flatagents 2.4.4__tar.gz → 2.5.0__tar.gz

{flatagents-2.4.4 → flatagents-2.5.0}/.gitignore

@@ -15,4 +15,5 @@ data
 logs
 
 .aider.tags.cache.v4/
+.speed-kills/
 auth.json

{flatagents-2.4.4 → flatagents-2.5.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: flatagents
-Version: 2.4.4
+Version: 2.5.0
 Summary: A lightweight framework for building LLM-powered agents.
 Project-URL: Homepage, https://github.com/memgrafter/flatagents
 Project-URL: Repository, https://github.com/memgrafter/flatagents

{flatagents-2.4.4 → flatagents-2.5.0}/flatagents/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "2.4.4"
+__version__ = "2.5.0"
 
 from .baseagent import (
     # Base agent (abstract, for multi-step agents)

{flatagents-2.4.4 → flatagents-2.5.0}/flatagents/assets/flatagent.d.ts

@@ -151,7 +151,7 @@
  * The profile field specifies which profile name to use as base.
  */
 
-export const SPEC_VERSION = "2.4.4";
+export const SPEC_VERSION = "2.5.0";
 
 export interface AgentWrapper {
   spec: "flatagent";

{flatagents-2.4.4 → flatagents-2.5.0}/flatagents/assets/flatagent.slim.d.ts

@@ -1,4 +1,4 @@
-export const SPEC_VERSION = "2.4.4";
+export const SPEC_VERSION = "2.5.0";
 export interface AgentWrapper {
     spec: "flatagent";
     spec_version: string;

{flatagents-2.4.4 → flatagents-2.5.0}/flatagents/assets/flatagents-runtime.d.ts

@@ -800,7 +800,7 @@ export interface BackendConfig {
     aws_region?: string;
 }
 
-export const SPEC_VERSION = "2.4.4";
+export const SPEC_VERSION = "2.5.0";
 
 export interface SDKRuntimeWrapper {
     spec: "flatagents-runtime";

{flatagents-2.4.4 → flatagents-2.5.0}/flatagents/assets/flatagents-runtime.schema.json

@@ -11,7 +11,7 @@
         },
         "spec_version": {
           "type": "string",
-          "const": "2.4.4"
+          "const": "2.5.0"
         },
         "execution_lock": {
           "$ref": "#/definitions/ExecutionLock"

{flatagents-2.4.4 → flatagents-2.5.0}/flatagents/assets/flatagents-runtime.slim.d.ts

@@ -252,7 +252,7 @@ export interface BackendConfig {
     dynamodb_table?: string;
     aws_region?: string;
 }
-export const SPEC_VERSION = "2.4.4";
+export const SPEC_VERSION = "2.5.0";
 export interface SDKRuntimeWrapper {
     spec: "flatagents-runtime";
     spec_version: typeof SPEC_VERSION;

{flatagents-2.4.4 → flatagents-2.5.0}/flatagents/assets/flatmachine.d.ts

@@ -319,7 +319,7 @@
  * config_hash         - Content-addressed machine config key for cross-SDK resume (v2.1.0)
  */
 
-export const SPEC_VERSION = "2.4.4";
+export const SPEC_VERSION = "2.5.0";
 
 export interface MachineWrapper {
   spec: "flatmachine";

{flatagents-2.4.4 → flatagents-2.5.0}/flatagents/assets/flatmachine.slim.d.ts

@@ -1,4 +1,4 @@
-export const SPEC_VERSION = "2.4.4";
+export const SPEC_VERSION = "2.5.0";
 export interface MachineWrapper {
     spec: "flatmachine";
     spec_version: string;

{flatagents-2.4.4 → flatagents-2.5.0}/flatagents/assets/profiles.d.ts

@@ -110,7 +110,7 @@
  * oauth             - OAuth settings (used by codex backend)
  */
 
-export const SPEC_VERSION = "2.4.4";
+export const SPEC_VERSION = "2.5.0";
 
 export interface ProfilesWrapper {
   spec: "flatprofiles";

{flatagents-2.4.4 → flatagents-2.5.0}/flatagents/assets/profiles.slim.d.ts

@@ -1,4 +1,4 @@
-export const SPEC_VERSION = "2.4.4";
+export const SPEC_VERSION = "2.5.0";
 export interface ProfilesWrapper {
     spec: "flatprofiles";
     spec_version: string;

{flatagents-2.4.4 → flatagents-2.5.0}/flatagents/flatagent.py

@@ -741,6 +741,13 @@ class FlatAgent:
             _external_tools = None
             tools_prompt = self._render_tool_prompt(_mcp_tools)
 
