contextpress 0.6.3__tar.gz → 0.6.4__tar.gz

{contextpress-0.6.3 → contextpress-0.6.4}/CHANGELOG.md

@@ -4,6 +4,16 @@ All notable changes to `contextpress` are recorded here.
 The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/)
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.6.4] - 2026-08-18
+
+- **OpenAI tool messages** — ``role: tool`` / ``function`` are valid; ``tool_calls``,
+  ``tool_call_id``, and ``name`` round-trip on dict messages.
+- Structure minifies JSON in ``tool_calls[].function.arguments`` and tool-result content.
+- Filler never drops empty assistant turns that only carry ``tool_calls``.
+- Budget removes an assistant ``tool_calls`` turn together with matching ``role=tool``
+  results (no orphan tool messages).
+- Example: `examples/openai_tools_compress.py`.
+
 ## [0.6.3] - 2026-08-12
 
 - **LangChain round-trip** — ``compress()`` maps remaining turns back onto their original

{contextpress-0.6.3 → contextpress-0.6.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: contextpress
-Version: 0.6.3
+Version: 0.6.4
 Summary: Deterministic context compression for LLM chat, RAG, and agent pipelines
 Project-URL: Homepage, https://github.com/Taha-azizi/contextpress
 Project-URL: Documentation, https://github.com/Taha-azizi/contextpress#readme
@@ -244,6 +244,8 @@ Description-Content-Type: text/markdown
 Deterministic context compression for LLM chat, RAG, and agent pipelines.
 Created and maintained by **[Taha Azizi](https://github.com/Taha-azizi)**.
 
+**Write-up:** [Introducing contextpress](https://pub.towardsai.net/introducing-contextpress-the-python-library-that-refactors-your-llm-context-c57965617edb) — Towards AI (Medium)
+
 ---
 
 ## Project Status
@@ -389,7 +391,7 @@ That script builds a long history and a tight `token_budget` so you can see turn
 
 - **chat** — Typical back-and-forth dialogue. Filler removal, repetition deduplication, resolution collapsing, recency weighting, and token budgets are tuned for conversational flow.
 - **rag_doc** — Document chunks or RAG context. Resolution is off; repetition compares all chunks; recency uses relevance to the latest user query instead of chat recency.
-- **agent** — Tool-using or task-oriented threads. Resolution can trigger on a single high-confidence completion signal; filler rules preserve tool-related turns when markers are present.
+- **agent** — Tool-using or task-oriented threads. Resolution can trigger on a single high-confidence completion signal; filler rules preserve tool-related turns when markers are present. OpenAI Chat Completions ``tool_calls`` / ``role: tool`` messages are first-class (0.6.4+).
 
 ```python
 ContextManager(type="chat")
@@ -398,6 +400,7 @@ ContextManager(type="agent")
 ```
 
 Runnable agent example: [`examples/agent_pipeline.py`](examples/agent_pipeline.py).
+OpenAI tools example: [`examples/openai_tools_compress.py`](examples/openai_tools_compress.py).
 
 ## Pipeline stages
 
@@ -406,7 +409,7 @@ Runnable agent example: [`examples/agent_pipeline.py`](examples/agent_pipeline.p
 3. **Repetition** — TF-IDF cosine similarity; keeps the more recent of similar turns.
 4. **Resolution** — Collapses agreed threads into a single `RESOLVED:` synthetic system turn (chat/agent only).
 5. **Recency** — Extractively compresses older turns (or low-relevance chunks in `rag_doc`) while preserving the latest context.
-6. **Budget** — Enforces a hard token limit with `tiktoken`, removing oldest turns first while protecting system prompts and recent turns.
+6. **Budget** — Enforces a hard token limit with `tiktoken`, removing oldest turns first while protecting system prompts and recent turns. Assistant ``tool_calls`` and matching ``role: tool`` results are dropped together (0.6.4+).
 
