Python: Implement annotation-based context compaction (#4469)

* Implement annotation-based context compaction Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Handle missing compaction attributes in BaseChatClient Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Fix CI typing and bandit issues Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Optimize incremental compaction annotation pass Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * refinement * Python: add ToolResultCompactionStrategy and CompactionProvider Add ToolResultCompactionStrategy that collapses older tool-call groups into short summary messages (e.g. [Tool calls: get_weather]) while keeping the most recent groups verbatim. This mirrors the .NET ToolResultCompactionStrategy from PR #4533. Add CompactionProvider as a context-provider that auto-applies compaction before each agent turn and stores compacted history in session state after each turn. Includes tests and samples for both features. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * refinement and alignment with dotnet PR * updated tool result compaction * updated tool result compaction * Python: add ToolResultCompactionStrategy, CompactionProvider, and skip_excluded - ToolResultCompactionStrategy collapses older tool-call groups into [Tool results: func_name: result] summaries with bidirectional tracing (same pattern as SummarizationStrategy). - CompactionProvider as BaseContextProvider with separate before_strategy and after_strategy parameters. before_strategy compacts loaded context; after_strategy compacts stored history via history_source_id. - InMemoryHistoryProvider gains skip_excluded flag to filter out messages marked as excluded by compaction strategies. - Tests, samples, and exports updated. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * fixed checks * fix mypy * Fix: ensure summary messages from both strategies get full compaction annotations SummarizationStrategy was not calling annotate_message_groups after inserting its summary message, so the summary lacked core group annotations (id, kind, index, has_reasoning, _excluded). Added the missing call. ToolResultCompactionStrategy already had it. Added tests verifying both strategies produce fully annotated summaries. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * updated propagation * fix mypy --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-06-16 21:04:09 +08:00 · 2026-03-11 20:23:00 +01:00
parent 565c0b1623
commit 3e03a305f6
29 changed files with 4397 additions and 205 deletions
@@ -0,0 +1,115 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+import asyncio
+from typing import Any
+
+from agent_framework import (
+    CharacterEstimatorTokenizer,
+    ChatResponse,
+    Message,
+    SelectiveToolCallCompactionStrategy,
+    SlidingWindowStrategy,
+    SummarizationStrategy,
+    TokenBudgetComposedStrategy,
+    annotate_message_groups,
+    apply_compaction,
+    included_token_count,
+)
+
+"""This sample demonstrates composed in-run compaction with a token budget.
+
+Key components:
+- TokenBudgetComposedStrategy
+- Sequential strategy composition
+- Summarization with a SupportsChatGetResponse-compatible summarizer client
+"""
+
+
+class BudgetSummaryClient:
+    async def get_response(
+        self,
+        messages: list[Message],
+        *,
+        stream: bool = False,
+        options: dict[str, Any] | None = None,
+        **kwargs: Any,
+    ) -> ChatResponse:
+        summary_text = f"Budget summary generated from {len(messages)} prompt messages."
+        return ChatResponse(messages=[Message(role="assistant", text=summary_text)])
+
+
+def _build_long_history() -> list[Message]:
+    history = [Message(role="system", text="You are a migration copilot.")]
+    for i in range(1, 8):
+        history.append(
+            Message(
+                role="user",
+                text=f"Iteration {i}: capture migration requirements and edge cases.",
+            )
+        )
+        history.append(
+            Message(
+                role="assistant",
+                text=(
+                    f"Iteration {i}: detailed plan with dependencies, rollback guidance, and testing details. "
+                    "This sentence is intentionally long to create token pressure."
+                ),
+            )
+        )
+    return history
+
+
+async def main() -> None:
+    # 1. Build synthetic history representing long-running in-run growth.
+    messages = _build_long_history()
+
+    # 2. Configure tokenizer and measure token count before compaction.
+    tokenizer = CharacterEstimatorTokenizer()
+    annotate_message_groups(messages, tokenizer=tokenizer)
+    budget_before = included_token_count(messages)
+
+    # 3. Configure composed strategy stack.
+    composed = TokenBudgetComposedStrategy(
+        token_budget=200,
+        tokenizer=tokenizer,
+        strategies=[
+            SelectiveToolCallCompactionStrategy(keep_last_tool_call_groups=0),
+            SummarizationStrategy(
+                client=BudgetSummaryClient(),
+                target_count=3,
+                threshold=3,
+            ),
+            SlidingWindowStrategy(keep_last_groups=4),
+        ],
+    )
+
+    # 4. Apply compaction and inspect the budget result.
+    projected = await apply_compaction(messages, strategy=composed, tokenizer=tokenizer)
+    budget_after = included_token_count(messages)
+
+    print(f"Projected messages after compaction: {len(projected)}")
+    print(f"Included token count before compaction: {budget_before}")
+    print(f"Included token count after compaction: {budget_after}")
+    print("Projected roles:", [m.role for m in projected])
+    print("Projected messages with token counts:")
+    for msg in projected:
+        group = msg.additional_properties.get("_group")
+        token_count = group.get("token_count") if isinstance(group, dict) else None
+        text_preview = msg.text[:80] if msg.text else "<non-text>"
+        print(f"- [{msg.role}] {text_preview} ({token_count} tokens)")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
+
+"""
+Sample output:
+Projected messages after compaction: 3
+Included token count before compaction: 793
+Included token count after compaction: 144
+Projected roles: ['system', 'user', 'assistant']
+Projected messages with token counts:
+- [system] You are a migration copilot. (35 tokens)
+- [user] Iteration 7: capture migration requirements and edge cases. (43 tokens)
+- [assistant] Iteration 7: detailed plan with dependencies, rollback guidance, and testing det (66 tokens)
+"""