Files
hermes-agent/hermes_cli/observability/shared_metrics_fields.py
teknium1 85c317b0a9 fix(telemetry): feature_adoption ignores Hermes-internal curator/skill work; age read never raises (M15, m21)
- `curator` latched on the default-on scheduled pass (outcome=success, all
  buckets 0), so it meant "idled through one curator interval". It now needs
  trigger=manual (the user ran it).
- `skills_created` and the v4 milestone `first_skill_created` latched when
  the background-review fork created a skill (provenance=agent_created).
  Both now require provenance != agent_created (foreground creates are
  stamped "learn" -> not agent_created, see tools/skill_usage.record_created).
- days_since_install_bucket caught only sqlite3.Error; a non-numeric
  sessions.started_at raised ValueError into the subscriber on every event,
  re-opened state.db each time and never latched. It now also catches
  OSError/TypeError/ValueError -> `unknown`.

Probes:
  probe_internal_adoption.py
    before: feature_adoption curator + skills_created, milestone first_skill_created
    after:  no feature_adoption rows, no milestone
  probe_statedb_errors.py textstamp
    before: ValueError warnings, 0 rows, settled 0 of 1
    after:  1 state.db read, feature_adoption {unknown, projects}, settled 1 of 1
Tests test_hermes_internal_work_never_latches_adoption and
test_unreadable_first_session_reads_unknown are RED on base.
2026-09-28 12:43:03 -07:00

