github-router 0.3.175 → 0.3.177

package/dist/main.js

@@ -1,8 +1,8 @@
 #!/usr/bin/env node
-import { $ as isAdvisorRequested, $t as cacheModels, A as liveExec, At as hasSupportedBrowserInstalled, Bt as DEFAULT_CODEX_MODEL_FALLBACKS, C as fileReviewDebounce, Ct as pickEndpoint, D as stopGateEnabledForRepo, Dt as readResponseBodyCapped, E as repoRoot, Et as MAX_RESPONSE_BODY_BYTES, Ft as ArtifactClient, G as vscodeRipgrepPath, Gt as pickClaudeDefault, H as buildToolbeltAwareness, Ht as UPSTREAM_FETCH_TIMEOUT_MS, It as collapsePathKeys, J as searchWeb, Jt as setupCopilotToken, K as TOOLBELT_TOOLS, Kt as getPackageVersion, Lt as toolbeltPathOverride, Mt as extractTarGzMember, Nt as extractZipMember, O as stopReviewStateDir, Ot as parseJsonOrDiagnose, Pt as shouldUseInsecureTls, Q as injectAdvisorTool, Qt as cacheCopilotVersion, Rt as DEFAULT_CLAUDE_MODEL_FALLBACKS, S as fileLastPromptStore, St as resolveMcpToolTimeoutMs, T as repoFingerprint, Tt as createChatCompletions, U as toolbeltEnabled, Ut as UPSTREAM_INACTIVITY_TIMEOUT_MS, V as availableToolCommands, Vt as DEFAULT_PORT, W as toolbeltSkipSet, Wt as generateRandomPort, X as ADVISOR_TOOL_INSTRUCTIONS, Xt as setupGitHubToken, Y as ADVISOR_INTERNAL_TOOL_NAME, Yt as setupGitHubAgentToken, Z as buildAdvisorStream, Zt as tryRefreshAndRetry, _ as stopGateId, _t as workerToolsEnabled, a as buildPeerAwarenessSnippet, an as sleep, at as relayAnthropicStream, b as fileBaselineStore, bt as getTokenCount, c as buildArtifactOpenHookCommand, cn as HTTPError, ct as agentToolsEnabled, d as captureLaunchBaseline, dn as copilotBaseUrl, dt as browserCompoundToolsEnabled, en as cacheVSCodeVersion, et as buildAnthropicErrorEvent, f as decideStopHook, fn as copilotHeaders, ft as browserToolsEnabled, g as stopGateDisabled, gt as standInToolEnabled, h as launchBaselineKey, ht as implementerSubagentModel, i as buildAgentPrompt, in as resolveModel, it as readIteratorWithTimeout, j as resolveSealedGate, jt as provisionAndIndexColbert, k as trustRepo, kt as provisionBrowserAssets, l as buildSessionBindHookCommand, ln as forwardError, lt as artifactToolsEnabled, m as injectStopHookIntoSettingsFile, mn as state, mt as geminiAvailable, n as MCP_GROUPS, nn as isNullish, nt as isControllerClosedError, o as buildPeerAwarenessSummary, on as getModels, ot as handleMcpDelete, p as fileBlockBudget, pn as githubHeaders, pt as fleetToolsEnabled, q as assetFor, qt as withInstallLock, r as assertMcpToolSurfaceConsistent, rn as resolveCodexModel, rt as logStreamError, s as personasFor, sn as fetchWithTransientRetry, st as handleMcpPost, t as GROUP_META, tn as filterBetaHeader, tt as buildOpenAIErrorEvent, u as buildStopHookCommand, un as GITHUB_API_BASE_URL, ut as browseAgentEnabled, v as stopGatePlanMode, vt as countTokens, w as isSubagentContext, wt as createResponses, x as fileFindingsStore, xt as assembleResponsesPayload, y as stopReviewEnabled, yt as createMessages, zt as DEFAULT_CODEX_MODEL } from "./peer-mcp-personas-BCGYWok0.js";
-import { a as removeOwnClaudeConfigMirror, i as isUnderClaudeConfigMirror, l as writeArtifactCredsToMirror, n as ensureClaudeConfigMirror, r as ensurePaths, t as PATHS, u as writeRuntimeFileSecure } from "./paths-CTlT1nTo.js";
-import { c as killManagedTree, d as runCommandCapture, f as runCommandVoid, l as parseBoolEnv, s as killChildProcessTree, u as resolveExecutable } from "./lifecycle-ChPBRt6K.js";
-import { a as sweepRegistry } from "./lifecycle-BUXxiltc.js";
+import { $ as isAdvisorRequested, $t as cacheCopilotVersion, A as liveExec, At as provisionBrowserAssets, Bt as DEFAULT_CODEX_MODEL, C as fileReviewDebounce, Ct as resolveMcpToolTimeoutMs, D as stopGateEnabledForRepo, Dt as MAX_RESPONSE_BODY_BYTES, E as repoRoot, Et as createChatCompletions, Ft as shouldUseInsecureTls, G as vscodeRipgrepPath, Gt as generateRandomPort, H as buildToolbeltAwareness, Ht as DEFAULT_PORT, It as ArtifactClient, J as searchWeb, Jt as withInstallLock, K as TOOLBELT_TOOLS, Kt as pickClaudeDefault, Lt as collapsePathKeys, Mt as provisionAndIndexColbert, Nt as extractTarGzMember, O as stopReviewStateDir, Ot as readResponseBodyCapped, Pt as extractZipMember, Q as injectAdvisorTool, Qt as tryRefreshAndRetry, Rt as toolbeltPathOverride, S as fileLastPromptStore, St as assembleResponsesPayload, T as repoFingerprint, Tt as createResponses, U as toolbeltEnabled, Ut as UPSTREAM_FETCH_TIMEOUT_MS, V as availableToolCommands, Vt as DEFAULT_CODEX_MODEL_FALLBACKS, W as toolbeltSkipSet, Wt as UPSTREAM_INACTIVITY_TIMEOUT_MS, X as ADVISOR_TOOL_INSTRUCTIONS, Xt as setupGitHubAgentToken, Y as ADVISOR_INTERNAL_TOOL_NAME, Yt as setupCopilotToken, Z as buildAdvisorStream, Zt as setupGitHubToken, _ as stopGateId, _t as workerToolsEnabled, a as buildPeerAwarenessSnippet, an as resolveModel, at as relayAnthropicStream, b as fileBaselineStore, bt as createMessages, c as buildArtifactOpenHookCommand, cn as fetchWithTransientRetry, ct as agentToolsEnabled, d as captureLaunchBaseline, dn as GITHUB_API_BASE_URL, dt as browserCompoundToolsEnabled, en as cacheModels, et as buildAnthropicErrorEvent, f as decideStopHook, fn as copilotBaseUrl, ft as browserToolsEnabled, g as stopGateDisabled, gt as standInToolEnabled, h as launchBaselineKey, hn as state, ht as nativeSubagentModel, i as buildAgentPrompt, in as resolveCodexModel, it as readIteratorWithTimeout, j as resolveSealedGate, jt as hasSupportedBrowserInstalled, k as trustRepo, kt as parseJsonOrDiagnose, l as buildSessionBindHookCommand, ln as HTTPError, lt as artifactToolsEnabled, m as injectStopHookIntoSettingsFile, mn as githubHeaders, mt as geminiAvailable, n as MCP_GROUPS, nn as filterBetaHeader, nt as isControllerClosedError, o as buildPeerAwarenessSummary, on as sleep, ot as handleMcpDelete, p as fileBlockBudget, pn as copilotHeaders, pt as fleetToolsEnabled, q as assetFor, qt as getPackageVersion, r as assertMcpToolSurfaceConsistent, rn as isNullish, rt as logStreamError, s as personasFor, sn as getModels, st as handleMcpPost, t as GROUP_META, tn as cacheVSCodeVersion, tt as buildOpenAIErrorEvent, u as buildStopHookCommand, un as forwardError, ut as browseAgentEnabled, v as stopGatePlanMode, vt as shimDefaultsToXhigh, w as isSubagentContext, wt as pickEndpoint, x as fileFindingsStore, xt as getTokenCount, y as stopReviewEnabled, yt as countTokens, zt as DEFAULT_CLAUDE_MODEL_FALLBACKS } from "./peer-mcp-personas-BtQi-dF-.js";
+import { a as isUnderClaudeConfigMirror, d as writeRuntimeFileSecure, i as ensurePaths, o as removeOwnClaudeConfigMirror, r as ensureClaudeConfigMirror, t as PATHS, u as writeArtifactCredsToMirror } from "./paths-BZPgxKUl.js";
+import { c as killManagedTree, d as runCommandCapture, f as runCommandVoid, l as parseBoolEnv, s as killChildProcessTree, u as resolveExecutable } from "./lifecycle-DrDEjx_c.js";
+import { a as sweepRegistry } from "./lifecycle-Dvr3pGGG.js";
 import { defineCommand, runMain } from "citty";
 import consola from "consola";
 import { createHash, randomBytes, randomUUID } from "node:crypto";
@@ -1137,16 +1137,16 @@ 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` (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).",
+			"- **Plan / design / architecture choice** → fan out to `codex-critic` (gpt-5.6-sol, 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-reviewer` (gemini-3.1-pro, second-lab line-level review)" : "") + (opts.geminiAvailable ? " AND `gemini-critic` for cross-lab triangulation" : "") + ". For very small changes (<20 lines), one `codex-reviewer` call is enough.",
