llm-relay 0.57.0 → 0.59.0

package/dist/candidates.d.ts

@@ -129,8 +129,8 @@ export interface Candidate {
     /**
      * What this deployment (or its provider, or its group) has STATED about itself — the learned
      * facts from `target-facts.ts`, each with the scope it applies at. The measurement kinds
-     * (`context-limit`, `rate-limit-rpm|rpd|tpm|tpd`) also carry `value`: the ceiling itself, as
-     * stated. Display-only — nothing here reorders or gates a candidate.
+     * (`context-limit`, `max-output`, `rate-limit-rpm|rpd|tpm|tpd`) also carry `value`: the ceiling
+     * itself, as stated. Display-only — nothing here reorders or gates a candidate.
      */
     facts: Array<{
         kind: string;

package/dist/cli.d.ts

@@ -260,11 +260,13 @@ export declare function renderCommand(invoke: {
 /**
  * `llm-relay lanes [--probe]` — what each cli lane's own tool says it serves.
  *
- * ⚠ `--probe` is the ONE place the relay runs a lane's command, and only as an explicit operator
- * action — same precedent as `pools --probe` sending real completions. The request path reads the
- * cached manifest and never spawns anything. Without `--probe` this just prints the cache.
+ * ⚠ `--probe` spawns lane commands as an explicit operator action — same precedent as
+ * `pools --probe` sending real completions. The request path reads the cached manifest and never
+ * spawns anything; outside it the only OTHER spawn site is the relay's background lane cadence
+ * (owner decision 2026-08-29, docs/quota-reprobe-design-2026-08-29.md). Without `--probe` this
+ * just prints the cache.
  */
-export declare function runLanes(): void;
+export declare function runLanes(): Promise<void>;
 /**
  * `llm-relay dispatch [lane]` — which lane to hand a delegated task to next.
  *

package/dist/cli.js

@@ -12,9 +12,10 @@ import { offloadState, setOffload } from "./offload.js";
 import { buildCandidates } from "./candidates.js";
 import { CREDENTIAL_LABEL_PATTERN, makeCredentialId } from "./credential-id.js";
 import { providerCredentialSlots, slotAllowsModel } from "./credential-fleet.js";
-import { loadLaneManifest, verifyModel } from "./lane-manifest.js";
+import { loadLaneManifest, rosterIsStale, verifyModel } from "./lane-manifest.js";
 import { probeLanes } from "./lane-probe.js";
-import { buildDispatch, normalizeCliCommand, specContextWindow, CONTEXT_TOKEN } from "./dispatch.js";
+import { buildDispatch, normalizeCliCommand, restoreExhaustedRows, specContextWindow, CONTEXT_TOKEN } from "./dispatch.js";
+import { loadExhaustedRows } from "./dispatch-exhaustion-persistence.js";
 import { detectHostRouting, parseHostRoutingState } from "./host-routing.js";
 import { contextWindowResolver, COST_CLASSES } from "./metadata.js";
 import { snapshotContextWindow } from "./tier-data.js";
@@ -1674,14 +1675,16 @@ export function renderCommand(invoke, shell = shellFor()) {
 /**
  * `llm-relay lanes [--probe]` — what each cli lane's own tool says it serves.
  *
- * ⚠ `--probe` is the ONE place the relay runs a lane's command, and only as an explicit operator
- * action — same precedent as `pools --probe` sending real completions. The request path reads the
- * cached manifest and never spawns anything. Without `--probe` this just prints the cache.
+ * ⚠ `--probe` spawns lane commands as an explicit operator action — same precedent as
+ * `pools --probe` sending real completions. The request path reads the cached manifest and never
+ * spawns anything; outside it the only OTHER spawn site is the relay's background lane cadence
+ * (owner decision 2026-08-29, docs/quota-reprobe-design-2026-08-29.md). Without `--probe` this
+ * just prints the cache.
  */
-export function runLanes() {
+export async function runLanes() {
     const cfg = loadOrExit();
     if (hasFlag("--probe")) {
-        for (const r of probeLanes(cfg)) {
+        for (const r of await probeLanes(cfg)) {
             process.stdout.write(r.ok
                 ? `  ✓ ${r.lane.padEnd(8)} ${String(r.modelCount).padStart(3)} models via \`${r.via}\`
 `
@@ -1697,8 +1700,11 @@ export function runLanes() {
         return;
     }
     for (const [lane, entry] of Object.entries(manifest.lanes)) {
+        const stale = rosterIsStale(entry)
+            ? "  ⚠ STALE — no longer evidence for eviction; run `llm-relay lanes --probe`"
+            : "";
         process.stdout.write(`
-${lane} — ${entry.models.length} models, probed ${entry.probedAt} via \`${entry.via}\`
+${lane} — ${entry.models.length} models, probed ${entry.probedAt} via \`${entry.via}\`${stale}
 `);
         for (const m of entry.models) {
             const supports = m.supports
@@ -1722,7 +1728,7 @@ ${lane} — ${entry.models.length} models, probed ${entry.probedAt} via \`${entr
         const model = i >= 0 ? rung.args[i + 1] : undefined;
         if (!model)
             continue;
-        const v = verifyModel(manifest, rung.command, model);
+        const v = verifyModel(manifest, rung.command, model, { args: rung.args });
         if (v.status === "not-servable" && !bad.includes(v.reason))
             bad.push(v.reason);
     }
@@ -1846,6 +1852,11 @@ export async function runDispatch(arg) {
             specContextWindow(l.spec, cfg, cachedContextWindow) !== null);
     const staleProxy = hostStale || windowStale;
     const live = staleProxy ? null : liveRaw;
+    // The durable half of exhaustion state is readable cold since dispatch-exhaustion persistence
+    // landed: restore still-future rows so the fallback ladder shows what the relay recorded.
+    // In-memory-only rows on a running relay still need the live answer — this narrows the gap, it
+    // does not close it. Read-only here: no listener is installed, so nothing writes back.
+    restoreExhaustedRows(cfg, loadExhaustedRows());
     const view = normalizeDispatchCommands(live ??
         buildDispatch(cfg, {
             ...(task ? { task } : {}),
@@ -2230,6 +2241,7 @@ const FACT_MEANING = {
     "credential-invalid": "the provider says this key is bad — every deployment behind it demoted",
     "rate-limited": "throughput throttled by the provider — demoted, never evicted",
     "context-limit": "a ceiling this deployment stated about itself — a measurement, never a demotion",
+    "max-output": "a ceiling this deployment stated about itself — a measurement, never a demotion",
     "rate-limit-rpm": "a ceiling this deployment stated about itself — a measurement, never a demotion",
     "rate-limit-rpd": "a ceiling this deployment stated about itself — a measurement, never a demotion",
     "rate-limit-tpm": "a ceiling this deployment stated about itself — a measurement, never a demotion",
@@ -3284,14 +3296,10 @@ export function main() {
         return;
     }
     if (arg2 === "lanes") {
-        try {
-            runLanes();
-        }
-        catch (e) {
-            process.stderr.write(`llm-relay lanes: ${e.message}
-`);
+        runLanes().catch((e) => {
+            process.stderr.write(`llm-relay lanes: ${e.message}\n`);
             process.exit(1);
-        }
+        });
         return;
     }
     if (arg2 === "dispatch") {