github-router 0.3.246 → 0.3.248

package/dist/{peer-mcp-personas-B4r1LmlQ.js → peer-mcp-personas-HWhxZUvf.js}

@@ -16338,25 +16338,25 @@ async function provisionColbert() {
 }
 async function provisionBinary(asset, dest) {
 	const sidecar = `${dest}.sha256`;
-	if (existsSync(dest) && await sidecarMatches(sidecar, asset.sha256)) return;
+	if (await installedArtifactIsIntact(sidecar, dest, asset.sha256)) return;
 	await mkdir(path.dirname(dest), { recursive: true });
 	const archive = await download(asset.url);
 	verifySha(archive, asset.sha256, "colgrep binary");
 	const member = await extractMember(asset, archive, "colgrep");
 	if (!member) throw new Error("colgrep binary not found in archive");
 	await atomicWrite(dest, member, true);
-	await writeFile(sidecar, asset.sha256).catch(() => {});
+	await writeFile(sidecar, sidecarContent(asset.sha256, member)).catch(() => {});
 }
 async function provisionOrt(asset, dest) {
 	const sidecar = `${dest}.sha256`;
-	if (existsSync(dest) && await sidecarMatches(sidecar, asset.sha256)) return;
+	if (await installedArtifactIsIntact(sidecar, dest, asset.sha256)) return;
 	await mkdir(path.dirname(dest), { recursive: true });
 	const archive = await download(asset.url);
 	verifySha(archive, asset.sha256, "ONNX Runtime");
 	const member = await extractMember(asset, archive, asset.member ?? "");
 	if (!member) throw new Error("ORT dylib not found in archive");
 	await atomicWrite(dest, member, true);
-	await writeFile(sidecar, asset.sha256).catch(() => {});
+	await writeFile(sidecar, sidecarContent(asset.sha256, member)).catch(() => {});
 	if (process$1.platform !== "win32" && asset.soname) {
 		const link = path.join(path.dirname(dest), asset.soname);
 		await rm(link, { force: true }).catch(() => {});
@@ -16431,13 +16431,44 @@ async function atomicWrite(dest, bytes, executable) {
 		}
 	}
 }
-async function sidecarMatches(sidecar, sha256) {
+/**
+* Is the artifact already installed AND still the bytes we installed?
+*
+* Sidecar format is two whitespace-separated hashes: the ARCHIVE sha (which
+* manifest revision this came from) and the INSTALLED-FILE sha (what we wrote
+* to disk). Both are required.
+*
+* The archive hash alone is not enough, and trusting it caused a real outage.
+* `existsSync(dest) && sidecarMatches(archiveSha)` treats any file at the path
+* as good, so when a lane-1 test with a leaked `node:os` mock overwrote the
+* real `colgrep.exe` with a 6-byte stub, provisioning kept short-circuiting on
+* the still-valid sidecar and never re-downloaded. The smoke test then failed
+* every launch, removed `.smoke-ok`, and semantic search reported "unavailable
+* on this host" forever. Only manually deleting the binary could recover it.
+*
+* A legacy single-hash sidecar means the installed bytes were never recorded,
+* so their provenance is unknown. That returns false deliberately: one extra
+* download self-heals every install already corrupted by the above, which is
+* the whole point of checking.
+*
+* This matches what the model files already do (they hash installed content
+* before reuse); the binary and ORT were the two that skipped it.