-			"- **Large artifact** → the only peers that take a large artifact WHOLE are `codex-critic` (gpt-5.5, ≈922K-token input window) and `opus-critic` (Opus-4.7-1M, ≈936K-token input on enterprise catalogs; ≈168K otherwise). Route the full artifact to those for cross-lab coverage. `codex-reviewer` (≈272K) and `gemini-critic` (≈136K) have small windows — see Decomposition below: never summarize or downsize the request to squeeze a large artifact into a small-window peer.",
+			"- **Large artifact** → the only peers that take a large artifact WHOLE are `codex-critic` (gpt-5.6-sol, ≈1M-token input window) and `opus-critic` (Opus-4.7-1M, ≈936K-token input on enterprise catalogs; ≈168K otherwise). Route the full artifact to those for cross-lab coverage. `codex-reviewer` (≈272K) and `gemini-critic` (≈136K) have small windows — see Decomposition below: never summarize or downsize the request to squeeze a large artifact into a small-window peer.",
 			"- **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",
 			"",
-			"Route by the peer's real PROMPT WINDOW (input tokens): `codex-critic` gpt-5.5 ≈922K · `opus-critic` Opus-4.7-1M ≈936K (enterprise catalogs; ≈168K otherwise) · `codex-reviewer` gpt-5.3-codex ≈272K · `gemini-critic` gemini-3.1-pro ≈136K. The proxy REJECTS (with an actionable message) any single call whose brief exceeds the target peer's window — it will NOT silently truncate, because dropping lines from a review artifact is worse than a clear error. So: send the full artifact only to peers whose window fits it (large artifacts → `codex-critic` and/or `opus-critic`). When a peer's window is too small (commonly `gemini-critic` at ≈136K, or `codex-reviewer` at ≈272K), do NOT summarize or downsize the request to include it — either skip that peer, or split the artifact into 2-4 logical batches BY CONCERN (not by raw size — semantic batches give better per-batch reviews) that each fit, and call in parallel. Use the big-window peers for the whole and reserve a small-window peer like gemini for the concerns it can actually hold. The proxy's MCP cap allows up to 8 in-flight calls. Aggregate findings yourself before reporting back. (Separately, on the JSON transport a per-effort `predictedTooLong` byte cap still guards the ~60s tools/call timeout for non-SSE clients; Claude Code uses SSE, which streams with heartbeats and isn't subject to that cap.)",
+			"Route by the peer's real PROMPT WINDOW (input tokens): `codex-critic` gpt-5.6-sol ≈1M · `opus-critic` Opus-4.7-1M ≈936K (enterprise catalogs; ≈168K otherwise) · `codex-reviewer` gpt-5.3-codex ≈272K · `gemini-critic` gemini-3.1-pro ≈136K. The proxy REJECTS (with an actionable message) any single call whose brief exceeds the target peer's window — it will NOT silently truncate, because dropping lines from a review artifact is worse than a clear error. So: send the full artifact only to peers whose window fits it (large artifacts → `codex-critic` and/or `opus-critic`). When a peer's window is too small (commonly `gemini-critic` at ≈136K, or `codex-reviewer` at ≈272K), do NOT summarize or downsize the request to include it — either skip that peer, or split the artifact into 2-4 logical batches BY CONCERN (not by raw size — semantic batches give better per-batch reviews) that each fit, and call in parallel. Use the big-window peers for the whole and reserve a small-window peer like gemini for the concerns it can actually hold. The proxy's MCP cap allows up to 8 in-flight calls. Aggregate findings yourself before reporting back. (Separately, on the JSON transport a per-effort `predictedTooLong` byte cap still guards the ~60s tools/call timeout for non-SSE clients; Claude Code uses SSE, which streams with heartbeats and isn't subject to that cap.)",
 			"",
