hermes-bsd/tests/gateway/test_compress_command.py

308 lines
12 KiB
Python
Raw Normal View History

"""Tests for gateway /compress user-facing messaging."""
from datetime import datetime
from unittest.mock import MagicMock, patch
import pytest
from gateway.config import GatewayConfig, Platform, PlatformConfig
from gateway.platforms.base import MessageEvent
from gateway.session import SessionEntry, SessionSource, build_session_key
def _make_source() -> SessionSource:
return SessionSource(
platform=Platform.TELEGRAM,
user_id="u1",
chat_id="c1",
user_name="tester",
chat_type="dm",
)
def _make_event(text: str = "/compress") -> MessageEvent:
return MessageEvent(text=text, source=_make_source(), message_id="m1")
def _make_history() -> list[dict[str, str]]:
return [
{"role": "user", "content": "one"},
{"role": "assistant", "content": "two"},
{"role": "user", "content": "three"},
{"role": "assistant", "content": "four"},
]
def _make_runner(history: list[dict[str, str]]):
from gateway.run import GatewayRunner
runner = object.__new__(GatewayRunner)
runner.config = GatewayConfig(
platforms={Platform.TELEGRAM: PlatformConfig(enabled=True, token="***")}
)
session_entry = SessionEntry(
session_key=build_session_key(_make_source()),
session_id="sess-1",
created_at=datetime.now(),
updated_at=datetime.now(),
platform=Platform.TELEGRAM,
chat_type="dm",
)
runner.session_store = MagicMock()
runner.session_store.get_or_create_session.return_value = session_entry
runner.session_store.load_transcript.return_value = history
runner.session_store.rewrite_transcript = MagicMock()
runner.session_store.update_session = MagicMock()
runner.session_store._save = MagicMock()
runner._session_db = None
return runner
@pytest.mark.asyncio
async def test_compress_command_reports_noop_without_success_banner():
history = _make_history()
runner = _make_runner(history)
agent_instance = MagicMock()
agent_instance.shutdown_memory_provider = MagicMock()
agent_instance.close = MagicMock()
fix(compression): include system prompt + tool schemas in token estimates (#18265) The user-visible /compress banner and the post-compression last_prompt_tokens writeback both counted only the raw message transcript (chars/4). With a 15KB system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of request pressure — a 234x gap. Two user-facing consequences: - Banner shows 'Compressing … (~45 tokens)…' while compression is actually firing on 10K+ tokens of real pressure, confusing users about why compression triggered (reported by @codecovenant on X; #6217). - Post-compression last_prompt_tokens writeback omits tool schemas, so the next should_compress() check compares real usage against a stale underestimate — compression triggers late, potentially past the model's context limit on small-context models (#14695). Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough() at every user-visible banner and at the post-compression writeback. estimate_request_tokens_rough() already existed for exactly this purpose and includes system prompt + tool schemas. Touched call sites: - run_agent.py: post-compression last_prompt_tokens writeback, post-tool call should_compress() fallback when provider usage is missing - cli.py: /compress banner + summary - gateway/run.py: gateway /compress banner + summary - tui_gateway/server.py: TUI /compress status + summary - acp_adapter/server.py: ACP /compact before/after Left intentionally alone: - Session-hygiene fallback and the 'no agent' /status path in gateway/run.py — no agent instance is in scope to query for system prompt/tools, and the existing 30-50% overestimate wobble on hygiene is safety-accepted. - Verbose-mode 'Request size' logging — informational only, already counts system prompt via api_messages[0]. Also relabels the feedback line from 'Rough transcript estimate' to 'Approx request size' so the metric label matches what it actually measures. Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217); user report @codecovenant on X (2026-04-30). Closes #14695 Closes #6217
2026-04-30 23:03:54 -07:00
agent_instance._cached_system_prompt = ""
agent_instance.tools = None
agent_instance.context_compressor.has_content_to_compress.return_value = True
agent_instance.session_id = "sess-1"
agent_instance._compress_context.return_value = (list(history), "")
fix(compression): include system prompt + tool schemas in token estimates (#18265) The user-visible /compress banner and the post-compression last_prompt_tokens writeback both counted only the raw message transcript (chars/4). With a 15KB system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of request pressure — a 234x gap. Two user-facing consequences: - Banner shows 'Compressing … (~45 tokens)…' while compression is actually firing on 10K+ tokens of real pressure, confusing users about why compression triggered (reported by @codecovenant on X; #6217). - Post-compression last_prompt_tokens writeback omits tool schemas, so the next should_compress() check compares real usage against a stale underestimate — compression triggers late, potentially past the model's context limit on small-context models (#14695). Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough() at every user-visible banner and at the post-compression writeback. estimate_request_tokens_rough() already existed for exactly this purpose and includes system prompt + tool schemas. Touched call sites: - run_agent.py: post-compression last_prompt_tokens writeback, post-tool call should_compress() fallback when provider usage is missing - cli.py: /compress banner + summary - gateway/run.py: gateway /compress banner + summary - tui_gateway/server.py: TUI /compress status + summary - acp_adapter/server.py: ACP /compact before/after Left intentionally alone: - Session-hygiene fallback and the 'no agent' /status path in gateway/run.py — no agent instance is in scope to query for system prompt/tools, and the existing 30-50% overestimate wobble on hygiene is safety-accepted. - Verbose-mode 'Request size' logging — informational only, already counts system prompt via api_messages[0]. Also relabels the feedback line from 'Rough transcript estimate' to 'Approx request size' so the metric label matches what it actually measures. Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217); user report @codecovenant on X (2026-04-30). Closes #14695 Closes #6217
2026-04-30 23:03:54 -07:00
def _estimate(messages, **_kwargs):
assert messages == history
return 100
with (
patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "test-key"}),
patch("gateway.run._resolve_gateway_model", return_value="test-model"),
patch("run_agent.AIAgent", return_value=agent_instance),
