 2048;
+const DEDUP_KEY_MAX_LEN = 200;
+const CREDENTIAL_RE = /\b(eyJ[A-Za-z0-9_-]{20,}(?:\.[A-Za-z0-9_-]+){0,2}|gh[opsu]_[A-Za-z0-9_]{20,}|Bearer\s+\S{20,})\b/g;
+const ALLOWED_TYPES = new Set([
+	"fatal",
+	"error",
+	"warn"
+]);
+function sanitize(line) {
+	return line.replace(CREDENTIAL_RE, "[REDACTED]");
+}
+function serializeArg(arg) {
+	if (typeof arg === "string") return arg;
+	if (arg instanceof Error) {
+		const parts = [arg.message];
+		if (arg.stack) parts.push(arg.stack);
+		return parts.join("\n");
+	}
+	return String(arg);
+}
+function formatLogLine(logObj) {
+	return sanitize(`${logObj.date.toISOString()} [${(logObj.type ?? "error").toUpperCase()}] ${logObj.args.map((a) => {
+		const s = serializeArg(a);
+		return s.length > ARG_MAX_LEN ? s.slice(0, ARG_MAX_LEN) + "…" : s;
+	}).join(" ").replace(/\r\n|\r|\n/g, "\\n")}\n`);
+}
+function makeDedupeKey(logObj) {
+	const firstArg = logObj.args.length > 0 ? serializeArg(logObj.args[0]) : "";
+	const key = `${logObj.type}:${firstArg}`;
+	return key.length > DEDUP_KEY_MAX_LEN ? key.slice(0, DEDUP_KEY_MAX_LEN) : key;
+}
+function rotateIfNeeded(filePath) {
+	let size;
+	try {
+		size = fs$1.statSync(filePath).size;
+	} catch {
+		return;
+	}
+	if (size <= MAX_LOG_BYTES) return;
+	try {
+		fs$1.renameSync(filePath, filePath + ".1");
+	} catch {}
+}
+var FileLogReporter = class {
+	filePath;
+	seen = /* @__PURE__ */ new Set();
+	writing = false;
+	constructor(filePath) {
+		this.filePath = filePath;
+		rotateIfNeeded(filePath);
+	}
+	log(logObj, _ctx) {
+		if (!ALLOWED_TYPES.has(logObj.type)) return;
+		if (this.writing) return;
+		const key = makeDedupeKey(logObj);
+		if (this.seen.has(key)) return;
+		if (this.seen.size >= DEDUP_MAX) this.seen.clear();
+		this.seen.add(key);
+		const line = formatLogLine(logObj);
+		this.writing = true;
+		try {
+			const fd = fs$1.openSync(this.filePath, "a", 384);
+			fs$1.writeSync(fd, line);
+			fs$1.closeSync(fd);
+		} catch {} finally {
+			this.writing = false;
+		}
+	}
+};
+const nullStream = new Writable({ write(_chunk, _encoding, cb) {
+	cb();
+} });
+/**
+* Switch consola to file-only mode for TUI sessions.
+* Removes the terminal reporter and installs a file reporter that
+* persists errors and warnings to disk with dedup and credential scrubbing.
+*
+* Also sinks consola's stdout/stderr streams as belt-and-suspenders:
+* even if a terminal reporter is re-added, it cannot write to the terminal.
+* Crash handlers that call process.stderr.write() directly are unaffected.
+* FileLogReporter uses fs.writeSync() directly and is also unaffected.
