fix(models): update context limits and error types for paid plans

.gitignore

@@ -4,10 +4,19 @@
 credentials.json
 state.json
 
+# Local Docker Compose overrides
+docker-compose.yaml
+docker-compose.local.yml
+docker-compose.local.yaml
+compose.local.yml
+compose.local.yaml
+
 # IDE
 .vscode/
 .idea/
 .shard/
+.agents/
+skills-lock.json
 
 # Python
 __pycache__/
@@ -30,4 +39,4 @@ requests/
 # Testing
 .pytest_cache/
 .coverage
-htmlcov/
+htmlcov/

kiro/config.py

@@ -273,20 +273,24 @@ def _get_raw_env_value(var_name: str, env_file: str = ".env") -> Optional[str]:
 # - Some models may not be available on your Kiro plan (e.g., Opus on free tier)
 # - New models released after this version won't appear here
 # - Update gateway regularly to get the latest model list
-FALLBACK_MODELS: List[Dict[str, str]] = [
-    {"modelId": "auto"},
-    {"modelId": "claude-sonnet-4"},
-    {"modelId": "claude-sonnet-4.5"},
-    {"modelId": "claude-sonnet-4.6"},
-    {"modelId": "claude-haiku-4.5"},
-    {"modelId": "claude-opus-4.5"},
-    {"modelId": "claude-opus-4.6"},
-    {"modelId": "claude-opus-4.7"},
-    {"modelId": "deepseek-3.2"},
-    {"modelId": "glm-5"},
-    {"modelId": "minimax-m2.1"},
-    {"modelId": "minimax-m2.5"},
-    {"modelId": "qwen3-coder-next"},
+FALLBACK_MODELS: List[Dict] = [
+    # 1M context: auto router and new Claude 4.x flagship models (Kiro docs)
+    {"modelId": "auto",              "tokenLimits": {"maxInputTokens": 1000000}},
+    {"modelId": "claude-sonnet-4.6", "tokenLimits": {"maxInputTokens": 1000000}},
+    {"modelId": "claude-opus-4.6",   "tokenLimits": {"maxInputTokens": 1000000}},
+    {"modelId": "claude-opus-4.7",   "tokenLimits": {"maxInputTokens": 1000000}},
+    {"modelId": "claude-opus-4.8",   "tokenLimits": {"maxInputTokens": 1000000}},
+    # 200K context: older Claude 4.x models and Haiku (Kiro docs)
+    {"modelId": "claude-sonnet-4",   "tokenLimits": {"maxInputTokens": 200000}},
+    {"modelId": "claude-sonnet-4.5", "tokenLimits": {"maxInputTokens": 200000}},
+    {"modelId": "claude-haiku-4.5",  "tokenLimits": {"maxInputTokens": 200000}},
+    {"modelId": "claude-opus-4.5",   "tokenLimits": {"maxInputTokens": 200000}},
+    # Non-Claude models
+    {"modelId": "deepseek-3.2",      "tokenLimits": {"maxInputTokens": 128000}},
+    {"modelId": "qwen3-coder-next",  "tokenLimits": {"maxInputTokens": 256000}},
+    {"modelId": "glm-5",             "tokenLimits": {"maxInputTokens": 200000}},
+    {"modelId": "minimax-m2.1",      "tokenLimits": {"maxInputTokens": 200000}},
+    {"modelId": "minimax-m2.5",      "tokenLimits": {"maxInputTokens": 200000}},
 ]
 
 # ==================================================================================================
@@ -297,7 +301,9 @@ def _get_raw_env_value(var_name: str, env_file: str = ".env") -> Optional[str]:
 MODEL_CACHE_TTL: int = 3600
 
 # Default maximum number of input tokens
-DEFAULT_MAX_INPUT_TOKENS: int = 200000
+# Set to 1M to match the highest-tier models (sonnet-4-6, opus-4-6/4-7/4-8) on paid plans.
+# Per-model overrides in FALLBACK_MODELS take precedence over this value.
+DEFAULT_MAX_INPUT_TOKENS: int = 1000000
 
 # ==================================================================================================
 # Tool Description Handling (Kiro API Limitations)
@@ -342,6 +348,19 @@ def _get_raw_env_value(var_name: str, env_file: str = ".env") -> Optional[str]:
 # Set to DEBUG for detailed troubleshooting
 LOG_LEVEL: str = os.getenv("LOG_LEVEL", "INFO").upper()
 