fix(compression): include system prompt + tool schemas in token estimates (#18265) The user-visible /compress banner and the post-compression last_prompt_tokens writeback both counted only the raw message transcript (chars/4). With a 15KB system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of request pressure — a 234x gap. Two user-facing consequences: - Banner shows 'Compressing … (~45 tokens)…' while compression is actually firing on 10K+ tokens of real pressure, confusing users about why compression triggered (reported by @codecovenant on X; #6217). - Post-compression last_prompt_tokens writeback omits tool schemas, so the next should_compress() check compares real usage against a stale underestimate — compression triggers late, potentially past the model's context limit on small-context models (#14695). Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough() at every user-visible banner and at the post-compression writeback. estimate_request_tokens_rough() already existed for exactly this purpose and includes system prompt + tool schemas. Touched call sites: - run_agent.py: post-compression last_prompt_tokens writeback, post-tool call should_compress() fallback when provider usage is missing - cli.py: /compress banner + summary - gateway/run.py: gateway /compress banner + summary - tui_gateway/server.py: TUI /compress status + summary - acp_adapter/server.py: ACP /compact before/after Left intentionally alone: - Session-hygiene fallback and the 'no agent' /status path in gateway/run.py — no agent instance is in scope to query for system prompt/tools, and the existing 30-50% overestimate wobble on hygiene is safety-accepted. - Verbose-mode 'Request size' logging — informational only, already counts system prompt via api_messages[0]. Also relabels the feedback line from 'Rough transcript estimate' to 'Approx request size' so the metric label matches what it actually measures. Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217); user report @codecovenant on X (2026-04-30). Closes #14695 Closes #6217
2026-04-30 23:03:54 -07:00
patch("agent.model_metadata.estimate_request_tokens_rough", side_effect=_estimate),
):
result = await runner._handle_compress_command(_make_event())
assert "No changes from compression" in result
assert "Compressed:" not in result
fix(compression): include system prompt + tool schemas in token estimates (#18265) The user-visible /compress banner and the post-compression last_prompt_tokens writeback both counted only the raw message transcript (chars/4). With a 15KB system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of request pressure — a 234x gap. Two user-facing consequences: - Banner shows 'Compressing … (~45 tokens)…' while compression is actually firing on 10K+ tokens of real pressure, confusing users about why compression triggered (reported by @codecovenant on X; #6217). - Post-compression last_prompt_tokens writeback omits tool schemas, so the next should_compress() check compares real usage against a stale underestimate — compression triggers late, potentially past the model's context limit on small-context models (#14695). Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough() at every user-visible banner and at the post-compression writeback. estimate_request_tokens_rough() already existed for exactly this purpose and includes system prompt + tool schemas. Touched call sites: - run_agent.py: post-compression last_prompt_tokens writeback, post-tool call should_compress() fallback when provider usage is missing - cli.py: /compress banner + summary - gateway/run.py: gateway /compress banner + summary - tui_gateway/server.py: TUI /compress status + summary - acp_adapter/server.py: ACP /compact before/after Left intentionally alone: - Session-hygiene fallback and the 'no agent' /status path in gateway/run.py — no agent instance is in scope to query for system prompt/tools, and the existing 30-50% overestimate wobble on hygiene is safety-accepted. - Verbose-mode 'Request size' logging — informational only, already counts system prompt via api_messages[0]. Also relabels the feedback line from 'Rough transcript estimate' to 'Approx request size' so the metric label matches what it actually measures. Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217); user report @codecovenant on X (2026-04-30). Closes #14695 Closes #6217
2026-04-30 23:03:54 -07:00
assert "Approx request size: ~100 tokens (unchanged)" in result
agent_instance.shutdown_memory_provider.assert_called_once()
agent_instance.close.assert_called_once()
@pytest.mark.asyncio
async def test_compress_command_explains_when_token_estimate_rises():
history = _make_history()
compressed = [
history[0],
{"role": "assistant", "content": "Dense summary that still counts as more tokens."},
history[-1],
]
runner = _make_runner(history)
agent_instance = MagicMock()
agent_instance.shutdown_memory_provider = MagicMock()
agent_instance.close = MagicMock()
fix(compression): include system prompt + tool schemas in token estimates (#18265) The user-visible /compress banner and the post-compression last_prompt_tokens writeback both counted only the raw message transcript (chars/4). With a 15KB system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of request pressure — a 234x gap. Two user-facing consequences: - Banner shows 'Compressing … (~45 tokens)…' while compression is actually firing on 10K+ tokens of real pressure, confusing users about why compression triggered (reported by @codecovenant on X; #6217). - Post-compression last_prompt_tokens writeback omits tool schemas, so the next should_compress() check compares real usage against a stale underestimate — compression triggers late, potentially past the model's context limit on small-context models (#14695). Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough() at every user-visible banner and at the post-compression writeback. estimate_request_tokens_rough() already existed for exactly this purpose and includes system prompt + tool schemas. Touched call sites: - run_agent.py: post-compression last_prompt_tokens writeback, post-tool call should_compress() fallback when provider usage is missing - cli.py: /compress banner + summary - gateway/run.py: gateway /compress banner + summary - tui_gateway/server.py: TUI /compress status + summary - acp_adapter/server.py: ACP /compact before/after Left intentionally alone: - Session-hygiene fallback and the 'no agent' /status path in gateway/run.py — no agent instance is in scope to query for system prompt/tools, and the existing 30-50% overestimate wobble on hygiene is safety-accepted. - Verbose-mode 'Request size' logging — informational only, already counts system prompt via api_messages[0]. Also relabels the feedback line from 'Rough transcript estimate' to 'Approx request size' so the metric label matches what it actually measures. Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217); user report @codecovenant on X (2026-04-30). Closes #14695 Closes #6217
2026-04-30 23:03:54 -07:00
agent_instance._cached_system_prompt = ""
agent_instance.tools = None
agent_instance.context_compressor.has_content_to_compress.return_value = True
agent_instance.session_id = "sess-1"
agent_instance._compress_context.return_value = (compressed, "")
fix(compression): include system prompt + tool schemas in token estimates (#18265) The user-visible /compress banner and the post-compression last_prompt_tokens writeback both counted only the raw message transcript (chars/4). With a 15KB system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of request pressure — a 234x gap. Two user-facing consequences: - Banner shows 'Compressing … (~45 tokens)…' while compression is actually firing on 10K+ tokens of real pressure, confusing users about why compression triggered (reported by @codecovenant on X; #6217). - Post-compression last_prompt_tokens writeback omits tool schemas, so the next should_compress() check compares real usage against a stale underestimate — compression triggers late, potentially past the model's context limit on small-context models (#14695). Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough() at every user-visible banner and at the post-compression writeback. estimate_request_tokens_rough() already existed for exactly this purpose and includes system prompt + tool schemas. Touched call sites: - run_agent.py: post-compression last_prompt_tokens writeback, post-tool call should_compress() fallback when provider usage is missing - cli.py: /compress banner + summary - gateway/run.py: gateway /compress banner + summary - tui_gateway/server.py: TUI /compress status + summary - acp_adapter/server.py: ACP /compact before/after Left intentionally alone: - Session-hygiene fallback and the 'no agent' /status path in gateway/run.py — no agent instance is in scope to query for system prompt/tools, and the existing 30-50% overestimate wobble on hygiene is safety-accepted. - Verbose-mode 'Request size' logging — informational only, already counts system prompt via api_messages[0]. Also relabels the feedback line from 'Rough transcript estimate' to 'Approx request size' so the metric label matches what it actually measures. Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217); user report @codecovenant on X (2026-04-30). Closes #14695 Closes #6217