+*/
+function enableFileLogging() {
+	const reporter = new FileLogReporter(PATHS.ERROR_LOG_PATH);
+	consola.options.throttle = 0;
+	consola.setReporters([reporter]);
+	consola.options.stdout = nullStream;
+	consola.options.stderr = nullStream;
+}
+
 //#endregion
 //#region src/lib/port.ts
 const DEFAULT_PORT = 8787;
-const DEFAULT_CODEX_MODEL = "gpt5.3-codex";
+const DEFAULT_CODEX_MODEL = "gpt-5.3-codex";
 const PORT_RANGE_MIN = 11e3;
 const PORT_RANGE_MAX = 65535;
 /** Generate a random port number in the range [11000, 65535]. */
@@ -442,6 +589,7 @@ function buildLaunchCommand(target) {
 			...target.extraArgs
 		] : [
 			"codex",
+			"--full-auto",
 			"-m",
 			target.model ?? DEFAULT_CODEX_MODEL,
 			...target.extraArgs
@@ -456,17 +604,26 @@ function launchChild(target, server$1) {
 	const { cmd, env } = buildLaunchCommand(target);
 	const executable = cmd[0];
 	if (!commandExists(executable)) {
-		consola.error(`"${executable}" not found on PATH. Install it first, then try again.`);
+		const msg = `"${executable}" not found on PATH. Install it first, then try again.`;
+		consola.error(msg);
+		process$1.stderr.write(msg + "\n");
 		process$1.exit(1);
 	}
 	let child;
 	try {
-		child = spawn(cmd[0], cmd.slice(1), {
+		if (process$1.platform === "win32") child = spawn(cmd.map((a) => a.includes(" ") ? `"${a}"` : a).join(" "), [], {
+			env,
+			stdio: "inherit",
+			shell: true
+		});
+		else child = spawn(cmd[0], cmd.slice(1), {
 			env,
 			stdio: "inherit"
 		});
 	} catch (error) {
-		consola.error(`Failed to launch ${executable}:`, error instanceof Error ? error.message : String(error));
+		const msg = `Failed to launch ${executable}: ${error instanceof Error ? error.message : String(error)}`;
+		consola.error(msg);
+		process$1.stderr.write(msg + "\n");
 		server$1.close(true).catch(() => {});
 		process$1.exit(1);
 	}
@@ -494,10 +651,61 @@ function launchChild(target, server$1) {
 	};
 	process$1.on("SIGINT", onSignal);
 	process$1.on("SIGTERM", onSignal);
-	child.on("exit", (exitCode) => {
-		cleanup().then(() => exit(exitCode ?? 0)).catch(() => exit(1));
+	child.on("exit", (exitCode, signal) => {
+		const code = exitCode ?? (signal ? 128 : 1);
+		cleanup().then(() => exit(code)).catch(() => exit(1));
+	});
+	child.on("error", () => {
+		cleanup().then(() => exit(1)).catch(() => exit(1));
 	});
-	child.on("error", () => exit(1));
+}
+
+//#endregion
+//#region src/lib/model-validation.ts
+const ENDPOINT_ALIASES = {
+	"/chat/completions": "/chat/completions",
+	"/v1/chat/completions": "/chat/completions",
+	"/responses": "/responses",
+	"/v1/responses": "/responses",
+	"/v1/messages": "/v1/messages"
+};
+/**
+* Check whether a model supports the given endpoint, based on cached
+* `supported_endpoints` metadata from the Copilot `/models` response.
+*
+* Returns `true` (allow) when:
+* - the model is not found in the cache (don't block unknown models)
+* - the model has no `supported_endpoints` field (backward-compat)
+* - the endpoint is listed in `supported_endpoints`
+*/
+function modelSupportsEndpoint(modelId, path$1) {
+	const endpoint = ENDPOINT_ALIASES[path$1] ?? path$1;
+	const model = state.models?.data.find((m) => m.id === modelId);
+	if (!model) return true;
+	const supported = model.supported_endpoints;
+	if (!supported || supported.length === 0) return true;
+	return supported.includes(endpoint);
+}
+/**
+* Log an error when a model is used on an endpoint it doesn't support.
+* Returns `true` if a mismatch was detected (for testing).
+*/
+function logEndpointMismatch(modelId, path$1) {
+	if (modelSupportsEndpoint(modelId, path$1)) return false;
+	const supported = (state.models?.data.find((m) => m.id === modelId))?.supported_endpoints ?? [];
+	consola.error(`Model "${modelId}" does not support ${path$1}. Supported endpoints: ${supported.join(", ")}`);
+	return true;
+}
+/**
+* Return model IDs that support the given endpoint.
