# Conflicts: # apps/desktop/e2e/archived-hidden-session-recoverable.spec.ts # apps/desktop/e2e/bot-chat-message-agent-friendly-name.spec.ts # apps/desktop/e2e/bot-mailbox-unreadable-ticket.spec.ts # apps/desktop/e2e/bot-mode-roster-localized.spec.ts # apps/desktop/e2e/bot-mode-row-click-mirrors-registry.spec.ts # apps/desktop/e2e/bot-mode-tab-shows-bot-name.spec.ts # apps/desktop/e2e/bot-roster-group-row-organisation.spec.ts # apps/desktop/e2e/bot-roster-ignores-infra-dirs.spec.ts # apps/desktop/e2e/bot-roster-timestamp-meta.spec.ts # apps/desktop/e2e/bot-roster-user-sections.spec.ts # apps/desktop/e2e/bot-routines-pane-narrow.spec.ts # apps/desktop/e2e/bot-row-open-recent-session.spec.ts # apps/desktop/e2e/bot-tile-ignores-ambient-composer-model.spec.ts # apps/desktop/e2e/group-composer-auto-grow.spec.ts # apps/desktop/e2e/group-create-gate-remote-roster.spec.ts # apps/desktop/e2e/group-prompt-renamed-primary-handle.spec.ts # apps/desktop/e2e/hosted-room-backend-continuity.spec.ts # apps/desktop/e2e/hosted-room-legacy-store-migration.spec.ts # apps/desktop/e2e/settings-scope-chips-bot-title.spec.ts # apps/desktop/e2e/worktree-branch-status.spec.ts # apps/desktop/electron/backend-probes.test.ts # apps/desktop/electron/connection-apply.test.ts # apps/desktop/electron/desktop-electron-pin.test.ts # apps/desktop/electron/desktop-uninstall.test.ts # apps/desktop/electron/gateway-file-download-transport.test.ts # apps/desktop/electron/gateway-stop-before-update.test.ts # apps/desktop/electron/github-api-auth.test.ts # apps/desktop/electron/registry-primary-profile-scope.test.ts # apps/desktop/electron/update-api-check.test.ts # apps/desktop/electron/update-handoff-marker.test.ts # apps/desktop/electron/venv-blocker-scan.test.ts # apps/desktop/scripts/after-extract.test.mjs # apps/desktop/scripts/local-pack-publish.test.mjs # apps/desktop/scripts/tasks-scroll.test.mjs # apps/desktop/src/app/settings/model-settings.test.tsx # apps/desktop/src/app/updates-overlay.blockers.test.tsx # apps/desktop/src/components/desktop-install-overlay.test.tsx # apps/desktop/src/lib/update-copy.test.ts # scripts/ci/check_os_marker_fakes.py # tests-js/desktop-mac-usage-descriptions.test.ts # tests-js/node-engine-alignment.test.ts # tests/agent/lsp/test_install_and_lint_fixes.py # tests/agent/test_command_token_source.py # tests/agent/test_compression_boundary_hook.py # tests/agent/test_create_openai_client_ssl_verify.py # tests/agent/test_custom_provider_ca_probes.py # tests/agent/test_endpoint_blackhole.py # tests/agent/test_estimator_parity.py # tests/agent/test_in_place_compaction.py # tests/agent/test_moa_loop_mode.py # tests/agent/test_model_metadata.py # tests/agent/test_skill_session_platform_gate.py # tests/agent/test_skill_utils.py # tests/agent/test_ssl_ca_guard.py # tests/computer_use/test_doctor.py # tests/cron/test_codex_execution_paths.py # tests/cron/test_cron_bot_chat_delivery.py # tests/cron/test_cron_script.py # tests/cron/test_media_delivery_parity.py # tests/cron/test_misfire_catchup.py # tests/cron/test_parallel_pool.py # tests/cron/test_recurring_eagain_redispatch.py # tests/gateway/test_choice_picker.py # tests/gateway/test_control_socket_windows_live.py # tests/gateway/test_dingtalk.py # tests/gateway/test_feishu.py # tests/gateway/test_feishu_onboard.py # tests/gateway/test_gateway_shutdown.py # tests/gateway/test_matrix.py # tests/gateway/test_model_command_custom_providers.py # tests/gateway/test_reasoning_command.py # tests/gateway/test_runtime_footer.py # tests/gateway/test_session.py # tests/gateway/test_session_hygiene.py # tests/gateway/test_status.py # tests/gateway/test_teams.py # tests/gateway/test_turn_lease.py # tests/gateway/test_whatsapp_connect.py # tests/hermes_cli/test_approvals_command.py # tests/hermes_cli/test_auth_store_lock_concurrent.py # tests/hermes_cli/test_backup.py # tests/hermes_cli/test_banner_git_state.py # tests/hermes_cli/test_certifi_repair.py # tests/hermes_cli/test_cmd_update.py # tests/hermes_cli/test_compat_manifest_targets.py # tests/hermes_cli/test_computer_use_cli.py # tests/hermes_cli/test_cpr_local_leak.py # tests/hermes_cli/test_dashboard_auth_gate.py # tests/hermes_cli/test_dashboard_procs_kill_grace.py # tests/hermes_cli/test_desktop_lifecycle_windows_live.py # tests/hermes_cli/test_doctor.py # tests/hermes_cli/test_doctor_command_install.py # tests/hermes_cli/test_fleet_config_migration_windows_live.py # tests/hermes_cli/test_gateway.py # tests/hermes_cli/test_gateway_platform_gating.py # tests/hermes_cli/test_gateway_restart_loop.py # tests/hermes_cli/test_gateway_task_probe.py # tests/hermes_cli/test_gateway_wsl.py # tests/hermes_cli/test_gui_command.py # tests/hermes_cli/test_install_cua_driver.py # tests/hermes_cli/test_kanban_db.py # tests/hermes_cli/test_lazy_command_exports.py # tests/hermes_cli/test_lazy_refresh_venv_repair.py # tests/hermes_cli/test_linux_desktop_entry.py # tests/hermes_cli/test_local_runtime.py # tests/hermes_cli/test_local_runtime_updates.py # tests/hermes_cli/test_managed_uv.py # tests/hermes_cli/test_mcp_reload_confirm_gate.py # tests/hermes_cli/test_nous_subscription.py # tests/hermes_cli/test_npm_engine.py # tests/hermes_cli/test_personality_none.py # tests/hermes_cli/test_pet_toggle.py # tests/hermes_cli/test_plan_reconciliation_windows_live.py # tests/hermes_cli/test_plugin_event_bus.py # tests/hermes_cli/test_plugin_manifest_v2.py # tests/hermes_cli/test_plugin_packs.py # tests/hermes_cli/test_plugins_cmd.py # tests/hermes_cli/test_plugins_cmd_enable_disable_nested.py # tests/hermes_cli/test_process_identity.py # tests/hermes_cli/test_profiles.py # tests/hermes_cli/test_profiles_sidebar_cache.py # tests/hermes_cli/test_pty_bridge.py # tests/hermes_cli/test_resolve_turn_limit.py # tests/hermes_cli/test_serve_runtime_inventory.py # tests/hermes_cli/test_session_vacuum_config.py # tests/hermes_cli/test_set_config_value.py # tests/hermes_cli/test_signal_handler_kanban_worker.py # tests/hermes_cli/test_slash_confirm_windows.py # tests/hermes_cli/test_stale_pid_guard.py # tests/hermes_cli/test_startup_fast_guards.py # tests/hermes_cli/test_status.py # tests/hermes_cli/test_telegram_managed_bot.py # tests/hermes_cli/test_tools_config.py # tests/hermes_cli/test_update_apply_shallow_count.py # tests/hermes_cli/test_update_autostash.py # tests/hermes_cli/test_update_concurrent_quarantine.py # tests/hermes_cli/test_update_fetch_failure_classifier.py # tests/hermes_cli/test_update_fleet_probe_resume_token.py # tests/hermes_cli/test_update_handoff_backend_reap.py # tests/hermes_cli/test_update_handoff_desktop_rebuild.py # tests/hermes_cli/test_update_head_moved_gate.py # tests/hermes_cli/test_update_host_obligation.py # tests/hermes_cli/test_update_import_guard.py # tests/hermes_cli/test_update_interrupted_recovery.py # tests/hermes_cli/test_update_inventory.py # tests/hermes_cli/test_update_launchd_unloaded_gateway.py # tests/hermes_cli/test_update_missing_configured_deps.py # tests/hermes_cli/test_update_modified_notice.py # tests/hermes_cli/test_update_multiplex_migration_hook.py # tests/hermes_cli/test_update_no_gateway_restart.py # tests/hermes_cli/test_update_orphan_backend_reap.py # tests/hermes_cli/test_update_parked_branch_guard.py # tests/hermes_cli/test_update_post_pull_syntax_guard.py # tests/hermes_cli/test_update_receipt.py # tests/hermes_cli/test_update_self_lock.py # tests/hermes_cli/test_update_shim_fail_closed.py # tests/hermes_cli/test_update_shim_self_lock.py # tests/hermes_cli/test_update_sqlite_remediation.py # tests/hermes_cli/test_update_stale_dashboard.py # tests/hermes_cli/test_update_stale_virtualenv.py # tests/hermes_cli/test_update_venv_health.py # tests/hermes_cli/test_update_venv_ownership_preflight.py # tests/hermes_cli/test_update_wedged_gateway.py # tests/hermes_cli/test_update_yes_flag.py # tests/hermes_cli/test_update_zip_two_phase.py # tests/hermes_cli/test_urllib_security.py # tests/hermes_cli/test_ux_messages_auth_config.py # tests/hermes_cli/test_ux_messages_startup.py # tests/hermes_cli/test_venv_holder_classifier.py # tests/hermes_cli/test_verify_console_scripts.py # tests/hermes_cli/test_verify_core_dependencies.py # tests/hermes_cli/test_web_server.py # tests/hermes_cli/test_web_server_console_ws.py # tests/hermes_cli/test_web_server_ws_ping.py # tests/hermes_cli/test_web_ui_build.py # tests/hermes_state/test_fts_rebuild_admission.py # tests/hermes_state/test_hermes_state.py # tests/plugins/memory/test_memory_lazy_install.py # tests/plugins/test_google_meet_plugin.py # tests/plugins/test_langfuse_plugin.py # tests/plugins/test_security_guidance_plugin.py # tests/plugins/test_transform_llm_output_hook.py # tests/scripts/desktop_update/test_desktop_update_windows_gateway_flag.py # tests/scripts/desktop_update/test_desktop_update_windows_python_handoff.py # tests/scripts/desktop_update/test_desktop_update_windows_timestamp.py # tests/scripts/install/test_install_clone_throttle_fallback.py # tests/scripts/install/test_install_lockfile_churn.py # tests/scripts/install/test_install_no_initial_commit.py # tests/scripts/install/test_install_sh_browser_install.py # tests/scripts/install/test_install_sh_node_prerelease.py # tests/scripts/install/test_install_sh_symlink_stomp.py # tests/scripts/install/test_install_sh_uv_lock_config.py # tests/scripts/install/test_install_unmerged_index.py # tests/scripts/test_contributor_map.py # tests/scripts/test_run_tests_parallel.py # tests/skills/test_competitor_news_monitor_skill.py # tests/skills/test_document_to_action_items_skill.py # tests/skills/test_google_workspace_setup.py # tests/skills/test_google_workspace_setup_deps.py # tests/skills/test_grounded_citations_skill.py # tests/skills/test_ip_as_logo_skill.py # tests/skills/test_live_dashboard_skill.py # tests/skills/test_mcp_oauth_remote_gateway_skill.py # tests/skills/test_office_document_skills.py # tests/skills/test_openclaw_migration.py # tests/skills/test_product_price_monitor_skill.py # tests/skills/test_scrollcraft_skill.py # tests/skills/test_setup_wizard_generator_skill.py # tests/skills/test_weekly_review_planning_skill.py # tests/test_engines_satisfiable.py # tests/test_fast_safe_load.py # tests/test_hermes_bootstrap.py # tests/test_hermes_constants.py # tests/test_hermes_logging.py # tests/test_managed_runtime_resolution.py # tests/test_model_tools_async_bridge.py # tests/test_packaging_build_guard.py # tests/test_packaging_metadata.py # tests/test_yaml_indent_consistency.py # tests/tools/test_approval_timeout_overflow.py # tests/tools/test_base_environment.py # tests/tools/test_bot_mode_dm.py # tests/tools/test_browser_chromium_check.py # tests/tools/test_browser_hardening.py # tests/tools/test_browser_homebrew_paths.py # tests/tools/test_browser_npx_warmup.py # tests/tools/test_browser_orphan_reaper.py # tests/tools/test_browser_real_profile.py # tests/tools/test_browser_use_cli.py # tests/tools/test_clipboard.py # tests/tools/test_code_execution.py # tests/tools/test_code_execution_modes.py # tests/tools/test_code_execution_windows_env.py # tests/tools/test_computer_use.py # tests/tools/test_delegate_liveness_timeout.py # tests/tools/test_execute_code_approval_cluster.py # tests/tools/test_execution_flag_detection.py # tests/tools/test_fal_common.py # tests/tools/test_file_operations.py # tests/tools/test_file_tools.py # tests/tools/test_file_tools_cwd_resolution.py # tests/tools/test_file_tools_live.py # tests/tools/test_lazy_deps.py # tests/tools/test_lazy_deps_durable_target.py # tests/tools/test_lazy_deps_managed.py # tests/tools/test_local_env_blocklist.py # tests/tools/test_local_tempdir.py # tests/tools/test_macos_protected_search.py # tests/tools/test_mcp_npx_cached_bin.py # tests/tools/test_oneshot_completion_linger.py # tests/tools/test_process_registry.py # tests/tools/test_read_file_schema_gating.py # tests/tools/test_skill_improvements.py # tests/tools/test_skills_sync.py # tests/tools/test_termux_api_detection.py # tests/tools/test_tirith_security.py # tests/tools/test_transcription_tools.py # tests/tools/test_tts_streaming.py # tests/tools/test_wake_word.py # tests/tui_gateway/test_compute_host_borrowed_lease.py # tests/tui_gateway/test_compute_host_turn_protocol.py # tests/tui_gateway/test_isolated_orphan_activity.py # tests/tui_gateway/test_protocol.py # tests/tui_gateway/test_slash_worker_profile_home.py # tests/tui_gateway/test_subprocess_encoding.py # tests/tui_gateway/test_tui_gateway_server.py # ui-tui/src/__tests__/terminalParity.test.ts # ui-tui/src/__tests__/termuxComposerLayout.test.ts # ui-tui/src/__tests__/textInputFastEcho.test.ts
247 lines
13 KiB
Python
247 lines
13 KiB
Python
"""Bedrock cache provenance and compressor budgets across disk-backed restarts.
|
|
|
|
AWS documents Grok 4.6's Bedrock context window as 500K, independently of
|
|
xAI's direct API catalog: https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-xai-grok-4-6.html
|
|
Only provider I/O is stubbed; cache/config readers and compressor are real.
