Files
hermes-agent/agent/turn_preflight.py

394 lines
21 KiB
Python

"""Pre-API preflight compression gate and post-tool compression for the turn loop.
Runs once per API call after the request pressure is measured: grow a managed llama.cpp
window (last resort), or compress when over threshold (deferring on noisy estimates, in
failure cooldown, or while the review fork's first request is pending), handle the
provider-proven overflow re-check fail-closed, and emit the deduped blocked/uncompressed
overflow warnings. Nothing here imports ``agent.conversation_loop`` at module level
(cycle); loop-internal helpers resolve lazily."""
from __future__ import annotations
import logging
from dataclasses import dataclass
from typing import Any, Dict, List, Optional
from agent.context_engine import automatic_compaction_status_message
from agent.conversation_compression import (
PRE_API_COMPRESSION_STATUS_TEMPLATE, _reset_read_dedup_caches, compression_blocked_transiently,
compression_skipped_due_to_lock, context_compression_timed_out,
conversation_history_after_compression, ensure_compression_feasibility_checked,
)
from agent.turn_context import _review_fork_first_request_pending
from agent.turn_context_compaction import (
_apply_grown_window, _blocked_compress_reason, _clear_overflow_warn, _refund_api_call,
_reanchor, _reset_retry_state_after_compaction,
)
logger = logging.getLogger("agent.conversation_loop")
@dataclass
class PreflightGateVerdict:
"""``action``: ``"fallthrough"`` (make the API call), ``"continue"`` (window grown or
history compacted — the call/budget was refunded, re-enter the turn loop and
re-measure), ``"break"`` (turn ends: compression timeout or non-actionable compaction
handoff — ``final_response``/``failed``/``_turn_exit_reason`` set) or ``"return"``
(``result`` is the turn's result dict). The other fields are the loop locals the gate
rebinds; ``_last_preflight_pressure`` is re-armed only by a compression pass."""
action: str
pending_moa_prepared_request: Any
messages: Any
active_system_prompt: Any
conversation_history: Any
api_call_count: Any
compression_attempts: Any
final_response: Any
failed: Any
_turn_exit_reason: Any
_compression_timeout_exhausted: Any
_preflight_compression_blocked: Any
_provider_overflow_recovery_pending: Any
_last_preflight_pressure: Any
result: Optional[Dict[str, Any]] = None
def run_preflight_compression(
agent: Any, v: PreflightGateVerdict, *, compressor: Any, request_pressure_tokens: int,
provider_overflow_preflight: bool, defer_preflight: Any, moa_prepared_request: Any,
system_message: Any, user_message: Any, max_compression_attempts: int, effective_task_id: Any,
) -> PreflightGateVerdict:
"""Mirror of the turn-prologue guard chain (defer on noisy estimate → skip in failure
cooldown → ``should_compress``), rebinding the loop locals on ``v`` and setting
``v.action``. A compression pass that never reaches the provider refunds the
call/budget in every branch (skip, re-run, timeout) so ``api_call_count`` never
over-reports; a lock/transient skip refunds the attempt and leaves the progress
blocker unarmed. A forced provider-overflow preflight that any gate blocks fails
closed (llama.cpp may silently truncate)."""
from agent.conversation_loop import (
_COMPRESSION_TIMEOUT_FINAL_RESPONSE, _HANDOFF_SKIP_FINAL_RESPONSE,
_compression_deferred_result, _maybe_grow_local_window, _provider_overflow_exhausted_result,
_should_skip_model_call_for_reference_handoff,
)
def _done(action: str, result: Optional[Dict[str, Any]] = None) -> PreflightGateVerdict:
v.action, v.result = action, result
return v
def _exhausted_result() -> Dict[str, Any]:
return _provider_overflow_exhausted_result(
agent, v.messages, v.conversation_history, v.api_call_count, request_pressure_tokens,
max_compression_attempts,
)
_compression_cooldown = getattr(
compressor, "get_active_compression_failure_cooldown", lambda: None
)()
_eligible = (
agent.compression_enabled
and len(v.messages) > 1
and v.compression_attempts < max_compression_attempts
)
if _eligible:
# Aux clamp must land before the first compaction fires on the main-window threshold (#114707).
ensure_compression_feasibility_checked(agent, request_pressure_tokens)
if (
_eligible
and not _review_fork_first_request_pending(agent)
and (not v._preflight_compression_blocked or provider_overflow_preflight)
and (not defer_preflight(request_pressure_tokens) or provider_overflow_preflight)
and not _compression_cooldown
and compressor.should_compress(request_pressure_tokens)
):
