failproofai 1.0.1-beta.1 → 1.0.2-beta.0

package/.next/standalone/sdk/python/tests/test_server_contract.py

@@ -0,0 +1,650 @@
+"""What the ingest server promotes, and what this SDK must therefore emit.
+
+The server pulls a fixed set of keys out of each event's payload and stores them
+as real indexed columns; everything else stays in the opaque `payload` blob.
+Those promoted columns are what every filter, facet and chart in the product
+reads. The extraction is total in the worst way:
+
+    fn ps(payload, key) -> Option<&str> { payload.get(key).and_then(|v| v.as_str()) }
+
+A missing key and a key holding the wrong JSON type are indistinguishable — both
+yield `None`, both store NULL, and the row still ingests with `200 OK`. So an
+event that spells `tool_name` as `toolName`, or sends it as a number, lands in
+storage looking successful and is invisible to every tool-name filter forever.
+
+This file pins the SDK's half. `PROMOTED` is frozen here because this repository
+cannot see the server, and `test_promoted_columns_match_the_server` re-derives
+the same list from a real AgentEye checkout when `FP_AGENTEYE_ROOT` points at
+one, so the frozen copy cannot quietly rot.
+"""
+import json
+import os
+import re
+from pathlib import Path
+
+import pytest
+
+from failproofai_sdk import _context, _events, _resolver, _schema
+from failproofai_sdk._events import EventNamespace
+
+# ─────────────────────────────────────────────────────────────────────────────
+# The promoted-column contract
+# ─────────────────────────────────────────────────────────────────────────────
+
+#: Payload keys the server lifts into an indexed column via `ps()` (string).
+PROMOTED_STRING = frozenset(
+    {
+        "tool_name",
+        "tool_call_id",
+        "hook_name",
+        "hook_id",
+        "input_id",
+        "pause_id",
+        "error_type",
+        "model",
+    }
+)
+
+#: Payload keys the server lifts via `pu32()` (unsigned, 32-bit).
+PROMOTED_NUMERIC = frozenset({"duration_ms", "input_tokens", "output_tokens"})
+
+PROMOTED = PROMOTED_STRING | PROMOTED_NUMERIC
+
+
+class _Recorder:
+    """Stands in for the writer so a call can be inspected without touching disk."""
+
+    def __init__(self):
+        self.entries = []
+
+    def submit(self, entry):
+        self.entries.append(entry)
+
+
+def _emit_everything():
+    """Call every public event method once, fully populated. Returns the payloads."""
+    recorder = _Recorder()
+    ns = EventNamespace(recorder)
+
+    ids = dict(session_id="s", agent_id="a")
+    ns.agent_start(**ids, goal="g", parent_id="p")
+    ns.agent_end(**ids, outcome="success", summary="s")
+    ns.agent_pause(**ids, pause_id="p1", reason="r", user_id="u")
+    ns.agent_resume(**ids, pause_id="p1", reason="r", user_id="u")
+    ns.tool_use(**ids, tool_name="bash", tool_call_id="tc", input={"cmd": "ls"})
+    ns.tool_result(**ids, tool_name="bash", tool_call_id="tc", output="o")
+    ns.model_request(**ids, model="m", messages=[], system=None, tools=[])
+    ns.model_response(**ids, model="m", stop_reason="end_turn", input_tokens=1, output_tokens=2, content="c", role="assistant")
+    ns.error(**ids, error_type="ValueError", message="m", traceback="t")
+    ns.hook_triggered(**ids, hook_name="h", hook_id="h1", trigger_event="PreToolUse", input={})
+    ns.hook_completed(**ids, hook_name="h", hook_id="h1", outcome="allow", output="o")
+    ns.human_wait(**ids, input_id="i1", prompt="p", options=["y"], reason="r")
+    ns.human_input(**ids, input_id="i1", response="y")
+    ns.human_pause(**ids, reason="r", user_id="u")
+    ns.human_interrupt(**ids, reason="r", user_id="u", at_step="1")
+    return recorder.entries
+
+
+def test_every_promoted_column_is_emitted_by_at_least_one_event():
+    """A promoted column nothing ever populates is a dead column, silently."""
