failproofai 1.0.1 → 1.0.2-beta.0

package/.next/standalone/sdk/python/failproofai_sdk/_events.py

@@ -0,0 +1,752 @@
+import logging
+from datetime import datetime, timezone
+from typing import Any
+
+from failproofai_sdk import _context
+
+from failproofai_sdk._schema import (
+    AgentEndEvent,
+    AgentPauseEvent,
+    AgentResumeEvent,
+    AgentStartEvent,
+    ErrorEvent,
+    HookCompletedEvent,
+    HookTriggeredEvent,
+    HumanInputEvent,
+    HumanInterruptEvent,
+    HumanPauseEvent,
+    HumanWaitEvent,
+    ModelRequestEvent,
+    ModelResponseEvent,
+    ToolResultEvent,
+    ToolUseEvent,
+)
+
+logger = logging.getLogger(__name__)
+
+# session_id and agent_id are explicit signature params on every method, so Python raises
+# TypeError before our validator runs if a caller tries to pass them as extra **fields.
+# timestamp and type are not in the signature, so they land in **fields and are caught here.
+_RESERVED = frozenset({"timestamp", "session_id", "agent_id", "type", "environment"})
+
+# Payload keys ingest lifts out of the JSON blob into unsigned 32-bit columns via
+# `pu32()`. Everything else is stored as-is and can be any shape, but these three
+# are read with a typed accessor that returns None on a mismatch — and a None
+# there is written as NULL under a 200 OK. Nothing is logged, nothing is
+# rejected, and the row still arrives, so the only symptom is a column that is
+# empty for some events and not others.
+#
+# `duration_ms` is refused outright on the four events that MEASURE it. These
+# checks cover the other way in: any of the three passed as a custom field on an
+# event that does not name it, plus `model_response`'s own two parameters, which
+# are the ones a caller is most likely to fill straight from a provider response.
+_PROMOTED_NUMERIC = frozenset({"duration_ms", "input_tokens", "output_tokens"})
+
+#: Payload keys ingest lifts into an INDEXED STRING column via `ps()`, which
+#: reads JSON strings and stores NULL for anything else — the same silent-at-200
+#: failure `_PROMOTED_NUMERIC` guards, on the eight keys that had no guard at
+#: all. `None` is the realistic way in: `tool_name=getattr(tool, "name", None)`
+#: is ordinary code, and it produced a row that is invisible to every tool-name
+#: filter and whose `tool_result` never pairs.
+_PROMOTED_STRING = frozenset(
+    {
+        "tool_name",
+        "tool_call_id",
+        "hook_name",
+        "hook_id",
+        "input_id",
+        "pause_id",
+        "error_type",
+        "model",
+    }
+)
+
+_U32_MAX = 2**32 - 1
+
+
+def _validate_promoted_numeric(name: str, value) -> None:
+    """Reject anything `pu32()` would silently turn into NULL.
+
+    Rejecting rather than coercing, and at the boundary rather than in the
+    writer, for the same reason `_validated_interval` does: this is the last
+    point where the caller still has a stack trace pointing at their own call.
+    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
+    who could fix the source of it.
+
+    `bool` is checked before `int` because it IS an int in Python, and `True`
+    would otherwise sail through and store as 1.
+    """
+    if value is None:
+        return
+    if isinstance(value, bool) or not isinstance(value, int):
+        raise ValueError(
+            f"{name} must be an int (the server reads it as an unsigned 32-bit "
+            f"integer and stores NULL for anything else), got {type(value).__name__}: {value!r}"
+        )
+    if not 0 <= value <= _U32_MAX:
+        raise ValueError(
+            f"{name} must be between 0 and {_U32_MAX} (an unsigned 32-bit "
+            f"integer), got {value!r}"
+        )
+
+
+def _validate_promoted_string(name: str, value) -> None:
+    """Reject anything `ps()` would silently turn into NULL.
+
+    Same reasoning as `_validate_promoted_numeric`, and the same boundary: this
+    is the last point where the caller still has a stack trace pointing at their
+    own call. `_build` copies the base dict verbatim, so a `None` here reached
+    the wire as an explicit JSON `null`, the row was accepted at 200 OK, and the
+    column was empty for some events and not others with nothing logged anywhere.
+    """
+    if value is None:
+        raise ValueError(
+            f"{name} must be a string (the server lifts it into an indexed column "
+            f"and stores NULL for anything else, so the event would be accepted at "
+            f"200 OK and be invisible to every filter on {name}), got None"
+        )
+    if not isinstance(value, str):
+        raise TypeError(
+            f"{name} must be a string (the server lifts it into an indexed column "
+            f"and stores NULL for anything else), got {type(value).__name__}: {value!r}"
+        )
+
+
+def _validate_identity(name: str, value) -> None:
+    """Reject an id the server will skip, at the point the caller can see it.