2026-04-30 23:03:54 -07:00
def _estimate(messages, **_kwargs):
if messages == history:
return 100
if messages == compressed:
return 120
raise AssertionError(f"unexpected transcript: {messages!r}")
with (
patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "test-key"}),
patch("gateway.run._resolve_gateway_model", return_value="test-model"),
patch("run_agent.AIAgent", return_value=agent_instance),
fix(compression): include system prompt + tool schemas in token estimates (#18265) The user-visible /compress banner and the post-compression last_prompt_tokens writeback both counted only the raw message transcript (chars/4). With a 15KB system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of request pressure — a 234x gap. Two user-facing consequences: - Banner shows 'Compressing … (~45 tokens)…' while compression is actually firing on 10K+ tokens of real pressure, confusing users about why compression triggered (reported by @codecovenant on X; #6217). - Post-compression last_prompt_tokens writeback omits tool schemas, so the next should_compress() check compares real usage against a stale underestimate — compression triggers late, potentially past the model's context limit on small-context models (#14695). Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough() at every user-visible banner and at the post-compression writeback. estimate_request_tokens_rough() already existed for exactly this purpose and includes system prompt + tool schemas. Touched call sites: - run_agent.py: post-compression last_prompt_tokens writeback, post-tool call should_compress() fallback when provider usage is missing - cli.py: /compress banner + summary - gateway/run.py: gateway /compress banner + summary - tui_gateway/server.py: TUI /compress status + summary - acp_adapter/server.py: ACP /compact before/after Left intentionally alone: - Session-hygiene fallback and the 'no agent' /status path in gateway/run.py — no agent instance is in scope to query for system prompt/tools, and the existing 30-50% overestimate wobble on hygiene is safety-accepted. - Verbose-mode 'Request size' logging — informational only, already counts system prompt via api_messages[0]. Also relabels the feedback line from 'Rough transcript estimate' to 'Approx request size' so the metric label matches what it actually measures. Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217); user report @codecovenant on X (2026-04-30). Closes #14695 Closes #6217
2026-04-30 23:03:54 -07:00
patch("agent.model_metadata.estimate_request_tokens_rough", side_effect=_estimate),
):
result = await runner._handle_compress_command(_make_event())
assert "Compressed: 4 → 3 messages" in result
fix(compression): include system prompt + tool schemas in token estimates (#18265) The user-visible /compress banner and the post-compression last_prompt_tokens writeback both counted only the raw message transcript (chars/4). With a 15KB system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of request pressure — a 234x gap. Two user-facing consequences: - Banner shows 'Compressing … (~45 tokens)…' while compression is actually firing on 10K+ tokens of real pressure, confusing users about why compression triggered (reported by @codecovenant on X; #6217). - Post-compression last_prompt_tokens writeback omits tool schemas, so the next should_compress() check compares real usage against a stale underestimate — compression triggers late, potentially past the model's context limit on small-context models (#14695). Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough() at every user-visible banner and at the post-compression writeback. estimate_request_tokens_rough() already existed for exactly this purpose and includes system prompt + tool schemas. Touched call sites: - run_agent.py: post-compression last_prompt_tokens writeback, post-tool call should_compress() fallback when provider usage is missing - cli.py: /compress banner + summary - gateway/run.py: gateway /compress banner + summary - tui_gateway/server.py: TUI /compress status + summary - acp_adapter/server.py: ACP /compact before/after Left intentionally alone: - Session-hygiene fallback and the 'no agent' /status path in gateway/run.py — no agent instance is in scope to query for system prompt/tools, and the existing 30-50% overestimate wobble on hygiene is safety-accepted. - Verbose-mode 'Request size' logging — informational only, already counts system prompt via api_messages[0]. Also relabels the feedback line from 'Rough transcript estimate' to 'Approx request size' so the metric label matches what it actually measures. Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217); user report @codecovenant on X (2026-04-30). Closes #14695 Closes #6217
2026-04-30 23:03:54 -07:00
assert "Approx request size: ~100 → ~120 tokens" in result
assert "denser summaries" in result
agent_instance.shutdown_memory_provider.assert_called_once()
agent_instance.close.assert_called_once()
@pytest.mark.asyncio
fix(compress): abort instead of dropping messages when summary LLM fails (#28102) When auxiliary compression's summary generation returns None (aux model errored, returned non-JSON, timed out, etc.) the compressor previously still dropped every middle message between compress_start..compress_end and replaced them with a static 'Summary generation was unavailable' placeholder. The session kept going but the user silently lost N turns of context for nothing. New behavior: on summary failure, compress() aborts entirely — returns the input messages unchanged and sets _last_compress_aborted=True. The existing _summary_failure_cooldown_until gate (30-60s) keeps the aux model from being burned on every turn. Auto-compress callers detect the no-op (len(after) == len(before)) and stop looping. The chat is 'frozen' at its current size until the next /compress or /new. Manual /compress (CLI + gateway) now passes force=True which clears the cooldown so users can retry immediately after an auto-abort. If the manual retry also fails, the user gets a visible warning telling them nothing was dropped and how to retry. - agent/context_compressor.py: compress() gains force= kwarg; failure branch sets _last_compress_aborted and returns messages unchanged instead of inserting placeholder. - run_agent.py: _compress_context() detects abort, surfaces warning, skips session-rotation entirely, returns messages unchanged. - cli.py + gateway/run.py: manual /compress paths pass force=True. - gateway/run.py: hygiene + /compress handlers detect _last_compress_aborted and emit the new 'Compression aborted' warning (gateway.compress.aborted) instead of the old 'N historical messages were removed' message. - locales/*.yaml: new gateway.compress.aborted key in all 16 locales. - tests: updated to assert the abort contract (messages preserved, compression_count not incremented, abort flag set, no placeholder leaked). New test_force_true_bypasses_failure_cooldown covers the manual-retry path.