 **Cost estimate** (0.6+, approximate list prices for planning):
 

{contextpress-0.6.3 → contextpress-0.6.4}/README.md

@@ -3,6 +3,8 @@
 Deterministic context compression for LLM chat, RAG, and agent pipelines.
 Created and maintained by **[Taha Azizi](https://github.com/Taha-azizi)**.
 
+**Write-up:** [Introducing contextpress](https://pub.towardsai.net/introducing-contextpress-the-python-library-that-refactors-your-llm-context-c57965617edb) — Towards AI (Medium)
+
 ---
 
 ## Project Status
@@ -148,7 +150,7 @@ That script builds a long history and a tight `token_budget` so you can see turn
 
 - **chat** — Typical back-and-forth dialogue. Filler removal, repetition deduplication, resolution collapsing, recency weighting, and token budgets are tuned for conversational flow.
 - **rag_doc** — Document chunks or RAG context. Resolution is off; repetition compares all chunks; recency uses relevance to the latest user query instead of chat recency.
-- **agent** — Tool-using or task-oriented threads. Resolution can trigger on a single high-confidence completion signal; filler rules preserve tool-related turns when markers are present.
+- **agent** — Tool-using or task-oriented threads. Resolution can trigger on a single high-confidence completion signal; filler rules preserve tool-related turns when markers are present. OpenAI Chat Completions ``tool_calls`` / ``role: tool`` messages are first-class (0.6.4+).
 
 ```python
 ContextManager(type="chat")
@@ -157,6 +159,7 @@ ContextManager(type="agent")
 ```
 
 Runnable agent example: [`examples/agent_pipeline.py`](examples/agent_pipeline.py).
+OpenAI tools example: [`examples/openai_tools_compress.py`](examples/openai_tools_compress.py).
 
 ## Pipeline stages
 
@@ -165,7 +168,7 @@ Runnable agent example: [`examples/agent_pipeline.py`](examples/agent_pipeline.p
 3. **Repetition** — TF-IDF cosine similarity; keeps the more recent of similar turns.
 4. **Resolution** — Collapses agreed threads into a single `RESOLVED:` synthetic system turn (chat/agent only).
 5. **Recency** — Extractively compresses older turns (or low-relevance chunks in `rag_doc`) while preserving the latest context.
-6. **Budget** — Enforces a hard token limit with `tiktoken`, removing oldest turns first while protecting system prompts and recent turns.
+6. **Budget** — Enforces a hard token limit with `tiktoken`, removing oldest turns first while protecting system prompts and recent turns. Assistant ``tool_calls`` and matching ``role: tool`` results are dropped together (0.6.4+).
 
 **Cost estimate** (0.6+, approximate list prices for planning):
 

{contextpress-0.6.3 → contextpress-0.6.4}/ROADMAP.md

@@ -41,5 +41,6 @@ deterministic Tier‑1 NLP for chat / RAG / agent **message histories**, with op
 | **0.6.1** | Wire estimated USD into `CompressionStats` / reports — shipped |
 | **0.6.2** | Agent-oriented fixtures for JSON/tool payloads; `summary()` report — shipped |
 | **0.6.3** | LangChain compress round-trip; `output_tokens` on cost stats / `summary()` — shipped |
+| **0.6.4** | OpenAI `tool_calls` / `role=tool` round-trip, JSON minify, budget pair integrity — shipped |
 
 Stay classical-NLP-first; keep optional LLM extras optional.

{contextpress-0.6.3 → contextpress-0.6.4}/contextpress/__init__.py

@@ -16,7 +16,7 @@ __all__ = [
     "CompressionResult",
     "CompressionStats",
 ]
-__version__ = "0.6.3"
+__version__ = "0.6.4"
 
 
 def __getattr__(name: str) -> Any:

{contextpress-0.6.3 → contextpress-0.6.4}/contextpress/models.py

@@ -25,7 +25,7 @@ class Turn:
     This is the canonical unit the entire pipeline operates on.
     """
 
-    role: str  # "user" | "assistant" | "system"
+    role: str  # "user" | "assistant" | "system" | "tool" | "function"
     content: str | list[ContentBlock]  # string for simple, list for multimodal
     timestamp: datetime | None = None
     metadata: dict[str, Any] = field(default_factory=dict)

