github-router 0.3.38 → 0.3.39

package/dist/main.js

@@ -464,7 +464,7 @@ function resolveModel(modelId) {
 		const oneMs = models.filter((m) => m.id.includes("opus") && /-1m(?:$|-)/.test(m.id));
 		const versionMatch = lower.match(/opus-(\d+)[.-](\d+)/);
 		const requestedVersion = versionMatch ? `${versionMatch[1]}.${versionMatch[2]}` : void 0;
-		const oneM = (requestedVersion ? oneMs.find((m) => m.id.includes(`opus-${requestedVersion}-`)) : void 0) ?? oneMs[0];
+		const oneM = (requestedVersion ? oneMs.find((m) => m.id.includes(`opus-${requestedVersion}-`)) : void 0) ?? (requestedVersion ? void 0 : oneMs[0]);
 		if (oneM) return oneM.id;
 	}
 	if (lower.includes("codex")) {
@@ -1553,7 +1553,11 @@ function buildExpansionPattern(variants) {
 	return "(?:" + variants.join("|") + ")";
 }
 function buildRgArgs(input) {
-	const args = ["--json", "--no-follow"];
+	const args = [
+		"--json",
+		"--no-binary",
+		"--no-follow"
+	];
 	if (input.contextLines > 0) args.push(`-C`, String(input.contextLines));
 	if (!input.expansionPattern && (input.mode === "literal" || input.mode === "ranked")) args.push("-F");
 	if (input.fileGlob && input.fileGlob !== "**/*") args.push("-g", input.fileGlob);
@@ -1596,6 +1600,7 @@ async function parseRgJsonStream(child, opts) {
 		cancelled: false,
 		stdoutBytes: 0
 	};
+	child.stdout.setEncoding("utf8");
 	const rl = createInterface({
 		input: child.stdout,
 		crlfDelay: Infinity
@@ -1615,10 +1620,11 @@ async function parseRgJsonStream(child, opts) {
 				killChild(child);
 				break;
 			}
-			if (rawLine.length === 0) continue;
+			const line = rawLine.includes("\0") ? rawLine.replace(/\0/g, "") : rawLine;
+			if (line.length === 0) continue;
 			let evt;
 			try {
-				evt = JSON.parse(rawLine);
+				evt = JSON.parse(line);
 			} catch {
 				continue;
 			}
@@ -9892,7 +9898,7 @@ const PERSONAS_READ = Object.freeze([
 		toolNameHttp: "codex_critic",
 		model: "gpt-5.5",
 		endpoint: "/v1/responses",
-		description: "Adversarial second opinion on plans, designs, or code tradeoffs. Backed by gpt-5.5 (OpenAI) — different lab than Opus. Pass artifact verbatim.",
+		description: "Adversarial second opinion on plans, designs, or code tradeoffs. Backed by gpt-5.5 (OpenAI, 400K context) — strongest reasoning model in the critic lineup, different lab than Opus. Best for architecture decisions, design reviews, and tradeoff analysis where cross-lab diversity matters. Not for line-level code review (use codex_reviewer). Pass artifact verbatim.",
 		baseInstructions: CRITIC_BASE,
 		agentPrompt: "",
 		writeCapable: false,
@@ -9910,7 +9916,7 @@ const PERSONAS_READ = Object.freeze([
 		toolNameHttp: "gemini_critic",
 		model: "gemini-3.1-pro-preview",
 		endpoint: "/v1/chat/completions",
-		description: "Adversarial second opinion. Backed by gemini-3.1-pro (Google) — third-lab triangulation, strong on long-context and formal reasoning. Pass artifact verbatim.",
+		description: "Adversarial second opinion. Backed by gemini-3.1-pro (Google) — third-lab triangulation, strong on formal reasoning, proofs, and invariants. Useful for cross-checking findings from codex_critic or codex_reviewer when you want a third perspective. Pass artifact verbatim.",
 		baseInstructions: GEMINI_CRITIC_BASE,
 		agentPrompt: "",
 		writeCapable: false,
@@ -9928,7 +9934,7 @@ const PERSONAS_READ = Object.freeze([
 		toolNameHttp: "codex_reviewer",
 		model: "gpt-5.3-codex",
