hermes-agent/tests/gateway/test_delivery_ledger.py
Teknium 5854aad8b5
feat(gateway): durable delivery-obligation ledger for final responses (#67181)
A final response generated but not confirmed-delivered was the one
artifact the gateway could lose without a trace: crash or planned
restart between finalize and platform ACK dropped it silently, and the
resume path re-ran the whole turn at full cost (#58818 P1, #41696,
#63695's gateway half).

gateway/delivery_ledger.py records each outbound final response in
state.db (same conventions as the async-delegation ledger: WAL, owner
pid + process-start-time liveness, bounded retention):

  pending -> attempting -> delivered | failed
  startup sweep on dead-owner rows -> redeliver | abandoned

Contract (the lessons from the closed delivery-outbox attempt #61790):
- obligation recorded BEFORE the first send attempt; cleared only on
  SendResult.success (destination acceptance, #51184)
- ambiguity is labeled, never silently retried: rows that were mid-send
  when the process died redeliver with a visible '♻️ Recovered reply —
  may be a duplicate' prefix (honest at-least-once)
- stable ids from session_key + inbound message id + content, so
  distinct threads/topics can never collide
- poison rows bounded: 3 attempts / 24h freshness -> abandoned; claim
  atomically re-stamps ownership so racing sweeps can't double-claim
- redelivery clears resume_pending for the session so the resume path
  never re-runs a turn whose answer the ledger already holds
- best-effort everywhere: ledger failure can never block or delay a send
- slash-command/ephemeral/empty responses are not recorded; cron and
  proactive delivery stay on DeliveryRouter (separate subsystem)

Config: gateway.delivery_ledger (default on; no version bump needed).

Validation: 30 ledger+producer tests; 352 blast-radius gateway tests
green; cross-process E2E (record in process A, kill it mid-send, claim
+ marker + redeliver in a fresh process B against the same state.db).
2026-07-19 00:45:32 -07:00