|
|
"""
|
|
|
|
import json
|
|
import os
|
|
import subprocess
|
|
import sys
|
|
import time
|
|
from types import SimpleNamespace
|
|
from unittest.mock import Mock
|
|
|
|
import pytest
|
|
import hermes_yaml as yaml
|
|
|
|
from agent import bedrock_adapter as ba
|
|
from agent import model_metadata as mm
|
|
from agent.context_compressor import ContextCompressor
|
|
|
|
|
|
@pytest.fixture(autouse=True)
|
|
def isolated_home(tmp_path, monkeypatch):
|
|
monkeypatch.setenv("HERMES_HOME", str(tmp_path))
|
|
mm._BEDROCK_PROBE_FAILURE_CACHE.clear()
|
|
monkeypatch.setattr(ba, "resolve_bedrock_region", lambda: "us-east-1")
|
|
yield tmp_path
|
|
mm._BEDROCK_PROBE_FAILURE_CACHE.clear()
|
|
|
|
|
|
@pytest.mark.parametrize("model", ["global.xai.grok-4.6"]) # prefix variants resolve identically
|
|
@pytest.mark.parametrize("base_url", ["", "https://bedrock-runtime.us-east-1.amazonaws.com"])
|
|
@pytest.mark.parametrize("legacy", [None, 128_000, 700_000])
|
|
@pytest.mark.parametrize("probed", [None, 128_000, 800_000])
|
|
def test_bedrock_resolution_migrates_ambiguous_cache_and_preserves_probe(
|
|
isolated_home, monkeypatch, model, base_url, legacy, probed,
|
|
):
|
|
"""Neither an old small nor large scalar proves where it came from.
|
|
|
|
A successful probe is authoritative, including below-table limits. Its
|
|
provenance must survive unrelated writes and an actual process restart.
|
|
"""
|
|
cache_url = base_url or "bedrock://"
|
|
key = mm._context_cache_key(model, cache_url)
|
|
cache_file = isolated_home / "context_length_cache.yaml"
|
|
if legacy is not None:
|
|
cache_file.write_text(yaml.safe_dump({"context_lengths": {key: legacy}}))
|
|
probe = Mock(return_value=probed)
|
|
monkeypatch.setattr(ba, "probe_bedrock_context_length", probe)
|
|
expected = probed if probed is not None else 500_000
|
|
compressor = ContextCompressor(model, provider="bedrock", base_url=base_url, quiet_mode=True)
|
|
assert compressor.context_length == expected
|
|
# Preserve the existing raise-only 75% floor for windows below 512K.
|
|
expected_threshold = int(expected * (0.75 if expected < 512_000 else 0.5))
|
|
assert compressor.threshold_tokens == expected_threshold
|
|
assert mm.get_model_context_length(model, provider="bedrock", base_url=base_url) == expected
|
|
probe.assert_called_once_with(model, "us-east-1")
|
|
assert mm.get_cached_context_length(model, cache_url, bedrock_confirmed=True) == probed