+# ==================================================================================================
+# Billing Attribution Strip
+# ==================================================================================================
+
+# Claude Code 2.1.x prepends a per-request system text block of the form
+# ``x-anthropic-billing-header: cc_version=...; cc_entrypoint=...; cch=<5hex>;``
+# where ``cch`` is a fresh random hex token. The Kiro gateway concatenates all
+# system text blocks before forwarding, so this random prefix would invalidate
+# any upstream prompt cache keyed on the prompt prefix. When enabled (default),
+# the gateway strips the leading attribution line / block before forwarding.
+# Disable only for A/B comparison or to debug attribution behavior.
+STRIP_BILLING_HEADER: bool = os.getenv("STRIP_BILLING_HEADER", "true").lower() in ("true", "1", "yes")
+
 # ==================================================================================================
 # First Token Timeout Settings (Streaming Retry)
 # ==================================================================================================
@@ -478,6 +497,26 @@ def _warn_timeout_configuration():
 FAKE_REASONING_INITIAL_BUFFER_SIZE: int = int(os.getenv("FAKE_REASONING_INITIAL_BUFFER_SIZE", "20"))
 
 
+# ==================================================================================================
+# Native Thinking Settings (Kiro/Claude Adaptive Thinking)
+# ==================================================================================================
+
+# Experimental native thinking pass-through for Kiro models that expose Claude adaptive thinking.
+#
+# Modes:
+# - "off": Disabled (default, preserves existing fake reasoning behavior)
+# - "auto": Enable only when the client explicitly requests a reasoning effort
+# - "force": Enable for supported models even when the client does not send an effort
+KIRO_NATIVE_THINKING_MODE: str = os.getenv("KIRO_NATIVE_THINKING_MODE", "off").lower()
+if KIRO_NATIVE_THINKING_MODE not in ("off", "auto", "force"):
+    KIRO_NATIVE_THINKING_MODE = "off"
+
+# Claude Opus 4.8/4.7 default to omitted thinking text unless display is explicitly summarized.
+KIRO_NATIVE_THINKING_DISPLAY: str = os.getenv("KIRO_NATIVE_THINKING_DISPLAY", "summarized").lower()
+if KIRO_NATIVE_THINKING_DISPLAY not in ("summarized", "omitted"):
+    KIRO_NATIVE_THINKING_DISPLAY = "summarized"
+
+
 # ==================================================================================================
 # Payload Size Guard Settings
 # ==================================================================================================

kiro/converters_anthropic.py

@@ -24,11 +24,12 @@
 to the unified format used by converters_core.py.
 """
 
+import re
 from typing import Any, Dict, List, Optional
 
 from loguru import logger
 
-from kiro.config import HIDDEN_MODELS
+from kiro.config import HIDDEN_MODELS, STRIP_BILLING_HEADER
 from kiro.model_resolver import get_model_id_for_kiro
 from kiro.models_anthropic import (
     AnthropicMessagesRequest,
@@ -39,6 +40,8 @@
     UnifiedMessage,
     UnifiedTool,
     ThinkingConfig,
+    build_native_thinking_config,
+    reasoning_effort_to_budget,
     build_kiro_payload,
     extract_text_content,
     extract_images_from_content,
@@ -75,6 +78,32 @@ def convert_anthropic_content_to_text(content: Any) -> str:
     return str(content) if content else ""
 
 
+_BILLING_HEADER_LINE_PATTERN = re.compile(
+    r"^x-anthropic-billing-header:[^\n]*\n?", re.IGNORECASE
+)
+
+
+def _strip_billing_attribution(text: str) -> str:
+    """
+    Remove Claude Code's per-request billing attribution string.
+
+    Claude Code (2.1.x) prepends a line like
+    ``x-anthropic-billing-header: cc_version=...; cc_entrypoint=...; cch=<5hex>;``
+    to the system prompt content. The ``cch`` segment is a fresh random hex per
+    request, which defeats any prompt cache keyed on the prompt prefix on
+    upstreams that do not understand the attribution header.
+
+    The line carries no semantic content for the model and is safe to remove
+    before forwarding to Kiro.
+    """
+    if not text or not STRIP_BILLING_HEADER:
+        return text
+    stripped = _BILLING_HEADER_LINE_PATTERN.sub("", text, count=1)
+    if stripped != text:
+        return stripped.lstrip("\n")
+    return text
+
+
 def extract_system_prompt(system: Any) -> str:
     """
     Extracts system prompt text from Anthropic system field.
@@ -86,6 +115,12 @@ def extract_system_prompt(system: Any) -> str:
     The second format is used for prompt caching with cache_control.
     We extract only the text, ignoring cache_control (not supported by Kiro).
 
+    Claude Code injects a per-request billing attribution block as the first
+    system text block (``x-anthropic-billing-header: ...; cch=<5hex>;``). The
+    random ``cch`` segment changes on every request and would invalidate any
+    upstream prompt cache keyed on the prompt prefix, so we drop attribution
+    blocks (and any attribution prefix on string-form prompts) here.
+
     Args:
         system: System prompt in string or list format
 
