[eric] agents: a process that will inject failures says so, and names a typo that armed nothing

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018foyDoK19jjbYdudfzQVkZ

backend/apps/agents/agents.py

@@ -8,6 +8,7 @@ from fastapi import HTTPException, Request
 from typeguard import typechecked
 
 from backend.apps.agents.agent_manager import agent_manager
+from backend.apps.agents.core.fault_injection import announce as p_announce_armed_faults
 from backend.apps.agents.core.models import AgentConfig, AgentSession, ApprovalResponse
 from backend.apps.agents.core.seq_log import seq_log
 from backend.apps.agents.manager.session.history_compaction import estimate_post_compact_input
@@ -24,6 +25,7 @@ p_group_meta_inflight: dict[tuple[str, str], asyncio.Future] = {}
 @asynccontextmanager
 async def agents_lifespan():
     logger.info("Agents sub-app starting")
+    p_announce_armed_faults()
     await agent_manager.reconcile_on_startup()
     await agent_manager.restore_all_sessions()
     # Off the critical path: crash-cut turns resume themselves once everything is hydrated.

backend/apps/agents/core/fault_injection.py

@@ -12,9 +12,12 @@ the drill reports a pass.
     OSW_FAULT=policy_block,auth_401 bash run.sh
 """
 
+import logging
 import os
 from typing import Set
 
+logger = logging.getLogger(__name__)
+
 # Every fault the harness knows. A name outside this set is a typo, not a feature.
 KNOWN_FAULTS: Set[str] = {
     "policy_block",      # the provider declines the request (ENG-383 failover, ENG-387 doors)
@@ -60,3 +63,19 @@ def unknown_faults() -> Set[str]:
     if not raw:
         return set()
     return {p.strip() for p in raw.split(",") if p.strip()} - KNOWN_FAULTS
+
+
+def announce() -> None:
+    """Say out loud, once at boot, that this process will inject failures, and name any word that
+    armed nothing. Silence here is the exact row-6 shape this module exists to kill: a typo arms
+    NOTHING, the drill then exercises the untouched happy path, and the pass is meaningless."""
+    raw = os.environ.get("OSW_FAULT", "")
+    if not raw.strip():
+        return
+    live = sorted({p.strip() for p in raw.split(",") if p.strip()} & KNOWN_FAULTS)
+    bogus = sorted(unknown_faults())
+    logger.warning(
+        "OSW_FAULT is set: this process will DELIBERATELY INJECT failures. "
+        f"armed={live or 'NOTHING'}"
+        + (f"; not a known fault, armed nothing: {bogus}" if bogus else "")
+    )

backend/tests/test_fault_injection.py

@@ -12,7 +12,7 @@ import pytest
 import builtins
 
 from backend.apps.agents.core.fault_injection import (
-    KNOWN_FAULTS, armed, armed_once, reset_fired, unknown_faults,
+    KNOWN_FAULTS, announce, armed, armed_once, reset_fired, unknown_faults,
 )
 from backend.apps.agents.core.error_classify import (
     has_auth_status, is_connection_lost, is_content_policy_block,
@@ -107,3 +107,23 @@ def test_every_known_fault_is_wired_somewhere():
         "a fault this build knows but wires nowhere is a guard that can never be drilled"
     for kind in WIRED_IN:
         assert p_block(kind).strip(), f"{kind} has an empty branch"
+
+
+def test_arming_is_announced_and_a_typo_is_named(monkeypatch, caplog):
+    # The harness built to kill row-6 silence had it: unknown_faults() existed and NOTHING called it,
+    # so a mistyped name armed nothing while the drill exercised the untouched happy path.
+    monkeypatch.setenv("OSW_FAULT", "policy_block,plicy_blok")
+    with caplog.at_level("WARNING"):
+        announce()
+    assert "policy_block" in caplog.text and "plicy_blok" in caplog.text
+    caplog.clear()
+    monkeypatch.delenv("OSW_FAULT")
+    with caplog.at_level("WARNING"):
+        announce()
+    assert caplog.text == "", "a shipped build must say nothing at all"
+
+
+def test_the_boot_path_actually_announces():
+    src = open("backend/apps/agents/agents.py").read()
+    assert "p_announce_armed_faults()" in src, \
+        "an announcement nothing calls is the silence it was written to prevent"