+*/
+async function installedArtifactIsIntact(sidecar, dest, archiveSha) {
 	try {
-		return (await readFile(sidecar, "utf8")).trim() === sha256;
+		if (!existsSync(dest)) return false;
+		const [recordedArchive, recordedInstalled] = (await readFile(sidecar, "utf8")).trim().split(/\s+/);
+		if (recordedArchive !== archiveSha) return false;
+		if (!recordedInstalled) return false;
+		return createHash("sha256").update(await readFile(dest)).digest("hex") === recordedInstalled;
 	} catch {
 		return false;
 	}
 }
+/** Sidecar content pairing the archive revision with the installed bytes. */
+function sidecarContent(archiveSha, installed) {
+	return `${archiveSha}\n${createHash("sha256").update(installed).digest("hex")}`;
+}
 /**
 * Post-provision smoke test. Runs ONE cheap colgrep invocation with the
 * EXACT isolating env the runner uses (`COLGREP_DATA_DIR`,
@@ -17689,18 +17720,30 @@ function lexicalSearchCodeMode(mode) {
 * Status-specific, actionable fallback hint. The semantic index isn't ready,
 * so the model got LEXICAL results (great for exact symbols, sparse for a
 * natural-language phrase since the lexical backend matches literally). Tell
-* it both levers: retry `mode:"semantic"` shortly (the index is self-healing
-* in the background) OR re-query now with specific symbol/keyword terms.
+* it both levers: retry `mode:"semantic"` (the index is self-healing in the
+* background) OR re-query now with specific symbol/keyword terms.
+*
+* "shortly" was too vague and cost a real recovery. A build takes MINUTES on a
+* large repo, and after a failed build the FIRST query is consumed triggering
+* the re-kick and still returns a fallback. So a caller who retried once,
+* seconds later, saw a second fallback and concluded the tool was broken
+* rather than mid-repair. Observed end to end on this repo: query, then
+* ~5 minutes of `building`, then `ready`. Naming the timescale and the
+* one-query-to-trigger behaviour is what turns "it's still broken" into
+* "it's coming back".
+*
+* The wording stays a range rather than a number: build time scales with
+* repository size, so a hard figure would be wrong for most callers.
 */
 const FALLBACK_GUIDANCE_MARKER = "retry mode:\"semantic\"";
-const FALLBACK_GUIDANCE = `${FALLBACK_GUIDANCE_MARKER} shortly, or re-query now with specific symbol/keyword terms`;
+const FALLBACK_GUIDANCE = `${FALLBACK_GUIDANCE_MARKER} in a few minutes (a build takes minutes on a large repo), or re-query now with specific symbol/keyword terms`;
 function fallbackNoticeFor(status) {
 	const tail = FALLBACK_GUIDANCE;
 	switch (status) {
-		case "building": return `semantic index is building; returned lexical keyword matches — ${tail}`;
-		case "stale": return `semantic index predates the current HEAD/tree (a background re-index was started); returned lexical keyword matches — ${tail}`;
-		case "unavailable": return `no semantic index for this workspace yet (a background build was started); returned lexical keyword matches — ${tail}`;
-		case "failed": return `semantic index unavailable (build failing — see proxy logs); returned lexical keyword matches — ${tail}`;
+		case "building": return `semantic index is building; returned lexical keyword matches. ${tail}`;
+		case "stale": return `semantic index predates the current HEAD/tree (a background re-index was started); returned lexical keyword matches. ${tail}`;
+		case "unavailable": return `no semantic index for this workspace yet (a background build was started); returned lexical keyword matches. ${tail}`;
+		case "failed": return `semantic index unavailable; this query started a background rebuild, so it returned lexical keyword matches. ${tail}`;
 		default: return "returned lexical results";
 	}
 }
@@ -23458,6 +23501,54 @@ function resolveModelAndThinking(opts) {
 	if (!clamp) clamp = allowed[0];
 	return mkOk(clamp);
 }
+/** Worker-usable models need a big enough window to be worth delegating to. */
+const CATALOG_MIN_CONTEXT = 2e5;
+/**
+* Derived view of the live catalog: every model a worker could actually be
+* pointed at, with the metadata needed to choose between them.
+*
+* DERIVED ONLY, and that is the whole design. A one-liner like "strong
+* reasoning, weak long-context recall" cannot be computed from catalog
+* metadata — it is editorial, it goes stale silently as vendors ship, and the
+* asymmetry is brutal: a MISSING characterization costs one suboptimal pick
+* the model recovers from, while a WRONG one misroutes invisibly at the call
+* site. So this ships facts and lets the caller judge.
