Fix CI: type errors, Windows test paths, drop upstream-only tests

Type Check (ty) fixes in new/overwritten TACO files:
  - tmux_session.py: widen logging_path to PurePath and allow None
    asciinema paths (EnvironmentPaths.agent_dir is PurePosixPath; TACO
    can run without asciinema recording); fix get_incremental_output's
    last_command default (str | None = None); guard AsciinemaHandler
    call with a not-None check on the local recording path.
  - terminus_2.py: annotate _run_subagent's message_history as
    list[dict[str, Any] | Message] to match LiteLLM.call; drop an
    obsolete # type: ignore comment.
  - output_filter.py: use walrus-assigned re.search so match.group is
    only called after a non-None check.
  - module/client.py: suppress openai SDK overload noise with a
    native # ty: ignore[no-matching-overload] on the 3 create() calls.

Test fixes / cleanups:
  - Fix Path(__file__).resolve().parents[N] in the 4 new compression
    tests (tests live at tests/unit/agents/terminus_2/, repo root is
    parents[4], not parents[3]).
  - Update the BibTeX citation in both READMEs to the official author
    list / @misc + eprint form.
  - Drop upstream-only tests that exercise agent features TACO does
    not ship (skills section, tmux send-keys chunking, _init_llm
    patching): the 3 TestTerminus2 classes in test_agent_skills.py,
    test_tmux_session.py, test_terminus_2_no_retry_on_cancelled.py.

Made-with: Cursor

Author: mod-eerf

README.md

@@ -88,11 +88,13 @@ All flags are passed on the CLI as `--ak <name>=<value>`.
 ## Citation
 
 ```bibtex
-@article{taco2026,
-  title   = {TACO: A Self-Evolving Framework for Efficient Terminal Agents
-             via Observational Context Compression},
-  author  = {TACO Team},
-  journal = {arXiv preprint arXiv:2604.19572},
-  year    = {2026}
+@misc{ren2026selfevolvingframeworkefficientterminal,
+      title={A Self-Evolving Framework for Efficient Terminal Agents via Observational Context Compression},
+      author={Jincheng Ren and Siwei Wu and Yizhi Li and Kang Zhu and Shu Xu and Boyu Feng and Ruibin Yuan and Wei Zhang and Riza Batista-Navarro and Jian Yang and Chenghua Lin},
+      year={2026},
+      eprint={2604.19572},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL},
+      url={https://arxiv.org/abs/2604.19572},
 }
 ```

src/harbor/agents/terminus_2/README.md

@@ -88,11 +88,13 @@ All flags are passed on the CLI as `--ak <name>=<value>`.
 ## Citation
 
 ```bibtex
-@article{taco2026,
-  title   = {TACO: A Self-Evolving Framework for Efficient Terminal Agents
-             via Observational Context Compression},
-  author  = {TACO Team},
-  journal = {arXiv preprint arXiv:2604.19572},
-  year    = {2026}
+@misc{ren2026selfevolvingframeworkefficientterminal,
+      title={A Self-Evolving Framework for Efficient Terminal Agents via Observational Context Compression},
+      author={Jincheng Ren and Siwei Wu and Yizhi Li and Kang Zhu and Shu Xu and Boyu Feng and Ruibin Yuan and Wei Zhang and Riza Batista-Navarro and Jian Yang and Chenghua Lin},
+      year={2026},
+      eprint={2604.19572},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL},
+      url={https://arxiv.org/abs/2604.19572},
 }
 ```

src/harbor/agents/terminus_2/module/client.py

@@ -294,7 +294,7 @@ def chat(
                     if key not in ['messages', 'reasoning_effort']:
                         completions_params[key] = value
 
-                response = self.client.completions.create(**completions_params)
+                response = self.client.completions.create(**completions_params)  # ty: ignore[no-matching-overload]
 
                 # Extract response content
                 if response.choices and len(response.choices) > 0:
@@ -305,7 +305,7 @@ def chat(
                     raise LLMClientError("Invalid response format: no choices in response")
             else:
                 # /v1/chat/completions endpoint (server applies chat template).
-                response = self.client.chat.completions.create(**request_params)
+                response = self.client.chat.completions.create(**request_params)  # ty: ignore[no-matching-overload]
 
                 # Extract response content
                 if response.choices and len(response.choices) > 0:
@@ -368,7 +368,7 @@ def chat_with_usage(
         request_params.update(kwargs)
 
         try:
-            response = self.client.chat.completions.create(**request_params)
+            response = self.client.chat.completions.create(**request_params)  # ty: ignore[no-matching-overload]
 
