pyagent-compress 0.1.0__tar.gz

pyagent_compress-0.1.0/.gitignore

@@ -0,0 +1,18 @@
+__pycache__/
+*.py[cod]
+*$py.class
+*.egg-info/
+dist/
+build/
+.eggs/
+*.egg
+.venv/
+venv/
+.mypy_cache/
+.ruff_cache/
+.pytest_cache/
+.coverage
+htmlcov/
+site/
+.env
+*.log

pyagent_compress-0.1.0/PKG-INFO

@@ -0,0 +1,53 @@
+Metadata-Version: 2.4
+Name: pyagent-compress
+Version: 0.1.0
+Summary: Inter-agent message compression and token budget management for multi-agent LLM systems
+License: MIT
+Keywords: LLM,agents,compression,efficiency,tokens
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Typing :: Typed
+Requires-Python: >=3.11
+Requires-Dist: pyagent-patterns>=0.1.0
+Provides-Extra: dev
+Requires-Dist: mypy>=1.10; extra == 'dev'
+Requires-Dist: pytest-asyncio>=0.23; extra == 'dev'
+Requires-Dist: pytest>=8.0; extra == 'dev'
+Requires-Dist: ruff>=0.5; extra == 'dev'
+Description-Content-Type: text/markdown
+
+# pyagent-compress
+
+**Inter-agent message compression and token budget management** for multi-agent LLM systems. Reduce token costs without losing key information.
+
+## Install
+
+```bash
+pip install pyagent-compress
+```
+
+## Components
+
+- **MessageCompressor** — Reduce message size by removing filler and ranking sentences
+- **AgentPruner** — Detect and remove non-contributing agents
+- **InteractionPruner** — Detect consensus and prune redundant rounds
+- **TokenBudget** — Enforce per-agent and per-workflow token limits
+- **CompressMiddleware** — Auto-compress agent outputs
+
+## Quick Example
+
+```python
+from pyagent_compress import MessageCompressor, TokenBudget
+
+compressor = MessageCompressor(target_ratio=0.5)
+result = compressor.compress("Let me think about this... Basically, revenue grew 15%.")
+print(f"Savings: {result.savings_pct:.0%}")
+
+budget = TokenBudget(workflow_limit=50_000, per_agent_limit=10_000)
+budget.consume("agent_a", 3000)
+```

pyagent_compress-0.1.0/README.md

@@ -0,0 +1,30 @@
+# pyagent-compress
+
+**Inter-agent message compression and token budget management** for multi-agent LLM systems. Reduce token costs without losing key information.
+
+## Install
+
+```bash
+pip install pyagent-compress
+```
+
+## Components
+
+- **MessageCompressor** — Reduce message size by removing filler and ranking sentences
+- **AgentPruner** — Detect and remove non-contributing agents
+- **InteractionPruner** — Detect consensus and prune redundant rounds
+- **TokenBudget** — Enforce per-agent and per-workflow token limits
+- **CompressMiddleware** — Auto-compress agent outputs
+
+## Quick Example
+
+```python
+from pyagent_compress import MessageCompressor, TokenBudget
+
+compressor = MessageCompressor(target_ratio=0.5)
+result = compressor.compress("Let me think about this... Basically, revenue grew 15%.")
+print(f"Savings: {result.savings_pct:.0%}")
+
+budget = TokenBudget(workflow_limit=50_000, per_agent_limit=10_000)
+budget.consume("agent_a", 3000)
+```

pyagent_compress-0.1.0/pyproject.toml

@@ -0,0 +1,29 @@
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[project]
+name = "pyagent-compress"
+version = "0.1.0"
+description = "Inter-agent message compression and token budget management for multi-agent LLM systems"
+readme = "README.md"
+requires-python = ">=3.11"
+license = {text = "MIT"}
+keywords = ["agents", "compression", "LLM", "tokens", "efficiency"]
+classifiers = [
+    "Development Status :: 3 - Alpha",
+    "Intended Audience :: Developers",
+    "License :: OSI Approved :: MIT License",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+    "Typing :: Typed",
+]
+dependencies = ["pyagent-patterns>=0.1.0"]
+
+[project.optional-dependencies]
+dev = ["pytest>=8.0", "pytest-asyncio>=0.23", "ruff>=0.5", "mypy>=1.10"]
+
+[tool.hatch.build.targets.wheel]
+packages = ["src/pyagent_compress"]