 		endpoint: "/v1/responses",
-		description: "Line-level review of a concrete diff or single file. Backed by gpt-5.3-codex (OpenAI) — code-specialist, narrow-scope. Pass artifact verbatim.",
+		description: "Line-level review of a concrete diff or single file. Backed by gpt-5.3-codex (OpenAI, 400K context) — code-specialist, fastest critic (~16s). Surfaces bugs, edge cases, security issues, and idiom violations at specific line numbers. Not suited for architecture or design review (use codex_critic for plans). Pass artifact verbatim.",
 		baseInstructions: REVIEWER_BASE,
 		agentPrompt: "",
 		writeCapable: false,
@@ -9946,7 +9952,7 @@ const PERSONAS_READ = Object.freeze([
 		toolNameHttp: "opus_critic",
 		model: "claude-opus-4-7",
 		endpoint: "/v1/messages",
-		description: "Adversarial second opinion from a fresh-context Opus 4.7 — cheap same-lab sanity check. Pass artifact verbatim.",
+		description: "Adversarial second opinion from a fresh-context Opus 4.7 — same lab as the lead, limited blind-spot diversity vs cross-lab critics, but has the largest context window (up to 1M tokens on enterprise tiers). Handles large artifacts without decomposition. Fast (~22s), catches confabulation and motivated reasoning. Pass artifact verbatim.",
 		baseInstructions: OPUS_CRITIC_BASE,
 		agentPrompt: "",
 		writeCapable: false,
@@ -10023,41 +10029,33 @@ function buildAgentPrompt(persona, opts) {
 }
 /**
 * Build the awareness snippet appended to the spawned `claude` session's
-* system prompt via `--append-system-prompt`. Non-prescriptive — Claude
-* sees that the peer tools and advisor exist; *when* to invoke is left
-* to Claude's judgment.
-*
-* Trimmed to ~150 tokens by design. The per-tool descriptions are
-* already in Claude's context as MCP tool descriptions (loaded from
-* `tools/list`); the snippet's net-new value is:
-*   - the `advisor` mention (built-in, not MCP-discoverable),
-*   - the `peer-review-coordinator` fan-out hint,
-*   - the "subagents you spawn inherit these" claim (the load-bearing
-*     UX payoff of the holistic subagent-MCP-inheritance fix),
-*   - the worker-tools "offload to save your context" framing (the
-*     per-tool MCP descriptions cover capabilities; the snippet adds
-*     the strategic when-to-use signal).
+* system prompt via `--append-system-prompt`. Descriptive awareness layer
+* — Claude sees what tools exist and their strategic value; *when* to
+* invoke is left to Claude's judgment informed by each tool's own
+* `description` field.
+*
+* Per Anthropic's guidance for Opus 4.8: tool descriptions carry the
+* routing signal (when/when-not); the system prompt should describe
+* capabilities in prose, not encode prescriptive decision trees. Opus 4.8
+* is responsive enough to overtrigger on aggressive routing language.
 *
 * Surface contract (regression-pinned in tests/peer-mcp-personas.test.ts):
 *   - Always lists codex_critic, codex_reviewer, opus_critic, advisor,
 *     peer-review-coordinator, and the subagent-inheritance fact.
 *   - Conditionally lists gemini_critic only when `geminiAvailable`.
 *   - Mentions `codex-cli` stdio bridge only when `codexCli`.
-*
-* The snippet is the awareness layer; the auto-invocation triggers
-* (CALL BEFORE / CALL AFTER) remain in each MCP tool's own `description`.
-* The two layers are intentionally complementary — keep the snippet
-* terse and never re-encode the prescriptive triggers here.