+*/
+function listModelsForEndpoint(path$1) {
+	const endpoint = ENDPOINT_ALIASES[path$1] ?? path$1;
+	return (state.models?.data ?? []).filter((m) => {
+		const supported = m.supported_endpoints;
+		if (!supported || supported.length === 0) return true;
+		return supported.includes(endpoint);
+	}).map((m) => m.id);
 }
 
 //#endregion
@@ -594,7 +802,7 @@ function formatTokens(n) {
 function formatTokenInfo(inputTokens, outputTokens, model) {
 	if (inputTokens === void 0) return void 0;
 	const parts = [];
-	const maxPrompt = model?.capabilities.limits.max_prompt_tokens;
+	const maxPrompt = model?.capabilities?.limits?.max_prompt_tokens;
 	if (maxPrompt) {
 		const pct = (inputTokens / maxPrompt * 100).toFixed(1);
 		parts.push(`in:${formatTokens(inputTokens)}/${formatTokens(maxPrompt)} (${pct}%)`);
@@ -714,7 +922,7 @@ const getEncodeChatFunction = async (encoding) => {
 * Get tokenizer type from model information
 */
 const getTokenizerFromModel = (model) => {
-	return model.capabilities.tokenizer || "o200k_base";
+	return model.capabilities?.tokenizer || "o200k_base";
 };
 /**
 * Get model-specific constants for token calculation
@@ -948,6 +1156,7 @@ async function handleCompletion$1(c) {
 	const resolvedModel = resolveModel(payload.model);
 	if (resolvedModel !== payload.model) payload.model = resolvedModel;
 	const selectedModel = state.models?.data.find((model) => model.id === payload.model);
+	logEndpointMismatch(payload.model, "/chat/completions");
 	let inputTokens;
 	try {
 		if (selectedModel) inputTokens = (await getTokenCount(payload, selectedModel)).input;
@@ -955,7 +1164,7 @@ async function handleCompletion$1(c) {
 	if (isNullish(payload.max_tokens)) {
 		payload = {
 			...payload,
-			max_tokens: selectedModel?.capabilities.limits.max_output_tokens
+			max_tokens: selectedModel?.capabilities?.limits?.max_output_tokens
 		};
 		if (debugEnabled) consola.debug("Set max_tokens to:", JSON.stringify(payload.max_tokens));
 	}
@@ -1366,7 +1575,9 @@ async function handleCompletion(c) {
 	if (state.manualApprove) await awaitApproval();
 	const betaHeaders = extractBetaHeaders(c);
 	const { body: resolvedBody, originalModel, resolvedModel } = resolveModelInBody(await processWebSearch(rawBody));
-	const selectedModel = state.models?.data.find((m) => m.id === (resolvedModel ?? originalModel));
+	const modelId = resolvedModel ?? originalModel;
+	const selectedModel = state.models?.data.find((m) => m.id === modelId);
+	if (modelId) logEndpointMismatch(modelId, "/v1/messages");
 	const effectiveBetas = applyDefaultBetas(betaHeaders, resolvedModel ?? originalModel);
 	let response;
 	try {
@@ -1591,12 +1802,9 @@ async function handleResponses(c) {
 	const resolvedModel = resolveModel(payload.model);
 	if (resolvedModel !== payload.model) payload.model = resolvedModel;
 	const selectedModel = state.models?.data.find((model) => model.id === payload.model);
+	logEndpointMismatch(payload.model, "/responses");
 	if (state.manualApprove) await awaitApproval();
 	await injectWebSearchIfNeeded(payload);
-	if (isNullish(payload.max_output_tokens)) {
-		payload.max_output_tokens = selectedModel?.capabilities.limits.max_output_tokens;
-		if (debugEnabled) consola.debug("Set max_output_tokens to:", JSON.stringify(payload.max_output_tokens));
-	}
 	const response = await createResponses(payload, selectedModel?.requestHeaders).catch(async (error) => {
 		if (error instanceof HTTPError) {
 			const errorBody = await error.response.clone().text().catch(() => "");
@@ -1673,31 +1881,99 @@ function extractUserQuery(input) {
 		}
 	}
 }
+/**
+* Compaction prompt used when GitHub Copilot API does not support
+* /responses/compact natively. Matches the prompt Codex CLI uses for
+* local (non-OpenAI) compaction.
+*/
+const COMPACTION_PROMPT = `You are performing a CONTEXT CHECKPOINT COMPACTION. Create a handoff summary for another LLM that will resume the task.
+
+Include:
+- Current progress and key decisions made
+- Important context, constraints, or user preferences
+- What remains to be done (clear next steps)
+- Any critical data, examples, or references needed to continue
+
+Be concise, structured, and focused on helping the next LLM seamlessly continue the work.`;
 async function handleResponsesCompact(c) {
 	const startTime = Date.now();
 	await checkRateLimit(state);
 	if (!state.copilotToken) throw new Error("Copilot token not found");
 	if (state.manualApprove) await awaitApproval();
-	const body = await c.req.text();
+	const body = await c.req.json();
 	const response = await fetch(`${copilotBaseUrl(state)}/responses/compact`, {
 		method: "POST",
 		headers: copilotHeaders(state),
-		body
+		body: JSON.stringify(body)
 	});
-	if (!response.ok) {
+	if (response.ok) {
 		logRequest({
 			method: "POST",
 			path: c.req.path,
-			status: response.status
+			status: 200
 		}, void 0, startTime);
-		throw new HTTPError("Copilot responses/compact request failed", response);
+		return c.json(await response.json());
+	}
+	if (response.status === 404) {
+		consola.debug("Copilot API does not support /responses/compact, using synthetic compaction");
+		return await syntheticCompact(c, body, startTime);
+	}
+	logRequest({
+		method: "POST",
+		path: c.req.path,
+		status: response.status
+	}, void 0, startTime);
+	throw new HTTPError("Copilot responses/compact request failed", response);
+}
+/**
+* Synthetic compaction: sends the conversation history to Copilot's
+* regular /responses endpoint with a compaction prompt appended,
+* then returns the model's summary in the compact response format.