2026-05-18 10:19:40 -07:00
async def test_compress_command_appends_warning_when_compression_aborts():
"""When the auxiliary summariser fails and the compressor ABORTS (returns
messages unchanged), /compress must append a visible warning to its
reply telling the user nothing was dropped and how to retry. Otherwise
the failure is silently logged and the user has no idea why nothing
happened."""
history = _make_history()
fix(compress): abort instead of dropping messages when summary LLM fails (#28102) When auxiliary compression's summary generation returns None (aux model errored, returned non-JSON, timed out, etc.) the compressor previously still dropped every middle message between compress_start..compress_end and replaced them with a static 'Summary generation was unavailable' placeholder. The session kept going but the user silently lost N turns of context for nothing. New behavior: on summary failure, compress() aborts entirely — returns the input messages unchanged and sets _last_compress_aborted=True. The existing _summary_failure_cooldown_until gate (30-60s) keeps the aux model from being burned on every turn. Auto-compress callers detect the no-op (len(after) == len(before)) and stop looping. The chat is 'frozen' at its current size until the next /compress or /new. Manual /compress (CLI + gateway) now passes force=True which clears the cooldown so users can retry immediately after an auto-abort. If the manual retry also fails, the user gets a visible warning telling them nothing was dropped and how to retry. - agent/context_compressor.py: compress() gains force= kwarg; failure branch sets _last_compress_aborted and returns messages unchanged instead of inserting placeholder. - run_agent.py: _compress_context() detects abort, surfaces warning, skips session-rotation entirely, returns messages unchanged. - cli.py + gateway/run.py: manual /compress paths pass force=True. - gateway/run.py: hygiene + /compress handlers detect _last_compress_aborted and emit the new 'Compression aborted' warning (gateway.compress.aborted) instead of the old 'N historical messages were removed' message. - locales/*.yaml: new gateway.compress.aborted key in all 16 locales. - tests: updated to assert the abort contract (messages preserved, compression_count not incremented, abort flag set, no placeholder leaked). New test_force_true_bypasses_failure_cooldown covers the manual-retry path.
2026-05-18 10:19:40 -07:00
# Abort path: compressor returns the input messages unchanged.
compressed = list(history)
runner = _make_runner(history)
agent_instance = MagicMock()
agent_instance.shutdown_memory_provider = MagicMock()
agent_instance.close = MagicMock()
fix(compression): include system prompt + tool schemas in token estimates (#18265) The user-visible /compress banner and the post-compression last_prompt_tokens writeback both counted only the raw message transcript (chars/4). With a 15KB system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of request pressure — a 234x gap. Two user-facing consequences: - Banner shows 'Compressing … (~45 tokens)…' while compression is actually firing on 10K+ tokens of real pressure, confusing users about why compression triggered (reported by @codecovenant on X; #6217). - Post-compression last_prompt_tokens writeback omits tool schemas, so the next should_compress() check compares real usage against a stale underestimate — compression triggers late, potentially past the model's context limit on small-context models (#14695). Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough() at every user-visible banner and at the post-compression writeback. estimate_request_tokens_rough() already existed for exactly this purpose and includes system prompt + tool schemas. Touched call sites: - run_agent.py: post-compression last_prompt_tokens writeback, post-tool call should_compress() fallback when provider usage is missing - cli.py: /compress banner + summary - gateway/run.py: gateway /compress banner + summary - tui_gateway/server.py: TUI /compress status + summary - acp_adapter/server.py: ACP /compact before/after Left intentionally alone: - Session-hygiene fallback and the 'no agent' /status path in gateway/run.py — no agent instance is in scope to query for system prompt/tools, and the existing 30-50% overestimate wobble on hygiene is safety-accepted. - Verbose-mode 'Request size' logging — informational only, already counts system prompt via api_messages[0]. Also relabels the feedback line from 'Rough transcript estimate' to 'Approx request size' so the metric label matches what it actually measures. Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217); user report @codecovenant on X (2026-04-30). Closes #14695 Closes #6217
2026-04-30 23:03:54 -07:00
agent_instance._cached_system_prompt = ""
agent_instance.tools = None
agent_instance.context_compressor.has_content_to_compress.return_value = True
fix(compress): abort instead of dropping messages when summary LLM fails (#28102) When auxiliary compression's summary generation returns None (aux model errored, returned non-JSON, timed out, etc.) the compressor previously still dropped every middle message between compress_start..compress_end and replaced them with a static 'Summary generation was unavailable' placeholder. The session kept going but the user silently lost N turns of context for nothing. New behavior: on summary failure, compress() aborts entirely — returns the input messages unchanged and sets _last_compress_aborted=True. The existing _summary_failure_cooldown_until gate (30-60s) keeps the aux model from being burned on every turn. Auto-compress callers detect the no-op (len(after) == len(before)) and stop looping. The chat is 'frozen' at its current size until the next /compress or /new. Manual /compress (CLI + gateway) now passes force=True which clears the cooldown so users can retry immediately after an auto-abort. If the manual retry also fails, the user gets a visible warning telling them nothing was dropped and how to retry. - agent/context_compressor.py: compress() gains force= kwarg; failure branch sets _last_compress_aborted and returns messages unchanged instead of inserting placeholder. - run_agent.py: _compress_context() detects abort, surfaces warning, skips session-rotation entirely, returns messages unchanged. - cli.py + gateway/run.py: manual /compress paths pass force=True. - gateway/run.py: hygiene + /compress handlers detect _last_compress_aborted and emit the new 'Compression aborted' warning (gateway.compress.aborted) instead of the old 'N historical messages were removed' message. - locales/*.yaml: new gateway.compress.aborted key in all 16 locales. - tests: updated to assert the abort contract (messages preserved, compression_count not incremented, abort flag set, no placeholder leaked). New test_force_true_bypasses_failure_cooldown covers the manual-retry path.
2026-05-18 10:19:40 -07:00
# Simulate compression aborting (force=True bypassed cooldown but the
# aux LLM is genuinely broken).
agent_instance.context_compressor._last_compress_aborted = True
agent_instance.context_compressor._last_summary_fallback_used = False
agent_instance.context_compressor._last_summary_dropped_count = 0
agent_instance.context_compressor._last_summary_error = (
"404 model not found: gemini-3-flash-preview"
)
agent_instance.session_id = "sess-1"
agent_instance._compress_context.return_value = (compressed, "")
fix(compression): include system prompt + tool schemas in token estimates (#18265) The user-visible /compress banner and the post-compression last_prompt_tokens writeback both counted only the raw message transcript (chars/4). With a 15KB system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of request pressure — a 234x gap. Two user-facing consequences: - Banner shows 'Compressing … (~45 tokens)…' while compression is actually firing on 10K+ tokens of real pressure, confusing users about why compression triggered (reported by @codecovenant on X; #6217). - Post-compression last_prompt_tokens writeback omits tool schemas, so the next should_compress() check compares real usage against a stale underestimate — compression triggers late, potentially past the model's context limit on small-context models (#14695). Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough() at every user-visible banner and at the post-compression writeback. estimate_request_tokens_rough() already existed for exactly this purpose and includes system prompt + tool schemas. Touched call sites: - run_agent.py: post-compression last_prompt_tokens writeback, post-tool call should_compress() fallback when provider usage is missing - cli.py: /compress banner + summary - gateway/run.py: gateway /compress banner + summary - tui_gateway/server.py: TUI /compress status + summary - acp_adapter/server.py: ACP /compact before/after Left intentionally alone: - Session-hygiene fallback and the 'no agent' /status path in gateway/run.py — no agent instance is in scope to query for system prompt/tools, and the existing 30-50% overestimate wobble on hygiene is safety-accepted. - Verbose-mode 'Request size' logging — informational only, already counts system prompt via api_messages[0]. Also relabels the feedback line from 'Rough transcript estimate' to 'Approx request size' so the metric label matches what it actually measures. Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217); user report @codecovenant on X (2026-04-30). Closes #14695 Closes #6217
2026-04-30 23:03:54 -07:00
def _estimate(messages, **_kwargs):
if messages == history:
return 100
if messages == compressed:
fix(compress): abort instead of dropping messages when summary LLM fails (#28102) When auxiliary compression's summary generation returns None (aux model errored, returned non-JSON, timed out, etc.) the compressor previously still dropped every middle message between compress_start..compress_end and replaced them with a static 'Summary generation was unavailable' placeholder. The session kept going but the user silently lost N turns of context for nothing. New behavior: on summary failure, compress() aborts entirely — returns the input messages unchanged and sets _last_compress_aborted=True. The existing _summary_failure_cooldown_until gate (30-60s) keeps the aux model from being burned on every turn. Auto-compress callers detect the no-op (len(after) == len(before)) and stop looping. The chat is 'frozen' at its current size until the next /compress or /new. Manual /compress (CLI + gateway) now passes force=True which clears the cooldown so users can retry immediately after an auto-abort. If the manual retry also fails, the user gets a visible warning telling them nothing was dropped and how to retry. - agent/context_compressor.py: compress() gains force= kwarg; failure branch sets _last_compress_aborted and returns messages unchanged instead of inserting placeholder. - run_agent.py: _compress_context() detects abort, surfaces warning, skips session-rotation entirely, returns messages unchanged. - cli.py + gateway/run.py: manual /compress paths pass force=True. - gateway/run.py: hygiene + /compress handlers detect _last_compress_aborted and emit the new 'Compression aborted' warning (gateway.compress.aborted) instead of the old 'N historical messages were removed' message. - locales/*.yaml: new gateway.compress.aborted key in all 16 locales. - tests: updated to assert the abort contract (messages preserved, compression_count not incremented, abort flag set, no placeholder leaked). New test_force_true_bypasses_failure_cooldown covers the manual-retry path.