pyagent_compress-0.1.0/src/pyagent_compress/__init__.py

@@ -0,0 +1,15 @@
+"""PyAgent Compress — inter-agent message compression and token budget management."""
+
+from pyagent_compress.budget import TokenBudget
+from pyagent_compress.compressor import MessageCompressor
+from pyagent_compress.middleware import CompressMiddleware
+from pyagent_compress.pruner import AgentPruner, InteractionPruner
+
+__all__ = [
+    "MessageCompressor",
+    "AgentPruner",
+    "InteractionPruner",
+    "TokenBudget",
+    "CompressMiddleware",
+]
+__version__ = "0.1.0"

pyagent_compress-0.1.0/src/pyagent_compress/budget.py

@@ -0,0 +1,118 @@
+"""TokenBudget: enforce per-agent and per-workflow token limits.
+
+Tracks token consumption across a workflow and raises BudgetExceeded
+when limits are hit, enabling graceful degradation.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+
+class BudgetExceeded(Exception):
+    """Raised when a token budget is exceeded."""
+
+    def __init__(self, agent: str, used: int, limit: int) -> None:
+        self.agent = agent
+        self.used = used
+        self.limit = limit
+        super().__init__(f"Agent '{agent}' exceeded budget: {used}/{limit} tokens")
+
+
+@dataclass
+class AgentBudget:
+    """Budget tracking for a single agent."""
+
+    name: str
+    limit: int
+    used: int = 0
+
+    @property
+    def remaining(self) -> int:
+        return max(0, self.limit - self.used)
+
+    @property
+    def utilization(self) -> float:
+        return self.used / self.limit if self.limit > 0 else 0.0
+
+    def consume(self, tokens: int, strict: bool = True) -> None:
+        """Consume tokens from budget.
+
+        Args:
+            tokens: Number of tokens consumed.
+            strict: If True, raises BudgetExceeded. If False, just tracks.
+        """
+        self.used += tokens
+        if strict and self.used > self.limit:
+            raise BudgetExceeded(self.name, self.used, self.limit)
+
+
+@dataclass
+class TokenBudget:
+    """Manage token budgets across a multi-agent workflow.
+
+    Args:
+        workflow_limit: Total token limit across all agents.
+        per_agent_limit: Default per-agent token limit.
+        strict: If True, raises BudgetExceeded on limit violations.
+    """
+
+    workflow_limit: int = 100_000
+    per_agent_limit: int = 20_000
+    strict: bool = True
+    _agents: dict[str, AgentBudget] = field(default_factory=dict)
+    _total_used: int = 0
+
+    def register_agent(self, name: str, limit: int | None = None) -> None:
+        """Register an agent with an optional custom limit."""
+        self._agents[name] = AgentBudget(name=name, limit=limit or self.per_agent_limit)
+
+    def consume(self, agent_name: str, tokens: int) -> None:
+        """Record token consumption for an agent.
+
+        Args:
+            agent_name: The consuming agent's name.
+            tokens: Number of tokens consumed.
+        """
+        if agent_name not in self._agents:
+            self.register_agent(agent_name)
+
+        self._agents[agent_name].consume(tokens, strict=self.strict)
+        self._total_used += tokens
+
+        if self.strict and self._total_used > self.workflow_limit:
+            raise BudgetExceeded("workflow", self._total_used, self.workflow_limit)
+
+    def remaining(self, agent_name: str | None = None) -> int:
+        """Get remaining tokens for an agent or the whole workflow."""
+        if agent_name:
+            budget = self._agents.get(agent_name)
+            return budget.remaining if budget else self.per_agent_limit
+        return max(0, self.workflow_limit - self._total_used)
+
+    @property
+    def total_used(self) -> int:
+        return self._total_used
+
+    @property
+    def workflow_utilization(self) -> float:
+        return self._total_used / self.workflow_limit if self.workflow_limit > 0 else 0.0
+
+    def summary(self) -> dict[str, dict[str, int | float]]:
+        """Return a summary of all agent budgets."""
+        result: dict[str, dict[str, int | float]] = {
+            "workflow": {
+                "limit": self.workflow_limit,
+                "used": self._total_used,
+                "remaining": self.remaining(),
+                "utilization": self.workflow_utilization,
+            }
+        }
+        for name, budget in self._agents.items():
+            result[name] = {
+                "limit": budget.limit,
+                "used": budget.used,
+                "remaining": budget.remaining,
+                "utilization": budget.utilization,
+            }
+        return result

pyagent_compress-0.1.0/src/pyagent_compress/compressor.py

@@ -0,0 +1,144 @@
+"""MessageCompressor: reduce inter-agent message size while preserving key information.
+
+Implements extractive compression: keeps the most important sentences
+from verbose LLM reasoning traces, stripping filler and repetition.
+
+Based on: "Cut the Crap: Economical Communication Pipeline for LLM-based MAS" (2024)
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+
+
+@dataclass(frozen=True)
+class CompressionResult:
+    """Result of message compression.
+
+    Attributes:
+        original: The original text.
+        compressed: The compressed text.
+        original_tokens: Estimated original token count.
+        compressed_tokens: Estimated compressed token count.
+        savings_pct: Percentage of tokens saved (0-1).
+    """
+
+    original: str
+    compressed: str
+    original_tokens: int
+    compressed_tokens: int
+
+    @property
+    def savings_pct(self) -> float:
+        if self.original_tokens == 0:
+            return 0.0
+        return 1.0 - (self.compressed_tokens / self.original_tokens)
+
+
+# Filler phrases commonly found in LLM reasoning traces
+_FILLER_PATTERNS = [
+    r"(?i)\b(let me think|let's think|okay so|well|basically|essentially|in other words)\b[,.]?\s*",
+    r"(?i)\b(as I mentioned|as we discussed|to summarize so far|in summary)\b[,.]?\s*",
+    r"(?i)\b(it's worth noting that|it's important to note that|I should mention that)\b\s*",
+    r"(?i)\b(first of all|secondly|thirdly|finally|in conclusion)\b[,.]?\s*",
+    r"(?i)\b(I think|I believe|in my opinion|from my perspective)\b[,.]?\s*",
+]
+
+
+class MessageCompressor:
+    """Compress inter-agent messages by removing filler and extracting key sentences.
+
+    Args:
+        target_ratio: Target compression ratio (0.3 = keep 30% of tokens).
+        min_sentence_length: Minimum characters for a sentence to be kept.
+        remove_filler: Whether to strip filler phrases.
+    """
+
+    def __init__(
+        self,
+        target_ratio: float = 0.5,
+        min_sentence_length: int = 20,
+        remove_filler: bool = True,
+    ) -> None:
+        self._target_ratio = target_ratio
+        self._min_sentence_length = min_sentence_length
+        self._remove_filler = remove_filler
+
+    def compress(self, text: str) -> CompressionResult:
+        """Compress a message text.
+
+        Strategy:
+        1. Remove filler phrases
+        2. Split into sentences
+        3. Score sentences by information density
+        4. Keep top sentences until target ratio met
+        """
+        original_tokens = len(text) // 4
+
+        if original_tokens <= 20:
+            # Too short to compress meaningfully
+            return CompressionResult(text, text, original_tokens, original_tokens)
+
+        working = text
+
+        # Step 1: Remove filler
+        if self._remove_filler:
+            for pattern in _FILLER_PATTERNS:
+                working = re.sub(pattern, "", working)
+
+        # Step 2: Split into sentences
+        sentences = re.split(r"(?<=[.!?])\s+", working)
+        sentences = [s.strip() for s in sentences if len(s.strip()) >= self._min_sentence_length]
+
+        if not sentences:
+            compressed_tokens = len(working) // 4
+            return CompressionResult(text, working.strip(), original_tokens, compressed_tokens)
+
+        # Step 3: Score sentences
+        scored = [(self._score_sentence(s), s) for s in sentences]
+        scored.sort(key=lambda x: x[0], reverse=True)
+
+        # Step 4: Keep top sentences until target ratio
+        target_tokens = int(original_tokens * self._target_ratio)
+        kept: list[str] = []
+        running_tokens = 0
+
+        for score, sentence in scored:
+            sent_tokens = len(sentence) // 4
+            if running_tokens + sent_tokens > target_tokens and kept:
+                break
+            kept.append(sentence)
+            running_tokens += sent_tokens
+
+        # Restore original order
+        ordered = [s for s in sentences if s in kept]
+        compressed = " ".join(ordered)
+        compressed_tokens = len(compressed) // 4
+
+        return CompressionResult(text, compressed, original_tokens, compressed_tokens)
+
+    @staticmethod
+    def _score_sentence(sentence: str) -> float:
+        """Score a sentence by information density (higher = more informative)."""
+        score = 0.0
+
+        # Longer sentences tend to carry more information (diminishing returns)
+        score += min(len(sentence.split()) / 20, 1.0) * 0.3
+
+        # Sentences with numbers/data are more informative
+        numbers = len(re.findall(r"\d+\.?\d*", sentence))
+        score += min(numbers / 3, 1.0) * 0.3
+
+        # Sentences with technical terms
+        technical = len(re.findall(
+            r"(?i)\b(result|conclusion|therefore|because|shows|indicates|found|"
+            r"evidence|data|percent|increase|decrease|significant)\b", sentence
+        ))
+        score += min(technical / 3, 1.0) * 0.2
+
+        # Shorter sentences with substance are preferred
+        if 10 <= len(sentence.split()) <= 25:
+            score += 0.2
+
+        return score

