From faa1c6f9728f3f8aef38f82241861199db7d4ac9 Mon Sep 17 00:00:00 2001
From: Peter Steinberger <steipete@gmail.com>
Date: Sun, 10 May 2026 09:46:22 +0100
Subject: [PATCH] fix: preserve custom provider context limits (#79911)

---
 CHANGELOG.md                               |  1 +
 src/commands/onboard-custom-config.test.ts |  7 ++++++-
 src/commands/onboard-custom-config.ts      | 17 +++++++++--------
 3 files changed, 16 insertions(+), 9 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index eb4cb77f3db..73938ac91b9 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -21,6 +21,7 @@ Docs: https://docs.openclaw.ai
 
 ### Fixes
 
+- CLI/onboarding: give non-Azure custom providers a safe generated context window and heal legacy 4k wizard entries without overwriting explicit valid small model limits, preventing first-turn compaction loops. Fixes #79428. (#79911) Thanks @Jefsky.
 - Ollama: stop native `/api/chat` requests from copying catalog `contextWindow` or `maxTokens` into `options.num_ctx` unless `params.num_ctx` is explicitly configured, avoiding pathological prompt-ingestion latency on local large-context models. Fixes #62267. Thanks @BenSHPD.
 - Ollama: keep the model idle watchdog enabled for `*:cloud` models routed through a local Ollama host, so cloud-backed tool-loop stalls fail over visibly instead of inheriting local-model no-idle behavior. Fixes #79350. Thanks @geek111.
 - Voice/Ollama: honor routed voice agent `tools.allow` for classic embedded voice responses, including empty allowlists, so no-tool Ollama agents do not receive tool schemas. Fixes #79506. Thanks @donkeykong91.
diff --git a/src/commands/onboard-custom-config.test.ts b/src/commands/onboard-custom-config.test.ts
index 576e9e7943d..e990ba9a102 100644
--- a/src/commands/onboard-custom-config.test.ts
+++ b/src/commands/onboard-custom-config.test.ts
@@ -117,10 +117,15 @@ describe("applyCustomApiConfig", () => {
       expectedContextWindow: CUSTOM_PROVIDER_DEFAULT_CONTEXT_WINDOW_TOKENS,
     },
     {
-      name: "raises context window below stable default compaction floor (#79428)",
+      name: "raises legacy generated hard-min context window (#79428)",
       existingContextWindow: CONTEXT_WINDOW_HARD_MIN_TOKENS,
       expectedContextWindow: CUSTOM_PROVIDER_DEFAULT_CONTEXT_WINDOW_TOKENS,
     },
+    {
+      name: "preserves explicit small context window when already valid",
+      existingContextWindow: 8192,
+      expectedContextWindow: 8192,
+    },
     {
       name: "preserves existing custom model context window when already above minimum",
       existingContextWindow: 131072,
diff --git a/src/commands/onboard-custom-config.ts b/src/commands/onboard-custom-config.ts
index ff800b8f7bd..f1255113dd9 100644
--- a/src/commands/onboard-custom-config.ts
+++ b/src/commands/onboard-custom-config.ts
@@ -15,10 +15,10 @@ import { normalizeAlias } from "./models/alias-name.js";
 
 /**
  * Wizard default for non-Azure custom APIs when context length is unknown.
- * Must exceed the compaction `reserveTokensFloor` default (20000 tokens) in
- * `agent-runner-memory.ts`, or the first turns enter an infinite compact loop (#79428).
+ * Mirrors the generic persisted custom-model catalog fallback and leaves enough
+ * room above the default compaction reserve floor in `pi-settings.ts`.
  */
-export const CUSTOM_PROVIDER_DEFAULT_CONTEXT_WINDOW_TOKENS = 32_768;
+export const CUSTOM_PROVIDER_DEFAULT_CONTEXT_WINDOW_TOKENS = 128_000;
 const DEFAULT_CONTEXT_WINDOW = CUSTOM_PROVIDER_DEFAULT_CONTEXT_WINDOW_TOKENS;
 const DEFAULT_MAX_TOKENS = 4096;
 // Azure OpenAI uses the Responses API which supports larger defaults
@@ -32,11 +32,12 @@ export type CustomModelImageInputInference = {
 
 function normalizeContextWindowForCustomModel(value: unknown): number {
   const parsed = typeof value === "number" && Number.isFinite(value) ? Math.floor(value) : 0;
-  const atLeastHardMin =
-    parsed >= CONTEXT_WINDOW_HARD_MIN_TOKENS
-      ? parsed
-      : CUSTOM_PROVIDER_DEFAULT_CONTEXT_WINDOW_TOKENS;
-  return Math.max(atLeastHardMin, CUSTOM_PROVIDER_DEFAULT_CONTEXT_WINDOW_TOKENS);
+  if (parsed <= 0 || parsed === CONTEXT_WINDOW_HARD_MIN_TOKENS) {
+    return CUSTOM_PROVIDER_DEFAULT_CONTEXT_WINDOW_TOKENS;
+  }
+  return parsed >= CONTEXT_WINDOW_HARD_MIN_TOKENS
+    ? parsed
+    : CUSTOM_PROVIDER_DEFAULT_CONTEXT_WINDOW_TOKENS;
 }
 
 function customModelInputs(supportsImageInput: boolean): CustomModelInput[] {