+    emitted = set()
+    for payload in _emit_everything():
+        emitted |= payload.keys()
+
+    missing = PROMOTED - emitted
+    assert not missing, (
+        f"no event emits {sorted(missing)}, so those indexed columns are "
+        "always NULL. Either an event method stopped sending the key, or the "
+        "server promotes something this SDK never produces."
+    )
+
+
+def test_promoted_values_have_the_json_type_the_extractor_accepts():
+    """`ps()` needs a JSON string; `pu32()` needs a non-negative integer.
+
+    A type mismatch is NOT an error on either side — it is a NULL column next to
+    a payload that still visibly contains the value, which is the single most
+    confusing shape this pipeline can produce.
+    """
+    for payload in _emit_everything():
+        for key in PROMOTED_STRING & payload.keys():
+            assert isinstance(payload[key], str), (
+                f"{payload['type']}.{key} is {type(payload[key]).__name__}; "
+                "ps() only reads JSON strings and would store NULL"
+            )
+        for key in PROMOTED_NUMERIC & payload.keys():
+            value = payload[key]
+            assert isinstance(value, int) and not isinstance(value, bool), (
+                f"{payload['type']}.{key} is {type(value).__name__}; emit a "
+                "whole integer so the value survives regardless of how the "
+                "server's numeric extractor is implemented"
+            )
+            assert value >= 0, f"{payload['type']}.{key} is negative; the column is unsigned"
+            assert value <= 0xFFFFFFFF, f"{payload['type']}.{key} overflows UInt32"
+
+
+def test_promoted_keys_never_collide_with_a_reserved_name():
+    """Reserved names are validated separately; a promoted key must not be one."""
+    assert not (PROMOTED & _events._RESERVED)
+
+
+@pytest.mark.parametrize("event_type", ["tool_result", "hook_completed", "human_input", "agent_resume"])
+def test_unpaired_events_omit_duration_rather_than_sending_zero(event_type):
+    """No start seen means unknown, and unknown must not be recorded as `0ms`.
+
+    Zero is a real, plottable duration. Emitting it for "we never saw the start"
+    poisons every latency percentile with values that were never measured.
+    """
+    recorder = _Recorder()
+    ns = EventNamespace(recorder)
+    ids = dict(session_id="s", agent_id="a")
+
+    if event_type == "tool_result":
+        ns.tool_result(**ids, tool_name="t", tool_call_id="never-started")
+    elif event_type == "hook_completed":
+        ns.hook_completed(**ids, hook_name="h", hook_id="never-started")
+    elif event_type == "human_input":
+        ns.human_input(**ids, input_id="never-started")
+    else:
+        ns.agent_resume(**ids, pause_id="never-started")
+
+    assert "duration_ms" not in recorder.entries[0]
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Reserved names — enforcement is SPLIT across two mechanisms, deliberately
+# ─────────────────────────────────────────────────────────────────────────────
+
+
+def test_reserved_set_is_exactly_the_five_ingest_reads_structurally():
+    assert _events._RESERVED == {"timestamp", "session_id", "agent_id", "type", "environment"}
+
+
+@pytest.mark.parametrize("name", ["session_id", "agent_id"])
+def test_identity_kwargs_raise_type_error_not_value_error(name):
+    """These are explicit parameters, so Python rejects the duplicate first.
+
+    Two exception types for one apparent rule. Documented rather than unified,
+    because unifying it means shadowing a real signature parameter to raise a
+    nicer error, and that trades a loud failure for a subtler one.
+    """
+    ns = EventNamespace(_Recorder())
+    kwargs = {"session_id": "s", "agent_id": "a"}
+    kwargs[name] = "duplicate"
+    with pytest.raises(TypeError):
+        ns.agent_start(**kwargs, **{name: "again"})
+
+
+@pytest.mark.parametrize("name", ["timestamp", "type", "environment"])
+def test_stamped_fields_raise_value_error_from_the_validator(name):
+    """These are not signature parameters, so they reach `_validate_fields`."""