# Managed local runtime: grow the context window before compressing (last
# resort). Only for a llamacpp provider at the supervised base_url.
_grown_window = _maybe_grow_local_window(agent, compressor, request_pressure_tokens)
if _grown_window:
# Bigger window granted: recalibrate and skip compression this pass. Never
# reached the provider — refund like the compression path does.
_apply_grown_window(agent, compressor, _grown_window)
v.api_call_count = _refund_api_call(agent, v.api_call_count)
return _done("continue")
if moa_prepared_request is not None:
v.pending_moa_prepared_request = moa_prepared_request
v.compression_attempts += 1
# Compression is running: reset the blocked-overflow warning dedup so a
# later blocked turn warns again.
_clear_overflow_warn(agent)
_threshold = int(getattr(compressor, "threshold_tokens", 0) or 0)
_context_length = int(getattr(compressor, "context_length", 0) or 0)
logger.info(
"Pre-API compression: ~%s request tokens >= %s threshold "
"(context=%s, attempt=%s/%s)",
f"{request_pressure_tokens:,}",
f"{_threshold:,}",
f"{_context_length:,}" if getattr(compressor, "context_length", 0) else "unknown",
v.compression_attempts,
max_compression_attempts,
)
_pre_api_status = automatic_compaction_status_message(
compressor,
phase="pre_api",
default_message=PRE_API_COMPRESSION_STATUS_TEMPLATE.format(
tokens=request_pressure_tokens
),
approx_tokens=request_pressure_tokens,
threshold_tokens=_threshold,
context_length=_context_length,
model=agent.model,
attempt=v.compression_attempts,
max_attempts=max_compression_attempts,
)
if _pre_api_status:
agent._emit_status(_pre_api_status)
v._last_preflight_pressure = request_pressure_tokens
_pre_api_input = v.messages
v.messages, v.active_system_prompt = agent._compress_context(
v.messages, system_message, approx_tokens=request_pressure_tokens,
task_id=effective_task_id,
)
if context_compression_timed_out(agent):
# Progress-aware timeout: never reached the provider — refund the
# call/budget and stop; an overflow retry would only re-compress.
v.api_call_count = _refund_api_call(agent, v.api_call_count)
v.final_response = _COMPRESSION_TIMEOUT_FINAL_RESPONSE
v.failed = True
v._compression_timeout_exhausted = True
v._turn_exit_reason = "context_compression_timeout"
return _done("break")
if v.messages is _pre_api_input and (
compression_skipped_due_to_lock(agent) or compression_blocked_transiently(agent)
):
# Temporary DEFER (lock held / cooldown), not evidence about compressibility:
# refund the attempt, leave the progress blocker unarmed and proceed.
v.compression_attempts -= 1
v._last_preflight_pressure = None
if v.pending_moa_prepared_request is moa_prepared_request:
v.pending_moa_prepared_request = None
else:
_reset_retry_state_after_compaction(agent)
# Re-baseline the flush cursor: rotation returns None (child flushes
# whole); in-place returns list(messages) — None would re-append
# persisted rows. See conversation_history_after_compression().
v.conversation_history = conversation_history_after_compression(
agent, v.messages, v.conversation_history
)
# Never reaches the provider on skip or re-run — refund the call/budget
# in BOTH cases, else budget leaks and api_call_count over-reports.
v.api_call_count = _refund_api_call(agent, v.api_call_count)
if _should_skip_model_call_for_reference_handoff(v.messages, user_message):
# Reference-only handoff must not become the active turn after a
# completed assistant response.
logger.info(
"Skipping post-compaction model call: reference-only "
"handoff would be the sole active user turn (#80622)"
)
if not v.final_response:
v.final_response = _HANDOFF_SKIP_FINAL_RESPONSE
v._turn_exit_reason = "compaction_handoff_not_actionable"
return _done("break")
return _done("continue")
elif provider_overflow_preflight and _compression_cooldown:
# Provider proved the request cannot fit and the compressor is unavailable:
# don't resend; let the next user turn retry after cooldown.
agent._persist_session(v.messages, v.conversation_history)
return _done("return", _compression_deferred_result(
agent, v.messages, v.api_call_count, reason="transient_block"
))
elif provider_overflow_preflight and v.compression_attempts >= max_compression_attempts:
# All recovery passes consumed and still over threshold: fail closed —
# llama.cpp may silently truncate an oversized retry.
return _done("return", _exhausted_result())
elif _eligible and not defer_preflight(request_pressure_tokens) and _compression_cooldown:
