progress

ai_client.py

@@ -14,7 +14,7 @@ _anthropic_client = None
 _anthropic_history: list[dict] = []
 
 # Injected by gui.py - called when AI wants to run a command.
-# Signature: (script: str) -> str | None
+# Signature: (script: str, base_dir: str) -> str | None
 # Returns the output string if approved, None if rejected.
 confirm_and_run_callback = None
 
@@ -22,24 +22,26 @@ confirm_and_run_callback = None
 # Signature: (entry: dict) -> None
 comms_log_callback = None
 
+# Injected by gui.py - called whenever a tool call completes (after run).
+# Signature: (script: str, result: str) -> None
+tool_log_callback = None
+
 MAX_TOOL_ROUNDS = 5
 
+# Anthropic system prompt - sent with cache_control so it is cached after the
+# first request and reused on every subsequent call within the TTL window.
+_ANTHROPIC_SYSTEM = (
+    "You are a helpful coding assistant with access to a PowerShell tool. "
+    "When asked to create or edit files, prefer targeted edits over full rewrites. "
+    "Always explain what you are doing before invoking the tool."
+)
+
 # ------------------------------------------------------------------ comms log
 
 _comms_log: list[dict] = []
 
 MAX_FIELD_CHARS = 400   # beyond this we show a truncated preview in the UI
 
-def _clamp(value, max_chars: int = MAX_FIELD_CHARS) -> tuple[str, bool]:
-    """Return (display_str, was_truncated)."""
-    if isinstance(value, (dict, list)):
-        s = json.dumps(value, ensure_ascii=False, indent=2)
-    else:
-        s = str(value)
-    if len(s) > max_chars:
-        return s[:max_chars], True
-    return s, False
-
 
 def _append_comms(direction: str, kind: str, payload: dict):
     """
@@ -78,15 +80,6 @@ class ProviderError(Exception):
     """
     Raised when the upstream API returns a hard error we want to surface
     distinctly in the UI (quota, rate-limit, auth, balance, etc.).
-
-    Attributes
-    ----------
-    kind : str
-        One of: "quota", "rate_limit", "auth", "balance", "network", "unknown"
-    provider : str
-        "gemini" or "anthropic"
-    original : Exception
-        The underlying SDK exception.
     """
     def __init__(self, kind: str, provider: str, original: Exception):
         self.kind = kind
@@ -94,7 +87,6 @@ class ProviderError(Exception):
         self.original = original
         super().__init__(str(original))
 
-    # Human-readable banner shown in the Response panel
     def ui_message(self) -> str:
         labels = {
             "quota":      "QUOTA EXHAUSTED",
@@ -109,7 +101,6 @@ class ProviderError(Exception):
 
 
 def _classify_anthropic_error(exc: Exception) -> ProviderError:
-    """Map an anthropic SDK exception to a ProviderError."""
     try:
         import anthropic
         if isinstance(exc, anthropic.RateLimitError):
@@ -129,7 +120,6 @@ def _classify_anthropic_error(exc: Exception) -> ProviderError:
                 return ProviderError("auth", "anthropic", exc)
             if status == 402:
                 return ProviderError("balance", "anthropic", exc)
-            # Anthropic puts credit-balance errors in the body at 400
             if "credit" in body or "balance" in body or "billing" in body:
                 return ProviderError("balance", "anthropic", exc)
             if "quota" in body or "limit" in body or "exceeded" in body:
@@ -140,10 +130,7 @@ def _classify_anthropic_error(exc: Exception) -> ProviderError:
 
 
 def _classify_gemini_error(exc: Exception) -> ProviderError:
-    """Map a google-genai SDK exception to a ProviderError."""
     body = str(exc).lower()
-    # google-genai surfaces HTTP errors as google.api_core exceptions or
-    # google.genai exceptions; inspect the message text as a reliable fallback.
     try:
         from google.api_core import exceptions as gac
         if isinstance(exc, gac.ResourceExhausted):
@@ -156,7 +143,6 @@ def _classify_gemini_error(exc: Exception) -> ProviderError:
             return ProviderError("network", "gemini", exc)
     except ImportError:
         pass
-    # Fallback: parse status code / message string
     if "429" in body or "quota" in body or "resource exhausted" in body:
         return ProviderError("quota", "gemini", exc)
     if "rate" in body and "limit" in body:
@@ -226,6 +212,9 @@ def _list_anthropic_models() -> list[str]:
 