+
+    `session_id` and `agent_id` are on every one of the 15 event types and are
+    what everything downstream groups by. Ingest requires each to be a JSON
+    string: hand it anything else and the row is SKIPPED — and the response is
+    `200 OK` with `{"accepted": 0, "skipped": 1}`, so nothing upstream learns.
+    The SDK reports success, the collector deletes the batch, and the event is
+    gone. Verified against the live server for int, null and object.
+
+    `None` is the realistic way in — an uninitialised variable, a dict lookup
+    that missed, an id threaded through a code path that forgot to set it. The
+    caller does not get a wrong number here, they get no data at all, and the
+    only symptom is a dashboard that is emptier than it should be.
+
+    Empty and whitespace-only are refused as well, and those the server DOES
+    accept. That is the worse outcome of the two: every event lands, grouped
+    under one blank id, so the data looks present and is silently merged.
+    """
+    if not isinstance(value, str):
+        # TypeError, not ValueError: the wrong TYPE was supplied, and this is
+        # also what a caller got before identity became optional — omitting a
+        # required keyword argument has always been a TypeError, so code that
+        # catches one keeps working.
+        raise TypeError(
+            f"{name} must be a str — the server skips any event whose {name} is "
+            f"not a JSON string, and answers 200 as though it stored it. "
+            f"Got {type(value).__name__}: {value!r}"
+        )
+    if not value.strip():
+        raise ValueError(
+            f"{name} must not be empty — the server accepts it, so every event "
+            f"sent this way is silently grouped under one blank id."
+        )
+
+def _resolve_identity(session_id, agent_id) -> "tuple[str, str]":
+    """Fill an omitted `session_id`/`agent_id` from the ambient scope, then validate.
+
+    Both arguments were required on all 15 methods, which meant threading them
+    through every function that might emit — the diff nobody wants to review, and
+    the reason the reference integration shipped a contextvars wrapper as markdown
+    for customers to paste in. `session()` / `agent()` bind them instead.
+
+    ORDER MATTERS. The validation runs on the RESOLVED value, not the argument.
+    Validating first would reject every ambient call; resolving without validating
+    would put the silent-skip back: ingest drops an event whose `session_id` is not
+    a JSON string and answers `200 OK` with `{"accepted":0,"skipped":1}`, so a run
+    with nothing bound would vanish rather than fail.
+
+    Called AFTER `_validate_fields`, deliberately. A reserved `**field` is a
+    fault in the call itself and reads identically from anywhere, so reporting it
+    first gives a stable, reproducible message; the identity error depends on
+    where the call was made from, and is the less useful of the two to hear when
+    both are true.
+
+    `agent_id` falls back to `DEFAULT_AGENT_ID` rather than raising — an event
+    emitted inside `session()` with no `agent()` around it lands somewhere sensible,
+    which is the convention the skill already teaches. `session_id` has no such
+    default: inventing one would scatter a run across as many sessions as it has
+    emit sites.
+    """
+    if session_id is None:
+        session_id = _context.session_id()
+    if agent_id is None:
+        agent_id = _context.agent_id()  # DEFAULT_AGENT_ID when no agent scope
+
+    if session_id is None:
+        # TypeError for the same reason: this is a missing required argument,
+        # which is what it literally was until the scopes made it optional.
+        raise TypeError(
+            "session_id is required and nothing is bound. Pass session_id=..., or "
+            "wrap the call in `with failproofai_sdk.session():` / "
+            "`with failproofai_sdk.agent(\"name\"):`. A new thread does not inherit "
+            "the ambient scope — hand work to it with "
+            "`failproofai_sdk.propagate(fn)`."
+        )
+    _validate_identity("session_id", session_id)
+    _validate_identity("agent_id", agent_id)
+    return session_id, agent_id
+
+
+def _measured_duration_ms(start_ts, end_ts) -> "int | None":
+    """Whole milliseconds between a paired start and end, or None.
+
+    The four paired events each computed this inline, and none of them applied
+    the range the SERVER enforces — the same range `_validate_promoted_numeric`
+    refuses a caller for. `pu32()` reads `duration_ms` as an unsigned 32-bit
+    integer and stores NULL for anything outside it, at `200 OK`, so an
+    out-of-range duration is not an error anywhere: the row lands with an empty
+    column and nothing says why.
+
+    Two ways to leave the range, both reachable without anything being wrong
+    with the caller:
+
+    * **over.** 2**32 ms is ~49.7 days. A `human_wait` answered after a long
+      weekend, or an `agent_pause` resumed a month later, is an ordinary
+      lifetime for these pairs, not an abuse of them.
+    * **under.** These are wall-clock readings from `datetime.now()`, so an NTP
+      step backwards between start and end yields a NEGATIVE interval. `round()`
+      keeps the sign, and a negative into an unsigned column is the same silent
+      NULL.
+
+    Omitted rather than clamped. A clamped 49.7 days is indistinguishable from a
+    measurement, and the whole reason `duration_ms` is computed here instead of
+    accepted from the caller is that a reported duration is unfalsifiable. An
+    absent field is at least honest, and the timestamps are still on both events
+    for anyone who wants to do the subtraction themselves.
+    """
+    if start_ts is None:
+        return None
+    ms = round((end_ts - start_ts).total_seconds() * 1000)
+    if not 0 <= ms <= _U32_MAX:
+        logger.warning(
+            "Failproof AI omitted duration_ms=%d: outside the unsigned 32-bit range "
+            "the server stores it in (0..%d). The event is unaffected.",
+            ms,
+            _U32_MAX,
+        )
+        return None
+    return ms
+
+
+# Hard cap on `_pending` correlation map size. Orphaned starts (a `tool_use` with
+# no `tool_result`, a `human_wait` the user never answers, etc.) would otherwise
+# grow this dict unbounded in a long-running process. At the cap we evict the
+# oldest entry FIFO — Python dicts preserve insertion order since 3.7.
+_PENDING_CAP = 10_000
+
+
+# Every pairing in `_pending` is keyed by what it pairs and by the SESSION it
+# belongs to — and deliberately NOT by the agent.
+#
+# The rule: include what makes the id unique, exclude what can legitimately
+# change between the start event and the end event.
+#
+# * KIND belongs in the key. Tool pairs keyed on the bare `tool_call_id` and hook
+#   pairs on the bare `hook_id` shared one flat keyspace, so a caller whose tool
+#   call and hook happened to share an id — not exotic, both are frequently the
+#   harness's own step id — got a `hook_completed` that consumed the `tool_use`
+#   timestamp, and then a `tool_result` with no duration at all.
+#
+# * SESSION belongs in the key. `_pending` lives on one process-wide namespace,
+#   so two sessions in one process — a supervisor running agents concurrently,
+#   the ordinary multi-agent shape — collided on any shared step id. Starting
+#   `step-1` in session A and then in session B overwrote A's timestamp; A's
+#   result reported B's interval and B's reported none.
+#
+# * AGENT DOES NOT. This was the tempting third component and it is wrong. Once a
+#   framework runs tools inside sub-agents, a `tool_use` opened under `planner`
+#   and closed under `worker` is routine — LangGraph and CrewAI both do it — and
+#   an agent-scoped key makes those pairs miss entirely, silently dropping
+#   `duration_ms` for exactly the nested runs that most need it. A session cannot
+#   change under a pair; an agent can.
+#
+# These are correlation keys only; they are never emitted and never leave the
+# process, so the shape changes no wire format. It only changes `duration_ms` in
+# the colliding cases, from a fabricated or missing value to a correct one.
+def _tool_key(session_id: str, tool_call_id: str) -> str:
+    return f"tool:{session_id}:{tool_call_id}"
+
+
+def _hook_key(session_id: str, hook_id: str) -> str:
+    return f"hook:{session_id}:{hook_id}"
+
+
+class EventNamespace:
+    def __init__(self, writer) -> None:
+        self._writer = writer
+        self._pending: dict[str, datetime] = {}
+
+    def _track_pending(self, key: str, ts: datetime) -> None:
+        # Evicting has to tolerate another thread doing the same thing, because
+        # this runs on the CALLER'S agent loop and a raise here is a crash in
+        # their code, not a lost measurement.
+        #
+        # `len()` then `next(iter())` then remove is a read-modify-write, and
+        # nothing serialises the three. Two threads arriving at the cap together
+        # pick the SAME victim, and the second `del` raised KeyError straight
+        # out of `event.tool_use()`. Reproduced at 24 crashes per 30_000 calls
+        # across 10 threads — and only once `_pending` is full, which is the
+        # long-running multi-agent process this cap exists for in the first
+        # place. `next(iter())` has two more shapes for the same reason:
+        # StopIteration if the dict was emptied, RuntimeError if it was resized
+        # between the iterator and the first step.
+        #
+        # No lock, deliberately. `_pending` operations are nanoseconds and a
+        # lock held at the instant of a `fork()` is inherited locked by a thread
+        # that does not exist in the child — the exact hazard `_writer` rebuilds
+        # its Event and lock to avoid. Tolerant operations have no such edge.