# Summary-LLM cooldown blocks compression: deduped warning only when over
# threshold (should_compress_info reason is None below it).
_block_reason = _blocked_compress_reason(compressor, request_pressure_tokens)
if _block_reason:
agent._warn_context_overflow_blocked(
_block_reason, request_pressure_tokens,
int(getattr(compressor, "threshold_tokens", 0) or 0),
)
elif not agent.compression_enabled and len(v.messages) > 1:
# Uncompressed session guard: compression is disabled, so warn (deduped) when
# the request exceeds the context window; the turn-context preflight re-arms.
_ctx_len = getattr(getattr(agent, "context_compressor", None), "context_length", None)
if isinstance(_ctx_len, int) and _ctx_len > 0 and request_pressure_tokens > _ctx_len:
_warn_fn = getattr(agent, "_warn_uncompressed_context_overflow", None)
if callable(_warn_fn):
_warn_fn(request_pressure_tokens, _ctx_len)
if provider_overflow_preflight:
# Any other gate blocking the forced preflight (e.g. uncompressible one-
# message request) must fail closed: the request is proven not to fit.
return _done("return", _exhausted_result())
return _done("fallthrough")
@dataclass
class PostToolCompressionVerdict:
"""``end_turn`` True → a reference-only compaction handoff would be the sole active
user turn: stop without another model call (``final_response`` /
``turn_exit_reason`` set)."""
end_turn: bool
messages: List[Dict[str, Any]]
active_system_prompt: Any
conversation_history: Any
compression_attempts: int
final_response: Any
turn_exit_reason: Any
current_turn_user_idx: int
def compress_after_tool_results(
agent: Any, *, messages: List[Dict[str, Any]], system_message: Any, user_message: Any,
active_system_prompt: Any, conversation_history: Any, compression_attempts: int,
max_compression_attempts: int, effective_task_id: Any, final_response: Any,
turn_exit_reason: Any, current_turn_user_idx: int,
) -> PostToolCompressionVerdict:
"""Post-tool-call compression decision. Pressure comes from API-reported
``prompt_tokens`` (a tight lower bound; thinking models inflate completion tokens),
``0`` right after compression (no real count yet), else the route-aware
overhead-inclusive estimate. Over threshold but blocked → deduped warning plus the
deterministic tool-result-only prune, committed only when the engine returns a NEW
list (never rebuild ``conversation_history`` for it)."""
from agent.conversation_loop import (
_HANDOFF_SKIP_FINAL_RESPONSE, _midturn_request_pressure_tokens,
_should_skip_model_call_for_reference_handoff,
)
from agent.model_metadata import estimate_request_tokens_rough
def _verdict(end_turn: bool) -> PostToolCompressionVerdict:
return PostToolCompressionVerdict(
end_turn=end_turn, messages=messages, active_system_prompt=active_system_prompt,
conversation_history=conversation_history, compression_attempts=compression_attempts,
final_response=final_response, turn_exit_reason=turn_exit_reason,
current_turn_user_idx=current_turn_user_idx,
)
_compressor = agent.context_compressor
# A new checkpoint must reach the provider before stale usage can trigger
# local compression, overflow warnings, or destructive tool-result pruning.
if bool(getattr(_compressor, "awaiting_real_usage_after_compression", False)):
return _verdict(False)
# Real usage decides: the anchor is the provider's last prompt count plus a rough delta for
# ONLY the tool results appended since (the raw last_prompt_tokens ignores them). Right after
# a compaction (-1 sentinel) there is no real count yet: never treat the schema-heavy rough
# figure as pressure. The whole-request rough estimate is the last resort (usage-less
# provider, post-disconnect, gateway restart), kept route-aware (#96995/#97602).
from agent.usage_anchor import anchored_context_tokens
_anchored = anchored_context_tokens(messages, getattr(agent, "_usage_anchor", None))
if _anchored is not None:
_real_tokens = _anchored
elif _compressor.last_prompt_tokens > 0:
# Only prompt_tokens: thinking models inflate completion_tokens with
# reasoning that uses no context → premature compression. (#12026)
_real_tokens = _compressor.last_prompt_tokens
elif _compressor.last_prompt_tokens == -1:
_real_tokens = 0
else:
_real_tokens = _midturn_request_pressure_tokens(
agent, messages, active_system_prompt or "",
estimate_request_tokens_rough(messages, tools=agent.tools or None),
)
if agent.compression_enabled and compression_attempts < max_compression_attempts:
ensure_compression_feasibility_checked(agent, _real_tokens)
if (
agent.compression_enabled
and compression_attempts < max_compression_attempts
and not bool(
getattr(_compressor, "awaiting_real_usage_after_compression", False)
)
and _compressor.should_compress(_real_tokens)
):
