fix: don't post non-actionable errors as GitHub review comments

src/kit/pr_review/agentic_reviewer.py

@@ -9,6 +9,7 @@
 from .base_reviewer import BaseReviewer
 from .config import LLMProvider, ReviewConfig
 from .diff_parser import DiffParser
+from .error_utils import is_non_actionable_error
 from .file_prioritizer import FilePrioritizer
 from .priority_filter import filter_review_by_priority
 
@@ -1094,10 +1095,14 @@ def review_pr_agentic(self, pr_input: str) -> str:
             # Generate final comment
             review_comment = self._generate_agentic_comment(pr_details, files, analysis)
 
-            # Post comment if configured to do so AND analysis completed successfully
+            # Post comment if configured to do so AND analysis completed successfully.
+            # Never post non-actionable infrastructure errors (token limits, 5xx, rate limits).
             if self.config.post_as_comment:
-                comment_result = self.post_pr_comment(owner, repo, pr_number, review_comment)
-                print(f"Posted comment: {comment_result['html_url']}")
+                if is_non_actionable_error(review_comment):
+                    print("Skipping GitHub comment: LLM provider returned a non-actionable error")
+                else:
+                    comment_result = self.post_pr_comment(owner, repo, pr_number, review_comment)
+                    print(f"Posted comment: {comment_result['html_url']}")
             else:
                 print("Comment posting disabled in configuration")
 

src/kit/pr_review/error_utils.py

@@ -0,0 +1,73 @@
+"""Utilities for classifying and handling non-actionable review errors."""
+
+# Prefixes emitted by _analyze_with_*_enhanced and agentic analysis methods
+# when the underlying LLM/provider call fails.
+_ERROR_PREFIXES = (
+    "error during enhanced llm analysis:",
+    "error during enhanced ollama analysis:",
+    "error during agentic analysis turn",
+)
+
+# Token / context-length limit patterns (provider-agnostic)
+_TOKEN_LIMIT_PATTERNS = (
+    "maximum context length",
+    "context_length_exceeded",
+    "context length exceeded",
+    "too many tokens",
+    "token limit",
+    "context window",
+    "prompt is too long",
+    "request too large",
+    "content_too_large",
+    "resource_exhausted",
+    "max_tokens",
+    "maximum number of tokens",
+    "input is too long",
+    "string_above_max_length",
+)
+
+# Upstream HTTP 5xx patterns
+_HTTP_5XX_PATTERNS = (
+    "500 internal server error",
+    "502 bad gateway",
+    "503 service unavailable",
+    "504 gateway timeout",
+    "overloaded_error",
+    "overloaded",
+    "internal server error",
+    "server_error",
+    "the server had an error",
+)
+
+# Rate-limit / 429 patterns
+_RATE_LIMIT_PATTERNS = (
+    "rate_limit",
+    "rate limit",
+    "429",
+    "too many requests",
+)
+
+
+def is_non_actionable_error(text: str) -> bool:
+    """Return True if *text* contains a non-actionable infrastructure error.
+
+    Non-actionable errors are transient provider/infrastructure failures that
+    a PR author cannot fix (token limits, upstream 5xx, rate limits).  These
+    should never be posted as GitHub review comments.
+
+    The check is intentionally conservative: it only fires when the text
+    contains one of the known error prefixes **and** a recognised
+    infrastructure-error pattern, so legitimate reviews that happen to
+    mention "500" or "token" are not affected.
+    """
+    if not text:
+        return False
+
+    lower = text.lower()
+
+    # Only consider text that was produced by an LLM analysis error path.
+    if not any(prefix in lower for prefix in _ERROR_PREFIXES):
+        return False
+
+    all_patterns = _TOKEN_LIMIT_PATTERNS + _HTTP_5XX_PATTERNS + _RATE_LIMIT_PATTERNS
+    return any(pattern in lower for pattern in all_patterns)

src/kit/pr_review/reviewer.py

@@ -13,6 +13,7 @@
 from .base_reviewer import BaseReviewer
 from .config import LLMProvider, ReviewConfig
 from .diff_parser import DiffParser, FileDiff
+from .error_utils import is_non_actionable_error
 from .file_prioritizer import FilePrioritizer
 from .priority_filter import filter_review_by_priority
 from .validator import validate_review_quality
@@ -478,11 +479,16 @@ def review_pr(self, pr_input: str) -> str:
                 basic_analysis = f"Quick analysis mode.\n\nFiles changed: {len(files)} files with {sum(f['additions'] for f in files)} additions and {sum(f['deletions'] for f in files)} deletions."
                 review_comment = self._generate_intelligent_comment(pr_details, files, basic_analysis)
 
-            # Post comment if configured to do so
+            # Post comment if configured to do so, but never post non-actionable
+            # infrastructure errors (token limits, upstream 5xx, rate limits).
             if self.config.post_as_comment:
-                comment_result = self.post_pr_comment(owner, repo, pr_number, review_comment)
-                if not quiet:
-                    print(f"Posted comment: {comment_result['html_url']}")
+                if is_non_actionable_error(review_comment):
+                    if not quiet:
+                        print("Skipping GitHub comment: LLM provider returned a non-actionable error")
+                else:
+                    comment_result = self.post_pr_comment(owner, repo, pr_number, review_comment)
+                    if not quiet:
+                        print(f"Posted comment: {comment_result['html_url']}")
 
             # Display cost summary
             if not quiet: