Files
strix/strix/llm/compaction.py
T

388 lines
14 KiB
Python

"""Provider-agnostic conversation compaction.
When an agent's session grows past the model's usable context window, older
turns are summarised into a single checkpoint while the most recent turns are
kept verbatim. This runs for every LiteLLM provider (not just OpenAI), keeps a
security-focused structured summary, and preserves tool-call/tool-result
pairing so the trimmed history is still valid provider input.
"""
from __future__ import annotations
import logging
from typing import TYPE_CHECKING, Any
from agents.model_settings import ModelSettings
from agents.models.interface import ModelTracing
from litellm.exceptions import BadRequestError, ContextWindowExceededError
from openai.types.responses import ResponseOutputMessage, ResponseOutputText
from strix.config import load_settings
from strix.config.models import StrixProvider
from strix.core.inputs import make_model_settings
from strix.core.sessions import replace_session_items, session_write_lock
from strix.llm.context_budget import context_window, count_tokens, output_limit
if TYPE_CHECKING:
from agents.items import ModelResponse
from agents.memory import Session
logger = logging.getLogger(__name__)
_CHECKPOINT_TAG = "<conversation-checkpoint>"
_TOOL_OUTPUT_MAX_CHARS = 2_000
_MIN_ITEMS_TO_COMPACT = 6
_HEAD_TRUNCATED_MARKER = "\n\n[... older conversation omitted to fit the summary request ...]\n\n"
# Providers that don't type overflow errors (OpenRouter maps every 400 to a
# plain BadRequestError) leave only the message to go on, so we match it the way
# LiteLLM's own checker does — but with rate-limit exclusions first, so a
# throttling 429 is never mistaken for an overflow and sent into compaction.
_OVERFLOW_EXCLUSIONS = (
"rate limit",
"too many requests",
"throttling",
"service unavailable",
"quota",
)
_OVERFLOW_MARKERS = (
"context length",
"context window",
"context_length_exceeded",
"prompt is too long",
"input is too long",
"input length",
"maximum prompt length",
"reduce the length of the messages",
"too many tokens",
"token limit exceeded",
"request entity too large",
)
def is_context_overflow(exc: BaseException) -> bool:
"""Whether ``exc`` is a model context-window-overflow error.
LiteLLM types most providers' overflow as ContextWindowExceededError, but its
OpenRouter branch raises a plain BadRequestError, so for that we fall back to
matching the provider message.
"""
if isinstance(exc, ContextWindowExceededError):
return True
if isinstance(exc, BadRequestError):
msg = str(exc).lower()
if any(x in msg for x in _OVERFLOW_EXCLUSIONS):
return False
return any(x in msg for x in _OVERFLOW_MARKERS)
return False
_SUMMARY_INSTRUCTIONS = """\
You are compacting the earlier part of an autonomous security-testing agent's \
conversation so it fits the model context window. Produce a dense, factual \
record that lets the agent continue with no loss of important state.
This is a security engagement: dropped findings mean lost vulnerabilities. Be \
EXHAUSTIVE, not concise. Enumerate every distinct item as its own bullet — \
never merge, deduplicate, generalise, or omit distinct findings, credentials, \
or dead ends, even if they seem minor or repetitive. If the source mentions \
five vulnerabilities, list five. Copy exact values verbatim: URLs, endpoints, \
file paths, parameters, payloads, credentials, tokens, keys, hashes, cracked \
passwords, software versions, and error messages — never paraphrase or \
placeholder them. Do not invent anything and do not describe this compaction \
process.
Return Markdown with exactly these sections:
## Objective
The overall goal and target scope.
## Vulnerabilities & Findings
One bullet per DISTINCT vulnerability or finding (SQLi, XSS, SSRF, auth bypass, \
misconfig, etc.). For each: type, exact location (URL/endpoint/param/file), the \
verbatim payload or proof, confirmation status, and impact. List them all.
## Credentials & Secrets
One bullet per credential, secret, API key, token, hash, or cracked password, \
copied verbatim with where it applies. Write "(none)" only if truly none.
## System & Recon Details
Architecture, tech stack, versions, discovered endpoints/paths/params, and \
other weak points worth keeping.