2026-05-18 10:19:40 -07:00
return 100
raise AssertionError(f"unexpected transcript: {messages!r}")
with (
patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "***"}),
patch("gateway.run._resolve_gateway_model", return_value="test-model"),
patch("run_agent.AIAgent", return_value=agent_instance),
fix(compression): include system prompt + tool schemas in token estimates (#18265) The user-visible /compress banner and the post-compression last_prompt_tokens writeback both counted only the raw message transcript (chars/4). With a 15KB system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of request pressure — a 234x gap. Two user-facing consequences: - Banner shows 'Compressing … (~45 tokens)…' while compression is actually firing on 10K+ tokens of real pressure, confusing users about why compression triggered (reported by @codecovenant on X; #6217). - Post-compression last_prompt_tokens writeback omits tool schemas, so the next should_compress() check compares real usage against a stale underestimate — compression triggers late, potentially past the model's context limit on small-context models (#14695). Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough() at every user-visible banner and at the post-compression writeback. estimate_request_tokens_rough() already existed for exactly this purpose and includes system prompt + tool schemas. Touched call sites: - run_agent.py: post-compression last_prompt_tokens writeback, post-tool call should_compress() fallback when provider usage is missing - cli.py: /compress banner + summary - gateway/run.py: gateway /compress banner + summary - tui_gateway/server.py: TUI /compress status + summary - acp_adapter/server.py: ACP /compact before/after Left intentionally alone: - Session-hygiene fallback and the 'no agent' /status path in gateway/run.py — no agent instance is in scope to query for system prompt/tools, and the existing 30-50% overestimate wobble on hygiene is safety-accepted. - Verbose-mode 'Request size' logging — informational only, already counts system prompt via api_messages[0]. Also relabels the feedback line from 'Rough transcript estimate' to 'Approx request size' so the metric label matches what it actually measures. Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217); user report @codecovenant on X (2026-04-30). Closes #14695 Closes #6217
2026-04-30 23:03:54 -07:00
patch("agent.model_metadata.estimate_request_tokens_rough", side_effect=_estimate),
):
result = await runner._handle_compress_command(_make_event())
fix(compress): abort instead of dropping messages when summary LLM fails (#28102) When auxiliary compression's summary generation returns None (aux model errored, returned non-JSON, timed out, etc.) the compressor previously still dropped every middle message between compress_start..compress_end and replaced them with a static 'Summary generation was unavailable' placeholder. The session kept going but the user silently lost N turns of context for nothing. New behavior: on summary failure, compress() aborts entirely — returns the input messages unchanged and sets _last_compress_aborted=True. The existing _summary_failure_cooldown_until gate (30-60s) keeps the aux model from being burned on every turn. Auto-compress callers detect the no-op (len(after) == len(before)) and stop looping. The chat is 'frozen' at its current size until the next /compress or /new. Manual /compress (CLI + gateway) now passes force=True which clears the cooldown so users can retry immediately after an auto-abort. If the manual retry also fails, the user gets a visible warning telling them nothing was dropped and how to retry. - agent/context_compressor.py: compress() gains force= kwarg; failure branch sets _last_compress_aborted and returns messages unchanged instead of inserting placeholder. - run_agent.py: _compress_context() detects abort, surfaces warning, skips session-rotation entirely, returns messages unchanged. - cli.py + gateway/run.py: manual /compress paths pass force=True. - gateway/run.py: hygiene + /compress handlers detect _last_compress_aborted and emit the new 'Compression aborted' warning (gateway.compress.aborted) instead of the old 'N historical messages were removed' message. - locales/*.yaml: new gateway.compress.aborted key in all 16 locales. - tests: updated to assert the abort contract (messages preserved, compression_count not incremented, abort flag set, no placeholder leaked). New test_force_true_bypasses_failure_cooldown covers the manual-retry path.
2026-05-18 10:19:40 -07:00
# A clearly-marked warning must be appended.
assert "⚠️" in result
fix(compress): abort instead of dropping messages when summary LLM fails (#28102) When auxiliary compression's summary generation returns None (aux model errored, returned non-JSON, timed out, etc.) the compressor previously still dropped every middle message between compress_start..compress_end and replaced them with a static 'Summary generation was unavailable' placeholder. The session kept going but the user silently lost N turns of context for nothing. New behavior: on summary failure, compress() aborts entirely — returns the input messages unchanged and sets _last_compress_aborted=True. The existing _summary_failure_cooldown_until gate (30-60s) keeps the aux model from being burned on every turn. Auto-compress callers detect the no-op (len(after) == len(before)) and stop looping. The chat is 'frozen' at its current size until the next /compress or /new. Manual /compress (CLI + gateway) now passes force=True which clears the cooldown so users can retry immediately after an auto-abort. If the manual retry also fails, the user gets a visible warning telling them nothing was dropped and how to retry. - agent/context_compressor.py: compress() gains force= kwarg; failure branch sets _last_compress_aborted and returns messages unchanged instead of inserting placeholder. - run_agent.py: _compress_context() detects abort, surfaces warning, skips session-rotation entirely, returns messages unchanged. - cli.py + gateway/run.py: manual /compress paths pass force=True. - gateway/run.py: hygiene + /compress handlers detect _last_compress_aborted and emit the new 'Compression aborted' warning (gateway.compress.aborted) instead of the old 'N historical messages were removed' message. - locales/*.yaml: new gateway.compress.aborted key in all 16 locales. - tests: updated to assert the abort contract (messages preserved, compression_count not incremented, abort flag set, no placeholder leaked). New test_force_true_bypasses_failure_cooldown covers the manual-retry path.
2026-05-18 10:19:40 -07:00
assert "Compression aborted" in result
# Underlying error must surface so users can fix their config.
assert "404 model not found" in result
fix(compress): abort instead of dropping messages when summary LLM fails (#28102) When auxiliary compression's summary generation returns None (aux model errored, returned non-JSON, timed out, etc.) the compressor previously still dropped every middle message between compress_start..compress_end and replaced them with a static 'Summary generation was unavailable' placeholder. The session kept going but the user silently lost N turns of context for nothing. New behavior: on summary failure, compress() aborts entirely — returns the input messages unchanged and sets _last_compress_aborted=True. The existing _summary_failure_cooldown_until gate (30-60s) keeps the aux model from being burned on every turn. Auto-compress callers detect the no-op (len(after) == len(before)) and stop looping. The chat is 'frozen' at its current size until the next /compress or /new. Manual /compress (CLI + gateway) now passes force=True which clears the cooldown so users can retry immediately after an auto-abort. If the manual retry also fails, the user gets a visible warning telling them nothing was dropped and how to retry. - agent/context_compressor.py: compress() gains force= kwarg; failure branch sets _last_compress_aborted and returns messages unchanged instead of inserting placeholder. - run_agent.py: _compress_context() detects abort, surfaces warning, skips session-rotation entirely, returns messages unchanged. - cli.py + gateway/run.py: manual /compress paths pass force=True. - gateway/run.py: hygiene + /compress handlers detect _last_compress_aborted and emit the new 'Compression aborted' warning (gateway.compress.aborted) instead of the old 'N historical messages were removed' message. - locales/*.yaml: new gateway.compress.aborted key in all 16 locales. - tests: updated to assert the abort contract (messages preserved, compression_count not incremented, abort flag set, no placeholder leaked). New test_force_true_bypasses_failure_cooldown covers the manual-retry path.
