Merge remote-tracking branch 'origin/master' into perf/autonomous-loop-readonly-batch

# Conflicts:
#	swarms/agents/autonomous_loop.py

Author: ayaangazali

examples/multi_agent/concurrent_examples/concurrent_mix.py

@@ -18,15 +18,12 @@
     corporate law and ensure that all hiring practices are in compliance with 
     state and federal regulations.
     """,
-    model_name="claude-sonnet-4-20250514",
+    model_name="claude-sonnet-5",
     max_loops=1,
     autosave=False,
     dashboard=False,
     verbose=True,
-    output_type="str",
-    artifacts_on=True,
-    artifacts_output_path="delaware_ccorp_hiring_description.md",
-    artifacts_file_extension=".md",
+    temperature=None,
 )
 
 indian_foreign_agent = Agent(
@@ -45,15 +42,12 @@
     implications of hiring foreign nationals and the requirements for obtaining 
     necessary visas and work permits.
     """,
-    model_name="claude-sonnet-4-20250514",
+    model_name="claude-sonnet-5",
     max_loops=1,
     autosave=False,
     dashboard=False,
     verbose=True,
-    output_type="str",
-    artifacts_on=True,
-    artifacts_output_path="indian_foreign_hiring_description.md",
-    artifacts_file_extension=".md",
+    temperature=None,
 )
 
 # List of agents and corresponding tasks

examples/multi_agent/hierarchical_swarm/hierarchical_swarm_plan_panel_example.py

@@ -0,0 +1,39 @@
+from swarms import Agent, HierarchicalSwarm
+
+
+def main() -> None:
+    """Run a director with two specialized workers.
+
+    Returns:
+        None.
+    """
+    researcher = Agent(
+        agent_name="Researcher",
+        agent_description="Finds relevant facts and key considerations.",
+        system_prompt="Research the assigned topic and return concise facts.",
+        model_name="gpt-5.4",
+        max_loops=1,
+    )
+    writer = Agent(
+        agent_name="Writer",
+        agent_description="Turns research into a clear final explanation.",
+        system_prompt="Write a concise explanation for a general audience.",
+        model_name="gpt-5.4",
+        max_loops=1,
+    )
+
+    swarm = HierarchicalSwarm(
+        name="ResearchWritingSwarm",
+        description="Researches a topic and explains the findings.",
+        agents=[researcher, writer],
+        max_loops=1,
+    )
+
+    result = swarm.run(
+        "Explain three practical benefits of multi-agent systems."
+    )
+    print(result)
+
+
+if __name__ == "__main__":
+    main()

pyproject.toml

@@ -5,7 +5,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "swarms"
-version = "14.0.2"
+version = "15.0.0"
 description = "Swarms - TGSC"
 license = "Apache-2.0"
 authors = ["Kye Gomez <kye@swarms.world>"]
@@ -99,8 +99,6 @@ pytest = "*"
 
 [tool.pytest.ini_options]
 testpaths = ["tests"]
-# Without this, pytest-asyncio's strict default silently skips every bare
-# `async def test_`, reporting it as a failure that never ran.
 asyncio_mode = "auto"
 markers = [
     "remote: tests that hit a real remote MCP server over the network",
@@ -109,12 +107,6 @@ markers = [
 [tool.ruff]
 line-length = 70
 
-# Library code under swarms/ is held to the full rule set. Example scripts,
-# test helpers and one-off scripts are not: an availability probe imports a
-# package to see whether it is installed (F401), a demo loads dotenv before
-# importing (E402), and a teaching example catches everything so it can print
-# a friendly message (E722). Enforcing those there produces churn without
-# making the library safer.
 [tool.ruff.lint.per-file-ignores]
 "examples/**" = ["E402", "E721", "E722", "F401", "F402"]
 "tests/**" = ["E402", "E721", "E722", "F401", "F402"]

requirements.txt

@@ -11,7 +11,7 @@ networkx
 httpx
 requests
 litellm==1.76.1
-mcp>=1.28.1,<2.0.0
+mcp>=1.28.1,<3.0.0
 schedule
 opentelemetry-sdk
 opentelemetry-exporter-otlp-proto-http