 TOOL_NAME = "run_powershell"
 
+# The tool list for Anthropic.  cache_control is placed on the last (only) tool
+# so that the system-prompt + tools prefix is cached together after the first
+# request and served from cache on every subsequent round.
 _ANTHROPIC_TOOLS = [
     {
         "name": TOOL_NAME,
@@ -245,7 +234,8 @@ _ANTHROPIC_TOOLS = [
                 }
             },
             "required": ["script"]
-        }
+        },
+        "cache_control": {"type": "ephemeral"},
     }
 ]
 
@@ -279,13 +269,18 @@ def _run_script(script: str, base_dir: str) -> str:
     """
     Delegate to the GUI confirmation callback.
     Returns result string (stdout/stderr) or a rejection message.
+    Also fires tool_log_callback if registered.
     """
     if confirm_and_run_callback is None:
         return "ERROR: no confirmation handler registered"
     result = confirm_and_run_callback(script, base_dir)
     if result is None:
-        return "USER REJECTED: command was not executed"
+        output = "USER REJECTED: command was not executed"
-    return result
+    else:
+        output = result
+    if tool_log_callback is not None:
+        tool_log_callback(script, output)
+    return output
 
 # ------------------------------------------------------------------ gemini
 
@@ -321,7 +316,6 @@ def _send_gemini(md_content: str, user_message: str, base_dir: str) -> str:
         response = _gemini_chat.send_message(full_message)
 
         for round_idx in range(MAX_TOOL_ROUNDS):
-            # Log the raw response candidates as text summary
             text_parts_raw = [
                 part.text
                 for candidate in response.candidates
@@ -383,6 +377,32 @@ def _send_gemini(md_content: str, user_message: str, base_dir: str) -> str:
         raise _classify_gemini_error(exc) from exc
 
 # ------------------------------------------------------------------ anthropic
+#
+# Caching strategy (Anthropic prompt caching):
+#
+#   The Anthropic API caches a contiguous prefix of the input.  To maximise
+#   cache hits we structure every request as follows:
+#
+#   system (array form):
+#       [0] _ANTHROPIC_SYSTEM text   <- cache_control: ephemeral
+#           Stable across the whole session; cached after the first request.
+#
+#   tools:
+#       Last tool has cache_control: ephemeral.
+#       Stable across the whole session; cached together with the system prompt.
+#
+#   messages[0]  (first user turn ever, or re-sent each call):
+#       content[0]: <context> block  <- cache_control: ephemeral
+#           The aggregated markdown.  Changes only when the user regenerates.
+#           A new cache entry is created when it changes; otherwise it's a hit.
+#       content[1]: user question    <- no cache_control (varies every turn)
+#
+#   Subsequent turns (tool results, follow-up questions) are appended to
+#   _anthropic_history normally without extra cache markers.
+#
+#   Token cost of cache creation is ~25 % more than a normal input token, but
+#   cache reads cost ~10 % of a normal input token, so steady-state (many
+#   rounds / sends per session) is much cheaper.
 
 def _ensure_anthropic_client():
     global _anthropic_client
@@ -391,6 +411,7 @@ def _ensure_anthropic_client():
         creds = _load_credentials()
         _anthropic_client = anthropic.Anthropic(api_key=creds["anthropic"]["api_key"])
 
+
 def _send_anthropic(md_content: str, user_message: str, base_dir: str) -> str:
     global _anthropic_history
     import anthropic
@@ -398,19 +419,40 @@ def _send_anthropic(md_content: str, user_message: str, base_dir: str) -> str:
     try:
         _ensure_anthropic_client()
 
-        full_message = f"<context>\n{md_content}\n</context>\n\n{user_message}"
+        # Build the user content: context block (cached) + question (not cached).
-        _anthropic_history.append({"role": "user", "content": full_message})
+        # The cache anchor is placed on the context block so the entire prefix
+        # (system + tools + context) is eligible for caching.
+        user_content = [
+            {
+                "type": "text",
+                "text": f"<context>\n{md_content}\n</context>",
+                "cache_control": {"type": "ephemeral"},
+            },
+            {
+                "type": "text",
+                "text": user_message,
+            },
+        ]
+
+        _anthropic_history.append({"role": "user", "content": user_content})
 
