Skip to content

Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.

24.6 KB
Raw Blame
"""Integration: Mechanism B (read maturation) cache behaviour, through the
REAL Anthropic handler, across a multi-turn session.

The held Read is cached with its turn: the client's breakpoints are forwarded
where the client put them, and the Read matures once its file quiesces even
though the provider-confirmed prefix has moved past it by then. The unit
tests in ``test_read_maturation.py`` call the manager in isolation with
``frozen_message_count=0``; the live test in ``test_live/`` does a
2-request hold->mature with no intermediate turns. Neither exercises the
realistic path where a held Read sits across several turns while the prefix
tracker advances ``frozen_message_count`` from the provider's reported cache
usage.

This is a regression test for that path. It drives the real handler with a
mocked upstream that echoes the cache usage Anthropic would report (caching
everything up to the breakpoint the handler chose, system blocks included),
so the prefix tracker advances exactly as in production. It then asserts,
directly on the FORWARDED bytes:

1. the client's breakpoint is forwarded where the client put it on every
   turn, including the turns that hold the Read;
2. the mechanism actually engages: the Read is held verbatim while the file
   is active, then matures into a CCR marker once it quiesces, in that order.

Note on cache-state isolation: the CCR store is persistent (SQLite at
~/.headroom/ccr_store.db by default) and shared across processes, so stale
entries from prior runs can perturb maturation timing. Run against a clean
store for deterministic results.
"""

from __future__ import annotations

import copy
import json

import pytest

pytest.importorskip("fastapi")

import httpx
from fastapi.testclient import TestClient

from headroom.proxy.server import ProxyConfig, create_app

MODEL = "claude-haiku-4-5-20251001"
SYSTEM = [
    {
        "type": "text",
        "text": "You are a coding assistant. Be terse. " * 200,
        "cache_control": {"type": "ephemeral"},
    }
]
READ_TOOL = {
    "name": "Read",
    "description": "Read a file",
    "input_schema": {
        "type": "object",
        "properties": {"file_path": {"type": "string"}},
        "required": ["file_path"],
    },
}
READ_ID = "toolu_r1"
# Big enough to dominate message tokens and clear the maturation min-size gate.
BIG = "".join(f"  {i}\tdef f_{i}(): return {i}  # line {i}\n" for i in range(700))


def _read_pair(tail: bool) -> list[dict]:
    tr = {"type": "tool_result", "tool_use_id": READ_ID, "content": BIG}
    if tail:
        tr["cache_control"] = {"type": "ephemeral"}
    return [
        {
            "role": "assistant",
            "content": [
                {
                    "type": "tool_use",
                    "id": READ_ID,
                    "name": "Read",
                    "input": {"file_path": "/x/foo.py"},
                }
            ],
        },
        {"role": "user", "content": [tr]},
    ]


def _quiet_pair(i: int, tail: bool) -> list[dict]:
    u = {"type": "text", "text": f"Unrelated question {i}: what is {i}+{i}?"}
    if tail:
        u["cache_control"] = {"type": "ephemeral"}
    return [
        {"role": "assistant", "content": [{"type": "text", "text": str(2 * i)}]},
        {"role": "user", "content": [u]},
    ]


def _convo(nquiet: int) -> list[dict]:
    """Read of /x/foo.py followed by ``nquiet`` turns that never touch it.
    The Claude-Code-style tail breakpoint rides the newest user block."""
    msgs: list[dict] = [{"role": "user", "content": [{"type": "text", "text": "Read /x/foo.py"}]}]
    msgs += _read_pair(tail=(nquiet == 0))
    for i in range(1, nquiet + 1):
        msgs += _quiet_pair(i, tail=(i == nquiet))
    return msgs


def _busy_convo(nquiet: int) -> list[dict]:
    """``_convo`` with a large JSON Bash output before the Read and on every
    quiet turn, as in a real coding session. The router compresses those, and
    a compression run is what marks the held Read as locally replayable, so
    the frozen count follows the provider's cache past it."""