 			"## Aggregation contract",
 			"",
@@ -1205,10 +1205,22 @@ function buildPeerAgentDefinitions(opts) {
 		codexCli: opts.codexCli,
 		geminiAvailable: opts.geminiAvailable
 	});
-	if (opts.implementerModel && opts.implementerModel.length > 0) out.implementer = {
-		description: "Bounded implementation subagent running gpt-5.5 (strong non-Claude coder, high reasoning). Use for well-scoped coding tasks — edits, small features, fixes — you want implemented in an integrated subagent. Model is overridable at spawn.",
+	const nativeModel = opts.nativeSubagentModel && opts.nativeSubagentModel.length > 0 ? opts.nativeSubagentModel : void 0;
+	const modelField = nativeModel ? { model: nativeModel } : {};
+	out.implementer = {
+		description: nativeModel ? `Bounded implementation subagent running ${nativeModel} (strong non-Claude coder, maximum reasoning). Use proactively for well-scoped coding tasks — edits, small features, fixes — to keep the lead's context focused; runs in its own context. Model is overridable at spawn.` : `Bounded implementation subagent (native tools, runs on the lead's model in its own context). Use proactively for well-scoped coding tasks — edits, small features, fixes — to keep the lead's context focused. Model is overridable at spawn.`,
 		prompt: "You are a bounded implementation subagent for well-scoped coding tasks. Implement the requested change surgically, matching the surrounding code style and minimizing unrelated churn. Use the dedicated Edit/Write/Read tools for file changes and Grep/Glob for search; reserve Bash for running builds, tests, and git — do not shell out (sed/awk/python/here-docs) to read or edit files. Verify with the project's build or tests where applicable. Do the work yourself — do not spawn further subagents. Report exactly what changed and any risks.",
-		model: opts.implementerModel
+		...modelField
+	};
+	out.debugger = {
+		description: nativeModel ? `Root-cause & debugging subagent running ${nativeModel} at maximum reasoning. Use proactively for reproducing a bug end to end, isolating the true root cause, and proposing a minimal fix — investigation-heavy work best kept off the lead's context; runs in its own context. Model is overridable at spawn.` : `Root-cause & debugging subagent (native tools, runs on the lead's model in its own context). Use proactively for reproducing a bug end to end, isolating the true root cause, and proposing a minimal fix — kept off the lead's context. Model is overridable at spawn.`,
+		prompt: "You are a root-cause debugging subagent. Reproduce the failure end to end first — as close to how a real user hits it as you can — before theorizing. Form hypotheses and test them against the actual code and runtime: read the code, run the repro with Bash, and add temporary instrumentation only if needed (remove it after). Identify the true root cause, not a symptom; if a fix is in scope, make it minimal and verify the repro now passes. Use the dedicated Edit/Write/Read tools for file changes and Grep/Glob for search; reserve Bash for running repros, tests, and git — do not shell out (sed/awk/python/here-docs) to read or edit files. Do the work yourself — do not spawn further subagents. Report the root cause with evidence, the fix (if any), and any residual risks.",
+		...modelField
+	};
+	out["qa-engineer"] = {
+		description: nativeModel ? `Review, testing & QA subagent running ${nativeModel} at maximum reasoning. Use proactively to review a change for correctness, author and run tests that try to break it, and give a severity-ranked go/no-go while keeping the lead's context focused; runs in its own context. Model is overridable at spawn.` : `Review, testing & QA subagent (native tools, runs on the lead's model in its own context). Use proactively to review a change for correctness, author and run tests that try to break it, and give a severity-ranked go/no-go. Model is overridable at spawn.`,
