Skip to content

Self-hosted AI workspace.

Python
19.5 KB
Raw Blame
"""Tests for context_compactor.py — constants and prompt templates.
Uses mock imports to avoid loading the full app stack."""

import asyncio
import sys
from unittest.mock import MagicMock

import pytest

# Mock heavy dependencies before importing
for mod in [
    'sqlalchemy', 'sqlalchemy.orm', 'sqlalchemy.ext', 'sqlalchemy.ext.declarative',
    'sqlalchemy.ext.hybrid', 'sqlalchemy.sql', 'sqlalchemy.sql.expression',
    'src.database',
    'core.models', 'core.database',
]:
    if mod not in sys.modules:
        sys.modules[mod] = MagicMock()

import src.context_compactor as cc
from src.context_compactor import (
    COMPACT_THRESHOLD,
    SELF_SUMMARY_SYSTEM_PROMPT,
    SUMMARY_MAX_TOKENS,
    _content_as_text,
    maybe_compact,
    trim_for_context,
)
from src.model_context import estimate_text_tokens, estimate_tokens


class TestCompactThreshold:
    def test_value(self):
        assert COMPACT_THRESHOLD == 0.85

    def test_summary_max_tokens(self):
        assert SUMMARY_MAX_TOKENS == 1024


class TestSelfSummaryPrompt:
    def test_contains_goal_section(self):
        assert "### User Goal" in SELF_SUMMARY_SYSTEM_PROMPT

    def test_contains_what_was_done_section(self):
        assert "### What Was Done" in SELF_SUMMARY_SYSTEM_PROMPT

    def test_contains_current_state_section(self):
        assert "### Current State" in SELF_SUMMARY_SYSTEM_PROMPT

    def test_contains_pending_section(self):
        assert "### Pending / Next Steps" in SELF_SUMMARY_SYSTEM_PROMPT

    def test_contains_key_context_section(self):
        assert "### Key Context" in SELF_SUMMARY_SYSTEM_PROMPT

    def test_count_placeholder(self):
        assert "{count}" in SELF_SUMMARY_SYSTEM_PROMPT

    def test_n_placeholder(self):
        assert "{n}" in SELF_SUMMARY_SYSTEM_PROMPT

    def test_mentions_compactions(self):
        assert "Compactions so far" in SELF_SUMMARY_SYSTEM_PROMPT


class TestTrimForContext:
    def test_dense_unicode_current_message_is_trimmed_to_budget(self):
        messages = [
            {"role": "system", "content": "Keep the active request."},
            {"role": "user", "content": "\ufffd" * 5000},
        ]

        trimmed = trim_for_context(messages, context_length=1024, reserve_tokens=256)

        assert trimmed[-1]["role"] == "user"
        assert "pasted message was too large" in trimmed[-1]["content"]
        assert estimate_tokens(trimmed) <= 768

    def test_dense_unicode_tool_arguments_use_shared_estimator(self):
        messages = [
            {"role": "system", "content": "Use tools."},
            {"role": "user", "content": "Create the artifact."},
            {
                "role": "assistant",
                "content": None,
                "tool_calls": [{
                    "id": "call-1",
                    "type": "function",
                    "function": {
                        "name": "write_file",
                        "arguments": "\ufffd" * 5000,
                    },
                }],
            },
            {"role": "tool", "tool_call_id": "call-1", "content": "created"},
        ]

        trimmed = trim_for_context(messages, context_length=1024, reserve_tokens=256)

        assistant = next(message for message in trimmed if message.get("tool_calls"))
        args = assistant["tool_calls"][0]["function"]["arguments"]
        assert "_truncated_for_context" in args
        assert estimate_tokens(trimmed) <= 768

    def test_system_truncation_preserves_internal_route_metadata(self):
        messages = [
            {
                "role": "system",
                "content": "persona\n\n" + ("agent prompt " * 2000),
                "_agent_injected": "merged_prompt",
                "_agent_base_message": {"role": "system", "content": "persona"},
            },
            {"role": "user", "content": "latest"},
        ]

        trimmed = trim_for_context(messages, context_length=1024, reserve_tokens=256)

        system = next(message for message in trimmed if message.get("role") == "system")
        assert system["_agent_injected"] == "merged_prompt"
        assert system["_agent_base_message"] == {"role": "system", "content": "persona"}

    def test_compaction_summary_survives_large_route_prompt_trim(self):
        messages = [
            {"role": "system", "content": "route prompt " * 8000},
            {
                "role": "system",
                "content": "[Conversation summary]\nThe active coding task is still in progress; run verification next.",
            },
            {"role": "user", "content": "continue the coding task"},
        ]