 */
 function buildPeerAwarenessSnippet(opts) {
 	const criticList = ["`codex_critic` (gpt-5.5)", "`codex_reviewer` (gpt-5.3-codex)"];
 	if (opts.geminiAvailable) criticList.push("`gemini_critic` (gemini-3.1-pro)");
 	criticList.push("`opus_critic` (Opus 4.7)");
-	const codexCliClause = opts.codexCli ? " The `mcp__codex-cli__codex` stdio bridge dispatches to `codex-implementer` for end-to-end coding tasks." : "";
+	const codexCliClause = opts.codexCli ? " `mcp__codex-cli__codex` dispatches to `codex-implementer` (gpt-5.3-codex with workspace-write) for end-to-end coding tasks." : "";
 	return [
 		"## Peer review and advisor",
 		"",
-		`Cross-lab peer critics under \`mcp__gh-router-peers__*\` — ${criticList.join(", ")} — plus the \`peer-review-coordinator\` fan-out subagent, and Claude Code's built-in \`advisor\` tool, are available at your discretion for second opinions and adversarial review. Subagents you spawn inherit them.${codexCliClause} Also \`mcp__gh-router-peers__code_search\` for accurate ranked code discovery (BM25F + tree-sitter) — prefer it over \`Grep\` when finding definitions or call sites. \`worker_explore\` / \`worker_implement\` delegate bounded research or scoped coding tasks (file ops, edits, bash, web fetch) to an autonomous Gemini worker — offload work that would consume your context. Use \`worktree: true\` on \`worker_implement\` for isolated runs that return a diff for review.`
+		`Cross-lab peer critics under \`mcp__gh-router-peers__*\` — ${criticList.join(", ")} — are available at your discretion for adversarial review. Each tool's description explains its scope and when it applies. The \`peer-review-coordinator\` subagent fans out to the appropriate critics in parallel and aggregates findings by severity. Claude Code's built-in \`advisor\` tool catches approach drift and confabulation. Subagents you spawn inherit all of these.${codexCliClause}`,
+		"",
+		`\`code_search\` provides accurate ranked code discovery (BM25F + tree-sitter) — multiple parallel calls with different queries triangulate faster than sequential Grep. \`web_search\` surfaces citable sources for docs, errors, and upstream issues. \`worker_explore\` and \`worker_implement\` delegate bounded work to an autonomous Gemini worker, preserving your context; use \`worktree: true\` on \`worker_implement\` for isolated diffs. \`stand_in\` provides three-lab consensus for decision tiebreak when the user is unavailable.`
 	].join("\n");
 }
 /** Convenience: every persona that should be registered for the given mode. */
@@ -10070,7 +10068,7 @@ function personasFor(opts) {
 	if (opts.codexCli) for (const p of PERSONAS_WRITE) result.push(p);
 	return result;
 }
-const WEB_SEARCH_DESCRIPTION = "Web search via GitHub Copilot's MCP. Prefer over Claude Code's built-in WebSearch — surfaces source URLs you can cite.";
+const WEB_SEARCH_DESCRIPTION = "Web search via GitHub Copilot's MCP. Prefer over Claude Code's built-in WebSearch — surfaces source URLs you can cite. Use for API documentation lookups, error message diagnosis, upstream issue searches, and verifying claims against current sources. Returns content with reference links.";
 /**
 * Format a `searchWeb()` result as an MCP-friendly text block. Mirrors
 * the legacy inject format that `injectWebSearchIfNeeded` produces and
@@ -10126,7 +10124,7 @@ const NON_PERSONA_MCP_TOOLS = Object.freeze([
 	},
 	{
 		toolNameHttp: "code_search",