2026-05-18 10:19:40 -07:00
# User must be told nothing was dropped — the whole point of the
# new behavior is no silent data loss.
assert "No messages were dropped" in result
agent_instance.shutdown_memory_provider.assert_called_once()
agent_instance.close.assert_called_once()
@pytest.mark.asyncio
async def test_compress_command_surfaces_aux_model_failure_even_when_recovered():
"""When the user's configured ``auxiliary.compression.model`` errors out
but compression recovers by retrying on the main model, /compress must
STILL inform the user. Silent recovery hides broken config the user
needs to fix."""
history = _make_history()
# Compressed transcript — normal successful compression, no placeholder.
compressed = [
history[0],
{"role": "assistant", "content": "summary via main model"},
history[-1],
]
runner = _make_runner(history)
agent_instance = MagicMock()
agent_instance.shutdown_memory_provider = MagicMock()
agent_instance.close = MagicMock()
fix(compression): include system prompt + tool schemas in token estimates (#18265) The user-visible /compress banner and the post-compression last_prompt_tokens writeback both counted only the raw message transcript (chars/4). With a 15KB system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of request pressure — a 234x gap. Two user-facing consequences: - Banner shows 'Compressing … (~45 tokens)…' while compression is actually firing on 10K+ tokens of real pressure, confusing users about why compression triggered (reported by @codecovenant on X; #6217). - Post-compression last_prompt_tokens writeback omits tool schemas, so the next should_compress() check compares real usage against a stale underestimate — compression triggers late, potentially past the model's context limit on small-context models (#14695). Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough() at every user-visible banner and at the post-compression writeback. estimate_request_tokens_rough() already existed for exactly this purpose and includes system prompt + tool schemas. Touched call sites: - run_agent.py: post-compression last_prompt_tokens writeback, post-tool call should_compress() fallback when provider usage is missing - cli.py: /compress banner + summary - gateway/run.py: gateway /compress banner + summary - tui_gateway/server.py: TUI /compress status + summary - acp_adapter/server.py: ACP /compact before/after Left intentionally alone: - Session-hygiene fallback and the 'no agent' /status path in gateway/run.py — no agent instance is in scope to query for system prompt/tools, and the existing 30-50% overestimate wobble on hygiene is safety-accepted. - Verbose-mode 'Request size' logging — informational only, already counts system prompt via api_messages[0]. Also relabels the feedback line from 'Rough transcript estimate' to 'Approx request size' so the metric label matches what it actually measures. Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217); user report @codecovenant on X (2026-04-30). Closes #14695 Closes #6217
2026-04-30 23:03:54 -07:00
agent_instance._cached_system_prompt = ""
agent_instance.tools = None
agent_instance.context_compressor.has_content_to_compress.return_value = True
# Fallback placeholder was NOT used — recovery succeeded.
fix(compress): abort instead of dropping messages when summary LLM fails (#28102) When auxiliary compression's summary generation returns None (aux model errored, returned non-JSON, timed out, etc.) the compressor previously still dropped every middle message between compress_start..compress_end and replaced them with a static 'Summary generation was unavailable' placeholder. The session kept going but the user silently lost N turns of context for nothing. New behavior: on summary failure, compress() aborts entirely — returns the input messages unchanged and sets _last_compress_aborted=True. The existing _summary_failure_cooldown_until gate (30-60s) keeps the aux model from being burned on every turn. Auto-compress callers detect the no-op (len(after) == len(before)) and stop looping. The chat is 'frozen' at its current size until the next /compress or /new. Manual /compress (CLI + gateway) now passes force=True which clears the cooldown so users can retry immediately after an auto-abort. If the manual retry also fails, the user gets a visible warning telling them nothing was dropped and how to retry. - agent/context_compressor.py: compress() gains force= kwarg; failure branch sets _last_compress_aborted and returns messages unchanged instead of inserting placeholder. - run_agent.py: _compress_context() detects abort, surfaces warning, skips session-rotation entirely, returns messages unchanged. - cli.py + gateway/run.py: manual /compress paths pass force=True. - gateway/run.py: hygiene + /compress handlers detect _last_compress_aborted and emit the new 'Compression aborted' warning (gateway.compress.aborted) instead of the old 'N historical messages were removed' message. - locales/*.yaml: new gateway.compress.aborted key in all 16 locales. - tests: updated to assert the abort contract (messages preserved, compression_count not incremented, abort flag set, no placeholder leaked). New test_force_true_bypasses_failure_cooldown covers the manual-retry path.