+    ns = EventNamespace(_Recorder())
+    with pytest.raises(ValueError, match="Reserved field names"):
+        ns.agent_start(session_id="s", agent_id="a", **{name: "x"})
+
+
+@pytest.mark.parametrize(
+    "call",
+    [
+        lambda ns: ns.tool_result(session_id="s", agent_id="a", tool_name="t", tool_call_id="c", duration_ms=5),
+        lambda ns: ns.hook_completed(session_id="s", agent_id="a", hook_name="h", hook_id="i", duration_ms=5),
+        lambda ns: ns.human_input(session_id="s", agent_id="a", input_id="i", duration_ms=5),
+        lambda ns: ns.agent_resume(session_id="s", agent_id="a", pause_id="p", duration_ms=5),
+    ],
+)
+def test_duration_ms_cannot_be_supplied_by_the_caller(call):
+    """It is measured, not reported — a caller-supplied value would be unfalsifiable."""
+    ns = EventNamespace(_Recorder())
+    with pytest.raises(ValueError, match="auto-computed"):
+        call(ns)
+
+
+def test_a_reserved_name_is_rejected_by_every_event_method():
+    """One method forgetting `_validate_fields` would let a caller rewrite `type`."""
+    ns = EventNamespace(_Recorder())
+    methods = [m for m in dir(ns) if not m.startswith("_") and callable(getattr(ns, m))]
+    assert len(methods) == 15, f"expected 15 event methods, found {len(methods)}: {sorted(methods)}"
+
+    required = {
+        "tool_use": dict(tool_name="t", tool_call_id="c"),
+        "tool_result": dict(tool_name="t", tool_call_id="c"),
+        "hook_triggered": dict(hook_name="h", hook_id="i"),
+        "hook_completed": dict(hook_name="h", hook_id="i"),
+        "error": dict(error_type="E", message="m"),
+        "human_wait": dict(input_id="i"),
+        "human_input": dict(input_id="i"),
+        "agent_pause": dict(pause_id="p"),
+        "agent_resume": dict(pause_id="p"),
+    }
+    for method in methods:
+        with pytest.raises(ValueError, match="Reserved field names"):
+            getattr(ns, method)(
+                session_id="s", agent_id="a", **required.get(method, {}), type="hijacked"
+            )
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# The strings the rename must never touch
+# ─────────────────────────────────────────────────────────────────────────────
+
+#: Renaming any of these desynchronises the SDK from the daemon that reads its
+#: spool, with no error on either side — events are simply written somewhere
+#: nothing watches. The Python import name and the PyPI distribution name were
+#: renamed to failproofai_sdk / failproofai-sdk; the wire and filesystem
+#: contract deliberately was not, exactly as the `fp` CLI kept `X-AgentEye-Org`
+#: and the `ae_session` cookie through its own rename.
+FROZEN_STRINGS = {
+    "AGENTEYE_ENVIRONMENT": "the environment label, read at import time",
+    "FAILPROOFAI_HOME": "moves the umbrella root, mirrored in fp-home.ts",
+    "custom-agents": "the DEFAULT spool root, mirrored in fp-home.ts and config.rs",
+}
+
+#: Deliberately NOT frozen, and each for its own reason.
+#:
+#: ``AGENTEYE_SPOOL_TO_FAILPROOFAI`` was the opt-in that selected the umbrella
+#: root. It also required that directory to already exist, and nothing ever
+#: created it, so the branch never once fired. It is retired rather than frozen:
+#: the umbrella is the default now, so anyone who exported it already has what
+#: they were asking for.
+#:
+#: ``.agenteye`` is no longer a literal this package must contain. It survives
+#: in prose and in `legacy_agenteye_dir()`, but freezing it would make this test
+#: pass on a comment — which is exactly how it passed while the variable above
+#: was being deleted.
