github-router 0.3.257 → 0.3.258

package/dist/{server-setup-CA0ZT-nq.js → server-setup-CJzxzYks.js}

@@ -1,4 +1,4 @@
-import { A as searchWeb, B as readIteratorWithTimeout, Et as toolbeltPathOverride, F as isAdvisorRequested, Ft as withInstallLock, H as handleMcpDelete, I as buildAnthropicErrorEvent, L as buildOpenAIErrorEvent, M as ADVISOR_TOOL_INSTRUCTIONS, Mt as UPSTREAM_INACTIVITY_TIMEOUT_MS, N as buildAdvisorStream, Nt as generateRandomPort, P as injectAdvisorTool, R as isControllerClosedError, U as handleMcpPost, V as relayAnthropicStream, W as agentToolsEnabled, at as countTokens, ct as assembleResponsesPayload, dt as createResponses, ft as createChatCompletions, ht as parseJsonOrDiagnose, it as shimDefaultsToXhigh, j as ADVISOR_INTERNAL_TOOL_NAME, jt as UPSTREAM_FETCH_TIMEOUT_MS, lt as resolveMcpToolTimeoutMs, mt as readResponseBodyCapped, ot as createMessages, pt as MAX_RESPONSE_BODY_BYTES, r as assertMcpToolSurfaceConsistent, st as getTokenCount, ut as pickEndpoint, w as toolbeltEnabled, z as logStreamError } from "./peer-mcp-personas-ByulAXNm.js";
+import { A as searchWeb, B as readIteratorWithTimeout, Et as toolbeltPathOverride, F as isAdvisorRequested, Ft as withOneMSuffix, H as handleMcpDelete, I as buildAnthropicErrorEvent, It as withInstallLock, L as buildOpenAIErrorEvent, M as ADVISOR_TOOL_INSTRUCTIONS, Mt as UPSTREAM_INACTIVITY_TIMEOUT_MS, N as buildAdvisorStream, Nt as generateRandomPort, P as injectAdvisorTool, R as isControllerClosedError, U as handleMcpPost, V as relayAnthropicStream, W as agentToolsEnabled, at as countTokens, ct as assembleResponsesPayload, dt as createResponses, ft as createChatCompletions, ht as parseJsonOrDiagnose, it as shimDefaultsToXhigh, j as ADVISOR_INTERNAL_TOOL_NAME, jt as UPSTREAM_FETCH_TIMEOUT_MS, lt as resolveMcpToolTimeoutMs, mt as readResponseBodyCapped, ot as createMessages, pt as MAX_RESPONSE_BODY_BYTES, r as assertMcpToolSurfaceConsistent, st as getTokenCount, ut as pickEndpoint, w as toolbeltEnabled, z as logStreamError } from "./peer-mcp-personas-DYT9RZBk.js";
 import { t as getPackageVersion } from "./version-_Q1WpsQp.js";
 import { i as ensurePaths, t as PATHS } from "./paths-C1cLZ7Ga.js";
 import { C as copilotHeaders, O as state, S as copilotBaseUrl, a as cacheCopilotVersion, c as filterBetaHeader, d as resolveModel, f as sleep, g as fetchWithTransientRetry, i as tryRefreshAndRetry, l as isNullish, n as setupGitHubAgentToken, o as cacheModels, r as setupGitHubToken, s as cacheVSCodeVersion, t as setupCopilotToken, v as HTTPError, y as forwardError } from "./token-R2ocFC3M.js";
@@ -882,7 +882,7 @@ function initProxyFromEnv() {
 //#endregion
 //#region package.json
 var name = "github-router";
-var version = "0.3.257";
+var version = "0.3.258";
 //#endregion
 //#region src/lib/approval.ts
 const awaitApproval = async () => {
@@ -1706,16 +1706,54 @@ function anthropicSseStreamFromEvents(events, opts) {
 * handler → shim → handler import cycle). `handler.ts` re-exports these for
 * backward compatibility with existing imports/tests.
 */
+/**
+* Copilot's reasoning-effort tiers, lowest to highest.
+*
+* Both ends were added after the fact and both are load-bearing:
+*
+* `none` is advertised by every gpt-5.x entry in the live catalog. While it was
+* missing here it was treated as an UNRECOGNIZED value, so a client asking for
+* the MINIMUM on a model that does not offer it (gemini advertises only
+* low/medium/high) was anchored at the unknown-value tier and clamped to
+* `high` — the maximum. Listing it makes the clamp resolve to `low` instead,
+* which is what "nearest supported tier" should always have meant.