compression_attempts += 1
# Compression is running: reset blocked-overflow warning dedup so a
# future blocked turn can warn again.
_clear_overflow_warn(agent)
agent._safe_print(" ⟳ compacting context…")
_post_tool_input = messages
# Pass overhead-aware _real_tokens, not last_prompt_tokens (0 in the
# no-usage fallback), so the overflow guard sees the true size.
messages, active_system_prompt = agent._compress_context(
messages, system_message, approx_tokens=_real_tokens, task_id=effective_task_id
)
if messages is _post_tool_input and compression_skipped_due_to_lock(agent):
# Lock-skip no-op is a temporary defer, not evidence about compressibility:
# refund so a lock-loser loop doesn't burn the budget toward exhausted.
# #69870 lock-skip / #97488 transient-block: this pass no-oped for a TEMPORARY reason (another
# path holds the compression lock, or a timed cooldown/backoff guard is active). That is a
# temporary DEFER, not evidence about compressibility — refund the attempt (it must not burn the
# shared overflow-recovery budget toward compression_exhausted → gateway auto-reset,
# #9893/#35809) and leave the insufficient-progress blocker unarmed. Proceed with the current
# request: if it truly does not fit, the provider's 413/overflow handler returns the soft
# compression_deferred result with that stronger signal.
# #69870 lock-skip: the provider proved the request does not fit, but this compression pass
# no-oped only because another path holds the session's compression lock. Temporary defer, not
# exhaustion — refund the attempt and end the turn softly so the gateway does NOT auto-reset the
# session (#9893/#35809).
# #97488 transient-block: compression no-oped because a timed guard (host-timeout cooldown /
# structural backoff) is active — a temporary defer, not evidence of incompressibility. Never
# classify it as compression_exhausted (gateway auto-reset).
# bypass_cooldown=True, # #100661 provider-proven overflow
# #97488: timed transient guard — defer, never exhaustion (gateway auto-reset).
# #69870 lock-skip: the provider proved the request does not fit, but this compression pass
# no-oped only because another path holds the session's compression lock. Temporary defer, not
# exhaustion — refund the attempt and end the turn softly so the gateway does NOT auto-reset the
# session (#9893/#35809).
# #97488 transient-block: a timed guard (host-timeout cooldown / structural backoff) no-oped
# this pass — defer softly, never compression_exhausted (which would auto-reset the session).
# #69870 lock-skip: this pass no-oped because another path holds the session's compression lock
# — a temporary defer, not evidence about compressibility.
compression_attempts -= 1
else:
conversation_history = conversation_history_after_compression(
agent, messages, conversation_history
)
if _should_skip_model_call_for_reference_handoff(messages, user_message):
logger.info(
"Skipping post-tool compaction model call: "
"reference-only handoff would be the sole "
"active user turn (#80622)"
)
if not final_response:
final_response = _HANDOFF_SKIP_FINAL_RESPONSE
turn_exit_reason = "compaction_handoff_not_actionable"
return _verdict(True)
current_turn_user_idx = _reanchor(agent, messages, user_message)
elif agent.compression_enabled:
# Over threshold but compression blocked (cooldown/anti-thrash): deduped
# warning so context can't silently overflow. ``attempts_spent`` names the
# attempts_exhausted lockout when the engine says RUN but the per-turn
# budget is spent (#101889).
_block_reason = _blocked_compress_reason(
_compressor, _real_tokens, attempts_spent=compression_attempts
)
if _block_reason:
agent._warn_context_overflow_blocked(
_block_reason, _real_tokens, int(getattr(_compressor, "threshold_tokens", 0) or 0)
)
# Proactive tool-result prune (deterministic, no LLM, keeps tail): no-op unless
# proactive_prune_tokens is exceeded; commits only past
# proactive_prune_min_reclaim_tokens so cache breaks stay episodic.
_prune = getattr(_compressor, "prune_tool_results_only", None)
if callable(_prune):
try:
_pruned_msgs, _pruned_n = _prune(messages, current_tokens=_real_tokens)
except Exception:
logger.debug("proactive tool-result prune failed; skipping", exc_info=True)
_pruned_msgs, _pruned_n = messages, 0
# Standard no-op caller contract: only commit when the engine returned a
# NEW list object with a non-zero count. Do NOT rebuild
# conversation_history: rows already carry _DB_PERSISTED_MARKER, and on a
# stale in-place flag the helper could seed unpersisted rows.
if _pruned_n and _pruned_msgs is not messages:
messages = _pruned_msgs
# A committed prune is a content-loss boundary like compaction: demoted skill_view /
# read_file bodies survive only as one-line markers, so the repeat-read dedup must
# stop answering "unchanged" for them or the reload the marker asks for is refused.
_reset_read_dedup_caches(effective_task_id, session_id=agent.session_id or "")
return _verdict(False)