+*/
+async function syntheticCompact(c, body, startTime) {
+	const input = Array.isArray(body.input) ? [...body.input] : [];
+	input.push({
+		type: "message",
+		role: "user",
+		content: [{
+			type: "input_text",
+			text: COMPACTION_PROMPT
+		}]
+	});
+	const payload = {
+		model: body.model,
+		input,
+		instructions: body.instructions,
+		stream: false,
+		store: false
+	};
+	let result;
+	try {
+		result = await createResponses(payload);
+	} catch (error) {
+		if (error instanceof HTTPError) logRequest({
+			method: "POST",
+			path: c.req.path,
+			status: error.response.status
+		}, void 0, startTime);
+		throw error;
 	}
 	logRequest({
 		method: "POST",
 		path: c.req.path,
 		status: 200
 	}, void 0, startTime);
-	return c.json(await response.json());
+	return c.json({
+		id: `resp_compact_${randomUUID().replace(/-/g, "").slice(0, 24)}`,
+		object: "response.compaction",
+		created_at: Math.floor(Date.now() / 1e3),
+		output: result.output,
+		usage: result.usage ?? {
+			input_tokens: 0,
+			output_tokens: 0,
+			total_tokens: 0
+		}
+	});
 }
 
 //#endregion
@@ -1799,7 +2075,7 @@ async function setupAndServe(options) {
 	} else await setupGitHubToken();
 	await setupCopilotToken();
 	await cacheModels();
-	consola.info(`Available models: \n${state.models?.data.map((model) => `- ${model.id}`).join("\n")}`);
+	consola.debug(`Available models: \n${state.models?.data.map((model) => `- ${model.id}`).join("\n")}`);
 	const serveOptions = {
 		fetch: server.fetch,
 		hostname: "127.0.0.1",
@@ -1978,13 +2254,22 @@ const claude = defineCommand({
 			consola.error("Failed to start server:", error instanceof Error ? error.message : error);
 			process$1.exit(1);
 		}
-		consola.success(`Server ready on ${serverUrl}, launching Claude Code...`);
-		consola.level = 1;
+		enableFileLogging();
+		let resolvedModel;
+		if (args.model) {
+			resolvedModel = resolveModel(args.model);
+			if (resolvedModel !== args.model) consola.info(`Model "${args.model}" resolved to "${resolvedModel}"`);
+			if (!state.models?.data.find((m) => m.id === resolvedModel)) {
+				const available = listModelsForEndpoint("/v1/messages");
+				consola.warn(`Model "${resolvedModel}" not found. Available claude models: ${available.join(", ")}`);
+			}
+		}
+		process$1.stderr.write(`Server ready on ${serverUrl}, launching Claude Code...\n`);
 		launchChild({
 			kind: "claude-code",
-			envVars: getClaudeCodeEnvVars(serverUrl, args.model),
+			envVars: getClaudeCodeEnvVars(serverUrl, resolvedModel ?? args.model),
 			extraArgs: args._ ?? [],
-			model: args.model
+			model: resolvedModel ?? args.model
 		}, server$1);
 	}
 });
@@ -2024,14 +2309,24 @@ const codex = defineCommand({
 			consola.error("Failed to start server:", error instanceof Error ? error.message : error);
 			process$1.exit(1);
 		}
-		const codexModel = args.model ?? DEFAULT_CODEX_MODEL;
-		consola.success(`Server ready on ${serverUrl}, launching Codex CLI (${codexModel})...`);
-		consola.level = 1;
+		const requestedModel = args.model ?? DEFAULT_CODEX_MODEL;
+		enableFileLogging();
+		const codexModel = resolveCodexModel(requestedModel);
+		if (codexModel !== requestedModel) consola.info(`Model "${requestedModel}" resolved to "${codexModel}"`);
+		const modelEntry = state.models?.data.find((m) => m.id === codexModel);
+		if (!modelEntry) {
+			const available = listModelsForEndpoint("/responses");
+			consola.warn(`Model "${codexModel}" not found. Available codex models: ${available.join(", ")}`);
+		} else {
+			const ctx = modelEntry.capabilities?.limits?.max_context_window_tokens;
+			if (ctx) consola.info(`Model context window: ${ctx.toLocaleString()} tokens`);
+		}
+		process$1.stderr.write(`Server ready on ${serverUrl}, launching Codex CLI (${codexModel})...\n`);
 		launchChild({
 			kind: "codex",
 			envVars: getCodexEnvVars(serverUrl),
 			extraArgs: args._ ?? [],
-			model: args.model
+			model: codexModel
 		}, server$1);
 	}
 });