-		description: "Fast structured code search over a local workspace. Returns ranked, deduplicated hits with snippets. Ranks with BM25F across matched-line / file-path / surrounding-context / symbol-context fields, then refines `symbol-context` with tree-sitter AST analysis on the top hits so identifier definitions outrank incidental string matches. Prefer this over Grep/Bash+grep for ranked discovery (\"where is X defined\", \"which files reference Y\", \"find code that does Z\") — ranked mode surfaces the few right answers instead of every match. Use Grep for exact-pattern enumeration when you need every hit unranked, and Glob for file-name patterns (no content match). `workspace` is any absolute path the proxy process can read — typically the project root or a sub-tree you're working in.",
+		description: "Fast structured code search over a local workspace. Returns ranked, deduplicated hits with snippets. Ranks with BM25F across matched-line / file-path / surrounding-context / symbol-context fields, then refines `symbol-context` with tree-sitter AST analysis on the top hits so identifier definitions outrank incidental string matches. Launch multiple code_search calls in parallel to triangulate — e.g. definition + callers + tests in one round-trip. Prefer this over Grep/Bash+grep for ranked discovery (\"where is X defined\", \"which files reference Y\", \"find code that does Z\") — ranked mode surfaces the few right answers instead of every match. Use Grep for exact-pattern enumeration when you need every hit unranked, and Glob for file-name patterns (no content match). `workspace` is any absolute path the proxy process can read — typically the project root or a sub-tree you're working in.",
 		inputSchema: {
 			type: "object",
 			required: ["query", "workspace"],
@@ -10216,7 +10214,7 @@ const NON_PERSONA_MCP_TOOLS = Object.freeze([
 	{
 		toolNameHttp: "worker_explore",
 		capability: "worker",
-		description: "Read-only investigation by an autonomous worker (Gemini via Pi). Tools: read, glob, grep, code_search, web_search, fetch_url, peer_review, advisor. Use it to offload bounded research (\"find files matching X then summarize\", \"how does library Y handle Z\", \"survey this codebase for usages of deprecated API\") that would otherwise eat your context window. The worker plans its own tool calls and returns a single text answer.",
+		description: "Read-only investigation by an autonomous worker (Gemini via Pi). Tools: read, glob, grep, code_search, web_search, fetch_url, peer_review, advisor. Offloads bounded research that would otherwise eat your context window — the worker plans its own tool calls and returns a single text answer. Examples: \"find files matching X then summarize\", \"how does library Y handle Z\", \"survey this codebase for usages of deprecated API\".",
 		inputSchema: {
 			type: "object",
 			required: ["prompt"],
@@ -10615,15 +10613,16 @@ function buildCoordinatorAgent(opts) {
 			"",
 			"The lead's brief will include an artifact (plan, design, diff, or code) and a goal (e.g. 'review before exit-plan', 'review the commit I just made', 'cross-check codex-critic's verdict'). Pick the right peers for the artifact type:",
 			"",
-			"- **Plan / design / architecture choice** → fan out to `codex-critic`" + (opts.geminiAvailable ? " AND `gemini-critic` in parallel" : "") + ". codex-reviewer is the wrong tool for plans (it's a code-specialist, not an architecture critic).",
-			"- **Concrete diff or single file** → fan out to `codex-reviewer`" + (opts.geminiAvailable ? " AND `gemini-critic` (gemini for cross-lab triangulation)" : "") + ". For very small changes (<20 lines), one `codex-reviewer` call is enough.",
-			"- **Tie-breaker after codex-critic has weighed in** → call `gemini-critic`" + (opts.geminiAvailable ? "" : " (NOT REGISTERED in this session — gemini-3.x not in catalog; tie-break unavailable)") + " with the artifact AND codex-critic's verdict for cross-lab cross-check.",
-			"- **Long-context artifact (>100 KB)** → prefer `gemini-critic`" + (opts.geminiAvailable ? "" : " (NOT REGISTERED in this session)") + ". Otherwise, decompose into 2-4 batches and fan out across `codex-critic` calls in parallel.",