{contextpress-0.6.3 → contextpress-0.6.4}/contextpress/normalizer.py

@@ -12,11 +12,14 @@ from typing import Any
 
 from contextpress.models import ContentBlock, Conversation, Turn
 
-_VALID_ROLES = frozenset({"user", "assistant", "system"})
+_VALID_ROLES = frozenset({"user", "assistant", "system", "tool", "function"})
+_TOOL_COPY_KEYS = ("tool_calls", "tool_call_id", "name")
 _LC_TYPE_MAP = {
     "human": "user",
     "ai": "assistant",
     "system": "system",
+    "tool": "tool",
+    "function": "function",
 }
 
 
@@ -201,6 +204,9 @@ def normalize_messages(
             ts = _parse_timestamp(d)
             content = d.get("content", "")
             meta: dict[str, Any] = {"_dict_index": i, "_original_dict": d}
+            for key in _TOOL_COPY_KEYS:
+                if key in d:
+                    meta[key] = copy.deepcopy(d[key])
 
             if isinstance(content, list):
                 blocks = _blocks_from_openai_style(content)
@@ -282,8 +288,17 @@ def denormalize_output(conversation: Conversation, ctx: dict[str, Any]) -> Any:
             base["role"] = t.role
         if isinstance(t.content, list):
             base["content"] = _blocks_to_openai_style(t.content)
+        elif (
+            t.content == ""
+            and isinstance(t.metadata.get("_original_dict"), dict)
+            and t.metadata["_original_dict"].get("content") is None
+        ):
+            base["content"] = None
         else:
             base["content"] = t.content
+        for key in _TOOL_COPY_KEYS:
+            if key in t.metadata:
+                base[key] = copy.deepcopy(t.metadata[key])
         out_dicts.append(base)
     return out_dicts
 

{contextpress-0.6.3 → contextpress-0.6.4}/contextpress/stats.py

@@ -10,6 +10,7 @@ import tiktoken
 from contextpress.costs import estimate_token_cost
 from contextpress.models import Conversation, Turn
 from contextpress.normalizer import extract_text_for_processing
+from contextpress.tools import tool_payload_text
 
 
 def get_encoding(model: str | None) -> tiktoken.Encoding:
@@ -26,6 +27,9 @@ def count_turn_tokens(turn: Turn, encoding: tiktoken.Encoding) -> int:
         body = turn.content
     else:
         body = extract_text_for_processing(turn)
+    extra = tool_payload_text(turn)
+    if extra:
+        body = f"{body}\n{extra}" if body else extra
     return len(encoding.encode(f"{turn.role}\n{body}"))
 
 

{contextpress-0.6.3 → contextpress-0.6.4}/contextpress/strategies/budget.py

@@ -8,6 +8,7 @@ import tiktoken
 from contextpress.models import Conversation, Turn
 from contextpress.stats import count_turn_tokens, get_encoding
 from contextpress.strategies.base import BaseStrategy
+from contextpress.tools import tool_group_indices
 
 
 def _truncate_system_turn(turn: Turn, encoding: tiktoken.Encoding, max_tokens: int) -> Turn:
@@ -73,9 +74,17 @@ class BudgetStrategy(BaseStrategy):
             keep = min(2, len(ns_positions))
             protected = set(ns_positions[-keep:]) if keep else set()
             removable = [i for i in ns_positions if i not in protected]
-            if removable:
-                turns.pop(removable[0])
-                n_removed += 1
+            group: list[int] | None = None
+            for idx in removable:
+                candidate = tool_group_indices(turns, idx)
+                if any(g in protected for g in candidate):
+                    continue
+                group = candidate
+                break
+            if group is not None:
+                for g in sorted(group, reverse=True):
+                    turns.pop(g)
+                    n_removed += 1
                 continue
 
