Expose retry method and improve token budgeting

mcp/mcp_http_server.py

@@ -90,12 +90,12 @@ async def handle_mcp_request(request: Request) -> Dict[str, Any]:
     try:
         if method == "ping":
             return mcp_respond(request_id, result={"ok": True, "ts": time.time()})
-        if method == "flowdex.infer":
+        elif method == "flowdex.infer":
             return mcp_respond(
                 request_id,
                 result=call_api("POST", "/infer", json_body=params),
             )
-        if method == "flowdex.infer.retry":
+        elif method == "flowdex.infer.retry":
             run_id = params.get("run_id")
             if not run_id:
                 raise BridgeError("run_id is required for flowdex.infer.retry")
@@ -112,17 +112,18 @@ async def handle_mcp_request(request: Request) -> Dict[str, Any]:
                     json_body=retry_payload,
                 ),
             )
-        if method == "flowdex.memory.get":
+        elif method == "flowdex.memory.get":
             return mcp_respond(
                 request_id,
                 result=call_api("GET", "/memory/get", params=params),
             )
-        if method == "flowdex.health":
+        elif method == "flowdex.health":
             return mcp_respond(
                 request_id,
                 result=call_api("GET", "/health", params=params),
             )
-        raise BridgeError(f"Unknown method: {method}")
+        else:
+            raise BridgeError(f"Unknown method: {method}")
     except BridgeError as exc:
         return mcp_respond(request_id, error=str(exc))
     except Exception as exc:  # pragma: no cover - defensive programming

server/app.py

@@ -6,14 +6,23 @@
 import time
 from collections import Counter
 from pathlib import Path
-from typing import Any, Dict, List, Optional
+from typing import Any, Dict, List, Optional, Tuple
 
 import redis
 import requests
 from fastapi import Depends, FastAPI, Header, HTTPException
 from pydantic import BaseModel, Field
 from redis.exceptions import RedisError
 
+try:
+    import tiktoken
+
+    _TOKEN_ENCODER = tiktoken.get_encoding("cl100k_base")
+    _HAS_TIKTOKEN = True
+except ImportError:  # pragma: no cover - optional dependency
+    _TOKEN_ENCODER = None
+    _HAS_TIKTOKEN = False
+
 CACHE_DIR = Path(os.environ.get("FLOWDEX_CACHE_DIR", ".flowdex_cache"))
 CACHE_DIR.mkdir(parents=True, exist_ok=True)
 
@@ -130,10 +139,48 @@ def json_patch(old: str, new: str) -> Dict[str, Any]:
     return {"common_prefix": prefix_len, "added": new[prefix_len:]}
 
 
-def truncate_to_budget(text: str, budget_chars: int) -> str:
-    if len(text) <= budget_chars:
-        return text
-    return text[-budget_chars:]
+def count_tokens(text: str) -> int:
+    """Count tokens using ``tiktoken`` when available, otherwise estimate."""
+
+    if not text:
+        return 0
+    if _HAS_TIKTOKEN and _TOKEN_ENCODER is not None:
+        return len(_TOKEN_ENCODER.encode(text))
+    return max(1, len(text) // 4)
+
+
+def truncate_to_token_budget(text: str, budget_tokens: int) -> Tuple[str, int]:
+    """Truncate ``text`` to fit the supplied token budget.
+
+    Returns a tuple of the truncated text and the number of tokens consumed.
+    """
+
+    if not text or budget_tokens <= 0:
+        return "", 0
+
+    if not (_HAS_TIKTOKEN and _TOKEN_ENCODER is not None):
+        char_budget = max(budget_tokens * 4, 0)
+        truncated_text = text[-char_budget:] if char_budget else ""
+        return truncated_text, count_tokens(truncated_text)
+
+    tokens = _TOKEN_ENCODER.encode(text)
+    if len(tokens) <= budget_tokens:
+        return text, len(tokens)
+
+    truncated_tokens = tokens[-budget_tokens:]
+    try:
+        truncated_text = _TOKEN_ENCODER.decode(truncated_tokens)
+    except Exception:  # pragma: no cover - defensive decoding fallback
+        start_char = max(
+            0,
+            len(text)
+            - int(len(text) * (budget_tokens / max(len(tokens), 1)) * 1.1),
+        )
+        truncated_text = text[start_char:]
+        truncated_tokens = _TOKEN_ENCODER.encode(truncated_text)[-budget_tokens:]
+        truncated_text = _TOKEN_ENCODER.decode(truncated_tokens)
+
+    return truncated_text, len(truncated_tokens)
 
 
 def _safe_json_dumps(value: Any) -> str:
@@ -147,7 +194,7 @@ def _safe_json_dumps(value: Any) -> str:
         return repr(value)
 
 
-def build_tool_context(tool_specs: List[Dict[str, Any]], budget_chars: int) -> str:
+def build_tool_context(tool_specs: List[Dict[str, Any]], max_length: Optional[int] = None) -> str:
     if not tool_specs:
         return ""
     serialized = []
@@ -167,7 +214,9 @@ def build_tool_context(tool_specs: List[Dict[str, Any]], budget_chars: int) -> s
             )
         )
     blob = "\n\n".join(serialized)
-    return truncate_to_budget(blob, budget_chars)
+    if max_length is None or len(blob) <= max_length:
+        return blob
+    return blob[-max_length:]
 
 
 TOKEN_PATTERN = re.compile(r"\b\w+\b", re.UNICODE)
@@ -259,9 +308,14 @@ def call_llm(prompt: Dict[str, Any], request: InferRequest, tool_specs: List[Dic
     if prompt.get("user"):
         segments.append(f"<user_input>\n{prompt['user']}\n</user_input>")
 