+#: ``AGENTEYE_HOME`` was frozen here as "the operator override both daemons
+#: honour". It no longer resolves the spool in THIS package: it could aim the
+#: SDK at any directory from the environment, including as an unintended side
+#: effect of exporting it for `agenteye-collector`, which is the one component
+#: that still reads it. `failproofaid` keeps watching `~/.agenteye/events`, so
+#: nothing already written there is stranded — but this SDK writes to the
+#: umbrella and nothing can redirect it off.
+#:
+#: It is RETIRED rather than dropped from the file, because the name still
+#: appears in `_resolver`'s prose explaining why it went away, and a frozen
+#: substring check would happily pass on that comment.
+RETIRED_STRINGS = ("AGENTEYE_SPOOL_TO_FAILPROOFAI", "AGENTEYE_HOME")
+
+PACKAGE_DIR = Path(_resolver.__file__).resolve().parent
+
+
+def test_the_spool_contract_strings_are_still_spelled_the_old_way():
+    """A well-meaning rename sweep is the realistic threat here, not a redesign."""
+    source = "\n".join(
+        p.read_text(encoding="utf-8") for p in sorted(PACKAGE_DIR.glob("*.py"))
+    )
+    for literal, why in FROZEN_STRINGS.items():
+        assert literal in source, (
+            f"{literal!r} is gone from the package — {why}. This is a wire and "
+            "filesystem contract with a daemon that is released separately, so "
+            "renaming it here strands every event this SDK writes. If the "
+            "rename is genuinely intended, change the daemon FIRST and keep "
+            "reading the old name for at least one release."
+        )
+
+
+def test_the_retired_opt_in_is_not_read_by_any_module():
+    """Freed, not merely unused — a leftover branch would contradict the docs.
+
+    Checked over `os.environ` lookups rather than the raw text, because the name
+    still appears in `_resolver`'s prose explaining why it went away, and a
+    substring check over source is how the frozen-strings test above was passing
+    for the wrong reason while the variable was being deleted.
+    """
+    import re
+
+    for path in sorted(PACKAGE_DIR.glob("*.py")):
+        source = path.read_text(encoding="utf-8")
+        for retired in RETIRED_STRINGS:
+            reads = re.findall(rf"environ(?:\.get)?[\(\[]\s*[\"']{retired}", source)
+            assert not reads, f"{path.name} still reads the retired {retired}"
+
+
+def test_the_spool_layout_is_events_and_failed_under_the_base_dir():
+    """The daemon watches `<base>/events` and quarantines to `<base>/failed`."""
+    writer_source = (PACKAGE_DIR / "_writer.py").read_text(encoding="utf-8")
+    assert '"events"' in writer_source
+
+
+def test_batches_are_published_by_atomic_rename_from_a_tmp_suffix():
+    """`.tmp` -> `.jsonl` via os.replace is what stops a half-written read.
+
+    The daemon takes any `.jsonl` in the directory as complete. Writing straight
+    to the final name would hand it a truncated file whose unparseable lines
+    ingest counts as `skipped` — 200 OK, events gone.
+    """
+    writer_source = (PACKAGE_DIR / "_writer.py").read_text(encoding="utf-8")
+    assert '.tmp' in writer_source
+    assert '.jsonl' in writer_source
+    assert "os.replace(tmp_path, final_path)" in writer_source, (
+        "the atomic publish is gone. shutil.move, Path.rename across devices, or "
+        "a plain write to the final name all reintroduce the torn-read window."
+    )
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Cross-check against a real AgentEye checkout, when one is available
+# ─────────────────────────────────────────────────────────────────────────────
+
+_AGENTEYE_ROOT = os.environ.get("FP_AGENTEYE_ROOT")
+_INGEST_RS = Path(_AGENTEYE_ROOT) / "server" / "src" / "routes" / "ingest.rs" if _AGENTEYE_ROOT else None
+
+requires_agenteye_checkout = pytest.mark.skipif(
+    _INGEST_RS is None or not _INGEST_RS.is_file(),
+    reason="set FP_AGENTEYE_ROOT to an AgentEye checkout to verify against real ingest.rs",
+)
+
+
+@requires_agenteye_checkout
+def test_promoted_columns_match_the_server():
+    """Re-derive PROMOTED from ingest.rs so the frozen copy above cannot rot."""
+    source = _INGEST_RS.read_text(encoding="utf-8")
+
+    actual_strings = set(re.findall(r'ps\(payload_value,\s*"([a-z_]+)"\)', source))
+    actual_numeric = set(re.findall(r'pu32\(payload_value,\s*"([a-z_]+)"\)', source))
+
+    # `model_name` is a documented fallback spelling the server also accepts;
+    # this SDK only ever emits `model`, which is the preferred one.
+    actual_strings.discard("model_name")
+
+    assert actual_strings == PROMOTED_STRING, (
+        "the server's promoted string columns drifted from this SDK's frozen "
+        f"copy. server={sorted(actual_strings)} sdk={sorted(PROMOTED_STRING)}"
+    )
+    assert PROMOTED_NUMERIC >= actual_numeric, (
+        "the server promotes a numeric column this SDK does not know about: "
+        f"{sorted(actual_numeric - PROMOTED_NUMERIC)}"
+    )
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Promoted numeric columns — the type has to be right, not just the key
+# ─────────────────────────────────────────────────────────────────────────────
+
+
+@pytest.mark.parametrize("name", sorted(_events._PROMOTED_NUMERIC))
+@pytest.mark.parametrize("bad", ["many", 12.5, True, -1, 2**32, [1], {"n": 1}])
+def test_a_promoted_numeric_is_refused_rather_than_stored_as_null(name, bad):
+    """`pu32()` returns None on a mismatch, and None is written as NULL at 200 OK.
+
+    Nothing is logged and nothing is rejected — the row still arrives — so the
+    only symptom is a column that is populated for some events and not others.
+    `_RESERVED` never covered this: it blocks five structural keys and lets every
+    other custom field through untouched, whatever type it carries.
+
+    Rejecting rather than coercing, because a float is a mistake worth hearing
+    about: the server drops it whole rather than rounding it, and rounding it
+    here would hide that from the one person able to fix the source.
+    """
+    ns = EventNamespace(_Recorder())
+    with pytest.raises(ValueError, match=name):
+        ns.agent_start(session_id="s", agent_id="a", **{name: bad})
+
+
+@pytest.mark.parametrize("name", ["input_tokens", "output_tokens"])
+def test_model_response_checks_its_own_two_token_counts(name):
+    """They are named parameters, so they never reach `_validate_fields`.
+
+    They are also the likeliest of the three to arrive wrong — a caller reading
+    them straight off a provider's usage object gets whatever that object holds.
+    """
+    ns = EventNamespace(_Recorder())
+    with pytest.raises(ValueError, match=name):
+        ns.model_response(session_id="s", agent_id="a", model="m", **{name: "1024"})
+
+
+@pytest.mark.parametrize("name", sorted(_events._PROMOTED_NUMERIC))
+def test_a_valid_promoted_numeric_still_goes_through_untouched(name):
+    recorder = _Recorder()
+    EventNamespace(recorder).agent_start(session_id="s", agent_id="a", **{name: 4096})
+    assert recorder.entries[0][name] == 4096
+
+
+@pytest.mark.parametrize("name", sorted(_events._PROMOTED_NUMERIC))
+def test_the_boundary_values_of_a_u32_are_accepted(name):
+    """0 and 2**32 - 1 are valid, and an off-by-one here silently rejects real data."""
+    recorder = _Recorder()
+    ns = EventNamespace(recorder)
+    ns.agent_start(session_id="s", agent_id="a", **{name: 0})
+    ns.agent_start(session_id="s", agent_id="a", **{name: 2**32 - 1})
+    assert [e[name] for e in recorder.entries] == [0, 2**32 - 1]
+
+
+def test_none_is_not_treated_as_a_bad_promoted_numeric():
+    """`model_response` passes its optionals straight through as None."""
+    recorder = _Recorder()
+    EventNamespace(recorder).model_response(
+        session_id="s", agent_id="a", model="m", input_tokens=None, output_tokens=None
+    )
+    assert "input_tokens" not in recorder.entries[0]
+
+
+def test_the_promoted_numeric_set_matches_the_one_ingest_lifts():
+    """Two hand-maintained lists, one in the SDK and one in this file's header.
+
+    If they drift, the validator stops covering a column that ingest still
+    promotes, and this file's own assertions stop describing the server.
+    """
+    assert _events._PROMOTED_NUMERIC == PROMOTED_NUMERIC
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# The MEASURED duration is bound by the same u32 range a caller is held to
+# ─────────────────────────────────────────────────────────────────────────────
+
+#: The four events that pair with an earlier one and compute `duration_ms`.
+#: Each takes a start call and an end call against the same id.
+PAIRED = {
+    "tool_result": (
+        lambda ns: ns.tool_use(session_id="s", agent_id="a", tool_name="t", tool_call_id="x"),
+        lambda ns: ns.tool_result(session_id="s", agent_id="a", tool_name="t", tool_call_id="x"),
+    ),
+    "hook_completed": (
+        lambda ns: ns.hook_triggered(session_id="s", agent_id="a", hook_name="h", hook_id="x"),
+        lambda ns: ns.hook_completed(session_id="s", agent_id="a", hook_name="h", hook_id="x"),
+    ),
+    "human_input": (
+        lambda ns: ns.human_wait(session_id="s", agent_id="a", input_id="x"),
+        lambda ns: ns.human_input(session_id="s", agent_id="a", input_id="x"),
+    ),
+    "agent_resume": (
+        lambda ns: ns.agent_pause(session_id="s", agent_id="a", pause_id="x"),
+        lambda ns: ns.agent_resume(session_id="s", agent_id="a", pause_id="x"),
+    ),
+}
+
+
+def _emit_pair_with_gap(event_type, gap_seconds, monkeypatch):
+    """Run one start/end pair with a controlled interval between them."""
+    import datetime as _dt
+
+    start_call, end_call = PAIRED[event_type]
+    recorder = _Recorder()
+    ns = EventNamespace(recorder)
+
+    base = _dt.datetime(2026, 1, 1, tzinfo=_dt.timezone.utc)
+    clock = {"now": base}
+    monkeypatch.setattr(EventNamespace, "_now", staticmethod(lambda: clock["now"]))
+
+    start_call(ns)
+    clock["now"] = base + _dt.timedelta(seconds=gap_seconds)
+    recorder.entries.clear()
+    end_call(ns)
+    return recorder.entries[0]
+
+
+@pytest.mark.parametrize("event_type", sorted(PAIRED))
+def test_a_pair_spanning_more_than_u32_milliseconds_omits_the_duration(event_type, monkeypatch):
+    """~49.7 days is an ordinary lifetime for these pairs, not an abuse.
+
+    A `human_wait` answered after a long weekend, an `agent_pause` resumed a
+    month later. Over the range, `pu32()` stores NULL at 200 OK — so an
+    unbounded value is not an error anywhere, just an empty column.
+    """
+    over = (2**32 / 1000) + 60  # comfortably past 2**32 ms
+    payload = _emit_pair_with_gap(event_type, over, monkeypatch)
+    assert "duration_ms" not in payload, (
+        f"{event_type} emitted {payload.get('duration_ms')}, which the server stores as NULL"
+    )
+
+
+@pytest.mark.parametrize("event_type", sorted(PAIRED))
+def test_a_backwards_clock_omits_the_duration_rather_than_going_negative(event_type, monkeypatch):
+    """`datetime.now()` is wall-clock, so an NTP step back yields a negative gap.