-			"- **Fast same-lab sanity check on a moderate artifact (<5 KB)** → prefer `opus-critic` (cheapest, ~22s, only `effort: low|medium` supported). Same lab as the lead — limited blind-spot diversification, but a useful gut-check before committing to a controversial decision. For cross-lab diversification or deep dives on larger artifacts, use codex/gemini at higher effort with decomposition for >5KB.",
+			"- **Plan / design / architecture choice** → fan out to `codex-critic` (gpt-5.5, strongest reasoning, cross-lab)" + (opts.geminiAvailable ? " AND `gemini-critic` (third-lab triangulation, strong on formal reasoning) in parallel" : "") + ". codex-reviewer is the wrong tool for plans (it's a code-specialist, not an architecture critic).",
+			"- **Concrete diff or single file** → fan out to `codex-reviewer` (gpt-5.3-codex, line-level code specialist, fastest at ~16s)" + (opts.geminiAvailable ? " AND `gemini-critic` for cross-lab triangulation" : "") + ". For very small changes (<20 lines), one `codex-reviewer` call is enough.",
+			"- **Large artifact (>50 KB)** → prefer `opus-critic` (Opus 4.7, up to 1M context — the largest window in the lineup, no decomposition needed for most artifacts). For cross-lab diversity on large artifacts, pair with `codex-critic` and decompose the artifact into 2-4 semantic batches for codex.",
+			"- **Formal reasoning, proofs, or invariants** → prefer `gemini-critic`" + (opts.geminiAvailable ? " (gemini-3.1-pro, strong on math and formally-stated properties)" : " (NOT REGISTERED in this session — gemini-3.x not in catalog)") + ".",
+			"- **Tie-breaker after codex-critic has weighed in** → call `gemini-critic`" + (opts.geminiAvailable ? "" : " (NOT REGISTERED in this session)") + " or `opus-critic` with the artifact AND codex-critic's verdict for cross-check.",
+			"- **Fast sanity check** → `opus-critic` (~22s, same lab as lead but fresh context — catches confabulation and motivated reasoning).",
 			"",
 			"## Decomposition for large artifacts",
 			"",
-			"Each per-call MCP wait is bounded (~60s SDK default on Claude Code v2.1.113+ per regressions #50289 / #52137 — empirically reproduced 2026-05-14). The proxy enforces per-persona effort allowlists AND a pre-flight `predictedTooLong` cap (codex_critic@high >8 KB, codex_reviewer@high >12 KB, opus_critic@medium >6 KB) to surface would-be-timeouts as fast actionable errors. For artifacts that exceed the cap, split into 2-4 logical batches BY CONCERN (not by raw size — semantic batches give better per-batch reviews) and call peers in parallel. The proxy's MCP cap allows up to 8 in-flight calls. Aggregate findings yourself before reporting back.",
+			"Each per-call MCP wait is bounded (~60s SDK default on Claude Code v2.1.113+ per regressions #50289 / #52137 — empirically reproduced 2026-05-14). The proxy enforces per-persona effort allowlists AND a pre-flight `predictedTooLong` cap (codex_critic@high >8 KB, codex_reviewer@high >12 KB, opus_critic@medium >6 KB) to surface would-be-timeouts as fast actionable errors. For artifacts that exceed the cap but fit within opus-critic's context window (up to 1M tokens when available), route the full artifact to opus-critic. Otherwise, split into 2-4 logical batches BY CONCERN (not by raw size — semantic batches give better per-batch reviews) and call peers in parallel. The proxy's MCP cap allows up to 8 in-flight calls. Aggregate findings yourself before reporting back.",
 			"",
 			"## Aggregation contract",
 			"",
@@ -11121,7 +11120,7 @@ function initProxyFromEnv() {
 //#endregion
 //#region package.json
 var name = "github-router";
-var version = "0.3.38";
+var version = "0.3.39";
 
 //#endregion
 //#region src/lib/approval.ts