llm-relay 0.57.0 → 0.58.0

package/dist/context-limits.d.ts

@@ -12,6 +12,10 @@
  * characters per token, and persisting it would put a guess into the one store whose whole value
  * is that it contains measurements. If nothing parses, nothing is learned. Same rule as
  * `resolveMetadata`: no rung may be a guess.
+ *
+ * The OUTPUT-token sibling lives here too (the `max-output` half below): a deployment that
+ * rejects an over-sized `max_tokens` often states its real output ceiling in the same kind of
+ * error body, under the same discipline.
  */
 /**
  * ⚠ **The storage moved; the PARSING is what this module is.** Ceilings now live in
@@ -48,7 +52,33 @@ export declare function observedContextLimit(provider: string, model: string, op
     path?: string;
     now?: number;
 }): number | null;
-/** Flush pending observations. Called on shutdown, like the other write-behind stores. */
+/** Retained for callers and docs that name the output ceiling's staleness window (30 days). */
+export declare const OBSERVED_MAX_OUTPUT_TTL_MS: number;
+/**
+ * Extract a stated output-token ceiling from an error body, or null.
+ *
+ * Returns null for anything it cannot read as an explicit maximum — "max_tokens is too large"
+ * alone proves the request overshot but states no ceiling. Same fail-safe as
+ * `parseStatedContextLimit`: if nothing parses, nothing is learned.
+ */
+export declare function parseStatedMaxOutput(body: string): number | null;
+/** Does this error body describe a `max_tokens` / output-cap rejection at all? */
+export declare function looksLikeMaxOutputError(body: string): boolean;
+/**
+ * Record an output ceiling a deployment stated about itself. Deployment scope — the message names
+ * the model, never the account. A fresh observation replaces an older one, same as the context
+ * half: the deployment is the authority on its own ceiling.
+ */
+export declare function recordObservedMaxOutput(provider: string, model: string, tokens: number, opts?: {
+    path?: string;
+    now?: number;
+}): void;
+/** The learned output ceiling for a deployment, or null when none was observed or it expired. */
+export declare function observedMaxOutput(provider: string, model: string, opts?: {
+    path?: string;
+    now?: number;
+}): number | null;
+/** Flush pending observations — BOTH halves; the store is shared. Called on shutdown. */
 export declare function flushObservedContextLimits(opts?: {
     path?: string;
 }): void;

package/dist/context-limits.js

@@ -13,6 +13,10 @@ import { factsFor, flushFacts, recordFact, resetFacts, FACT_TTL_MS } from "./tar
  * characters per token, and persisting it would put a guess into the one store whose whole value
  * is that it contains measurements. If nothing parses, nothing is learned. Same rule as
  * `resolveMetadata`: no rung may be a guess.
+ *
+ * The OUTPUT-token sibling lives here too (the `max-output` half below): a deployment that
+ * rejects an over-sized `max_tokens` often states its real output ceiling in the same kind of
+ * error body, under the same discipline.
  */
 /**
  * ⚠ **The storage moved; the PARSING is what this module is.** Ceilings now live in
@@ -92,7 +96,88 @@ export function observedContextLimit(provider, model, opts = {}) {
     }
     return null;
 }
-/** Flush pending observations. Called on shutdown, like the other write-behind stores. */
+// ── Max-output ceilings — the OUTPUT-token sibling of the context parser above ─────────────────
+//
+// The groq case that motivated this half: 13 requests answered 400 with "`max_tokens` must be
+// less than or equal to `8192`…" and no store could hold the stated value, so the failure
+// repeated on every walk that landed there. Same store, same TTL convention, same fail-safe:
+// only an EXPLICITLY stated maximum is recorded, a miss learns nothing, and the fact is
+// DISPLAY-ONLY — nothing clamps, refuses or routes on it
+// (docs/max-output-caps-design-2026-08-29.md).