+        #
+        # The cost is that the cap is approximate under contention: several
+        # threads may insert after each evicting once. That is the same trade
+        # `EventWriter.submit` already makes — a backstop against unbounded
+        # growth, not an exact quota.
+        if len(self._pending) >= _PENDING_CAP:
+            try:
+                oldest = next(iter(self._pending))
+            except (StopIteration, RuntimeError):  # emptied or resized under us
+                pass
+            else:
+                self._pending.pop(oldest, None)  # may already be gone
+        self._pending[key] = ts
+
+    def _validate_fields(self, fields: dict) -> None:
+        bad = _RESERVED & fields.keys()
+        if bad:
+            raise ValueError(f"Reserved field names cannot be used as custom fields: {sorted(bad)}")
+        for name in _PROMOTED_NUMERIC & fields.keys():
+            _validate_promoted_numeric(name, fields[name])
+        for name in _PROMOTED_STRING & fields.keys():
+            _validate_promoted_string(name, fields[name])
+
+    @staticmethod
+    def _now() -> datetime:
+        return datetime.now(timezone.utc)
+
+    @staticmethod
+    def _fmt_ts(dt: datetime) -> str:
+        return dt.strftime("%Y-%m-%dT%H:%M:%S.%f") + "Z"
+
+    def tool_use(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        tool_name: str,
+        tool_call_id: str,
+        input: dict | None = None,
+        **fields,
+    ) -> None:
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        self._track_pending(_tool_key(session_id, tool_call_id), ts)
+        self._writer.submit(
+            ToolUseEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                tool_name=tool_name,
+                tool_call_id=tool_call_id,
+                input=input,
+                extra_fields=fields,
+            ).to_dict()
+        )
+
+    def tool_result(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        tool_name: str,
+        tool_call_id: str,
+        output: Any | None = None,
+        error: str | None = None,
+        **fields,
+    ) -> None:
+        if "duration_ms" in fields:
+            raise ValueError("duration_ms is auto-computed by the SDK and cannot be passed by the caller")
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        start_ts = self._pending.pop(_tool_key(session_id, tool_call_id), None)
+        duration_ms = _measured_duration_ms(start_ts, ts)
+        self._writer.submit(
+            ToolResultEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                tool_name=tool_name,
+                tool_call_id=tool_call_id,
+                output=output,
+                error=error,
+                duration_ms=duration_ms,
+                extra_fields=fields,
+            ).to_dict()
+        )
+
+    def model_request(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        model: str | None = None,
+        messages: list[dict] | None = None,
+        system: Any | None = None,
+        tools: list[dict] | None = None,
+        request_id: str | None = None,
+        **fields,
+    ) -> None:
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        self._writer.submit(
+            ModelRequestEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                model=model,
+                messages=messages,
+                system=system,
+                tools=tools,
+                request_id=request_id,
+                extra_fields=fields,
+            ).to_dict()
+        )
+
+    def model_response(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        model: str | None = None,
+        stop_reason: str | None = None,
+        input_tokens: int | None = None,
+        output_tokens: int | None = None,
+        content: Any | None = None,
+        role: str | None = None,
+        request_id: str | None = None,
+        **fields,
+    ) -> None:
+        # Named parameters, so they never reach `_validate_fields`. They are also
+        # the likeliest of the three to arrive wrong: a caller reading them off a
+        # provider's usage object gets whatever that object holds.
