Restore legacy prefix matching for context window patterns

Commit 3694c44 switched CONTEXT_WINDOWS matching from substring to
fnmatch, which broke suffixed model names (deepseek-v4-flash,
gpt-5-pro, qwen3.5-32b, kimi-k2.7-0613 ...) by falling through to the
wildcard fallback (128k).  Restore legacy behavior by adding a
trailing '*' to explicit entries so first-match-wins still hits the
specific size before general fallbacks.

Author: beacoder

python_agent_harness/config.py

@@ -17,22 +17,25 @@
 CONTEXT_TRIGGER = 0.70
 
 # Entries are matched in order (first match wins): put more specific
-# patterns before general ones. Supports wildcards (*).
+# patterns before general ones. Supports wildcards (*).  Trailing `*`
+# on explicit entries preserves prefix matching for suffixed model
+# names (e.g. "deepseek-v4-flash" -> deepseek-v4*), matching the
+# legacy substring behavior.
 CONTEXT_WINDOWS: list[tuple[str, int]] = [
-    ("gpt-5-mini", 128_000),
-    ("gpt-5", 400_000),
-    ("gpt-oss-120b", 128_000),
-    ("claude", 200_000),
-    ("deepseek-v3", 128_000),
-    ("deepseek-v4", 1_000_000),
-    ("qwen3.5", 131_072),
-    ("qwen3.6", 262_144),
-    ("qwen3.8", 262_144),
-    ("qwen3", 131_072),
-    ("glm-5.2", 1_000_000),
-    ("glm-5.1", 128_000),
-    ("kimi-k2.7", 256_000),
-    ("kimi", 128_000),
+    ("gpt-5-mini*", 128_000),
+    ("gpt-5*", 400_000),
+    ("gpt-oss-120b*", 128_000),
+    ("claude*", 200_000),
+    ("deepseek-v3*", 128_000),
+    ("deepseek-v4*", 1_000_000),
+    ("qwen3.5*", 131_072),
+    ("qwen3.6*", 262_144),
+    ("qwen3.8*", 262_144),
+    ("qwen3*", 131_072),
+    ("glm-5.2*", 1_000_000),
+    ("glm-5.1*", 128_000),
+    ("kimi-k2.7*", 256_000),
+    ("kimi*", 128_000),
     # Wildcard fallbacks for unknown models
     ("gpt-*", 128_000),
     ("claude-*", 200_000),

tests/test_token_estimator.py

@@ -26,8 +26,13 @@ def test_cjk_estimate(self):
 
     def test_context_window_matching(self):
         self.assertEqual(context_window_for("deepseek-v4"), 1_000_000)
+        self.assertEqual(context_window_for("deepseek-v4-flash"), 1_000_000)
+        self.assertEqual(context_window_for("deepseek-v4-pro"), 1_000_000)
         self.assertEqual(context_window_for("gpt-5-mini"), 128_000)
+        self.assertEqual(context_window_for("gpt-5-pro"), 400_000)
         self.assertEqual(context_window_for("claude-sonnet"), 200_000)
+        self.assertEqual(context_window_for("qwen3.5-32b"), 131_072)
+        self.assertEqual(context_window_for("kimi-k2.7-0613"), 256_000)
         self.assertEqual(context_window_for("unknown-model"), 128_000)
 
     def test_calibrator(self):