hermes-agent/tests/agent/test_compaction_anti_thrash.py
ethernet 890bbbda1f Merge remote-tracking branch 'origin/main' into ethie/pm-clean
# 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
2026-09-23 07:02:44 -04:00

228 lines
10 KiB
Python

"""The anti-thrashing guard must actually fire.
Two defects made it dead code:
1. Effectiveness was scored as ``(current_tokens - estimate(compressed)) / current_tokens``.
``current_tokens`` is the provider's FULL prompt (system prompt + tool schemas +
messages); ``estimate(compressed)`` covers messages only. The mixed basis reported
~96% savings on every pass, so ``_ineffective_compression_count`` reset each time.
2. Even scored correctly, message shrinkage is the wrong yardstick. ``should_compress()``
trips on the full prompt, but compaction can only shrink messages. When the
incompressible floor alone meets the threshold, each pass can shrink messages by a
healthy margin, reset the counter, and still leave the prompt over the line -- so the
next turn compacts again, forever.
Together these made a mis-sized context window present as a hung CLI rather than a
warning. Effectiveness is now scored against the goal -- did the prompt get under the
threshold? -- judged in ``update_from_response()``, the one place that sees the
provider's real prompt count for the just-compacted conversation.
Two subtleties this pins:
* The verdict must NOT live in ``should_compress()``. That runs twice per turn with
two different measures (a rough preflight estimate and the real post-response count,
#36718); the rough one can dip below the threshold and reset the strike every turn,
re-opening the loop. ``test_rough_preflight_reading_does_not_reopen_the_loop``.
* It must not be computed analytically as ``floor = current_tokens - rough_estimate``.
That subtracts an estimate from a real count, absorbs the tokenizer skew into "floor",
and disables compaction on a healthy session.
``test_no_false_positive_under_tokenizer_skew``.
"""
import time
import pytest
from agent.context_compressor import ContextCompressor
def _compressor(threshold_tokens: int) -> ContextCompressor:
cc = ContextCompressor(
model="test-model",
threshold_percent=0.75,
protect_first_n=5,
protect_last_n=20,
quiet_mode=True,
config_context_length=40960,
provider="test",
)
cc.threshold_tokens = threshold_tokens # pin; don't couple to window math
cc._generate_summary = lambda *a, **k: "Summary of earlier turns."
return cc
def _messages(n: int, size: int = 1500) -> list:
msgs = [{"role": "system", "content": "sys"}]
for i in range(n):
role = "user" if i % 2 == 0 else "assistant"
msgs.append({"role": role, "content": f"m{i} " + "z" * size})
return msgs
def _turn(cc, msgs, real_prompt_tokens):
"""One agent turn as conversation_loop drives it.
should_compress() gates on the real prompt count; if it fires, compress()
runs and the provider then reports real usage for the shorter conversation.
The anti-thrashing verdict is judged in update_from_response() against that
real count -- not inside should_compress(), which is called twice per turn
with two different measures. Returns (msgs, did_compact).
"""
if not cc.should_compress(real_prompt_tokens):
return msgs, False
msgs = cc.compress(msgs, current_tokens=real_prompt_tokens)
cc._verify_compaction_cleared_threshold = True
cc.update_from_response({"prompt_tokens": real_prompt_tokens})
return msgs, True
class TestSavingsBasis:
def test_savings_does_not_depend_on_current_tokens(self):
"""Savings is a property of the messages, not of the caller's token count.
Before the fix, passing the full-prompt count made an identical compaction
report a wildly higher savings percentage.
"""
msgs = _messages(14)
a = _compressor(threshold_tokens=24_576)
a.compress([m.copy() for m in msgs], current_tokens=100_000)
with_full_prompt = a._last_compression_savings_pct
b = _compressor(threshold_tokens=24_576)
b.compress([m.copy() for m in msgs], current_tokens=None)
messages_only = b._last_compression_savings_pct
assert with_full_prompt == pytest.approx(messages_only, abs=0.01), (
"savings must be scored on a messages-vs-messages basis; got "
f"{with_full_prompt:.1f}% with a full-prompt count vs "
f"{messages_only:.1f}% without"
)
def test_no_op_compaction_is_not_reported_as_huge_savings(self):
cc = _compressor(threshold_tokens=24_576)
msgs = _messages(4, size=10) # below the minimum compressible size
cc.compress(msgs, current_tokens=100_000)
assert cc._last_compression_savings_pct < 50
class TestFutilityGuard:
def test_stops_when_floor_alone_meets_threshold(self):
"""Incompressible floor >= threshold -> shrinking messages cannot help."""