+        _validate_promoted_numeric("input_tokens", input_tokens)
+        _validate_promoted_numeric("output_tokens", output_tokens)
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        self._writer.submit(
+            ModelResponseEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                model=model,
+                stop_reason=stop_reason,
+                input_tokens=input_tokens,
+                output_tokens=output_tokens,
+                content=content,
+                role=role,
+                request_id=request_id,
+                extra_fields=fields,
+            ).to_dict()
+        )
+
+    def agent_start(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        goal: str | None = None,
+        parent_id: str | None = None,
+        **fields,
+    ) -> None:
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        self._writer.submit(
+            AgentStartEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                goal=goal,
+                parent_id=parent_id,
+                extra_fields=fields,
+            ).to_dict()
+        )
+
+    def agent_end(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        outcome: str | None = None,
+        summary: str | None = None,
+        **fields,
+    ) -> None:
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        self._writer.submit(
+            AgentEndEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                outcome=outcome,
+                summary=summary,
+                extra_fields=fields,
+            ).to_dict()
+        )
+
+    def agent_pause(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        pause_id: str,
+        reason: str | None = None,
+        user_id: str | None = None,
+        **fields,
+    ) -> None:
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        self._track_pending(f"pause:{session_id}:{pause_id}", ts)
+        self._writer.submit(
+            AgentPauseEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                pause_id=pause_id,
+                reason=reason,
+                user_id=user_id,
+                extra_fields=fields,
+            ).to_dict()
+        )
+
+    def agent_resume(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        pause_id: str,
+        reason: str | None = None,
+        user_id: str | None = None,
+        **fields,
+    ) -> None:
+        if "duration_ms" in fields:
+            raise ValueError("duration_ms is auto-computed by the SDK and cannot be passed by the caller")
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        start_ts = self._pending.pop(f"pause:{session_id}:{pause_id}", None)
+        duration_ms = _measured_duration_ms(start_ts, ts)
+        self._writer.submit(
+            AgentResumeEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                pause_id=pause_id,
+                duration_ms=duration_ms,
+                reason=reason,
+                user_id=user_id,
+                extra_fields=fields,
+            ).to_dict()
+        )
+
+    def hook_triggered(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        hook_name: str,
+        hook_id: str,
+        trigger_event: str | None = None,
+        input: Any | None = None,
+        **fields,
+    ) -> None:
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        self._track_pending(_hook_key(session_id, hook_id), ts)
+        self._writer.submit(
+            HookTriggeredEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                hook_name=hook_name,
+                hook_id=hook_id,
+                trigger_event=trigger_event,
+                input=input,
+                extra_fields=fields,
+            ).to_dict()
+        )
+
+    def hook_completed(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        hook_name: str,
+        hook_id: str,
+        outcome: str | None = None,
+        output: Any | None = None,
+        error: str | None = None,
+        **fields,
+    ) -> None:
+        if "duration_ms" in fields:
+            raise ValueError("duration_ms is auto-computed by the SDK and cannot be passed by the caller")
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        start_ts = self._pending.pop(_hook_key(session_id, hook_id), None)
+        duration_ms = _measured_duration_ms(start_ts, ts)
+        self._writer.submit(
+            HookCompletedEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                hook_name=hook_name,
+                hook_id=hook_id,
+                outcome=outcome,
+                output=output,
+                error=error,
+                duration_ms=duration_ms,
+                extra_fields=fields,
+            ).to_dict()
+        )
+
+    def error(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        error_type: str,
+        message: str,
+        traceback: str | None = None,
+        **fields,
+    ) -> None:
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        self._writer.submit(
+            ErrorEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                error_type=error_type,
+                message=message,
+                traceback=traceback,
+                extra_fields=fields,
+            ).to_dict()
+        )
+
+    def human_wait(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        input_id: str,
+        prompt: str | None = None,
+        options: list[str] | None = None,
+        reason: str | None = None,
+        **fields,
+    ) -> None:
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        self._track_pending(f"human:{session_id}:{input_id}", ts)
+        self._writer.submit(
+            HumanWaitEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                input_id=input_id,
+                prompt=prompt,
+                options=options,
+                reason=reason,
+                extra_fields=fields,
+            ).to_dict()
+        )
+
+    def human_input(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        input_id: str,
+        response: str | None = None,
+        **fields,
+    ) -> None:
+        if "duration_ms" in fields:
+            raise ValueError("duration_ms is auto-computed by the SDK and cannot be passed by the caller")
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        start_ts = self._pending.pop(f"human:{session_id}:{input_id}", None)
+        duration_ms = _measured_duration_ms(start_ts, ts)
+        self._writer.submit(
+            HumanInputEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                input_id=input_id,
+                response=response,
+                duration_ms=duration_ms,
+                extra_fields=fields,
+            ).to_dict()
+        )
+
+    def human_pause(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        reason: str | None = None,
+        user_id: str | None = None,
+        **fields,
+    ) -> None:
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        self._writer.submit(
+            HumanPauseEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                reason=reason,
+                user_id=user_id,
+                extra_fields=fields,
+            ).to_dict()
+        )
+
+    def human_interrupt(
+        self,
+        *,
+        session_id: str | None = None,
+        agent_id: str | None = None,
+        reason: str | None = None,
+        user_id: str | None = None,
+        at_step: str | None = None,
+        **fields,
+    ) -> None:
+        self._validate_fields(fields)
+        session_id, agent_id = _resolve_identity(session_id, agent_id)
+        ts = self._now()
+        self._writer.submit(
+            HumanInterruptEvent(
+                timestamp=self._fmt_ts(ts),
+                session_id=session_id,
+                agent_id=agent_id,
+                reason=reason,
+                user_id=user_id,
+                at_step=at_step,
+                extra_fields=fields,
+            ).to_dict()
+        )