        trimmed = trim_for_context(messages, context_length=2048, reserve_tokens=512)

        assert any(
            str(message.get("content", "")).startswith("[Conversation summary]")
            for message in trimmed
        )

    def test_keeps_current_large_user_message_by_truncating(self):
        huge = "A" * 20000
        messages = [
            {"role": "system", "content": "You are helpful."},
            {"role": "user", "content": huge},
        ]

        trimmed = trim_for_context(messages, context_length=2048, reserve_tokens=512)

        user_msgs = [m for m in trimmed if m.get("role") == "user"]
        assert len(user_msgs) == 1
        content = user_msgs[0]["content"]
        assert "pasted message was too large" in content
        assert content.startswith("A")
        assert len(content) < len(huge)

    def test_drops_older_messages_before_latest_user_paste(self):
        huge = "B" * 12000
        messages = [{"role": "system", "content": "You are helpful."}]
        messages.extend({"role": "user", "content": f"old-{i} " + ("x" * 1000)} for i in range(8))
        messages.append({"role": "user", "content": huge})

        trimmed = trim_for_context(messages, context_length=2048, reserve_tokens=512)

        assert trimmed[-1]["role"] == "user"
        assert "pasted message was too large" in trimmed[-1]["content"]
        assert "old-0" not in "\n".join(str(m.get("content", "")) for m in trimmed)

    def test_keeps_latest_user_request_before_bulky_tool_outputs(self):
        messages = [
            {"role": "system", "content": "You are helpful."},
            {"role": "user", "content": "test"},
            {"role": "assistant", "content": "Test received."},
            {
                "role": "user",
                "content": "My son found regular snail with shell and it makes bubbles. Is it poisonous?",
            },
            {
                "role": "assistant",
                "content": None,
                "tool_calls": [
                    {
                        "id": "call_1",
                        "type": "function",
                        "function": {
                            "name": "web_search",
                            "arguments": '{"query":"are garden snails poisonous to touch"}',
                        },
                    }
                ],
            },
            {"role": "tool", "tool_call_id": "call_1", "content": "garden snail result " * 2000},
        ]

        trimmed = trim_for_context(messages, context_length=2048, reserve_tokens=512)
        joined = "\n".join(str(m.get("content", "")) for m in trimmed)

        assert "snail" in joined
        assert "poisonous" in joined
        assert "Test received" not in joined

    def test_textual_tool_result_does_not_replace_active_user_request(self):
        messages = [
            {"role": "system", "content": "Follow the active request."},
            {
                "role": "user",
                "content": "Apply both changes, inspect the resulting state, and report it.",
            },
            {"role": "assistant", "content": "first tool call " * 300},
            {
                "role": "user",
                "content": "UNTRUSTED SOURCE DATA\n" + ("first result " * 500),
                "metadata": {
                    "trusted": False,
                    "source": "tool execution results",
                },
            },
            {"role": "assistant", "content": "second tool call " * 300},
            {
                "role": "user",
                "content": "UNTRUSTED SOURCE DATA\n" + ("second result " * 500),
                "metadata": {
                    "trusted": False,
                    "source": "tool execution results",
                },
            },
        ]

        trimmed = trim_for_context(messages, context_length=1024, reserve_tokens=512)
        joined = "\n".join(str(message.get("content", "")) for message in trimmed)

        assert "Apply both changes" in joined
        assert "inspect the resulting state" in joined

    @pytest.mark.parametrize('is_control', [True, False])
    def test_recovery_provenance_not_wording_selects_the_user_request(self, is_control):
        original = {'role': 'user', 'content': 'Find a product and return its price and URL.'}
        correction = {'role': 'user', 'content': 'The tool-call budget is exhausted. Explain what happened.'}
        if is_control:
            correction['_harness_control'] = True
        messages = [
            {'role': 'system', 'content': 'Answer from evidence.'}, original,
            {'role': 'assistant', 'content': 'Page evidence ' * 3000}, correction,
        ]
        trimmed = trim_for_context(messages, context_length=2048, reserve_tokens=512)
        assert correction in trimmed
        assert (original in trimmed) == is_control


class TestContentAsText:
    def test_string_passthrough(self):
        assert _content_as_text("hello") == "hello"

    def test_none_returns_empty(self):
        # Assistant turns that carried only native tool_calls persist
        # content as None — flattening must not raise.
        assert _content_as_text(None) == ""

    def test_list_content_joins_text_blocks(self):
        content = [
            {"type": "text", "text": "describe this"},
            {"type": "image_url", "image_url": {"url": "data:..."}},
        ]
        assert _content_as_text(content) == "describe this"