cc = _compressor(threshold_tokens=24_576)
msgs = _messages(14)
# Full prompt is far above threshold and dominated by system + tools,
# so the real count never drops no matter how the message list shrinks.
real_prompt_tokens = 33_564
fired = 0
for _ in range(6):
msgs, did = _turn(cc, msgs, real_prompt_tokens)
if did:
fired += 1
msgs.append({"role": "user", "content": "next " + "w" * 4000})
assert cc._ineffective_compression_count >= 2
assert not cc.should_compress(real_prompt_tokens), (
"compaction that cannot clear the threshold must stop"
)
assert fired <= 3, f"expected the loop to break early, compacted {fired}x"
def test_effective_compaction_still_resets_the_counter(self):
"""A compaction that gets the prompt under the threshold is not thrashing."""
cc = _compressor(threshold_tokens=750_000)
msgs = _messages(120, size=2000)
before = len(msgs)
out = cc.compress(msgs, current_tokens=None)
assert len(out) < before, "a long transcript must still compact"
assert cc._last_compression_savings_pct > 10
assert cc._last_compression_made_progress is True
assert cc._ineffective_compression_count == 0
def test_no_false_positive_under_tokenizer_skew(self):
"""The provider's real count exceeds the rough estimate; that is not a floor.
An analytic `floor = real_prompt - rough_estimate` misreads tokenizer skew
as incompressible overhead and disables compaction on a healthy session.
The check must compare the caller's own measure on both sides.
"""
from agent.context_compressor import estimate_messages_tokens_rough
skew, floor = 1.6, 30_000
cc = _compressor(threshold_tokens=150_000)
msgs = _messages(160, size=2500)
real_prompt = floor + int(skew * estimate_messages_tokens_rough(msgs))
assert cc.should_compress(real_prompt)
msgs = cc.compress(msgs, current_tokens=real_prompt)
real_after = floor + int(skew * estimate_messages_tokens_rough(msgs))
cc._verify_compaction_cleared_threshold = True
cc.update_from_response({"prompt_tokens": real_after}) # provider's real count
assert real_after < cc.threshold_tokens, "compaction really did clear it"
assert cc._ineffective_compression_count == 0, (
"tokenizer skew must not be mistaken for an incompressible floor"
)
def test_model_switch_resets_and_persists_fallback_streak(self, tmp_path):
from hermes_state import SessionDB
db = SessionDB(db_path=tmp_path / "state.db")
db.create_session("s1", source="cli")
cc = _compressor(threshold_tokens=24_576)
cc.bind_session_state(db, "s1")
cc.record_completed_compaction(used_fallback=True)
cc.update_model("next-model", 100_000)
assert cc._fallback_compression_streak == 0
assert db.get_compression_fallback_streak("s1") == 0
class TestMinimumMessagesBranch:
def test_too_few_messages_defers_via_structural_backoff(self):
"""A structurally impossible compaction must not strike the breaker.
#93022 — too-few-messages is a transcript-shape fact, not evidence
of an incompressible floor: striking it punished unrelated later
failures. Instead the branch arms the structural no-op backoff so
retries are deferred without burning anti-thrash strikes.
"""
cc = _compressor(threshold_tokens=1)
msgs = _messages(3, size=10)
before = cc._ineffective_compression_count
out = cc.compress(msgs, current_tokens=100_000)
assert len(out) == len(msgs), "nothing should have been compressed"
assert cc._last_compression_made_progress is False
assert cc._ineffective_compression_count == before, (
"structural no-op must leave the strike counter untouched"
)
assert cc._structural_no_op_backoff_until > time.monotonic(), (
"structural no-op must arm the retry backoff"
)
assert cc._compression_block_reason().startswith("structural_backoff")
class TestRejectedCompactionStrike:
"""#88568 — a would-grow refusal must count as an ineffective strike.
The anti-growth guard correctly keeps the original transcript, but the
rejection used to leave ``_ineffective_compression_count`` untouched, so
the breaker never latched and automatic compression retried the SAME
unchanged transcript on every turn.
"""
def test_two_rejections_stop_further_automatic_compression(self):
cc = _compressor(threshold_tokens=1)
cc.record_rejected_compaction()
cc.record_rejected_compaction()
# The latch consumers key on the counter itself (>= 2 blocks);
# pin the counter and the recovery-clock arming side effect.
assert cc._ineffective_compression_count >= 2
def test_rejection_does_not_arm_real_usage_verification(self):
"""Nothing was committed, so the next response must not be scored
against the pre-rejection transcript (that verdict belongs to
committed compactions only)."""
cc = _compressor(threshold_tokens=1)
cc.record_rejected_compaction()
assert cc._verify_compaction_cleared_threshold is False