             # Last resort: truncate system (see behavior contract note on invariant 1)

{contextpress-0.6.3 → contextpress-0.6.4}/contextpress/strategies/filler.py

@@ -6,6 +6,7 @@ import re
 from contextpress.models import Conversation, Turn
 from contextpress.normalizer import apply_text_to_turn, extract_text_for_processing
 from contextpress.strategies.base import BaseStrategy
+from contextpress.tools import has_tool_marker
 
 # Curated lists — longer phrases first for safe replacement order
 FILLER_PHRASES = [
@@ -70,16 +71,6 @@ ACKNOWLEDGEMENT_PHRASES = [
     "thank you for that",
 ]
 
-_TOOL_MARKERS = ("tool_calls", "tool_call", "tool_use", "tool_result", "<tool", "[tool")
-
-
-def _has_tool_marker(turn: Turn) -> bool:
-    meta = turn.metadata or {}
-    if any(k in meta for k in ("tool_calls", "tool_call", "tool_use", "tool_result")):
-        return True
-    text = extract_text_for_processing(turn).lower()
-    return any(m in text for m in _TOOL_MARKERS)
-
 
 def _build_filler_pattern() -> re.Pattern[str]:
     pattern_parts: list[str] = []
@@ -155,7 +146,7 @@ class FillerStrategy(BaseStrategy):
                     new_turns.append(nt)
                 continue
 
-            if self.conv_type == "agent" and _has_tool_marker(turn):
+            if has_tool_marker(turn):
                 new_turns.append(copy.deepcopy(turn))
                 continue
 

{contextpress-0.6.3 → contextpress-0.6.4}/contextpress/strategies/structure.py

@@ -9,6 +9,7 @@ import re
 from contextpress.models import Conversation, Turn
 from contextpress.normalizer import apply_text_to_turn, extract_text_for_processing
 from contextpress.strategies.base import BaseStrategy
+from contextpress.tools import minify_tool_fields
 
 _CODE_FENCE = re.compile(r"(```[\s\S]*?```)", re.MULTILINE)
 _MULTI_BLANK = re.compile(r"\n{3,}")
@@ -85,9 +86,10 @@ class StructureStrategy(BaseStrategy):
             text = extract_text_for_processing(turn)
             compacted = compact_structure_text(text, aggressiveness=self.aggressiveness)
             if compacted != text:
-                new_turns.append(apply_text_to_turn(turn, compacted))
+                nt = apply_text_to_turn(turn, compacted)
             else:
-                new_turns.append(copy.deepcopy(turn))
+                nt = copy.deepcopy(turn)
+            new_turns.append(minify_tool_fields(nt))
         return Conversation(
             turns=new_turns,
             type=conversation.type,