pyagent_compress-0.1.0/src/pyagent_compress/middleware.py

@@ -0,0 +1,102 @@
+"""CompressMiddleware: inject compression between pattern stages.
+
+Wraps agents so that their outputs are automatically compressed
+before being passed to the next agent, reducing inter-agent token transfer.
+"""
+
+from __future__ import annotations
+
+from pyagent_patterns.base import Agent, Message
+from pyagent_compress.budget import TokenBudget
+from pyagent_compress.compressor import MessageCompressor
+
+
+class CompressedAgent(Agent):
+    """Agent wrapper that compresses output messages.
+
+    Args:
+        agent: The original agent.
+        compressor: MessageCompressor instance.
+        budget: Optional TokenBudget for tracking.
+    """
+
+    def __init__(
+        self,
+        agent: Agent,
+        compressor: MessageCompressor,
+        budget: TokenBudget | None = None,
+    ) -> None:
+        super().__init__(
+            name=agent.name,
+            llm=agent.llm,
+            system_prompt=agent.system_prompt,
+            description=agent.description,
+        )
+        self._original = agent
+        self._compressor = compressor
+        self._budget = budget
+        self.compression_log: list[dict[str, int | float]] = []
+
+    async def run(self, messages: list[Message]) -> Message:
+        """Run the agent and compress the output."""
+        result = await self._original.run(messages)
+
+        # Compress the output
+        compressed = self._compressor.compress(result.content)
+        self.compression_log.append({
+            "original_tokens": compressed.original_tokens,
+            "compressed_tokens": compressed.compressed_tokens,
+            "savings_pct": compressed.savings_pct,
+        })
+
+        # Track budget if available
+        if self._budget:
+            self._budget.consume(self._original.name, compressed.compressed_tokens)
+
+        return Message(
+            role=result.role,
+            content=compressed.compressed,
+            name=result.name,
+            metadata={
+                **result.metadata,
+                "compressed": True,
+                "original_tokens": compressed.original_tokens,
+                "compressed_tokens": compressed.compressed_tokens,
+                "savings_pct": compressed.savings_pct,
+            },
+        )
+
+
+class CompressMiddleware:
+    """Middleware that wraps agents with automatic output compression.
+
+    Usage:
+        middleware = CompressMiddleware(target_ratio=0.5)
+        compressed_agent = middleware.wrap(my_agent)
+
+    Args:
+        compressor: Optional MessageCompressor. Created with defaults if None.
+        budget: Optional TokenBudget for workflow-wide tracking.
+        target_ratio: Target compression ratio (used if compressor not provided).
+    """
+
+    def __init__(
+        self,
+        compressor: MessageCompressor | None = None,
+        budget: TokenBudget | None = None,
+        target_ratio: float = 0.5,
+    ) -> None:
+        self._compressor = compressor or MessageCompressor(target_ratio=target_ratio)
+        self._budget = budget
+
+    def wrap(self, agent: Agent) -> CompressedAgent:
+        """Wrap an agent with compression."""
+        return CompressedAgent(
+            agent=agent,
+            compressor=self._compressor,
+            budget=self._budget,
+        )
+
+    def wrap_all(self, agents: list[Agent]) -> list[CompressedAgent]:
+        """Wrap multiple agents."""
+        return [self.wrap(a) for a in agents]