-    tool_context = build_tool_context(tool_specs, request.budget.tools)
+    tool_context_blob = build_tool_context(tool_specs, 100_000)
+    tool_context, tool_tokens = truncate_to_token_budget(
+        tool_context_blob, request.budget.tools
+    )
     if tool_context:
         segments.append(f"<tools>\n{tool_context}\n</tools>")
+    if "tokens_used" in prompt:
+        prompt["tokens_used"]["tools"] = tool_tokens
 
     user_message = "\n\n".join(seg for seg in segments if seg)
     if not user_message.strip():
@@ -417,9 +471,9 @@ def run_inference(req: InferRequest) -> Dict[str, Any]:
     patch = json_patch(prior_text, context_blob)
     prior_cache_path.write_text(context_blob)
 
-    sys_b = truncate_to_budget(req.system_prompt, req.budget.system)
-    ctx_b = truncate_to_budget(context_blob, req.budget.context)
-    usr_b = truncate_to_budget(req.user_input, req.budget.user)
+    sys_b, sys_tokens = truncate_to_token_budget(req.system_prompt, req.budget.system)
+    ctx_b, ctx_tokens = truncate_to_token_budget(context_blob, req.budget.context)
+    usr_b, usr_tokens = truncate_to_token_budget(req.user_input, req.budget.user)
 
     tool_names: List[str] = []
     tool_specs: List[Dict[str, Any]] = []
@@ -445,6 +499,11 @@ def run_inference(req: InferRequest) -> Dict[str, Any]:
         "model": req.model,
         "max_tokens": req.max_tokens,
         "retrievals": retrieved_contexts,
+        "tokens_used": {
+            "system": sys_tokens,
+            "context": ctx_tokens,
+            "user": usr_tokens,
+        },
     }
 
     output = call_llm(prompt, req, tool_specs, retrieved_contexts)
@@ -463,6 +522,7 @@ def run_inference(req: InferRequest) -> Dict[str, Any]:
         "output": output,
         "budgets": req.budget.dict(),
         "used_context_chars": len(ctx_b),
+        "used_tokens": prompt["tokens_used"],
         "retrievals": retrieved_contexts,
         "tools_considered": tool_names,
     }
@@ -501,18 +561,19 @@ def infer_retry(
     except Exception as exc:
         raise HTTPException(400, detail=f"Failed to parse original request: {exc}") from exc
 
-    new_system_prompt = retry_req.system_prompt_override or original_req.system_prompt
-    new_user_input = retry_req.user_input_override or original_req.user_input
-
-    error_prefix = (
-        "--- Previous Attempt Failed ---\n"
-        "The previous attempt to solve this task failed with the following error:\n"
-        f"<error>\n{retry_req.error_context}\n</error>\n"
-        "Please analyze this error and provide a new solution.\n"
-        "--- Original System Prompt ---\n"
+    original_req.system_prompt = (
+        retry_req.system_prompt_override or original_req.system_prompt
     )
 
-    original_req.system_prompt = error_prefix + new_system_prompt
-    original_req.user_input = new_user_input
+    original_user_input = retry_req.user_input_override or original_req.user_input
+    error_context_block = (
+        "--- Context: Previous Attempt Failed ---\n"
+        "The previous attempt to solve the task failed. Analyze the following error "
+        "and provide a corrected solution based on the original request.\n"
+        f"<error_details>\n{retry_req.error_context}\n</error_details>\n"
+        "--- Original User Input ---\n"
+    )
+    original_req.user_input = error_context_block + original_user_input
 
+    print(f"Retrying run {run_id}. Added error context to user input.")
     return run_inference(original_req)

server/requirements.txt

@@ -4,3 +4,4 @@ pydantic
 python-dotenv
 redis
 requests
+tiktoken

server/tests/test_utils.py

@@ -10,8 +10,9 @@
     _safe_json_dumps,
     _tokenize,
     build_tool_context,
+    count_tokens,
     json_patch,
-    truncate_to_budget,
+    truncate_to_token_budget,
 )
 
 
@@ -27,9 +28,19 @@ def test_json_patch(old: str, new: str, expected: dict[str, object]) -> None:
     assert json_patch(old, new) == expected
 
 
-def test_truncate_to_budget() -> None:
-    assert truncate_to_budget("abcdef", 10) == "abcdef"
-    assert truncate_to_budget("abcdef", 3) == "def"
+def test_truncate_to_token_budget() -> None:
+    text = "abcdef" * 5
+    truncated, tokens = truncate_to_token_budget(text, 3)
+    assert truncated.endswith("def")
+    assert tokens <= 3
+
+    truncated_full, tokens_full = truncate_to_token_budget(text, 100)
+    assert truncated_full == text
+    assert tokens_full >= count_tokens(text)
+
+    empty_text, empty_tokens = truncate_to_token_budget("", 10)
+    assert empty_text == ""
+    assert empty_tokens == 0
 
 
 @pytest.mark.parametrize(
@@ -51,7 +62,7 @@ def test_build_tool_context_truncates_and_serializes() -> None:
         {"name": "beta", "description": "second", "cost": "low", "schema": complex_schema},
     ]
 
-    context = build_tool_context(tool_specs, budget_chars=500)
+    context = build_tool_context(tool_specs, max_length=500)
     assert "Tool: alpha" in context
     assert "Schema: {\"x\": \"int\"}" in context
     # Non-serializable value should fallback to repr