mycode-sdk 0.7.5__tar.gz → 0.8.0__tar.gz

{mycode_sdk-0.7.5 → mycode_sdk-0.8.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mycode-sdk
-Version: 0.7.5
+Version: 0.8.0
 Summary: Lightweight Python SDK for building AI agents.
 Project-URL: Homepage, https://github.com/legibet/mycode
 Project-URL: Repository, https://github.com/legibet/mycode

{mycode_sdk-0.7.5 → mycode_sdk-0.8.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "mycode-sdk"
-version = "0.7.5"
+version = "0.8.0"
 description = "Lightweight Python SDK for building AI agents."
 readme = "README.md"
 requires-python = ">=3.12"

{mycode_sdk-0.7.5 → mycode_sdk-0.8.0}/src/mycode/agent.py

@@ -23,6 +23,7 @@ from mycode.messages import (
     ConversationMessage,
     build_message,
     flatten_message_text,
+    text_block,
     tool_result_block,
     user_text_message,
 )
@@ -30,10 +31,13 @@ from mycode.models import infer_provider_from_model, resolve_model_metadata
 from mycode.providers import get_provider_adapter
 from mycode.providers.base import ProviderAdapter, ProviderRequest, ProviderStreamEvent
 from mycode.session import (
+    COMPACT_ACK,
     COMPACT_SUMMARY_PROMPT,
+    CONTINUATION_FOOTER,
+    CONTINUATION_HEADER,
     DEFAULT_COMPACT_THRESHOLD,
+    TRANSCRIPT_HINT,
     SessionStore,
-    apply_compact,
     build_compact_event,
     should_compact,
 )
@@ -500,7 +504,7 @@ class Agent:
                 provider=self.provider,
                 model=self.model,
                 session_id=self.session_id,
-                messages=self.messages,
+                messages=self._project_for_provider(self.messages),
                 system=self.system,
                 tools=self.tools.definitions,
                 max_tokens=self.max_tokens,
@@ -633,7 +637,12 @@ class Agent:
                 try:
                     async for event in self._compact(adapter, persist):
                         yield event
-                except (Exception, asyncio.CancelledError):
+                except asyncio.CancelledError:
+                    yield Event("error", {"message": "cancelled"})
+                    return
+                except Exception:
+                    # Best-effort: transient failures retry next threshold check;
+                    # persistent ones surface from phase 1 of the next turn.
                     logger.warning(
                         "Context compaction failed, continuing without compaction",
                         exc_info=True,
@@ -678,17 +687,52 @@ class Agent:
     # Context compaction
     # ------------------------------------------------------------------
 
+    def _project_for_provider(
+        self,
+        messages: list[ConversationMessage],
+    ) -> list[ConversationMessage]:
+        """Replace pre-compact history with a summary continuation."""
+
+        last_compact = -1
+        for i, message in enumerate(messages):
+            if message.get("role") == "compact":
+                last_compact = i
+
+        if last_compact < 0:
+            return messages
+
+        summary_text = ""
+        for block in messages[last_compact].get("content") or []:
+            if isinstance(block, dict) and block.get("type") == "text":
+                summary_text = str(block.get("text") or "")
+                break
+
+        tail = [m for m in messages[last_compact + 1 :] if m.get("role") != "compact"]
+        # No tail or assistant-led tail = mid-loop; append a resume instruction
+        # and skip the ack. A user-led tail needs the ack to keep alternation.
+        continue_now = not tail or tail[0].get("role") == "assistant"
+
+        parts = [CONTINUATION_HEADER, summary_text]
+        if self._store and self.session_id:
+            parts.append(TRANSCRIPT_HINT.format(path=self._store.messages_path(self.session_id)))
+        if continue_now:
+            parts.append(CONTINUATION_FOOTER)
+
+        projected = [build_message("user", [text_block("\n\n".join(parts))])]
+        if not continue_now:
+            projected.append(build_message("assistant", [text_block(COMPACT_ACK)]))
+        projected.extend(tail)
+        return projected
+
     async def _compact(
         self,
         adapter: ProviderAdapter,
         persist: PersistCallback,
     ) -> AsyncIterator[Event]:
-        """Generate a conversation summary and replace in-memory messages."""
-
-        compacted_count = len(self.messages)
+        """Generate a conversation summary and append a compact marker."""
 