contextpress-0.6.4/contextpress/tools.py

@@ -0,0 +1,181 @@
+"""Helpers for OpenAI-style tool_calls / role=tool messages (0.6.4+)."""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+from contextpress.models import Turn
+from contextpress.normalizer import extract_text_for_processing
+
+_TOOL_META_KEYS = ("tool_calls", "tool_call", "tool_use", "tool_result", "tool_call_id")
+_TEXT_MARKERS = ("tool_calls", "tool_call", "tool_use", "tool_result", "<tool", "[tool")
+_RESULT_ROLES = frozenset({"tool", "function"})
+
+
+def _layers(turn: Turn) -> list[dict[str, Any]]:
+    meta = turn.metadata or {}
+    layers: list[dict[str, Any]] = [meta]
+    orig = meta.get("_original_dict")
+    if isinstance(orig, dict) and orig is not meta:
+        layers.append(orig)
+    return layers
+
+
+def has_tool_marker(turn: Turn) -> bool:
+    """True if this turn is a tool call, tool result, or has a textual tool marker."""
+    if turn.role in _RESULT_ROLES:
+        return True
+    for layer in _layers(turn):
+        if any(k in layer for k in _TOOL_META_KEYS):
+            return True
+    text = extract_text_for_processing(turn).lower()
+    return any(m in text for m in _TEXT_MARKERS)
+
+
+def assistant_tool_call_ids(turn: Turn) -> list[str]:
+    ids: list[str] = []
+    seen: set[str] = set()
+    for layer in _layers(turn):
+        calls = layer.get("tool_calls")
+        if not isinstance(calls, list):
+            continue
+        for call in calls:
+            if not isinstance(call, dict):
+                continue
+            cid = call.get("id")
+            if cid is None:
+                continue
+            sid = str(cid)
+            if sid not in seen:
+                seen.add(sid)
+                ids.append(sid)
+        break
+    return ids
+
+
+def result_tool_call_id(turn: Turn) -> str | None:
+    for layer in _layers(turn):
+        tid = layer.get("tool_call_id")
+        if tid is not None:
+            return str(tid)
+    return None
+
+
+def tool_payload_text(turn: Turn) -> str:
+    """Extra text for token counting (tool_calls JSON, not already in content)."""
+    for layer in _layers(turn):
+        calls = layer.get("tool_calls")
+        if isinstance(calls, list) and calls:
+            return json.dumps(calls, ensure_ascii=False)
+    return ""
+
+
+def _minify_json_string(text: str) -> str:
+    s = text.strip()
+    if not s or s[0] not in "{[":
+        return text
+    try:
+        obj = json.loads(s)
+    except (json.JSONDecodeError, TypeError, ValueError):
+        return text
+    return json.dumps(obj, separators=(",", ":"), ensure_ascii=False)
+
+
+def minify_tool_fields(turn: Turn) -> Turn:
+    """Minify JSON in tool_calls[].function.arguments; write back to metadata."""
+    meta = turn.metadata
+    if not meta:
+        return turn
+    changed = False
+    new_meta = dict(meta)
+
+    def _compact_calls(calls: Any) -> Any:
+        nonlocal changed
+        if not isinstance(calls, list):
+            return calls
+        out = []
+        for call in calls:
+            if not isinstance(call, dict):
+                out.append(call)
+                continue
+            c = dict(call)
+            fn = c.get("function")
+            if isinstance(fn, dict):
+                fn = dict(fn)
+                args = fn.get("arguments")
+                if isinstance(args, str):
+                    mini = _minify_json_string(args)
+                    if mini != args:
+                        fn["arguments"] = mini
+                        changed = True
+                c["function"] = fn
+            out.append(c)
+        return out
+
+    if "tool_calls" in new_meta:
+        new_meta["tool_calls"] = _compact_calls(new_meta["tool_calls"])
+    orig = new_meta.get("_original_dict")
+    if isinstance(orig, dict):
+        orig = dict(orig)
+        if "tool_calls" in orig:
+            orig["tool_calls"] = _compact_calls(orig["tool_calls"])
+        if turn.role in _RESULT_ROLES and isinstance(orig.get("content"), str):
+            target = turn.content if isinstance(turn.content, str) else orig["content"]
+            mini = _minify_json_string(target)
+            if mini != orig["content"]:
+                orig["content"] = mini
+                changed = True
+        new_meta["_original_dict"] = orig
+
+    if not changed:
+        return turn
+    return Turn(
+        role=turn.role,
+        content=turn.content,
+        timestamp=turn.timestamp,
+        metadata=new_meta,
+        importance=turn.importance,
+        resolved=turn.resolved,
+        compressed=True,
+        original_content=(
+            turn.original_content if turn.original_content is not None else turn.content
+        ),
+    )
+
+
+def tool_group_indices(turns: list[Turn], index: int) -> list[int]:
+    """Indices of an assistant tool_calls turn plus following matching tool results.
+
+    If ``index`` is a tool result, walk back to the parent assistant first.
+    """
+    if index < 0 or index >= len(turns):
+        return [index]
+    t = turns[index]
+    if t.role in _RESULT_ROLES:
+        k = index - 1
+        while k >= 0:
+            if turns[k].role == "assistant" and assistant_tool_call_ids(turns[k]):
+                rid = result_tool_call_id(t)
+                ids = set(assistant_tool_call_ids(turns[k]))
+                if rid is None or rid in ids:
+                    return tool_group_indices(turns, k)
+                break
+            if turns[k].role not in _RESULT_ROLES:
+                break
+            k -= 1
+        return [index]
+
+    ids = set(assistant_tool_call_ids(t))
+    if t.role != "assistant" or not ids:
+        return [index]
+    group = [index]
+    j = index + 1
+    while j < len(turns) and turns[j].role in _RESULT_ROLES:
+        rid = result_tool_call_id(turns[j])
+        if rid is None or rid in ids:
+            group.append(j)
+        else:
+            break
+        j += 1
+    return group