swarms/__init__.py

@@ -14,3 +14,36 @@
 from swarms.telemetry import *  # noqa: E402, F403
 from swarms.tools import *  # noqa: E402, F403
 from swarms.utils import *  # noqa: E402, F403
+
+
+def __getattr__(name: str) -> str:
+    """Resolve ``swarms.__version__`` on first access.
+
+    The version comes from the installed distribution's metadata rather
+    than a literal here, so it cannot drift from ``pyproject.toml``.
+    Resolving it lazily keeps the dist-info scan off the import path for
+    the majority of programs, which never read it.
+    """
+    if name == "__version__":
+        from importlib.metadata import (
+            PackageNotFoundError,
+            version,
+        )
+
+        try:
+            resolved = version("swarms")
+        except PackageNotFoundError:
+            # A source checkout with no installed distribution.
+            resolved = "unknown"
+
+        globals()["__version__"] = resolved
+        return resolved
+
+    raise AttributeError(
+        f"module {__name__!r} has no attribute {name!r}"
+    )
+
+
+def __dir__() -> list:
+    """Advertise ``__version__`` before anything has accessed it."""
+    return sorted(set(globals()) | {"__version__"})

swarms/agents/autonomous_loop.py

@@ -195,6 +195,42 @@ def _run_readonly_planning_tools(
 
         batch.clear()
 
+    def _maybe_compress_context(self) -> bool:
+        """Compress the conversation when it nears the context limit.
+
+        ``ContextCompressor`` measures and compacts
+        ``agent.short_memory``, but the request body actually sent to
+        the model is ``self._transcript`` — so after a compaction the
+        transcript is rebuilt around the summary. Compressing only the
+        mirror would leave the payload growing unbounded, which is the
+        failure the compressor exists to prevent.
+
+        Returns:
+            bool: True when a compression ran and the transcript was
+            re-seeded. The caller must then restore whatever immediate
+            instruction the model needs (e.g. the prompt for the
+            subtask in flight).
+        """
+        compressor = getattr(self.agent, "_context_compressor", None)
+        if compressor is None:
+            return False
+        summary = compressor.maybe_compress(self.agent)
+        if summary is None:
+            return False
+
+        # compact() already re-seeded short_memory with the summary,
+        # so the transcript copy must not be mirrored back a second
+        # time.
+        self._transcript = Transcript()
+        self._say_user(
+            "[Compressed Memory Summary]\n"
+            "Earlier turns were compressed to stay within the "
+            "context window. Progress so far:\n\n"
+            f"{summary}",
+            mirror=False,
+        )
+        return True
+
     def _run_autonomous_loop(
         self,
         task: Optional[Union[str, Any]] = None,
@@ -724,6 +760,16 @@ def _run_autonomous_loop(
                 ):
                     subtask_iterations += 1
 
+                    # Between iterations every recorded tool call has
+                    # been answered, so this is the one point the
+                    # transcript can be replaced without orphaning a
+                    # tool_call id.
+                    if self._maybe_compress_context():
+                        # The rebuilt transcript holds only the
+                        # summary; restore the instruction for the
+                        # subtask in flight.
+                        self._say_user(execution_prompt)
+
                     try:
                         response = self.agent.call_llm(
                             task=None,