         # Ask the same provider for a summary — no tools, just text generation.
-        compact_messages = list(self.messages) + [user_text_message(COMPACT_SUMMARY_PROMPT)]
+        compact_messages = self._project_for_provider(self.messages) + [user_text_message(COMPACT_SUMMARY_PROMPT)]
         request = ProviderRequest(
             provider=self.provider,
             model=self.model,
@@ -711,34 +755,22 @@ class Agent:
                     summary_message = msg
 
         if not summary_message:
-            logger.warning("Compaction produced no response")
-            return
+            raise ValueError("compaction produced no response")
 
         summary_text = flatten_message_text(summary_message, include_thinking=False)
         if not summary_text:
-            logger.warning("Compaction produced empty summary")
-            return
+            raise ValueError("compaction produced empty summary")
 
         summary_total_tokens = (summary_message.get("meta") or {}).get("total_tokens")
         compact_event = build_compact_event(
             summary_text,
             provider=self.provider,
             model=self.model,
-            compacted_count=compacted_count,
             total_tokens=summary_total_tokens,
         )
 
         # Persist the compact event (append-only — original messages stay in JSONL).
         await persist(compact_event)
-
-        # Rebuild in-memory messages from the compact event.
         self.messages.append(compact_event)
-        self.messages = apply_compact(self.messages)
-
-        yield Event(
-            "compact",
-            {
-                "message": f"Context compacted ({compacted_count} messages → summary)",
-                "compacted_count": compacted_count,
-            },
-        )
+
+        yield Event("compact", {})

{mycode_sdk-0.7.5 → mycode_sdk-0.8.0}/src/mycode/models_catalog.json

@@ -794,6 +794,13 @@
       "supports_pdf_input": true,
       "supports_reasoning": true
     },
+    "gpt-5.5-pro": {
+      "context_window": 1050000,
+      "max_output_tokens": 128000,
+      "supports_image_input": true,
+      "supports_pdf_input": true,
+      "supports_reasoning": true
+    },
     "gpt-image-1": {
       "context_window": 0,
       "max_output_tokens": 0,
@@ -1545,6 +1552,13 @@
       "supports_pdf_input": false,
       "supports_reasoning": true
     },
+    "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free": {
+      "context_window": 256000,
+      "max_output_tokens": 65536,
+      "supports_image_input": true,
+      "supports_pdf_input": false,
+      "supports_reasoning": true
+    },
     "nvidia/nemotron-3-super-120b-a12b": {
       "context_window": 262144,
       "max_output_tokens": 262144,
@@ -1755,6 +1769,13 @@
       "supports_pdf_input": true,
       "supports_reasoning": true
     },
+    "openai/gpt-5.5-pro": {
+      "context_window": 1050000,
+      "max_output_tokens": 128000,
+      "supports_image_input": true,
+      "supports_pdf_input": true,
+      "supports_reasoning": true
+    },
     "openai/gpt-oss-120b": {
       "context_window": 131072,
       "max_output_tokens": 32768,

{mycode_sdk-0.7.5 → mycode_sdk-0.8.0}/src/mycode/session.py

@@ -24,7 +24,7 @@ from mycode.messages import ConversationMessage, build_message, flatten_message_
 # Session format and compacting defaults
 # ---------------------------------------------------------------------
 
-MESSAGE_FORMAT_VERSION = 6
+MESSAGE_FORMAT_VERSION = 7
 DEFAULT_COMPACT_THRESHOLD = 0.8
 DEFAULT_SESSION_TITLE = "New chat"
 
@@ -35,26 +35,38 @@ capture everything needed to continue the work seamlessly.
 