contextpress-0.6.4/examples/openai_tools_compress.py

@@ -0,0 +1,57 @@
+"""OpenAI Chat Completions tool_calls round-trip (0.6.4+)."""
+
+from __future__ import annotations
+
+import json
+
+from contextpress import ContextManager
+
+arguments = json.dumps(
+    {"service": "api-v2", "environment": "staging", "limit": 20},
+    indent=2,
+)
+tool_result = json.dumps(
+    {
+        "service": "api-v2",
+        "events": [{"id": "evt-001", "version": "2.4.1", "status": "healthy"}],
+    },
+    indent=2,
+)
+
+messages = [
+    {"role": "system", "content": "You are a deploy agent with function tools."},
+    {"role": "user", "content": "Find recent staging deploys for api-v2."},
+    {
+        "role": "assistant",
+        "content": None,
+        "tool_calls": [
+            {
+                "id": "call_search_1",
+                "type": "function",
+                "function": {"name": "search_deploys", "arguments": arguments},
+            }
+        ],
+    },
+    {
+        "role": "tool",
+        "tool_call_id": "call_search_1",
+        "name": "search_deploys",
+        "content": tool_result,
+    },
+    {"role": "assistant", "content": "Staging is healthy on 2.4.1."},
+]
+
+cm = ContextManager(type="agent", compression="medium", cost_provider="openai")
+result = cm.compress(messages, token_budget=None, return_stats=True)
+
+print(result.summary())
+print()
+for m in result.messages:
+    role = m["role"]
+    if m.get("tool_calls"):
+        args = m["tool_calls"][0]["function"]["arguments"]
+        print(f"{role} tool_calls arguments:", args[:80])
+    elif role == "tool":
+        print(f"{role} ({m.get('tool_call_id')}):", str(m.get("content", ""))[:80])
+    else:
+        print(f"{role}:", str(m.get("content", ""))[:80])

{contextpress-0.6.3 → contextpress-0.6.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "contextpress"
-version = "0.6.3"
+version = "0.6.4"
 description = "Deterministic context compression for LLM chat, RAG, and agent pipelines"
 readme = "README.md"
 license = { file = "LICENSE" }

contextpress-0.6.4/tests/fixtures/chats/12_openai_tools.json

@@ -0,0 +1,34 @@
+{
+  "id": "12_openai_tools",
+  "type": "agent",
+  "source": "synthetic (OpenAI Chat Completions tool_calls + role=tool)",
+  "messages": [
+    {"role": "system", "content": "You are a deploy agent with function tools."},
+    {"role": "user", "content": "Find recent staging deploys for api-v2."},
+    {
+      "role": "assistant",
+      "content": null,
+      "tool_calls": [
+        {
+          "id": "call_search_1",
+          "type": "function",
+          "function": {
+            "name": "search_deploys",
+            "arguments": "{\n  \"service\": \"api-v2\",\n  \"environment\": \"staging\",\n  \"limit\": 20\n}"
+          }
+        }
+      ]
+    },
+    {
+      "role": "tool",
+      "tool_call_id": "call_search_1",
+      "name": "search_deploys",
+      "content": "{\n  \"service\": \"api-v2\",\n  \"environment\": \"staging\",\n  \"events\": [\n    {\n      \"id\": \"evt-001\",\n      \"version\": \"2.4.1\",\n      \"status\": \"healthy\"\n    }\n  ]\n}"
+    },
+    {
+      "role": "assistant",
+      "content": "Staging is healthy on 2.4.1."
+    },
+    {"role": "user", "content": "Schedule a follow-up check."}
+  ]
+}