+*
+* It exists because the hardcoded chains cannot discover anything. Models are
+* live in the catalog that appear nowhere in `src/` — nobody evaluated them
+* because nothing surfaced them. That is a DISCOVERABILITY gap, not a
+* capability gap, which is also why the per-mode and per-agent defaults are
+* deliberately left alone: they encode cross-lab DECORRELATION policy, not
+* just quality. `reviewerModel()` must differ from the implementer's lab so a
+* model never reviews its own output, and no capability table can express
+* "must differ from whoever produced this". `vendor` is included precisely so
+* a caller can reason about lab diversity without being handed a ranking.
+*
+* Efforts are clamped to WORKER_THINKING_LEVELS: nine live models advertise a
+* `max` tier above `xhigh` that the worker layer filters out, so showing the
+* raw array would advertise an effort no worker can request.
+*/
+function buildCatalogView() {
+	const rows = [];
+	for (const model of state.models?.data ?? []) {
+		const supports = model.capabilities?.supports;
+		const limits = model.capabilities?.limits;
+		if (supports?.tool_calls !== true) continue;
+		const ctx = limits?.max_context_window_tokens ?? 0;
+		if (ctx < CATALOG_MIN_CONTEXT) continue;
+		const efforts = (supports.reasoning_effort ?? []).filter((effort) => WORKER_THINKING_LEVELS.includes(effort));
+		if (efforts.length === 0) continue;
+		rows.push({
+			id: model.id,
+			vendor: model.vendor,
+			ctx,
+			...limits?.max_output_tokens ? { maxOut: limits.max_output_tokens } : {},
+			efforts,
+			...model.model_picker_price_category ? { cost: model.model_picker_price_category } : {}
+		});
+	}
+	return rows.sort((a, b) => a.id.localeCompare(b.id));
+}
 //#endregion
 //#region src/lib/worker-agent/session-defaults.ts
 const MODES = Object.freeze([
@@ -34085,10 +34176,11 @@ const NON_PERSONA_MCP_TOOLS = Object.freeze([
 					thinking
 				});
 			}
-			const table = Object.fromEntries(WORKER_MODES.map((workerMode) => [workerMode, resolveModeDefaults(workerMode)]));
+			const body = { ...Object.fromEntries(WORKER_MODES.map((workerMode) => [workerMode, resolveModeDefaults(workerMode)])) };
+			if (!clearAll && !clear && mode === void 0) body.catalog = buildCatalogView();
 			return { content: [{
 				type: "text",
-				text: JSON.stringify(table)
+				text: JSON.stringify(body)
 			}] };
 		}
 	},
@@ -34972,4 +35064,4 @@ function enumerateInjectedMcpToolNames(groupKeys, opts = {}) {
 //#endregion
 export { searchWeb as $, DEFAULT_CLAUDE_MODEL_FALLBACKS as $t, trustRepo as A, state as An, getTokenCount as At, TEST_DEFAULT_MODEL as B, hasSupportedBrowserInstalled as Bt, fileLastPromptStore as C, fetchWithTransientRetry as Cn, scoutModel as Ct, repoRoot as D, copilotBaseUrl as Dn, shimDefaultsToXhigh as Dt, repoFingerprint as E, GITHUB_API_BASE_URL as En, workerToolsEnabled as Et, EXPLORE_DEFAULT_MODEL as F, createChatCompletions as Ft, buildEnv as G, CONDENSED_OPERATING_SEQUENCE as Gt, resolveModeDefaults as H, extractTarGzMember as Ht, EXPLORE_DEFAULT_THINKING as