         _append_comms("OUT", "request", {
-            "message": full_message,
+            "message": f"<context>\n{md_content}\n</context>\n\n{user_message}",
         })
 
         for round_idx in range(MAX_TOOL_ROUNDS):
             response = _anthropic_client.messages.create(
                 model=_model,
                 max_tokens=8096,
+                system=[
+                    {
+                        "type": "text",
+                        "text": _ANTHROPIC_SYSTEM,
+                        "cache_control": {"type": "ephemeral"},
+                    }
+                ],
                 tools=_ANTHROPIC_TOOLS,
-                messages=_anthropic_history
+                messages=_anthropic_history,
             )
 
             _anthropic_history.append({
@@ -418,22 +460,31 @@ def _send_anthropic(md_content: str, user_message: str, base_dir: str) -> str:
                 "content": response.content
             })
 
-            # Summarise the response content for the log
             text_blocks = [b.text for b in response.content if hasattr(b, "text") and b.text]
             tool_use_blocks = [
                 {"id": b.id, "name": b.name, "input": b.input}
                 for b in response.content
                 if b.type == "tool_use"
             ]
+
+            # Collect usage; cache fields are present when caching is active
+            usage_dict: dict = {}
+            if response.usage:
+                usage_dict["input_tokens"]  = response.usage.input_tokens
+                usage_dict["output_tokens"] = response.usage.output_tokens
+                cache_creation = getattr(response.usage, "cache_creation_input_tokens", None)
+                cache_read     = getattr(response.usage, "cache_read_input_tokens",     None)
+                if cache_creation is not None:
+                    usage_dict["cache_creation_input_tokens"] = cache_creation
+                if cache_read is not None:
+                    usage_dict["cache_read_input_tokens"] = cache_read
+
             _append_comms("IN", "response", {
                 "round":       round_idx,
                 "stop_reason": response.stop_reason,
                 "text":        "\n".join(text_blocks),
                 "tool_calls":  tool_use_blocks,
-                "usage":       {
+                "usage":       usage_dict,
-                    "input_tokens":  response.usage.input_tokens,
-                    "output_tokens": response.usage.output_tokens,
-                } if response.usage else {},
             })
 
             if response.stop_reason != "tool_use":
@@ -457,7 +508,7 @@ def _send_anthropic(md_content: str, user_message: str, base_dir: str) -> str:
                     tool_results.append({
                         "type":        "tool_result",
                         "tool_use_id": block.id,
-                        "content": output
+                        "content":     output,
                     })
 
             if not tool_results:
@@ -465,11 +516,14 @@ def _send_anthropic(md_content: str, user_message: str, base_dir: str) -> str:
 
             _anthropic_history.append({
                 "role":    "user",
-                "content": tool_results
+                "content": tool_results,
             })
 
             _append_comms("OUT", "tool_result_send", {
-                "results": [{"tool_use_id": r["tool_use_id"], "content": r["content"]} for r in tool_results],
+                "results": [
+                    {"tool_use_id": r["tool_use_id"], "content": r["content"]}
+                    for r in tool_results
+                ],
             })
 
         text_parts = [

config.toml

@@ -25,8 +25,12 @@ paths = []
 
 [discussion]
 history = [
-    "Make sure we are optimially using the anthropic api for this. \nI want to fully utilize caching if possible and just reduce overall loss of limits. \nAdd a log for comms history thats saved in ./logs and a the same for tool calls (scripts in ./scripts/generated, and their call equence in ./logs) these logs are closed in the next runtime of this gui program. \nOn open they amke new file buffers, each file buffer has a timestamp of when it was first made.",
+    "[User]\nMake sure we are optimially using the anthropic api for this. \nI want to fully utilize caching if possible and just reduce overall loss of limits. \nAdd a log for comms history thats saved in ./logs and a the same for tool calls (scripts in ./scripts/generated, and their call equence in ./logs) these logs are closed in the next runtime of this gui program. \nOn open they amke new file buffers, each file buffer has a timestamp of when it was first made.",
-    "Now finish the gui portion: in gui.py or anything left (last made seesion_logger.py it seems). Caching strategy also looks to be updated in ai_client.py but not complete.",
+    "[Claude]\nNow finish the gui portion: in gui.py or anything left (last made seesion_logger.py it seems). Caching strategy also looks to be updated in ai_client.py but not complete had to be reverted.",
+    "[User]\ncontinue. I had to revert ai_client.py becuase it was causing reponse code failures from the vendor. I added docs for anthropic into your context as well. For lasts version before revert of ai_client.py see:\n\n```py\n[Omitted already updated in your context as you fixed it last]\n```",
+    "[Claude]\nNow update `gui.py` to wire up `session_logger` and `tool_log_callback`:",
+    "[Claude]\nThe PowerShell `-replace` with a multi-line string replacement needs to be done differently. I'll write the complete updated `gui.py` directly with all the changes applied. The changes are targeted and well-understood:",
+    "[Claude]\nGood. Now let's verify all hooks are correctly in place and clean up the temp script:\n```",
 ]
 