{contextpress-0.6.3 → contextpress-0.6.4}/tests/fixtures/chats/README.md

@@ -16,3 +16,4 @@ No network required — used by `tests/test_fixture_chats.py`.
 | `09_agent_tool_json.json` | agent | Large pretty-printed tool JSON |
 | `10_agent_repeated_logs.json` | agent | Repeated log lines in tool output |
 | `11_agent_mixed.json` | agent | Tool call + result + follow-up thread |
+| `12_openai_tools.json` | agent | OpenAI `tool_calls` + `role=tool` |

{contextpress-0.6.3 → contextpress-0.6.4}/tests/test_fixture_chats.py

@@ -58,7 +58,7 @@ def test_fixture_recommend_preset(path: Path):
 
 
 def test_fixture_catalog_not_empty():
-    assert len(FIXTURES) >= 11
+    assert len(FIXTURES) >= 12
 
 
 def test_c3_ack_may_weaken_resolution_signal():

{contextpress-0.6.3 → contextpress-0.6.4}/tests/test_v062.py

@@ -95,7 +95,11 @@ def test_agent_repeated_logs_dedupes_lines():
     cm = ContextManager(type="agent", compression="medium")
     result = cm.compress(data["messages"], token_budget=None, return_stats=True)
     assert "structure" in result.stats.stages_run
-    tool_turn = next(m for m in result.messages if m.get("role") == "user" and "Tool result" in str(m.get("content", "")))
+    tool_turn = next(
+        m
+        for m in result.messages
+        if m.get("role") == "user" and "Tool result" in str(m.get("content", ""))
+    )
     content = str(tool_turn["content"])
     assert content.count("payment gateway timeout") == 1
 