|
|
# Ordinary cache updates must not erase another entry's provenance.
|
|
mm.save_context_length("other-model", "https://other.example/v1", 64_000)
|
|
mm._invalidate_cached_context_length("other-model", "https://other.example/v1")
|
|
if probed is None:
|
|
assert yaml.safe_load(cache_file.read_text())["context_lengths"].get(key) == legacy
|
|
else:
|
|
script = '''
|
|
import json, sys
|
|
from agent import bedrock_adapter as ba
|
|
from agent.context_compressor import ContextCompressor
|
|
ba.probe_bedrock_context_length = lambda *a, **k: (_ for _ in ()).throw(AssertionError("reprobed persisted success"))
|
|
c = ContextCompressor(sys.argv[1], provider="bedrock", base_url=sys.argv[2], quiet_mode=True)
|
|
print(json.dumps([c.context_length, c.threshold_tokens]))
|
|
'''
|
|
result = subprocess.run([sys.executable, "-c", script, model, base_url],
|
|
env=dict(os.environ), text=True, capture_output=True, check=True)
|
|
assert json.loads(result.stdout) == [expected, expected_threshold]
|
|
|
|
|
|
@pytest.mark.parametrize("base_url", ["", "https://bedrock-runtime.us-east-1.amazonaws.com"])
|
|
@pytest.mark.parametrize("retry", ["ttl", "invalidate", "restart", "profile", "endpoint"])
|
|
def test_failure_memo_retry_scope_and_expiry(isolated_home, monkeypatch, base_url, retry):
|
|
model = "global.xai.grok-4.6"
|
|
cache_url = base_url or "bedrock://"
|
|
probe = Mock(side_effect=[None, 96_000])
|
|
monkeypatch.setattr(ba, "probe_bedrock_context_length", probe)
|
|
assert mm.get_model_context_length(model, provider="bedrock", base_url=base_url) == 500_000
|
|
assert mm.get_model_context_length(model, provider="bedrock", base_url=base_url) == 500_000
|
|
assert probe.call_count == 1
|
|
assert not (isolated_home / "context_length_cache.yaml").exists()
|
|
if retry == "ttl":
|
|
expired = time.monotonic() - mm._BEDROCK_PROBE_FAILURE_TTL_SECONDS - 1
|
|
for key in mm._BEDROCK_PROBE_FAILURE_CACHE:
|
|
mm._BEDROCK_PROBE_FAILURE_CACHE[key] = expired
|
|
elif retry == "invalidate":
|
|
mm._invalidate_cached_context_length(model, cache_url)
|
|
elif retry == "restart":
|
|
mm._BEDROCK_PROBE_FAILURE_CACHE.clear()
|
|
elif retry == "profile":
|
|
new_home = isolated_home / "other-profile"
|
|
new_home.mkdir()
|
|
monkeypatch.setenv("HERMES_HOME", str(new_home))
|
|
else:
|
|
base_url = "https://bedrock-runtime.us-east-1.amazonaws.com/other"
|
|
assert mm.get_model_context_length(model, provider="bedrock", base_url=base_url) == 96_000
|
|
assert probe.call_count == 2
|
|
if retry == "ttl":
|
|
assert not mm._BEDROCK_PROBE_FAILURE_CACHE
|
|
|
|
|
|
@pytest.mark.parametrize("base_url", ["", "https://bedrock-runtime.us-east-1.amazonaws.com"])
|
|
@pytest.mark.parametrize("override", ["argument", "config"])
|
|
def test_explicit_context_override_and_compressor_caps_win(isolated_home, monkeypatch, base_url, override):
|
|
model = "us.xai.grok-4.6"
|
|
probe = Mock(side_effect=AssertionError("explicit override must not probe"))
|
|
monkeypatch.setattr(ba, "probe_bedrock_context_length", probe)
|
|
explicit_context = 80_000 if override == "argument" else None
|
|
if override == "config":
|
|
(isolated_home / "config.yaml").write_text(yaml.safe_dump({
|
|
"model_overrides": {"bedrock": {model: {"context_window": 80_000}}},
|
|
}))
|
|
compressor = ContextCompressor(model, provider="bedrock", base_url=base_url,
|
|
threshold_tokens_cap=30_000, max_tokens=10_000,
|
|
quiet_mode=True, config_context_length=explicit_context)
|
|
assert compressor.context_length == 80_000
|
|
assert compressor.threshold_tokens == 30_000
|
|
assert not (isolated_home / "context_length_cache.yaml").exists()
|
|
probe.assert_not_called()
|
|
|
|
|
|
def test_unknown_model_fallback_and_host_inference(monkeypatch):
|
|
probe = Mock(return_value=None)
|
|
monkeypatch.setattr(ba, "probe_bedrock_context_length", probe)