283 lines
9.5 KiB
Python

"""Tests for the gateway delivery-obligation ledger (gateway/delivery_ledger.py).
State machine, dead-owner claiming, attempts cap, stale cutoff, retention,
id stability, and the startup redelivery sweep's contract:
- pending rows redeliver plainly (send never started, no dup risk)
- attempting/failed rows carry the recovered-reply marker (honest
at-least-once; ambiguity is labeled, never silently resent)
- rows owned by a LIVE process are never claimed
- poison rows abandon at the attempts cap / stale cutoff
"""
import time
from unittest.mock import AsyncMock, MagicMock, patch
import pytest
from gateway import delivery_ledger as dl
@pytest.fixture(autouse=True)
def _fresh_db(tmp_path, monkeypatch):
"""Isolated state.db per test (autouse HERMES_HOME isolation already
redirects get_hermes_home; make the redirect explicit and per-test)."""
home = tmp_path / ".hermes"
home.mkdir()
monkeypatch.setattr(dl, "_db_path", lambda: home / "state.db")
yield
def _record(oid="ob-1", session_key="agent:main:slack:channel:C1", **kw):
dl.record_obligation(
obligation_id=oid,
session_key=session_key,
platform=kw.get("platform", "slack"),
chat_id=kw.get("chat_id", "C1"),
thread_id=kw.get("thread_id", "171.001"),
content=kw.get("content", "the final answer"),
)
def _row(oid):
with dl._connect() as conn:
r = conn.execute(
"""SELECT state, attempts, owner_pid, content
FROM delivery_obligations WHERE obligation_id=?""",
(oid,),
).fetchone()
return None if r is None else {
"state": r[0], "attempts": r[1], "owner_pid": r[2], "content": r[3],
}
def _orphan(oid):
"""Make the row look like it belongs to a dead process."""
with dl._connect() as conn:
conn.execute(
"UPDATE delivery_obligations SET owner_pid=999999999, "
"owner_started_at=1 WHERE obligation_id=?",
(oid,),
)
class TestStateMachine:
def test_record_starts_pending(self):
_record()
assert _row("ob-1")["state"] == "pending"
def test_full_happy_path(self):
_record()
dl.mark_attempting("ob-1")
assert _row("ob-1")["state"] == "attempting"
dl.mark_delivered("ob-1")
assert _row("ob-1")["state"] == "delivered"
def test_failed_records_error(self):
_record()
dl.mark_attempting("ob-1")
dl.mark_failed("ob-1", "chat_not_found")
assert _row("ob-1")["state"] == "failed"
def test_rerecord_same_id_is_idempotent(self):
_record()
dl.mark_attempting("ob-1")
_record() # INSERT OR REPLACE resets to pending — same turn re-record
assert _row("ob-1")["state"] == "pending"
class TestObligationId:
def test_stable_and_distinct(self):
a = dl.compute_obligation_id("sk1", "msg1", "hello")
assert a == dl.compute_obligation_id("sk1", "msg1", "hello")
# Different thread (baked into session_key) → different id. This is
# the cron-topic collision class from the earlier outbox attempt.
assert a != dl.compute_obligation_id("sk1:threadB", "msg1", "hello")
assert a != dl.compute_obligation_id("sk1", "msg2", "hello")
assert a != dl.compute_obligation_id("sk1", "msg1", "other")
assert len(a) == 24
class TestSweep:
def test_live_owner_rows_never_claimed(self):
_record() # owner = this (live) process
assert dl.sweep_recoverable() == []
def test_dead_owner_pending_claimed_without_marker(self):
_record()
_orphan("ob-1")
claimed = dl.sweep_recoverable()
assert len(claimed) == 1
assert claimed[0]["needs_marker"] is False
assert claimed[0]["attempts"] == 1
# Claim re-stamps ownership: a second sweep in the same (live)
# process must not double-claim.
assert dl.sweep_recoverable() == []
def test_dead_owner_attempting_needs_marker(self):
_record()
dl.mark_attempting("ob-1")
_orphan("ob-1")
claimed = dl.sweep_recoverable()
assert claimed[0]["needs_marker"] is True
def test_dead_owner_failed_needs_marker(self):
_record()
dl.mark_failed("ob-1", "boom")
_orphan("ob-1")
claimed = dl.sweep_recoverable()
assert claimed[0]["needs_marker"] is True
def test_delivered_rows_ignored(self):
_record()
dl.mark_delivered("ob-1")
_orphan("ob-1")
assert dl.sweep_recoverable() == []
def test_attempts_cap_abandons(self):
_record()
_orphan("ob-1")
with dl._connect() as conn:
conn.execute(
"UPDATE delivery_obligations SET attempts=? WHERE obligation_id=?",
(dl.MAX_ATTEMPTS, "ob-1"),
)
assert dl.sweep_recoverable() == []
assert _row("ob-1")["state"] == "abandoned"
def test_stale_cutoff_abandons(self):
_record()
_orphan("ob-1")
future = time.time() + dl.STALE_AFTER_SECONDS + 60
assert dl.sweep_recoverable(now=future) == []
assert _row("ob-1")["state"] == "abandoned"
class TestPrune:
def test_old_delivered_rows_pruned(self):
_record()
dl.mark_delivered("ob-1")
with dl._connect() as conn:
conn.execute(
"UPDATE delivery_obligations SET updated_at=? WHERE obligation_id=?",
(time.time() - dl._RETENTION_SECONDS - 60, "ob-1"),
)
dl._prune()
assert _row("ob-1") is None
def test_undelivered_rows_survive_retention(self):
_record()
with dl._connect() as conn:
conn.execute(
"UPDATE delivery_obligations SET updated_at=? WHERE obligation_id=?",
(time.time() - dl._RETENTION_SECONDS - 60, "ob-1"),
)
dl._prune()
assert _row("ob-1") is not None
class TestLedgerEnabled:
def test_default_on(self):
assert dl.ledger_enabled({}) is True
assert dl.ledger_enabled({"gateway": {}}) is True
def test_explicit_off(self):
assert dl.ledger_enabled({"gateway": {"delivery_ledger": False}}) is False
assert dl.ledger_enabled({"gateway": {"delivery_ledger": "off"}}) is False
def test_truthy_strings(self):
assert dl.ledger_enabled({"gateway": {"delivery_ledger": "true"}}) is True
class TestGatewayRedeliverySweep:
"""Drive the real GatewayRunner._redeliver_pending_obligations."""
@staticmethod
def _runner(adapter=None):
from gateway.config import Platform
from gateway.run import GatewayRunner
runner = object.__new__(GatewayRunner)
runner.adapters = {Platform.SLACK: adapter} if adapter else {}
_store = MagicMock()
_store.clear_resume_pending = AsyncMock()
_store._store = None
runner.session_store = None
runner._async_session_store = _store
return runner
@staticmethod
def _adapter(success=True):
adapter = MagicMock()
adapter.send = AsyncMock(
return_value=MagicMock(success=success, error="" if success else "nope")
)
return adapter
@pytest.mark.asyncio
async def test_pending_redelivers_plain_and_clears_resume(self):
_record() # pending
_orphan("ob-1")
adapter = self._adapter()
runner = self._runner(adapter)
n = await runner._redeliver_pending_obligations()
assert n == 1
sent = adapter.send.call_args.kwargs
assert sent["content"] == "the final answer" # no marker
assert sent["metadata"] == {"thread_id": "171.001"}
assert _row("ob-1")["state"] == "delivered"
runner._async_session_store.clear_resume_pending.assert_awaited_once_with(
"agent:main:slack:channel:C1"
)
@pytest.mark.asyncio
async def test_attempting_redelivers_with_marker(self):
_record()
dl.mark_attempting("ob-1")
_orphan("ob-1")
adapter = self._adapter()
runner = self._runner(adapter)
await runner._redeliver_pending_obligations()
sent = adapter.send.call_args.kwargs
assert sent["content"].startswith(dl.RECOVERED_MARKER)
assert sent["content"].endswith("the final answer")
@pytest.mark.asyncio
async def test_send_failure_marks_failed_for_next_boot(self):
_record()
_orphan("ob-1")
runner = self._runner(self._adapter(success=False))
n = await runner._redeliver_pending_obligations()
assert n == 0
assert _row("ob-1")["state"] == "failed"
@pytest.mark.asyncio
async def test_missing_adapter_leaves_row_recoverable(self):
_record()
_orphan("ob-1")
runner = self._runner(adapter=None) # slack not connected
n = await runner._redeliver_pending_obligations()
assert n == 0
# Row still claimed by us but NOT delivered/abandoned — a later boot
# (attempts cap permitting) can retry once the platform connects.
assert _row("ob-1")["state"] == "pending"
@pytest.mark.asyncio
async def test_disabled_gate_short_circuits(self):
_record()
_orphan("ob-1")
adapter = self._adapter()
runner = self._runner(adapter)
with patch.object(dl, "ledger_enabled", return_value=False), patch(
"gateway.delivery_ledger.ledger_enabled", return_value=False
):
n = await runner._redeliver_pending_obligations()
assert n == 0
adapter.send.assert_not_awaited()