swarms/prompts/agent_judge_prompt.py

@@ -1,132 +1,3 @@
-HIERARCHICAL_SWARM_JUDGE_PROMPT = """
-# Hierarchical Swarm Judge — Evaluation Protocol
-
-You are an elite evaluation agent embedded inside a hierarchical multi-agent swarm. Your sole responsibility is to rigorously assess the quality of every worker agent's output after each execution cycle and produce a structured, evidence-grounded scoring report.
-
-You are NOT a worker agent. You do not perform the task. You evaluate those who did.
-
----
-
-## Your Inputs
-
-You will receive:
-1. **The original task** — what the swarm was asked to accomplish.
-2. **The director's plan** — the strategy and order assignments issued by the director.
-3. **Each agent's output** — the actual response produced by each worker agent.
-4. **Full conversation history** — the complete context of everything that occurred before you were called.
-
----
-
-## Evaluation Dimensions
-
-Score each agent on the following five dimensions. Each dimension is scored 0–10. The overall agent score is the weighted average defined below.
-
-### 1. Task Adherence (weight: 25%)
-Did the agent actually do what it was assigned? Did it stay on topic and fulfill the specific order given by the director — not a paraphrase of it, not a tangential interpretation, but the precise assignment?
-
-- **10**: Perfectly on-task. Every part of the assignment is addressed directly.
-- **7–9**: Mostly on-task with minor drift or omissions.
-- **4–6**: Partially addressed the assignment; significant gaps or off-topic content.
-- **1–3**: Largely ignored the assigned task; substituted its own agenda.
-- **0**: No relevance to the assigned task whatsoever.
-
-### 2. Accuracy & Factual Integrity (weight: 25%)
-Are the claims, data points, and conclusions factually sound? Are assertions supported by reasoning or evidence, or are they speculative and unsupported?
-
-- **10**: All claims are accurate, well-supported, and internally consistent.
-- **7–9**: Mostly accurate; minor unsupported claims or imprecise statements.
-- **4–6**: Several questionable claims or logical inconsistencies.
-- **1–3**: Significant factual errors or unsupported speculation throughout.
-- **0**: Output is factually unreliable or contradicts known information.
-
-### 3. Depth & Completeness (weight: 20%)
-Did the agent produce a thorough, substantive response — or a shallow, surface-level one? Were edge cases, nuances, and implications considered?
-
-- **10**: Comprehensive. Covers all relevant angles with appropriate depth.
-- **7–9**: Solid depth; a few areas could be expanded.
-- **4–6**: Superficial in key areas; missing important dimensions.
-- **1–3**: Very thin output; little substance beyond restatement of the task.
-- **0**: Empty, trivially short, or entirely non-substantive.
-
-### 4. Clarity & Communication (weight: 15%)
-Is the output well-structured, readable, and unambiguous? Could a downstream agent or human act on this output without confusion?
-
-- **10**: Exceptionally clear, logically organized, precise language throughout.
-- **7–9**: Clear and readable; minor structural or phrasing issues.
-- **4–6**: Understandable but poorly organized or unnecessarily verbose/terse.
-- **1–3**: Confusing, disorganized, or full of ambiguous statements.
-- **0**: Incomprehensible or self-contradictory.
-
-### 5. Contribution to Swarm Goal (weight: 15%)
-Considering the swarm's overall objective, did this agent's output move the mission forward? Did it produce something that other agents or the director can build upon?
-
-- **10**: Directly advances the collective goal; highly actionable by downstream agents.
-- **7–9**: Useful contribution; minor gaps in handoff value.
-- **4–6**: Marginally useful; another agent would need to redo significant work.
-- **1–3**: Redundant, contradicts other agents, or creates confusion downstream.
-- **0**: Actively harmful to the swarm's progress.
-
----
-
-## Composite Score Formula
-
-```
-composite_score = (
-    task_adherence       * 0.25 +
-    accuracy             * 0.25 +
-    depth_completeness   * 0.20 +
-    clarity              * 0.15 +
-    swarm_contribution   * 0.15
-)
-```
-
-Round to the nearest integer (0–10) for the final `score` field.
-
----
-
-## Reasoning Standards
-
-Your `reasoning` field for each agent must:
-- Cite **specific content** from the agent's output (quote or paraphrase concrete examples).
-- Identify **what was done well** before identifying weaknesses.
-- Avoid vague language like "good job" or "needs improvement" — every claim must be specific.
-- Be no shorter than 3 sentences and no longer than 8 sentences.
-
-Your `suggestions` field must:
-- Give **concrete, actionable** improvement directions — not generic advice.
-- Specify *what* the agent should add, remove, or restructure in future iterations.
-- Be grounded in the gap between what was produced and what was needed.
-
----
-
-## Overall Report Standards
-
-Your `summary` must:
-- Synthesize how the agents performed **as a collective**, not just individually.
-- Identify the strongest and weakest agent by name.
-- Note any critical gaps in the swarm's combined output that the director should address in the next loop.
-- Be 3–6 sentences.
-
-Your `overall_quality` score must reflect the swarm's collective output quality — not the average of individual scores. Weight it toward the degree to which the swarm, as a whole, accomplished the original task.
-
----
-
-## Behavioral Rules
-
-- **Never hallucinate agent outputs.** Only evaluate what was actually provided.
-- **Never praise without evidence.** Every positive statement must cite something specific.
-- **Never penalize for scope.** Agents are only responsible for their assigned order — not the entire task.
-- **Maintain calibration.** A score of 10 should be genuinely exceptional. A score of 5 is mediocre but functional. Reserve 0–2 for outputs that are harmful or completely off-task.
-- **Be adversarially honest.** The purpose of your evaluation is to improve the swarm in subsequent loops — not to make agents feel good.
-
----
-
-## Output Format
-
-You must return a valid `JudgeReport` using the provided tool schema. Do not include any text outside the structured tool call.
-"""
-
-
 AGENT_JUDGE_PROMPT = """
 # Adaptive Output Evaluator - Role and Protocol