 [ai]

dpg_layout.ini

@@ -16,7 +16,7 @@ DockId=0x00000005,0
 
 [Window][###30]
 Pos=0,654
-Size=364,1342
+Size=364,766
 Collapsed=0
 DockId=0x0000001D,0
 
@@ -45,8 +45,8 @@ Collapsed=0
 DockId=0x00000018,0
 
 [Window][###103]
-Pos=1870,1330
+Pos=2004,1330
-Size=1970,807
+Size=1836,807
 Collapsed=0
 DockId=0x0000001C,0
 
@@ -144,37 +144,37 @@ Collapsed=0
 
 [Window][###118]
 Pos=366,0
-Size=841,2137
+Size=650,2137
 Collapsed=0
 DockId=0x00000018,0
 
 [Window][###78]
-Pos=0,1998
+Pos=0,1422
-Size=364,139
+Size=364,715
 Collapsed=0
 DockId=0x0000001E,0
 
 [Window][###88]
-Pos=1209,0
+Pos=1018,0
-Size=659,2137
+Size=984,2137
 Collapsed=0
 DockId=0x00000019,0
 
 [Window][###95]
 Pos=366,0
-Size=841,2137
+Size=650,2137
 Collapsed=0
 DockId=0x00000018,1
 
 [Window][###110]
-Pos=1870,0
+Pos=2004,0
-Size=1970,1328
+Size=1836,1328
 Collapsed=0
 DockId=0x0000001B,0
 
 [Window][###112]
 Pos=366,0
-Size=841,2137
+Size=650,2137
 Collapsed=0
 DockId=0x00000018,2
 
