From 46ce1a03a902af5b973c9d75921e794b174680ac Mon Sep 17 00:00:00 2001
From: Maciej Szeszko <mszeszko@meta.com>
Date: Mon, 11 May 2026 17:45:43 -0700
Subject: [PATCH] Cap Claude thinking budget (#14731)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Summary:
### Motivation
Claude's auto review workflow classifies the PR as complex and sets `MAX_THINKING_TOKENS`=**32000** in `.github/workflows/ai-review-analysis.yml:534`. The Claude action then sends a request where `thinking.budget_tokens` is **32000**, but the request `max_tokens` is not greater than that, so Anthropic rejects it before the review starts ([example](https://github.com/facebook/rocksdb/actions/runs/25691112276/job/75427435133)).

### Fix
Reduce the "complex" thinking budget from **32000** to **24000** tokens and clamps manual overrides to the same ceiling, preventing the thinking budget from exceeding or equalling the action's effective `max_tokens`. 24k is still generous — enough for deep reasoning on complex PRs while guaranteeing the model can emit the formatted review.

Pull Request resolved: https://github.com/facebook/rocksdb/pull/14731

Reviewed By: xingbowang

Differential Revision: D104749426

Pulled By: mszeszko-meta

fbshipit-source-id: bf94d0f50e7c2c9bc9e4d4cdffd61087d739018a
---
 .github/workflows/ai-review-analysis.yml | 22 ++++++++++++++++++----
 .github/workflows/claude-review.yml      |  2 +-
 2 files changed, 19 insertions(+), 5 deletions(-)

diff --git a/.github/workflows/ai-review-analysis.yml b/.github/workflows/ai-review-analysis.yml
index 7cad645926..0e0250ca1a 100644
--- a/.github/workflows/ai-review-analysis.yml
+++ b/.github/workflows/ai-review-analysis.yml
@@ -531,7 +531,9 @@ jobs:
               core.warning(
                 `Could not parse classifier output: ${error.message}`);
             }
-            const budget = label === 'simple' ? '16000' : '32000';
+            // Keep Claude's thinking budget below the action's effective
+            // max_tokens so the API has room for the final review text.
+            const budget = label === 'simple' ? '16000' : '24000';
             core.info(
               `Complexity classified as: ${label} ` +
               `-> MAX_THINKING_TOKENS=${budget}`
@@ -1115,9 +1117,19 @@ jobs:
             }
             const override = (process.env.OVERRIDE || '').trim();
             if (override && /^\d+$/.test(override)) {
-              core.info(`Using thinking_budget override: ${override}`);
+              const maxSafeBudget = 24000;
+              const requested = parseInt(override, 10);
+              const budget = Math.min(requested, maxSafeBudget).toString();
+              if (budget !== override) {
+                core.info(
+                  `Clamping thinking_budget override from ${override} ` +
+                  `to ${budget} so max_tokens can exceed thinking tokens.`
+                );
+              } else {
+                core.info(`Using thinking_budget override: ${budget}`);
+              }
               core.setOutput('label', 'override');
-              core.setOutput('tokens', override);
+              core.setOutput('tokens', budget);
               return;
             }
             let label = 'complex';
@@ -1135,7 +1147,9 @@ jobs:
               core.warning(
                 `Could not parse classifier output: ${error.message}`);
             }
-            const budget = label === 'simple' ? '16000' : '32000';
+            // Keep Claude's thinking budget below the action's effective
+            // max_tokens so the API has room for the final review text.
+            const budget = label === 'simple' ? '16000' : '24000';
             core.info(
               `Complexity classified as: ${label} ` +
               `-> MAX_THINKING_TOKENS=${budget}`
diff --git a/.github/workflows/claude-review.yml b/.github/workflows/claude-review.yml
index 7087330d9f..85acf93d4f 100644
--- a/.github/workflows/claude-review.yml
+++ b/.github/workflows/claude-review.yml
@@ -34,7 +34,7 @@ on:
           - claude-sonnet-4-6
         default: claude-opus-4-6
       thinking_budget:
-        description: Override MAX_THINKING_TOKENS (blank = auto-classify)
+        description: Override MAX_THINKING_TOKENS (blank = auto-classify, capped at 24000)
         required: false
         type: string
         default: ""