             # Extract content
             content = ""

src/harbor/agents/terminus_2/output_filter.py

@@ -362,11 +362,9 @@ def _extract_progress(
             extracted.append("Reading package lists")
         elif "Building dependency tree" in tail:
             extracted.append("Building dependency tree")
-        elif re.search(r"Setting up\s+(\S+)", tail):
-            match = re.search(r"Setting up\s+(\S+)", tail)
+        elif (match := re.search(r"Setting up\s+(\S+)", tail)) is not None:
             extracted.append(f"Setting up {match.group(1)}")
-        elif re.search(r"Unpacking\s+(\S+)", tail):
-            match = re.search(r"Unpacking\s+(\S+)", tail)
+        elif (match := re.search(r"Unpacking\s+(\S+)", tail)) is not None:
             extracted.append(f"Unpacking {match.group(1)}")
 
         if extracted:

src/harbor/agents/terminus_2/terminus_2.py

@@ -6,7 +6,9 @@
 from dataclasses import dataclass
 from datetime import datetime, timezone
 from pathlib import Path
-from typing import Literal
+from typing import Any, Literal
+
+from litellm import Message
 
 from tenacity import retry, stop_after_attempt
 
@@ -896,7 +898,7 @@ def _track_api_request_time(self, start_time: float) -> None:
     async def _run_subagent(
         self,
         prompt: str,
-        message_history: list[dict],
+        message_history: list[dict[str, Any] | Message],
         steps: list[Step],
         session_id: str,
         agent_name: str,
@@ -1330,7 +1332,7 @@ async def _query_llm(
 
             if hasattr(self._parser, "salvage_truncated_response"):
                 salvaged_response, has_multiple_blocks = (
-                    self._parser.salvage_truncated_response(truncated_response)  # type: ignore
+                    self._parser.salvage_truncated_response(truncated_response)
                 )
 
             if salvaged_response:

src/harbor/agents/terminus_2/tmux_session.py

@@ -2,7 +2,7 @@
 import re
 import shlex
 import time
-from pathlib import Path
+from pathlib import Path, PurePath
 
 from harbor.agents.terminus_2.asciinema_handler import AsciinemaHandler
 from harbor.environments.base import BaseEnvironment
@@ -25,9 +25,9 @@ def __init__(
         self,
         session_name: str,
         environment: BaseEnvironment,
-        logging_path: Path,
-        local_asciinema_recording_path: Path,
-        remote_asciinema_recording_path: Path,
+        logging_path: PurePath,
+        local_asciinema_recording_path: Path | None,
+        remote_asciinema_recording_path: PurePath | None,
         pane_width: int = 160,
         pane_height: int = 40,
     ):
@@ -392,8 +392,9 @@ async def stop(self) -> None:
 
             await asyncio.sleep(0.5)
 
-            # Merge markers into the recording
-            if self._markers:
+            # Merge markers into the recording. remote and local recording paths
+            # are always set or unset as a pair, so the local path is non-None here.
+            if self._markers and self._local_asciinema_recording_path is not None:
                 self._logger.debug(
                     f"Merging {len(self._markers)} markers into recording"
                 )
@@ -571,7 +572,7 @@ async def _find_new_content(self, current_buffer: str) -> str | None:
             return current_buffer[idx:]
         return None
 
-    async def get_incremental_output(self, last_command: str = None) -> str:
+    async def get_incremental_output(self, last_command: str | None = None) -> str:
         """
         Get either new terminal output since last call, or current screen if
         unable to determine.

tests/integration/test_terminus_2_no_retry_on_cancelled.py

@@ -303,135 +303,3 @@ async def _exec(command, timeout_sec=None):
 