    def bash(tid: str) -> list[dict]:
        out = json.dumps([{"path": f"src/{tid}_{j}.py", "size": j} for j in range(300)])
        return [
            {
                "role": "assistant",
                "content": [
                    {"type": "tool_use", "id": tid, "name": "Bash", "input": {"command": "ls"}}
                ],
            },
            {
                "role": "user",
                "content": [{"type": "tool_result", "tool_use_id": tid, "content": out}],
            },
        ]

    msgs: list[dict] = [{"role": "user", "content": [{"type": "text", "text": "Read /x/foo.py"}]}]
    msgs += bash("toolu_b0") + _read_pair(tail=False)
    for i in range(1, nquiet + 1):
        msgs += bash(f"toolu_b{i}")
    msgs[-1]["content"][-1]["cache_control"] = {"type": "ephemeral"}
    return msgs


def _breakpoint_index(messages: list[dict]) -> int:
    """Index of the last message carrying a cache_control block (-1 if none).
    Anthropic caches everything up to AND INCLUDING this message."""
    bp = -1
    for i, m in enumerate(messages):
        c = m.get("content")
        if isinstance(c, list) and any(isinstance(b, dict) and "cache_control" in b for b in c):
            bp = i
    return bp


def _read_result_content(message: dict) -> str | None:
    c = message.get("content")
    if isinstance(c, list):
        for b in c:
            if (
                isinstance(b, dict)
                and b.get("type") == "tool_result"
                and b.get("tool_use_id") == READ_ID
            ):
                return b.get("content")
    return None


def _est_tokens(message: dict) -> int:
    return max(1, len(str(message.get("content", ""))) // 4)


def test_held_read_keeps_client_breakpoint_then_matures(monkeypatch):
    # Isolate the CCR store: it is persistent (SQLite) and shared across
    # processes by default, so stale entries from other runs would perturb
    # maturation timing and make this test non-deterministic. The in-memory
    # backend gives a pristine store per test.
    from headroom.cache.compression_store import reset_compression_store

    monkeypatch.setenv("HEADROOM_CCR_BACKEND", "memory")
    reset_compression_store()

    # Match the real proxy: cache machinery ON (the prefix tracker + compression
    # cache are what maturation's hold/frozen-count logic depends on). Disabling
    # them masks the behavior under test.
    config = ProxyConfig(
        optimize=True,
        read_maturation=True,
        mode="token",
        cache_enabled=True,
        rate_limit_enabled=False,
        cost_tracking_enabled=False,
        log_requests=False,
    )
    # The provider confirms no cache (as after a lapse), so nothing is frozen
    # and a quiet Read may mature. Once the provider has cached it, it stays
    # verbatim instead: test_read_inside_the_cached_prefix_stays_verbatim.
    forwarded = _drive_session(config, n_turns=7, session_id="nobust-1", cache_hits=False)

    assert forwarded, "no requests were forwarded"

    # Per-turn classification of the Read's forwarded form.
    held_verbatim = []  # turns where the Read is forwarded verbatim (held)
    matured = []  # turns where the Read has become a CCR marker
    for turn, msgs in enumerate(forwarded):
        for m in msgs:
            content = _read_result_content(m)
            if content is None:
                continue
            if content == BIG:
                held_verbatim.append(turn)
            elif "Retrieve original: hash=" in content:
                matured.append(turn)

    # INVARIANT 1: the client's breakpoint is forwarded where the client put
    # it. This used to be the opposite ("the verbatim Read is never
    # cache-written"): the handler moved the breakpoint in front of the held
    # Read, which dropped Claude Code's tail breakpoint for the whole hold and
    # re-sent the Read and everything after it uncached on every held turn, to
    # save one cache write of the Read. The held Read is now cached with its
    # turn.
    moved = [
        turn
        for turn, msgs in enumerate(forwarded)
        if _breakpoint_index(msgs) != _breakpoint_index(_convo(turn))
    ]
    assert not moved, f"client cache breakpoint moved or dropped on turn(s) {moved}"