2026-05-18 10:19:40 -07:00
agent_instance.context_compressor._last_compress_aborted = False
agent_instance.context_compressor._last_summary_fallback_used = False
agent_instance.context_compressor._last_summary_dropped_count = 0
agent_instance.context_compressor._last_summary_error = None
# But the configured aux model DID fail before the retry succeeded.
agent_instance.context_compressor._last_aux_model_failure_model = (
"gemini-3-flash-preview"
)
agent_instance.context_compressor._last_aux_model_failure_error = (
"404 model not found: gemini-3-flash-preview"
)
agent_instance.session_id = "sess-1"
agent_instance._compress_context.return_value = (compressed, "")
fix(compression): include system prompt + tool schemas in token estimates (#18265) The user-visible /compress banner and the post-compression last_prompt_tokens writeback both counted only the raw message transcript (chars/4). With a 15KB system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of request pressure — a 234x gap. Two user-facing consequences: - Banner shows 'Compressing … (~45 tokens)…' while compression is actually firing on 10K+ tokens of real pressure, confusing users about why compression triggered (reported by @codecovenant on X; #6217). - Post-compression last_prompt_tokens writeback omits tool schemas, so the next should_compress() check compares real usage against a stale underestimate — compression triggers late, potentially past the model's context limit on small-context models (#14695). Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough() at every user-visible banner and at the post-compression writeback. estimate_request_tokens_rough() already existed for exactly this purpose and includes system prompt + tool schemas. Touched call sites: - run_agent.py: post-compression last_prompt_tokens writeback, post-tool call should_compress() fallback when provider usage is missing - cli.py: /compress banner + summary - gateway/run.py: gateway /compress banner + summary - tui_gateway/server.py: TUI /compress status + summary - acp_adapter/server.py: ACP /compact before/after Left intentionally alone: - Session-hygiene fallback and the 'no agent' /status path in gateway/run.py — no agent instance is in scope to query for system prompt/tools, and the existing 30-50% overestimate wobble on hygiene is safety-accepted. - Verbose-mode 'Request size' logging — informational only, already counts system prompt via api_messages[0]. Also relabels the feedback line from 'Rough transcript estimate' to 'Approx request size' so the metric label matches what it actually measures. Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217); user report @codecovenant on X (2026-04-30). Closes #14695 Closes #6217
2026-04-30 23:03:54 -07:00
def _estimate(messages, **_kwargs):
if messages == history:
return 100
if messages == compressed:
return 60
raise AssertionError(f"unexpected transcript: {messages!r}")
with (
patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "***"}),
patch("gateway.run._resolve_gateway_model", return_value="test-model"),
patch("run_agent.AIAgent", return_value=agent_instance),
fix(compression): include system prompt + tool schemas in token estimates (#18265) The user-visible /compress banner and the post-compression last_prompt_tokens writeback both counted only the raw message transcript (chars/4). With a 15KB system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of request pressure — a 234x gap. Two user-facing consequences: - Banner shows 'Compressing … (~45 tokens)…' while compression is actually firing on 10K+ tokens of real pressure, confusing users about why compression triggered (reported by @codecovenant on X; #6217). - Post-compression last_prompt_tokens writeback omits tool schemas, so the next should_compress() check compares real usage against a stale underestimate — compression triggers late, potentially past the model's context limit on small-context models (#14695). Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough() at every user-visible banner and at the post-compression writeback. estimate_request_tokens_rough() already existed for exactly this purpose and includes system prompt + tool schemas. Touched call sites: - run_agent.py: post-compression last_prompt_tokens writeback, post-tool call should_compress() fallback when provider usage is missing - cli.py: /compress banner + summary - gateway/run.py: gateway /compress banner + summary - tui_gateway/server.py: TUI /compress status + summary - acp_adapter/server.py: ACP /compact before/after Left intentionally alone: - Session-hygiene fallback and the 'no agent' /status path in gateway/run.py — no agent instance is in scope to query for system prompt/tools, and the existing 30-50% overestimate wobble on hygiene is safety-accepted. - Verbose-mode 'Request size' logging — informational only, already counts system prompt via api_messages[0]. Also relabels the feedback line from 'Rough transcript estimate' to 'Approx request size' so the metric label matches what it actually measures. Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217); user report @codecovenant on X (2026-04-30). Closes #14695 Closes #6217
2026-04-30 23:03:54 -07:00
patch("agent.model_metadata.estimate_request_tokens_rough", side_effect=_estimate),
):
result = await runner._handle_compress_command(_make_event())
# Compression succeeded
assert "Compressed:" in result
# No ⚠️ warning (that's reserved for dropped-turns case)
assert "⚠️" not in result
# But there IS an info note about the broken aux model
assert "" in result
assert "gemini-3-flash-preview" in result
assert "404" in result
assert "auxiliary.compression.model" in result
# The user's context is explicitly called out as intact
assert "intact" in result
agent_instance.shutdown_memory_provider.assert_called_once()
agent_instance.close.assert_called_once()
@pytest.mark.asyncio
async def test_compress_command_passes_session_db_and_persists_rotated_session():
"""session_db must be wired into the /compress temp agent so that
_compress_context can actually rotate the session and persist the
compressed transcript without it compression is a silent no-op."""
history = _make_history()
compressed = [
history[0],
{"role": "assistant", "content": "compressed summary"},
history[-1],
]
runner = _make_runner(history)
runner._session_db = object()
agent_instance = MagicMock()
agent_instance.shutdown_memory_provider = MagicMock()
agent_instance.close = MagicMock()
agent_instance._cached_system_prompt = ""
agent_instance.tools = None
agent_instance.context_compressor.has_content_to_compress.return_value = True
agent_instance.compression_in_place = False
agent_instance.session_id = "sess-1"
def _compress(messages, *_args, **_kwargs):
agent_instance.session_id = "sess-2"
return compressed, ""
agent_instance._compress_context.side_effect = _compress
def _estimate(messages, **_kwargs):
if messages == history:
return 100
if messages == compressed:
return 60
raise AssertionError(f"unexpected transcript: {messages!r}")
with (
patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "***"}),
patch("gateway.run._resolve_gateway_model", return_value="test-model"),
patch("run_agent.AIAgent", return_value=agent_instance) as mock_agent_cls,
patch("agent.model_metadata.estimate_request_tokens_rough", side_effect=_estimate),
):
result = await runner._handle_compress_command(_make_event())
assert "Compressed:" in result
mock_agent_cls.assert_called_once()
assert mock_agent_cls.call_args.kwargs["session_db"] is runner._session_db
runner.session_store._save.assert_called_once()
runner.session_store.rewrite_transcript.assert_called_once_with(
"sess-2", compressed
)
runner.session_store.update_session.assert_called_once_with(
build_session_key(_make_source()), last_prompt_tokens=0
)
agent_instance.shutdown_memory_provider.assert_called_once()
agent_instance.close.assert_called_once()