+/** Retained for callers and docs that name the output ceiling's staleness window (30 days). */
+export const OBSERVED_MAX_OUTPUT_TTL_MS = FACT_TTL_MS["max-output"];
+/**
+ * Patterns that carry an explicitly stated `max_tokens` ceiling. Every pattern must capture the
+ * MAXIMUM, never the requested count — same rule as `STATED_LIMIT_PATTERNS`, for the same reason:
+ * the requested number is the larger one, and persisting it would overstate the real ceiling.
+ */
+const STATED_MAX_OUTPUT_PATTERNS = [
+    // Groq: "`max_tokens` must be less than or equal to `8192`, the maximum value for `max_tokens`
+    // is less than the `context_window` for this model" — the ceiling is the number after the
+    // comparator, and the tail clause carries no number at all. Also TGI's "`max_new_tokens` must
+    // be <= 4096".
+    /max_(?:completion_|output_|new_)?tokens`?\s*must be (?:less than or equal to|<=)\s*`?(\d[\d,_]*)/i,
+    // OpenAI: "max_tokens is too large: 40000. This model supports at most 16384 completion tokens."
+    /supports?\s+at most\s+(\d[\d,_]*)\s+(?:completion|output)\s+tokens/i,
+    // Anthropic-style: "max_tokens: 40000 > 8192, which is the maximum allowed number of output
+    // tokens…" — anchored to the maximum wording, so a bare comparison never matches. Two fixed
+    // terminator forms rather than one optional-comma alternation: the gap class excludes ">" and
+    // each terminator starts unambiguously, which keeps the scan linear.
+    /max_tokens[^.\n>]{0,24}>\s*(\d[\d,_]*)\s+maximum\b/i,
+    /max_tokens[^.\n>]{0,24}>\s*(\d[\d,_]*),\s*which is the maximum/i,
+    // OpenAI-compatible front-ends that restate the field: "the maximum value for max_tokens is 8192".
+    /maximum value for\s+`?max_(?:completion_|output_|new_)?tokens`?\s*(?:is|:)\s*`?(\d[\d,_]*)/i,
+];
+/**
+ * Extract a stated output-token ceiling from an error body, or null.
+ *
+ * Returns null for anything it cannot read as an explicit maximum — "max_tokens is too large"
+ * alone proves the request overshot but states no ceiling. Same fail-safe as
+ * `parseStatedContextLimit`: if nothing parses, nothing is learned.
+ */
+export function parseStatedMaxOutput(body) {
+    if (typeof body !== "string" || body.length === 0)
+        return null;
+    // Bodies are small; this bound only stops a pathological one from driving the regex engine.
+    const text = body.length > 8192 ? body.slice(0, 8192) : body;
+    for (const re of STATED_MAX_OUTPUT_PATTERNS) {
+        const m = re.exec(text);
+        if (!m?.[1])
+            continue;
+        const n = Number(m[1].replace(/[,_]/g, ""));
+        if (Number.isFinite(n) && n > 0 && n <= MAX_CREDIBLE_TOKENS)
+            return Math.floor(n);
+    }
+    return null;
+}
+/** Does this error body describe a `max_tokens` / output-cap rejection at all? */
+export function looksLikeMaxOutputError(body) {
+    return /max_(?:completion_|output_|new_)?tokens|(?:completion|output)\s+tokens/i.test(body);
+}
+/**
+ * Record an output ceiling a deployment stated about itself. Deployment scope — the message names
+ * the model, never the account. A fresh observation replaces an older one, same as the context
+ * half: the deployment is the authority on its own ceiling.
+ */
+export function recordObservedMaxOutput(provider, model, tokens, opts = {}) {
+    if (!Number.isFinite(tokens) || tokens <= 0 || tokens > MAX_CREDIBLE_TOKENS)
+        return;
+    recordFact("max-output", { kind: "deployment", provider, model }, {
+        ...(opts.path !== undefined ? { path: opts.path } : {}),
+        ...(opts.now !== undefined ? { now: opts.now } : {}),
+        value: Math.floor(tokens),
+    });
+}
+/** The learned output ceiling for a deployment, or null when none was observed or it expired. */
+export function observedMaxOutput(provider, model, opts = {}) {
+    // Output ceilings are deployment-scoped measurements, not credential-scoped facts.