    # INVARIANT 2 (mechanism actually engages): the Read is held verbatim while
    # the file is active, then matures once it quiesces. Guards against a
    # vacuous pass where maturation silently no-ops.
    assert held_verbatim, "expected the fresh Read to be held verbatim on early turns"
    assert matured, "expected the Read to mature into a CCR marker after quiescing"
    # The matured marker only appears AFTER the verbatim hold (ordering).
    assert min(matured) > max(held_verbatim), (
        f"maturation must follow the hold: held={held_verbatim} matured={matured}"
    )


def _drive_session(
    config,
    n_turns: int,
    session_id: str,
    saved_out: list[int] | None = None,
    convo=_convo,
    cache_hits: bool = True,
    tools: list[dict] | None = None,
) -> list[list[dict]]:
    """Drive ``n_turns`` cumulative turns of ``convo`` through the real handler
    with a mocked upstream; return the forwarded message arrays per turn. When
    ``saved_out`` is given, each turn's emitted ``x-headroom-tokens-saved`` is
    appended to it. The upstream caches up to the handler's breakpoint, or
    reports no cache at all when ``cache_hits`` is False."""
    app = create_app(config)
    forwarded: list[list[dict]] = []
    with TestClient(app) as client:
        proxy = client.app.state.proxy
        original_retry = proxy._retry_request

        async def _mock_upstream(method, url, headers, body, stream=False, **kwargs):
            msgs = body.get("messages", []) or []
            forwarded.append(copy.deepcopy(msgs))
            bp = _breakpoint_index(msgs)
            sys_tokens = sum(
                max(1, len(str(b.get("text", ""))) // 4)
                for b in (body.get("system") or [])
                if isinstance(b, dict)
            )
            cached = sys_tokens + (sum(_est_tokens(m) for m in msgs[: bp + 1]) if bp >= 0 else 0)
            if not cache_hits:
                cached = 0
            return httpx.Response(
                200,
                json={
                    "id": "msg_x",
                    "type": "message",
                    "role": "assistant",
                    "content": [{"type": "text", "text": "ok"}],
                    "usage": {
                        "input_tokens": 20,
                        "output_tokens": 2,
                        "cache_read_input_tokens": cached,
                        "cache_creation_input_tokens": 0,
                    },
                },
            )

        proxy._retry_request = _mock_upstream
        try:
            for n in range(n_turns):
                r = client.post(
                    "/v1/messages",
                    headers={
                        "x-api-key": "test-key",
                        "anthropic-version": "2023-06-01",
                        "x-headroom-session-id": session_id,
                        "content-type": "application/json",
                    },
                    json={
                        "model": MODEL,
                        "max_tokens": 20,
                        "system": SYSTEM,
                        "tools": tools or [READ_TOOL],
                        "messages": convo(n),
                    },
                )
                assert r.status_code == 200, f"turn {n}: {r.text[:300]}"
                if saved_out is not None:
                    saved_out.append(int(r.headers.get("x-headroom-tokens-saved", 0)))
        finally:
            proxy._retry_request = original_retry
    return forwarded


def _first_matured_turn(forwarded: list[list[dict]]) -> int | None:
    """The first turn index whose forwarded Read is a CCR marker."""
    for turn, msgs in enumerate(forwarded):
        for m in msgs:
            content = _read_result_content(m)
            if content and "Retrieve original: hash=" in content:
                return turn
    return None


def test_read_cached_with_its_turn_stays_verbatim(monkeypatch):
    """In a session whose tool output compresses, the frozen count follows the
    provider's cache past the held Read from the next request on. The Read is
    then left verbatim for good: maturing it would re-write everything after it
    at the cache-write rate, which costs more than the Read saves."""
    from headroom.cache.compression_store import reset_compression_store
    from headroom.transforms.read_maturation import ReadMaturationManager

    monkeypatch.setenv("HEADROOM_CCR_BACKEND", "memory")
    reset_compression_store()

    frozen_seen: list[int] = []
    apply = ReadMaturationManager.apply

    def _recording_apply(self, messages, frozen_message_count=0):
        frozen_seen.append(frozen_message_count)
        return apply(self, messages, frozen_message_count=frozen_message_count)

    monkeypatch.setattr(ReadMaturationManager, "apply", _recording_apply)

    config = ProxyConfig(
        optimize=True,
        read_maturation=True,
        mode="token",
        cache_enabled=True,
        rate_limit_enabled=False,
        cost_tracking_enabled=False,
        log_requests=False,
    )
    forwarded = _drive_session(config, n_turns=7, session_id="cached-read-1", convo=_busy_convo)

    moved = [
        turn
        for turn, msgs in enumerate(forwarded)
        if _breakpoint_index(msgs) != _breakpoint_index(_busy_convo(turn))
    ]
    assert not moved, f"client cache breakpoint moved or dropped on turn(s) {moved}"
    read_at = 4  # the Read's tool_result message in _busy_convo
    assert all(f > read_at for f in frozen_seen[1:]), (
        f"fixture must cache the held Read: frozen counts {frozen_seen}"
    )
    assert _first_matured_turn(forwarded) is None, "a cached Read was matured"


DEFERRED_TOOL = {
    "name": "mcp__notes__search",
    "description": "Search notes",
    "defer_loading": True,
    "input_schema": {"type": "object", "properties": {"q": {"type": "string"}}},
}


def _tool_search_convo(nquiet: int) -> list[dict]:
    """``_busy_convo`` after a ToolSearch whose result holds only tool_reference
    blocks, as Claude Code loads deferred tools. Such textless results are
    stable; tests must not rely on them stopping the local replay window."""
    msgs = _busy_convo(nquiet)
    search = [
        {
            "role": "assistant",
            "content": [
                {
                    "type": "tool_use",
                    "id": "toolu_ts",
                    "name": "ToolSearch",
                    "input": {"query": "notes"},
                }
            ],
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "tool_result",
                    "tool_use_id": "toolu_ts",
                    "content": [{"type": "tool_reference", "tool_name": DEFERRED_TOOL["name"]}],
                }
            ],
        },
    ]
    return msgs[:1] + search + msgs[1:]


def _without_cache_control(messages: list[dict]) -> list[dict]:
    return json.loads(
        json.dumps(messages),
        object_hook=lambda d: {k: v for k, v in d.items() if k != "cache_control"},
    )


def _prefix_rewrites(forwarded: list[list[dict]]) -> list[int]:
    """Turns whose forwarded messages do not start with the previous turn's.
    The mock caches each turn through its last message, so any such turn
    re-writes the provider cache. Breakpoints move every turn and are ignored."""
    rewrites = []
    for turn in range(1, len(forwarded)):
        prev = _without_cache_control(forwarded[turn - 1])
        if _without_cache_control(forwarded[turn])[: len(prev)] != prev:
            rewrites.append(turn)
    return rewrites


def _record_pipeline_freeze(monkeypatch) -> list[int]:
    from headroom.proxy import session_engine

    seen: list[int] = []
    prepare = session_engine.prepare_turn

    def _recording_prepare(*args, **kwargs):
        prep = prepare(*args, **kwargs)
        seen.append(prep.frozen_message_count)
        return prep

    monkeypatch.setattr(session_engine, "prepare_turn", _recording_prepare)
    return seen


@pytest.mark.parametrize(
    ("convo", "tools", "pipeline_stop"),
    [
        # Both input shapes exercise the same explicitly bounded local cache.
        (_convo, [READ_TOOL], 2),
        (_tool_search_convo, [READ_TOOL, DEFERRED_TOOL], 2),
    ],
    ids=["quiet", "tool-reference"],
)
@pytest.mark.parametrize("cache_hits", [True, False], ids=["provider-cached", "provider-uncached"])
def test_read_inside_the_cached_prefix_stays_verbatim(
    monkeypatch, convo, tools, pipeline_stop, cache_hits
):
    """By the time the Read's file quiesces the provider has cached it, so
    maturing it would re-write the cache from the Read onward. The pipeline's
    own freeze stops below the Read here; maturation must follow the
    provider-confirmed floor that finalize_turn replays instead. The uncached
    control must still mature, so retaining the Read cannot pass vacuously."""
    from headroom.cache.compression_cache import CompressionCache
    from headroom.cache.compression_store import reset_compression_store
    from headroom.transforms.read_maturation import ReadMaturationManager

    monkeypatch.setenv("HEADROOM_CCR_BACKEND", "memory")
    reset_compression_store()
    # Explicitly model a local replay window shorter than the provider's
    # confirmed prefix (e.g. after local cache eviction). Keep real preparation
    # and tracking; do not depend on any particular tool-result shape being a
    # cache miss. Textless ToolSearch results are legitimately stable now.
    compute_frozen = CompressionCache.compute_frozen_count

    def _bounded_local_replay(self, messages):
        return min(compute_frozen(self, messages), pipeline_stop)

    monkeypatch.setattr(CompressionCache, "compute_frozen_count", _bounded_local_replay)
    pipeline_freeze = _record_pipeline_freeze(monkeypatch)
    maturation_freeze: list[int] = []
    apply = ReadMaturationManager.apply

    def _recording_apply(self, messages, frozen_message_count=0):
        maturation_freeze.append(frozen_message_count)
        return apply(self, messages, frozen_message_count=frozen_message_count)

    monkeypatch.setattr(ReadMaturationManager, "apply", _recording_apply)
    config = ProxyConfig(
        optimize=True,
        read_maturation=True,
        mode="token",
        cache_enabled=True,
        rate_limit_enabled=False,
        cost_tracking_enabled=False,
        log_requests=False,
    )

    forwarded = _drive_session(
        config,
        n_turns=8,
        session_id="cached-read-2",
        convo=convo,
        tools=tools,
        cache_hits=cache_hits,
    )

    assert pipeline_freeze and max(pipeline_freeze) <= pipeline_stop, (
        f"fixture must keep the pipeline freeze below the Read: {pipeline_freeze}"
    )
    assert len(forwarded) == 8, "every turn must reach the upstream"
    read_at = next(i for i, msg in enumerate(convo(0)) if _read_result_content(msg) is not None)
    assert pipeline_stop <= read_at, "the local replay window must exclude the held Read"
    if not cache_hits:
        assert maturation_freeze and not any(maturation_freeze)
        assert _first_matured_turn(forwarded) is not None, "uncached control never matured"
        return
    assert len(maturation_freeze) == 8 and all(f > read_at for f in maturation_freeze[1:]), (
        f"provider-confirmed maturation floor must protect the Read: {maturation_freeze}"
    )
    assert all(_read_result_content(msgs[read_at]) == BIG for msgs in forwarded), (
        "the held Read must remain present and verbatim on every turn"
    )
    assert _first_matured_turn(forwarded) is None, "a cached Read was matured"
    assert not _prefix_rewrites(forwarded), "the forwarded cached prefix changed"


def test_quiesce_turns_config_is_honored(monkeypatch):
    """`quiesce_turns` must be runtime-configurable end-to-end: a fresh Read of
    /x/foo.py matures `quiesce_turns` quiet turns after it appears (the convo
    builds one quiet assistant turn per step, and the Read sits at assistant
    turn 1). With quiesce_turns=2 it must mature at turn 2 — not the built-in
    default of 5. Currently the handler hardcodes ReadMaturationConfig(enabled=
    True), ignoring the configured value, so this fails (matures at 5)."""
    from headroom.cache.compression_store import reset_compression_store

    monkeypatch.setenv("HEADROOM_CCR_BACKEND", "memory")
    reset_compression_store()

    config = ProxyConfig(
        optimize=True,
        read_maturation=True,
        read_maturation_quiesce_turns=2,
        mode="token",
        cache_enabled=True,
        rate_limit_enabled=False,
        cost_tracking_enabled=False,
        log_requests=False,
    )
    forwarded = _drive_session(config, n_turns=4, session_id="quiesce-cfg-1", cache_hits=False)
    first = _first_matured_turn(forwarded)
    assert first == 2, (
        f"expected the Read to mature at turn 2 with quiesce_turns=2, "
        f"but first matured at turn {first} (handler ignored the configured value)"
    )


def test_read_maturation_knobs_from_env(monkeypatch):
    """Operators must be able to tune maturation via env vars (the pilot
    playbook says 'pick quiesce_turns')."""
    from headroom.proxy.server import _MULTI_WORKER_CONFIG_ENV, _proxy_config_from_env

    # _proxy_config_from_env short-circuits on a prebuilt multi-worker JSON
    # config and ignores the HEADROOM_* vars entirely. Clear it so this test
    # actually exercises the env-var parsing path it claims to (and isn't
    # poisoned by a leaked HEADROOM_PROXY_CONFIG_JSON from another test).
    monkeypatch.delenv(_MULTI_WORKER_CONFIG_ENV, raising=False)

    monkeypatch.setenv("HEADROOM_READ_MATURATION", "1")
    monkeypatch.setenv("HEADROOM_ROLLOUT_CHANNEL", "beta")
    monkeypatch.setenv("HEADROOM_READ_MATURATION_QUIESCE_TURNS", "3")
    monkeypatch.setenv("HEADROOM_READ_MATURATION_MAX_HOLD_TURNS", "10")
    monkeypatch.setenv("HEADROOM_READ_MATURATION_MIN_SIZE_BYTES", "4096")

    cfg = _proxy_config_from_env()

    assert cfg.read_maturation is True
    assert cfg.read_maturation_quiesce_turns == 3
    assert cfg.read_maturation_max_hold_turns == 10
    assert cfg.read_maturation_min_size_bytes == 4096


def test_read_maturation_env_cannot_bypass_stable_rollout(monkeypatch):
    """Every env-driven server composition root must enforce the beta gate."""
    from headroom.proxy.server import _MULTI_WORKER_CONFIG_ENV, _proxy_config_from_env

    monkeypatch.delenv(_MULTI_WORKER_CONFIG_ENV, raising=False)
    monkeypatch.setenv("HEADROOM_READ_MATURATION", "1")
    monkeypatch.setenv("HEADROOM_ROLLOUT_CHANNEL", "stable")

    cfg = _proxy_config_from_env()

    assert cfg.read_maturation is False
    assert cfg.rollout is not None
    assert cfg.rollout.decision("read_maturation").reason.value == "blocked_by_channel"


def test_replayed_marker_is_not_rebooked_in_the_emitted_savings(monkeypatch):
    """The savings a request EMITS must be first-appearance.

    Subtracting the replay debt inside the maturation block is not enough on
    its own: the handler recomputes ``tokens_saved`` from the plain
    original-vs-optimized diff twice more before the outcome is recorded (the
    pre-send hook recount and the consistency recount), and either one throws
    the adjustment away — so every replay turn re-books a removal that was
    already booked when the Read matured. Assert on the emitted header, which
    is the same figure the outcome and PERF line carry, not on the manager.
    """
    from headroom.cache.compression_store import reset_compression_store

    monkeypatch.setenv("HEADROOM_CCR_BACKEND", "memory")
    reset_compression_store()

    config = ProxyConfig(
        optimize=True,
        read_maturation=True,
        read_maturation_quiesce_turns=2,
        mode="token",
        cache_enabled=True,
        rate_limit_enabled=False,
        cost_tracking_enabled=False,
        log_requests=False,
    )
    saved: list[int] = []
    forwarded = _drive_session(
        config, n_turns=5, session_id="first-appearance-1", saved_out=saved, cache_hits=False
    )

    first = _first_matured_turn(forwarded)
    assert first is not None, "the Read never matured, so there is nothing to replay"
    replays = saved[first + 1 :]
    assert replays, "expected at least one replay turn after maturation"

    # The turn that matures the Read books its removal, once.
    assert saved[first] > 0, "the maturing turn booked nothing"
    # Later turns forward the same marker instead of the verbatim Read, so the
    # wire diff is just as large — but it is the SAME removal, already booked.
    assert max(replays) < saved[first] * 0.2, (
        f"replayed marker re-booked through the final accounting path: "
        f"matured turn saved {saved[first]}, replay turns saved {replays}"
    )