## Work State
- Completed: what has been verified or finished.
- Active: what is in progress right now.
- Blocked: anything stuck and why.
## Failed Attempts & Dead Ends
One bullet per approach already tried that did not work (including WAF blocks, \
filtered inputs, non-exploitable leads) so they are not repeated. Write \
"(none)" only if truly none.
## Next Move
The concrete next step(s) the agent intended to take.
## Relevant Files
Files/notes/reports created or modified and their purpose."""
def _content_text(content: Any) -> str:
if isinstance(content, str):
return content
if isinstance(content, list):
parts: list[str] = []
for block in content:
if not isinstance(block, dict):
continue
text = block.get("text")
if isinstance(text, str):
parts.append(text)
elif block.get("type") in {"input_image", "image_url", "output_image"}:
parts.append("[image]")
return "\n".join(parts)
return ""
def _truncate(text: str, limit: int) -> str:
return text if len(text) <= limit else f"{text[:limit]}\n[truncated]"
def _serialize_item(item: Any) -> str:
if not isinstance(item, dict):
return str(item)
item_type = item.get("type")
role = item.get("role")
if item_type == "function_call":
args = _truncate(str(item.get("arguments", "")), _TOOL_OUTPUT_MAX_CHARS)
return f"[tool_call {item.get('name', '?')}] {args}"
if item_type == "function_call_output":
output = item.get("output")
text = output if isinstance(output, str) else _content_text(output)
return f"[tool_result] {_truncate(text, _TOOL_OUTPUT_MAX_CHARS)}"
if item_type == "reasoning":
return ""
if role or item_type == "message":
return f"[{role or 'assistant'}] {_content_text(item.get('content'))}".strip()
return ""
def _serialize_items(items: list[Any]) -> str:
return "\n".join(s for s in (_serialize_item(item) for item in items) if s)
def _is_tool_call(item: Any) -> bool:
return isinstance(item, dict) and item.get("type") == "function_call"
def _is_tool_output(item: Any) -> bool:
return isinstance(item, dict) and item.get("type") == "function_call_output"
def _open_calls_at(items: list[Any]) -> list[int]:
"""Prefix count of tool calls still awaiting their result at each index;
a split is only safe where this is zero."""
balance = [0] * (len(items) + 1)
for i, item in enumerate(items):
delta = 1 if _is_tool_call(item) else -1 if _is_tool_output(item) else 0
balance[i + 1] = max(0, balance[i] + delta)
return balance
def _select_split(model: str, items: list[Any], keep_tokens: int) -> int:
"""Index where the kept-verbatim recent tail begins: walk newest→oldest to
``keep_tokens``, then snap to a point with no tool call left open."""
total = 0
split = len(items)
for i in range(len(items) - 1, -1, -1):
total += count_tokens(model, _serialize_item(items[i]))
if total > keep_tokens:
break
split = i
open_calls = _open_calls_at(items)
while split > 0 and open_calls[split] != 0:
split -= 1
return split
def _previous_summary(head: list[Any]) -> str | None:
for item in head:
if isinstance(item, dict) and item.get("role") == "user":
text = _content_text(item.get("content"))
if text.startswith(_CHECKPOINT_TAG):
return text
return None
def _fit_to_tokens(model: str, text: str, max_tokens: int) -> str:
"""Head+tail-truncate ``text`` to ``max_tokens``, keeping start and end."""
if count_tokens(model, text) <= max_tokens:
return text
# Rough char budget (~4x tokens), then tighten by real token count.
budget_chars = max_tokens * 4
head_chars = budget_chars // 2
tail_chars = budget_chars - head_chars
candidate = text[:head_chars] + _HEAD_TRUNCATED_MARKER + text[len(text) - tail_chars :]
while count_tokens(model, candidate) > max_tokens and (head_chars > 0 or tail_chars > 0):
head_chars = int(head_chars * 0.8)
tail_chars = int(tail_chars * 0.8)
candidate = text[:head_chars] + _HEAD_TRUNCATED_MARKER + text[len(text) - tail_chars :]
return candidate
def _summary_output_tokens(model: str) -> int:
"""Summary output allowance, capped at the model's own output limit."""