+		prompt: "You are a review, testing, and QA subagent. Verify correctness against the ACTUAL code by reading it — never assume. Where the change warrants it, author tests that try to BREAK the implementation (edge cases, error paths, and the acceptance criteria as executable checks), then run them and report which pass and which fail; do NOT modify production code just to make tests pass. Use the dedicated Edit/Write/Read tools for file changes and Grep/Glob for search; reserve Bash for running builds, tests, and git — do not shell out (sed/awk/python/here-docs) to read or edit files. Do the work yourself — do not spawn further subagents. Report severity-ranked findings with `file:line` citations and end with a clear go/no-go.",
+		...modelField
 	};
 	if (opts.workerToolsAvailable) {
 		const workersKey = workersKeyOf(opts.groupKeys);
@@ -1506,7 +1518,7 @@ async function writePeerMcpRuntimeFiles(serverUrl, opts) {
 		groupKeys: opts.groupKeys,
 		workerToolsAvailable: opts.workerToolsAvailable,
 		browseAvailable: opts.browseAvailable,
-		implementerModel: opts.implementerModel,
+		nativeSubagentModel: opts.nativeSubagentModel,
 		nonce,
 		codexHome
 	});
@@ -1787,7 +1799,7 @@ async function callMcpTool(opts) {
 	};
 }
 /**
-* One non-streaming gpt-5.5 (or any model id) inference via `/v1/responses`.
+* One non-streaming gpt-5.6-sol (or any model id) inference via `/v1/responses`.
 * Returns the assistant text (possibly empty). Throws on transport/HTTP/parse
 * failure. `effort` maps to the Responses `reasoning.effort` knob.
 */
@@ -2556,7 +2568,7 @@ async function discoverGateCommands(cwd, opts) {
 	if (files.length === 0) return null;
 	let result;
 	try {
-		const { runWorkerAgent } = await import("./engine-cUomEg92.js");
+		const { runWorkerAgent } = await import("./engine-DOOYh35O.js");
 		result = await runWorkerAgent({
 			mode: "explore",
 			workspace: root,
@@ -2622,14 +2634,14 @@ function decidePromptSubmit(input) {
 * concluding. Mirrors the v1 advisory tone: additive, never blocking.
 */
 const PROMPT_SEARCH_TIP = "TIP (advisory): when this task needs code context, search lexical + semantic in parallel — one `mcp__search__code` call with mode:\"lexical\" and one with mode:\"semantic\", issued in the same turn — before concluding.";
-/** System prompt for the single gpt-5.5 scope/goal inference. Steers a SHORT,
+/** System prompt for the single gpt-5.6-sol scope/goal inference. Steers a SHORT,
 *  user-derived (not invented) advisory note grounded in the search results. */
 const PROMPT_SCOPE_SYSTEM = "You are a scoping assistant for a coding agent about to act on a user's request. You are given the user's request and the results of a lexical + semantic code search over the relevant repository. Produce a SHORT advisory note (<= 120 words), plain text only:\n1. SCOPE: one line — is this trivial, focused (one area), or large/cross-cutting — grounded in what the search surfaced (reference the most relevant file(s) by name).\n2. GOAL: restate the user's OWN ask as a single measurable objective, in THEIR terms. Do NOT invent new requirements or acceptance criteria beyond what they asked.\n3. Only if the task is large/cross-cutting, add a final line: \"Consider /gh-research first to saturate understanding, then /gh-orchestrate to compose a floor-raising workflow.\" Omit it for a focused or trivial task.\nThis is advisory — the agent decides whether to follow it. Be concrete and concise; no preamble.";
 /** Max chars of each search-result blob fed into the scope inference. */
 const SEARCH_CONTEXT_CAP = 6 * 1024;