pyagent_compress-0.1.0/src/pyagent_compress/pruner.py

@@ -0,0 +1,161 @@
+"""AgentPruner and InteractionPruner: reduce multi-agent overhead.
+
+AgentPruner: detect and eliminate non-contributing agents mid-execution.
+InteractionPruner: skip communication rounds when consensus is reached early.
+
+Based on: arxiv:2503.18891 "AgentDropout: Dynamic Agent Elimination for Token-Efficient MAS"
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from pyagent_patterns.base import Message
+
+
+@dataclass(frozen=True)
+class ContributionScore:
+    """How much an agent is contributing to the conversation."""
+
+    agent_name: str
+    score: float  # 0-1
+    unique_info: float  # 0-1 — how much unique information vs repetition
+    message_count: int
+
+
+class AgentPruner:
+    """Detect and flag non-contributing agents for removal.
+
+    Scores each agent's contribution based on:
+    - Unique information (not repeated from other agents)
+    - Response diversity (different from own previous responses)
+    - Task relevance (overlap with original task keywords)
+
+    Args:
+        min_contribution: Minimum contribution score (0-1) to keep an agent.
+        window_size: Number of recent messages to analyze per agent.
+    """
+
+    def __init__(self, min_contribution: float = 0.3, window_size: int = 5) -> None:
+        self._min_contribution = min_contribution
+        self._window_size = window_size
+
+    def score_agents(
+        self,
+        messages: list[Message],
+        task: str,
+    ) -> list[ContributionScore]:
+        """Score each agent's contribution from message history."""
+        # Group messages by agent name
+        by_agent: dict[str, list[str]] = {}
+        for msg in messages:
+            name = msg.name or "unknown"
+            by_agent.setdefault(name, []).append(msg.content)
+
+        all_contents = [m.content for m in messages]
+        task_words = set(task.lower().split())
+
+        scores: list[ContributionScore] = []
+        for agent_name, contents in by_agent.items():
+            recent = contents[-self._window_size :]
+
+            # Unique info: how much of this agent's content is NOT in other agents' content
+            other_text = " ".join(
+                c for name, cs in by_agent.items() if name != agent_name for c in cs
+            ).lower()
+            unique_words = set()
+            for content in recent:
+                for word in content.lower().split():
+                    if word not in other_text and len(word) > 3:
+                        unique_words.add(word)
+
+            total_words = sum(len(c.split()) for c in recent)
+            unique_ratio = len(unique_words) / max(total_words, 1)
+
+            # Task relevance
+            agent_words = set(" ".join(recent).lower().split())
+            relevance = len(agent_words & task_words) / max(len(task_words), 1)
+
+            # Self-diversity (are recent messages different from each other?)
+            if len(recent) >= 2:
+                diversity = 1.0 - self._similarity(recent[-1], recent[-2])
+            else:
+                diversity = 1.0
+
+            score = 0.4 * unique_ratio + 0.3 * relevance + 0.3 * diversity
+            scores.append(ContributionScore(
+                agent_name=agent_name,
+                score=min(score, 1.0),
+                unique_info=unique_ratio,
+                message_count=len(contents),
+            ))
+
+        return scores
+
+    def should_prune(self, scores: list[ContributionScore]) -> list[str]:
+        """Return agent names that should be pruned (below min contribution)."""
+        return [s.agent_name for s in scores if s.score < self._min_contribution]
+
+    @staticmethod
+    def _similarity(a: str, b: str) -> float:
+        """Simple word-overlap similarity between two texts."""
+        words_a = set(a.lower().split())
+        words_b = set(b.lower().split())
+        if not words_a or not words_b:
+            return 0.0
+        return len(words_a & words_b) / len(words_a | words_b)
+
+
+class InteractionPruner:
+    """Detect early consensus and skip remaining communication rounds.
+
+    Analyzes agent outputs to determine if consensus has been reached,
+    allowing patterns like Debate or Voting to terminate early.
+
+    Args:
+        consensus_threshold: Minimum similarity between agent outputs
+            to consider consensus reached (0-1).
+        min_rounds: Minimum rounds before early termination is allowed.
+    """
+
+    def __init__(
+        self,
+        consensus_threshold: float = 0.7,
+        min_rounds: int = 1,
+    ) -> None:
+        self._threshold = consensus_threshold
+        self._min_rounds = min_rounds
+
+    def has_consensus(self, outputs: list[str], current_round: int) -> bool:
+        """Check if agents have reached consensus.
+
+        Args:
+            outputs: Current round's outputs from all agents.
+            current_round: The current round number (1-indexed).
+
+        Returns:
+            True if consensus is reached and we can stop early.
+        """
+        if current_round < self._min_rounds:
+            return False
+
+        if len(outputs) < 2:
+            return True
+
+        # Check pairwise similarity
+        similarities: list[float] = []
+        for i in range(len(outputs)):
+            for j in range(i + 1, len(outputs)):
+                sim = self._similarity(outputs[i], outputs[j])
+                similarities.append(sim)
+
+        avg_similarity = sum(similarities) / len(similarities) if similarities else 0.0
+        return avg_similarity >= self._threshold
+
+    @staticmethod
+    def _similarity(a: str, b: str) -> float:
+        words_a = set(a.lower().split())
+        words_b = set(b.lower().split())
+        if not words_a or not words_b:
+            return 0.0
+        return len(words_a & words_b) / len(words_a | words_b)