     env.exec = AsyncMock(side_effect=_exec)
     return env
-
-
-class TestTerminus2Skills:
-    """Test _build_skills_section() for Terminus2."""
-
-    def _make_agent(self, temp_dir, skills_dir=None):
-        from harbor.agents.terminus_2.terminus_2 import Terminus2
-
-        return Terminus2(
-            logs_dir=temp_dir,
-            model_name="anthropic/claude-sonnet-4-5",
-            skills_dir=skills_dir,
-        )
-
-    async def test_no_skills_dir_returns_none(self, temp_dir):
-        agent = self._make_agent(temp_dir)
-        env = _make_mock_environment()
-        assert await agent._build_skills_section(env) is None
-
-    async def test_nonexistent_skills_dir_returns_none(self, temp_dir):
-        agent = self._make_agent(temp_dir, skills_dir="/nonexistent/path")
-        env = _make_mock_environment(is_dir=False)
-        assert await agent._build_skills_section(env) is None
-
-    async def test_empty_skills_dir_returns_none(self, temp_dir):
-        agent = self._make_agent(temp_dir, skills_dir="/skills")
-        env = _make_mock_environment(skills=None)
-        assert await agent._build_skills_section(env) is None
-
-    async def test_skills_dir_with_valid_skill(self, temp_dir):
-        content = SKILL_FRONTMATTER.format(
-            name="greet", description="Say hello to the user.", body="Do it."
-        )
-        agent = self._make_agent(temp_dir, skills_dir="/skills")
-        env = _make_mock_environment(skills={"greet": content})
-        result = await agent._build_skills_section(env)
-        assert result is not None
-        assert "<available_skills>" in result
-        assert "<name>greet</name>" in result
-        assert "<description>Say hello to the user.</description>" in result
-        assert "<location>/skills/greet/SKILL.md</location>" in result
-
-    async def test_multiple_skills_sorted(self, temp_dir):
-        skills = {
-            "zeta": SKILL_FRONTMATTER.format(
-                name="zeta", description="Zeta skill.", body=""
-            ),
-            "alpha": SKILL_FRONTMATTER.format(
-                name="alpha", description="Alpha skill.", body=""
-            ),
-            "mid": SKILL_FRONTMATTER.format(
-                name="mid", description="Mid skill.", body=""
-            ),
-        }
-        agent = self._make_agent(temp_dir, skills_dir="/skills")
-        env = _make_mock_environment(skills=skills)
-        result = await agent._build_skills_section(env)
-        assert result is not None
-        alpha_pos = result.index("<name>alpha</name>")
-        mid_pos = result.index("<name>mid</name>")
-        zeta_pos = result.index("<name>zeta</name>")
-        assert alpha_pos < mid_pos < zeta_pos
-
-    async def test_skips_invalid_frontmatter(self, temp_dir):
-        """SKILL.md without valid YAML frontmatter is ignored."""
-        agent = self._make_agent(temp_dir, skills_dir="/skills")
-        env = _make_mock_environment(skills={"bad-skill": "No frontmatter here."})
-        assert await agent._build_skills_section(env) is None
-
-
-class TestTerminus2ParseSkillFrontmatter:
-    """Test _parse_skill_frontmatter() directly."""
-
-    def test_valid_frontmatter(self):
-        from harbor.agents.terminus_2.terminus_2 import Terminus2
-
-        content = "---\nname: my-skill\ndescription: Does things.\n---\nBody.\n"
-        result = Terminus2._parse_skill_frontmatter(content)
-        assert result == {"name": "my-skill", "description": "Does things."}
-
-    def test_missing_name(self):
-        from harbor.agents.terminus_2.terminus_2 import Terminus2
-
-        content = "---\ndescription: No name field.\n---\nBody.\n"
-        assert Terminus2._parse_skill_frontmatter(content) is None
-
-    def test_no_frontmatter_delimiter(self):
-        from harbor.agents.terminus_2.terminus_2 import Terminus2
-
-        assert (
-            Terminus2._parse_skill_frontmatter("Just markdown, no frontmatter.") is None
-        )
-
-    def test_frontmatter_with_dashes_in_yaml_value(self):
-        """Ensure --- inside a YAML value does not break frontmatter parsing."""
-        from harbor.agents.terminus_2.terminus_2 import Terminus2
-
-        content = '---\nname: my-skill\ndescription: "Use --- to separate sections"\n---\nBody.\n'
-        result = Terminus2._parse_skill_frontmatter(content)
-        assert result is not None
-        assert result["name"] == "my-skill"
-        assert result["description"] == "Use --- to separate sections"
-
-
-class TestTerminus2SkillsXmlEscaping:
-    """Test that XML special characters are properly escaped in skills output."""
-
-    def _make_agent(self, temp_dir, skills_dir=None):
-        from harbor.agents.terminus_2.terminus_2 import Terminus2
-
-        return Terminus2(
-            logs_dir=temp_dir,
-            model_name="anthropic/claude-sonnet-4-5",
-            skills_dir=skills_dir,
-        )
-
-    async def test_xml_special_chars_escaped(self, temp_dir):
-        """Skill name/description with <, >, & must be escaped in XML output."""
-        content = SKILL_FRONTMATTER.format(
-            name="A<B>&C",
-            description='Use <tag> & "quotes"',
-            body="Body.",
-        )
-        agent = self._make_agent(temp_dir, skills_dir="/skills")
-        env = _make_mock_environment(skills={"special": content})
-        result = await agent._build_skills_section(env)
-        assert result is not None
-        assert "<available_skills>" in result
-        # Raw < > & must not appear unescaped inside text content
-        assert "A&lt;B&gt;&amp;C" in result
-        assert "&lt;tag&gt;" in result
-        assert "&amp;" in result