I, MAX_RESPONSE_BODY_BYTES as It, toolbeltEnabled as J, ArtifactClient as Jt, availableToolCommands as K, DEFINITION_OF_GREATNESS as Kt, IMPLEMENT_DEFAULT_MODEL as L, readResponseBodyCapped as Lt, resolveSealedGate as M, resolveMcpToolTimeoutMs as Mt, BROWSE_DEFAULT_MODEL as N, pickEndpoint as Nt, stopGateEnabledForRepo as O, copilotHeaders as On, countTokens as Ot, DEFAULT_MODEL as P, createResponses as Pt, assetFor as Q, toolbeltPathOverride as Qt, PLAN_DEFAULT_MODEL as R, parseJsonOrDiagnose as Rt, fileFindingsStore as S, getGitHubUser as Sn, reviewerModel as St, isSubagentContext as T, forwardError as Tn, standInToolEnabled as Tt, resolveWorkerRunOpts as U, extractZipMember as Ut, appendPlanReminder as V, provisionAndIndexColbert as Vt, runWorkerAgent as W, warmTreeSitterPool as Wt, vscodeRipgrepPath as X, buildWorkspaceHeaderJson as Xt, toolbeltSkipSet as Y, buildWorkspaceHeaderHelperCommand as Yt, TOOLBELT_TOOLS$1 as Z, collapsePathKeys as Zt, stopGateDisabled as _, isNullish as _n, browserCompoundToolsEnabled as _t, buildPeerAwarenessSnippet as a, generateRandomPort as an, buildAnthropicErrorEvent as at, stopReviewEnabled as b, sleep$3 as bn, geminiAvailable as bt, personasFor as c, withInstallLock as cn, logStreamError as ct, buildStopHookCommand as d, setupGitHubToken as dn, handleMcpDelete as dt, DEFAULT_CODEX_MODEL as en, ADVISOR_INTERNAL_TOOL_NAME as et, captureLaunchBaseline as f, tryRefreshAndRetry as fn, handleMcpPost as ft, launchBaselineKey as g, filterBetaHeader as gn, browseAgentEnabled as gt, injectStopHookIntoSettingsFile as h, cacheVSCodeVersion as hn, brainstormModel as ht, buildAgentPrompt as i, UPSTREAM_INACTIVITY_TIMEOUT_MS as in, isAdvisorRequested as it, liveExec as j, assembleResponsesPayload as jt, stopReviewStateDir as k, githubHeaders as kn, createMessages as kt, buildArtifactOpenHookCommand as l, setupCopilotToken as ln, readIteratorWithTimeout as lt, fileBlockBudget as m, cacheModels as mn, artifactToolsEnabled as mt, MCP_GROUPS as n, DEFAULT_PORT as nn, buildAdvisorStream as nt, buildPeerAwarenessSummary as o, pickClaudeDefault as on, buildOpenAIErrorEvent as ot, decideStopHook as p, cacheCopilotVersion as pn, agentToolsEnabled as pt, buildToolbeltAwareness as q, shouldUseInsecureTls as qt, assertMcpToolSurfaceConsistent as r, UPSTREAM_FETCH_TIMEOUT_MS as rn, injectAdvisorTool as rt, enumerateInjectedMcpToolNames as s, getPackageVersion as sn, isControllerClosedError as st, GROUP_META as t, DEFAULT_CODEX_MODEL_FALLBACKS as tn, ADVISOR_TOOL_INSTRUCTIONS as tt, buildSessionBindHookCommand as u, setupGitHubAgentToken as un, relayAnthropicStream as ut, stopGateId as v, resolveCodexModel as vn, browserToolsEnabled as vt, fileReviewDebounce as w, HTTPError as wn, scribeModel as wt, fileBaselineStore as x, getModels as xn, nativeSubagentModel as xt, stopGatePlanMode as y, resolveModel as yn, fleetToolsEnabled as yt, REVIEW_DEFAULT_MODEL as z, provisionBrowserAssets as zt };
 
-//# sourceMappingURL=peer-mcp-personas-B4r1LmlQ.js.map
+//# sourceMappingURL=peer-mcp-personas-HWhxZUvf.js.map