@@ -96,18 +131,25 @@ def extract_system_prompt(system: Any) -> str:
         return ""
 
     if isinstance(system, str):
-        return system
+        return _strip_billing_attribution(system)
 
     if isinstance(system, list):
         text_parts = []
         for block in system:
+            text: Optional[str] = None
             if isinstance(block, dict):
-                # Handle {"type": "text", "text": "...", "cache_control": {...}}
                 if block.get("type") == "text":
-                    text_parts.append(block.get("text", ""))
+                    text = block.get("text", "")
             elif hasattr(block, "type") and block.type == "text":
-                # Handle Pydantic model
-                text_parts.append(getattr(block, "text", ""))
+                text = getattr(block, "text", "")
+            if text is None:
+                continue
+            # Drop pure-attribution blocks; strip leading attribution line
+            # from mixed blocks (defensive).
+            stripped = _strip_billing_attribution(text)
+            if not stripped.strip():
+                continue
+            text_parts.append(stripped)
         return "\n".join(text_parts)
 
     return str(system)
@@ -376,6 +418,7 @@ def extract_thinking_config_from_anthropic(request: AnthropicMessagesRequest) ->
 
     Handles thinking parameter:
     - {"type": "enabled", "budget_tokens": N} → enabled with budget
+    - {"type": "adaptive", "effort": "max"} → enabled with effort-based budget
     - {"type": "disabled"} → disabled
     - None → enabled with default budget
 
@@ -400,6 +443,11 @@ def extract_thinking_config_from_anthropic(request: AnthropicMessagesRequest) ->
         >>> request.thinking = {"type": "enabled", "budget_tokens": 8000}
         >>> extract_thinking_config_from_anthropic(request)
         ThinkingConfig(enabled=True, budget_tokens=8000)
+
+        >>> # Adaptive effort translated to gateway fake thinking budget
+        >>> request.thinking = {"type": "adaptive", "effort": "max"}
+        >>> extract_thinking_config_from_anthropic(request)
+        ThinkingConfig(enabled=True, budget_tokens=4096)
     """
     if not request.thinking:
         # No thinking specified → use defaults
@@ -422,6 +470,31 @@ def extract_thinking_config_from_anthropic(request: AnthropicMessagesRequest) ->
             logger.debug(f"Extracted thinking config from Anthropic: type='enabled', budget={budget}")
         return ThinkingConfig(enabled=True, budget_tokens=budget)
 
+    if thinking_type == "adaptive":
+        effort = request.thinking.get("effort")
+        if not effort:
+            logger.debug("Extracted adaptive thinking config from Anthropic without effort")
+            return ThinkingConfig(enabled=True, budget_tokens=None)
+
+        if effort == "none":
+            logger.debug("Extracted adaptive thinking config from Anthropic: effort='none'")
+            return ThinkingConfig(enabled=False, budget_tokens=None)
+
+        try:
+            budget = reasoning_effort_to_budget(request.max_tokens, effort)
+        except ValueError:
+            logger.warning(
+                f"Unsupported Anthropic adaptive thinking effort '{effort}'. "
+                "Using default fake thinking budget."
+            )
+            return ThinkingConfig(enabled=True, budget_tokens=None)
+
+        logger.debug(
+            f"Extracted adaptive thinking config from Anthropic: effort='{effort}', "
+            f"max_tokens={request.max_tokens}, budget={budget}"
+        )
+        return ThinkingConfig(enabled=True, budget_tokens=budget)
+
     # Unknown type → use defaults
     return ThinkingConfig(enabled=True, budget_tokens=None)
 
@@ -466,12 +539,24 @@ def anthropic_to_kiro(
 
     # Extract thinking configuration from thinking parameter
     thinking_config = extract_thinking_config_from_anthropic(request)
+    native_effort: Optional[str] = None
+    native_display: Optional[str] = None
+    if isinstance(request.thinking, dict) and request.thinking.get("type") == "adaptive":
+        native_effort = request.thinking.get("effort") or "high"
+        native_display = request.thinking.get("display")
+    native_thinking_config = build_native_thinking_config(model_id, native_effort)
+    if native_display in ("summarized", "omitted"):
+        native_thinking_config.display = native_display
+    if native_thinking_config.enabled:
+        # Native adaptive thinking supersedes fake tag injection for this request.
+        thinking_config = ThinkingConfig(enabled=False, budget_tokens=None)
 
     logger.debug(
         f"Converting Anthropic request: model={request.model} -> {model_id}, "
         f"messages={len(unified_messages)}, tools={len(unified_tools) if unified_tools else 0}, "
         f"system_prompt_length={len(system_prompt)}, "
-        f"thinking_enabled={thinking_config.enabled}, thinking_budget={thinking_config.budget_tokens}"
+        f"thinking_enabled={thinking_config.enabled}, thinking_budget={thinking_config.budget_tokens}, "
+        f"native_thinking_enabled={native_thinking_config.enabled}, native_effort={native_thinking_config.effort}"
     )
 
     # Use core function to build payload
@@ -483,6 +568,7 @@ def anthropic_to_kiro(
         conversation_id=conversation_id,
         profile_arn=profile_arn,
         thinking_config=thinking_config,
+        native_thinking_config=native_thinking_config,
     )
 
     return result.payload