202 lines
8.3 KiB
Python

"""Bounded field builders for the decision-data shared metrics.
Sessions, install milestones, setup completion, token volumes, compression, model switches,
fallbacks, slash commands and extension installs. Every builder takes RAW runtime values and
returns only closed-enum, bucketed or public-catalog dimensions (see shared_metrics_contract).
"""
from __future__ import annotations
from typing import Any
from . import shared_metrics_catalog as catalog
from . import shared_metrics_contract as contract
from .shared_metrics_contract import (
MODEL_IDENTIFIER_MAX_LENGTH, PROVIDER_IDENTIFIER_MAX_LENGTH, _bucket, _metric_identifier, _norm,
)
from .shared_metrics_signals import user_created_skill
_MINUTE_MS = 60_000
_SESSION_DURATION_THRESHOLDS = (
(_MINUTE_MS, "lt_1m"), (5 * _MINUTE_MS, "1m_to_5m"), (30 * _MINUTE_MS, "5m_to_30m"),
(120 * _MINUTE_MS, "30m_to_2h"), (480 * _MINUTE_MS, "2h_to_8h"),
)
_HOUR_S = 3_600
_INSTALL_AGE_THRESHOLDS = (
(_HOUR_S, "lt_1h"), (24 * _HOUR_S, "1h_to_1d"), (7 * 24 * _HOUR_S, "1d_to_7d"),
(30 * 24 * _HOUR_S, "7d_to_30d"), (90 * 24 * _HOUR_S, "30d_to_90d"),
)
_TTFT_THRESHOLDS = (
(0.5, "lt_500ms"), (1.0, "500ms_to_1s"), (2.0, "1s_to_2s"), (5.0, "2s_to_5s"), (15.0, "5s_to_15s"),
)
_CONTEXT_FILL_THRESHOLDS = ((50, "lt_50"), (75, "50_to_75"), (90, "75_to_90"), (100, "90_to_100"))
# A session this long is the "real use" signal (vs a one-shot try-out).
LONG_SESSION_TURNS = 6
def _number(value: Any) -> float | None:
return contract._non_negative_number(value)
def provider_identifier(value: Any) -> str:
return catalog.provider_metric_name(value)
def session_fields(
start_fields: dict[str, str], *, turns: int, failed_turns: int, last_outcome: str, active_ms: int,
) -> dict[str, str]:
"""One closed session: where it ran, how many turns, how long it was active, how it ended."""
return {
"active_duration_bucket": _bucket(max(0, active_ms), _SESSION_DURATION_THRESHOLDS, "gte_8h"),
"entrypoint": start_fields.get("entrypoint", "unknown"),
"execution_surface": start_fields.get("execution_surface", "unknown"),
"failed_turn_count_bucket": contract.count_bucket(failed_turns),
"last_outcome": last_outcome if last_outcome in contract.TASK_OUTCOMES else "unknown",
"platform": start_fields.get("platform", "none"),
"turn_count_bucket": contract.size_bucket(turns),
}
def install_age_bucket(age_seconds: Any) -> str:
age = _number(age_seconds)
return "unknown" if age is None else _bucket(age, _INSTALL_AGE_THRESHOLDS, "gte_90d")
def ttft_bucket(event: dict[str, Any]) -> str:
"""Time to first streamed chunk for one primary call; ``not_streamed`` when nothing streamed."""
started, first = _number(event.get("started_at")), _number(event.get("first_chunk_at"))
if started is None:
return "unknown"
if first is None:
return "not_streamed"
return _bucket(max(0.0, first - started), _TTFT_THRESHOLDS, "gte_15s")
def context_fill_bucket(tokens_before: Any, context_length: Any) -> str:
tokens, length = _number(tokens_before), _number(context_length)
if tokens is None or not length:
return "unknown"
return _bucket(100 * tokens / length, _CONTEXT_FILL_THRESHOLDS, "gte_100")
_USAGE_KEYS = {
"input": "input_tokens", "output": "output_tokens", "cache_read": "cache_read_tokens",
"cache_write": "cache_write_tokens", "reasoning": "reasoning_tokens",
}
def model_token_fields(
usage: Any, *, model: Any, provider: Any, call_role: str, aux_task: Any = None,
) -> dict[str, Any] | None:
"""Token-usage mark data (dimensions + integer amounts); None when there is no usage."""
if not isinstance(usage, dict):
return None
amounts = {}
for token_type, key in _USAGE_KEYS.items():
value = usage.get(key)
amounts[token_type] = value if isinstance(value, int) and not isinstance(value, bool) and value > 0 else 0
if not any(amounts.values()):
return None
return {
"aux_task": catalog.aux_task_metric_name(aux_task) if call_role == "auxiliary" else "none",
"call_role": call_role,
"model": catalog.model_metric_name(
model, provider_identifier(provider), max_length=MODEL_IDENTIFIER_MAX_LENGTH,
),
"provider": provider_identifier(provider),
**amounts,
}
def setup_completed_fields(*, surface: Any, provider: Any) -> dict[str, str]:
value = _norm(surface)
return {
"provider": provider_identifier(provider) if provider else "none",
"surface": value if value in contract.SETUP_SURFACES else "other",
}
_COMPRESSION_TRIGGERS = {
**dict.fromkeys(("auto", "threshold", "preflight", "auto_threshold"), "auto"),
**dict.fromkeys(("overflow", "context_overflow", "overflow_error", "error"), "overflow"),
**dict.fromkeys(("manual", "user", "command", "slash"), "manual"),
}
_COMPRESSION_OUTCOMES = {
**dict.fromkeys(("success", "ok", "compressed", "completed"), "success"),
**dict.fromkeys(("skipped", "noop", "not_needed"), "skipped"),
}
def compression_fields(*, trigger: Any, outcome: Any, tokens_before: Any, context_length: Any) -> dict[str, str]:
return {
"context_fill_bucket": context_fill_bucket(tokens_before, context_length),
"outcome": _COMPRESSION_OUTCOMES.get(_norm(outcome), "failed"),
"trigger": _COMPRESSION_TRIGGERS.get(_norm(trigger), "other"),
}
def model_switch_fields(*, from_provider: Any, to_provider: Any, surface: Any) -> dict[str, str]:
return {
"execution_surface": contract.execution_surface({"execution_surface": surface}),
"from_provider": provider_identifier(from_provider),
"to_provider": provider_identifier(to_provider),
}
def fallback_fields(*, from_provider: Any, to_provider: Any, reason: Any) -> dict[str, str]:
value = getattr(reason, "value", reason)
return {
"error_class": contract.model_error_class({"reason": value}),
"from_provider": provider_identifier(from_provider),
"to_provider": provider_identifier(to_provider),
}
def slash_command_fields(*, command: Any, surface: Any) -> dict[str, str]:
return {
"command": catalog.slash_command_metric_name(command),
"execution_surface": contract.execution_surface({"execution_surface": surface}),
}
def extension_install_fields(*, kind: Any, source: Any, name: Any, outcome: Any) -> dict[str, str] | None:
kind_value = _norm(kind)
if kind_value not in contract.EXTENSION_KINDS:
return None
source_value = _norm(source)
return {
"kind": kind_value,
"name": catalog.extension_metric_name(kind_value, name),
"outcome": "success" if _norm(outcome) in {"success", "ok", "installed"} else "failed",
"source": source_value if source_value in contract.EXTENSION_SOURCES else "other",
}
def _long_session(d: dict[str, str]) -> bool:
return d.get("turn_count_bucket") not in {"0", "1", "2", "3_to_5"}
# metric -> [(milestone, predicate over its dimensions)]
_MILESTONE_RULES = {
contract.TASK_STARTED_METRIC: (
("first_task_started", lambda d: True),
("first_gateway_message", lambda d: d.get("platform") not in {None, "none"}),
("first_scheduled_task", lambda d: d.get("entrypoint") == "scheduled_task"),
),
contract.TASK_FINISHED_METRIC: (("first_task_success", lambda d: d.get("outcome") == "success"),),
contract.TOOL_USAGE_METRIC: (
("first_tool_success", lambda d: d.get("outcome") == "success"),
("first_mcp_tool_success", lambda d: d.get("outcome") == "success" and d.get("tool_name") == "mcp"),
("first_delegation", lambda d: d.get("outcome") == "success" and d.get("tool_name") == "delegate_task"),
),
contract.SKILL_LIFECYCLE_METRIC: (("first_skill_created", user_created_skill),),
contract.SKILL_LOAD_METRIC: (("first_skill_reused", lambda d: d.get("reuse_state") == "reused"),),
contract.SESSION_METRIC: (("first_long_session", _long_session),),
contract.SETUP_COMPLETED_METRIC: (("setup_completed", lambda d: True),),
}
def milestones_for(metric_name: str, dimensions: dict[str, str]) -> tuple[str, ...]:
"""Install milestones a recorded counter reaches (each is latched once per install by the store)."""
return tuple(name for name, reached in _MILESTONE_RULES.get(metric_name, ()) if reached(dimensions))