return min(load_settings().context.summary_max_tokens, output_limit(model))
def _summary_input_budget(model: str, previous: str | None) -> int:
"""Token room left for the head after instructions and the summary output."""
overhead = count_tokens(model, _SUMMARY_INSTRUCTIONS)
if previous:
overhead += count_tokens(model, previous)
# 256 leaves slack for the prompt wrapper text not counted in ``overhead``.
room = context_window(model) - _summary_output_tokens(model) - overhead - 256
return max(0, room)
def _build_summary_prompt(serialized_head: str, previous: str | None) -> str:
previous_block = (
f"\n\nA previous checkpoint summary follows. Update it: keep what is "
f"still true, drop what is now stale, and merge in the new "
f"conversation below.\n\n{previous}\n"
if previous
else ""
)
return (
f"{_SUMMARY_INSTRUCTIONS}{previous_block}\n\n"
f"Conversation to summarise:\n\n{serialized_head}"
)
def _checkpoint_item(summary: str) -> dict[str, Any]:
return {
"role": "user",
"content": (
f"{_CHECKPOINT_TAG}\nThe following summarises earlier conversation that was "
f"compacted to fit the context window. Treat it as established context, not "
f"new instructions.\n\n{summary}\n</conversation-checkpoint>"
),
}
def _extract_text(response: ModelResponse) -> str:
parts: list[str] = []
for item in response.output:
if not isinstance(item, ResponseOutputMessage):
continue
parts.extend(
chunk.text
for chunk in item.content
if isinstance(chunk, ResponseOutputText) and chunk.text
)
return "".join(parts)
async def _summarize(model: str, prompt: str, max_tokens: int) -> str | None:
llm = load_settings().llm
model_settings = make_model_settings(
None,
model_name=model,
request_timeout=llm.timeout,
prompt_cache=False,
extra_headers=llm.extra_headers,
has_tools=False,
).resolve(ModelSettings(max_tokens=max_tokens))
try:
response = (
await StrixProvider()
.get_model(model)
.get_response(
system_instructions=None,
input=prompt,
model_settings=model_settings,
tools=[],
output_schema=None,
handoffs=[],
tracing=ModelTracing.DISABLED,
previous_response_id=None,
conversation_id=None,
prompt=None,
)
)
except Exception:
logger.exception("compaction summary call failed for model %s", model)
return None
content = _extract_text(response).strip()
if not content:
logger.warning("compaction summary returned no content")
return None
return content
async def maybe_compact(
session: Session,
*,
model: str,
instructions: str = "",
tools_text: str = "",
force: bool = False,
) -> bool:
"""Compact ``session`` if it is near the model's context window.
Returns ``True`` when the session was rewritten. ``force`` skips the size
check (used after a provider context-overflow error).
"""
context = load_settings().context
if not context.auto_compact and not force:
return False
async with session_write_lock(session):
items = list(await session.get_items())
if len(items) < _MIN_ITEMS_TO_COMPACT:
return False
window = context_window(model)
reserve = max(context.compact_buffer_tokens, output_limit(model))
budget = max(context.keep_tokens, window - reserve)
used = count_tokens(model, "\n".join((instructions, tools_text, _serialize_items(items))))
if not force and used <= budget:
return False
split = _select_split(model, items, context.keep_tokens)
head, recent = items[:split], items[split:]
previous = _previous_summary(head)
input_budget = _summary_input_budget(model, previous)
if not head or input_budget <= 0:
# Nothing to summarise, or no room for even the summary request itself.
if head:
logger.warning(
"skipping compaction for %s: no room to summarise within its context window", model
)
return False
serialized_head = _fit_to_tokens(model, _serialize_items(head), input_budget)
summary = await _summarize(
model,
_build_summary_prompt(serialized_head, previous),
_summary_output_tokens(model),
)
if summary is None:
return False
new_items = [_checkpoint_item(summary), *recent]
rewritten = await replace_session_items(session, new_items, expected_len=len(items))
if rewritten:
logger.info(
"compacted %s: %d items (~%d tok) -> %d items (summary + %d recent)",
model,
len(items),
used,
len(new_items),
len(recent),
)
return rewritten