+    for (const fact of factsFor(provider, null, model, opts)) {
+        if (fact.kind === "max-output" && typeof fact.value === "number")
+            return fact.value;
+    }
+    return null;
+}
+/** Flush pending observations — BOTH halves; the store is shared. Called on shutdown. */
 export function flushObservedContextLimits(opts = {}) {
     flushFacts(opts);
 }

package/dist/context-limits.js.map

@@ -1 +1 @@
-{"version":3,"file":"context-limits.js","sourceRoot":"","sources":["../src/context-limits.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,QAAQ,EAAE,UAAU,EAAE,UAAU,EAAE,UAAU,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAE9F;;;;;;;;;;;;;;GAcG;AAEH;;;;;;;;;GASG;AAEH,8EAA8E;AAC9E,MAAM,CAAC,MAAM,qBAAqB,GAAG,WAAW,CAAC,eAAe,CAAC,CAAC;AAElE,6EAA6E;AAC7E,MAAM,mBAAmB,GAAG,WAAW,CAAC;AAExC;;;;;GAKG;AACH,MAAM,qBAAqB,GAAa;IACtC,+FAA+F;IAC/F,mDAAmD;IACnD,wEAAwE;IACxE,qCAAqC;IACrC,qFAAqF;IACrF,iFAAiF;IACjF,wDAAwD;IACxD,4EAA4E;CAC7E,CAAC;AAEF;;;;;GAKG;AACH,MAAM,UAAU,uBAAuB,CAAC,IAAY;IAClD,IAAI,OAAO,IAAI,KAAK,QAAQ,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IAC/D,4FAA4F;IAC5F,MAAM,IAAI,GAAG,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC;IAC7D,KAAK,MAAM,EAAE,IAAI,qBAAqB,EAAE,CAAC;QACvC,MAAM,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACxB,IAAI,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;YAAE,SAAS;QACtB,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,CAAC;QAC5C,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,mBAAmB;YAAE,OAAO,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IACpF,CAAC;IACD,OAAO,IAAI,CAAC;AACd,CAAC;AAED,uEAAuE;AACvE,MAAM,UAAU,2BAA2B,CAAC,IAAY;IACtD,OAAO,2EAA2E,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAChG,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,0BAA0B,CACxC,QAAgB,EAChB,KAAa,EACb,MAAc,EACd,OAAwC,EAAE;IAE1C,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,MAAM,IAAI,CAAC,IAAI,MAAM,GAAG,mBAAmB;QAAE,OAAO;IACpF,UAAU,CAAC,eAAe,EAAE,EAAE,IAAI,EAAE,YAAY,EAAE,QAAQ,EAAE,KAAK,EAAE,EAAE;QACnE,GAAG,CAAC,IAAI,CAAC,IAAI,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACvD,GAAG,CAAC,IAAI,CAAC,GAAG,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACpD,KAAK,EAAE,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC;KAC1B,CAAC,CAAC;AACL,CAAC;AAED,8FAA8F;AAC9F,MAAM,UAAU,oBAAoB,CAClC,QAAgB,EAChB,KAAa,EACb,OAAwC,EAAE;IAE1C,oFAAoF;IACpF,KAAK,MAAM,IAAI,IAAI,QAAQ,CAAC,QAAQ,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,CAAC;QACzD,IAAI,IAAI,CAAC,IAAI,KAAK,eAAe,IAAI,OAAO,IAAI,CAAC,KAAK,KAAK,QAAQ;YAAE,OAAO,IAAI,CAAC,KAAK,CAAC;IACzF,CAAC;IACD,OAAO,IAAI,CAAC;AACd,CAAC;AAED,0FAA0F;AAC1F,MAAM,UAAU,0BAA0B,CAAC,OAA0B,EAAE;IACrE,UAAU,CAAC,IAAI,CAAC,CAAC;AACnB,CAAC;AAED,gFAAgF;AAChF,MAAM,UAAU,0BAA0B;IACxC,UAAU,EAAE,CAAC;AACf,CAAC"}