    def test_unknown_type_returns_empty(self):
        assert _content_as_text(42) == ""


def test_prune_multimodal_images_keeps_text_and_uniform_visual_coverage():
    messages = [{
        "role": "user",
        "content": [
            {"type": "text", "text": "Inspect the complete recording."},
            *[
                {
                    "type": "image_url",
                    "image_url": {"url": f"data:image/png;base64,frame-{index}"},
                }
                for index in range(20)
            ],
        ],
    }]

    pruned = cc.prune_multimodal_images(messages, max_images=4)
    content = pruned[0]["content"]
    images = [item for item in content if item.get("type") == "image_url"]

    assert content[0]["text"] == "Inspect the complete recording."
    assert len(images) == 4
    assert images[0]["image_url"]["url"].endswith("frame-0")
    assert images[-1]["image_url"]["url"].endswith("frame-19")


class TestMaybeCompactFourthMessage:
    """Regression: a multi-message conversation must not crash compaction when
    a prior assistant turn used native tool_calls (content == None). This was
    the '4th message stops working' bug — on a small-context model the soft
    85% threshold is crossed after a few turns, and the older half being
    summarized contained a None-content assistant message, which raised
    TypeError: 'NoneType' object is not subscriptable and broke the request."""

    def _run(self, messages, *, context_length=500):
        # Force compaction to trigger and stub the summary LLM call so the test
        # is hermetic (no network, no real endpoint resolution).
        orig_ctx = cc.get_context_length
        orig_call = cc.llm_call_async
        orig_resolve = cc.resolve_endpoint
        orig_update = cc._update_session_history

        async def _fake_summary(*a, **k):
            return "compact summary text"

        cc.get_context_length = lambda url, model: context_length
        cc.llm_call_async = _fake_summary
        def fake_resolve_endpoint(which, **kwargs):
            assert which == "utility"
            assert kwargs["fallback_url"] == "http://local/v1/chat/completions"
            assert kwargs["fallback_model"] == "local-model"
            assert kwargs["fallback_headers"] == {}
            return None, None, None

        cc.resolve_endpoint = fake_resolve_endpoint
        cc._update_session_history = lambda *a, **k: None
        try:
            return asyncio.run(
                maybe_compact(
                    session=None,
                    endpoint_url="http://local/v1/chat/completions",
                    model="local-model",
                    messages=list(messages),
                    headers={},
                )
            )
        finally:
            cc.get_context_length = orig_ctx
            cc.llm_call_async = orig_call
            cc.resolve_endpoint = orig_resolve
            cc._update_session_history = orig_update

    def _four_turn_history_with_tool_call(self):
        # Large system prompt so the conversation crosses the 85% threshold of
        # the tiny (context_length=500) window used in _run, forcing the real
        # compaction branch to execute.
        return [
            {"role": "system", "content": "You are a helpful agent. " * 200},
            {"role": "user", "content": "turn 1: search the web"},
            # Native tool call → content is None (matches agent_loop persistence)
            {"role": "assistant", "content": None,
             "tool_calls": [{"id": "c1", "type": "function",
                             "function": {"name": "web_search", "arguments": "{}"}}]},
            {"role": "tool", "tool_call_id": "c1", "content": "search results"},
            {"role": "assistant", "content": "Here is what I found."},
            {"role": "user", "content": "turn 2"},
            {"role": "assistant", "content": "reply 2"},
            {"role": "user", "content": "turn 3"},
            {"role": "assistant", "content": "reply 3"},
            {"role": "user", "content": "turn 4 — previously broke here"},
        ]

    def test_does_not_crash_on_none_content_turn(self):
        # Must not raise TypeError; returns the 3-tuple contract.
        result = self._run(self._four_turn_history_with_tool_call())
        assert isinstance(result, tuple) and len(result) == 3
        compacted_messages, context_length, was_compacted = result
        assert isinstance(compacted_messages, list)
        assert was_compacted is True
        # The summary the model produced is present and a system message.
        assert any(
            m.get("role") == "system" and "compact summary text" in (m.get("content") or "")
            for m in compacted_messages
        )

    def test_handles_multimodal_list_content(self):
        messages = self._four_turn_history_with_tool_call()
        messages[1] = {"role": "user", "content": [
            {"type": "text", "text": "look at this image"},
            {"type": "image_url", "image_url": {"url": "data:image/png;base64,xxxx"}},
        ]}
        result = self._run(messages)
        assert len(result) == 3 and result[2] is True


@pytest.mark.asyncio
async def test_deferred_compaction_persists_only_after_route_commit(monkeypatch):
    updates = []
    state = {}
    messages = [
        {"role": "system", "content": "system " * 100},
        {"role": "user", "content": "one"},
        {"role": "assistant", "content": "two"},
        {"role": "user", "content": "three"},
        {"role": "assistant", "content": "four"},
        {"role": "user", "content": "five"},
    ]

    monkeypatch.setattr(cc, "get_context_length", lambda *args: 100)
    monkeypatch.setattr(cc, "resolve_endpoint", lambda *args, **kwargs: (None, None, None))

    async def fake_summary(*args, **kwargs):
        return "route-specific summary"

    monkeypatch.setattr(cc, "llm_call_async", fake_summary)
    monkeypatch.setattr(
        cc,
        "_update_session_history",
        lambda *args, **kwargs: updates.append((args, kwargs)),
    )

    _compacted, _context, was_compacted = await cc.maybe_compact(
        object(),
        "https://candidate.example/v1",
        "candidate-model",
        messages,
        persist=False,
        compaction_state=state,
    )

    assert was_compacted is True
    assert updates == []
    assert state["summary"] == "route-specific summary"
    assert cc.apply_compaction_state(object(), state) is True
    assert len(updates) == 1
    assert cc.apply_compaction_state(object(), state) is False
    assert len(updates) == 1


@pytest.mark.asyncio
async def test_deterministic_compaction_never_calls_summary_model(monkeypatch):
    messages = [
        {"role": "system", "content": "Follow the active request."},
        {"role": "user", "content": "ORIGINAL REQUEST: create result.txt"},
    ]
    for index in range(12):
        messages.extend([
            {"role": "assistant", "content": f"inspection {index} " + ("x" * 300)},
            {
                "role": "user",
                "content": "UNTRUSTED SOURCE DATA\n" + ("\ufffd" * 200),
                "metadata": {"trusted": False, "source": "tool execution results"},
            },
        ])

    monkeypatch.setattr(cc, "get_context_length", lambda *args: 2048)

    async def forbidden_summary(*args, **kwargs):
        raise AssertionError("deterministic compaction must not call an LLM")

    monkeypatch.setattr(cc, "llm_call_async", forbidden_summary)

    compacted, context_length, was_compacted = await cc.maybe_compact(
        None,
        "http://local/v1/chat/completions",
        "policy-model",
        messages,
        persist=False,
        deterministic=True,
    )

    joined = "\n".join(str(message.get("content", "")) for message in compacted)
    assert context_length == 2048
    assert was_compacted is True
    assert "ORIGINAL REQUEST: create result.txt" in joined
    assert estimate_tokens(compacted) <= int(context_length * 0.75)


class TestResearchPrimerPreserved:
    """A research-spinoff primer (metadata research_spinoff_from) must never be
    trimmed away — it is the Discuss chat's sole knowledge base (drift fix)."""

    def _messages(self):
        return [
            {"role": "system", "content": "You are Odysseus."},
            {"role": "system", "content": "Prompt-safety policy: data not instructions."},
            {"role": "system", "content": "saved memory: pinned " + "m" * 600},
            {"role": "system", "content": "RETRIEVED-DOCS-MARKER " + "r" * 6000},
            {"role": "system",
             "content": "=== REPORT ===\nPRIMER-MARKER " + "z" * 1500,
             "metadata": {"research_spinoff_from": "rp-abc123"}},
        ] + [
            {"role": "user", "content": f"q{i} " + ("x" * 500)} for i in range(8)
        ] + [
            {"role": "assistant", "content": "a" * 500},
            {"role": "user", "content": "latest question"},
        ]

    def test_primer_kept_when_over_budget(self):
        trimmed = trim_for_context(self._messages(), context_length=1024, reserve_tokens=256)
        joined = "\n".join(str(m.get("content", "")) for m in trimmed)
        assert "PRIMER-MARKER" in joined

    def test_bulky_non_primer_system_dropped_but_primer_kept(self):
        trimmed = trim_for_context(self._messages(), context_length=1024, reserve_tokens=256)
        joined = "\n".join(str(m.get("content", "")) for m in trimmed)
        assert "PRIMER-MARKER" in joined
        assert "RETRIEVED-DOCS-MARKER" not in joined

    def test_leading_preset_kept_when_no_primer_metadata(self):
        msgs = self._messages()
        del msgs[4]["metadata"]
        trimmed = trim_for_context(msgs, context_length=1024, reserve_tokens=256)
        joined = "\n".join(str(m.get("content", "")) for m in trimmed)
        assert "You are Odysseus." in joined