+
+    `round()` keeps the sign, and a negative into an unsigned column is the same
+    silent NULL as an oversized one.
+    """
+    payload = _emit_pair_with_gap(event_type, -5, monkeypatch)
+    assert "duration_ms" not in payload, f"{event_type} emitted a negative duration"
+
+
+@pytest.mark.parametrize("event_type", sorted(PAIRED))
+def test_an_ordinary_gap_still_produces_a_duration(event_type, monkeypatch):
+    """The bound must not swallow the normal case it exists to protect."""
+    payload = _emit_pair_with_gap(event_type, 1.5, monkeypatch)
+    assert payload["duration_ms"] == 1500
+    assert isinstance(payload["duration_ms"], int)
+
+
+@pytest.mark.parametrize("event_type", sorted(PAIRED))
+def test_the_upper_boundary_itself_is_kept(event_type, monkeypatch):
+    """Exactly 2**32 - 1 ms is representable; an off-by-one here drops real data."""
+    payload = _emit_pair_with_gap(event_type, (2**32 - 1) / 1000, monkeypatch)
+    assert payload["duration_ms"] == 2**32 - 1
+
+
+def test_the_measured_and_the_caller_supplied_paths_enforce_the_same_range():
+    """One range, two entry points. They drifted once already.
+
+    `_validate_promoted_numeric` refused a caller anything outside 0..2**32-1
+    while the SDK's own computation was unbounded — the same field, held to two
+    different standards depending on who produced it.
+    """
+    import datetime as _dt
+
+    base = _dt.datetime(2026, 1, 1, tzinfo=_dt.timezone.utc)
+    over = base + _dt.timedelta(milliseconds=2**32)
+    assert _events._measured_duration_ms(base, over) is None
+    with pytest.raises(ValueError):
+        _events._validate_promoted_numeric("duration_ms", 2**32)
+
+    ok = base + _dt.timedelta(milliseconds=2**32 - 1)
+    assert _events._measured_duration_ms(base, ok) == 2**32 - 1
+    _events._validate_promoted_numeric("duration_ms", 2**32 - 1)
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# session_id / agent_id — on every event, and skipped silently when wrong
+# ─────────────────────────────────────────────────────────────────────────────
+
+#: Every public event method, with the arguments it needs besides the two ids.
+ALL_METHODS = [
+    ("agent_start", {}), ("agent_end", {}),
+    ("agent_pause", {"pause_id": "p"}), ("agent_resume", {"pause_id": "p"}),
+    ("tool_use", {"tool_name": "t", "tool_call_id": "c"}),
+    ("tool_result", {"tool_name": "t", "tool_call_id": "c"}),
+    ("model_request", {}), ("model_response", {}),
+    ("hook_triggered", {"hook_name": "h", "hook_id": "i"}),
+    ("hook_completed", {"hook_name": "h", "hook_id": "i"}),
+    ("error", {"error_type": "E", "message": "m"}),
+    ("human_wait", {"input_id": "i"}), ("human_input", {"input_id": "i"}),
+    ("human_pause", {}), ("human_interrupt", {}),
+]
+
+
+def test_the_method_list_here_covers_every_public_event():
+    """Or a method added later silently skips the check below."""
+    public = {
+        n for n in dir(EventNamespace)
+        if not n.startswith("_") and callable(getattr(EventNamespace, n))
+    }
+    assert {m for m, _ in ALL_METHODS} == public
+
+
+@pytest.mark.parametrize("method,extra", ALL_METHODS, ids=[m for m, _ in ALL_METHODS])
+@pytest.mark.parametrize("bad", [None, 123, {"x": 1}, ["a"], b"bytes"], ids=lambda v: type(v).__name__)
+def test_a_non_string_session_id_is_refused_on_every_event(method, extra, bad):
+    """Ingest SKIPS these and answers 200 — `{"accepted":0,"skipped":1}`.