|
|
base_url = "https://bedrock-runtime.us-east-1.amazonaws.com"
|
|
assert mm.get_model_context_length("unknown.future-model", base_url=base_url) == ba.BEDROCK_DEFAULT_CONTEXT_LENGTH
|
|
assert mm.get_model_context_length("xai.grok-4.6", base_url=base_url) == 500_000
|
|
assert mm.get_cached_context_length("unknown.future-model", base_url) is None
|
|
|
|
|
|
@pytest.mark.parametrize("base_url", ["", "https://bedrock-runtime.us-east-1.amazonaws.com"])
|
|
@pytest.mark.parametrize("writer", ["overflow", "usage"])
|
|
def test_provider_confirmed_writers_survive_restart(monkeypatch, base_url, writer):
|
|
from agent.turn_overflow import _adopt_provider_context_limit
|
|
from agent.turn_usage import record_response_usage
|
|
|
|
model = "global.xai.grok-4.6"
|
|
compressor = ContextCompressor(model, base_url=base_url, provider="bedrock", quiet_mode=True)
|
|
compressor.context_length = 500_000
|
|
agent = SimpleNamespace(model=model, provider="bedrock", api_mode="bedrock", base_url=base_url,
|
|
context_compressor=compressor, _buffer_vprint=lambda *a: None,
|
|
_safe_print=lambda *a: None, log_prefix="", client=None,
|
|
_session_db=None, verbose_logging=False, quiet_mode=True,
|
|
session_api_calls=0, session_estimated_cost_usd=0)
|
|
if writer == "overflow":
|
|
assert _adopt_provider_context_limit(SimpleNamespace(agent=agent),
|
|
"maximum context length is 96000 tokens", 500_000) == 96_000
|
|
else:
|
|
compressor.context_length = 96_000
|
|
compressor._context_probed = compressor._context_probe_persistable = True
|
|
for name in ("prompt", "completion", "total", "input", "output", "cache_read", "cache_write", "reasoning"):
|
|
setattr(agent, f"session_{name}_tokens", 0)
|
|
response = SimpleNamespace(usage={"input_tokens": 100, "output_tokens": 5})
|
|
record_response_usage(agent, response, messages=[{"role": "user", "content": "hi"}],
|
|
api_call_count=1, api_duration=0.1, compression_attempts=0, max_compression_attempts=3)
|
|
assert mm.get_cached_context_length(model, base_url or "bedrock://") == 96_000
|
|
script = '''
|
|
import sys
|
|
from agent import model_metadata as mm, bedrock_adapter as ba
|
|
ba.probe_bedrock_context_length = lambda *a, **k: (_ for _ in ()).throw(AssertionError("lost provider limit"))
|
|
assert mm.get_model_context_length(sys.argv[1], base_url=sys.argv[2], provider="bedrock") == 96000
|
|
'''
|
|
subprocess.run([sys.executable, "-c", script, model, base_url], check=True, capture_output=True, text=True)
|
|
|
|
|
|
@pytest.mark.parametrize("base_url", ["", "https://bedrock-runtime.us-east-1.amazonaws.com"])
|
|
def test_readonly_legacy_cache_does_not_reset_probe_cooldown(isolated_home, monkeypatch, base_url):
|
|
model = "global.xai.grok-4.6"
|
|
key = mm._context_cache_key(model, base_url or "bedrock://")
|
|
cache_file = isolated_home / "context_length_cache.yaml"
|
|
cache_file.write_text(yaml.safe_dump({"context_lengths": {key: 128_000}}))
|
|
probe = Mock(return_value=None)
|
|
monkeypatch.setattr(ba, "probe_bedrock_context_length", probe)
|
|
monkeypatch.setattr(mm, "_write_context_cache", Mock(side_effect=OSError("read-only")))
|
|
for _ in range(3):
|
|
assert mm.get_model_context_length(model, provider="bedrock", base_url=base_url) == 500_000
|
|
assert probe.call_count == 1
|
|
|
|
|
|
@pytest.mark.parametrize("rewrite", [None, 96_000, 120_000])
|
|
def test_provenance_is_backward_readable_and_generic_writes_clear_it(isolated_home, monkeypatch, rewrite):
|
|
model, base_url = "xai.grok-4.6", "https://bedrock-runtime.us-east-1.amazonaws.com"
|
|
probe = Mock(side_effect=[96_000, 110_000])
|
|
monkeypatch.setattr(ba, "probe_bedrock_context_length", probe)
|
|
assert mm.get_model_context_length(model, base_url=base_url) == 96_000
|
|
raw = yaml.safe_load((isolated_home / "context_length_cache.yaml").read_text())