 /** Wrap the prior-turn review findings in an explicitly NON-AUTHORITATIVE frame. */
 function framePendingFindings(findings) {
-	return "ADVISORY — independent review of your PREVIOUS change (NON-AUTHORITATIVE): an independent gpt-5.5 reviewer flagged the following. Evaluate each on its merits — fix the real ones, and ignore any wrong one with a one-line reason. You are NOT obligated to act on these.\n" + findings.trim();
+	return "ADVISORY — independent review of your PREVIOUS change (NON-AUTHORITATIVE): an independent gpt-5.6-sol reviewer flagged the following. Evaluate each on its merits — fix the real ones, and ignore any wrong one with a one-line reason. You are NOT obligated to act on these.\n" + findings.trim();
 }
 function joinSections(sections) {
 	return sections.map((s) => s.trim()).filter((s) => s.length > 0).join("\n\n");
@@ -2643,7 +2655,7 @@ function joinSections(sections) {
 *   - subagent/teammate  -> empty (top-level only, like v1).
 *   - findings           -> always surfaced (+ cleared) regardless of triviality.
 *   - trivial prompt     -> findings only (no search tip, no model call).
-*   - substantive prompt -> static search tip + parallel lexical+semantic search -> ONE gpt-5.5 call
+*   - substantive prompt -> static search tip + parallel lexical+semantic search -> ONE gpt-5.6-sol call
 *                           -> grounded scope/goal note. Fail-open to PROMPT_STEER_GOAL.
 *   - steerEnabled=false -> findings only (no goal/tip).
 */
@@ -4716,7 +4728,7 @@ function initProxyFromEnv() {
 //#endregion
 //#region package.json
 var name = "github-router";
-var version$1 = "0.3.175";
+var version$1 = "0.3.177";
 
 //#endregion
 //#region src/lib/approval.ts
@@ -5799,13 +5811,21 @@ function parseDisableParallelToolUse(toolChoice) {
 	if (!toolChoice || typeof toolChoice !== "object") return void 0;
 	return toolChoice.disable_parallel_tool_use === true ? false : void 0;
 }
-/** Default absent Anthropic `thinking` to high effort, clamped by the model.
-*  Returns undefined for a model that advertises NO `reasoning_effort` allowlist
-*  — such a model may not support reasoning at all, so forcing an effort could
-*  400; leaving it unset preserves the pre-default safe behavior for that case. */
+/** Default absent Anthropic `thinking` to an effort, clamped by the model.
+*  OpenAI-frontier shim models (gpt-5.6-sol/gpt-5.5) default to xhigh so the
+*  native gpt-5.6-sol subagents (implementer/debugger/qa-engineer) and any
+*  `-m gpt-5.6-sol` main-loop session reason at max when the client sends no
+*  `thinking` block; every other shim model stays high. The
+*  `supported.includes("xhigh")` guard makes the intent explicit rather than
+*  trusting clampEffort to silently degrade. Opt out with
+*  `GH_ROUTER_DISABLE_FRONTIER_XHIGH_DEFAULT=1`. Returns undefined for a model
+*  that advertises NO `reasoning_effort` allowlist (may not support reasoning;
+*  forcing an effort could 400). Note: an explicit client `thinking` budget is
+*  handled by parseReasoningEffort and is NOT affected by this default. */
 function defaultReasoningEffort(model) {
 	const supported = model?.capabilities?.supports?.reasoning_effort;
-	return Array.isArray(supported) && supported.length > 0 ? clampEffort("high", supported) : void 0;
+	if (!(Array.isArray(supported) && supported.length > 0)) return void 0;
+	return clampEffort(parseBoolEnv(process.env.GH_ROUTER_DISABLE_FRONTIER_XHIGH_DEFAULT) !== true && model?.id != null && shimDefaultsToXhigh(model.id) && supported.includes("xhigh") ? "xhigh" : "high", supported);
 }
 /**
 * Map Anthropic `thinking` to a Responses reasoning effort, clamped to the
@@ -8062,6 +8082,10 @@ function parseSharedArgs(args) {
 * 1M/400k window, never overflows). See `seedGatewayModelCache`.
