fix(compression): respect user-configured threshold below 0.50 on small-ctx models

agent/context_compressor.py

@@ -1181,9 +1181,20 @@ def _effective_threshold_percent(
         e.g. Codex gpt-5.5's 85%) always wins; only lower values are raised.
         Large-context models keep the configured value — at 512K+ the default
         50% trigger already leaves ample post-compaction headroom.
+
+        The floor is a safety net for the *default* threshold (0.50), not a
+        hard override of explicit user configuration.  When a user sets
+        ``threshold: 0.3`` to stay within TPM rate limits (e.g. 100K TPM on
+        a 160K-ctx reasoning model), forcing 75% prevents compression from
+        ever firing and causes cascading 429s.  Respect the user's value when
+        it is below the default — they chose it deliberately.
         """
         if context_length and context_length < _SMALL_CTX_WINDOW_LIMIT:
-            return max(threshold_percent, _SMALL_CTX_THRESHOLD_PERCENT)
+            # Only apply the floor when the configured value looks like the
+            # default (>= 0.50).  A value below 0.50 signals deliberate user
+            # configuration — raising it would defeat the purpose.
+            if threshold_percent >= 0.50:
+                return max(threshold_percent, _SMALL_CTX_THRESHOLD_PERCENT)
         return threshold_percent
 
     @staticmethod

agent/transports/codex.py

@@ -265,6 +265,12 @@ def build_kwargs(
         # the cache-scope routing headers below. Falls back to session_id when
         # there is no static content to hash.
         cache_key = _content_cache_key(instructions, response_tools) or session_id
+        # The backend rejects prompt_cache_key > 64 chars with HTTP 400.
+        # When _content_cache_key falls back to session_id the key can be
+        # arbitrarily long (e.g. "paperclip:company:<uuid>:agent:<uuid>" at
+        # 97 chars).  Hash over-length keys so they always fit.
+        if cache_key and len(cache_key) > 64:
+            cache_key = f"pck_{hashlib.sha256(cache_key.encode('utf-8', 'replace')).hexdigest()[:24]}"
         # xAI Responses takes prompt_cache_key in extra_body (set further
         # down); GitHub Models opts out of cache-key routing entirely.
         if not is_github_responses and not is_xai_responses and cache_key: