ouroboros/tests/test_transcript_prefix.py

"""The transcript is append-only between the sends of one loop execution.

OpenAI-family caches (Codex backend, OpenAI API, OpenRouter->OpenAI) reuse a
previous request only when it is a byte-prefix of the next, so a transient
trailing message or an in-place rewrite of an already-sent message discards
the whole conversation cache (issue #906, measured 2026-09-14).  The unit
cases pin the recorder; the loop case pins the invariant on the real
``run_llm_loop`` with the acceptance observation actually injected.
"""
from __future__ import annotations

import copy
from types import SimpleNamespace

from ouroboros import loop, transcript_prefix
from ouroboros.transcript_prefix import CHECKPOINT_KIND, message_digest, observe_send, sanction_rewrite

# The heaviest real harness that injects the per-round acceptance observation.
from tests.test_acceptance_async_loop import ANSWER, call, full_loop, keep  # noqa: F401

# ---------------------------------------------------------------------------
# message_digest: content identity, not envelope shape
# ---------------------------------------------------------------------------

def test_cache_control_marker_does_not_change_the_digest():
    plain = {"role": "user", "content": "do the work"}
    marked = {"role": "user", "content": [
        {"type": "text", "text": "do the work", "cache_control": {"type": "ephemeral"}}]}
    assert message_digest(plain) == message_digest(marked)


def test_block_and_string_tool_results_share_one_digest():
    as_string = {"role": "tool", "tool_call_id": "t1", "content": "result body"}
    as_blocks = {"role": "tool", "tool_call_id": "t1",
                 "content": [{"type": "text", "text": "result "}, {"type": "text", "text": "body"}]}
    assert message_digest(as_string) == message_digest(as_blocks)


def test_private_custody_keys_do_not_change_the_digest():
    bare = {"role": "user", "content": "observation"}
    carried = {"role": "user", "content": "observation", "acceptance_observation": True,
               "nativeContinuation": {"id": "x"}, "_context_capsule": {"a": 1},
               "_caption": "cap", "_source_path": "/tmp/x.png"}
    assert message_digest(bare) == message_digest(carried)


def test_role_text_and_tool_calls_do_change_the_digest():
    base = {"role": "user", "content": "a"}
    assert message_digest(base) != message_digest({"role": "assistant", "content": "a"})
    assert message_digest(base) != message_digest({"role": "user", "content": "b"})
    calling = {"role": "assistant", "content": "", "tool_calls": [call("read_file", {"path": "x"}, "c1")]}
    other = {"role": "assistant", "content": "", "tool_calls": [call("read_file", {"path": "y"}, "c1")]}
    assert message_digest(calling) != message_digest(other)
    assert message_digest(calling) != message_digest({"role": "assistant", "content": ""})
    assert message_digest({"role": "tool", "tool_call_id": "a", "content": "r"}) != message_digest(
        {"role": "tool", "tool_call_id": "b", "content": "r"})


# ---------------------------------------------------------------------------
# observe_send: which break, where, and whether it was sanctioned
# ---------------------------------------------------------------------------

SYSTEM = {"role": "system", "content": "Complete the owner task."}
TASK = {"role": "user", "content": "Prepare the report."}
ASSISTANT = {"role": "assistant", "content": "working"}


def test_first_send_and_pure_append_record_nothing():
    slot = SimpleNamespace()
    assert observe_send(slot, [SYSTEM, TASK], round_idx=1) is None
    assert observe_send(slot, [SYSTEM, TASK, ASSISTANT], round_idx=2) is None
    assert observe_send(slot, [SYSTEM, TASK, ASSISTANT], round_idx=3) is None
    assert getattr(slot, transcript_prefix.DIGEST_ATTR) == [
        message_digest(m) for m in (SYSTEM, TASK, ASSISTANT)]


def test_replaced_last_message_is_a_tail_replacement():
    slot = SimpleNamespace()
    observe_send(slot, [SYSTEM, TASK, ASSISTANT], round_idx=1)
    fact = observe_send(slot, [SYSTEM, TASK, {"role": "assistant", "content": "other"}], round_idx=2)
    assert fact == {"checkpoint_kind": CHECKPOINT_KIND, "round": 2, "index": 2, "kind": "tail_replaced",
                    "previous_messages": 3, "current_messages": 3, "sanctioned_by": None}


def test_changed_middle_message_is_a_rewrite():
    slot = SimpleNamespace()
    observe_send(slot, [SYSTEM, TASK, ASSISTANT], round_idx=1)
    fact = observe_send(
        slot, [SYSTEM, {"role": "user", "content": "other task"}, ASSISTANT, TASK], round_idx=2)
    assert fact["kind"] == "rewritten" and fact["index"] == 1
    assert (fact["previous_messages"], fact["current_messages"]) == (3, 4)


def test_changed_system_message_is_a_system_rewrite():
    slot = SimpleNamespace()
    observe_send(slot, [SYSTEM, TASK, ASSISTANT], round_idx=1)
    fact = observe_send(slot, [{"role": "system", "content": "reprojected"}, TASK, ASSISTANT], round_idx=4)
    assert fact["kind"] == "system_rewritten" and fact["index"] == 0 and fact["round"] == 4


def test_shorter_transcript_with_an_equal_prefix_is_a_shrink():
    slot = SimpleNamespace()
    observe_send(slot, [SYSTEM, TASK, ASSISTANT], round_idx=1)
    fact = observe_send(slot, [SYSTEM, TASK], round_idx=2)
    assert fact["kind"] == "shrunk" and fact["index"] == 3
    assert (fact["previous_messages"], fact["current_messages"]) == (3, 2)


def test_sanction_stamp_is_consumed_by_the_next_observation():
    slot = SimpleNamespace()
    observe_send(slot, [SYSTEM, TASK, ASSISTANT], round_idx=1)
    sanction_rewrite(slot, "compaction")
    stamped = observe_send(slot, [SYSTEM, TASK, {"role": "assistant", "content": "other"}], round_idx=2)
    assert stamped["sanctioned_by"] == "compaction"
    later = observe_send(slot, [SYSTEM, TASK, {"role": "assistant", "content": "third"}], round_idx=3)
    assert later["sanctioned_by"] is None, "the stamp is one-shot"


def test_a_compaction_stamp_never_explains_a_system_rewrite():
    slot = SimpleNamespace()
    observe_send(slot, [SYSTEM, TASK, ASSISTANT], round_idx=1)
    sanction_rewrite(slot, "compaction")
    fact = observe_send(slot, [{"role": "system", "content": "reprojected"}, TASK, ASSISTANT], round_idx=2)
    assert fact["kind"] == "system_rewritten" and fact["sanctioned_by"] is None
    assert getattr(slot, transcript_prefix.SANCTION_ATTR) is None, "the stamp is still consumed"


def test_sanctioned_by_rides_through_unchanged():
    slot = SimpleNamespace()
    observe_send(slot, [SYSTEM, TASK, ASSISTANT], round_idx=1)
    fact = observe_send(slot, [SYSTEM, TASK], round_idx=2, sanctioned_by="compaction")
    assert fact["sanctioned_by"] == "compaction"


# ---------------------------------------------------------------------------
# The loop-level invariant
# ---------------------------------------------------------------------------

def _describe(message):
    calls = transcript_prefix._tool_call_identity(message.get("tool_calls"))
    return (f"role={message.get('role')!r} tool_calls={[c[1] for c in calls]!r} "
            f"text={transcript_prefix._plain_text(message.get('content'))[:120]!r}")


def _assert_prefix_chain(sends):
    for step in range(1, len(sends)):
        previous, current = sends[step - 1], sends[step]
        for index, (before, after) in enumerate(zip(previous, current)):
            assert message_digest(before) == message_digest(after), (
                f"send {step} is not a prefix extension of send {step - 1}: message {index} of "
                f"{len(previous)} was rewritten\n  before: {_describe(before)}\n  after:  {_describe(after)}")
        assert len(current) >= len(previous), (
            f"send {step} dropped messages: {len(previous)} -> {len(current)}; "
            f"last surviving message {_describe(current[-1])}")


def _prefix_breaks(events):
    rows = []
    for item in list(events.queue):
        data = item.get("data") if isinstance(item, dict) else None
        if isinstance(data, dict) and data.get("type") == "task_checkpoint" \
                and data.get("checkpoint_kind") == CHECKPOINT_KIND:
            rows.append(data)
    return rows


def _four_tool_rounds(f):
    """Script: nominate + one effect, then a read per round, then finalize."""
    def main(_llm, messages, *_args, **_kwargs):
        f.model_inputs.append(copy.deepcopy(messages))
        f.model_step += 1
        if f.model_step == 1:
            return {"content": "The report is ready; requesting its review.", "tool_calls": [
                call("task_acceptance_review", {"claim": ANSWER}, "nominate"),
                call("write_file", {"root": "task_drive", "path": "proof.txt",
                                    "content": "effect"}, "seed"),
            ]}, 0.0
        if f.model_step in (2, 3, 4):
            if f.model_step == 2:
                assert f.entered.wait(5), f.progress
            return {"content": "", "tool_calls": [
                call("read_file", {"root": "task_drive", "path": "proof.txt"},
                     f"read-{f.model_step}")]}, 0.0
        assert f.model_step < 9, f.progress
        return keep(f), 0.0
    return main


def test_every_send_of_one_execution_extends_the_previous_send(full_loop, monkeypatch):  # noqa: F811 -- imported pytest fixture
    f = full_loop
    monkeypatch.setattr(loop, "call_llm_with_retry", _four_tool_rounds(f))

    result, usage, _trace = f.run()

    assert result == ANSWER, (result, f.progress)
    assert len(f.model_inputs) >= 5, f.model_inputs
    _assert_prefix_chain(f.model_inputs)
    assert _prefix_breaks(f.events) == []
    assert "prompt_prefix_breaks" not in usage


# ---------------------------------------------------------------------------
# One predicate for every producer that appends behind a sent row
# ---------------------------------------------------------------------------

def test_a_sent_tail_row_is_never_merged_into_whatever_produced_it():
    """The #929 carve-out generalized: the acceptance observation, an owner
    follow-up, a task message and a roster note are all just rows that went out
    in the previous send; with the slot in hand the digest list decides, and a
    slot-less producer still keeps the observation marker as its stand-in."""
    from ouroboros.loop_messages import _append_or_merge_user_content
    from ouroboros.transcript_prefix import sent_in_previous_send

    slot = SimpleNamespace()
    messages = [SYSTEM, TASK, {"role": "user", "content": "[Message from my human]: first"}]
    _append_or_merge_user_content(messages, "merged before any send", slot=slot)
    assert len(messages) == 3 and "merged before any send" in messages[-1]["content"]
    observe_send(slot, messages, round_idx=1)
    assert sent_in_previous_send(slot, messages[-1]) is True
    _append_or_merge_user_content(messages, "[Message from independent task x]\nhello", slot=slot)
    assert len(messages) == 4, "a sent tail is byte-frozen: the new content is its own row"
    assert messages[2]["content"].endswith("merged before any send")
    _append_or_merge_user_content(messages, "same round follow-up", slot=slot)
    assert len(messages) == 4 and messages[-1]["content"].endswith("same round follow-up")
    # The slot-less arm: an observation row is refused on its marker alone.
    tail = {"role": "user", "content": "[ACCEPTANCE_SUBJECT_OBSERVATION]", "acceptance_observation": True}
    plain = [SYSTEM, TASK, tail]
    _append_or_merge_user_content(plain, "owner words")
    assert len(plain) == 4 and plain[2] is tail


def test_the_roster_note_rides_the_real_loop_as_an_append(full_loop, monkeypatch):  # noqa: F811 -- imported pytest fixture
    """A host-listed independent root appears once, as a fresh tail row, and the
    transcript stays a prefix chain across every send (no tail_replaced)."""
    from ouroboros.utils import atomic_write_json, utc_now_iso

    f = full_loop
    (f.ctx.drive_root / "state").mkdir(parents=True, exist_ok=True)
    atomic_write_json(f.ctx.drive_root / "state" / "queue_snapshot.json", {
        "ts": utc_now_iso(), "running": [
            {"id": "peer-root", "task": {"id": "peer-root", "title": "Peer work", "chat_id": 0}},
        ], "pending": [], "worker_total": 1,
    })
    monkeypatch.setattr(loop, "call_llm_with_retry", _four_tool_rounds(f))

    result, usage, _trace = f.run()

    assert result == ANSWER, (result, f.progress)
    _assert_prefix_chain(f.model_inputs)
    assert _prefix_breaks(f.events) == []
    notes = [m for m in f.model_inputs[-1]
             if m.get("role") == "user" and "[INDEPENDENT_ROOTS]" in transcript_prefix._plain_text(m.get("content"))]
    assert len(notes) == 1, "the roster changed once, so exactly one note was appended"
    assert "- peer-root · Peer work · chat=0 · running" in transcript_prefix._plain_text(notes[0]["content"])
    assert "prompt_prefix_breaks" not in usage


def _fake_compaction(messages, *_args, **_kwargs):
    """Stand in for the summarizer behind both compaction seams: rewrite the
    first tool result in place and report the pass as applied."""
    rebuilt = list(messages)
    for index, message in enumerate(rebuilt):
        if message.get("role") == "tool":
            rebuilt[index] = {**message, "content": "[compacted tool result]"}
            break
    receipt = SimpleNamespace(status="applied", checkpoint_ref="ckpt-1", reclaimed_tokens=10, goal_reached=True)
    return rebuilt, receipt, {"prompt_tokens": 1, "completion_tokens": 1}


def _after_step(f, step, hook):
    """Run ``hook`` once the scripted model has answered round ``step``."""
    inner = _four_tool_rounds(f)

    def main(*args, **kwargs):
        out = inner(*args, **kwargs)
        if f.model_step == step:
            hook()
        return out
    return main


def _assert_one_sanctioned_rewrite_at_round_three(f, usage):
    breaks = _prefix_breaks(f.events)
    assert [b["kind"] for b in breaks] == ["rewritten"], breaks
    assert breaks[0]["sanctioned_by"] == "compaction" and breaks[0]["round"] == 3, breaks
    # A sanctioned rewrite is disclosed, never counted as an unexplained break.
    assert "prompt_prefix_breaks" not in usage
    # The rounds after the compaction extend the compacted transcript again.
    _assert_prefix_chain(f.model_inputs[2:])


def test_manual_compaction_is_recorded_as_a_sanctioned_break_on_its_round(full_loop, monkeypatch):  # noqa: F811 -- imported pytest fixture
    f = full_loop
    monkeypatch.setattr(loop, "compact_tool_history_llm", _fake_compaction)
    # The owner-requested reclaim is picked up by _run_round_compaction at the top of round 3.
    monkeypatch.setattr(loop, "call_llm_with_retry", _after_step(
        f, 2, lambda: setattr(f.ctx, "_pending_compaction", 1)))

    result, usage, _trace = f.run()

    assert result == ANSWER, (result, f.progress)
    _assert_one_sanctioned_rewrite_at_round_three(f, usage)


def test_automatic_reclaim_inside_the_model_call_is_a_sanctioned_break_on_its_round(full_loop, monkeypatch):  # noqa: F811 -- imported pytest fixture
    f = full_loop
    monkeypatch.setattr(loop, "call_llm_with_retry", _four_tool_rounds(f))
    monkeypatch.setattr(loop, "compact_tool_history_llm", _fake_compaction)
    fired = []

    def measure(ctx, *, automatic_pass_used):
        # ContextFit decides "reclaim_once" exactly once, at round 3, the way the
        # automatic reclaim runs inside _call_round_model after the round began.
        if ctx.round_idx != 3 or fired or automatic_pass_used:
            return None
        fired.append(True)
        measurement = SimpleNamespace(route_fp="fp", round_id="r3", measurement_basis="cold_estimate",
                                      measurement_density=1.0, reclaim_goal_tokens=100)
        return SimpleNamespace(action="reclaim_once", measurement=measurement, automatic_pass_used=False)

    monkeypatch.setattr(loop, "_measure_round_main_fit", measure)

    result, usage, _trace = f.run()

    assert result == ANSWER, (result, f.progress)
    assert fired, "the automatic reclaim did not run"
    _assert_one_sanctioned_rewrite_at_round_three(f, usage)