contextpress-0.6.4/tests/test_v064.py

@@ -0,0 +1,175 @@
+"""0.6.4 — OpenAI tool_calls / role=tool round-trip."""
+
+from __future__ import annotations
+
+import json
+import warnings
+from pathlib import Path
+
+from contextpress import ContextManager
+from contextpress.models import Conversation, Turn
+from contextpress.normalizer import normalize_messages
+from contextpress.strategies.budget import BudgetStrategy
+from contextpress.tools import has_tool_marker
+
+FIXTURE = Path(__file__).parent / "fixtures" / "chats" / "12_openai_tools.json"
+
+
+def _pretty_args() -> str:
+    return json.dumps({"service": "api-v2", "environment": "staging", "limit": 20}, indent=2)
+
+
+def _openai_tool_thread() -> list[dict]:
+    return [
+        {"role": "system", "content": "You are a deploy agent."},
+        {"role": "user", "content": "Find staging deploys."},
+        {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [
+                {
+                    "id": "call_search_1",
+                    "type": "function",
+                    "function": {"name": "search_deploys", "arguments": _pretty_args()},
+                }
+            ],
+        },
+        {
+            "role": "tool",
+            "tool_call_id": "call_search_1",
+            "name": "search_deploys",
+            "content": json.dumps({"ok": True, "version": "2.4.1"}, indent=2),
+        },
+        {"role": "assistant", "content": "Staging is healthy on 2.4.1."},
+        {"role": "user", "content": "Thanks, that is all."},
+    ]
+
+
+def _call_and_result_ids(messages: list[dict]) -> tuple[set[str], set[str]]:
+    calls: set[str] = set()
+    results: set[str] = set()
+    for m in messages:
+        if m.get("role") == "assistant":
+            for c in m.get("tool_calls") or []:
+                if isinstance(c, dict) and c.get("id"):
+                    calls.add(str(c["id"]))
+        if m.get("role") == "tool" and m.get("tool_call_id"):
+            results.add(str(m["tool_call_id"]))
+    return calls, results
+
+
+def test_tool_role_no_unknown_warning():
+    with warnings.catch_warnings(record=True) as caught:
+        warnings.simplefilter("always")
+        conv, _ = normalize_messages(_openai_tool_thread())
+    assert conv.turns[3].role == "tool"
+    assert not any("unknown role" in str(w.message).lower() for w in caught)
+
+
+def test_tool_keys_copied_to_metadata():
+    conv, _ = normalize_messages(_openai_tool_thread())
+    asst = conv.turns[2]
+    tool = conv.turns[3]
+    assert "tool_calls" in asst.metadata
+    assert asst.metadata["tool_calls"][0]["id"] == "call_search_1"
+    assert tool.metadata.get("tool_call_id") == "call_search_1"
+    assert tool.metadata.get("name") == "search_deploys"
+    assert has_tool_marker(asst)
+    assert has_tool_marker(tool)
+
+
+def test_has_tool_marker_via_original_dict_only():
+    t = Turn(
+        role="assistant",
+        content="",
+        metadata={
+            "_original_dict": {
+                "role": "assistant",
+                "content": None,
+                "tool_calls": [{"id": "x"}],
+            }
+        },
+    )
+    assert has_tool_marker(t)
+
+
+def test_empty_tool_calls_assistant_not_dropped():
+    messages = [
+        {"role": "user", "content": "search now"},
+        {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [
+                {
+                    "id": "call_1",
+                    "type": "function",
+                    "function": {"name": "search", "arguments": "{}"},
+                }
+            ],
+        },
+    ]
+    cm = ContextManager(type="agent", compression="high")
+    out = cm.compress(messages, token_budget=None)
+    assert any(m.get("tool_calls") for m in out)
+    assert out[1]["content"] is None
+
+
+def test_roundtrip_keeps_tool_keys_and_minifies_json():
+    messages = _openai_tool_thread()
+    original = json.loads(json.dumps(messages))
+    cm = ContextManager(type="agent", compression="medium")
+    out = cm.compress(messages, token_budget=None)
+    assert messages == original
+    asst = next(m for m in out if m.get("tool_calls"))
+    tool = next(m for m in out if m.get("role") == "tool")
+    args = asst["tool_calls"][0]["function"]["arguments"]
+    assert "\n" not in args
+    assert json.loads(args)["service"] == "api-v2"
+    assert "\n" not in tool["content"]
+    assert json.loads(tool["content"])["ok"] is True
+    assert tool["tool_call_id"] == "call_search_1"
+    assert tool["name"] == "search_deploys"
+
+
+def test_budget_does_not_orphan_tool_result():
+    padding = "deploy status note " * 40
+    turns = [
+        Turn(
+            role="assistant",
+            content="",
+            metadata={
+                "tool_calls": [
+                    {
+                        "id": "call_old",
+                        "type": "function",
+                        "function": {"name": "search", "arguments": "{}"},
+                    }
+                ]
+            },
+        ),
+        Turn(role="tool", content=padding, metadata={"tool_call_id": "call_old"}),
+        Turn(role="user", content="what is the latest status?"),
+        Turn(role="assistant", content="Latest is healthy."),
+    ]
+    c = Conversation(turns=turns, type="agent")
+    out = BudgetStrategy(token_budget=40, model=None).process(c)
+    remaining = [(t.role, (t.metadata or {}).get("tool_call_id")) for t in out.turns]
+    tool_ids = {t.metadata.get("tool_call_id") for t in out.turns if t.role == "tool"}
+    call_ids = set()
+    for t in out.turns:
+        for call in (t.metadata or {}).get("tool_calls") or []:
+            if isinstance(call, dict) and call.get("id"):
+                call_ids.add(call["id"])
+    assert tool_ids <= call_ids
+    assert remaining[-1][0] == "assistant"
+
+
+def test_fixture_openai_tools_compresses():
+    data = json.loads(FIXTURE.read_text(encoding="utf-8"))
+    cm = ContextManager(type="agent", compression="medium")
+    result = cm.compress(data["messages"], token_budget=None, return_stats=True)
+    calls, results = _call_and_result_ids(result.messages)
+    assert results <= calls
+    assert result.stats.tokens_after <= result.stats.tokens_before
+    asst = next(m for m in result.messages if m.get("tool_calls"))
+    assert "\n" not in asst["tool_calls"][0]["function"]["arguments"]