|
|
# Old readers take this value directly into arithmetic. Metadata is additive.
|
|
assert raw["context_lengths"][mm._context_cache_key(model, base_url)] + 1 == 96_001
|
|
if rewrite is not None:
|
|
mm.save_context_length(model, base_url, rewrite)
|
|
assert mm.get_model_context_length(model, base_url=base_url) == 110_000
|
|
else:
|
|
assert mm.get_model_context_length(model, base_url=base_url) == 96_000
|
|
|
|
|
|
@pytest.mark.parametrize("lengths", [True, [128_000], "bad"])
|
|
def test_malformed_lengths_do_not_block_provider_persistence(isolated_home, monkeypatch, lengths):
|
|
(isolated_home / "context_length_cache.yaml").write_text(yaml.safe_dump({"context_lengths": lengths}))
|
|
monkeypatch.setattr(ba, "probe_bedrock_context_length", lambda *a: 96_000)
|
|
assert mm.get_model_context_length("xai.grok-4.6", provider="bedrock") == 96_000
|
|
assert mm.get_cached_context_length("xai.grok-4.6", "bedrock://") == 96_000
|
|
|
|
|
|
@pytest.mark.parametrize("marker", [True, "96000", 128_000, [96_000], {"source": "probe"}])
|
|
def test_malformed_or_mismatched_provenance_requires_revalidation(isolated_home, monkeypatch, marker):
|
|
model, base_url = "xai.grok-4.6", "bedrock://"
|
|
key = mm._context_cache_key(model, base_url)
|
|
(isolated_home / "context_length_cache.yaml").write_text(yaml.safe_dump({
|
|
"context_lengths": {key: 96_000}, "bedrock_confirmed_v1": {key: marker},
|
|
}))
|
|
probe = Mock(return_value=100_000)
|
|
monkeypatch.setattr(ba, "probe_bedrock_context_length", probe)
|
|
assert mm.get_model_context_length(model, provider="bedrock") == 100_000
|
|
probe.assert_called_once()
|
|
|
|
|
|
def test_context_local_profile_memos_do_not_cross_and_expired_rows_are_pruned(isolated_home, monkeypatch):
|
|
from hermes_constants import set_hermes_home_override, reset_hermes_home_override
|
|
|
|
probe = Mock(side_effect=[None, None, 96_000])
|
|
monkeypatch.setattr(ba, "probe_bedrock_context_length", probe)
|
|
model = "global.xai.grok-4.6"
|
|
assert mm.get_model_context_length(model, provider="bedrock") == 500_000
|
|
# Leave a different model's expired row: lookup must prune it too.
|
|
assert mm.get_model_context_length("unknown.future", provider="bedrock") == ba.BEDROCK_DEFAULT_CONTEXT_LENGTH
|
|
for key in mm._BEDROCK_PROBE_FAILURE_CACHE:
|
|
if "unknown.future" in key:
|
|
mm._BEDROCK_PROBE_FAILURE_CACHE[key] = time.monotonic() - mm._BEDROCK_PROBE_FAILURE_TTL_SECONDS - 1
|
|
token = set_hermes_home_override(isolated_home / "routed-profile")
|
|
try:
|
|
assert mm.get_model_context_length(model, provider="bedrock") == 96_000
|
|
finally:
|
|
reset_hermes_home_override(token)
|
|
assert probe.call_count == 3
|
|
assert all("unknown.future" not in key for key in mm._BEDROCK_PROBE_FAILURE_CACHE)
|
|
assert mm.get_model_context_length(model, provider="bedrock") == 500_000
|
|
assert probe.call_count == 3
|