+*
+* `max` is advertised by `gpt-5.6-sol`, `gpt-5.6-terra`, `gpt-5.6-luna`,
+* `claude-opus-5` and the 4.7/4.8 Opus lines, and Claude Code's effort picker
+* offers it for any model whose entry allows it. Listing it lets an explicit
+* selection pass through, and lets `clampEffort` land on it for a model that
+* advertises nothing lower.
+*
+* `bucketEffort` deliberately reaches neither end — see below.
+*/
 const EFFORT_ORDER = [
+	"none",
 	"low",
 	"medium",
 	"high",
-	"xhigh"
+	"xhigh",
+	"max"
 ];
+/** Anchor for an effort value that is not a recognized tier at all.
+*
+*  Deliberately NOT the top of `EFFORT_ORDER`. Callers reach this only when the
+*  incoming value is unrecognized, which is a guess — and resolving a guess to
+*  the most expensive tier a model advertises would silently spend more than the
+*  caller could have meant. Anchoring here and clamping DOWN keeps the behavior
+*  identical to before `max` joined the ladder, while `max` stays reachable by
+*  explicit, valid selection. */
+const UNKNOWN_EFFORT_ANCHOR = "xhigh";
 /**
 * Bucket a thinking budget into a Copilot reasoning-effort string.
 * `<2000`→low, `<8000`→medium, `<24000`→high, else→xhigh.
 * Defaults missing/non-numeric budgets to 8000 ("high").
+*
+* The ceiling stays at `xhigh` even though `max` exists: Anthropic's
+* `budget_tokens` is unbounded above, so any threshold chosen for a `max`
+* bucket would silently re-tier existing callers whose budgets already map to
+* `xhigh`. `max` is reachable only by explicit selection
+* (`output_config.effort`), which is an unambiguous request rather than an
+* inference from a token count.
 */
 function bucketEffort(budget) {
 	const n = typeof budget === "number" && Number.isFinite(budget) ? budget : 8e3;
@@ -2049,6 +2087,45 @@ function parseReasoningEffort(thinking, model) {
 	return Array.isArray(supported) && supported.length > 0 ? clampEffort(bucketed, supported) : bucketed;
 }
 /**
+* Client-selected reasoning effort in Copilot shape (`output_config.effort`) —
+* what Claude Code's effort picker puts on the wire.
+*
+* This is the highest-precedence signal, matching the rule the passthrough path
+* already documents ("client-supplied `output_config.effort` always wins",
+* `routes/messages/handler.ts`). Before it was read here the field survived
+* every preprocessing step and was then simply dropped by the shim, which made
+* the picker a no-op on non-Claude models in three distinct ways: alone it fell
+* through to the `high` default, alongside a `thinking` budget it lost to the
+* bucketed budget, and alongside `thinking:{type:"adaptive"}` it produced NO
+* reasoning field at all (the adaptive shape is neither `undefined` nor
+* `enabled`, so both of the other branches decline it).
+*
+* `clampOutputConfigEffortInPlace` has normally already clamped this against the
+* model's allowlist upstream, but the clamp is repeated here so the shim is
+* correct when called directly, and so an unrecognized level (a future Copilot
+* tier) resolves the same way it does on the passthrough path: anchored at
+* `UNKNOWN_EFFORT_ANCHOR` and clamped DOWN to the highest the model actually
+* advertises, never up.
+*
+* When the model advertises NO allowlist there is nothing to validate against,
+* so only a recognized tier is honored and anything else declines to the
+* `thinking`/default chain. Passing an arbitrary string through would let a
+* typo become a synthesized `reasoning.effort` on a payload we build, which is
+* the 400 that `defaultReasoningEffort` already refuses to risk by returning
+* undefined in exactly this case.
+*/
+function parseOutputConfigEffort(body, model) {
+	const oc = body.output_config;
+	if (!oc || typeof oc !== "object") return void 0;
+	const effort = oc.effort;
+	if (typeof effort !== "string" || effort === "") return void 0;
+	const known = EFFORT_ORDER.includes(effort);
+	const supported = model?.capabilities?.supports?.reasoning_effort;
+	if (!(Array.isArray(supported) && supported.length > 0)) return known ? effort : void 0;
+	if (supported.includes(effort)) return effort;
+	return clampEffort(known ? effort : UNKNOWN_EFFORT_ANCHOR, supported);
+}
+/**
 * Parse an already-JSON-parsed Anthropic Messages body into the neutral shape.