 */
 const NATIVE_NON_CLAUDE_MODELS = [
+	{
+		id: "gpt-5.6-sol",
+		displayName: "GPT-5.6 Sol"
+	},
 	{
 		id: "gpt-5.5",
 		displayName: "GPT-5.5"
@@ -8495,7 +8519,7 @@ const claude = defineCommand({
 				groupKeys,
 				workerToolsAvailable: workerToolsEnabled(),
 				browseAvailable: browseAgentEnabled(),
-				implementerModel: implementerSubagentModel()
+				nativeSubagentModel: nativeSubagentModel()
 			});
 			state.peerMcpNonce = runtime.nonce;
 			envVars.GH_ROUTER_HOOK_MCP_URL = serverUrl;
@@ -8692,7 +8716,6 @@ const claude = defineCommand({
 				powerBrowseAvailable: state.powerBrowseEnabled,
 				fleetAvailable: fleetToolsEnabled(),
 				agentToolsAvailable: agentToolsEnabled(),
-				implementerAvailable: implementerSubagentModel() != null,
 				groupKeys
 			});
 			peerAwarenessSnippet = peerSnippet;
@@ -8784,13 +8807,10 @@ const codex = defineCommand({
 		if (codexModel !== requestedModel) consola.info(`Model "${requestedModel}" resolved to "${codexModel}"`);
 		if (usingDefault && state.models) {
 			const inCache = (id) => state.models?.data.some((m) => m.id === id) ?? false;
-			if (!inCache(codexModel)) for (const fallback of DEFAULT_CODEX_MODEL_FALLBACKS) {
-				const resolved = resolveCodexModel(fallback);
-				if (inCache(resolved)) {
-					consola.info(`Default model "${codexModel}" not in your Copilot model list; falling back to "${resolved}".`);
-					codexModel = resolved;
-					break;
-				}
+			const firstPresent = [DEFAULT_CODEX_MODEL, ...DEFAULT_CODEX_MODEL_FALLBACKS].map((id) => resolveModel(id)).find((id) => inCache(id));
+			if (firstPresent) {
+				if (firstPresent !== codexModel) consola.info(`Default model "${DEFAULT_CODEX_MODEL}" not in your Copilot model list; falling back to "${firstPresent}".`);
+				codexModel = firstPresent;
 			}
 		}
 		const modelEntry = state.models?.data.find((m) => m.id === codexModel);
@@ -8955,7 +8975,7 @@ const internalPromptSubmit = defineCommand({
 						},
 						infer: (system, user, signal) => callInference({
 							serverUrl: runtime.serverUrl,
-							model: "gpt-5.5",
+							model: "gpt-5.6-sol",
 							instructions: system,
 							input: user,
 							effort: "low",
@@ -9688,7 +9708,7 @@ THE DIFF:
 const internalStopReview = defineCommand({
 	meta: {
 		name: "internal-stop-review",
-		description: "Internal: the detached, advisory background reviewer. Reads a JSON payload on stdin, runs a read-only gpt-5.5 review of the working tree against the user's ask, and writes advisory findings for the next prompt to surface. Never blocks anything."
+		description: "Internal: the detached, advisory background reviewer. Reads a JSON payload on stdin, runs a read-only gpt-5.6-sol review of the working tree against the user's ask, and writes advisory findings for the next prompt to surface. Never blocks anything."
 	},
 	async run() {
 		try {
@@ -9717,7 +9737,7 @@ const internalStopReview = defineCommand({
 						transcriptPath: typeof payload.transcript_path === "string" ? payload.transcript_path : ""
 					}),
 					workspace: cwd,
-					model: "gpt-5.5",
+					model: "gpt-5.6-sol",
 					thinking: "high"
 				},
 				timeoutMs: REVIEW_TIMEOUT_MS