 Include:
 
-1. **User Requests**: Every distinct request or instruction the user gave, \
+1. **Task and Intent**: Describe the user's overall goal — what is being \
+built, fixed, or investigated, and why.
+2. **Decisions and Constraints**: List the decisions made, constraints \
+discovered, and approaches chosen or rejected, with the reasoning behind \
+each.
+3. **User Requests**: Every distinct request or instruction the user gave, \
 in chronological order. Preserve the user's original wording for ambiguous \
 or nuanced requests.
-2. **Completed Work**: What was accomplished — files created, modified, or \
-deleted; bugs fixed; features added. Include file paths and function names.
-3. **Current State**: The exact state of the work right now — what is working, \
-what is broken, what is partially done.
-4. **Key Decisions**: Important decisions made, constraints discovered, \
-approaches chosen or rejected, and why.
-5. **Next Steps**: What remains to be done, any work that was in progress \
-when this summary was generated.
+4. **Files and Changes**: Enumerate every file read, modified, or created \
+— paths, what changed, and any code snippets the next turn will need to \
+reason about, quoted verbatim.
+5. **Errors and Fixes**: List errors encountered with the original message \
+verbatim, the cause if known, and the resolution — or that it remains open.
+6. **Current State**: What is verified working, what is known broken, what \
+is in progress.
+7. **Next Step**: The next step to take, with a direct quote from the most \
+recent conversation showing where the work left off.
 
 Rules:
-- Be specific: include file paths, function names, error messages, and \
-concrete details.
+- Be specific: reproduce file paths, function names, error messages, and \
+other identifiers verbatim — never paraphrase them.
 - Do not add suggestions or opinions — only summarize what happened.
 - Keep it concise but complete.\
 """
 
-_COMPACT_ACK = "Understood. I have the context from the conversation summary and will continue the work."
+CONTINUATION_HEADER = "This session is being continued from a previous conversation that was compacted to fit the context window. The summary below covers the earlier portion of the conversation."
+
+TRANSCRIPT_HINT = "For verbatim details not captured in this summary (exact code snippets, error messages, or earlier output), read the original conversation log at: {path}"
+
+CONTINUATION_FOOTER = 'Resume directly from where the work left off. Do not acknowledge this summary, do not recap, and do not preface with "I\'ll continue" or similar.'
+
+COMPACT_ACK = "Acknowledged."
 
 
 # ---------------------------------------------------------------------
@@ -88,7 +100,6 @@ def build_compact_event(
     *,
     provider: str,
     model: str,
-    compacted_count: int,
     total_tokens: int | None = None,
 ) -> ConversationMessage:
     """Build the compact event stored in session JSONL."""
@@ -96,43 +107,12 @@ def build_compact_event(
     meta: dict[str, Any] = {
         "provider": provider,
         "model": model,
-        "compacted_count": compacted_count,
     }
     if total_tokens is not None:
         meta["total_tokens"] = total_tokens
     return build_message("compact", [text_block(summary_text)], meta=meta)
 
 
-def apply_compact(messages: list[ConversationMessage]) -> list[ConversationMessage]:
-    """Replace the latest compact event with a summary + synthetic ack."""
-
-    # Only the newest compact event matters. Older history before it is no
-    # longer visible once the summary replaces that earlier conversation.
-    last_compact_index: int | None = None
-    for index, message in enumerate(messages):
-        if message.get("role") == "compact":
-            last_compact_index = index
-
-    if last_compact_index is None:
-        return messages
-
-    summary_text = ""
-    for block in messages[last_compact_index].get("content") or []:
-        if isinstance(block, dict) and block.get("type") == "text":
-            summary_text = str(block.get("text") or "")
-            break
-
-    return [
-        build_message(
-            "user",
-            [text_block(f"[Conversation Summary]\n\n{summary_text}")],
-            meta={"synthetic": True},
-        ),
-        build_message("assistant", [text_block(_COMPACT_ACK)], meta={"synthetic": True}),
-        *messages[last_compact_index + 1 :],
-    ]
-
-
 def build_rewind_event(rewind_to: int) -> ConversationMessage:
     """Build a rewind marker to append to session JSONL."""
 
@@ -312,13 +292,11 @@ class SessionStore:
         except FileNotFoundError:
             pass
 
-        # Replay order defines the visible conversation state.
-        # 1) compact rewrites older history into one summary view
-        # 2) rewind truncates that visible list by message index
+        # Visible state = raw JSONL minus rewound tails. `compact` markers
+        # stay inline; the agent substitutes them when calling the provider.
         # Orphan tool_use blocks (e.g. left open by a server crash) are
         # closed by the provider adapter at replay time, not here.
-        visible_messages = apply_compact(raw_messages)
-        visible_messages = apply_rewind(visible_messages)
+        visible_messages = apply_rewind(raw_messages)
 
         return {"session": self._summary(session_id, meta), "messages": visible_messages}