+        # Session identity is transport metadata, not prompt input.
+        # Strip it from input_data so continuation calls (messages=..., no fresh
+        # user input) are not misclassified as having new user content.
+        codex_session_id = None
+        if self._backend == "codex" and "session_id" in input_data:
+            codex_session_id = input_data.pop("session_id")
+
         # Render prompts
         system_prompt = self._render_system_prompt(input_data, tools_prompt=tools_prompt, tools=_mcp_tools)
         user_prompt = self._render_user_prompt(input_data, tools_prompt=tools_prompt, tools=_mcp_tools)
@@ -815,10 +822,10 @@ class FlatAgent:
             if key not in _KNOWN_MODEL_FIELDS and key not in params and value is not None:
                 params[key] = value
 
-        # Forward session_id from input_data into params for the Codex backend so
-        # prompt_cache_key is set, enabling KV-cache hits across continuation turns.
-        if self._backend == "codex" and "session_id" in input_data and "session_id" not in params:
-            params["session_id"] = input_data["session_id"]
+        # Forward machine-provided session_id into Codex params so prompt_cache_key
+        # is set for this request.
+        if self._backend == "codex" and codex_session_id and "session_id" not in params:
+            params["session_id"] = codex_session_id
 
         # Add tools if available
         if _external_tools:
@@ -884,6 +891,22 @@ class FlatAgent:
                 finish_reason = self._extract_finish_reason(response)
                 if finish_reason:
                     monitor.metrics["finish_reason"] = finish_reason.value
+
+                # Debug session continuity for codex prompt caching.
+                if self._backend == "codex":
+                    request_meta = getattr(response, "_request_meta", {}) or {}
+                    request_headers = request_meta.get("headers") if isinstance(request_meta, dict) else {}
+                    if not isinstance(request_headers, dict):
+                        request_headers = {}
+                    session_header = request_headers.get("session_id")
+                    prompt_cache_key = request_meta.get("prompt_cache_key") if isinstance(request_meta, dict) else None
+                    logger.info(
+                        "Codex request debug: session_id=%s prompt_cache_key=%s finish_reason=%s cache_read_tokens=%s",
+                        session_header,
+                        prompt_cache_key,
+                        finish_reason.value if finish_reason else None,
+                        usage_info.cache_read_tokens if usage_info else 0,
+                    )
                 
                 # Extract rate limit headers from successful response
                 response_headers = extract_headers_from_response(response)

{flatagents-2.4.4 → flatagents-2.5.0}/flatagents/providers/openai_codex_client.py

@@ -147,6 +147,7 @@ class CodexClient:
                 "url": request_url,
                 "headers": self._redact_request_headers(headers),
                 "retries_used": retries_used,
+                "prompt_cache_key": session_id,
             }
             return result
         except CodexHTTPError as first_error:
@@ -184,6 +185,7 @@ class CodexClient:
                 "url": request_url,
                 "headers": self._redact_request_headers(retry_headers),
                 "retries_used": retries_used,
+                "prompt_cache_key": session_id,
             }
             return result
 
@@ -365,14 +367,32 @@ class CodexClient:
                 response_obj = event.get("response") if isinstance(event.get("response"), dict) else {}
                 result.status = str(response_obj.get("status")) if response_obj.get("status") else None
                 usage_obj = response_obj.get("usage") if isinstance(response_obj.get("usage"), dict) else {}
-                input_tokens = int(usage_obj.get("input_tokens") or 0)
-                output_tokens = int(usage_obj.get("output_tokens") or 0)
-                total_tokens = int(usage_obj.get("total_tokens") or (input_tokens + output_tokens))
-                cached_tokens = int(
+                current_input_tokens = int(usage_obj.get("input_tokens") or 0)
+                current_output_tokens = int(usage_obj.get("output_tokens") or 0)
+                current_total_tokens = int(
+                    usage_obj.get("total_tokens") or (current_input_tokens + current_output_tokens)
+                )
+                current_cached_tokens = int(
                     (usage_obj.get("input_tokens_details") or {}).get("cached_tokens")
                     if isinstance(usage_obj.get("input_tokens_details"), dict)
                     else 0
                 )
+
+                # Some payloads emit both response.completed and response.done with
+                # different usage fidelity. Preserve best-known values instead of
+                # letting a later sparse event overwrite cache metrics with zeros.
+                previous_usage = result.usage
+                if previous_usage is not None:
+                    input_tokens = max(previous_usage.input_tokens, current_input_tokens)
+                    output_tokens = max(previous_usage.output_tokens, current_output_tokens)
+                    total_tokens = max(previous_usage.total_tokens, current_total_tokens)
+                    cached_tokens = max(previous_usage.cached_tokens, current_cached_tokens)
+                else:
+                    input_tokens = current_input_tokens
+                    output_tokens = current_output_tokens
+                    total_tokens = current_total_tokens
+                    cached_tokens = current_cached_tokens
+
                 result.usage = CodexUsage(
                     input_tokens=input_tokens,
                     output_tokens=output_tokens,

{flatagents-2.4.4 → flatagents-2.5.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "flatagents"
-version = "2.4.4"
+version = "2.5.0"
 description = "A lightweight framework for building LLM-powered agents."
 readme = "README.md"
 license = "Apache-2.0"