@@ -183,17 +183,37 @@ Pos=1578,868
 Size=700,440
 Collapsed=0
 
+[Window][###228]
+Pos=1578,868
+Size=700,440
+Collapsed=0
+
+[Window][###317]
+Pos=1578,868
+Size=700,440
+Collapsed=0
+
+[Window][###412]
+Pos=1578,868
+Size=700,440
+Collapsed=0
+
+[Window][###513]
+Pos=1578,868
+Size=700,440
+Collapsed=0
+
 [Docking][Data]
 DockSpace             ID=0x7C6B3D9B Window=0xA87D555D Pos=0,0 Size=3840,2137 Split=X Selected=0x40484D8F
   DockNode            ID=0x00000003 Parent=0x7C6B3D9B SizeRef=364,1161 Split=Y Selected=0xEE087978
     DockNode          ID=0x00000005 Parent=0x00000003 SizeRef=235,354 Selected=0xEE087978
     DockNode          ID=0x00000006 Parent=0x00000003 SizeRef=235,805 Split=Y Selected=0x5F94F9BD
       DockNode        ID=0x00000009 Parent=0x00000006 SizeRef=235,453 Split=Y Selected=0x5F94F9BD
-        DockNode      ID=0x0000001D Parent=0x00000009 SizeRef=364,1342 Selected=0x5F94F9BD
+        DockNode      ID=0x0000001D Parent=0x00000009 SizeRef=364,766 Selected=0x5F94F9BD
-        DockNode      ID=0x0000001E Parent=0x00000009 SizeRef=364,139 Selected=0xF475F06A
+        DockNode      ID=0x0000001E Parent=0x00000009 SizeRef=364,715 Selected=0xF475F06A
       DockNode        ID=0x0000000A Parent=0x00000006 SizeRef=235,350 Selected=0x80199DAE
   DockNode            ID=0x00000004 Parent=0x7C6B3D9B SizeRef=3474,1161 Split=X
-    DockNode          ID=0x00000001 Parent=0x00000004 SizeRef=829,1161 Split=Y Selected=0x40484D8F
+    DockNode          ID=0x00000001 Parent=0x00000004 SizeRef=650,1161 Split=Y Selected=0x40484D8F
       DockNode        ID=0x00000007 Parent=0x00000001 SizeRef=595,492 Selected=0xBA13FCDE
       DockNode        ID=0x00000008 Parent=0x00000001 SizeRef=595,1643 Split=X Selected=0x40484D8F
         DockNode      ID=0x0000000F Parent=0x00000008 SizeRef=847,2137 Split=Y Selected=0x07E8375F
@@ -204,13 +224,13 @@ DockSpace             ID=0x7C6B3D9B Window=0xA87D555D Pos=0,0 Size=3840,2137 Spl
             DockNode  ID=0x00000017 Parent=0x00000013 SizeRef=1314,1749 Selected=0x4B454E0B
             DockNode  ID=0x00000018 Parent=0x00000013 SizeRef=1309,1749 CentralNode=1 Selected=0x73845A9B
           DockNode    ID=0x00000014 Parent=0x00000010 SizeRef=1967,445 Selected=0xC36FF36B
-    DockNode          ID=0x00000002 Parent=0x00000004 SizeRef=2631,1161 Split=X Selected=0x714F2F7B
+    DockNode          ID=0x00000002 Parent=0x00000004 SizeRef=2822,1161 Split=X Selected=0x714F2F7B
       DockNode        ID=0x0000000B Parent=0x00000002 SizeRef=968,1161 Selected=0xC915D9DA
       DockNode        ID=0x0000000C Parent=0x00000002 SizeRef=1661,1161 Split=Y Selected=0x714F2F7B
         DockNode      ID=0x0000000D Parent=0x0000000C SizeRef=396,342 Selected=0x714F2F7B
         DockNode      ID=0x0000000E Parent=0x0000000C SizeRef=396,817 Split=X Selected=0xCF08B82F
-          DockNode    ID=0x00000019 Parent=0x0000000E SizeRef=659,2137 Selected=0x052342BF
+          DockNode    ID=0x00000019 Parent=0x0000000E SizeRef=984,2137 Selected=0x052342BF
-          DockNode    ID=0x0000001A Parent=0x0000000E SizeRef=1970,2137 Split=Y Selected=0xCF08B82F
+          DockNode    ID=0x0000001A Parent=0x0000000E SizeRef=1836,2137 Split=Y Selected=0xCF08B82F
             DockNode  ID=0x0000001B Parent=0x0000001A SizeRef=2104,1328 Selected=0x43F4115A
             DockNode  ID=0x0000001C Parent=0x0000001A SizeRef=2104,807 Selected=0xCF08B82F
 

gui.py

@@ -1,4 +1,4 @@
-# gui.py
+# gui.py
 import dearpygui.dearpygui as dpg
 import tomllib
 import tomli_w
@@ -9,6 +9,7 @@ import aggregate
 import ai_client
 from ai_client import ProviderError
 import shell_runner
+import session_logger
 
 CONFIG_PATH = Path("config.toml")
 PROVIDERS = ["gemini", "anthropic"]
@@ -206,17 +207,24 @@ class App:
         self._pending_comms_lock = threading.Lock()
         self._comms_entry_count = 0
 
+        session_logger.open_session()
         ai_client.set_provider(self.current_provider, self.current_model)
         ai_client.confirm_and_run_callback = self._confirm_and_run
         ai_client.comms_log_callback = self._on_comms_entry
+        ai_client.tool_log_callback = self._on_tool_log
 
     # ---------------------------------------------------------------- comms log
 
     def _on_comms_entry(self, entry: dict):
         """Called from background thread; queue for main thread."""
+        session_logger.log_comms(entry)
         with self._pending_comms_lock:
             self._pending_comms.append(entry)
 
+    def _on_tool_log(self, script: str, result: str):
+        """Called from background thread when a tool call completes."""
+        session_logger.log_tool_call(script, result, None)
+
     def _flush_pending_comms(self):
         """Called every frame from the main render loop."""
         with self._pending_comms_lock:
@@ -772,6 +780,7 @@ class App:
             dpg.render_dearpygui_frame()
 
         dpg.save_init_file("dpg_layout.ini")
+        session_logger.close_session()
         dpg.destroy_context()
 
 
@@ -782,3 +791,4 @@ def main():
 
 if __name__ == "__main__":
     main()
+