 * `resolvedModel` is the catalog id the request will run on; `model` its
 * catalog entry (for the reasoning-effort allowlist).
@@ -2068,7 +2145,7 @@ function parseAnthropicRequest(body, resolvedModel, model) {
 		tools,
 		toolChoice: parseToolChoice(body.tool_choice),
 		parallelToolCalls: parseDisableParallelToolUse(body.tool_choice),
-		reasoningEffort: body.thinking === void 0 ? defaultReasoningEffort(model) : parseReasoningEffort(body.thinking, model),
+		reasoningEffort: parseOutputConfigEffort(body, model) ?? (body.thinking === void 0 ? defaultReasoningEffort(model) : parseReasoningEffort(body.thinking, model)),
 		maxOutputTokens: maxTokens,
 		stopSequences: stopSequences && stopSequences.length > 0 ? stopSequences : void 0,
 		stream: body.stream === true
@@ -3405,7 +3482,7 @@ function clampOutputConfigEffortInPlace(body, model) {
 	const current = oc.effort;
 	if (typeof current !== "string") return false;
 	if (supported.includes(current)) return false;
-	const clamped = clampEffort(EFFORT_ORDER.includes(current) ? current : "xhigh", supported);
+	const clamped = clampEffort(EFFORT_ORDER.includes(current) ? current : UNKNOWN_EFFORT_ANCHOR, supported);
 	if (clamped === current) return false;
 	oc.effort = clamped;
 	return true;
@@ -4557,9 +4634,9 @@ function parseSharedArgs(args) {
 *
 * Display labels only: the gateway-model cache schema Claude Code reads is
 * `{id, display_name?}` per model — there is NO per-model context-window
-* field, so context accounting for a selected row uses Claude Code's
-* default window (safe under-accounting: it compacts earlier than the real
-* 1M/400k window, never overflows). See `seedGatewayModelCache`.
+* field. Context accounting is instead driven off the id itself, via the
+* `[1m]` literal-bracket suffix `nativeSelectableModelsInCatalog` attaches to
+* every row the catalog says serves >=1M. See `withOneMSuffix`.
 */
 const NATIVE_NON_CLAUDE_MODELS = [
 	{
@@ -4591,13 +4668,22 @@ const NATIVE_NON_CLAUDE_MODELS = [
 * dropped — the caller then neither enables discovery nor writes a cache
 * for it, and lesser tiers see the unchanged picker. Pure (reads
 * `state.models`), so it is unit-testable without side effects.
+*
+* The projected id carries a `[1m]` suffix when the catalog advertises a
+* >=1M window for it, because Claude Code budgets a gateway-discovered row
+* at its 200K default otherwise — `gpt-5.6-sol` (1,050,000) would compact at
+* roughly a fifth of its real window. `withOneMSuffix` is catalog-gated, so
+* `gpt-5.3-codex` (400K) stays bare and keeps the conservative accounting;
+* over-budgeting it would trade premature compaction for a hard overflow.
+* The lookup below still keys off the BARE id — the decoration is applied
+* only to the value handed to Claude Code.
 */
 function nativeSelectableModelsInCatalog() {
 	const catalog = state.models?.data;
 	if (!catalog || catalog.length === 0) return [];
 	const present = new Set(catalog.map((m) => m.id));
 	return NATIVE_NON_CLAUDE_MODELS.filter((m) => present.has(m.id)).map((m) => ({
-		id: m.id,
+		id: withOneMSuffix(m.id),
 		display_name: m.displayName
 	}));
 }
@@ -4773,4 +4859,4 @@ function getCodexEnvVars(serverUrl) {
 //#endregion
 export { sharedServerArgs as a, listModelsForEndpoint as c, updateClaude as d, setupAndServe as i, runSelfUpdate as l, getCodexEnvVars as n, startKeepAwake as o, parseSharedArgs as r, stopKeepAwake as s, getClaudeCodeEnvVars as t, checkClaudeVersion as u };
 
-//# sourceMappingURL=server-setup-CA0ZT-nq.js.map
+//# sourceMappingURL=server-setup-CJzxzYks.js.map