+{"version":3,"file":"context-limits.js","sourceRoot":"","sources":["../src/context-limits.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,QAAQ,EAAE,UAAU,EAAE,UAAU,EAAE,UAAU,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAE9F;;;;;;;;;;;;;;;;;;GAkBG;AAEH;;;;;;;;;GASG;AAEH,8EAA8E;AAC9E,MAAM,CAAC,MAAM,qBAAqB,GAAG,WAAW,CAAC,eAAe,CAAC,CAAC;AAElE,6EAA6E;AAC7E,MAAM,mBAAmB,GAAG,WAAW,CAAC;AAExC;;;;;GAKG;AACH,MAAM,qBAAqB,GAAa;IACtC,+FAA+F;IAC/F,mDAAmD;IACnD,wEAAwE;IACxE,qCAAqC;IACrC,qFAAqF;IACrF,iFAAiF;IACjF,wDAAwD;IACxD,4EAA4E;CAC7E,CAAC;AAEF;;;;;GAKG;AACH,MAAM,UAAU,uBAAuB,CAAC,IAAY;IAClD,IAAI,OAAO,IAAI,KAAK,QAAQ,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IAC/D,4FAA4F;IAC5F,MAAM,IAAI,GAAG,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC;IAC7D,KAAK,MAAM,EAAE,IAAI,qBAAqB,EAAE,CAAC;QACvC,MAAM,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACxB,IAAI,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;YAAE,SAAS;QACtB,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,CAAC;QAC5C,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,mBAAmB;YAAE,OAAO,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IACpF,CAAC;IACD,OAAO,IAAI,CAAC;AACd,CAAC;AAED,uEAAuE;AACvE,MAAM,UAAU,2BAA2B,CAAC,IAAY;IACtD,OAAO,2EAA2E,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAChG,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,0BAA0B,CACxC,QAAgB,EAChB,KAAa,EACb,MAAc,EACd,OAAwC,EAAE;IAE1C,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,MAAM,IAAI,CAAC,IAAI,MAAM,GAAG,mBAAmB;QAAE,OAAO;IACpF,UAAU,CAAC,eAAe,EAAE,EAAE,IAAI,EAAE,YAAY,EAAE,QAAQ,EAAE,KAAK,EAAE,EAAE;QACnE,GAAG,CAAC,IAAI,CAAC,IAAI,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACvD,GAAG,CAAC,IAAI,CAAC,GAAG,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACpD,KAAK,EAAE,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC;KAC1B,CAAC,CAAC;AACL,CAAC;AAED,8FAA8F;AAC9F,MAAM,UAAU,oBAAoB,CAClC,QAAgB,EAChB,KAAa,EACb,OAAwC,EAAE;IAE1C,oFAAoF;IACpF,KAAK,MAAM,IAAI,IAAI,QAAQ,CAAC,QAAQ,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,CAAC;QACzD,IAAI,IAAI,CAAC,IAAI,KAAK,eAAe,IAAI,OAAO,IAAI,CAAC,KAAK,KAAK,QAAQ;YAAE,OAAO,IAAI,CAAC,KAAK,CAAC;IACzF,CAAC;IACD,OAAO,IAAI,CAAC;AACd,CAAC;AAED,kGAAkG;AAClG,EAAE;AACF,8FAA8F;AAC9F,0FAA0F;AAC1F,6FAA6F;AAC7F,wFAAwF;AACxF,yDAAyD;AACzD,+CAA+C;AAE/C,+FAA+F;AAC/F,MAAM,CAAC,MAAM,0BAA0B,GAAG,WAAW,CAAC,YAAY,CAAC,CAAC;AAEpE;;;;GAIG;AACH,MAAM,0BAA0B,GAAa;IAC3C,+FAA+F;IAC/F,0FAA0F;IAC1F,8FAA8F;IAC9F,eAAe;IACf,mGAAmG;IACnG,iGAAiG;IACjG,qEAAqE;IACrE,4FAA4F;IAC5F,4FAA4F;IAC5F,8FAA8F;IAC9F,qEAAqE;IACrE,qDAAqD;IACrD,iEAAiE;IACjE,mGAAmG;IACnG,6FAA6F;CAC9F,CAAC;AAEF;;;;;;GAMG;AACH,MAAM,UAAU,oBAAoB,CAAC,IAAY;IAC/C,IAAI,OAAO,IAAI,KAAK,QAAQ,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IAC/D,4FAA4F;IAC5F,MAAM,IAAI,GAAG,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC;IAC7D,KAAK,MAAM,EAAE,IAAI,0BAA0B,EAAE,CAAC;QAC5C,MAAM,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACxB,IAAI,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;YAAE,SAAS;QACtB,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,CAAC;QAC5C,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,mBAAmB;YAAE,OAAO,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IACpF,CAAC;IACD,OAAO,IAAI,CAAC;AACd,CAAC;AAED,kFAAkF;AAClF,MAAM,UAAU,uBAAuB,CAAC,IAAY;IAClD,OAAO,yEAAyE,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAC9F,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,uBAAuB,CACrC,QAAgB,EAChB,KAAa,EACb,MAAc,EACd,OAAwC,EAAE;IAE1C,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,MAAM,IAAI,CAAC,IAAI,MAAM,GAAG,mBAAmB;QAAE,OAAO;IACpF,UAAU,CAAC,YAAY,EAAE,EAAE,IAAI,EAAE,YAAY,EAAE,QAAQ,EAAE,KAAK,EAAE,EAAE;QAChE,GAAG,CAAC,IAAI,CAAC,IAAI,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACvD,GAAG,CAAC,IAAI,CAAC,GAAG,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACpD,KAAK,EAAE,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC;KAC1B,CAAC,CAAC;AACL,CAAC;AAED,iGAAiG;AACjG,MAAM,UAAU,iBAAiB,CAC/B,QAAgB,EAChB,KAAa,EACb,OAAwC,EAAE;IAE1C,mFAAmF;IACnF,KAAK,MAAM,IAAI,IAAI,QAAQ,CAAC,QAAQ,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,CAAC;QACzD,IAAI,IAAI,CAAC,IAAI,KAAK,YAAY,IAAI,OAAO,IAAI,CAAC,KAAK,KAAK,QAAQ;YAAE,OAAO,IAAI,CAAC,KAAK,CAAC;IACtF,CAAC;IACD,OAAO,IAAI,CAAC;AACd,CAAC;AAED,yFAAyF;AACzF,MAAM,UAAU,0BAA0B,CAAC,OAA0B,EAAE;IACrE,UAAU,CAAC,IAAI,CAAC,CAAC;AACnB,CAAC;AAED,gFAAgF;AAChF,MAAM,UAAU,0BAA0B;IACxC,UAAU,EAAE,CAAC;AACf,CAAC"}

package/dist/server.js

@@ -35,7 +35,7 @@ import { specOfTarget } from "./benchmarks.js";
 import { extractQuotaObservations } from "./quota-observation.js";
 import { materializeDynamicPools } from "./dynamic-pools.js";
 import { baseLog } from "./request-log.js";
-import { looksLikeContextLengthError, parseStatedContextLimit, recordObservedContextLimit } from "./context-limits.js";
+import { looksLikeContextLengthError, looksLikeMaxOutputError, parseStatedContextLimit, parseStatedMaxOutput, recordObservedContextLimit, recordObservedMaxOutput, } from "./context-limits.js";
 import { looksLikeRateLimitError, parseStatedRateLimit, recordObservedRateLimit } from "./rate-limits.js";
 import { clearFacts, cooldownUntil, factsFor, isCostBlocked, recordFact } from "./target-facts.js";
 import { createQuotaDemotionFn, quotaDemotionLabel } from "./quota-demotion.js";
@@ -1138,12 +1138,7 @@ async function handle(req, res, cfg, h) {
         const controller = new AbortController();
         const callerController = new AbortController();
         const timer = setTimeout(() => controller.abort(), target.timeoutMs);
-        const onResClose = () => {
-            if (!res.writableEnded) {
-                callerController.abort();
-                controller.abort();
-            }
-        };
+        const onResClose = abortOnClientClose(res, callerController, controller);
         res.on("close", onResClose);
         let attempt;
         const usage = createUsageAccumulator();
@@ -2329,6 +2324,19 @@ function beginHealthAttempt(h, resolvedAttempt, started, trace, usage, accountin
         completed: false,
     };
 }
+/**
+ * The candidate loops' client-gone reaction, shared by both fronts: when the client connection
+ * closes before the response is finished, abort the caller-facing work and the in-flight backend
+ * fetch together. One definition, so the two loops cannot drift.