+
+    Verified against the live server. Nothing upstream learns: the SDK reports
+    success, the collector deletes the batch, and the event is gone. `None` is
+    the realistic way in — an uninitialised variable, a lookup that missed.
+    """
+    ns = EventNamespace(_Recorder())
+    with pytest.raises(TypeError, match="session_id"):
+        getattr(ns, method)(session_id=bad, agent_id="a", **extra)
+
+
+@pytest.mark.parametrize("method,extra", ALL_METHODS, ids=[m for m, _ in ALL_METHODS])
+def test_a_non_string_agent_id_is_refused_on_every_event(method, extra):
+    """`None` is no longer invalid — it means "resolve from scope" (below).
+
+    `TypeError`, because the wrong TYPE was supplied — and because omitting a
+    required keyword argument was always a TypeError, so code catching one keeps
+    working now that identity is optional.
+    """
+    ns = EventNamespace(_Recorder())
+    with pytest.raises(TypeError, match="agent_id"):
+        getattr(ns, method)(session_id="s", agent_id=12345, **extra)
+
+
+@pytest.mark.parametrize("method,extra", ALL_METHODS, ids=[m for m, _ in ALL_METHODS])
+def test_an_omitted_agent_id_resolves_rather_than_raising(method, extra):
+    """Omitting it is the ergonomic path the scopes exist for, not an error.
+
+    With no `agent()` scope open it lands on `DEFAULT_AGENT_ID` — the convention
+    the skill already teaches — rather than raising. Inventing a *session* that
+    way would scatter one run across as many sessions as it has emit sites, so
+    only the agent has a default.
+    """
+    recorder = _Recorder()
+    getattr(EventNamespace(recorder), method)(session_id="s", **extra)
+    assert recorder.entries[0]["agent_id"] == _context.DEFAULT_AGENT_ID
+
+
+def test_an_unresolvable_session_id_raises_rather_than_emitting():
+    """Nothing passed and nothing bound must be loud.
+
+    Ingest skips an event whose `session_id` is not a JSON string and answers
+    `200 OK` with `{"accepted":0,"skipped":1}`, so emitting here would lose the
+    event with no error on either side — the exact failure the validation exists
+    to prevent, reached through the new ergonomic path instead of a bad argument.
+    """
+    ns = EventNamespace(_Recorder())
+    with pytest.raises(TypeError, match="session_id is required"):
+        ns.agent_start()
+
+
+def test_an_ambient_scope_satisfies_both_ids():
+    import failproofai_sdk
+
+    recorder = _Recorder()
+    ns = EventNamespace(recorder)
+    with failproofai_sdk.session("sess-x"):
+        with failproofai_sdk.agent("planner"):
+            ns.agent_start(goal="from ambient scope")
+    entry = recorder.entries[-1]
+    assert entry["session_id"] == "sess-x"
+    assert entry["agent_id"] == "planner"
+
+
+@pytest.mark.parametrize("blank", ["", "   ", "\t\n"], ids=["empty", "spaces", "whitespace"])
+def test_a_blank_identity_is_refused_although_the_server_accepts_it(blank):
+    """The worse of the two outcomes, which is why it is refused too.
+
+    A skipped event is at least absent. A blank id is ACCEPTED by the server, so
+    every event sent that way lands and is silently grouped under one id — the
+    data looks present and is quietly merged across unrelated runs.
+    """
+    ns = EventNamespace(_Recorder())
+    with pytest.raises(ValueError, match="empty"):
+        ns.agent_start(session_id=blank, agent_id="a")
+    with pytest.raises(ValueError, match="empty"):
+        ns.agent_start(session_id="s", agent_id=blank)
+
+
+def test_ordinary_identities_are_untouched():
+    recorder = _Recorder()
+    EventNamespace(recorder).agent_start(session_id="run-001", agent_id="planner")
+    assert recorder.entries[0]["session_id"] == "run-001"
+    assert recorder.entries[0]["agent_id"] == "planner"