pyagent_compress-0.1.0/tests/test_compress.py

@@ -0,0 +1,98 @@
+"""Tests for pyagent-compress."""
+
+from __future__ import annotations
+
+import pytest
+
+from pyagent_patterns.base import Agent, Message, MockLLM, Role
+from pyagent_compress.budget import BudgetExceeded, TokenBudget
+from pyagent_compress.compressor import MessageCompressor
+from pyagent_compress.middleware import CompressMiddleware
+from pyagent_compress.pruner import AgentPruner, InteractionPruner
+
+
+def test_compressor_reduces_tokens():
+    compressor = MessageCompressor(target_ratio=0.5)
+    text = (
+        "Let me think about this carefully. Basically, the analysis shows that "
+        "revenue increased by 15% year-over-year. The data indicates a strong "
+        "upward trend in Q3 earnings. In other words, the company is performing "
+        "well above market expectations. It's worth noting that the profit margin "
+        "expanded to 23%, which is significant compared to the industry average of 18%. "
+        "The conclusion is that this represents a solid buy opportunity."
+    )
+    result = compressor.compress(text)
+    assert result.compressed_tokens < result.original_tokens
+    assert result.savings_pct > 0
+
+
+def test_compressor_short_text_unchanged():
+    compressor = MessageCompressor()
+    result = compressor.compress("Short text")
+    assert result.compressed == result.original
+
+
+def test_agent_pruner_scores():
+    pruner = AgentPruner(min_contribution=0.3)
+    messages = [
+        Message(role=Role.ASSISTANT, content="Unique analysis of market trends", name="analyst"),
+        Message(role=Role.ASSISTANT, content="Unique analysis of market trends", name="copycat"),
+        Message(role=Role.ASSISTANT, content="Different risk assessment with new data", name="risk"),
+    ]
+    scores = pruner.score_agents(messages, "analyze market trends")
+    assert len(scores) == 3
+
+
+def test_interaction_pruner_consensus():
+    pruner = InteractionPruner(consensus_threshold=0.5, min_rounds=1)
+
+    # Same outputs → consensus
+    assert pruner.has_consensus(["The answer is 42", "The answer is 42"], current_round=1) is True
+
+    # Very different outputs → no consensus
+    assert pruner.has_consensus(
+        ["Completely different analysis about stocks",
+         "An unrelated discussion about weather patterns"],
+        current_round=1,
+    ) is False
+
+
+def test_interaction_pruner_min_rounds():
+    pruner = InteractionPruner(consensus_threshold=0.5, min_rounds=2)
+    # Round 1: should not trigger even if consensus
+    assert pruner.has_consensus(["Same", "Same"], current_round=1) is False
+    # Round 2: now can trigger
+    assert pruner.has_consensus(["Same answer", "Same answer"], current_round=2) is True
+
+
+def test_token_budget_tracking():
+    budget = TokenBudget(workflow_limit=10000, per_agent_limit=5000, strict=False)
+    budget.consume("agent_a", 1000)
+    budget.consume("agent_b", 2000)
+    assert budget.total_used == 3000
+    assert budget.remaining() == 7000
+    assert budget.remaining("agent_a") == 4000
+
+
+def test_token_budget_strict_raises():
+    budget = TokenBudget(workflow_limit=100, per_agent_limit=50, strict=True)
+    budget.consume("agent_a", 40)
+    with pytest.raises(BudgetExceeded):
+        budget.consume("agent_a", 20)  # Exceeds per-agent limit of 50
+
+
+@pytest.mark.asyncio
+async def test_compress_middleware():
+    llm = MockLLM(responses=[
+        "Let me think about this carefully. Basically, the revenue data shows "
+        "a 15% increase year-over-year, which is significant. The profit margin "
+        "expanded to 23%. In conclusion, this is a buy signal."
+    ])
+    agent = Agent("analyst", llm)
+
+    middleware = CompressMiddleware(target_ratio=0.5)
+    compressed = middleware.wrap(agent)
+
+    result = await compressed.run([Message.user("Analyze AAPL")])
+    assert result.metadata.get("compressed") is True
+    assert result.metadata.get("savings_pct", 0) >= 0