+ */
+function abortOnClientClose(res, callerController, controller) {
+    return () => {
+        if (!res.writableEnded) {
+            callerController.abort();
+            controller.abort();
+        }
+    };
+}
 /**
  * Learn a deployment's real context ceiling from an error it just returned.
  *
@@ -2360,6 +2368,37 @@ function observeContextLimit(status, target, body) {
         // Learning is best-effort and never in the request's way.
     }
 }
+/**
+ * Learn a deployment's stated OUTPUT-token ceiling (`max_tokens`) from an error it just returned.
+ *
+ * ⚠ Called from BOTH request paths beside `observeContextLimit`, for the same reason that helper
+ * documents: a learning loop wired into only one front would silently know nothing about half the
+ * traffic (docs/pool-failover.md). Called with bytes already buffered by
+ * `inspectCandidateResponse`, so it never tees or consumes the Response that may still become the
+ * client's terminal real error.
+ *
+ * Display-only by owner decision (docs/max-output-caps-design-2026-08-29.md): the fact renders in
+ * `llm-relay candidates`; nothing clamps, refuses, or routes on it.
+ */
+function observeMaxOutput(status, target, body) {
+    // A max_tokens rejection is a request-validation 400 (or a 413 variant); same gate as the
+    // context observer beside it.
+    if (status !== 400 && status !== 413)
+        return;
+    if (target.model === undefined)
+        return;
+    try {
+        if (!looksLikeMaxOutputError(body))
+            return;
+        const stated = parseStatedMaxOutput(body);
+        if (stated === null)
+            return;
+        recordObservedMaxOutput(target.provider, target.model, stated);
+    }
+    catch {
+        // Learning is best-effort and never in the request's way.
+    }
+}
 /**
  * Record the durable half of provider-stated quota observations as learned rate-limit facts.
  *
@@ -2625,6 +2664,7 @@ async function inspectCandidateResponse(res, attempt, retryAfterMs) {
     }
     const body = bytes.toString("utf8");
     observeContextLimit(status, attempt.target, body);
+    observeMaxOutput(status, attempt.target, body);
     observeRateLimit(attempt, status, body);
     const eligibility = carriesEligibilityFact(status) && body
         ? observeEligibility(attempt, status, retryAfterMs, body)
@@ -2833,12 +2873,7 @@ async function openAiFrontPath(res, credentialWalk, credentialTrace, ctx, h) {
         const controller = new AbortController();
         const callerController = new AbortController();
         const timer = setTimeout(() => controller.abort(), target.timeoutMs);
-        const onResClose = () => {
-            if (!res.writableEnded) {
-                callerController.abort();
-                controller.abort();
-            }
-        };
+        const onResClose = abortOnClientClose(res, callerController, controller);
         res.on("close", onResClose);
         let attempt;
         const usage = createUsageAccumulator();